wip script for lowering dlrm training

.github/workflows/nightly.yml

@@ -51,11 +51,11 @@ jobs:
       run: |
         ./setup_venv.ps1
         $env:SHARK_PACKAGE_VERSION=${{ env.package_version }}
-        pip wheel -v -w dist . --pre -f https://download.pytorch.org/whl/nightly/cpu -f https://llvm.github.io/torch-mlir/package-index/ -f https://nod-ai.github.io/SRT/pip-release-links.html
+        pip wheel -v -w dist . --pre -f https://download.pytorch.org/whl/nightly/cpu -f https://llvm.github.io/torch-mlir/package-index/ -f https://nod-ai.github.io/SHARK-Runtime/pip-release-links.html
         python process_skipfiles.py
         pyinstaller .\apps\stable_diffusion\shark_sd.spec
         mv ./dist/nodai_shark_studio.exe ./dist/nodai_shark_studio_${{ env.package_version_ }}.exe
-        signtool sign /f c:\g\shark_02152023.cer /fd certHash /csp "eToken Base Cryptographic Provider" /k "${{ secrets.CI_CERT }}" ./dist/nodai_shark_studio_${{ env.package_version_ }}.exe
+        signtool sign /f c:\g\shark_02152023.cer /csp "eToken Base Cryptographic Provider" /k "${{ secrets.CI_CERT }}" ./dist/nodai_shark_studio_${{ env.package_version_ }}.exe
   
     - name: Upload Release Assets
       id: upload-release-assets
@@ -104,7 +104,7 @@ jobs:
         echo "DATE=$(date +'%Y-%m-%d')" >> $GITHUB_ENV
         python -m pip install --upgrade pip
         python -m pip install flake8 pytest toml
-        if [ -f requirements.txt ]; then pip install -r requirements.txt -f https://llvm.github.io/torch-mlir/package-index/ -f https://nod-ai.github.io/SRT/pip-release-links.html; fi
+        if [ -f requirements.txt ]; then pip install -r requirements.txt -f https://llvm.github.io/torch-mlir/package-index/ -f https://nod-ai.github.io/SHARK-Runtime/pip-release-links.html; fi
     - name: Lint with flake8
       run: |
         # stop the build if there are Python syntax errors or undefined names
@@ -144,7 +144,7 @@ jobs:
         source shark.venv/bin/activate
         package_version="$(printf '%(%Y%m%d)T.${{ github.run_number }}')"
         SHARK_PACKAGE_VERSION=${package_version} \
-        pip wheel -v -w wheelhouse . --pre -f https://download.pytorch.org/whl/nightly/torch -f https://llvm.github.io/torch-mlir/package-index/ -f https://nod-ai.github.io/SRT/pip-release-links.html
+        pip wheel -v -w wheelhouse . --pre -f https://download.pytorch.org/whl/nightly/torch -f https://llvm.github.io/torch-mlir/package-index/ -f https://nod-ai.github.io/SHARK-Runtime/pip-release-links.html
         # Install the built wheel
         pip install ./wheelhouse/nodai*
         # Validate the Models

.github/workflows/test-models.yml

@@ -137,8 +137,7 @@ jobs:
         source shark.venv/bin/activate
         echo $PATH
         pip list | grep -E "torch|iree"
-      # disabled due to a low-visibility memory issue with pytest on macos.
-      # pytest --ci --ci_sha=${SHORT_SHA} --local_tank_cache="/Volumes/builder/anush/shark_cache" --tank_url="gs://shark_tank/nightly/" -k metal
+        pytest --ci --ci_sha=${SHORT_SHA} --local_tank_cache="/Volumes/builder/anush/shark_cache" --tank_url="gs://shark_tank/nightly/" -k metal
 
     - name: Validate Vulkan Models (a100)
       if: matrix.suite == 'vulkan' && matrix.os == 'a100'

.gitignore

@@ -182,7 +182,7 @@ generated_imgs/
 
 # Custom model related artefacts
 variants.json
-/models/
+models/
 
 # models folder
 apps/stable_diffusion/web/models/
@@ -193,12 +193,3 @@ stencil_annotator/
 # For DocuChat
 apps/language_models/langchain/user_path/
 db_dir_UserData
-
-# Embeded browser cache and other
-apps/stable_diffusion/web/EBWebView/
-
-# Llama2 tokenizer configs
-llama2_tokenizer_configs/
-
-# Webview2 runtime artefacts
-EBWebView/

.gitmodules

@@ -1,4 +1,4 @@
 [submodule "inference/thirdparty/shark-runtime"]
 	path = inference/thirdparty/shark-runtime
-	url =https://github.com/nod-ai/SRT.git
+	url =https://github.com/nod-ai/SHARK-Runtime.git
 	branch = shark-06032022

README.md

@@ -10,7 +10,7 @@ High Performance Machine Learning Distribution
   <summary>Prerequisites - Drivers </summary>
   
 #### Install your Windows hardware drivers
-* [AMD RDNA Users] Download the latest driver (23.2.1 is the oldest supported) [here](https://www.amd.com/en/support).
+* [AMD RDNA Users] Download the latest driver [here](https://www.amd.com/en/support/kb/release-notes/rn-rad-win-23-2-1).
 * [macOS Users] Download and install the 1.3.216 Vulkan SDK from [here](https://sdk.lunarg.com/sdk/download/1.3.216.0/mac/vulkansdk-macos-1.3.216.0.dmg). Newer versions of the SDK will not work. 
 * [Nvidia Users] Download and install the latest CUDA / Vulkan drivers from [here](https://developer.nvidia.com/cuda-downloads)
   
@@ -170,7 +170,7 @@ python -m pip install --upgrade pip
 This step pip installs SHARK and related packages on Linux Python 3.8, 3.10 and 3.11 and macOS / Windows Python 3.11
 
 ```shell
-pip install nodai-shark -f https://nod-ai.github.io/SHARK/package-index/ -f https://llvm.github.io/torch-mlir/package-index/ -f  https://nod-ai.github.io/SRT/pip-release-links.html --extra-index-url https://download.pytorch.org/whl/nightly/cpu
+pip install nodai-shark -f https://nod-ai.github.io/SHARK/package-index/ -f https://llvm.github.io/torch-mlir/package-index/ -f  https://nod-ai.github.io/SHARK-Runtime/pip-release-links.html --extra-index-url https://download.pytorch.org/whl/nightly/cpu
 ```
 
 ### Run shark tank model tests.
@@ -254,6 +254,7 @@ if you want to instead incorporate this into a python script, you can pass the `
 ```
 shark_module = SharkInference(
         mlir_model,
+        func_name,
         device=args.device,
         mlir_dialect="tm_tensor",
         dispatch_benchmarks="all",
@@ -296,7 +297,7 @@ torch_mlir, func_name = mlir_importer.import_mlir(tracing_required=True)
 # SharkInference accepts mlir in linalg, mhlo, and tosa dialect.
 
 from shark.shark_inference import SharkInference
-shark_module = SharkInference(torch_mlir, device="cpu", mlir_dialect="linalg")
+shark_module = SharkInference(torch_mlir, func_name, device="cpu", mlir_dialect="linalg")
 shark_module.compile()
 result = shark_module.forward((input))
 
@@ -319,17 +320,12 @@ mhlo_ir = r"""builtin.module  {
 
 arg0 = np.ones((1, 4)).astype(np.float32)
 arg1 = np.ones((4, 1)).astype(np.float32)
-shark_module = SharkInference(mhlo_ir, device="cpu", mlir_dialect="mhlo")
+shark_module = SharkInference(mhlo_ir, func_name="forward", device="cpu", mlir_dialect="mhlo")
 shark_module.compile()
 result = shark_module.forward((arg0, arg1))
 ```
 </details>
 
-## Examples Using the REST API
-
-* [Setting up SHARK for use with Blender](./docs/shark_sd_blender.md)
-* [Setting up SHARK for use with Koboldcpp](./docs/shark_sd_koboldcpp.md)
-
 ## Supported and Validated Models
 
 SHARK is maintained to support the latest innovations in ML Models: 

apps/language_models/langchain/expanded_pipelines.py

@@ -1,3 +1,4 @@
+"""Load question answering chains."""
 from __future__ import annotations
 from typing import (
     Any,
@@ -10,34 +11,23 @@ from typing import (
     Union,
     Protocol,
 )
-import inspect
-import json
-import warnings
-from pathlib import Path
-import yaml
-from abc import ABC, abstractmethod
-import langchain
 from langchain.base_language import BaseLanguageModel
 from langchain.callbacks.base import BaseCallbackManager
 from langchain.chains.question_answering import stuff_prompt
 from langchain.prompts.base import BasePromptTemplate
 from langchain.docstore.document import Document
+from abc import ABC, abstractmethod
+from langchain.chains.base import Chain
 from langchain.callbacks.manager import (
     CallbackManager,
     CallbackManagerForChainRun,
     Callbacks,
 )
-from langchain.load.serializable import Serializable
-from langchain.schema import RUN_KEY, BaseMemory, RunInfo
 from langchain.input import get_colored_text
 from langchain.load.dump import dumpd
 from langchain.prompts.prompt import PromptTemplate
 from langchain.schema import LLMResult, PromptValue
-from pydantic import Extra, Field, root_validator, validator
-
-
-def _get_verbosity() -> bool:
-    return langchain.verbose
+from pydantic import Extra, Field, root_validator
 
 
 def format_document(doc: Document, prompt: BasePromptTemplate) -> str:
@@ -58,413 +48,6 @@ def format_document(doc: Document, prompt: BasePromptTemplate) -> str:
     return prompt.format(**document_info)
 
 
-class Chain(Serializable, ABC):
-    """Base interface that all chains should implement."""
-
-    memory: Optional[BaseMemory] = None
-    callbacks: Callbacks = Field(default=None, exclude=True)
-    callback_manager: Optional[BaseCallbackManager] = Field(
-        default=None, exclude=True
-    )
-    verbose: bool = Field(
-        default_factory=_get_verbosity
-    )  # Whether to print the response text
-    tags: Optional[List[str]] = None
-
-    class Config:
-        """Configuration for this pydantic object."""
-
-        arbitrary_types_allowed = True
-
-    @property
-    def _chain_type(self) -> str:
-        raise NotImplementedError("Saving not supported for this chain type.")
-
-    @root_validator()
-    def raise_deprecation(cls, values: Dict) -> Dict:
-        """Raise deprecation warning if callback_manager is used."""
-        if values.get("callback_manager") is not None:
-            warnings.warn(
-                "callback_manager is deprecated. Please use callbacks instead.",
-                DeprecationWarning,
-            )
-            values["callbacks"] = values.pop("callback_manager", None)
-        return values
-
-    @validator("verbose", pre=True, always=True)
-    def set_verbose(cls, verbose: Optional[bool]) -> bool:
-        """If verbose is None, set it.
-
-        This allows users to pass in None as verbose to access the global setting.
-        """
-        if verbose is None:
-            return _get_verbosity()
-        else:
-            return verbose
-
-    @property
-    @abstractmethod
-    def input_keys(self) -> List[str]:
-        """Input keys this chain expects."""
-
-    @property
-    @abstractmethod
-    def output_keys(self) -> List[str]:
-        """Output keys this chain expects."""
-
-    def _validate_inputs(self, inputs: Dict[str, Any]) -> None:
-        """Check that all inputs are present."""
-        missing_keys = set(self.input_keys).difference(inputs)
-        if missing_keys:
-            raise ValueError(f"Missing some input keys: {missing_keys}")
-
-    def _validate_outputs(self, outputs: Dict[str, Any]) -> None:
-        missing_keys = set(self.output_keys).difference(outputs)
-        if missing_keys:
-            raise ValueError(f"Missing some output keys: {missing_keys}")
-
-    @abstractmethod
-    def _call(
-        self,
-        inputs: Dict[str, Any],
-        run_manager: Optional[CallbackManagerForChainRun] = None,
-    ) -> Dict[str, Any]:
-        """Run the logic of this chain and return the output."""
-
-    def __call__(
-        self,
-        inputs: Union[Dict[str, Any], Any],
-        return_only_outputs: bool = False,
-        callbacks: Callbacks = None,
-        *,
-        tags: Optional[List[str]] = None,
-        include_run_info: bool = False,
-    ) -> Dict[str, Any]:
-        """Run the logic of this chain and add to output if desired.
-
-        Args:
-            inputs: Dictionary of inputs, or single input if chain expects
-                only one param.
-            return_only_outputs: boolean for whether to return only outputs in the
-                response. If True, only new keys generated by this chain will be
-                returned. If False, both input keys and new keys generated by this
-                chain will be returned. Defaults to False.
-            callbacks: Callbacks to use for this chain run. If not provided, will
-                use the callbacks provided to the chain.
-            include_run_info: Whether to include run info in the response. Defaults
-                to False.
-        """
-        input_docs = inputs["input_documents"]
-        missing_keys = set(self.input_keys).difference(inputs)
-        if missing_keys:
-            raise ValueError(f"Missing some input keys: {missing_keys}")
-
-        callback_manager = CallbackManager.configure(
-            callbacks, self.callbacks, self.verbose, tags, self.tags
-        )
-        run_manager = callback_manager.on_chain_start(
-            dumpd(self),
-            inputs,
-        )
-
-        if "is_first" in inputs.keys() and not inputs["is_first"]:
-            run_manager_ = run_manager
-            input_list = [inputs]
-            stop = None
-            prompts = []
-            for inputs in input_list:
-                selected_inputs = {
-                    k: inputs[k] for k in self.prompt.input_variables
-                }
-                prompt = self.prompt.format_prompt(**selected_inputs)
-                _colored_text = get_colored_text(prompt.to_string(), "green")
-                _text = "Prompt after formatting:\n" + _colored_text
-                if run_manager_:
-                    run_manager_.on_text(_text, end="\n", verbose=self.verbose)
-                if "stop" in inputs and inputs["stop"] != stop:
-                    raise ValueError(
-                        "If `stop` is present in any inputs, should be present in all."
-                    )
-                prompts.append(prompt)
-
-            prompt_strings = [p.to_string() for p in prompts]
-            prompts = prompt_strings
-            callbacks = run_manager_.get_child() if run_manager_ else None
-            tags = None
-
-            """Run the LLM on the given prompt and input."""
-            # If string is passed in directly no errors will be raised but outputs will
-            # not make sense.
-            if not isinstance(prompts, list):
-                raise ValueError(
-                    "Argument 'prompts' is expected to be of type List[str], received"
-                    f" argument of type {type(prompts)}."
-                )
-            params = self.llm.dict()
-            params["stop"] = stop
-            options = {"stop": stop}
-            disregard_cache = self.llm.cache is not None and not self.llm.cache
-            callback_manager = CallbackManager.configure(
-                callbacks,
-                self.llm.callbacks,
-                self.llm.verbose,
-                tags,
-                self.llm.tags,
-            )
-            if langchain.llm_cache is None or disregard_cache:
-                # This happens when langchain.cache is None, but self.cache is True
-                if self.llm.cache is not None and self.cache:
-                    raise ValueError(
-                        "Asked to cache, but no cache found at `langchain.cache`."
-                    )
-                run_manager_ = callback_manager.on_llm_start(
-                    dumpd(self),
-                    prompts,
-                    invocation_params=params,
-                    options=options,
-                )
-
-                generations = []
-                for prompt in prompts:
-                    inputs_ = prompt
-                    num_workers = None
-                    batch_size = None
-
-                    if num_workers is None:
-                        if self.llm.pipeline._num_workers is None:
-                            num_workers = 0
-                        else:
-                            num_workers = self.llm.pipeline._num_workers
-                    if batch_size is None:
-                        if self.llm.pipeline._batch_size is None:
-                            batch_size = 1
-                        else:
-                            batch_size = self.llm.pipeline._batch_size
-
-                    preprocess_params = {}
-                    generate_kwargs = {}
-                    preprocess_params.update(generate_kwargs)
-                    forward_params = generate_kwargs
-                    postprocess_params = {}
-                    # Fuse __init__ params and __call__ params without modifying the __init__ ones.
-                    preprocess_params = {
-                        **self.llm.pipeline._preprocess_params,
-                        **preprocess_params,
-                    }
-                    forward_params = {
-                        **self.llm.pipeline._forward_params,
-                        **forward_params,
-                    }
-                    postprocess_params = {
-                        **self.llm.pipeline._postprocess_params,
-                        **postprocess_params,
-                    }
-
-                    self.llm.pipeline.call_count += 1
-                    if (
-                        self.llm.pipeline.call_count > 10
-                        and self.llm.pipeline.framework == "pt"
-                        and self.llm.pipeline.device.type == "cuda"
-                    ):
-                        warnings.warn(
-                            "You seem to be using the pipelines sequentially on GPU. In order to maximize efficiency please use a"
-                            " dataset",
-                            UserWarning,
-                        )
-
-                    model_inputs = self.llm.pipeline.preprocess(
-                        inputs_, **preprocess_params
-                    )
-                    model_outputs = self.llm.pipeline.forward(
-                        model_inputs, **forward_params
-                    )
-                    model_outputs["process"] = False
-                    return model_outputs
-                output = LLMResult(generations=generations)
-                run_manager_.on_llm_end(output)
-                if run_manager_:
-                    output.run = RunInfo(run_id=run_manager_.run_id)
-                response = output
-
-            outputs = [
-                # Get the text of the top generated string.
-                {self.output_key: generation[0].text}
-                for generation in response.generations
-            ][0]
-            run_manager.on_chain_end(outputs)
-            final_outputs: Dict[str, Any] = self.prep_outputs(
-                inputs, outputs, return_only_outputs
-            )
-            if include_run_info:
-                final_outputs[RUN_KEY] = RunInfo(run_id=run_manager.run_id)
-            return final_outputs
-        else:
-            _run_manager = (
-                run_manager or CallbackManagerForChainRun.get_noop_manager()
-            )
-            docs = inputs[self.input_key]
-            # Other keys are assumed to be needed for LLM prediction
-            other_keys = {
-                k: v for k, v in inputs.items() if k != self.input_key
-            }
-            doc_strings = [
-                format_document(doc, self.document_prompt) for doc in docs
-            ]
-            # Join the documents together to put them in the prompt.
-            inputs = {
-                k: v
-                for k, v in other_keys.items()
-                if k in self.llm_chain.prompt.input_variables
-            }
-            inputs[self.document_variable_name] = self.document_separator.join(
-                doc_strings
-            )
-            inputs["is_first"] = False
-            inputs["input_documents"] = input_docs
-
-            # Call predict on the LLM.
-            output = self.llm_chain(inputs, callbacks=_run_manager.get_child())
-            if "process" in output.keys() and not output["process"]:
-                return output
-            output = output[self.llm_chain.output_key]
-            extra_return_dict = {}
-        extra_return_dict[self.output_key] = output
-        outputs = extra_return_dict
-        run_manager.on_chain_end(outputs)
-        final_outputs: Dict[str, Any] = self.prep_outputs(
-            inputs, outputs, return_only_outputs
-        )
-        if include_run_info:
-            final_outputs[RUN_KEY] = RunInfo(run_id=run_manager.run_id)
-        return final_outputs
-
-    def prep_outputs(
-        self,
-        inputs: Dict[str, str],
-        outputs: Dict[str, str],
-        return_only_outputs: bool = False,
-    ) -> Dict[str, str]:
-        """Validate and prep outputs."""
-        self._validate_outputs(outputs)
-        if self.memory is not None:
-            self.memory.save_context(inputs, outputs)
-        if return_only_outputs:
-            return outputs
-        else:
-            return {**inputs, **outputs}
-
-    def prep_inputs(
-        self, inputs: Union[Dict[str, Any], Any]
-    ) -> Dict[str, str]:
-        """Validate and prep inputs."""
-        if not isinstance(inputs, dict):
-            _input_keys = set(self.input_keys)
-            if self.memory is not None:
-                # If there are multiple input keys, but some get set by memory so that
-                # only one is not set, we can still figure out which key it is.
-                _input_keys = _input_keys.difference(
-                    self.memory.memory_variables
-                )
-            if len(_input_keys) != 1:
-                raise ValueError(
-                    f"A single string input was passed in, but this chain expects "
-                    f"multiple inputs ({_input_keys}). When a chain expects "
-                    f"multiple inputs, please call it by passing in a dictionary, "
-                    "eg `chain({'foo': 1, 'bar': 2})`"
-                )
-            inputs = {list(_input_keys)[0]: inputs}
-        if self.memory is not None:
-            external_context = self.memory.load_memory_variables(inputs)
-            inputs = dict(inputs, **external_context)
-        self._validate_inputs(inputs)
-        return inputs
-
-    def apply(
-        self, input_list: List[Dict[str, Any]], callbacks: Callbacks = None
-    ) -> List[Dict[str, str]]:
-        """Call the chain on all inputs in the list."""
-        return [self(inputs, callbacks=callbacks) for inputs in input_list]
-
-    def run(
-        self,
-        *args: Any,
-        callbacks: Callbacks = None,
-        tags: Optional[List[str]] = None,
-        **kwargs: Any,
-    ) -> str:
-        """Run the chain as text in, text out or multiple variables, text out."""
-        if len(self.output_keys) != 1:
-            raise ValueError(
-                f"`run` not supported when there is not exactly "
-                f"one output key. Got {self.output_keys}."
-            )
-
-        if args and not kwargs:
-            if len(args) != 1:
-                raise ValueError(
-                    "`run` supports only one positional argument."
-                )
-            return self(args[0], callbacks=callbacks, tags=tags)[
-                self.output_keys[0]
-            ]
-
-        if kwargs and not args:
-            return self(kwargs, callbacks=callbacks, tags=tags)[
-                self.output_keys[0]
-            ]
-
-        if not kwargs and not args:
-            raise ValueError(
-                "`run` supported with either positional arguments or keyword arguments,"
-                " but none were provided."
-            )
-
-        raise ValueError(
-            f"`run` supported with either positional arguments or keyword arguments"
-            f" but not both. Got args: {args} and kwargs: {kwargs}."
-        )
-
-    def dict(self, **kwargs: Any) -> Dict:
-        """Return dictionary representation of chain."""
-        if self.memory is not None:
-            raise ValueError("Saving of memory is not yet supported.")
-        _dict = super().dict()
-        _dict["_type"] = self._chain_type
-        return _dict
-
-    def save(self, file_path: Union[Path, str]) -> None:
-        """Save the chain.
-
-        Args:
-            file_path: Path to file to save the chain to.
-
-        Example:
-        .. code-block:: python
-
-            chain.save(file_path="path/chain.yaml")
-        """
-        # Convert file to Path object.
-        if isinstance(file_path, str):
-            save_path = Path(file_path)
-        else:
-            save_path = file_path
-
-        directory_path = save_path.parent
-        directory_path.mkdir(parents=True, exist_ok=True)
-
-        # Fetch dictionary to save
-        chain_dict = self.dict()
-
-        if save_path.suffix == ".json":
-            with open(file_path, "w") as f:
-                json.dump(chain_dict, f, indent=4)
-        elif save_path.suffix == ".yaml":
-            with open(file_path, "w") as f:
-                yaml.dump(chain_dict, f, default_flow_style=False)
-        else:
-            raise ValueError(f"{save_path} must be json or yaml")
-
-
 class BaseCombineDocumentsChain(Chain, ABC):
     """Base interface for chains combining documents."""
 
@@ -496,6 +79,12 @@ class BaseCombineDocumentsChain(Chain, ABC):
         """
         return None
 
+    @abstractmethod
+    def combine_docs(
+        self, docs: List[Document], **kwargs: Any
+    ) -> Tuple[str, dict]:
+        """Combine documents into a single string."""
+
     def _call(
         self,
         inputs: Dict[str, List[Document]],
@@ -507,49 +96,13 @@ class BaseCombineDocumentsChain(Chain, ABC):
         docs = inputs[self.input_key]
         # Other keys are assumed to be needed for LLM prediction
         other_keys = {k: v for k, v in inputs.items() if k != self.input_key}
-        doc_strings = [
-            format_document(doc, self.document_prompt) for doc in docs
-        ]
-        # Join the documents together to put them in the prompt.
-        inputs = {
-            k: v
-            for k, v in other_keys.items()
-            if k in self.llm_chain.prompt.input_variables
-        }
-        inputs[self.document_variable_name] = self.document_separator.join(
-            doc_strings
+        output, extra_return_dict = self.combine_docs(
+            docs, callbacks=_run_manager.get_child(), **other_keys
         )
-
-        # Call predict on the LLM.
-        output, extra_return_dict = (
-            self.llm_chain(inputs, callbacks=_run_manager.get_child())[
-                self.llm_chain.output_key
-            ],
-            {},
-        )
-
         extra_return_dict[self.output_key] = output
         return extra_return_dict
 
 
-from pydantic import BaseModel
-
-
-class Generation(Serializable):
-    """Output of a single generation."""
-
-    text: str
-    """Generated text output."""
-
-    generation_info: Optional[Dict[str, Any]] = None
-    """Raw generation info response from the provider"""
-    """May include things like reason for finishing (e.g. in OpenAI)"""
-    # TODO: add log probs
-
-
-VALID_TASKS = ("text2text-generation", "text-generation", "summarization")
-
-
 class LLMChain(Chain):
     """Chain to run queries against LLMs.
 
@@ -600,13 +153,21 @@ class LLMChain(Chain):
         inputs: Dict[str, Any],
         run_manager: Optional[CallbackManagerForChainRun] = None,
     ) -> Dict[str, str]:
-        prompts, stop = self.prep_prompts([inputs], run_manager=run_manager)
-        response = self.llm.generate_prompt(
+        response = self.generate([inputs], run_manager=run_manager)
+        return self.create_outputs(response)[0]
+
+    def generate(
+        self,
+        input_list: List[Dict[str, Any]],
+        run_manager: Optional[CallbackManagerForChainRun] = None,
+    ) -> LLMResult:
+        """Generate LLM result from inputs."""
+        prompts, stop = self.prep_prompts(input_list, run_manager=run_manager)
+        return self.llm.generate_prompt(
             prompts,
             stop,
             callbacks=run_manager.get_child() if run_manager else None,
         )
-        return self.create_outputs(response)[0]
 
     def prep_prompts(
         self,
@@ -662,6 +223,23 @@ class LLMChain(Chain):
             for generation in response.generations
         ]
 
+    def predict(self, callbacks: Callbacks = None, **kwargs: Any) -> str:
+        """Format prompt with kwargs and pass to LLM.
+
+        Args:
+            callbacks: Callbacks to pass to LLMChain
+            **kwargs: Keys to pass to prompt template.
+
+        Returns:
+            Completion from LLM.
+
+        Example:
+            .. code-block:: python
+
+                completion = llm.predict(adjective="funny")
+        """
+        return self(kwargs, callbacks=callbacks)[self.output_key]
+
     def predict_and_parse(
         self, callbacks: Callbacks = None, **kwargs: Any
     ) -> Union[str, List[str], Dict[str, Any]]:
@@ -772,6 +350,14 @@ class StuffDocumentsChain(BaseCombineDocumentsChain):
         prompt = self.llm_chain.prompt.format(**inputs)
         return self.llm_chain.llm.get_num_tokens(prompt)
 
+    def combine_docs(
+        self, docs: List[Document], callbacks: Callbacks = None, **kwargs: Any
+    ) -> Tuple[str, dict]:
+        """Stuff all documents into one prompt and pass to LLM."""
+        inputs = self._get_inputs(docs, **kwargs)
+        # Call predict on the LLM.
+        return self.llm_chain.predict(callbacks=callbacks, **inputs), {}
+
     @property
     def _chain_type(self) -> str:
         return "stuff_documents_chain"

apps/language_models/langchain/gen.py

@@ -1129,7 +1129,7 @@ class Langchain:
                 max_time=max_time,
                 num_return_sequences=num_return_sequences,
             )
-            out = run_qa_db(
+            for r in run_qa_db(
                 query=instruction,
                 iinput=iinput,
                 context=context,
@@ -1170,8 +1170,689 @@ class Langchain:
                 auto_reduce_chunks=auto_reduce_chunks,
                 max_chunks=max_chunks,
                 device=self.device,
+            ):
+                (
+                    outr,
+                    extra,
+                ) = r  # doesn't accumulate, new answer every yield, so only save that full answer
+                yield dict(response=outr, sources=extra)
+            if save_dir:
+                extra_dict = gen_hyper_langchain.copy()
+                extra_dict.update(
+                    prompt_type=prompt_type,
+                    inference_server=inference_server,
+                    langchain_mode=langchain_mode,
+                    langchain_action=langchain_action,
+                    document_choice=document_choice,
+                    num_prompt_tokens=num_prompt_tokens,
+                    instruction=instruction,
+                    iinput=iinput,
+                    context=context,
+                )
+                save_generate_output(
+                    prompt=prompt,
+                    output=outr,
+                    base_model=base_model,
+                    save_dir=save_dir,
+                    where_from="run_qa_db",
+                    extra_dict=extra_dict,
+                )
+                if verbose:
+                    print(
+                        "Post-Generate Langchain: %s decoded_output: %s"
+                        % (str(datetime.now()), len(outr) if outr else -1),
+                        flush=True,
+                    )
+            if outr or base_model in non_hf_types:
+                # if got no response (e.g. not showing sources and got no sources,
+                # so nothing to give to LLM), then slip through and ask LLM
+                # Or if llama/gptj, then just return since they had no response and can't go down below code path
+                # clear before return, since .then() never done if from API
+                clear_torch_cache()
+                return
+
+        if inference_server.startswith(
+            "openai"
+        ) or inference_server.startswith("http"):
+            if inference_server.startswith("openai"):
+                import openai
+
+                where_from = "openai_client"
+
+                openai.api_key = os.getenv("OPENAI_API_KEY")
+                stop_sequences = list(
+                    set(prompter.terminate_response + [prompter.PreResponse])
+                )
+                stop_sequences = [x for x in stop_sequences if x]
+                # OpenAI will complain if ask for too many new tokens, takes it as min in some sense, wrongly so.
+                max_new_tokens_openai = min(
+                    max_new_tokens, model_max_length - num_prompt_tokens
+                )
+                gen_server_kwargs = dict(
+                    temperature=temperature if do_sample else 0,
+                    max_tokens=max_new_tokens_openai,
+                    top_p=top_p if do_sample else 1,
+                    frequency_penalty=0,
+                    n=num_return_sequences,
+                    presence_penalty=1.07
+                    - repetition_penalty
+                    + 0.6,  # so good default
+                )
+                if inference_server == "openai":
+                    response = openai.Completion.create(
+                        model=base_model,
+                        prompt=prompt,
+                        **gen_server_kwargs,
+                        stop=stop_sequences,
+                        stream=stream_output,
+                    )
+                    if not stream_output:
+                        text = response["choices"][0]["text"]
+                        yield dict(
+                            response=prompter.get_response(
+                                prompt + text,
+                                prompt=prompt,
+                                sanitize_bot_response=sanitize_bot_response,
+                            ),
+                            sources="",
+                        )
+                    else:
+                        collected_events = []
+                        text = ""
+                        for event in response:
+                            collected_events.append(
+                                event
+                            )  # save the event response
+                            event_text = event["choices"][0][
+                                "text"
+                            ]  # extract the text
+                            text += event_text  # append the text
+                            yield dict(
+                                response=prompter.get_response(
+                                    prompt + text,
+                                    prompt=prompt,
+                                    sanitize_bot_response=sanitize_bot_response,
+                                ),
+                                sources="",
+                            )
+                elif inference_server == "openai_chat":
+                    response = openai.ChatCompletion.create(
+                        model=base_model,
+                        messages=[
+                            {
+                                "role": "system",
+                                "content": "You are a helpful assistant.",
+                            },
+                            {
+                                "role": "user",
+                                "content": prompt,
+                            },
+                        ],
+                        stream=stream_output,
+                        **gen_server_kwargs,
+                    )
+                    if not stream_output:
+                        text = response["choices"][0]["message"]["content"]
+                        yield dict(
+                            response=prompter.get_response(
+                                prompt + text,
+                                prompt=prompt,
+                                sanitize_bot_response=sanitize_bot_response,
+                            ),
+                            sources="",
+                        )
+                    else:
+                        text = ""
+                        for chunk in response:
+                            delta = chunk["choices"][0]["delta"]
+                            if "content" in delta:
+                                text += delta["content"]
+                                yield dict(
+                                    response=prompter.get_response(
+                                        prompt + text,
+                                        prompt=prompt,
+                                        sanitize_bot_response=sanitize_bot_response,
+                                    ),
+                                    sources="",
+                                )
+                else:
+                    raise RuntimeError(
+                        "No such OpenAI mode: %s" % inference_server
+                    )
+            elif inference_server.startswith("http"):
+                inference_server, headers = get_hf_server(inference_server)
+                from gradio_utils.grclient import GradioClient
+                from text_generation import Client as HFClient
+
+                if isinstance(model, GradioClient):
+                    gr_client = model
+                    hf_client = None
+                elif isinstance(model, HFClient):
+                    gr_client = None
+                    hf_client = model
+                else:
+                    (
+                        inference_server,
+                        gr_client,
+                        hf_client,
+                    ) = self.get_client_from_inference_server(
+                        inference_server, base_model=base_model
+                    )
+
+                # quick sanity check to avoid long timeouts, just see if can reach server
+                requests.get(
+                    inference_server,
+                    timeout=int(os.getenv("REQUEST_TIMEOUT_FAST", "10")),
+                )
+
+                if gr_client is not None:
+                    # Note: h2oGPT gradio server could handle input token size issues for prompt,
+                    # but best to handle here so send less data to server
+
+                    chat_client = False
+                    where_from = "gr_client"
+                    client_langchain_mode = "Disabled"
+                    client_langchain_action = LangChainAction.QUERY.value
+                    gen_server_kwargs = dict(
+                        temperature=temperature,
+                        top_p=top_p,
+                        top_k=top_k,
+                        num_beams=num_beams,
+                        max_new_tokens=max_new_tokens,
+                        min_new_tokens=min_new_tokens,
+                        early_stopping=early_stopping,
+                        max_time=max_time,
+                        repetition_penalty=repetition_penalty,
+                        num_return_sequences=num_return_sequences,
+                        do_sample=do_sample,
+                        chat=chat_client,
+                    )
+                    # account for gradio into gradio that handles prompting, avoid duplicating prompter prompt injection
+                    if prompt_type in [
+                        None,
+                        "",
+                        PromptType.plain.name,
+                        PromptType.plain.value,
+                        str(PromptType.plain.value),
+                    ]:
+                        # if our prompt is plain, assume either correct or gradio server knows different prompt type,
+                        # so pass empty prompt_Type
+                        gr_prompt_type = ""
+                        gr_prompt_dict = ""
+                        gr_prompt = prompt  # already prepared prompt
+                        gr_context = ""
+                        gr_iinput = ""
+                    else:
+                        # if already have prompt_type that is not plain, None, or '', then already applied some prompting
+                        #  But assume server can handle prompting, and need to avoid double-up.
+                        #  Also assume server can do better job of using stopping.py to stop early, so avoid local prompting, let server handle
+                        #  So avoid "prompt" and let gradio server reconstruct from prompt_type we passed
+                        # Note it's ok that prompter.get_response() has prompt+text, prompt=prompt passed,
+                        #  because just means extra processing and removal of prompt, but that has no human-bot prompting doesn't matter
+                        #  since those won't appear
+                        gr_context = context
+                        gr_prompt = instruction
+                        gr_iinput = iinput
+                        gr_prompt_type = prompt_type
+                        gr_prompt_dict = prompt_dict
+                    client_kwargs = dict(
+                        instruction=gr_prompt
+                        if chat_client
+                        else "",  # only for chat=True
+                        iinput=gr_iinput,  # only for chat=True
+                        context=gr_context,
+                        # streaming output is supported, loops over and outputs each generation in streaming mode
+                        # but leave stream_output=False for simple input/output mode
+                        stream_output=stream_output,
+                        **gen_server_kwargs,
+                        prompt_type=gr_prompt_type,
+                        prompt_dict=gr_prompt_dict,
+                        instruction_nochat=gr_prompt
+                        if not chat_client
+                        else "",
+                        iinput_nochat=gr_iinput,  # only for chat=False
+                        langchain_mode=client_langchain_mode,
+                        langchain_action=client_langchain_action,
+                        top_k_docs=top_k_docs,
+                        chunk=chunk,
+                        chunk_size=chunk_size,
+                        document_choice=[DocumentChoices.All_Relevant.name],
+                    )
+                    api_name = "/submit_nochat_api"  # NOTE: like submit_nochat but stable API for string dict passing
+                    if not stream_output:
+                        res = gr_client.predict(
+                            str(dict(client_kwargs)), api_name=api_name
+                        )
+                        res_dict = ast.literal_eval(res)
+                        text = res_dict["response"]
+                        sources = res_dict["sources"]
+                        yield dict(
+                            response=prompter.get_response(
+                                prompt + text,
+                                prompt=prompt,
+                                sanitize_bot_response=sanitize_bot_response,
+                            ),
+                            sources=sources,
+                        )
+                    else:
+                        job = gr_client.submit(
+                            str(dict(client_kwargs)), api_name=api_name
+                        )
+                        text = ""
+                        sources = ""
+                        res_dict = dict(response=text, sources=sources)
+                        while not job.done():
+                            outputs_list = job.communicator.job.outputs
+                            if outputs_list:
+                                res = job.communicator.job.outputs[-1]
+                                res_dict = ast.literal_eval(res)
+                                text = res_dict["response"]
+                                sources = res_dict["sources"]
+                                if gr_prompt_type == "plain":
+                                    # then gradio server passes back full prompt + text
+                                    prompt_and_text = text
+                                else:
+                                    prompt_and_text = prompt + text
+                                yield dict(
+                                    response=prompter.get_response(
+                                        prompt_and_text,
+                                        prompt=prompt,
+                                        sanitize_bot_response=sanitize_bot_response,
+                                    ),
+                                    sources=sources,
+                                )
+                            time.sleep(0.01)
+                        # ensure get last output to avoid race
+                        res_all = job.outputs()
+                        if len(res_all) > 0:
+                            res = res_all[-1]
+                            res_dict = ast.literal_eval(res)
+                            text = res_dict["response"]
+                            sources = res_dict["sources"]
+                        else:
+                            # go with old text if last call didn't work
+                            e = job.future._exception
+                            if e is not None:
+                                stre = str(e)
+                                strex = "".join(
+                                    traceback.format_tb(e.__traceback__)
+                                )
+                            else:
+                                stre = ""
+                                strex = ""
+
+                            print(
+                                "Bad final response: %s %s %s %s %s: %s %s"
+                                % (
+                                    base_model,
+                                    inference_server,
+                                    res_all,
+                                    prompt,
+                                    text,
+                                    stre,
+                                    strex,
+                                ),
+                                flush=True,
+                            )
+                        if gr_prompt_type == "plain":
+                            # then gradio server passes back full prompt + text
+                            prompt_and_text = text
+                        else:
+                            prompt_and_text = prompt + text
+                        yield dict(
+                            response=prompter.get_response(
+                                prompt_and_text,
+                                prompt=prompt,
+                                sanitize_bot_response=sanitize_bot_response,
+                            ),
+                            sources=sources,
+                        )
+                elif hf_client:
+                    # HF inference server needs control over input tokens
+                    where_from = "hf_client"
+
+                    # prompt must include all human-bot like tokens, already added by prompt
+                    # https://github.com/huggingface/text-generation-inference/tree/main/clients/python#types
+                    stop_sequences = list(
+                        set(
+                            prompter.terminate_response
+                            + [prompter.PreResponse]
+                        )
+                    )
+                    stop_sequences = [x for x in stop_sequences if x]
+                    gen_server_kwargs = dict(
+                        do_sample=do_sample,
+                        max_new_tokens=max_new_tokens,
+                        # best_of=None,
+                        repetition_penalty=repetition_penalty,
+                        return_full_text=True,
+                        seed=SEED,
+                        stop_sequences=stop_sequences,
+                        temperature=temperature,
+                        top_k=top_k,
+                        top_p=top_p,
+                        # truncate=False,  # behaves oddly
+                        # typical_p=top_p,
+                        # watermark=False,
+                        # decoder_input_details=False,
+                    )
+                    # work-around for timeout at constructor time, will be issue if multi-threading,
+                    # so just do something reasonable or max_time if larger
+                    # lower bound because client is re-used if multi-threading
+                    hf_client.timeout = max(300, max_time)
+                    if not stream_output:
+                        text = hf_client.generate(
+                            prompt, **gen_server_kwargs
+                        ).generated_text
+                        yield dict(
+                            response=prompter.get_response(
+                                text,
+                                prompt=prompt,
+                                sanitize_bot_response=sanitize_bot_response,
+                            ),
+                            sources="",
+                        )
+                    else:
+                        text = ""
+                        for response in hf_client.generate_stream(
+                            prompt, **gen_server_kwargs
+                        ):
+                            if not response.token.special:
+                                # stop_sequences
+                                text_chunk = response.token.text
+                                text += text_chunk
+                                yield dict(
+                                    response=prompter.get_response(
+                                        prompt + text,
+                                        prompt=prompt,
+                                        sanitize_bot_response=sanitize_bot_response,
+                                    ),
+                                    sources="",
+                                )
+                else:
+                    raise RuntimeError(
+                        "Failed to get client: %s" % inference_server
+                    )
+            else:
+                raise RuntimeError(
+                    "No such inference_server  %s" % inference_server
+                )
+
+            if save_dir and text:
+                # save prompt + new text
+                extra_dict = gen_server_kwargs.copy()
+                extra_dict.update(
+                    dict(
+                        inference_server=inference_server,
+                        num_prompt_tokens=num_prompt_tokens,
+                    )
+                )
+                save_generate_output(
+                    prompt=prompt,
+                    output=text,
+                    base_model=base_model,
+                    save_dir=save_dir,
+                    where_from=where_from,
+                    extra_dict=extra_dict,
+                )
+            return
+        else:
+            assert not inference_server, (
+                "inferene_server=%s not supported" % inference_server
             )
-            return out
+
+        if isinstance(tokenizer, str):
+            # pipeline
+            if tokenizer == "summarization":
+                key = "summary_text"
+            else:
+                raise RuntimeError("No such task type %s" % tokenizer)
+            # NOTE: uses max_length only
+            yield dict(
+                response=model(prompt, max_length=max_new_tokens)[0][key],
+                sources="",
+            )
+
+        if "mbart-" in base_model.lower():
+            assert src_lang is not None
+            tokenizer.src_lang = self.languages_covered()[src_lang]
+
+        stopping_criteria = get_stopping(
+            prompt_type,
+            prompt_dict,
+            tokenizer,
+            self.device,
+            model_max_length=tokenizer.model_max_length,
+        )
+
+        print(prompt)
+        # exit(0)
+        inputs = tokenizer(prompt, return_tensors="pt")
+        if debug and len(inputs["input_ids"]) > 0:
+            print("input_ids length", len(inputs["input_ids"][0]), flush=True)
+        input_ids = inputs["input_ids"].to(self.device)
+        # CRITICAL LIMIT else will fail
+        max_max_tokens = tokenizer.model_max_length
+        max_input_tokens = max_max_tokens - min_new_tokens
+        # NOTE: Don't limit up front due to max_new_tokens, let go up to max or reach max_max_tokens in stopping.py
+        input_ids = input_ids[:, -max_input_tokens:]
+        # required for falcon if multiple threads or asyncio accesses to model during generation
+        if use_cache is None:
+            use_cache = False if "falcon" in base_model else True
+        gen_config_kwargs = dict(
+            temperature=float(temperature),
+            top_p=float(top_p),
+            top_k=top_k,
+            num_beams=num_beams,
+            do_sample=do_sample,
+            repetition_penalty=float(repetition_penalty),
+            num_return_sequences=num_return_sequences,
+            renormalize_logits=True,
+            remove_invalid_values=True,
+            use_cache=use_cache,
+        )
+        token_ids = [
+            "eos_token_id",
+            "pad_token_id",
+            "bos_token_id",
+            "cls_token_id",
+            "sep_token_id",
+        ]
+        for token_id in token_ids:
+            if (
+                hasattr(tokenizer, token_id)
+                and getattr(tokenizer, token_id) is not None
+            ):
+                gen_config_kwargs.update(
+                    {token_id: getattr(tokenizer, token_id)}
+                )
+        generation_config = GenerationConfig(**gen_config_kwargs)
+
+        gen_kwargs = dict(
+            input_ids=input_ids,
+            generation_config=generation_config,
+            return_dict_in_generate=True,
+            output_scores=True,
+            max_new_tokens=max_new_tokens,  # prompt + new
+            min_new_tokens=min_new_tokens,  # prompt + new
+            early_stopping=early_stopping,  # False, True, "never"
+            max_time=max_time,
+            stopping_criteria=stopping_criteria,
+        )
+        if "gpt2" in base_model.lower():
+            gen_kwargs.update(
+                dict(
+                    bos_token_id=tokenizer.bos_token_id,
+                    pad_token_id=tokenizer.eos_token_id,
+                )
+            )
+        elif "mbart-" in base_model.lower():
+            assert tgt_lang is not None
+            tgt_lang = self.languages_covered()[tgt_lang]
+            gen_kwargs.update(
+                dict(forced_bos_token_id=tokenizer.lang_code_to_id[tgt_lang])
+            )
+        else:
+            token_ids = ["eos_token_id", "bos_token_id", "pad_token_id"]
+            for token_id in token_ids:
+                if (
+                    hasattr(tokenizer, token_id)
+                    and getattr(tokenizer, token_id) is not None
+                ):
+                    gen_kwargs.update({token_id: getattr(tokenizer, token_id)})
+
+        decoder_kwargs = dict(
+            skip_special_tokens=True, clean_up_tokenization_spaces=True
+        )
+
+        decoder = functools.partial(tokenizer.decode, **decoder_kwargs)
+        decoder_raw_kwargs = dict(
+            skip_special_tokens=False, clean_up_tokenization_spaces=True
+        )
+
+        decoder_raw = functools.partial(tokenizer.decode, **decoder_raw_kwargs)
+
+        with torch.no_grad():
+            have_lora_weights = lora_weights not in [no_lora_str, "", None]
+            context_class_cast = (
+                NullContext
+                if self.device == "cpu"
+                or have_lora_weights
+                or self.device == "mps"
+                else torch.autocast
+            )
+            with context_class_cast(self.device):
+                # protection for gradio not keeping track of closed users,
+                # else hit bitsandbytes lack of thread safety:
+                # https://github.com/h2oai/h2ogpt/issues/104
+                # but only makes sense if concurrency_count == 1
+                context_class = NullContext  # if concurrency_count > 1 else filelock.FileLock
+                if verbose:
+                    print("Pre-Generate: %s" % str(datetime.now()), flush=True)
+                decoded_output = None
+                with context_class("generate.lock"):
+                    if verbose:
+                        print("Generate: %s" % str(datetime.now()), flush=True)
+                    # decoded tokenized prompt can deviate from prompt due to special characters
+                    inputs_decoded = decoder(input_ids[0])
+                    inputs_decoded_raw = decoder_raw(input_ids[0])
+                    if inputs_decoded == prompt:
+                        # normal
+                        pass
+                    elif inputs_decoded.lstrip() == prompt.lstrip():
+                        # sometimes extra space in front, make prompt same for prompt removal
+                        prompt = inputs_decoded
+                    elif inputs_decoded_raw == prompt:
+                        # some models specify special tokens that are part of normal prompt, so can't skip them
+                        inputs_decoded = prompt = inputs_decoded_raw
+                        decoder = decoder_raw
+                        decoder_kwargs = decoder_raw_kwargs
+                    elif inputs_decoded_raw.replace("<unk> ", "").replace(
+                        "<unk>", ""
+                    ).replace("\n", " ").replace(" ", "") == prompt.replace(
+                        "\n", " "
+                    ).replace(
+                        " ", ""
+                    ):
+                        inputs_decoded = prompt = inputs_decoded_raw
+                        decoder = decoder_raw
+                        decoder_kwargs = decoder_raw_kwargs
+                    else:
+                        if verbose:
+                            print(
+                                "WARNING: Special characters in prompt",
+                                flush=True,
+                            )
+                    if stream_output:
+                        skip_prompt = False
+                        streamer = H2OTextIteratorStreamer(
+                            tokenizer,
+                            skip_prompt=skip_prompt,
+                            block=False,
+                            **decoder_kwargs,
+                        )
+                        gen_kwargs.update(dict(streamer=streamer))
+                        target = wrapped_partial(
+                            self.generate_with_exceptions,
+                            model.generate,
+                            prompt=prompt,
+                            inputs_decoded=inputs_decoded,
+                            raise_generate_gpu_exceptions=raise_generate_gpu_exceptions,
+                            **gen_kwargs,
+                        )
+                        bucket = queue.Queue()
+                        thread = EThread(
+                            target=target, streamer=streamer, bucket=bucket
+                        )
+                        thread.start()
+                        outputs = ""
+                        try:
+                            for new_text in streamer:
+                                if bucket.qsize() > 0 or thread.exc:
+                                    thread.join()
+                                outputs += new_text
+                                yield dict(
+                                    response=prompter.get_response(
+                                        outputs,
+                                        prompt=inputs_decoded,
+                                        sanitize_bot_response=sanitize_bot_response,
+                                    ),
+                                    sources="",
+                                )
+                        except BaseException:
+                            # if any exception, raise that exception if was from thread, first
+                            if thread.exc:
+                                raise thread.exc
+                            raise
+                        finally:
+                            # clear before return, since .then() never done if from API
+                            clear_torch_cache()
+                            # in case no exception and didn't join with thread yet, then join
+                            if not thread.exc:
+                                thread.join()
+                        # in case raise StopIteration or broke queue loop in streamer, but still have exception
+                        if thread.exc:
+                            raise thread.exc
+                        decoded_output = outputs
+                    else:
+                        try:
+                            outputs = model.generate(**gen_kwargs)
+                        finally:
+                            clear_torch_cache()  # has to be here for API submit_nochat_api since.then() not called
+                        outputs = [decoder(s) for s in outputs.sequences]
+                        yield dict(
+                            response=prompter.get_response(
+                                outputs,
+                                prompt=inputs_decoded,
+                                sanitize_bot_response=sanitize_bot_response,
+                            ),
+                            sources="",
+                        )
+                        if outputs and len(outputs) >= 1:
+                            decoded_output = prompt + outputs[0]
+                    if save_dir and decoded_output:
+                        extra_dict = gen_config_kwargs.copy()
+                        extra_dict.update(
+                            dict(num_prompt_tokens=num_prompt_tokens)
+                        )
+                        save_generate_output(
+                            prompt=prompt,
+                            output=decoded_output,
+                            base_model=base_model,
+                            save_dir=save_dir,
+                            where_from="evaluate_%s" % str(stream_output),
+                            extra_dict=gen_config_kwargs,
+                        )
+                if verbose:
+                    print(
+                        "Post-Generate: %s decoded_output: %s"
+                        % (
+                            str(datetime.now()),
+                            len(decoded_output) if decoded_output else -1,
+                        ),
+                        flush=True,
+                    )
+                return outputs[0]
 
     inputs_list_names = list(inspect.signature(evaluate).parameters)
     global inputs_kwargs_list

apps/language_models/langchain/gpt_langchain.py

@@ -436,7 +436,7 @@ class GradioInference(LLM):
     chat_client: bool = False
 
     return_full_text: bool = True
-    stream_output: bool = Field(False, alias="stream")
+    stream: bool = False
     sanitize_bot_response: bool = False
 
     prompter: Any = None
@@ -481,7 +481,7 @@ class GradioInference(LLM):
         # so server should get prompt_type or '', not plain
         # This is good, so gradio server can also handle stopping.py conditions
         # this is different than TGI server that uses prompter to inject prompt_type prompting
-        stream_output = self.stream_output
+        stream_output = self.stream
         gr_client = self.client
         client_langchain_mode = "Disabled"
         client_langchain_action = LangChainAction.QUERY.value
@@ -596,7 +596,7 @@ class H2OHuggingFaceTextGenInference(HuggingFaceTextGenInference):
     inference_server_url: str = ""
     timeout: int = 300
     headers: dict = None
-    stream_output: bool = Field(False, alias="stream")
+    stream: bool = False
     sanitize_bot_response: bool = False
     prompter: Any = None
     tokenizer: Any = None
@@ -663,7 +663,7 @@ class H2OHuggingFaceTextGenInference(HuggingFaceTextGenInference):
         # lower bound because client is re-used if multi-threading
         self.client.timeout = max(300, self.timeout)
 
-        if not self.stream_output:
+        if not self.stream:
             res = self.client.generate(
                 prompt,
                 **gen_server_kwargs,
@@ -852,7 +852,7 @@ def get_llm(
                 top_p=top_p,
                 # typical_p=top_p,
                 callbacks=callbacks if stream_output else None,
-                stream_output=stream_output,
+                stream=stream_output,
                 prompter=prompter,
                 tokenizer=tokenizer,
                 client=hf_client,
@@ -2510,7 +2510,8 @@ def _run_qa_db(
         formatted_doc_chunks = "\n\n".join(
             [get_url(x) + "\n\n" + x.page_content for x in docs]
         )
-        return formatted_doc_chunks, ""
+        yield formatted_doc_chunks, ""
+        return
     if not docs and langchain_action in [
         LangChainAction.SUMMARIZE_MAP.value,
         LangChainAction.SUMMARIZE_ALL.value,
@@ -2522,7 +2523,8 @@ def _run_qa_db(
             else "No documents to summarize."
         )
         extra = ""
-        return ret, extra
+        yield ret, extra
+        return
     if not docs and langchain_mode not in [
         LangChainMode.DISABLED.value,
         LangChainMode.CHAT_LLM.value,
@@ -2534,7 +2536,8 @@ def _run_qa_db(
             else "No documents to query."
         )
         extra = ""
-        return ret, extra
+        yield ret, extra
+        return
 
     if chain is None and model_name not in non_hf_types:
         # here if no docs at all and not HF type
@@ -2554,7 +2557,22 @@ def _run_qa_db(
         )
         with context_class_cast(args.device):
             answer = chain()
-            return answer
+
+    if not use_context:
+        ret = answer["output_text"]
+        extra = ""
+        yield ret, extra
+    elif answer is not None:
+        ret, extra = get_sources_answer(
+            query,
+            answer,
+            scores,
+            show_rank,
+            answer_with_sources,
+            verbose=verbose,
+        )
+        yield ret, extra
+    return
 
 
 def get_similarity_chain(

apps/language_models/langchain/h2oai_pipeline.py

@@ -3,11 +3,13 @@ from apps.stable_diffusion.src.utils.utils import _compile_module
 from io import BytesIO
 import torch_mlir
 
+from transformers import TextGenerationPipeline
+from transformers.pipelines.text_generation import ReturnType
+
 from stopping import get_stopping
 from prompter import Prompter, PromptType
 
-from transformers import TextGenerationPipeline
-from transformers.pipelines.text_generation import ReturnType
+
 from transformers.generation import (
     GenerationConfig,
     LogitsProcessorList,
@@ -20,7 +22,7 @@ import gc
 from pathlib import Path
 from shark.shark_inference import SharkInference
 from shark.shark_downloader import download_public_file
-from shark.shark_importer import import_with_fx, save_mlir
+from shark.shark_importer import import_with_fx
 from apps.stable_diffusion.src import args
 
 # Brevitas
@@ -29,8 +31,14 @@ from brevitas_examples.llm.llm_quant.quantize import quantize_model
 from brevitas_examples.llm.llm_quant.run_utils import get_model_impl
 
 
-# fmt: off
-def quant〇matmul_rhs_group_quant〡shape(lhs: List[int], rhs: List[int], rhs_scale: List[int], rhs_zero_point: List[int], rhs_bit_width: int, rhs_group_size: int) -> List[int]:
+def brevitas〇matmul_rhs_group_quant〡shape(
+    lhs: List[int],
+    rhs: List[int],
+    rhs_scale: List[int],
+    rhs_zero_point: List[int],
+    rhs_bit_width: int,
+    rhs_group_size: int,
+) -> List[int]:
     if len(lhs) == 3 and len(rhs) == 2:
         return [lhs[0], lhs[1], rhs[0]]
     elif len(lhs) == 2 and len(rhs) == 2:
@@ -39,21 +47,30 @@ def quant〇matmul_rhs_group_quant〡shape(lhs: List[int], rhs: List[int], rhs_s
         raise ValueError("Input shapes not supported.")
 
 
-def quant〇matmul_rhs_group_quant〡dtype(lhs_rank_dtype: Tuple[int, int], rhs_rank_dtype: Tuple[int, int], rhs_scale_rank_dtype: Tuple[int, int], rhs_zero_point_rank_dtype: Tuple[int, int], rhs_bit_width: int, rhs_group_size: int) -> int:
+def brevitas〇matmul_rhs_group_quant〡dtype(
+    lhs_rank_dtype: Tuple[int, int],
+    rhs_rank_dtype: Tuple[int, int],
+    rhs_scale_rank_dtype: Tuple[int, int],
+    rhs_zero_point_rank_dtype: Tuple[int, int],
+    rhs_bit_width: int,
+    rhs_group_size: int,
+) -> int:
     # output dtype is the dtype of the lhs float input
     lhs_rank, lhs_dtype = lhs_rank_dtype
     return lhs_dtype
 
 
-def quant〇matmul_rhs_group_quant〡has_value_semantics(lhs, rhs, rhs_scale, rhs_zero_point, rhs_bit_width, rhs_group_size) -> None:
+def brevitas〇matmul_rhs_group_quant〡has_value_semantics(
+    lhs, rhs, rhs_scale, rhs_zero_point, rhs_bit_width, rhs_group_size
+) -> None:
     return
 
 
 brevitas_matmul_rhs_group_quant_library = [
-    quant〇matmul_rhs_group_quant〡shape,
-    quant〇matmul_rhs_group_quant〡dtype,
-    quant〇matmul_rhs_group_quant〡has_value_semantics]
-# fmt: on
+    brevitas〇matmul_rhs_group_quant〡shape,
+    brevitas〇matmul_rhs_group_quant〡dtype,
+    brevitas〇matmul_rhs_group_quant〡has_value_semantics,
+]
 
 global_device = "cuda"
 global_precision = "fp16"
@@ -229,7 +246,7 @@ class H2OGPTSHARKModel(torch.nn.Module):
                 ts_graph,
                 [*h2ogptCompileInput],
                 output_type=torch_mlir.OutputType.TORCH,
-                backend_legal_ops=["quant.matmul_rhs_group_quant"],
+                backend_legal_ops=["brevitas.matmul_rhs_group_quant"],
                 extra_library=brevitas_matmul_rhs_group_quant_library,
                 use_tracing=False,
                 verbose=False,
@@ -237,7 +254,7 @@ class H2OGPTSHARKModel(torch.nn.Module):
             print(f"[DEBUG] converting torch to linalg")
             run_pipeline_with_repro_report(
                 module,
-                "builtin.module(func.func(torch-unpack-quant-tensor),func.func(torch-convert-custom-quant-op),torch-backend-to-linalg-on-tensors-backend-pipeline)",
+                "builtin.module(func.func(torch-unpack-torch-tensor),torch-backend-to-linalg-on-tensors-backend-pipeline)",
                 description="Lowering Torch Backend IR -> Linalg-on-Tensors Backend IR",
             )
         else:
@@ -256,11 +273,6 @@ class H2OGPTSHARKModel(torch.nn.Module):
         bytecode = bytecode_stream.getvalue()
         del module
 
-        bytecode = save_mlir(
-            bytecode,
-            model_name=f"h2ogpt_{precision}",
-            frontend="torch",
-        )
         return bytecode
 
     def forward(self, input_ids, attention_mask):
@@ -273,215 +285,7 @@ class H2OGPTSHARKModel(torch.nn.Module):
         return result
 
 
-def decode_tokens(tokenizer, res_tokens):
-    for i in range(len(res_tokens)):
-        if type(res_tokens[i]) != int:
-            res_tokens[i] = int(res_tokens[i][0])
-
-    res_str = tokenizer.decode(res_tokens, skip_special_tokens=True)
-    return res_str
-
-
-def generate_token(h2ogpt_shark_model, model, tokenizer, **generate_kwargs):
-    del generate_kwargs["max_time"]
-    generate_kwargs["input_ids"] = generate_kwargs["input_ids"].to(
-        device=tensor_device
-    )
-    generate_kwargs["attention_mask"] = generate_kwargs["attention_mask"].to(
-        device=tensor_device
-    )
-    truncated_input_ids = []
-    stopping_criteria = generate_kwargs["stopping_criteria"]
-
-    generation_config_ = GenerationConfig.from_model_config(model.config)
-    generation_config = copy.deepcopy(generation_config_)
-    model_kwargs = generation_config.update(**generate_kwargs)
-
-    logits_processor = LogitsProcessorList()
-    stopping_criteria = (
-        stopping_criteria
-        if stopping_criteria is not None
-        else StoppingCriteriaList()
-    )
-
-    eos_token_id = generation_config.eos_token_id
-    generation_config.pad_token_id = eos_token_id
-
-    (
-        inputs_tensor,
-        model_input_name,
-        model_kwargs,
-    ) = model._prepare_model_inputs(
-        None, generation_config.bos_token_id, model_kwargs
-    )
-
-    model_kwargs["output_attentions"] = generation_config.output_attentions
-    model_kwargs[
-        "output_hidden_states"
-    ] = generation_config.output_hidden_states
-    model_kwargs["use_cache"] = generation_config.use_cache
-
-    input_ids = (
-        inputs_tensor
-        if model_input_name == "input_ids"
-        else model_kwargs.pop("input_ids")
-    )
-
-    input_ids_seq_length = input_ids.shape[-1]
-
-    generation_config.max_length = (
-        generation_config.max_new_tokens + input_ids_seq_length
-    )
-
-    logits_processor = model._get_logits_processor(
-        generation_config=generation_config,
-        input_ids_seq_length=input_ids_seq_length,
-        encoder_input_ids=inputs_tensor,
-        prefix_allowed_tokens_fn=None,
-        logits_processor=logits_processor,
-    )
-
-    stopping_criteria = model._get_stopping_criteria(
-        generation_config=generation_config,
-        stopping_criteria=stopping_criteria,
-    )
-
-    logits_warper = model._get_logits_warper(generation_config)
-
-    (
-        input_ids,
-        model_kwargs,
-    ) = model._expand_inputs_for_generation(
-        input_ids=input_ids,
-        expand_size=generation_config.num_return_sequences,  # 1
-        is_encoder_decoder=model.config.is_encoder_decoder,  # False
-        **model_kwargs,
-    )
-
-    if isinstance(eos_token_id, int):
-        eos_token_id = [eos_token_id]
-    eos_token_id_tensor = (
-        torch.tensor(eos_token_id).to(device=tensor_device)
-        if eos_token_id is not None
-        else None
-    )
-
-    pad_token_id = generation_config.pad_token_id
-    eos_token_id = eos_token_id
-
-    output_scores = generation_config.output_scores  # False
-    return_dict_in_generate = (
-        generation_config.return_dict_in_generate  # False
-    )
-
-    # init attention / hidden states / scores tuples
-    scores = () if (return_dict_in_generate and output_scores) else None
-
-    # keep track of which sequences are already finished
-    unfinished_sequences = torch.ones(
-        input_ids.shape[0],
-        dtype=torch.long,
-        device=input_ids.device,
-    )
-
-    timesRan = 0
-    import time
-
-    start = time.time()
-    print("\n")
-
-    res_tokens = []
-    while True:
-        model_inputs = model.prepare_inputs_for_generation(
-            input_ids, **model_kwargs
-        )
-
-        outputs = h2ogpt_shark_model.forward(
-            model_inputs["input_ids"], model_inputs["attention_mask"]
-        )
-
-        if args.precision == "fp16":
-            outputs = outputs.to(dtype=torch.float32)
-        next_token_logits = outputs
-
-        # pre-process distribution
-        next_token_scores = logits_processor(input_ids, next_token_logits)
-        next_token_scores = logits_warper(input_ids, next_token_scores)
-
-        # sample
-        probs = torch.nn.functional.softmax(next_token_scores, dim=-1)
-
-        next_token = torch.multinomial(probs, num_samples=1).squeeze(1)
-
-        # finished sentences should have their next token be a padding token
-        if eos_token_id is not None:
-            if pad_token_id is None:
-                raise ValueError(
-                    "If `eos_token_id` is defined, make sure that `pad_token_id` is defined."
-                )
-            next_token = next_token * unfinished_sequences + pad_token_id * (
-                1 - unfinished_sequences
-            )
-
-        input_ids = torch.cat([input_ids, next_token[:, None]], dim=-1)
-
-        model_kwargs["past_key_values"] = None
-        if "attention_mask" in model_kwargs:
-            attention_mask = model_kwargs["attention_mask"]
-            model_kwargs["attention_mask"] = torch.cat(
-                [
-                    attention_mask,
-                    attention_mask.new_ones((attention_mask.shape[0], 1)),
-                ],
-                dim=-1,
-            )
-
-        truncated_input_ids.append(input_ids[:, 0])
-        input_ids = input_ids[:, 1:]
-        model_kwargs["attention_mask"] = model_kwargs["attention_mask"][:, 1:]
-
-        new_word = tokenizer.decode(
-            next_token.cpu().numpy(),
-            add_special_tokens=False,
-            skip_special_tokens=True,
-            clean_up_tokenization_spaces=True,
-        )
-
-        res_tokens.append(next_token)
-        if new_word == "<0x0A>":
-            print("\n", end="", flush=True)
-        else:
-            print(f"{new_word}", end=" ", flush=True)
-
-        part_str = decode_tokens(tokenizer, res_tokens)
-        yield part_str
-
-        # if eos_token was found in one sentence, set sentence to finished
-        if eos_token_id_tensor is not None:
-            unfinished_sequences = unfinished_sequences.mul(
-                next_token.tile(eos_token_id_tensor.shape[0], 1)
-                .ne(eos_token_id_tensor.unsqueeze(1))
-                .prod(dim=0)
-            )
-            # stop when each sentence is finished
-            if unfinished_sequences.max() == 0 or stopping_criteria(
-                input_ids, scores
-            ):
-                break
-        timesRan = timesRan + 1
-
-    end = time.time()
-    print(
-        "\n\nTime taken is {:.2f} seconds/token\n".format(
-            (end - start) / timesRan
-        )
-    )
-
-    torch.cuda.empty_cache()
-    gc.collect()
-
-    res_str = decode_tokens(tokenizer, res_tokens)
-    yield res_str
+h2ogpt_model = H2OGPTSHARKModel()
 
 
 def pad_or_truncate_inputs(
@@ -694,6 +498,233 @@ class H2OTextGenerationPipeline(TextGenerationPipeline):
             )
         return records
 
+    def generate_new_token(self):
+        model_inputs = self.model.prepare_inputs_for_generation(
+            self.input_ids, **self.model_kwargs
+        )
+
+        outputs = h2ogpt_model.forward(
+            model_inputs["input_ids"], model_inputs["attention_mask"]
+        )
+
+        if args.precision == "fp16":
+            outputs = outputs.to(dtype=torch.float32)
+        next_token_logits = outputs
+
+        # pre-process distribution
+        next_token_scores = self.logits_processor(
+            self.input_ids, next_token_logits
+        )
+        next_token_scores = self.logits_warper(
+            self.input_ids, next_token_scores
+        )
+
+        # sample
+        probs = torch.nn.functional.softmax(next_token_scores, dim=-1)
+
+        next_token = torch.multinomial(probs, num_samples=1).squeeze(1)
+
+        # finished sentences should have their next token be a padding token
+        if self.eos_token_id is not None:
+            if self.pad_token_id is None:
+                raise ValueError(
+                    "If `eos_token_id` is defined, make sure that `pad_token_id` is defined."
+                )
+            next_token = (
+                next_token * self.unfinished_sequences
+                + self.pad_token_id * (1 - self.unfinished_sequences)
+            )
+
+        self.input_ids = torch.cat(
+            [self.input_ids, next_token[:, None]], dim=-1
+        )
+
+        self.model_kwargs["past_key_values"] = None
+        if "attention_mask" in self.model_kwargs:
+            attention_mask = self.model_kwargs["attention_mask"]
+            self.model_kwargs["attention_mask"] = torch.cat(
+                [
+                    attention_mask,
+                    attention_mask.new_ones((attention_mask.shape[0], 1)),
+                ],
+                dim=-1,
+            )
+
+        self.truncated_input_ids.append(self.input_ids[:, 0])
+        self.input_ids = self.input_ids[:, 1:]
+        self.model_kwargs["attention_mask"] = self.model_kwargs[
+            "attention_mask"
+        ][:, 1:]
+
+        return next_token
+
+    def generate_token(self, **generate_kwargs):
+        del generate_kwargs["max_time"]
+        self.truncated_input_ids = []
+
+        generation_config_ = GenerationConfig.from_model_config(
+            self.model.config
+        )
+        generation_config = copy.deepcopy(generation_config_)
+        self.model_kwargs = generation_config.update(**generate_kwargs)
+
+        logits_processor = LogitsProcessorList()
+        self.stopping_criteria = (
+            self.stopping_criteria
+            if self.stopping_criteria is not None
+            else StoppingCriteriaList()
+        )
+
+        eos_token_id = generation_config.eos_token_id
+        generation_config.pad_token_id = eos_token_id
+
+        (
+            inputs_tensor,
+            model_input_name,
+            self.model_kwargs,
+        ) = self.model._prepare_model_inputs(
+            None, generation_config.bos_token_id, self.model_kwargs
+        )
+        batch_size = inputs_tensor.shape[0]
+
+        self.model_kwargs[
+            "output_attentions"
+        ] = generation_config.output_attentions
+        self.model_kwargs[
+            "output_hidden_states"
+        ] = generation_config.output_hidden_states
+        self.model_kwargs["use_cache"] = generation_config.use_cache
+
+        self.input_ids = (
+            inputs_tensor
+            if model_input_name == "input_ids"
+            else self.model_kwargs.pop("input_ids")
+        )
+
+        input_ids_seq_length = self.input_ids.shape[-1]
+
+        generation_config.max_length = (
+            generation_config.max_new_tokens + input_ids_seq_length
+        )
+
+        self.logits_processor = self.model._get_logits_processor(
+            generation_config=generation_config,
+            input_ids_seq_length=input_ids_seq_length,
+            encoder_input_ids=inputs_tensor,
+            prefix_allowed_tokens_fn=None,
+            logits_processor=logits_processor,
+        )
+
+        self.stopping_criteria = self.model._get_stopping_criteria(
+            generation_config=generation_config,
+            stopping_criteria=self.stopping_criteria,
+        )
+
+        self.logits_warper = self.model._get_logits_warper(generation_config)
+
+        (
+            self.input_ids,
+            self.model_kwargs,
+        ) = self.model._expand_inputs_for_generation(
+            input_ids=self.input_ids,
+            expand_size=generation_config.num_return_sequences,  # 1
+            is_encoder_decoder=self.model.config.is_encoder_decoder,  # False
+            **self.model_kwargs,
+        )
+
+        if isinstance(eos_token_id, int):
+            eos_token_id = [eos_token_id]
+        self.eos_token_id_tensor = (
+            torch.tensor(eos_token_id).to(device=tensor_device)
+            if eos_token_id is not None
+            else None
+        )
+
+        self.pad_token_id = generation_config.pad_token_id
+        self.eos_token_id = eos_token_id
+
+        output_scores = generation_config.output_scores  # False
+        output_attentions = generation_config.output_attentions  # False
+        output_hidden_states = generation_config.output_hidden_states  # False
+        return_dict_in_generate = (
+            generation_config.return_dict_in_generate  # False
+        )
+
+        # init attention / hidden states / scores tuples
+        self.scores = (
+            () if (return_dict_in_generate and output_scores) else None
+        )
+        decoder_attentions = (
+            () if (return_dict_in_generate and output_attentions) else None
+        )
+        cross_attentions = (
+            () if (return_dict_in_generate and output_attentions) else None
+        )
+        decoder_hidden_states = (
+            () if (return_dict_in_generate and output_hidden_states) else None
+        )
+
+        # keep track of which sequences are already finished
+        self.unfinished_sequences = torch.ones(
+            self.input_ids.shape[0],
+            dtype=torch.long,
+            device=self.input_ids.device,
+        )
+
+        timesRan = 0
+        import time
+
+        start = time.time()
+        print("\n")
+
+        while True:
+            next_token = self.generate_new_token()
+            new_word = self.tokenizer.decode(
+                next_token.cpu().numpy(),
+                add_special_tokens=False,
+                skip_special_tokens=True,
+                clean_up_tokenization_spaces=True,
+            )
+
+            print(f"{new_word}", end="", flush=True)
+
+            # if eos_token was found in one sentence, set sentence to finished
+            if self.eos_token_id_tensor is not None:
+                self.unfinished_sequences = self.unfinished_sequences.mul(
+                    next_token.tile(self.eos_token_id_tensor.shape[0], 1)
+                    .ne(self.eos_token_id_tensor.unsqueeze(1))
+                    .prod(dim=0)
+                )
+                # stop when each sentence is finished
+                if (
+                    self.unfinished_sequences.max() == 0
+                    or self.stopping_criteria(self.input_ids, self.scores)
+                ):
+                    break
+            timesRan = timesRan + 1
+
+        end = time.time()
+        print(
+            "\n\nTime taken is {:.2f} seconds/token\n".format(
+                (end - start) / timesRan
+            )
+        )
+
+        self.input_ids = torch.cat(
+            [
+                torch.tensor(self.truncated_input_ids)
+                .to(device=tensor_device)
+                .unsqueeze(dim=0),
+                self.input_ids,
+            ],
+            dim=-1,
+        )
+
+        torch.cuda.empty_cache()
+        gc.collect()
+
+        return self.input_ids
+
     def _forward(self, model_inputs, **generate_kwargs):
         if self.can_stop:
             stopping_criteria = get_stopping(
@@ -753,13 +784,19 @@ class H2OTextGenerationPipeline(TextGenerationPipeline):
         input_ids, attention_mask = pad_or_truncate_inputs(
             input_ids, attention_mask, max_padding_length=max_padding_length
         )
+        self.stopping_criteria = generate_kwargs["stopping_criteria"]
 
-        return_dict = {
-            "model": self.model,
-            "tokenizer": self.tokenizer,
+        generated_sequence = self.generate_token(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            **generate_kwargs,
+        )
+        out_b = generated_sequence.shape[0]
+        generated_sequence = generated_sequence.reshape(
+            in_b, out_b // in_b, *generated_sequence.shape[1:]
+        )
+        return {
+            "generated_sequence": generated_sequence,
             "input_ids": input_ids,
-            "attention_mask": attention_mask,
-            "attention_mask": attention_mask,
+            "prompt_text": prompt_text,
         }
-        return_dict = {**return_dict, **generate_kwargs}
-        return return_dict

apps/language_models/langchain/langchain_requirements.txt

@@ -65,8 +65,8 @@ tiktoken==0.4.0
 openai==0.27.8
 
 # optional for chat with PDF
-langchain==0.0.329
-pypdf==3.17.0
+langchain==0.0.202
+pypdf==3.12.2
 # avoid textract, requires old six
 #textract==1.6.5
 

apps/language_models/langchain/make_db.py

@@ -1,4 +1,5 @@
 import os
+import fire
 
 from gpt_langchain import (
     path_to_docs,
@@ -201,3 +202,7 @@ def make_db_main(
     if verbose:
         print("DONE", flush=True)
     return db, collection_name
+
+
+if __name__ == "__main__":
+    fire.Fire(make_db_main)

apps/language_models/scripts/llama_ir_conversion_utils.py

@@ -1,442 +0,0 @@
-from pathlib import Path
-import argparse
-from argparse import RawTextHelpFormatter
-import re, gc
-
-"""
-    This script can be used as a standalone utility to convert IRs to dynamic + combine them.
-    Following are the various ways this script can be used :-
-        a. To convert a single Linalg IR to dynamic IR:
-            --dynamic --first_ir_path=<PATH TO FIRST IR>
-        b. To convert two Linalg IRs to dynamic IR:
-            --dynamic --first_ir_path=<PATH TO SECOND IR> --first_ir_path=<PATH TO SECOND IR>
-        c. To combine two Linalg IRs into one:
-            --combine --first_ir_path=<PATH TO FIRST IR> --second_ir_path=<PATH TO SECOND IR>
-        d. To convert both IRs into dynamic as well as combine the IRs:
-            --dynamic --combine --first_ir_path=<PATH TO FIRST IR> --second_ir_path=<PATH TO SECOND IR>
-
-    NOTE: For dynamic you'll also need to provide the following set of flags:-
-           i. For First Llama : --dynamic_input_size (DEFAULT: 19)
-          ii. For Second Llama: --model_name (DEFAULT: llama2_7b)
-                                --precision (DEFAULT: 'int4')
-          You may use --save_dynamic to also save the dynamic IR in option d above.
-          Else for option a. and b. the dynamic IR(s) will get saved by default.
-"""
-
-
-def combine_mlir_scripts(
-    first_vicuna_mlir,
-    second_vicuna_mlir,
-    output_name,
-    return_ir=True,
-):
-    print(f"[DEBUG] combining first and second mlir")
-    print(f"[DEBUG] output_name = {output_name}")
-    maps1 = []
-    maps2 = []
-    constants = set()
-    f1 = []
-    f2 = []
-
-    print(f"[DEBUG] processing first vicuna mlir")
-    first_vicuna_mlir = first_vicuna_mlir.splitlines()
-    while first_vicuna_mlir:
-        line = first_vicuna_mlir.pop(0)
-        if re.search("#map\d*\s*=", line):
-            maps1.append(line)
-        elif re.search("arith.constant", line):
-            constants.add(line)
-        elif not re.search("module", line):
-            line = re.sub("forward", "first_vicuna_forward", line)
-            f1.append(line)
-    f1 = f1[:-1]
-    del first_vicuna_mlir
-    gc.collect()
-
-    for i, map_line in enumerate(maps1):
-        map_var = map_line.split(" ")[0]
-        map_line = re.sub(f"{map_var}(?!\d)", map_var + "_0", map_line)
-        maps1[i] = map_line
-        f1 = [
-            re.sub(f"{map_var}(?!\d)", map_var + "_0", func_line)
-            for func_line in f1
-        ]
-
-    print(f"[DEBUG] processing second vicuna mlir")
-    second_vicuna_mlir = second_vicuna_mlir.splitlines()
-    while second_vicuna_mlir:
-        line = second_vicuna_mlir.pop(0)
-        if re.search("#map\d*\s*=", line):
-            maps2.append(line)
-        elif "global_seed" in line:
-            continue
-        elif re.search("arith.constant", line):
-            constants.add(line)
-        elif not re.search("module", line):
-            line = re.sub("forward", "second_vicuna_forward", line)
-            f2.append(line)
-    f2 = f2[:-1]
-    del second_vicuna_mlir
-    gc.collect()
-
-    for i, map_line in enumerate(maps2):
-        map_var = map_line.split(" ")[0]
-        map_line = re.sub(f"{map_var}(?!\d)", map_var + "_1", map_line)
-        maps2[i] = map_line
-        f2 = [
-            re.sub(f"{map_var}(?!\d)", map_var + "_1", func_line)
-            for func_line in f2
-        ]
-
-    module_start = 'module attributes {torch.debug_module_name = "_lambda"} {'
-    module_end = "}"
-
-    global_vars = []
-    vnames = []
-    global_var_loading1 = []
-    global_var_loading2 = []
-
-    print(f"[DEBUG] processing constants")
-    counter = 0
-    constants = list(constants)
-    while constants:
-        constant = constants.pop(0)
-        vname, vbody = constant.split("=")
-        vname = re.sub("%", "", vname)
-        vname = vname.strip()
-        vbody = re.sub("arith.constant", "", vbody)
-        vbody = vbody.strip()
-        if len(vbody.split(":")) < 2:
-            print(constant)
-        vdtype = vbody.split(":")[-1].strip()
-        fixed_vdtype = vdtype
-        if "c1_i64" in vname:
-            print(constant)
-            counter += 1
-        if counter == 2:
-            counter = 0
-            print("detected duplicate")
-            continue
-        vnames.append(vname)
-        if "true" not in vname:
-            global_vars.append(
-                f"ml_program.global private @{vname}({vbody}) : {fixed_vdtype}"
-            )
-            global_var_loading1.append(
-                f"\t\t%{vname} = ml_program.global_load_const @{vname} : {fixed_vdtype}"
-            )
-            global_var_loading2.append(
-                f"\t\t%{vname} = ml_program.global_load_const @{vname} : {fixed_vdtype}"
-            )
-        else:
-            global_vars.append(
-                f"ml_program.global private @{vname}({vbody}) : i1"
-            )
-            global_var_loading1.append(
-                f"\t\t%{vname} = ml_program.global_load_const @{vname} : i1"
-            )
-            global_var_loading2.append(
-                f"\t\t%{vname} = ml_program.global_load_const @{vname} : i1"
-            )
-
-    new_f1, new_f2 = [], []
-
-    print(f"[DEBUG] processing f1")
-    for line in f1:
-        if "func.func" in line:
-            new_f1.append(line)
-            for global_var in global_var_loading1:
-                new_f1.append(global_var)
-        else:
-            new_f1.append(line)
-
-    print(f"[DEBUG] processing f2")
-    for line in f2:
-        if "func.func" in line:
-            new_f2.append(line)
-            for global_var in global_var_loading2:
-                if (
-                    "c20_i64 = arith.addi %dim_i64, %c1_i64 : i64"
-                    in global_var
-                ):
-                    print(global_var)
-                new_f2.append(global_var)
-        else:
-            new_f2.append(line)
-
-    f1 = new_f1
-    f2 = new_f2
-
-    del new_f1
-    del new_f2
-    gc.collect()
-
-    print(
-        [
-            "c20_i64 = arith.addi %dim_i64, %c1_i64 : i64" in x
-            for x in [maps1, maps2, global_vars, f1, f2]
-        ]
-    )
-
-    # doing it this way rather than assembling the whole string
-    # to prevent OOM with 64GiB RAM when encoding the file.
-
-    print(f"[DEBUG] Saving mlir to {output_name}")
-    with open(output_name, "w+") as f_:
-        f_.writelines(line + "\n" for line in maps1)
-        f_.writelines(line + "\n" for line in maps2)
-        f_.writelines(line + "\n" for line in [module_start])
-        f_.writelines(line + "\n" for line in global_vars)
-        f_.writelines(line + "\n" for line in f1)
-        f_.writelines(line + "\n" for line in f2)
-        f_.writelines(line + "\n" for line in [module_end])
-
-    del maps1
-    del maps2
-    del module_start
-    del global_vars
-    del f1
-    del f2
-    del module_end
-    gc.collect()
-
-    if return_ir:
-        print(f"[DEBUG] Reading combined mlir back in")
-        with open(output_name, "rb") as f:
-            return f.read()
-
-
-def write_in_dynamic_inputs0(module, dynamic_input_size):
-    print("[DEBUG] writing dynamic inputs to first vicuna")
-    # Current solution for ensuring mlir files support dynamic inputs
-    # TODO: find a more elegant way to implement this
-    new_lines = []
-    module = module.splitlines()
-    while module:
-        line = module.pop(0)
-        line = re.sub(f"{dynamic_input_size}x", "?x", line)
-        if "?x" in line:
-            line = re.sub("tensor.empty\(\)", "tensor.empty(%dim)", line)
-        line = re.sub(f" {dynamic_input_size},", " %dim,", line)
-        if "tensor.empty" in line and "?x?" in line:
-            line = re.sub(
-                "tensor.empty\(%dim\)", "tensor.empty(%dim, %dim)", line
-            )
-        if "arith.cmpi" in line:
-            line = re.sub(f"c{dynamic_input_size}", "dim", line)
-        if "%0 = tensor.empty(%dim) : tensor<?xi64>" in line:
-            new_lines.append("%dim = tensor.dim %arg0, %c1 : tensor<1x?xi64>")
-        if "%dim = tensor.dim %arg0, %c1 : tensor<1x?xi64>" in line:
-            continue
-
-        new_lines.append(line)
-    return "\n".join(new_lines)
-
-
-def write_in_dynamic_inputs1(module, model_name, precision):
-    print("[DEBUG] writing dynamic inputs to second vicuna")
-
-    def remove_constant_dim(line):
-        if "c19_i64" in line:
-            line = re.sub("c19_i64", "dim_i64", line)
-        if "19x" in line:
-            line = re.sub("19x", "?x", line)
-            line = re.sub("tensor.empty\(\)", "tensor.empty(%dim)", line)
-        if "tensor.empty" in line and "?x?" in line:
-            line = re.sub(
-                "tensor.empty\(%dim\)",
-                "tensor.empty(%dim, %dim)",
-                line,
-            )
-        if "arith.cmpi" in line:
-            line = re.sub("c19", "dim", line)
-        if " 19," in line:
-            line = re.sub(" 19,", " %dim,", line)
-        if "x20x" in line or "<20x" in line:
-            line = re.sub("20x", "?x", line)
-            line = re.sub("tensor.empty\(\)", "tensor.empty(%dimp1)", line)
-        if " 20," in line:
-            line = re.sub(" 20,", " %dimp1,", line)
-        return line
-
-    module = module.splitlines()
-    new_lines = []
-
-    # Using a while loop and the pop method to avoid creating a copy of module
-    if "llama2_13b" in model_name:
-        pkv_tensor_shape = "tensor<1x40x?x128x"
-    elif "llama2_70b" in model_name:
-        pkv_tensor_shape = "tensor<1x8x?x128x"
-    else:
-        pkv_tensor_shape = "tensor<1x32x?x128x"
-    if precision in ["fp16", "int4", "int8"]:
-        pkv_tensor_shape += "f16>"
-    else:
-        pkv_tensor_shape += "f32>"
-
-    while module:
-        line = module.pop(0)
-        if "%c19_i64 = arith.constant 19 : i64" in line:
-            new_lines.append("%c2 = arith.constant 2 : index")
-            new_lines.append(
-                f"%dim_4_int = tensor.dim %arg1, %c2 : {pkv_tensor_shape}"
-            )
-            new_lines.append(
-                "%dim_i64 = arith.index_cast %dim_4_int : index to i64"
-            )
-            continue
-        if "%c2 = arith.constant 2 : index" in line:
-            continue
-        if "%c20_i64 = arith.constant 20 : i64" in line:
-            new_lines.append("%c1_i64 = arith.constant 1 : i64")
-            new_lines.append("%c20_i64 = arith.addi %dim_i64, %c1_i64 : i64")
-            new_lines.append(
-                "%dimp1 = arith.index_cast %c20_i64 : i64 to index"
-            )
-            continue
-        line = remove_constant_dim(line)
-        new_lines.append(line)
-
-    return "\n".join(new_lines)
-
-
-def save_dynamic_ir(ir_to_save, output_file):
-    if not ir_to_save:
-        return
-    # We only get string output from the dynamic conversion utility.
-    from contextlib import redirect_stdout
-
-    with open(output_file, "w") as f:
-        with redirect_stdout(f):
-            print(ir_to_save)
-
-
-if __name__ == "__main__":
-    parser = argparse.ArgumentParser(
-        prog="llama ir utility",
-        description="\tThis script can be used as a standalone utility to convert IRs to dynamic + combine them.\n"
-        + "\tFollowing are the various ways this script can be used :-\n"
-        + "\t\ta. To convert a single Linalg IR to dynamic IR:\n"
-        + "\t\t\t--dynamic --first_ir_path=<PATH TO FIRST IR>\n"
-        + "\t\tb. To convert two Linalg IRs to dynamic IR:\n"
-        + "\t\t\t--dynamic --first_ir_path=<PATH TO SECOND IR> --first_ir_path=<PATH TO SECOND IR>\n"
-        + "\t\tc. To combine two Linalg IRs into one:\n"
-        + "\t\t\t--combine --first_ir_path=<PATH TO FIRST IR> --second_ir_path=<PATH TO SECOND IR>\n"
-        + "\t\td. To convert both IRs into dynamic as well as combine the IRs:\n"
-        + "\t\t\t--dynamic --combine --first_ir_path=<PATH TO FIRST IR> --second_ir_path=<PATH TO SECOND IR>\n\n"
-        + "\tNOTE: For dynamic you'll also need to provide the following set of flags:-\n"
-        + "\t\t i. For First Llama : --dynamic_input_size (DEFAULT: 19)\n"
-        + "\t\tii. For Second Llama: --model_name (DEFAULT: llama2_7b)\n"
-        + "\t\t\t--precision (DEFAULT: 'int4')\n"
-        + "\t      You may use --save_dynamic to also save the dynamic IR in option d above.\n"
-        + "\t      Else for option a. and b. the dynamic IR(s) will get saved by default.\n",
-        formatter_class=RawTextHelpFormatter,
-    )
-    parser.add_argument(
-        "--precision",
-        "-p",
-        default="int4",
-        choices=["fp32", "fp16", "int8", "int4"],
-        help="Precision of the concerned IR",
-    )
-    parser.add_argument(
-        "--model_name",
-        type=str,
-        default="llama2_7b",
-        choices=["vicuna", "llama2_7b", "llama2_13b", "llama2_70b"],
-        help="Specify which model to run.",
-    )
-    parser.add_argument(
-        "--first_ir_path",
-        default=None,
-        help="path to first llama mlir file",
-    )
-    parser.add_argument(
-        "--second_ir_path",
-        default=None,
-        help="path to second llama mlir file",
-    )
-    parser.add_argument(
-        "--dynamic_input_size",
-        type=int,
-        default=19,
-        help="Specify the static input size to replace with dynamic dim.",
-    )
-    parser.add_argument(
-        "--dynamic",
-        default=False,
-        action=argparse.BooleanOptionalAction,
-        help="Converts the IR(s) to dynamic",
-    )
-    parser.add_argument(
-        "--save_dynamic",
-        default=False,
-        action=argparse.BooleanOptionalAction,
-        help="Save the individual IR(s) after converting to dynamic",
-    )
-    parser.add_argument(
-        "--combine",
-        default=False,
-        action=argparse.BooleanOptionalAction,
-        help="Converts the IR(s) to dynamic",
-    )
-
-    args, unknown = parser.parse_known_args()
-
-    dynamic = args.dynamic
-    combine = args.combine
-    assert (
-        dynamic or combine
-    ), "neither `dynamic` nor `combine` flag is turned on"
-    first_ir_path = args.first_ir_path
-    second_ir_path = args.second_ir_path
-    assert first_ir_path or second_ir_path, "no input ir has been provided"
-    if combine:
-        assert (
-            first_ir_path and second_ir_path
-        ), "you will need to provide both IRs to combine"
-    precision = args.precision
-    model_name = args.model_name
-    dynamic_input_size = args.dynamic_input_size
-    save_dynamic = args.save_dynamic
-
-    print(f"Dynamic conversion utility is turned {'ON' if dynamic else 'OFF'}")
-    print(f"Combining IR utility is turned {'ON' if combine else 'OFF'}")
-
-    if dynamic and not combine:
-        save_dynamic = True
-
-    first_ir = None
-    first_dynamic_ir_name = None
-    second_ir = None
-    second_dynamic_ir_name = None
-    if first_ir_path:
-        first_dynamic_ir_name = f"{Path(first_ir_path).stem}_dynamic"
-        with open(first_ir_path, "r") as f:
-            first_ir = f.read()
-    if second_ir_path:
-        second_dynamic_ir_name = f"{Path(second_ir_path).stem}_dynamic"
-        with open(second_ir_path, "r") as f:
-            second_ir = f.read()
-    if dynamic:
-        first_ir = (
-            write_in_dynamic_inputs0(first_ir, dynamic_input_size)
-            if first_ir
-            else None
-        )
-        second_ir = (
-            write_in_dynamic_inputs1(second_ir, model_name, precision)
-            if second_ir
-            else None
-        )
-        if save_dynamic:
-            save_dynamic_ir(first_ir, f"{first_dynamic_ir_name}.mlir")
-            save_dynamic_ir(second_ir, f"{second_dynamic_ir_name}.mlir")
-
-    if combine:
-        combine_mlir_scripts(
-            first_ir,
-            second_ir,
-            f"{model_name}_{precision}.mlir",
-            return_ir=False,
-        )

apps/language_models/scripts/stablelm.py

@@ -46,7 +46,6 @@ def compile_stableLM(
     model_vmfb_name,
     device="cuda",
     precision="fp32",
-    debug=False,
 ):
     from shark.shark_inference import SharkInference
 
@@ -93,7 +92,7 @@ def compile_stableLM(
     shark_module.compile()
 
     path = shark_module.save_module(
-        vmfb_path.parent.absolute(), vmfb_path.stem, debug=debug
+        vmfb_path.parent.absolute(), vmfb_path.stem
     )
     print("Saved vmfb at ", str(path))
 

apps/language_models/shark_llama_cli.spec

@@ -1,94 +0,0 @@
-# -*- mode: python ; coding: utf-8 -*-
-from PyInstaller.utils.hooks import collect_data_files
-from PyInstaller.utils.hooks import collect_submodules
-from PyInstaller.utils.hooks import copy_metadata
-
-import sys ; sys.setrecursionlimit(sys.getrecursionlimit() * 5)
-
-datas = []
-datas += collect_data_files('torch')
-datas += copy_metadata('torch')
-datas += copy_metadata('tqdm')
-datas += copy_metadata('regex')
-datas += copy_metadata('requests')
-datas += copy_metadata('packaging')
-datas += copy_metadata('filelock')
-datas += copy_metadata('numpy')
-datas += copy_metadata('tokenizers')
-datas += copy_metadata('importlib_metadata')
-datas += copy_metadata('torch-mlir')
-datas += copy_metadata('omegaconf')
-datas += copy_metadata('safetensors')
-datas += copy_metadata('huggingface-hub')
-datas += copy_metadata('sentencepiece')
-datas += copy_metadata("pyyaml")
-datas += collect_data_files("tokenizers")
-datas += collect_data_files("tiktoken")
-datas += collect_data_files("accelerate")
-datas += collect_data_files('diffusers')
-datas += collect_data_files('transformers')
-datas += collect_data_files('opencv-python')
-datas += collect_data_files('pytorch_lightning')
-datas += collect_data_files('skimage')
-datas += collect_data_files('gradio')
-datas += collect_data_files('gradio_client')
-datas += collect_data_files('iree')
-datas += collect_data_files('google-cloud-storage')
-datas += collect_data_files('py-cpuinfo')
-datas += collect_data_files("shark", include_py_files=True)
-datas += collect_data_files("timm", include_py_files=True)
-datas += collect_data_files("tqdm")
-datas += collect_data_files("tkinter")
-datas += collect_data_files("webview")
-datas += collect_data_files("sentencepiece")
-datas += collect_data_files("jsonschema")
-datas += collect_data_files("jsonschema_specifications")
-datas += collect_data_files("cpuinfo")
-datas += collect_data_files("langchain")
-
-binaries = []
-
-block_cipher = None
-
-hiddenimports = ['shark', 'shark.shark_inference', 'apps']
-hiddenimports += [x for x in collect_submodules("skimage") if "tests" not in x]
-hiddenimports += [x for x in collect_submodules("iree") if "tests" not in x]
-
-a = Analysis(
-    ['scripts/vicuna.py'],
-    pathex=['.'],
-    binaries=binaries,
-    datas=datas,
-    hiddenimports=hiddenimports,
-    hookspath=[],
-    hooksconfig={},
-    runtime_hooks=[],
-    excludes=[],
-    win_no_prefer_redirects=False,
-    win_private_assemblies=False,
-    cipher=block_cipher,
-    noarchive=False,
-)
-pyz = PYZ(a.pure, a.zipped_data, cipher=block_cipher)
-
-exe = EXE(
-    pyz,
-    a.scripts,
-    a.binaries,
-    a.zipfiles,
-    a.datas,
-    [],
-    name='shark_llama_cli',
-    debug=False,
-    bootloader_ignore_signals=False,
-    strip=False,
-    upx=True,
-    upx_exclude=[],
-    runtime_tmpdir=None,
-    console=True,
-    disable_windowed_traceback=False,
-    argv_emulation=False,
-    target_arch=None,
-    codesign_identity=None,
-    entitlements_file=None,
-)

apps/language_models/src/model_wrappers/falcon_sharded_model.py

@@ -1,598 +0,0 @@
-import torch
-from typing import Optional, Tuple
-
-
-class WordEmbeddingsLayer(torch.nn.Module):
-    def __init__(self, word_embedding_layer):
-        super().__init__()
-        self.model = word_embedding_layer
-
-    def forward(self, input_ids):
-        output = self.model.forward(input=input_ids)
-        return output
-
-
-class CompiledWordEmbeddingsLayer(torch.nn.Module):
-    def __init__(self, compiled_word_embedding_layer):
-        super().__init__()
-        self.model = compiled_word_embedding_layer
-
-    def forward(self, input_ids):
-        input_ids = input_ids.detach().numpy()
-        new_input_ids = self.model("forward", input_ids)
-        new_input_ids = new_input_ids.reshape(
-            [1, new_input_ids.shape[0], new_input_ids.shape[1]]
-        )
-        return torch.tensor(new_input_ids)
-
-
-class LNFEmbeddingLayer(torch.nn.Module):
-    def __init__(self, ln_f):
-        super().__init__()
-        self.model = ln_f
-
-    def forward(self, hidden_states):
-        output = self.model.forward(input=hidden_states)
-        return output
-
-
-class CompiledLNFEmbeddingLayer(torch.nn.Module):
-    def __init__(self, ln_f):
-        super().__init__()
-        self.model = ln_f
-
-    def forward(self, hidden_states):
-        hidden_states = hidden_states.detach().numpy()
-        new_hidden_states = self.model("forward", (hidden_states,))
-
-        return torch.tensor(new_hidden_states)
-
-
-class LMHeadEmbeddingLayer(torch.nn.Module):
-    def __init__(self, embedding_layer):
-        super().__init__()
-        self.model = embedding_layer
-
-    def forward(self, hidden_states):
-        output = self.model.forward(input=hidden_states)
-        return output
-
-
-class CompiledLMHeadEmbeddingLayer(torch.nn.Module):
-    def __init__(self, lm_head):
-        super().__init__()
-        self.model = lm_head
-
-    def forward(self, hidden_states):
-        hidden_states = hidden_states.detach().numpy()
-        new_hidden_states = self.model("forward", (hidden_states,))
-        return torch.tensor(new_hidden_states)
-
-
-class DecoderLayer(torch.nn.Module):
-    def __init__(self, decoder_layer_model, falcon_variant):
-        super().__init__()
-        self.model = decoder_layer_model
-
-    def forward(self, hidden_states, attention_mask):
-        output = self.model.forward(
-            hidden_states=hidden_states,
-            alibi=None,
-            attention_mask=attention_mask,
-            use_cache=True,
-        )
-        return (output[0], output[1][0], output[1][1])
-
-
-class CompiledDecoderLayer(torch.nn.Module):
-    def __init__(
-        self, layer_id, device_idx, falcon_variant, device, precision
-    ):
-        super().__init__()
-        self.layer_id = layer_id
-        self.device_index = device_idx
-        self.falcon_variant = falcon_variant
-        self.device = device
-        self.precision = precision
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        attention_mask: torch.Tensor,
-        alibi: torch.Tensor = None,
-        layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
-        head_mask: Optional[torch.Tensor] = None,
-        use_cache: bool = False,
-        output_attentions: bool = False,
-    ):
-        import gc
-
-        torch.cuda.empty_cache()
-        gc.collect()
-        from pathlib import Path
-        from apps.language_models.utils import get_vmfb_from_path
-
-        self.falcon_vmfb_path = Path(
-            f"falcon_{self.falcon_variant}_layer_{self.layer_id}_{self.precision}_{self.device}.vmfb"
-        )
-        print("vmfb path for layer: ", self.falcon_vmfb_path)
-        self.model = get_vmfb_from_path(
-            self.falcon_vmfb_path,
-            self.device,
-            "linalg",
-            device_id=self.device_index,
-        )
-        if self.model is None:
-            raise ValueError("Layer vmfb not found")
-
-        hidden_states = hidden_states.to(torch.float32).detach().numpy()
-        attention_mask = attention_mask.detach().numpy()
-
-        if alibi is not None or layer_past is not None:
-            raise ValueError("Past Key Values and alibi should be None")
-        else:
-            new_hidden_states, pkv1, pkv2 = self.model(
-                "forward",
-                (
-                    hidden_states,
-                    attention_mask,
-                ),
-            )
-            del self.model
-
-            return tuple(
-                [
-                    torch.tensor(new_hidden_states),
-                    tuple(
-                        [
-                            torch.tensor(pkv1),
-                            torch.tensor(pkv2),
-                        ]
-                    ),
-                ]
-            )
-
-
-class EightDecoderLayer(torch.nn.Module):
-    def __init__(self, decoder_layer_model, falcon_variant):
-        super().__init__()
-        self.model = decoder_layer_model
-        self.falcon_variant = falcon_variant
-
-    def forward(self, hidden_states, attention_mask):
-        new_pkvs = []
-        for layer in self.model:
-            outputs = layer(
-                hidden_states=hidden_states,
-                alibi=None,
-                attention_mask=attention_mask,
-                use_cache=True,
-            )
-            hidden_states = outputs[0]
-            new_pkvs.append(
-                (
-                    outputs[-1][0],
-                    outputs[-1][1],
-                )
-            )
-        if self.falcon_variant == "7b":
-            (
-                (new_pkv00, new_pkv01),
-                (new_pkv10, new_pkv11),
-                (new_pkv20, new_pkv21),
-                (new_pkv30, new_pkv31),
-                (new_pkv40, new_pkv41),
-                (new_pkv50, new_pkv51),
-                (new_pkv60, new_pkv61),
-                (new_pkv70, new_pkv71),
-            ) = new_pkvs
-            result = (
-                hidden_states,
-                new_pkv00,
-                new_pkv01,
-                new_pkv10,
-                new_pkv11,
-                new_pkv20,
-                new_pkv21,
-                new_pkv30,
-                new_pkv31,
-                new_pkv40,
-                new_pkv41,
-                new_pkv50,
-                new_pkv51,
-                new_pkv60,
-                new_pkv61,
-                new_pkv70,
-                new_pkv71,
-            )
-        elif self.falcon_variant == "40b":
-            (
-                (new_pkv00, new_pkv01),
-                (new_pkv10, new_pkv11),
-                (new_pkv20, new_pkv21),
-                (new_pkv30, new_pkv31),
-                (new_pkv40, new_pkv41),
-                (new_pkv50, new_pkv51),
-                (new_pkv60, new_pkv61),
-                (new_pkv70, new_pkv71),
-                (new_pkv80, new_pkv81),
-                (new_pkv90, new_pkv91),
-                (new_pkv100, new_pkv101),
-                (new_pkv110, new_pkv111),
-                (new_pkv120, new_pkv121),
-                (new_pkv130, new_pkv131),
-                (new_pkv140, new_pkv141),
-            ) = new_pkvs
-            result = (
-                hidden_states,
-                new_pkv00,
-                new_pkv01,
-                new_pkv10,
-                new_pkv11,
-                new_pkv20,
-                new_pkv21,
-                new_pkv30,
-                new_pkv31,
-                new_pkv40,
-                new_pkv41,
-                new_pkv50,
-                new_pkv51,
-                new_pkv60,
-                new_pkv61,
-                new_pkv70,
-                new_pkv71,
-                new_pkv80,
-                new_pkv81,
-                new_pkv90,
-                new_pkv91,
-                new_pkv100,
-                new_pkv101,
-                new_pkv110,
-                new_pkv111,
-                new_pkv120,
-                new_pkv121,
-                new_pkv130,
-                new_pkv131,
-                new_pkv140,
-                new_pkv141,
-            )
-        elif self.falcon_variant == "180b":
-            (
-                (new_pkv00, new_pkv01),
-                (new_pkv10, new_pkv11),
-                (new_pkv20, new_pkv21),
-                (new_pkv30, new_pkv31),
-                (new_pkv40, new_pkv41),
-                (new_pkv50, new_pkv51),
-                (new_pkv60, new_pkv61),
-                (new_pkv70, new_pkv71),
-                (new_pkv80, new_pkv81),
-                (new_pkv90, new_pkv91),
-                (new_pkv100, new_pkv101),
-                (new_pkv110, new_pkv111),
-                (new_pkv120, new_pkv121),
-                (new_pkv130, new_pkv131),
-                (new_pkv140, new_pkv141),
-                (new_pkv150, new_pkv151),
-                (new_pkv160, new_pkv161),
-                (new_pkv170, new_pkv171),
-                (new_pkv180, new_pkv181),
-                (new_pkv190, new_pkv191),
-            ) = new_pkvs
-            result = (
-                hidden_states,
-                new_pkv00,
-                new_pkv01,
-                new_pkv10,
-                new_pkv11,
-                new_pkv20,
-                new_pkv21,
-                new_pkv30,
-                new_pkv31,
-                new_pkv40,
-                new_pkv41,
-                new_pkv50,
-                new_pkv51,
-                new_pkv60,
-                new_pkv61,
-                new_pkv70,
-                new_pkv71,
-                new_pkv80,
-                new_pkv81,
-                new_pkv90,
-                new_pkv91,
-                new_pkv100,
-                new_pkv101,
-                new_pkv110,
-                new_pkv111,
-                new_pkv120,
-                new_pkv121,
-                new_pkv130,
-                new_pkv131,
-                new_pkv140,
-                new_pkv141,
-                new_pkv150,
-                new_pkv151,
-                new_pkv160,
-                new_pkv161,
-                new_pkv170,
-                new_pkv171,
-                new_pkv180,
-                new_pkv181,
-                new_pkv190,
-                new_pkv191,
-            )
-        else:
-            raise ValueError(
-                "Unsupported Falcon variant: ", self.falcon_variant
-            )
-        return result
-
-
-class CompiledEightDecoderLayer(torch.nn.Module):
-    def __init__(
-        self, layer_id, device_idx, falcon_variant, device, precision
-    ):
-        super().__init__()
-        self.layer_id = layer_id
-        self.device_index = device_idx
-        self.falcon_variant = falcon_variant
-        self.device = device
-        self.precision = precision
-
-    def forward(
-        self,
-        hidden_states: torch.Tensor,
-        attention_mask: torch.Tensor,
-        alibi: torch.Tensor = None,
-        layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
-        head_mask: Optional[torch.Tensor] = None,
-        use_cache: bool = False,
-        output_attentions: bool = False,
-    ):
-        import gc
-
-        torch.cuda.empty_cache()
-        gc.collect()
-        from pathlib import Path
-        from apps.language_models.utils import get_vmfb_from_path
-
-        self.falcon_vmfb_path = Path(
-            f"falcon_{self.falcon_variant}_layer_{self.layer_id}_{self.precision}_{self.device}.vmfb"
-        )
-        print("vmfb path for layer: ", self.falcon_vmfb_path)
-        self.model = get_vmfb_from_path(
-            self.falcon_vmfb_path,
-            self.device,
-            "linalg",
-            device_id=self.device_index,
-        )
-        if self.model is None:
-            raise ValueError("Layer vmfb not found")
-
-        hidden_states = hidden_states.to(torch.float32).detach().numpy()
-        attention_mask = attention_mask.detach().numpy()
-
-        if alibi is not None or layer_past is not None:
-            raise ValueError("Past Key Values and alibi should be None")
-        else:
-            output = self.model(
-                "forward",
-                (
-                    hidden_states,
-                    attention_mask,
-                ),
-            )
-            del self.model
-
-        if self.falcon_variant == "7b":
-            result = (
-                torch.tensor(output[0]),
-                (
-                    torch.tensor(output[1]),
-                    torch.tensor(output[2]),
-                ),
-                (
-                    torch.tensor(output[3]),
-                    torch.tensor(output[4]),
-                ),
-                (
-                    torch.tensor(output[5]),
-                    torch.tensor(output[6]),
-                ),
-                (
-                    torch.tensor(output[7]),
-                    torch.tensor(output[8]),
-                ),
-                (
-                    torch.tensor(output[9]),
-                    torch.tensor(output[10]),
-                ),
-                (
-                    torch.tensor(output[11]),
-                    torch.tensor(output[12]),
-                ),
-                (
-                    torch.tensor(output[13]),
-                    torch.tensor(output[14]),
-                ),
-                (
-                    torch.tensor(output[15]),
-                    torch.tensor(output[16]),
-                ),
-            )
-        elif self.falcon_variant == "40b":
-            result = (
-                torch.tensor(output[0]),
-                (
-                    torch.tensor(output[1]),
-                    torch.tensor(output[2]),
-                ),
-                (
-                    torch.tensor(output[3]),
-                    torch.tensor(output[4]),
-                ),
-                (
-                    torch.tensor(output[5]),
-                    torch.tensor(output[6]),
-                ),
-                (
-                    torch.tensor(output[7]),
-                    torch.tensor(output[8]),
-                ),
-                (
-                    torch.tensor(output[9]),
-                    torch.tensor(output[10]),
-                ),
-                (
-                    torch.tensor(output[11]),
-                    torch.tensor(output[12]),
-                ),
-                (
-                    torch.tensor(output[13]),
-                    torch.tensor(output[14]),
-                ),
-                (
-                    torch.tensor(output[15]),
-                    torch.tensor(output[16]),
-                ),
-                (
-                    torch.tensor(output[17]),
-                    torch.tensor(output[18]),
-                ),
-                (
-                    torch.tensor(output[19]),
-                    torch.tensor(output[20]),
-                ),
-                (
-                    torch.tensor(output[21]),
-                    torch.tensor(output[22]),
-                ),
-                (
-                    torch.tensor(output[23]),
-                    torch.tensor(output[24]),
-                ),
-                (
-                    torch.tensor(output[25]),
-                    torch.tensor(output[26]),
-                ),
-                (
-                    torch.tensor(output[27]),
-                    torch.tensor(output[28]),
-                ),
-                (
-                    torch.tensor(output[29]),
-                    torch.tensor(output[30]),
-                ),
-            )
-        elif self.falcon_variant == "180b":
-            result = (
-                torch.tensor(output[0]),
-                (
-                    torch.tensor(output[1]),
-                    torch.tensor(output[2]),
-                ),
-                (
-                    torch.tensor(output[3]),
-                    torch.tensor(output[4]),
-                ),
-                (
-                    torch.tensor(output[5]),
-                    torch.tensor(output[6]),
-                ),
-                (
-                    torch.tensor(output[7]),
-                    torch.tensor(output[8]),
-                ),
-                (
-                    torch.tensor(output[9]),
-                    torch.tensor(output[10]),
-                ),
-                (
-                    torch.tensor(output[11]),
-                    torch.tensor(output[12]),
-                ),
-                (
-                    torch.tensor(output[13]),
-                    torch.tensor(output[14]),
-                ),
-                (
-                    torch.tensor(output[15]),
-                    torch.tensor(output[16]),
-                ),
-                (
-                    torch.tensor(output[17]),
-                    torch.tensor(output[18]),
-                ),
-                (
-                    torch.tensor(output[19]),
-                    torch.tensor(output[20]),
-                ),
-                (
-                    torch.tensor(output[21]),
-                    torch.tensor(output[22]),
-                ),
-                (
-                    torch.tensor(output[23]),
-                    torch.tensor(output[24]),
-                ),
-                (
-                    torch.tensor(output[25]),
-                    torch.tensor(output[26]),
-                ),
-                (
-                    torch.tensor(output[27]),
-                    torch.tensor(output[28]),
-                ),
-                (
-                    torch.tensor(output[29]),
-                    torch.tensor(output[30]),
-                ),
-                (
-                    torch.tensor(output[31]),
-                    torch.tensor(output[32]),
-                ),
-                (
-                    torch.tensor(output[33]),
-                    torch.tensor(output[34]),
-                ),
-                (
-                    torch.tensor(output[35]),
-                    torch.tensor(output[36]),
-                ),
-                (
-                    torch.tensor(output[37]),
-                    torch.tensor(output[38]),
-                ),
-                (
-                    torch.tensor(output[39]),
-                    torch.tensor(output[40]),
-                ),
-            )
-        else:
-            raise ValueError(
-                "Unsupported Falcon variant: ", self.falcon_variant
-            )
-        return result
-
-
-class ShardedFalconModel:
-    def __init__(self, model, layers, word_embeddings, ln_f, lm_head):
-        super().__init__()
-        self.model = model
-        self.model.transformer.h = torch.nn.modules.container.ModuleList(
-            layers
-        )
-        self.model.transformer.word_embeddings = word_embeddings
-        self.model.transformer.ln_f = ln_f
-        self.model.lm_head = lm_head
-
-    def forward(
-        self,
-        input_ids,
-        attention_mask=None,
-    ):
-        return self.model.forward(
-            input_ids=input_ids,
-            attention_mask=attention_mask,
-        ).logits[:, -1, :]

apps/language_models/src/model_wrappers/vicuna4.py

@@ -1,876 +0,0 @@
-import argparse
-import json
-import re
-from io import BytesIO
-from pathlib import Path
-from tqdm import tqdm
-from typing import List, Optional, Tuple, Union
-import numpy as np
-import iree.runtime
-import itertools
-import subprocess
-
-import torch
-import torch_mlir
-from torch_mlir import TensorPlaceholder
-from torch_mlir.compiler_utils import run_pipeline_with_repro_report
-from transformers import (
-    AutoTokenizer,
-    AutoModelForCausalLM,
-    LlamaPreTrainedModel,
-)
-from transformers.modeling_outputs import (
-    BaseModelOutputWithPast,
-    CausalLMOutputWithPast,
-    SequenceClassifierOutputWithPast,
-)
-from transformers.modeling_utils import PreTrainedModel
-from transformers.utils import (
-    add_start_docstrings,
-    add_start_docstrings_to_model_forward,
-    logging,
-    replace_return_docstrings,
-)
-
-from apps.language_models.src.pipelines.SharkLLMBase import SharkLLMBase
-from apps.language_models.src.model_wrappers.vicuna_sharded_model import (
-    FirstVicunaLayer,
-    SecondVicunaLayer,
-    CompiledVicunaLayer,
-    ShardedVicunaModel,
-    LMHead,
-    LMHeadCompiled,
-    VicunaEmbedding,
-    VicunaEmbeddingCompiled,
-    VicunaNorm,
-    VicunaNormCompiled,
-)
-from apps.language_models.src.model_wrappers.vicuna_model import (
-    FirstVicuna,
-    SecondVicuna7B,
-)
-from apps.language_models.utils import (
-    get_vmfb_from_path,
-)
-from shark.shark_downloader import download_public_file
-from shark.shark_importer import get_f16_inputs
-from shark.shark_inference import SharkInference
-
-from transformers.models.llama.configuration_llama import LlamaConfig
-from transformers.models.llama.modeling_llama import (
-    LlamaDecoderLayer,
-    LlamaRMSNorm,
-    _make_causal_mask,
-    _expand_mask,
-)
-from torch import nn
-from time import time
-
-
-class LlamaModel(LlamaPreTrainedModel):
-    """
-    Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`]
-
-    Args:
-        config: LlamaConfig
-    """
-
-    def __init__(self, config: LlamaConfig):
-        super().__init__(config)
-        self.padding_idx = config.pad_token_id
-        self.vocab_size = config.vocab_size
-
-        self.embed_tokens = nn.Embedding(
-            config.vocab_size, config.hidden_size, self.padding_idx
-        )
-        self.layers = nn.ModuleList(
-            [
-                LlamaDecoderLayer(config)
-                for _ in range(config.num_hidden_layers)
-            ]
-        )
-        self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
-
-        self.gradient_checkpointing = False
-        # Initialize weights and apply final processing
-        self.post_init()
-
-    def get_input_embeddings(self):
-        return self.embed_tokens
-
-    def set_input_embeddings(self, value):
-        self.embed_tokens = value
-
-    # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask
-    def _prepare_decoder_attention_mask(
-        self,
-        attention_mask,
-        input_shape,
-        inputs_embeds,
-        past_key_values_length,
-    ):
-        # create causal mask
-        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
-        combined_attention_mask = None
-        if input_shape[-1] > 1:
-            combined_attention_mask = _make_causal_mask(
-                input_shape,
-                inputs_embeds.dtype,
-                device=inputs_embeds.device,
-                past_key_values_length=past_key_values_length,
-            )
-
-        if attention_mask is not None:
-            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
-            expanded_attn_mask = _expand_mask(
-                attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]
-            ).to(inputs_embeds.device)
-            combined_attention_mask = (
-                expanded_attn_mask
-                if combined_attention_mask is None
-                else expanded_attn_mask + combined_attention_mask
-            )
-
-        return combined_attention_mask
-
-    def forward(
-        self,
-        input_ids: torch.LongTensor = None,
-        attention_mask: Optional[torch.Tensor] = None,
-        position_ids: Optional[torch.LongTensor] = None,
-        past_key_values: Optional[List[torch.FloatTensor]] = None,
-        inputs_embeds: Optional[torch.FloatTensor] = None,
-        use_cache: Optional[bool] = None,
-        output_attentions: Optional[bool] = None,
-        output_hidden_states: Optional[bool] = None,
-        return_dict: Optional[bool] = None,
-    ):
-        t1 = time()
-        output_attentions = (
-            output_attentions
-            if output_attentions is not None
-            else self.config.output_attentions
-        )
-        output_hidden_states = (
-            output_hidden_states
-            if output_hidden_states is not None
-            else self.config.output_hidden_states
-        )
-        use_cache = (
-            use_cache if use_cache is not None else self.config.use_cache
-        )
-
-        return_dict = (
-            return_dict
-            if return_dict is not None
-            else self.config.use_return_dict
-        )
-
-        # retrieve input_ids and inputs_embeds
-        if input_ids is not None and inputs_embeds is not None:
-            raise ValueError(
-                "You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time"
-            )
-        elif input_ids is not None:
-            batch_size, seq_length = input_ids.shape
-        elif inputs_embeds is not None:
-            batch_size, seq_length, _ = inputs_embeds.shape
-        else:
-            raise ValueError(
-                "You have to specify either decoder_input_ids or decoder_inputs_embeds"
-            )
-
-        seq_length_with_past = seq_length
-        past_key_values_length = 0
-
-        if past_key_values is not None:
-            past_key_values_length = past_key_values[0][0].shape[2]
-            seq_length_with_past = (
-                seq_length_with_past + past_key_values_length
-            )
-
-        if position_ids is None:
-            device = (
-                input_ids.device
-                if input_ids is not None
-                else inputs_embeds.device
-            )
-            position_ids = torch.arange(
-                past_key_values_length,
-                seq_length + past_key_values_length,
-                dtype=torch.long,
-                device=device,
-            )
-            position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
-        else:
-            position_ids = position_ids.view(-1, seq_length).long()
-
-        if inputs_embeds is None:
-            inputs_embeds = self.embed_tokens(input_ids)
-        # embed positions
-        if attention_mask is None:
-            attention_mask = torch.ones(
-                (batch_size, seq_length_with_past),
-                dtype=torch.bool,
-                device=inputs_embeds.device,
-            )
-
-        attention_mask = self._prepare_decoder_attention_mask(
-            attention_mask,
-            (batch_size, seq_length),
-            inputs_embeds,
-            past_key_values_length,
-        )
-
-        hidden_states = inputs_embeds
-
-        # decoder layers
-        all_hidden_states = () if output_hidden_states else None
-        all_self_attns = () if output_attentions else None
-        next_decoder_cache = () if use_cache else None
-
-        for idx, decoder_layer in enumerate(self.compressedlayers):
-            if output_hidden_states:
-                all_hidden_states += (hidden_states,)
-
-            past_key_value = (
-                past_key_values[8 * idx : 8 * (idx + 1)]
-                if past_key_values is not None
-                else None
-            )
-
-            if self.gradient_checkpointing and self.training:
-
-                def create_custom_forward(module):
-                    def custom_forward(*inputs):
-                        # None for past_key_value
-                        return module(*inputs, output_attentions, None)
-
-                    return custom_forward
-
-                layer_outputs = torch.utils.checkpoint.checkpoint(
-                    create_custom_forward(decoder_layer),
-                    hidden_states,
-                    attention_mask,
-                    position_ids,
-                    None,
-                )
-            else:
-                layer_outputs = decoder_layer.forward(
-                    hidden_states,
-                    attention_mask=attention_mask,
-                    position_ids=position_ids,
-                    past_key_value=past_key_value,
-                    output_attentions=output_attentions,
-                    use_cache=use_cache,
-                )
-
-            hidden_states = layer_outputs[0]
-
-            if use_cache:
-                next_decoder_cache += (layer_outputs[1:],)
-
-            if output_attentions:
-                all_self_attns += (layer_outputs[1],)
-
-        try:
-            hidden_states = np.asarray(hidden_states, hidden_states.dtype)
-        except:
-            _ = 10
-
-        hidden_states = self.norm(hidden_states)
-
-        # add hidden states from the last decoder layer
-        if output_hidden_states:
-            all_hidden_states += (hidden_states,)
-
-        next_cache = next_decoder_cache if use_cache else None
-        next_cache = tuple(itertools.chain.from_iterable(next_cache))
-        print(f"Token generated in {time() - t1} seconds")
-        if not return_dict:
-            return tuple(
-                v
-                for v in [
-                    hidden_states,
-                    next_cache,
-                    all_hidden_states,
-                    all_self_attns,
-                ]
-                if v is not None
-            )
-        return BaseModelOutputWithPast(
-            last_hidden_state=hidden_states,
-            past_key_values=next_cache,
-            hidden_states=all_hidden_states,
-            attentions=all_self_attns,
-        )
-
-
-class EightLayerLayerSV(torch.nn.Module):
-    def __init__(self, layers):
-        super().__init__()
-        assert len(layers) == 8
-        self.layers = layers
-
-    def forward(
-        self,
-        hidden_states,
-        attention_mask,
-        position_ids,
-        pkv00,
-        pkv01,
-        pkv10,
-        pkv11,
-        pkv20,
-        pkv21,
-        pkv30,
-        pkv31,
-        pkv40,
-        pkv41,
-        pkv50,
-        pkv51,
-        pkv60,
-        pkv61,
-        pkv70,
-        pkv71,
-    ):
-        pkvs = [
-            (pkv00, pkv01),
-            (pkv10, pkv11),
-            (pkv20, pkv21),
-            (pkv30, pkv31),
-            (pkv40, pkv41),
-            (pkv50, pkv51),
-            (pkv60, pkv61),
-            (pkv70, pkv71),
-        ]
-        new_pkvs = []
-        for layer, pkv in zip(self.layers, pkvs):
-            outputs = layer(
-                hidden_states,
-                attention_mask=attention_mask,
-                position_ids=position_ids,
-                past_key_value=(
-                    pkv[0],
-                    pkv[1],
-                ),
-                use_cache=True,
-            )
-
-            hidden_states = outputs[0]
-            new_pkvs.append(
-                (
-                    outputs[-1][0],
-                    outputs[-1][1],
-                )
-            )
-        (
-            (new_pkv00, new_pkv01),
-            (new_pkv10, new_pkv11),
-            (new_pkv20, new_pkv21),
-            (new_pkv30, new_pkv31),
-            (new_pkv40, new_pkv41),
-            (new_pkv50, new_pkv51),
-            (new_pkv60, new_pkv61),
-            (new_pkv70, new_pkv71),
-        ) = new_pkvs
-        return (
-            hidden_states,
-            new_pkv00,
-            new_pkv01,
-            new_pkv10,
-            new_pkv11,
-            new_pkv20,
-            new_pkv21,
-            new_pkv30,
-            new_pkv31,
-            new_pkv40,
-            new_pkv41,
-            new_pkv50,
-            new_pkv51,
-            new_pkv60,
-            new_pkv61,
-            new_pkv70,
-            new_pkv71,
-        )
-
-
-class EightLayerLayerFV(torch.nn.Module):
-    def __init__(self, layers):
-        super().__init__()
-        assert len(layers) == 8
-        self.layers = layers
-
-    def forward(self, hidden_states, attention_mask, position_ids):
-        new_pkvs = []
-        for layer in self.layers:
-            outputs = layer(
-                hidden_states,
-                attention_mask=attention_mask,
-                position_ids=position_ids,
-                past_key_value=None,
-                use_cache=True,
-            )
-
-            hidden_states = outputs[0]
-            new_pkvs.append(
-                (
-                    outputs[-1][0],
-                    outputs[-1][1],
-                )
-            )
-        (
-            (new_pkv00, new_pkv01),
-            (new_pkv10, new_pkv11),
-            (new_pkv20, new_pkv21),
-            (new_pkv30, new_pkv31),
-            (new_pkv40, new_pkv41),
-            (new_pkv50, new_pkv51),
-            (new_pkv60, new_pkv61),
-            (new_pkv70, new_pkv71),
-        ) = new_pkvs
-        return (
-            hidden_states,
-            new_pkv00,
-            new_pkv01,
-            new_pkv10,
-            new_pkv11,
-            new_pkv20,
-            new_pkv21,
-            new_pkv30,
-            new_pkv31,
-            new_pkv40,
-            new_pkv41,
-            new_pkv50,
-            new_pkv51,
-            new_pkv60,
-            new_pkv61,
-            new_pkv70,
-            new_pkv71,
-        )
-
-
-class CompiledEightLayerLayerSV(torch.nn.Module):
-    def __init__(self, model):
-        super().__init__()
-        self.model = model
-
-    def forward(
-        self,
-        hidden_states,
-        attention_mask,
-        position_ids,
-        past_key_value,
-        output_attentions=False,
-        use_cache=True,
-    ):
-        hidden_states = hidden_states.detach()
-        attention_mask = attention_mask.detach()
-        position_ids = position_ids.detach()
-        (
-            (pkv00, pkv01),
-            (pkv10, pkv11),
-            (pkv20, pkv21),
-            (pkv30, pkv31),
-            (pkv40, pkv41),
-            (pkv50, pkv51),
-            (pkv60, pkv61),
-            (pkv70, pkv71),
-        ) = past_key_value
-        pkv00 = pkv00.detatch()
-        pkv01 = pkv01.detatch()
-        pkv10 = pkv10.detatch()
-        pkv11 = pkv11.detatch()
-        pkv20 = pkv20.detatch()
-        pkv21 = pkv21.detatch()
-        pkv30 = pkv30.detatch()
-        pkv31 = pkv31.detatch()
-        pkv40 = pkv40.detatch()
-        pkv41 = pkv41.detatch()
-        pkv50 = pkv50.detatch()
-        pkv51 = pkv51.detatch()
-        pkv60 = pkv60.detatch()
-        pkv61 = pkv61.detatch()
-        pkv70 = pkv70.detatch()
-        pkv71 = pkv71.detatch()
-
-        output = self.model(
-            "forward",
-            (
-                hidden_states,
-                attention_mask,
-                position_ids,
-                pkv00,
-                pkv01,
-                pkv10,
-                pkv11,
-                pkv20,
-                pkv21,
-                pkv30,
-                pkv31,
-                pkv40,
-                pkv41,
-                pkv50,
-                pkv51,
-                pkv60,
-                pkv61,
-                pkv70,
-                pkv71,
-            ),
-            send_to_host=False,
-        )
-        return (
-            output[0],
-            (output[1][0], output[1][1]),
-            (output[2][0], output[2][1]),
-            (output[3][0], output[3][1]),
-            (output[4][0], output[4][1]),
-            (output[5][0], output[5][1]),
-            (output[6][0], output[6][1]),
-            (output[7][0], output[7][1]),
-            (output[8][0], output[8][1]),
-        )
-
-
-def forward_compressed(
-    self,
-    input_ids: torch.LongTensor = None,
-    attention_mask: Optional[torch.Tensor] = None,
-    position_ids: Optional[torch.LongTensor] = None,
-    past_key_values: Optional[List[torch.FloatTensor]] = None,
-    inputs_embeds: Optional[torch.FloatTensor] = None,
-    use_cache: Optional[bool] = None,
-    output_attentions: Optional[bool] = None,
-    output_hidden_states: Optional[bool] = None,
-    return_dict: Optional[bool] = None,
-):
-    output_attentions = (
-        output_attentions
-        if output_attentions is not None
-        else self.config.output_attentions
-    )
-    output_hidden_states = (
-        output_hidden_states
-        if output_hidden_states is not None
-        else self.config.output_hidden_states
-    )
-    use_cache = use_cache if use_cache is not None else self.config.use_cache
-
-    return_dict = (
-        return_dict if return_dict is not None else self.config.use_return_dict
-    )
-
-    # retrieve input_ids and inputs_embeds
-    if input_ids is not None and inputs_embeds is not None:
-        raise ValueError(
-            "You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time"
-        )
-    elif input_ids is not None:
-        batch_size, seq_length = input_ids.shape
-    elif inputs_embeds is not None:
-        batch_size, seq_length, _ = inputs_embeds.shape
-    else:
-        raise ValueError(
-            "You have to specify either decoder_input_ids or decoder_inputs_embeds"
-        )
-
-    seq_length_with_past = seq_length
-    past_key_values_length = 0
-
-    if past_key_values is not None:
-        past_key_values_length = past_key_values[0][0].shape[2]
-        seq_length_with_past = seq_length_with_past + past_key_values_length
-
-    if position_ids is None:
-        device = (
-            input_ids.device if input_ids is not None else inputs_embeds.device
-        )
-        position_ids = torch.arange(
-            past_key_values_length,
-            seq_length + past_key_values_length,
-            dtype=torch.long,
-            device=device,
-        )
-        position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
-    else:
-        position_ids = position_ids.view(-1, seq_length).long()
-
-    if inputs_embeds is None:
-        inputs_embeds = self.embed_tokens(input_ids)
-    # embed positions
-    if attention_mask is None:
-        attention_mask = torch.ones(
-            (batch_size, seq_length_with_past),
-            dtype=torch.bool,
-            device=inputs_embeds.device,
-        )
-    attention_mask = self._prepare_decoder_attention_mask(
-        attention_mask,
-        (batch_size, seq_length),
-        inputs_embeds,
-        past_key_values_length,
-    )
-
-    hidden_states = inputs_embeds
-
-    # decoder layers
-    all_hidden_states = () if output_hidden_states else None
-    all_self_attns = () if output_attentions else None
-    next_decoder_cache = () if use_cache else None
-
-    for idx, decoder_layer in enumerate(self.compressedlayers):
-        if output_hidden_states:
-            all_hidden_states += (hidden_states,)
-
-        past_key_value = (
-            past_key_values[8 * idx : 8 * (idx + 1)]
-            if past_key_values is not None
-            else None
-        )
-
-        if self.gradient_checkpointing and self.training:
-
-            def create_custom_forward(module):
-                def custom_forward(*inputs):
-                    # None for past_key_value
-                    return module(*inputs, output_attentions, None)
-
-                return custom_forward
-
-            layer_outputs = torch.utils.checkpoint.checkpoint(
-                create_custom_forward(decoder_layer),
-                hidden_states,
-                attention_mask,
-                position_ids,
-                None,
-            )
-        else:
-            layer_outputs = decoder_layer(
-                hidden_states,
-                attention_mask=attention_mask,
-                position_ids=position_ids,
-                past_key_value=past_key_value,
-                output_attentions=output_attentions,
-                use_cache=use_cache,
-            )
-
-        hidden_states = layer_outputs[0]
-
-        if use_cache:
-            next_decoder_cache += (
-                layer_outputs[2 if output_attentions else 1],
-            )
-
-        if output_attentions:
-            all_self_attns += (layer_outputs[1],)
-
-    hidden_states = self.norm(hidden_states)
-
-    # add hidden states from the last decoder layer
-    if output_hidden_states:
-        all_hidden_states += (hidden_states,)
-
-    next_cache = next_decoder_cache if use_cache else None
-    if not return_dict:
-        return tuple(
-            v
-            for v in [
-                hidden_states,
-                next_cache,
-                all_hidden_states,
-                all_self_attns,
-            ]
-            if v is not None
-        )
-    return BaseModelOutputWithPast(
-        last_hidden_state=hidden_states,
-        past_key_values=next_cache,
-        hidden_states=all_hidden_states,
-        attentions=all_self_attns,
-    )
-
-
-class CompiledEightLayerLayer(torch.nn.Module):
-    def __init__(self, model):
-        super().__init__()
-        self.model = model
-
-    def forward(
-        self,
-        hidden_states,
-        attention_mask,
-        position_ids,
-        past_key_value=None,
-        output_attentions=False,
-        use_cache=True,
-    ):
-        t2 = time()
-        if past_key_value is None:
-            try:
-                hidden_states = np.asarray(hidden_states, hidden_states.dtype)
-            except:
-                pass
-            attention_mask = attention_mask.detach()
-            position_ids = position_ids.detach()
-            t1 = time()
-
-            output = self.model(
-                "first_vicuna_forward",
-                (hidden_states, attention_mask, position_ids),
-                send_to_host=False,
-            )
-            output2 = (
-                output[0],
-                (
-                    output[1],
-                    output[2],
-                ),
-                (
-                    output[3],
-                    output[4],
-                ),
-                (
-                    output[5],
-                    output[6],
-                ),
-                (
-                    output[7],
-                    output[8],
-                ),
-                (
-                    output[9],
-                    output[10],
-                ),
-                (
-                    output[11],
-                    output[12],
-                ),
-                (
-                    output[13],
-                    output[14],
-                ),
-                (
-                    output[15],
-                    output[16],
-                ),
-            )
-            return output2
-        else:
-            (
-                (pkv00, pkv01),
-                (pkv10, pkv11),
-                (pkv20, pkv21),
-                (pkv30, pkv31),
-                (pkv40, pkv41),
-                (pkv50, pkv51),
-                (pkv60, pkv61),
-                (pkv70, pkv71),
-            ) = past_key_value
-
-            try:
-                hidden_states = hidden_states.detach()
-                attention_mask = attention_mask.detach()
-                position_ids = position_ids.detach()
-                pkv00 = pkv00.detach()
-                pkv01 = pkv01.detach()
-                pkv10 = pkv10.detach()
-                pkv11 = pkv11.detach()
-                pkv20 = pkv20.detach()
-                pkv21 = pkv21.detach()
-                pkv30 = pkv30.detach()
-                pkv31 = pkv31.detach()
-                pkv40 = pkv40.detach()
-                pkv41 = pkv41.detach()
-                pkv50 = pkv50.detach()
-                pkv51 = pkv51.detach()
-                pkv60 = pkv60.detach()
-                pkv61 = pkv61.detach()
-                pkv70 = pkv70.detach()
-                pkv71 = pkv71.detach()
-            except:
-                x = 10
-
-            t1 = time()
-            if type(hidden_states) == iree.runtime.array_interop.DeviceArray:
-                hidden_states = np.array(hidden_states, hidden_states.dtype)
-                hidden_states = torch.tensor(hidden_states)
-                hidden_states = hidden_states.detach()
-
-            output = self.model(
-                "second_vicuna_forward",
-                (
-                    hidden_states,
-                    attention_mask,
-                    position_ids,
-                    pkv00,
-                    pkv01,
-                    pkv10,
-                    pkv11,
-                    pkv20,
-                    pkv21,
-                    pkv30,
-                    pkv31,
-                    pkv40,
-                    pkv41,
-                    pkv50,
-                    pkv51,
-                    pkv60,
-                    pkv61,
-                    pkv70,
-                    pkv71,
-                ),
-                send_to_host=False,
-            )
-            print(f"{time() - t1}")
-            del pkv00
-            del pkv01
-            del pkv10
-            del pkv11
-            del pkv20
-            del pkv21
-            del pkv30
-            del pkv31
-            del pkv40
-            del pkv41
-            del pkv50
-            del pkv51
-            del pkv60
-            del pkv61
-            del pkv70
-            del pkv71
-            output2 = (
-                output[0],
-                (
-                    output[1],
-                    output[2],
-                ),
-                (
-                    output[3],
-                    output[4],
-                ),
-                (
-                    output[5],
-                    output[6],
-                ),
-                (
-                    output[7],
-                    output[8],
-                ),
-                (
-                    output[9],
-                    output[10],
-                ),
-                (
-                    output[11],
-                    output[12],
-                ),
-                (
-                    output[13],
-                    output[14],
-                ),
-                (
-                    output[15],
-                    output[16],
-                ),
-            )
-            return output2

apps/language_models/src/model_wrappers/vicuna_model.py

@@ -1,13 +1,15 @@
 import torch
 from transformers import AutoModelForCausalLM
 
+from brevitas_examples.llm.llm_quant.quantize import quantize_model
+from brevitas_examples.llm.llm_quant.run_utils import get_model_impl
+
 
 class FirstVicuna(torch.nn.Module):
     def __init__(
         self,
         model_path,
         precision="fp32",
-        accumulates="fp32",
         weight_group_size=128,
         model_name="vicuna",
         hf_auth_token: str = None,
@@ -16,24 +18,15 @@ class FirstVicuna(torch.nn.Module):
         kwargs = {"torch_dtype": torch.float32}
         if "llama2" in model_name:
             kwargs["use_auth_token"] = hf_auth_token
-        self.accumulates = (
-            torch.float32 if accumulates == "fp32" else torch.float16
-        )
         self.model = AutoModelForCausalLM.from_pretrained(
             model_path, low_cpu_mem_usage=True, **kwargs
         )
-        print(f"[DEBUG] model_path : {model_path}")
         if precision in ["int4", "int8"]:
-            from brevitas_examples.llm.llm_quant.quantize import quantize_model
-            from brevitas_examples.llm.llm_quant.run_utils import (
-                get_model_impl,
-            )
-
             print("First Vicuna applying weight quantization..")
             weight_bit_width = 4 if precision == "int4" else 8
             quantize_model(
                 get_model_impl(self.model).layers,
-                dtype=self.accumulates,
+                dtype=torch.float32,
                 weight_bit_width=weight_bit_width,
                 weight_param_method="stats",
                 weight_scale_precision="float",
@@ -47,556 +40,6 @@ class FirstVicuna(torch.nn.Module):
     def forward(self, input_ids):
         op = self.model(input_ids=input_ids, use_cache=True)
         return_vals = []
-        token = torch.argmax(op.logits[:, -1, :], dim=1)
-        return_vals.append(token)
-
-        temp_past_key_values = op.past_key_values
-        for item in temp_past_key_values:
-            return_vals.append(item[0])
-            return_vals.append(item[1])
-        return tuple(return_vals)
-
-
-class SecondVicuna7B(torch.nn.Module):
-    def __init__(
-        self,
-        model_path,
-        precision="fp32",
-        accumulates="fp32",
-        weight_group_size=128,
-        model_name="vicuna",
-        hf_auth_token: str = None,
-    ):
-        super().__init__()
-        kwargs = {"torch_dtype": torch.float32}
-        if "llama2" in model_name:
-            kwargs["use_auth_token"] = hf_auth_token
-        self.model = AutoModelForCausalLM.from_pretrained(
-            model_path, low_cpu_mem_usage=True, **kwargs
-        )
-        self.accumulates = (
-            torch.float32 if accumulates == "fp32" else torch.float16
-        )
-        print(f"[DEBUG] model_path : {model_path}")
-        if precision in ["int4", "int8"]:
-            from brevitas_examples.llm.llm_quant.quantize import quantize_model
-            from brevitas_examples.llm.llm_quant.run_utils import (
-                get_model_impl,
-            )
-
-            print("Second Vicuna applying weight quantization..")
-            weight_bit_width = 4 if precision == "int4" else 8
-            quantize_model(
-                get_model_impl(self.model).layers,
-                dtype=self.accumulates,
-                weight_bit_width=weight_bit_width,
-                weight_param_method="stats",
-                weight_scale_precision="float",
-                weight_quant_type="asym",
-                weight_quant_granularity="per_group",
-                weight_group_size=weight_group_size,
-                quantize_weight_zero_point=False,
-            )
-            print("Weight quantization applied.")
-
-    def forward(
-        self,
-        i0,
-        i1,
-        i2,
-        i3,
-        i4,
-        i5,
-        i6,
-        i7,
-        i8,
-        i9,
-        i10,
-        i11,
-        i12,
-        i13,
-        i14,
-        i15,
-        i16,
-        i17,
-        i18,
-        i19,
-        i20,
-        i21,
-        i22,
-        i23,
-        i24,
-        i25,
-        i26,
-        i27,
-        i28,
-        i29,
-        i30,
-        i31,
-        i32,
-        i33,
-        i34,
-        i35,
-        i36,
-        i37,
-        i38,
-        i39,
-        i40,
-        i41,
-        i42,
-        i43,
-        i44,
-        i45,
-        i46,
-        i47,
-        i48,
-        i49,
-        i50,
-        i51,
-        i52,
-        i53,
-        i54,
-        i55,
-        i56,
-        i57,
-        i58,
-        i59,
-        i60,
-        i61,
-        i62,
-        i63,
-        i64,
-    ):
-        token = i0
-        past_key_values = (
-            (i1, i2),
-            (
-                i3,
-                i4,
-            ),
-            (
-                i5,
-                i6,
-            ),
-            (
-                i7,
-                i8,
-            ),
-            (
-                i9,
-                i10,
-            ),
-            (
-                i11,
-                i12,
-            ),
-            (
-                i13,
-                i14,
-            ),
-            (
-                i15,
-                i16,
-            ),
-            (
-                i17,
-                i18,
-            ),
-            (
-                i19,
-                i20,
-            ),
-            (
-                i21,
-                i22,
-            ),
-            (
-                i23,
-                i24,
-            ),
-            (
-                i25,
-                i26,
-            ),
-            (
-                i27,
-                i28,
-            ),
-            (
-                i29,
-                i30,
-            ),
-            (
-                i31,
-                i32,
-            ),
-            (
-                i33,
-                i34,
-            ),
-            (
-                i35,
-                i36,
-            ),
-            (
-                i37,
-                i38,
-            ),
-            (
-                i39,
-                i40,
-            ),
-            (
-                i41,
-                i42,
-            ),
-            (
-                i43,
-                i44,
-            ),
-            (
-                i45,
-                i46,
-            ),
-            (
-                i47,
-                i48,
-            ),
-            (
-                i49,
-                i50,
-            ),
-            (
-                i51,
-                i52,
-            ),
-            (
-                i53,
-                i54,
-            ),
-            (
-                i55,
-                i56,
-            ),
-            (
-                i57,
-                i58,
-            ),
-            (
-                i59,
-                i60,
-            ),
-            (
-                i61,
-                i62,
-            ),
-            (
-                i63,
-                i64,
-            ),
-        )
-        op = self.model(
-            input_ids=token, use_cache=True, past_key_values=past_key_values
-        )
-        return_vals = []
-        token = torch.argmax(op.logits[:, -1, :], dim=1)
-        return_vals.append(token)
-        temp_past_key_values = op.past_key_values
-        for item in temp_past_key_values:
-            return_vals.append(item[0])
-            return_vals.append(item[1])
-        return tuple(return_vals)
-
-
-class SecondVicuna13B(torch.nn.Module):
-    def __init__(
-        self,
-        model_path,
-        precision="int8",
-        accumulates="fp32",
-        weight_group_size=128,
-        model_name="vicuna",
-        hf_auth_token: str = None,
-    ):
-        super().__init__()
-        kwargs = {"torch_dtype": torch.float32}
-        if "llama2" in model_name:
-            kwargs["use_auth_token"] = hf_auth_token
-        self.model = AutoModelForCausalLM.from_pretrained(
-            model_path, low_cpu_mem_usage=True, **kwargs
-        )
-        self.accumulates = (
-            torch.float32 if accumulates == "fp32" else torch.float16
-        )
-        if precision in ["int4", "int8"]:
-            from brevitas_examples.llm.llm_quant.quantize import quantize_model
-            from brevitas_examples.llm.llm_quant.run_utils import (
-                get_model_impl,
-            )
-
-            print("Second Vicuna applying weight quantization..")
-            weight_bit_width = 4 if precision == "int4" else 8
-            quantize_model(
-                get_model_impl(self.model).layers,
-                dtype=self.accumulates,
-                weight_bit_width=weight_bit_width,
-                weight_param_method="stats",
-                weight_scale_precision="float",
-                weight_quant_type="asym",
-                weight_quant_granularity="per_group",
-                weight_group_size=weight_group_size,
-                quantize_weight_zero_point=False,
-            )
-            print("Weight quantization applied.")
-
-    def forward(
-        self,
-        i0,
-        i1,
-        i2,
-        i3,
-        i4,
-        i5,
-        i6,
-        i7,
-        i8,
-        i9,
-        i10,
-        i11,
-        i12,
-        i13,
-        i14,
-        i15,
-        i16,
-        i17,
-        i18,
-        i19,
-        i20,
-        i21,
-        i22,
-        i23,
-        i24,
-        i25,
-        i26,
-        i27,
-        i28,
-        i29,
-        i30,
-        i31,
-        i32,
-        i33,
-        i34,
-        i35,
-        i36,
-        i37,
-        i38,
-        i39,
-        i40,
-        i41,
-        i42,
-        i43,
-        i44,
-        i45,
-        i46,
-        i47,
-        i48,
-        i49,
-        i50,
-        i51,
-        i52,
-        i53,
-        i54,
-        i55,
-        i56,
-        i57,
-        i58,
-        i59,
-        i60,
-        i61,
-        i62,
-        i63,
-        i64,
-        i65,
-        i66,
-        i67,
-        i68,
-        i69,
-        i70,
-        i71,
-        i72,
-        i73,
-        i74,
-        i75,
-        i76,
-        i77,
-        i78,
-        i79,
-        i80,
-    ):
-        token = i0
-        past_key_values = (
-            (i1, i2),
-            (
-                i3,
-                i4,
-            ),
-            (
-                i5,
-                i6,
-            ),
-            (
-                i7,
-                i8,
-            ),
-            (
-                i9,
-                i10,
-            ),
-            (
-                i11,
-                i12,
-            ),
-            (
-                i13,
-                i14,
-            ),
-            (
-                i15,
-                i16,
-            ),
-            (
-                i17,
-                i18,
-            ),
-            (
-                i19,
-                i20,
-            ),
-            (
-                i21,
-                i22,
-            ),
-            (
-                i23,
-                i24,
-            ),
-            (
-                i25,
-                i26,
-            ),
-            (
-                i27,
-                i28,
-            ),
-            (
-                i29,
-                i30,
-            ),
-            (
-                i31,
-                i32,
-            ),
-            (
-                i33,
-                i34,
-            ),
-            (
-                i35,
-                i36,
-            ),
-            (
-                i37,
-                i38,
-            ),
-            (
-                i39,
-                i40,
-            ),
-            (
-                i41,
-                i42,
-            ),
-            (
-                i43,
-                i44,
-            ),
-            (
-                i45,
-                i46,
-            ),
-            (
-                i47,
-                i48,
-            ),
-            (
-                i49,
-                i50,
-            ),
-            (
-                i51,
-                i52,
-            ),
-            (
-                i53,
-                i54,
-            ),
-            (
-                i55,
-                i56,
-            ),
-            (
-                i57,
-                i58,
-            ),
-            (
-                i59,
-                i60,
-            ),
-            (
-                i61,
-                i62,
-            ),
-            (
-                i63,
-                i64,
-            ),
-            (
-                i65,
-                i66,
-            ),
-            (
-                i67,
-                i68,
-            ),
-            (
-                i69,
-                i70,
-            ),
-            (
-                i71,
-                i72,
-            ),
-            (
-                i73,
-                i74,
-            ),
-            (
-                i75,
-                i76,
-            ),
-            (
-                i77,
-                i78,
-            ),
-            (
-                i79,
-                i80,
-            ),
-        )
-        op = self.model(
-            input_ids=token, use_cache=True, past_key_values=past_key_values
-        )
-        return_vals = []
         return_vals.append(op.logits)
         temp_past_key_values = op.past_key_values
         for item in temp_past_key_values:
@@ -605,12 +48,11 @@ class SecondVicuna13B(torch.nn.Module):
         return tuple(return_vals)
 
 
-class SecondVicuna70B(torch.nn.Module):
+class SecondVicuna(torch.nn.Module):
     def __init__(
         self,
         model_path,
         precision="fp32",
-        accumulates="fp32",
         weight_group_size=128,
         model_name="vicuna",
         hf_auth_token: str = None,
@@ -622,21 +64,12 @@ class SecondVicuna70B(torch.nn.Module):
         self.model = AutoModelForCausalLM.from_pretrained(
             model_path, low_cpu_mem_usage=True, **kwargs
         )
-        self.accumulates = (
-            torch.float32 if accumulates == "fp32" else torch.float16
-        )
-        print(f"[DEBUG] model_path : {model_path}")
         if precision in ["int4", "int8"]:
-            from brevitas_examples.llm.llm_quant.quantize import quantize_model
-            from brevitas_examples.llm.llm_quant.run_utils import (
-                get_model_impl,
-            )
-
             print("Second Vicuna applying weight quantization..")
             weight_bit_width = 4 if precision == "int4" else 8
             quantize_model(
                 get_model_impl(self.model).layers,
-                dtype=self.accumulates,
+                dtype=torch.float32,
                 weight_bit_width=weight_bit_width,
                 weight_param_method="stats",
                 weight_scale_precision="float",
@@ -714,103 +147,9 @@ class SecondVicuna70B(torch.nn.Module):
         i62,
         i63,
         i64,
-        i65,
-        i66,
-        i67,
-        i68,
-        i69,
-        i70,
-        i71,
-        i72,
-        i73,
-        i74,
-        i75,
-        i76,
-        i77,
-        i78,
-        i79,
-        i80,
-        i81,
-        i82,
-        i83,
-        i84,
-        i85,
-        i86,
-        i87,
-        i88,
-        i89,
-        i90,
-        i91,
-        i92,
-        i93,
-        i94,
-        i95,
-        i96,
-        i97,
-        i98,
-        i99,
-        i100,
-        i101,
-        i102,
-        i103,
-        i104,
-        i105,
-        i106,
-        i107,
-        i108,
-        i109,
-        i110,
-        i111,
-        i112,
-        i113,
-        i114,
-        i115,
-        i116,
-        i117,
-        i118,
-        i119,
-        i120,
-        i121,
-        i122,
-        i123,
-        i124,
-        i125,
-        i126,
-        i127,
-        i128,
-        i129,
-        i130,
-        i131,
-        i132,
-        i133,
-        i134,
-        i135,
-        i136,
-        i137,
-        i138,
-        i139,
-        i140,
-        i141,
-        i142,
-        i143,
-        i144,
-        i145,
-        i146,
-        i147,
-        i148,
-        i149,
-        i150,
-        i151,
-        i152,
-        i153,
-        i154,
-        i155,
-        i156,
-        i157,
-        i158,
-        i159,
-        i160,
     ):
+        # input_ids = input_tuple[0]
+        # input_tuple = torch.unbind(pkv, dim=0)
         token = i0
         past_key_values = (
             (i1, i2),
@@ -938,198 +277,6 @@ class SecondVicuna70B(torch.nn.Module):
                 i63,
                 i64,
             ),
-            (
-                i65,
-                i66,
-            ),
-            (
-                i67,
-                i68,
-            ),
-            (
-                i69,
-                i70,
-            ),
-            (
-                i71,
-                i72,
-            ),
-            (
-                i73,
-                i74,
-            ),
-            (
-                i75,
-                i76,
-            ),
-            (
-                i77,
-                i78,
-            ),
-            (
-                i79,
-                i80,
-            ),
-            (
-                i81,
-                i82,
-            ),
-            (
-                i83,
-                i84,
-            ),
-            (
-                i85,
-                i86,
-            ),
-            (
-                i87,
-                i88,
-            ),
-            (
-                i89,
-                i90,
-            ),
-            (
-                i91,
-                i92,
-            ),
-            (
-                i93,
-                i94,
-            ),
-            (
-                i95,
-                i96,
-            ),
-            (
-                i97,
-                i98,
-            ),
-            (
-                i99,
-                i100,
-            ),
-            (
-                i101,
-                i102,
-            ),
-            (
-                i103,
-                i104,
-            ),
-            (
-                i105,
-                i106,
-            ),
-            (
-                i107,
-                i108,
-            ),
-            (
-                i109,
-                i110,
-            ),
-            (
-                i111,
-                i112,
-            ),
-            (
-                i113,
-                i114,
-            ),
-            (
-                i115,
-                i116,
-            ),
-            (
-                i117,
-                i118,
-            ),
-            (
-                i119,
-                i120,
-            ),
-            (
-                i121,
-                i122,
-            ),
-            (
-                i123,
-                i124,
-            ),
-            (
-                i125,
-                i126,
-            ),
-            (
-                i127,
-                i128,
-            ),
-            (
-                i129,
-                i130,
-            ),
-            (
-                i131,
-                i132,
-            ),
-            (
-                i133,
-                i134,
-            ),
-            (
-                i135,
-                i136,
-            ),
-            (
-                i137,
-                i138,
-            ),
-            (
-                i139,
-                i140,
-            ),
-            (
-                i141,
-                i142,
-            ),
-            (
-                i143,
-                i144,
-            ),
-            (
-                i145,
-                i146,
-            ),
-            (
-                i147,
-                i148,
-            ),
-            (
-                i149,
-                i150,
-            ),
-            (
-                i151,
-                i152,
-            ),
-            (
-                i153,
-                i154,
-            ),
-            (
-                i155,
-                i156,
-            ),
-            (
-                i157,
-                i158,
-            ),
-            (
-                i159,
-                i160,
-            ),
         )
         op = self.model(
             input_ids=token, use_cache=True, past_key_values=past_key_values
@@ -1151,17 +298,15 @@ class CombinedModel(torch.nn.Module):
     ):
         super().__init__()
         self.first_vicuna = FirstVicuna(first_vicuna_model_path)
-        # NOT using this path for 13B currently, hence using `SecondVicuna7B`.
-        self.second_vicuna = SecondVicuna7B(second_vicuna_model_path)
+        self.second_vicuna = SecondVicuna(second_vicuna_model_path)
 
     def forward(self, input_ids):
-        first_output = self.first_vicuna(input_ids=input_ids)
-        # generate second vicuna
-        compilation_input_ids = torch.zeros([1, 1], dtype=torch.int64)
-        pkv = tuple(
-            (torch.zeros([1, 32, 19, 128], dtype=torch.float32))
-            for _ in range(64)
-        )
-        secondVicunaCompileInput = (compilation_input_ids,) + pkv
-        second_output = self.second_vicuna(*secondVicunaCompileInput)
+        first_output = self.first_vicuna(input_ids=input_ids, use_cache=True)
+        logits = first_output[0]
+        pkv = first_output[1:]
+
+        token = torch.argmax(torch.tensor(logits)[:, -1, :], dim=1)
+        token = token.to(torch.int64).reshape([1, 1])
+        secondVicunaInput = (token,) + tuple(pkv)
+        second_output = self.second_vicuna(secondVicunaInput)
         return second_output

apps/language_models/src/model_wrappers/vicuna_sharded_model.py

@@ -66,7 +66,7 @@ class ShardedVicunaModel(torch.nn.Module):
     def __init__(self, model, layers, lmhead, embedding, norm):
         super().__init__()
         self.model = model
-        # assert len(layers) == len(model.model.layers)
+        assert len(layers) == len(model.model.layers)
         self.model.model.config.use_cache = True
         self.model.model.config.output_attentions = False
         self.layers = layers
@@ -132,10 +132,7 @@ class VicunaNormCompiled(torch.nn.Module):
         self.model = shark_module
 
     def forward(self, hidden_states):
-        try:
-            hidden_states.detach()
-        except:
-            pass
+        hidden_states.detach()
         output = self.model("forward", (hidden_states,))
         output = torch.tensor(output)
         return output

apps/language_models/src/pipelines/SharkLLMBase.py

@@ -3,10 +3,7 @@ from abc import ABC, abstractmethod
 
 class SharkLLMBase(ABC):
     def __init__(
-        self,
-        model_name,
-        hf_model_path=None,
-        max_num_tokens=512,
+        self, model_name, hf_model_path=None, max_num_tokens=512
     ) -> None:
         self.model_name = model_name
         self.hf_model_path = hf_model_path

apps/language_models/src/pipelines/falcon_pipeline.py

@@ -1,17 +1,4 @@
 from apps.language_models.src.model_wrappers.falcon_model import FalconModel
-from apps.language_models.src.model_wrappers.falcon_sharded_model import (
-    WordEmbeddingsLayer,
-    CompiledWordEmbeddingsLayer,
-    LNFEmbeddingLayer,
-    CompiledLNFEmbeddingLayer,
-    LMHeadEmbeddingLayer,
-    CompiledLMHeadEmbeddingLayer,
-    DecoderLayer,
-    EightDecoderLayer,
-    CompiledDecoderLayer,
-    CompiledEightDecoderLayer,
-    ShardedFalconModel,
-)
 from apps.language_models.src.pipelines.SharkLLMBase import SharkLLMBase
 from apps.language_models.utils import (
     get_vmfb_from_path,
@@ -20,39 +7,30 @@ from io import BytesIO
 from pathlib import Path
 from contextlib import redirect_stdout
 from shark.shark_downloader import download_public_file
-from shark.shark_importer import import_with_fx, save_mlir
+from shark.shark_importer import import_with_fx
 from shark.shark_inference import SharkInference
-from transformers import AutoTokenizer, AutoModelForCausalLM, GPTQConfig
+from transformers import AutoTokenizer, AutoModelForCausalLM
 from transformers.generation import (
     GenerationConfig,
     LogitsProcessorList,
     StoppingCriteriaList,
 )
 import copy
-import time
+
 import re
 import torch
 import torch_mlir
 import os
 import argparse
-import gc
 
 parser = argparse.ArgumentParser(
     prog="falcon runner",
     description="runs a falcon model",
 )
 
+parser.add_argument("--falcon_variant_to_use", default="7b", help="7b, 40b")
 parser.add_argument(
-    "--falcon_variant_to_use", default="7b", help="7b, 40b, 180b"
-)
-parser.add_argument(
-    "--compressed",
-    default=False,
-    action=argparse.BooleanOptionalAction,
-    help="Do the compression of sharded layers",
-)
-parser.add_argument(
-    "--precision", "-p", default="fp16", choices=["fp32", "fp16", "int4"]
+    "--precision", "-p", default="fp16", help="fp32, fp16, int8, int4"
 )
 parser.add_argument("--device", "-d", default="cuda", help="vulkan, cpu, cuda")
 parser.add_argument(
@@ -71,7 +49,7 @@ parser.add_argument(
 )
 parser.add_argument(
     "--load_mlir_from_shark_tank",
-    default=True,
+    default=False,
     action=argparse.BooleanOptionalAction,
     help="download precompile mlir from shark tank",
 )
@@ -81,585 +59,32 @@ parser.add_argument(
     action=argparse.BooleanOptionalAction,
     help="Run model in cli mode",
 )
-parser.add_argument(
-    "--hf_auth_token",
-    type=str,
-    default=None,
-    help="Specify your own huggingface authentication token for falcon-180B model.",
-)
-parser.add_argument(
-    "-s",
-    "--sharded",
-    default=False,
-    action=argparse.BooleanOptionalAction,
-    help="Run model as sharded",
-)
 
 
-class ShardedFalcon(SharkLLMBase):
+class Falcon(SharkLLMBase):
     def __init__(
         self,
         model_name,
-        hf_model_path="tiiuae/falcon-7b-instruct",
-        hf_auth_token: str = None,
+        hf_model_path,
         max_num_tokens=150,
         device="cuda",
         precision="fp32",
         falcon_mlir_path=None,
         falcon_vmfb_path=None,
-        debug=False,
     ) -> None:
         super().__init__(model_name, hf_model_path, max_num_tokens)
-        print("hf_model_path: ", self.hf_model_path)
-
-        if (
-            "180b" in self.model_name
-            and precision != "int4"
-            and hf_auth_token == None
-        ):
-            raise ValueError(
-                """ HF auth token required for falcon-180b. Pass it using
-                --hf_auth_token flag. You can ask for the access to the model
-                here: https://huggingface.co/tiiuae/falcon-180B-chat."""
-            )
-        self.hf_auth_token = hf_auth_token
         self.max_padding_length = 100
         self.device = device
         self.precision = precision
         self.falcon_vmfb_path = falcon_vmfb_path
         self.falcon_mlir_path = falcon_mlir_path
-        self.debug = debug
         self.tokenizer = self.get_tokenizer()
-        self.src_model = self.get_src_model()
-        self.shark_model = self.compile(compressed=args.compressed)
-
-    def get_tokenizer(self):
-        tokenizer = AutoTokenizer.from_pretrained(
-            self.hf_model_path,
-            trust_remote_code=True,
-            token=self.hf_auth_token,
-        )
-        tokenizer.padding_side = "left"
-        tokenizer.pad_token_id = 11
-        return tokenizer
-
-    def get_src_model(self):
-        print("Loading src model: ", self.model_name)
-        kwargs = {
-            "torch_dtype": torch.float,
-            "trust_remote_code": True,
-            "token": self.hf_auth_token,
-        }
-        if self.precision == "int4":
-            quantization_config = GPTQConfig(bits=4, disable_exllama=True)
-            kwargs["quantization_config"] = quantization_config
-            kwargs["load_gptq_on_cpu"] = True
-            kwargs["device_map"] = "cpu"
-        falcon_model = AutoModelForCausalLM.from_pretrained(
-            self.hf_model_path, **kwargs
-        )
-        if self.precision == "int4":
-            falcon_model = falcon_model.to(torch.float32)
-        return falcon_model
-
-    def compile_layer(
-        self, layer, falconCompileInput, layer_id, device_idx=None
-    ):
-        self.falcon_mlir_path = Path(
-            f"falcon_{args.falcon_variant_to_use}_layer_{layer_id}_{self.precision}.mlir"
-        )
-        self.falcon_vmfb_path = Path(
-            f"falcon_{args.falcon_variant_to_use}_layer_{layer_id}_{self.precision}_{self.device}.vmfb"
-        )
-
-        if args.use_precompiled_model:
-            if not self.falcon_vmfb_path.exists():
-                # Downloading VMFB from shark_tank
-                print(f"[DEBUG] Trying to download vmfb from shark_tank")
-                download_public_file(
-                    f"gs://shark_tank/falcon/sharded/falcon_{args.falcon_variant_to_use}/vmfb/"
-                    + str(self.falcon_vmfb_path),
-                    self.falcon_vmfb_path.absolute(),
-                    single_file=True,
-                )
-            vmfb = get_vmfb_from_path(
-                self.falcon_vmfb_path,
-                self.device,
-                "linalg",
-                device_id=device_idx,
-            )
-            if vmfb is not None:
-                return vmfb, device_idx
-
-        print(f"[DEBUG] vmfb not found at {self.falcon_vmfb_path.absolute()}")
-        if self.falcon_mlir_path.exists():
-            print(f"[DEBUG] mlir found at {self.falcon_mlir_path.absolute()}")
-            with open(self.falcon_mlir_path, "rb") as f:
-                bytecode = f.read()
-        else:
-            mlir_generated = False
-            print(
-                f"[DEBUG] mlir not found at {self.falcon_mlir_path.absolute()}"
-            )
-            if args.load_mlir_from_shark_tank:
-                # Downloading MLIR from shark_tank
-                print(f"[DEBUG] Trying to download mlir from shark_tank")
-                download_public_file(
-                    f"gs://shark_tank/falcon/sharded/falcon_{args.falcon_variant_to_use}/mlir/"
-                    + str(self.falcon_mlir_path),
-                    self.falcon_mlir_path.absolute(),
-                    single_file=True,
-                )
-                if self.falcon_mlir_path.exists():
-                    print(
-                        f"[DEBUG] mlir found at {self.falcon_mlir_path.absolute()}"
-                    )
-                    with open(self.falcon_mlir_path, "rb") as f:
-                        bytecode = f.read()
-                    mlir_generated = True
-
-            if not mlir_generated:
-                print(f"[DEBUG] generating MLIR locally")
-                if layer_id == "word_embeddings":
-                    f16_input_mask = [False]
-                elif layer_id in ["ln_f", "lm_head"]:
-                    f16_input_mask = [True]
-                elif "_" in layer_id or type(layer_id) == int:
-                    f16_input_mask = [True, False]
-                else:
-                    raise ValueError("Unsupported layer: ", layer_id)
-
-                print(f"[DEBUG] generating torchscript graph")
-                ts_graph = import_with_fx(
-                    layer,
-                    falconCompileInput,
-                    is_f16=True,
-                    f16_input_mask=f16_input_mask,
-                    mlir_type="torchscript",
-                    is_gptq=True,
-                )
-                del layer
-
-                print(f"[DEBUG] generating torch mlir")
-                module = torch_mlir.compile(
-                    ts_graph,
-                    falconCompileInput,
-                    torch_mlir.OutputType.LINALG_ON_TENSORS,
-                    use_tracing=False,
-                    verbose=False,
-                )
-                del ts_graph
-
-                print(f"[DEBUG] converting to bytecode")
-                bytecode_stream = BytesIO()
-                module.operation.write_bytecode(bytecode_stream)
-                bytecode = bytecode_stream.getvalue()
-                del module
-
-                f_ = open(self.falcon_mlir_path, "wb")
-                f_.write(bytecode)
-                print("Saved falcon mlir at ", str(self.falcon_mlir_path))
-                f_.close()
-                del bytecode
-
-        shark_module = SharkInference(
-            mlir_module=self.falcon_mlir_path,
-            device=self.device,
-            mlir_dialect="linalg",
-            device_idx=device_idx,
-        )
-        path = shark_module.save_module(
-            self.falcon_vmfb_path.parent.absolute(),
-            self.falcon_vmfb_path.stem,
-            extra_args=[
-                "--iree-vm-target-truncate-unsupported-floats",
-                "--iree-codegen-check-ir-before-llvm-conversion=false",
-                "--iree-vm-bytecode-module-output-format=flatbuffer-binary",
-            ]
-            + [
-                "--iree-llvmcpu-use-fast-min-max-ops",
-            ]
-            if self.precision == "int4"
-            else [],
-            debug=self.debug,
-        )
-        print("Saved falcon vmfb at ", str(path))
-        shark_module.load_module(path)
-
-        return shark_module, device_idx
-
-    def compile(self, compressed=False):
-        sample_input_ids = torch.zeros([100], dtype=torch.int64)
-        sample_attention_mask = torch.zeros(
-            [1, 1, 100, 100], dtype=torch.float32
-        )
-        num_group_layers = 1
-        if "7b" in self.model_name:
-            num_in_features = 4544
-            if compressed:
-                num_group_layers = 8
-        elif "40b" in self.model_name:
-            num_in_features = 8192
-            if compressed:
-                num_group_layers = 15
-        else:
-            num_in_features = 14848
-            sample_attention_mask = sample_attention_mask.to(dtype=torch.bool)
-            if compressed:
-                num_group_layers = 20
-
-        sample_hidden_states = torch.zeros(
-            [1, 100, num_in_features], dtype=torch.float32
-        )
-
-        # Determine number of available devices
-        num_devices = 1
-        if self.device == "rocm":
-            import iree.runtime as ireert
-
-            haldriver = ireert.get_driver(self.device)
-            num_devices = len(haldriver.query_available_devices())
-
-        lm_head = LMHeadEmbeddingLayer(self.src_model.lm_head)
-        print("Compiling Layer lm_head")
-        shark_lm_head, _ = self.compile_layer(
-            lm_head,
-            [sample_hidden_states],
-            "lm_head",
-            device_idx=0 % num_devices if self.device == "rocm" else None,
-        )
-        shark_lm_head = CompiledLMHeadEmbeddingLayer(shark_lm_head)
-
-        word_embedding = WordEmbeddingsLayer(
-            self.src_model.transformer.word_embeddings
-        )
-        print("Compiling Layer word_embeddings")
-        shark_word_embedding, _ = self.compile_layer(
-            word_embedding,
-            [sample_input_ids],
-            "word_embeddings",
-            device_idx=1 % num_devices if self.device == "rocm" else None,
-        )
-        shark_word_embedding = CompiledWordEmbeddingsLayer(
-            shark_word_embedding
-        )
-
-        ln_f = LNFEmbeddingLayer(self.src_model.transformer.ln_f)
-        print("Compiling Layer ln_f")
-        shark_ln_f, _ = self.compile_layer(
-            ln_f,
-            [sample_hidden_states],
-            "ln_f",
-            device_idx=2 % num_devices if self.device == "rocm" else None,
-        )
-        shark_ln_f = CompiledLNFEmbeddingLayer(shark_ln_f)
-
-        shark_layers = []
-        for i in range(
-            int(len(self.src_model.transformer.h) / num_group_layers)
-        ):
-            device_idx = i % num_devices if self.device == "rocm" else None
-            layer_id = i
-            pytorch_class = DecoderLayer
-            compiled_class = CompiledDecoderLayer
-            if compressed:
-                layer_id = (
-                    str(i * num_group_layers)
-                    + "_"
-                    + str((i + 1) * num_group_layers)
-                )
-                pytorch_class = EightDecoderLayer
-                compiled_class = CompiledEightDecoderLayer
-
-            print("Compiling Layer {}".format(layer_id))
-            if compressed:
-                layer_i = self.src_model.transformer.h[
-                    i * num_group_layers : (i + 1) * num_group_layers
-                ]
-            else:
-                layer_i = self.src_model.transformer.h[i]
-
-            pytorch_layer_i = pytorch_class(
-                layer_i, args.falcon_variant_to_use
-            )
-            shark_module, device_idx = self.compile_layer(
-                pytorch_layer_i,
-                [sample_hidden_states, sample_attention_mask],
-                layer_id,
-                device_idx=device_idx,
-            )
-            del shark_module
-            shark_layer_i = compiled_class(
-                layer_id,
-                device_idx,
-                args.falcon_variant_to_use,
-                self.device,
-                self.precision,
-            )
-            shark_layers.append(shark_layer_i)
-
-        sharded_model = ShardedFalconModel(
-            self.src_model,
-            shark_layers,
-            shark_word_embedding,
-            shark_ln_f,
-            shark_lm_head,
-        )
-        return sharded_model
-
-    def generate(self, prompt):
-        model_inputs = self.tokenizer(
-            prompt,
-            padding="max_length",
-            max_length=self.max_padding_length,
-            add_special_tokens=False,
-            return_tensors="pt",
-        )
-        model_inputs["prompt_text"] = prompt
-
-        input_ids = model_inputs["input_ids"]
-        attention_mask = model_inputs.get("attention_mask", None)
-
-        # Allow empty prompts
-        if input_ids.shape[1] == 0:
-            input_ids = None
-            attention_mask = None
-
-        generate_kwargs = {
-            "max_length": self.max_num_tokens,
-            "do_sample": True,
-            "top_k": 10,
-            "num_return_sequences": 1,
-            "eos_token_id": 11,
-        }
-        generate_kwargs["input_ids"] = input_ids
-        generate_kwargs["attention_mask"] = attention_mask
-        generation_config_ = GenerationConfig.from_model_config(
-            self.src_model.config
-        )
-        generation_config = copy.deepcopy(generation_config_)
-        model_kwargs = generation_config.update(**generate_kwargs)
-
-        logits_processor = LogitsProcessorList()
-        stopping_criteria = StoppingCriteriaList()
-
-        eos_token_id = generation_config.eos_token_id
-        generation_config.pad_token_id = eos_token_id
-
-        (
-            inputs_tensor,
-            model_input_name,
-            model_kwargs,
-        ) = self.src_model._prepare_model_inputs(
-            None, generation_config.bos_token_id, model_kwargs
-        )
-
-        model_kwargs["output_attentions"] = generation_config.output_attentions
-        model_kwargs[
-            "output_hidden_states"
-        ] = generation_config.output_hidden_states
-        model_kwargs["use_cache"] = generation_config.use_cache
-
-        input_ids = (
-            inputs_tensor
-            if model_input_name == "input_ids"
-            else model_kwargs.pop("input_ids")
-        )
-
-        self.logits_processor = self.src_model._get_logits_processor(
-            generation_config=generation_config,
-            input_ids_seq_length=input_ids.shape[-1],
-            encoder_input_ids=inputs_tensor,
-            prefix_allowed_tokens_fn=None,
-            logits_processor=logits_processor,
-        )
-
-        self.stopping_criteria = self.src_model._get_stopping_criteria(
-            generation_config=generation_config,
-            stopping_criteria=stopping_criteria,
-        )
-
-        self.logits_warper = self.src_model._get_logits_warper(
-            generation_config
-        )
-
-        (
-            self.input_ids,
-            self.model_kwargs,
-        ) = self.src_model._expand_inputs_for_generation(
-            input_ids=input_ids,
-            expand_size=generation_config.num_return_sequences,  # 1
-            is_encoder_decoder=self.src_model.config.is_encoder_decoder,  # False
-            **model_kwargs,
-        )
-
-        if isinstance(eos_token_id, int):
-            eos_token_id = [eos_token_id]
-        self.eos_token_id_tensor = (
-            torch.tensor(eos_token_id) if eos_token_id is not None else None
-        )
-
-        self.pad_token_id = generation_config.pad_token_id
-        self.eos_token_id = eos_token_id
-
-        output_scores = generation_config.output_scores  # False
-        return_dict_in_generate = (
-            generation_config.return_dict_in_generate  # False
-        )
-
-        # init attention / hidden states / scores tuples
-        self.scores = (
-            () if (return_dict_in_generate and output_scores) else None
-        )
-
-        # keep track of which sequences are already finished
-        self.unfinished_sequences = torch.ones(
-            input_ids.shape[0], dtype=torch.long, device=input_ids.device
-        )
-
-        all_text = prompt
-
-        start = time.time()
-        count = 0
-        for i in range(self.max_num_tokens - 1):
-            count = count + 1
-
-            next_token = self.generate_new_token()
-            new_word = self.tokenizer.decode(
-                next_token.cpu().numpy(),
-                add_special_tokens=False,
-                skip_special_tokens=True,
-                clean_up_tokenization_spaces=True,
-            )
-
-            all_text = all_text + new_word
-
-            print(f"{new_word}", end="", flush=True)
-            print(f"{all_text}", end="", flush=True)
-
-            # if eos_token was found in one sentence, set sentence to finished
-            if self.eos_token_id_tensor is not None:
-                self.unfinished_sequences = self.unfinished_sequences.mul(
-                    next_token.tile(self.eos_token_id_tensor.shape[0], 1)
-                    .ne(self.eos_token_id_tensor.unsqueeze(1))
-                    .prod(dim=0)
-                )
-                # stop when each sentence is finished
-                if (
-                    self.unfinished_sequences.max() == 0
-                    or self.stopping_criteria(input_ids, self.scores)
-                ):
-                    break
-
-        end = time.time()
-        print(
-            "\n\nTime taken is {:.2f} seconds/token\n".format(
-                (end - start) / count
-            )
-        )
-
-        torch.cuda.empty_cache()
-        gc.collect()
-
-        return all_text
-
-    def generate_new_token(self):
-        model_inputs = self.src_model.prepare_inputs_for_generation(
-            self.input_ids, **self.model_kwargs
-        )
-        outputs = self.shark_model.forward(
-            input_ids=model_inputs["input_ids"],
-            attention_mask=model_inputs["attention_mask"],
-        )
-        if self.precision in ["fp16", "int4"]:
-            outputs = outputs.to(dtype=torch.float32)
-        next_token_logits = outputs
-
-        # pre-process distribution
-        next_token_scores = self.logits_processor(
-            self.input_ids, next_token_logits
-        )
-        next_token_scores = self.logits_warper(
-            self.input_ids, next_token_scores
-        )
-
-        # sample
-        probs = torch.nn.functional.softmax(next_token_scores, dim=-1)
-
-        next_token = torch.multinomial(probs, num_samples=1).squeeze(1)
-
-        # finished sentences should have their next token be a padding token
-        if self.eos_token_id is not None:
-            if self.pad_token_id is None:
-                raise ValueError(
-                    "If `eos_token_id` is defined, make sure that `pad_token_id` is defined."
-                )
-            next_token = (
-                next_token * self.unfinished_sequences
-                + self.pad_token_id * (1 - self.unfinished_sequences)
-            )
-
-        self.input_ids = torch.cat(
-            [self.input_ids, next_token[:, None]], dim=-1
-        )
-
-        self.model_kwargs["past_key_values"] = None
-        if "attention_mask" in self.model_kwargs:
-            attention_mask = self.model_kwargs["attention_mask"]
-            self.model_kwargs["attention_mask"] = torch.cat(
-                [
-                    attention_mask,
-                    attention_mask.new_ones((attention_mask.shape[0], 1)),
-                ],
-                dim=-1,
-            )
-
-        self.input_ids = self.input_ids[:, 1:]
-        self.model_kwargs["attention_mask"] = self.model_kwargs[
-            "attention_mask"
-        ][:, 1:]
-
-        return next_token
-
-
-class UnshardedFalcon(SharkLLMBase):
-    def __init__(
-        self,
-        model_name,
-        hf_model_path="tiiuae/falcon-7b-instruct",
-        hf_auth_token: str = "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",
-        max_num_tokens=150,
-        device="cuda",
-        precision="fp32",
-        falcon_mlir_path=None,
-        falcon_vmfb_path=None,
-        debug=False,
-    ) -> None:
-        super().__init__(model_name, hf_model_path, max_num_tokens)
-        print("hf_model_path: ", self.hf_model_path)
-
-        if "180b" in self.model_name and hf_auth_token == None:
-            raise ValueError(
-                """ HF auth token required for falcon-180b. Pass it using
-                --hf_auth_token flag. You can ask for the access to the model
-                here: https://huggingface.co/tiiuae/falcon-180B-chat."""
-            )
-        self.hf_auth_token = hf_auth_token
-        self.max_padding_length = 100
-        self.device = device
-        self.precision = precision
-        self.falcon_vmfb_path = falcon_vmfb_path
-        self.falcon_mlir_path = falcon_mlir_path
-        self.debug = debug
-        self.tokenizer = self.get_tokenizer()
-        self.src_model = self.get_src_model()
         self.shark_model = self.compile()
+        self.src_model = self.get_src_model()
 
     def get_tokenizer(self):
         tokenizer = AutoTokenizer.from_pretrained(
-            self.hf_model_path,
-            trust_remote_code=True,
-            token=self.hf_auth_token,
+            self.hf_model_path, trust_remote_code=True
         )
         tokenizer.padding_side = "left"
         tokenizer.pad_token_id = 11
@@ -667,24 +92,13 @@ class UnshardedFalcon(SharkLLMBase):
 
     def get_src_model(self):
         print("Loading src model: ", self.model_name)
-        kwargs = {
-            "torch_dtype": torch.float,
-            "trust_remote_code": True,
-            "token": self.hf_auth_token,
-        }
-        if self.precision == "int4":
-            quantization_config = GPTQConfig(bits=4, disable_exllama=True)
-            kwargs["quantization_config"] = quantization_config
-            kwargs["load_gptq_on_cpu"] = True
-            kwargs["device_map"] = "cpu"
+        kwargs = {"torch_dtype": torch.float, "trust_remote_code": True}
         falcon_model = AutoModelForCausalLM.from_pretrained(
             self.hf_model_path, **kwargs
         )
-        if self.precision == "int4":
-            falcon_model = falcon_model.to(torch.float32)
         return falcon_model
 
-    def compile(self):
+    def compile_falcon(self):
         if args.use_precompiled_model:
             if not self.falcon_vmfb_path.exists():
                 # Downloading VMFB from shark_tank
@@ -706,37 +120,37 @@ class UnshardedFalcon(SharkLLMBase):
             if vmfb is not None:
                 return vmfb
 
-        print(f"[DEBUG] vmfb not found at {self.falcon_vmfb_path.absolute()}")
+        print(
+            f"[DEBUG] vmfb not found at {self.falcon_vmfb_path.absolute()}. Trying to work with"
+            f"[DEBUG] mlir path { self.falcon_mlir_path} {'exists' if self.falcon_mlir_path.exists() else 'does not exist'}"
+        )
         if self.falcon_mlir_path.exists():
-            print(f"[DEBUG] mlir found at {self.falcon_mlir_path.absolute()}")
             with open(self.falcon_mlir_path, "rb") as f:
                 bytecode = f.read()
         else:
             mlir_generated = False
-            print(
-                f"[DEBUG] mlir not found at {self.falcon_mlir_path.absolute()}"
+            # Downloading MLIR from shark_tank
+            download_public_file(
+                "gs://shark_tank/falcon/"
+                + "falcon_"
+                + args.falcon_variant_to_use
+                + "_"
+                + self.precision
+                + ".mlir",
+                self.falcon_mlir_path.absolute(),
+                single_file=True,
             )
-            if args.load_mlir_from_shark_tank:
-                # Downloading MLIR from shark_tank
-                print(f"[DEBUG] Trying to download mlir from shark_tank")
-                download_public_file(
-                    "gs://shark_tank/falcon/"
-                    + "falcon_"
-                    + args.falcon_variant_to_use
-                    + "_"
-                    + self.precision
-                    + ".mlir",
-                    self.falcon_mlir_path.absolute(),
-                    single_file=True,
+            if self.falcon_mlir_path.exists():
+                with open(self.falcon_mlir_path, "rb") as f:
+                    bytecode = f.read()
+                mlir_generated = True
+            else:
+                raise ValueError(
+                    f"MLIR not found at {self.falcon_mlir_path.absolute()}"
+                    " after downloading! Please check path and try again"
                 )
-                if self.falcon_mlir_path.exists():
-                    print(
-                        f"[DEBUG] mlir found at {self.falcon_mlir_path.absolute()}"
-                    )
-                    mlir_generated = True
 
             if not mlir_generated:
-                print(f"[DEBUG] generating MLIR locally")
                 compilation_input_ids = torch.randint(
                     low=1, high=10000, size=(1, 100)
                 )
@@ -753,10 +167,9 @@ class UnshardedFalcon(SharkLLMBase):
                 ts_graph = import_with_fx(
                     model,
                     falconCompileInput,
-                    is_f16=self.precision in ["fp16", "int4"],
+                    is_f16=self.precision == "fp16",
                     f16_input_mask=[False, False],
                     mlir_type="torchscript",
-                    is_gptq=self.precision == "int4",
                 )
                 del model
                 print(f"[DEBUG] generating torch mlir")
@@ -776,37 +189,35 @@ class UnshardedFalcon(SharkLLMBase):
                 bytecode = bytecode_stream.getvalue()
                 del module
 
-                f_ = open(self.falcon_mlir_path, "wb")
-                f_.write(bytecode)
-                print("Saved falcon mlir at ", str(self.falcon_mlir_path))
+                print(f"[DEBUG] writing mlir to file")
+                with open(f"{self.model_name}.mlir", "wb") as f_:
+                    with redirect_stdout(f_):
+                        print(module.operation.get_asm())
                 f_.close()
-                del bytecode
 
         shark_module = SharkInference(
-            mlir_module=self.falcon_mlir_path,
-            device=self.device,
-            mlir_dialect="linalg",
+            mlir_module=bytecode, device=self.device, mlir_dialect="linalg"
         )
         path = shark_module.save_module(
             self.falcon_vmfb_path.parent.absolute(),
             self.falcon_vmfb_path.stem,
             extra_args=[
+                "--iree-hal-dump-executable-sources-to=ies",
                 "--iree-vm-target-truncate-unsupported-floats",
                 "--iree-codegen-check-ir-before-llvm-conversion=false",
                 "--iree-vm-bytecode-module-output-format=flatbuffer-binary",
-            ]
-            + [
-                "--iree-llvmcpu-use-fast-min-max-ops",
-            ]
-            if self.precision == "int4"
-            else [],
-            debug=self.debug,
+                "--iree-spirv-index-bits=64",
+            ],
         )
         print("Saved falcon vmfb at ", str(path))
         shark_module.load_module(path)
 
         return shark_module
 
+    def compile(self):
+        falcon_shark_model = self.compile_falcon()
+        return falcon_shark_model
+
     def generate(self, prompt):
         model_inputs = self.tokenizer(
             prompt,
@@ -934,11 +345,7 @@ class UnshardedFalcon(SharkLLMBase):
 
         all_text = prompt
 
-        start = time.time()
-        count = 0
         for i in range(self.max_num_tokens - 1):
-            count = count + 1
-
             next_token = self.generate_new_token()
             new_word = self.tokenizer.decode(
                 next_token.cpu().numpy(),
@@ -965,13 +372,6 @@ class UnshardedFalcon(SharkLLMBase):
                 ):
                     break
 
-        end = time.time()
-        print(
-            "\n\nTime taken is {:.2f} seconds/token\n".format(
-                (end - start) / count
-            )
-        )
-
         torch.cuda.empty_cache()
         gc.collect()
 
@@ -987,7 +387,7 @@ class UnshardedFalcon(SharkLLMBase):
                 (model_inputs["input_ids"], model_inputs["attention_mask"]),
             )
         )
-        if self.precision in ["fp16", "int4"]:
+        if self.precision == "fp16":
             outputs = outputs.to(dtype=torch.float32)
         next_token_logits = outputs
 
@@ -1066,39 +466,18 @@ if __name__ == "__main__":
         else Path(args.falcon_vmfb_path)
     )
 
-    if args.precision == "int4":
-        if args.falcon_variant_to_use == "180b":
-            hf_model_path_value = "TheBloke/Falcon-180B-Chat-GPTQ"
-        else:
-            hf_model_path_value = (
-                "TheBloke/falcon-"
-                + args.falcon_variant_to_use
-                + "-instruct-GPTQ"
-            )
-    else:
-        if args.falcon_variant_to_use == "180b":
-            hf_model_path_value = "tiiuae/falcon-180B-chat"
-        else:
-            hf_model_path_value = (
-                "tiiuae/falcon-" + args.falcon_variant_to_use + "-instruct"
-            )
+    falcon = Falcon(
+        "falcon_" + args.falcon_variant_to_use,
+        hf_model_path="tiiuae/falcon-"
+        + args.falcon_variant_to_use
+        + "-instruct",
+        device=args.device,
+        precision=args.precision,
+        falcon_mlir_path=falcon_mlir_path,
+        falcon_vmfb_path=falcon_vmfb_path,
+    )
 
-    if not args.sharded:
-        falcon = UnshardedFalcon(
-            model_name="falcon_" + args.falcon_variant_to_use,
-            hf_model_path=hf_model_path_value,
-            device=args.device,
-            precision=args.precision,
-            falcon_mlir_path=falcon_mlir_path,
-            falcon_vmfb_path=falcon_vmfb_path,
-        )
-    else:
-        falcon = ShardedFalcon(
-            model_name="falcon_" + args.falcon_variant_to_use,
-            hf_model_path=hf_model_path_value,
-            device=args.device,
-            precision=args.precision,
-        )
+    import gc
 
     default_prompt_text = "Girafatron is obsessed with giraffes, the most glorious animal on the face of this Earth. Giraftron believes all other animals are irrelevant when compared to the glorious majesty of the giraffe.\nDaniel: Hello, Girafatron!\nGirafatron:"
     continue_execution = True
@@ -1118,11 +497,7 @@ if __name__ == "__main__":
             prompt = input("Please enter the prompt text: ")
         print("\nPrompt Text: ", prompt)
 
-        prompt_template = f"""A helpful assistant who helps the user with any questions asked.
-        User: {prompt}
-        Assistant:"""
-
-        res_str = falcon.generate(prompt_template)
+        res_str = falcon.generate(prompt)
         torch.cuda.empty_cache()
         gc.collect()
         print(

apps/language_models/src/pipelines/minigpt4_pipeline.py

@@ -126,7 +126,7 @@ def is_url(input_url):
 import os
 import tempfile
 from shark.shark_inference import SharkInference
-from shark.shark_importer import import_with_fx, save_mlir
+from shark.shark_importer import import_with_fx
 import torch
 import torch_mlir
 from torch_mlir.compiler_utils import run_pipeline_with_repro_report
@@ -136,8 +136,7 @@ from brevitas_examples.llm.llm_quant.quantize import quantize_model
 from brevitas_examples.llm.llm_quant.run_utils import get_model_impl
 
 
-# fmt: off
-def quant〇matmul_rhs_group_quant〡shape(lhs: List[int], rhs: List[int], rhs_scale: List[int], rhs_zero_point: List[int], rhs_bit_width: int, rhs_group_size: int) -> List[int]:
+def brevitas〇matmul_rhs_group_quant〡shape(lhs: List[int], rhs: List[int], rhs_scale: List[int], rhs_zero_point: List[int], rhs_bit_width: int, rhs_group_size: int) -> List[int]:
     if len(lhs) == 3 and len(rhs) == 2:
         return [lhs[0], lhs[1], rhs[0]]
     elif len(lhs) == 2 and len(rhs) == 2:
@@ -146,21 +145,20 @@ def quant〇matmul_rhs_group_quant〡shape(lhs: List[int], rhs: List[int], rhs_s
         raise ValueError("Input shapes not supported.")
 
 
-def quant〇matmul_rhs_group_quant〡dtype(lhs_rank_dtype: Tuple[int, int], rhs_rank_dtype: Tuple[int, int], rhs_scale_rank_dtype: Tuple[int, int], rhs_zero_point_rank_dtype: Tuple[int, int], rhs_bit_width: int, rhs_group_size: int) -> int:
+def brevitas〇matmul_rhs_group_quant〡dtype(lhs_rank_dtype: Tuple[int, int], rhs_rank_dtype: Tuple[int, int], rhs_scale_rank_dtype: Tuple[int, int], rhs_zero_point_rank_dtype: Tuple[int, int], rhs_bit_width: int, rhs_group_size: int) -> int:
     # output dtype is the dtype of the lhs float input
     lhs_rank, lhs_dtype = lhs_rank_dtype
     return lhs_dtype
 
 
-def quant〇matmul_rhs_group_quant〡has_value_semantics(lhs, rhs, rhs_scale, rhs_zero_point, rhs_bit_width, rhs_group_size) -> None:
+def brevitas〇matmul_rhs_group_quant〡has_value_semantics(lhs, rhs, rhs_scale, rhs_zero_point, rhs_bit_width, rhs_group_size) -> None:
     return
 
 
 brevitas_matmul_rhs_group_quant_library = [
-    quant〇matmul_rhs_group_quant〡shape,
-    quant〇matmul_rhs_group_quant〡dtype,
-    quant〇matmul_rhs_group_quant〡has_value_semantics]
-# fmt: on
+    brevitas〇matmul_rhs_group_quant〡shape,
+    brevitas〇matmul_rhs_group_quant〡dtype,
+    brevitas〇matmul_rhs_group_quant〡has_value_semantics]
 
 
 def load_vmfb(extended_model_name, device, mlir_dialect, extra_args=[]):
@@ -178,7 +176,7 @@ def load_vmfb(extended_model_name, device, mlir_dialect, extra_args=[]):
 
 
 def compile_module(
-    shark_module, extended_model_name, generate_vmfb, extra_args=[], debug=False,
+    shark_module, extended_model_name, generate_vmfb, extra_args=[]
 ):
     if generate_vmfb:
         vmfb_path = os.path.join(os.getcwd(), extended_model_name + ".vmfb")
@@ -190,7 +188,7 @@ def compile_module(
                 "No vmfb found. Compiling and saving to {}".format(vmfb_path)
             )
             path = shark_module.save_module(
-                os.getcwd(), extended_model_name, extra_args, debug=debug
+                os.getcwd(), extended_model_name, extra_args
             )
             shark_module.load_module(path, extra_args=extra_args)
     else:
@@ -199,7 +197,7 @@ def compile_module(
 
 
 def compile_int_precision(
-    model, inputs, precision, device, generate_vmfb, extended_model_name, debug=False
+    model, inputs, precision, device, generate_vmfb, extended_model_name
 ):
     torchscript_module = import_with_fx(
         model,
@@ -211,7 +209,7 @@ def compile_int_precision(
         torchscript_module,
         inputs,
         output_type="torch",
-        backend_legal_ops=["quant.matmul_rhs_group_quant"],
+        backend_legal_ops=["brevitas.matmul_rhs_group_quant"],
         extra_library=brevitas_matmul_rhs_group_quant_library,
         use_tracing=False,
         verbose=False,
@@ -219,7 +217,7 @@ def compile_int_precision(
     print(f"[DEBUG] converting torch to linalg")
     run_pipeline_with_repro_report(
         mlir_module,
-        "builtin.module(func.func(torch-unpack-quant-tensor),func.func(torch-convert-custom-quant-op),torch-backend-to-linalg-on-tensors-backend-pipeline)",
+        "builtin.module(func.func(torch-unpack-torch-tensor),torch-backend-to-linalg-on-tensors-backend-pipeline)",
         description="Lowering Torch Backend IR -> Linalg-on-Tensors Backend IR",
     )
     from contextlib import redirect_stdout
@@ -235,12 +233,6 @@ def compile_int_precision(
     mlir_module = BytesIO(mlir_module)
     bytecode = mlir_module.read()
     print(f"Elided IR written for {extended_model_name}")
-    bytecode = save_mlir(
-        bytecode,
-        model_name=extended_model_name,
-        frontend="torch",
-        dir=os.getcwd(),
-    )
     return bytecode
     shark_module = SharkInference(
         mlir_module=bytecode, device=device, mlir_dialect="tm_tensor"
@@ -257,7 +249,6 @@ def compile_int_precision(
             extended_model_name=extended_model_name,
             generate_vmfb=generate_vmfb,
             extra_args=extra_args,
-            debug=debug,
         ),
         bytecode,
     )
@@ -301,7 +292,6 @@ def shark_compile_through_fx_int(
         device,
         generate_or_load_vmfb,
         extended_model_name,
-        debug,
     )
     extra_args = [
         "--iree-hal-dump-executable-sources-to=ies",

apps/language_models/src/pipelines/stablelm_pipeline.py

@@ -32,13 +32,11 @@ class SharkStableLM(SharkLLMBase):
         max_num_tokens=512,
         device="cuda",
         precision="fp32",
-        debug="False",
     ) -> None:
         super().__init__(model_name, hf_model_path, max_num_tokens)
         self.max_sequence_len = 256
         self.device = device
         self.precision = precision
-        self.debug = debug
         self.tokenizer = self.get_tokenizer()
         self.shark_model = self.compile()
 
@@ -113,7 +111,7 @@ class SharkStableLM(SharkLLMBase):
         shark_module.compile()
 
         path = shark_module.save_module(
-            vmfb_path.parent.absolute(), vmfb_path.stem, debug=self.debug
+            vmfb_path.parent.absolute(), vmfb_path.stem
         )
         print("Saved vmfb at ", str(path))
 

apps/language_models/utils.py

@@ -8,7 +8,7 @@ from shark.shark_downloader import download_public_file
 
 # expects a Path / str as arg
 # returns None if path not found or SharkInference module
-def get_vmfb_from_path(vmfb_path, device, mlir_dialect, device_id=None):
+def get_vmfb_from_path(vmfb_path, device, mlir_dialect):
     if not isinstance(vmfb_path, Path):
         vmfb_path = Path(vmfb_path)
 
@@ -20,7 +20,7 @@ def get_vmfb_from_path(vmfb_path, device, mlir_dialect, device_id=None):
     print("Loading vmfb from: ", vmfb_path)
     print("Device from get_vmfb_from_path - ", device)
     shark_module = SharkInference(
-        None, device=device, mlir_dialect=mlir_dialect, device_idx=device_id
+        None, device=device, mlir_dialect=mlir_dialect
     )
     shark_module.load_module(vmfb_path)
     print("Successfully loaded vmfb")
@@ -28,13 +28,7 @@ def get_vmfb_from_path(vmfb_path, device, mlir_dialect, device_id=None):
 
 
 def get_vmfb_from_config(
-    shark_container,
-    model,
-    precision,
-    device,
-    vmfb_path,
-    padding=None,
-    device_id=None,
+    shark_container, model, precision, device, vmfb_path, padding=None
 ):
     vmfb_url = (
         f"gs://shark_tank/{shark_container}/{model}_{precision}_{device}"
@@ -43,6 +37,4 @@ def get_vmfb_from_config(
         vmfb_url = vmfb_url + f"_{padding}"
     vmfb_url = vmfb_url + ".vmfb"
     download_public_file(vmfb_url, vmfb_path.absolute(), single_file=True)
-    return get_vmfb_from_path(
-        vmfb_path, device, "tm_tensor", device_id=device_id
-    )
+    return get_vmfb_from_path(vmfb_path, device, "tm_tensor")

apps/shark_studio/api/llm.py

@@ -1,91 +0,0 @@
-from turbine_models.custom_models import stateless_llama
-from shark.iree_utils.compile_utils import get_iree_compiled_module
-from apps.shark_studio.api.utils import get_resource_path
-import iree.runtime as ireert
-import gc
-import torch
-
-llm_model_map = {
-    "llama2_7b": {
-        "initializer": stateless_llama.export_transformer_model,
-        "hf_model_name": "meta-llama/Llama-2-7b-chat-hf",
-        "stop_token": 2,
-        "max_tokens": 4096,
-    }
-}
-
-
-class LanguageModel:
-    def __init__(
-        self, model_name, hf_auth_token=None, device=None, precision="fp32"
-    ):
-        print(llm_model_map[model_name])
-        self.hf_model_name = llm_model_map[model_name]["hf_model_name"]
-        self.torch_ir, self.tokenizer = llm_model_map[model_name][
-            "initializer"
-        ](self.hf_model_name, hf_auth_token, compile_to="torch")
-        self.tempfile_name = get_resource_path("llm.torch.tempfile")
-        with open(self.tempfile_name, "w+") as f:
-            f.write(self.torch_ir)
-        del self.torch_ir
-        gc.collect()
-
-        self.device = device
-        self.precision = precision
-        self.max_tokens = llm_model_map[model_name]["max_tokens"]
-        self.iree_module_dict = None
-        self.compile()
-
-    def compile(self) -> None:
-        # this comes with keys: "vmfb", "config", and "temp_file_to_unlink".
-        self.iree_module_dict = get_iree_compiled_module(
-            self.tempfile_name, device=self.device, frontend="torch"
-        )
-        # TODO: delete the temp file
-
-    def chat(self, prompt):
-        history = []
-        for iter in range(self.max_tokens):
-            input_tensor = self.tokenizer(
-                prompt, return_tensors="pt"
-            ).input_ids
-            device_inputs = [
-                ireert.asdevicearray(
-                    self.iree_module_dict["config"], input_tensor
-                )
-            ]
-            if iter == 0:
-                token = torch.tensor(
-                    self.iree_module_dict["vmfb"]["run_initialize"](
-                        *device_inputs
-                    ).to_host()[0][0]
-                )
-            else:
-                token = torch.tensor(
-                    self.iree_module_dict["vmfb"]["run_forward"](
-                        *device_inputs
-                    ).to_host()[0][0]
-                )
-
-            history.append(token)
-            yield self.tokenizer.decode(history)
-
-            if token == llm_model_map["llama2_7b"]["stop_token"]:
-                break
-
-        for i in range(len(history)):
-            if type(history[i]) != int:
-                history[i] = int(history[i])
-        result_output = self.tokenizer.decode(history)
-        yield result_output
-
-
-if __name__ == "__main__":
-    lm = LanguageModel(
-        "llama2_7b",
-        hf_auth_token="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",
-        device="cpu-task",
-    )
-    print("model loaded")
-    for i in lm.chat("Hello, I am a robot."):
-        print(i)

apps/shark_studio/api/utils.py

@@ -1,14 +0,0 @@
-import os
-import sys
-
-
-def get_available_devices():
-    return ["cpu-task"]
-
-
-def get_resource_path(relative_path):
-    """Get absolute path to resource, works for dev and for PyInstaller"""
-    base_path = getattr(
-        sys, "_MEIPASS", os.path.dirname(os.path.abspath(__file__))
-    )
-    return os.path.join(base_path, relative_path)

apps/shark_studio/web/index.py

@@ -1,428 +0,0 @@
-from multiprocessing import Process, freeze_support
-import os
-import sys
-import logging
-from ui.chat import chat_element
-
-if sys.platform == "darwin":
-    os.environ["DYLD_LIBRARY_PATH"] = "/usr/local/lib"
-    # import before IREE to avoid MLIR library issues
-    import torch_mlir
-
-# import PIL, transformers, sentencepiece  # ensures inclusion in pysintaller exe generation
-# from apps.stable_diffusion.src import args, clear_all
-# import apps.stable_diffusion.web.utils.global_obj as global_obj
-
-
-def launch_app(address):
-    from tkinter import Tk
-    import webview
-
-    window = Tk()
-
-    # get screen width and height of display and make it more reasonably
-    # sized as we aren't making it full-screen or maximized
-    width = int(window.winfo_screenwidth() * 0.81)
-    height = int(window.winfo_screenheight() * 0.91)
-    webview.create_window(
-        "SHARK AI Studio",
-        url=address,
-        width=width,
-        height=height,
-        text_select=True,
-    )
-    webview.start(private_mode=False, storage_path=os.getcwd())
-
-
-if __name__ == "__main__":
-    # if args.debug:
-    logging.basicConfig(level=logging.DEBUG)
-    # required to do multiprocessing in a pyinstaller freeze
-    freeze_support()
-    #    if args.api or "api" in args.ui.split(","):
-    #        from apps.stable_diffusion.web.ui import (
-    #            txt2img_api,
-    #            img2img_api,
-    #            upscaler_api,
-    #            inpaint_api,
-    #            outpaint_api,
-    #            llm_chat_api,
-    #        )
-    #
-    #        from fastapi import FastAPI, APIRouter
-    #        import uvicorn
-    #
-    #        # init global sd pipeline and config
-    #        global_obj._init()
-    #
-    #        app = FastAPI()
-    #        app.add_api_route("/sdapi/v1/txt2img", txt2img_api, methods=["post"])
-    #        app.add_api_route("/sdapi/v1/img2img", img2img_api, methods=["post"])
-    #        app.add_api_route("/sdapi/v1/inpaint", inpaint_api, methods=["post"])
-    #        app.add_api_route("/sdapi/v1/outpaint", outpaint_api, methods=["post"])
-    #        app.add_api_route("/sdapi/v1/upscaler", upscaler_api, methods=["post"])
-    #
-    #        # chat APIs needed for compatibility with multiple extensions using OpenAI API
-    #        app.add_api_route(
-    #            "/v1/chat/completions", llm_chat_api, methods=["post"]
-    #        )
-    #        app.add_api_route("/v1/completions", llm_chat_api, methods=["post"])
-    #        app.add_api_route("/chat/completions", llm_chat_api, methods=["post"])
-    #        app.add_api_route("/completions", llm_chat_api, methods=["post"])
-    #        app.add_api_route(
-    #            "/v1/engines/codegen/completions", llm_chat_api, methods=["post"]
-    #        )
-    #        app.include_router(APIRouter())
-    #        uvicorn.run(app, host="0.0.0.0", port=args.server_port)
-    #        sys.exit(0)
-    #
-    # Setup to use shark_tmp for gradio's temporary image files and clear any
-    # existing temporary images there if they exist. Then we can import gradio.
-    # It has to be in this order or gradio ignores what we've set up.
-    # from apps.stable_diffusion.web.utils.gradio_configs import (
-    #    config_gradio_tmp_imgs_folder,
-    # )
-
-    # config_gradio_tmp_imgs_folder()
-    import gradio as gr
-
-    # Create custom models folders if they don't exist
-    # from apps.stable_diffusion.web.ui.utils import create_custom_models_folders
-
-    # create_custom_models_folders()
-
-    def resource_path(relative_path):
-        """Get absolute path to resource, works for dev and for PyInstaller"""
-        base_path = getattr(
-            sys, "_MEIPASS", os.path.dirname(os.path.abspath(__file__))
-        )
-        return os.path.join(base_path, relative_path)
-
-    dark_theme = resource_path("ui/css/sd_dark_theme.css")
-
-    # from apps.stable_diffusion.web.ui import (
-    # txt2img_web,
-    # txt2img_custom_model,
-    # txt2img_gallery,
-    # txt2img_png_info_img,
-    # txt2img_status,
-    # txt2img_sendto_img2img,
-    # txt2img_sendto_inpaint,
-    # txt2img_sendto_outpaint,
-    # txt2img_sendto_upscaler,
-    ## h2ogpt_upload,
-    ## h2ogpt_web,
-    # img2img_web,
-    # img2img_custom_model,
-    # img2img_gallery,
-    # img2img_init_image,
-    # img2img_status,
-    # img2img_sendto_inpaint,
-    # img2img_sendto_outpaint,
-    # img2img_sendto_upscaler,
-    # inpaint_web,
-    # inpaint_custom_model,
-    # inpaint_gallery,
-    # inpaint_init_image,
-    # inpaint_status,
-    # inpaint_sendto_img2img,
-    # inpaint_sendto_outpaint,
-    # inpaint_sendto_upscaler,
-    # outpaint_web,
-    # outpaint_custom_model,
-    # outpaint_gallery,
-    # outpaint_init_image,
-    # outpaint_status,
-    # outpaint_sendto_img2img,
-    # outpaint_sendto_inpaint,
-    # outpaint_sendto_upscaler,
-    # upscaler_web,
-    # upscaler_custom_model,
-    # upscaler_gallery,
-    # upscaler_init_image,
-    # upscaler_status,
-    # upscaler_sendto_img2img,
-    # upscaler_sendto_inpaint,
-    # upscaler_sendto_outpaint,
-    ##  lora_train_web,
-    ##  model_web,
-    ##  model_config_web,
-    # hf_models,
-    # modelmanager_sendto_txt2img,
-    # modelmanager_sendto_img2img,
-    # modelmanager_sendto_inpaint,
-    # modelmanager_sendto_outpaint,
-    # modelmanager_sendto_upscaler,
-    # stablelm_chat,
-    # minigpt4_web,
-    # outputgallery_web,
-    # outputgallery_tab_select,
-    # outputgallery_watch,
-    # outputgallery_filename,
-    # outputgallery_sendto_txt2img,
-    # outputgallery_sendto_img2img,
-    # outputgallery_sendto_inpaint,
-    # outputgallery_sendto_outpaint,
-    # outputgallery_sendto_upscaler,
-    # )
-
-    # init global sd pipeline and config
-    # global_obj._init()
-
-    def register_button_click(button, selectedid, inputs, outputs):
-        button.click(
-            lambda x: (
-                x[0]["name"] if len(x) != 0 else None,
-                gr.Tabs.update(selected=selectedid),
-            ),
-            inputs,
-            outputs,
-        )
-
-    def register_modelmanager_button(button, selectedid, inputs, outputs):
-        button.click(
-            lambda x: (
-                "None",
-                x,
-                gr.Tabs.update(selected=selectedid),
-            ),
-            inputs,
-            outputs,
-        )
-
-    def register_outputgallery_button(button, selectedid, inputs, outputs):
-        button.click(
-            lambda x: (
-                x,
-                gr.Tabs.update(selected=selectedid),
-            ),
-            inputs,
-            outputs,
-        )
-
-    with gr.Blocks(
-        css=dark_theme, analytics_enabled=False, title="Stable Diffusion"
-    ) as sd_web:
-        with gr.Tabs() as tabs:
-            # NOTE: If adding, removing, or re-ordering tabs, make sure that they
-            # have a unique id that doesn't clash with any of the other tabs,
-            # and that the order in the code here is the order they should
-            # appear in the ui, as the id value doesn't determine the order.
-
-            # Where possible, avoid changing the id of any tab that is the
-            # destination of one of the 'send to' buttons. If you do have to change
-            # that id, make sure you update the relevant register_button_click calls
-            # further down with the new id.
-            # with gr.TabItem(label="Text-to-Image", id=0):
-            #    txt2img_web.render()
-            # with gr.TabItem(label="Image-to-Image", id=1):
-            #    img2img_web.render()
-            # with gr.TabItem(label="Inpainting", id=2):
-            #    inpaint_web.render()
-            # with gr.TabItem(label="Outpainting", id=3):
-            #    outpaint_web.render()
-            # with gr.TabItem(label="Upscaler", id=4):
-            #    upscaler_web.render()
-            # if args.output_gallery:
-            #    with gr.TabItem(label="Output Gallery", id=5) as og_tab:
-            #        outputgallery_web.render()
-
-            #    # extra output gallery configuration
-            #    outputgallery_tab_select(og_tab.select)
-            #    outputgallery_watch(
-            #        [
-            #            txt2img_status,
-            #            img2img_status,
-            #            inpaint_status,
-            #            outpaint_status,
-            #            upscaler_status,
-            #        ]
-            #    )
-            ##  with gr.TabItem(label="Model Manager", id=6):
-            ##      model_web.render()
-            ##  with gr.TabItem(label="LoRA Training (Experimental)", id=7):
-            ##      lora_train_web.render()
-            with gr.TabItem(label="Chat Bot", id=0):
-                chat_element.render()
-            ##  with gr.TabItem(
-            ##      label="Generate Sharding Config (Experimental)", id=9
-            ##  ):
-            ##      model_config_web.render()
-            # with gr.TabItem(label="MultiModal (Experimental)", id=10):
-            #    minigpt4_web.render()
-            # with gr.TabItem(label="DocuChat Upload", id=11):
-            #     h2ogpt_upload.render()
-            # with gr.TabItem(label="DocuChat(Experimental)", id=12):
-            #     h2ogpt_web.render()
-
-        # send to buttons
-        # register_button_click(
-        #    txt2img_sendto_img2img,
-        #    1,
-        #    [txt2img_gallery],
-        #    [img2img_init_image, tabs],
-        # )
-        # register_button_click(
-        #    txt2img_sendto_inpaint,
-        #    2,
-        #    [txt2img_gallery],
-        #    [inpaint_init_image, tabs],
-        # )
-        # register_button_click(
-        #    txt2img_sendto_outpaint,
-        #    3,
-        #    [txt2img_gallery],
-        #    [outpaint_init_image, tabs],
-        # )
-        # register_button_click(
-        #    txt2img_sendto_upscaler,
-        #    4,
-        #    [txt2img_gallery],
-        #    [upscaler_init_image, tabs],
-        # )
-        # register_button_click(
-        #    img2img_sendto_inpaint,
-        #    2,
-        #    [img2img_gallery],
-        #    [inpaint_init_image, tabs],
-        # )
-        # register_button_click(
-        #    img2img_sendto_outpaint,
-        #    3,
-        #    [img2img_gallery],
-        #    [outpaint_init_image, tabs],
-        # )
-        # register_button_click(
-        #    img2img_sendto_upscaler,
-        #    4,
-        #    [img2img_gallery],
-        #    [upscaler_init_image, tabs],
-        # )
-        # register_button_click(
-        #    inpaint_sendto_img2img,
-        #    1,
-        #    [inpaint_gallery],
-        #    [img2img_init_image, tabs],
-        # )
-        # register_button_click(
-        #    inpaint_sendto_outpaint,
-        #    3,
-        #    [inpaint_gallery],
-        #    [outpaint_init_image, tabs],
-        # )
-        # register_button_click(
-        #    inpaint_sendto_upscaler,
-        #    4,
-        #    [inpaint_gallery],
-        #    [upscaler_init_image, tabs],
-        # )
-        # register_button_click(
-        #    outpaint_sendto_img2img,
-        #    1,
-        #    [outpaint_gallery],
-        #    [img2img_init_image, tabs],
-        # )
-        # register_button_click(
-        #    outpaint_sendto_inpaint,
-        #    2,
-        #    [outpaint_gallery],
-        #    [inpaint_init_image, tabs],
-        # )
-        # register_button_click(
-        #    outpaint_sendto_upscaler,
-        #    4,
-        #    [outpaint_gallery],
-        #    [upscaler_init_image, tabs],
-        # )
-        # register_button_click(
-        #    upscaler_sendto_img2img,
-        #    1,
-        #    [upscaler_gallery],
-        #    [img2img_init_image, tabs],
-        # )
-        # register_button_click(
-        #    upscaler_sendto_inpaint,
-        #    2,
-        #    [upscaler_gallery],
-        #    [inpaint_init_image, tabs],
-        # )
-        # register_button_click(
-        #    upscaler_sendto_outpaint,
-        #    3,
-        #    [upscaler_gallery],
-        #    [outpaint_init_image, tabs],
-        # )
-        # if args.output_gallery:
-        #    register_outputgallery_button(
-        #        outputgallery_sendto_txt2img,
-        #        0,
-        #        [outputgallery_filename],
-        #        [txt2img_png_info_img, tabs],
-        #    )
-        #    register_outputgallery_button(
-        #        outputgallery_sendto_img2img,
-        #        1,
-        #        [outputgallery_filename],
-        #        [img2img_init_image, tabs],
-        #    )
-        #    register_outputgallery_button(
-        #        outputgallery_sendto_inpaint,
-        #        2,
-        #        [outputgallery_filename],
-        #        [inpaint_init_image, tabs],
-        #    )
-        #    register_outputgallery_button(
-        #        outputgallery_sendto_outpaint,
-        #        3,
-        #        [outputgallery_filename],
-        #        [outpaint_init_image, tabs],
-        #    )
-        #    register_outputgallery_button(
-        #        outputgallery_sendto_upscaler,
-        #        4,
-        #        [outputgallery_filename],
-        #        [upscaler_init_image, tabs],
-        #    )
-        # register_modelmanager_button(
-        #    modelmanager_sendto_txt2img,
-        #    0,
-        #    [hf_models],
-        #    [txt2img_custom_model, tabs],
-        # )
-        # register_modelmanager_button(
-        #    modelmanager_sendto_img2img,
-        #    1,
-        #    [hf_models],
-        #    [img2img_custom_model, tabs],
-        # )
-        # register_modelmanager_button(
-        #    modelmanager_sendto_inpaint,
-        #    2,
-        #    [hf_models],
-        #    [inpaint_custom_model, tabs],
-        # )
-        # register_modelmanager_button(
-        #    modelmanager_sendto_outpaint,
-        #    3,
-        #    [hf_models],
-        #    [outpaint_custom_model, tabs],
-        # )
-        # register_modelmanager_button(
-        #    modelmanager_sendto_upscaler,
-        #    4,
-        #    [hf_models],
-        #    [upscaler_custom_model, tabs],
-        # )
-
-    sd_web.queue()
-    # if args.ui == "app":
-    #    t = Process(
-    #        target=launch_app, args=[f"http://localhost:{args.server_port}"]
-    #    )
-    #    t.start()
-    sd_web.launch(
-        share=True,
-        inbrowser=True,
-        server_name="0.0.0.0",
-        server_port=11911,  # args.server_port,
-    )

apps/shark_studio/web/ui/chat.py

@@ -1,517 +0,0 @@
-import gradio as gr
-import os
-from pathlib import Path
-from datetime import datetime as dt
-import json
-import sys
-from apps.shark_studio.api.utils import (
-    get_available_devices,
-)
-from apps.shark_studio.api.llm import (
-    llm_model_map,
-    LanguageModel,
-)
-
-
-def user(message, history):
-    # Append the user's message to the conversation history
-    return "", history + [[message, ""]]
-
-
-language_model = None
-
-
-# NOTE: Each `model_name` should have its own start message
-start_message = {
-    "llama2_7b": (
-        "You are a helpful, respectful and honest assistant. Always answer "
-        "as helpfully as possible, while being safe. Your answers should not "
-        "include any harmful, unethical, racist, sexist, toxic, dangerous, or "
-        "illegal content. Please ensure that your responses are socially "
-        "unbiased and positive in nature. If a question does not make any "
-        "sense, or is not factually coherent, explain why instead of "
-        "answering something not correct. If you don't know the answer "
-        "to a question, please don't share false information."
-    ),
-    "llama2_13b": (
-        "You are a helpful, respectful and honest assistant. Always answer "
-        "as helpfully as possible, while being safe. Your answers should not "
-        "include any harmful, unethical, racist, sexist, toxic, dangerous, or "
-        "illegal content. Please ensure that your responses are socially "
-        "unbiased and positive in nature. If a question does not make any "
-        "sense, or is not factually coherent, explain why instead of "
-        "answering something not correct. If you don't know the answer "
-        "to a question, please don't share false information."
-    ),
-    "llama2_70b": (
-        "You are a helpful, respectful and honest assistant. Always answer "
-        "as helpfully as possible, while being safe. Your answers should not "
-        "include any harmful, unethical, racist, sexist, toxic, dangerous, or "
-        "illegal content. Please ensure that your responses are socially "
-        "unbiased and positive in nature. If a question does not make any "
-        "sense, or is not factually coherent, explain why instead of "
-        "answering something not correct. If you don't know the answer "
-        "to a question, please don't share false information."
-    ),
-    "vicuna": (
-        "A chat between a curious user and an artificial intelligence "
-        "assistant. The assistant gives helpful, detailed, and "
-        "polite answers to the user's questions.\n"
-    ),
-}
-
-
-def create_prompt(model_name, history, prompt_prefix):
-    return ""
-    system_message = ""
-    if prompt_prefix:
-        system_message = start_message[model_name]
-
-    if "llama2" in model_name:
-        B_INST, E_INST = "[INST]", "[/INST]"
-        B_SYS, E_SYS = "<<SYS>>\n", "\n<</SYS>>\n\n"
-        conversation = "".join(
-            [f"{B_INST} {item[0]} {E_INST} {item[1]} " for item in history[1:]]
-        )
-        if prompt_prefix:
-            msg = f"{B_INST} {B_SYS}{system_message}{E_SYS}{history[0][0]} {E_INST} {history[0][1]} {conversation}"
-        else:
-            msg = f"{B_INST} {history[0][0]} {E_INST} {history[0][1]} {conversation}"
-    elif model_name in ["vicuna"]:
-        conversation = "".join(
-            [
-                "".join(["<|USER|>" + item[0], "<|ASSISTANT|>" + item[1]])
-                for item in history
-            ]
-        )
-        msg = system_message + conversation
-        msg = msg.strip()
-    else:
-        conversation = "".join(
-            ["".join([item[0], item[1]]) for item in history]
-        )
-        msg = system_message + conversation
-        msg = msg.strip()
-    return msg
-
-
-def get_default_config():
-    return False
-    import torch
-    from transformers import AutoTokenizer
-
-    hf_model_path = "TheBloke/vicuna-7B-1.1-HF"
-    tokenizer = AutoTokenizer.from_pretrained(hf_model_path, use_fast=False)
-    compilation_prompt = "".join(["0" for _ in range(17)])
-    compilation_input_ids = tokenizer(
-        compilation_prompt,
-        return_tensors="pt",
-    ).input_ids
-    compilation_input_ids = torch.tensor(compilation_input_ids).reshape(
-        [1, 19]
-    )
-    firstVicunaCompileInput = (compilation_input_ids,)
-    from apps.language_models.src.model_wrappers.vicuna_model import (
-        CombinedModel,
-    )
-    from shark.shark_generate_model_config import GenerateConfigFile
-
-    model = CombinedModel()
-    c = GenerateConfigFile(model, 1, ["gpu_id"], firstVicunaCompileInput)
-    c.split_into_layers()
-
-
-# model_vmfb_key = ""
-
-
-def chat_fn(
-    prompt_prefix,
-    history,
-    model,
-    device,
-    precision,
-    download_vmfb,
-    config_file,
-    cli=False,
-    progress=gr.Progress(),
-):
-    global language_model
-    if language_model is None:
-        language_model = LanguageModel(
-            model, device=device, precision=precision
-        )
-
-    language_model.chat(prompt_prefix)
-    return "", ""
-    global past_key_values
-    global model_vmfb_key
-
-    device_id = None
-    model_name, model_path = list(map(str.strip, model.split("=>")))
-    if "cuda" in device:
-        device = "cuda"
-    elif "sync" in device:
-        device = "cpu-sync"
-    elif "task" in device:
-        device = "cpu-task"
-    elif "vulkan" in device:
-        device_id = int(device.split("://")[1])
-        device = "vulkan"
-    elif "rocm" in device:
-        device = "rocm"
-    else:
-        print("unrecognized device")
-
-    from apps.language_models.scripts.vicuna import ShardedVicuna
-    from apps.language_models.scripts.vicuna import UnshardedVicuna
-    from apps.stable_diffusion.src import args
-
-    new_model_vmfb_key = f"{model_name}#{model_path}#{device}#{device_id}#{precision}#{download_vmfb}"
-    if vicuna_model is None or new_model_vmfb_key != model_vmfb_key:
-        model_vmfb_key = new_model_vmfb_key
-        max_toks = 128 if model_name == "codegen" else 512
-
-        # get iree flags that need to be overridden, from commandline args
-        _extra_args = []
-        # vulkan target triple
-        vulkan_target_triple = args.iree_vulkan_target_triple
-        from shark.iree_utils.vulkan_utils import (
-            get_all_vulkan_devices,
-            get_vulkan_target_triple,
-        )
-
-        if device == "vulkan":
-            vulkaninfo_list = get_all_vulkan_devices()
-            if vulkan_target_triple == "":
-                # We already have the device_id extracted via WebUI, so we directly use
-                # that to find the target triple.
-                vulkan_target_triple = get_vulkan_target_triple(
-                    vulkaninfo_list[device_id]
-                )
-            _extra_args.append(
-                f"-iree-vulkan-target-triple={vulkan_target_triple}"
-            )
-            if "rdna" in vulkan_target_triple:
-                flags_to_add = [
-                    "--iree-spirv-index-bits=64",
-                ]
-                _extra_args = _extra_args + flags_to_add
-
-            if device_id is None:
-                id = 0
-                for device in vulkaninfo_list:
-                    target_triple = get_vulkan_target_triple(
-                        vulkaninfo_list[id]
-                    )
-                    if target_triple == vulkan_target_triple:
-                        device_id = id
-                        break
-                    id += 1
-
-                assert (
-                    device_id
-                ), f"no vulkan hardware for target-triple '{vulkan_target_triple}' exists"
-            print(f"Will use vulkan target triple : {vulkan_target_triple}")
-
-        elif "rocm" in device:
-            # add iree rocm flags
-            _extra_args.append(
-                f"--iree-rocm-target-chip={args.iree_rocm_target_chip}"
-            )
-            print(f"extra args = {_extra_args}")
-
-        if model_name == "vicuna4":
-            vicuna_model = ShardedVicuna(
-                model_name,
-                hf_model_path=model_path,
-                device=device,
-                precision=precision,
-                max_num_tokens=max_toks,
-                compressed=True,
-                extra_args_cmd=_extra_args,
-            )
-        else:
-            #  if config_file is None:
-            vicuna_model = UnshardedVicuna(
-                model_name,
-                hf_model_path=model_path,
-                hf_auth_token=args.hf_auth_token,
-                device=device,
-                vulkan_target_triple=vulkan_target_triple,
-                precision=precision,
-                max_num_tokens=max_toks,
-                download_vmfb=download_vmfb,
-                load_mlir_from_shark_tank=True,
-                extra_args_cmd=_extra_args,
-                device_id=device_id,
-            )
-
-    if vicuna_model is None:
-        sys.exit("Unable to instantiate the model object, exiting.")
-
-    prompt = create_prompt(model_name, history, prompt_prefix)
-
-    partial_text = ""
-    token_count = 0
-    total_time_ms = 0.001  # In order to avoid divide by zero error
-    prefill_time = 0
-    is_first = True
-    for text, msg, exec_time in progress.tqdm(
-        vicuna_model.generate(prompt, cli=cli),
-        desc="generating response",
-    ):
-        if msg is None:
-            if is_first:
-                prefill_time = exec_time
-                is_first = False
-            else:
-                total_time_ms += exec_time
-                token_count += 1
-            partial_text += text + " "
-            history[-1][1] = partial_text
-            yield history, f"Prefill: {prefill_time:.2f}"
-        elif "formatted" in msg:
-            history[-1][1] = text
-            tokens_per_sec = (token_count / total_time_ms) * 1000
-            yield history, f"Prefill: {prefill_time:.2f} seconds\n Decode: {tokens_per_sec:.2f} tokens/sec"
-        else:
-            sys.exit(
-                "unexpected message from the vicuna generate call, exiting."
-            )
-
-    return history, ""
-
-
-def llm_chat_api(InputData: dict):
-    return None
-    print(f"Input keys : {InputData.keys()}")
-    # print(f"model : {InputData['model']}")
-    is_chat_completion_api = (
-        "messages" in InputData.keys()
-    )  # else it is the legacy `completion` api
-    # For Debugging input data from API
-    # if is_chat_completion_api:
-    #     print(f"message -> role : {InputData['messages'][0]['role']}")
-    #     print(f"message -> content : {InputData['messages'][0]['content']}")
-    # else:
-    #     print(f"prompt : {InputData['prompt']}")
-    # print(f"max_tokens : {InputData['max_tokens']}") # Default to 128 for now
-    global vicuna_model
-    model_name = (
-        InputData["model"] if "model" in InputData.keys() else "codegen"
-    )
-    model_path = llm_model_map[model_name]
-    device = "cpu-task"
-    precision = "fp16"
-    max_toks = (
-        None
-        if "max_tokens" not in InputData.keys()
-        else InputData["max_tokens"]
-    )
-    if max_toks is None:
-        max_toks = 128 if model_name == "codegen" else 512
-
-    # make it working for codegen first
-    from apps.language_models.scripts.vicuna import (
-        UnshardedVicuna,
-    )
-
-    device_id = None
-    if vicuna_model == 0:
-        if "cuda" in device:
-            device = "cuda"
-        elif "sync" in device:
-            device = "cpu-sync"
-        elif "task" in device:
-            device = "cpu-task"
-        elif "vulkan" in device:
-            device_id = int(device.split("://")[1])
-            device = "vulkan"
-        else:
-            print("unrecognized device")
-
-        vicuna_model = UnshardedVicuna(
-            model_name,
-            hf_model_path=model_path,
-            device=device,
-            precision=precision,
-            max_num_tokens=max_toks,
-            download_vmfb=True,
-            load_mlir_from_shark_tank=True,
-            device_id=device_id,
-        )
-
-    # TODO: add role dict for different models
-    if is_chat_completion_api:
-        # TODO: add funtionality for multiple messages
-        prompt = create_prompt(
-            model_name, [(InputData["messages"][0]["content"], "")]
-        )
-    else:
-        prompt = InputData["prompt"]
-    print("prompt = ", prompt)
-
-    res = vicuna_model.generate(prompt)
-    res_op = None
-    for op in res:
-        res_op = op
-
-    if is_chat_completion_api:
-        choices = [
-            {
-                "index": 0,
-                "message": {
-                    "role": "assistant",
-                    "content": res_op,  # since we are yeilding the result
-                },
-                "finish_reason": "stop",  # or length
-            }
-        ]
-    else:
-        choices = [
-            {
-                "text": res_op,
-                "index": 0,
-                "logprobs": None,
-                "finish_reason": "stop",  # or length
-            }
-        ]
-    end_time = dt.now().strftime("%Y%m%d%H%M%S%f")
-    return {
-        "id": end_time,
-        "object": "chat.completion"
-        if is_chat_completion_api
-        else "text_completion",
-        "created": int(end_time),
-        "choices": choices,
-    }
-
-
-def view_json_file(file_obj):
-    content = ""
-    with open(file_obj.name, "r") as fopen:
-        content = fopen.read()
-    return content
-
-
-with gr.Blocks(title="Chat") as chat_element:
-    with gr.Row():
-        model_choices = list(llm_model_map.keys())
-        model = gr.Dropdown(
-            label="Select Model",
-            value=model_choices[0],
-            choices=model_choices,
-            allow_custom_value=True,
-        )
-        supported_devices = get_available_devices()
-        enabled = True
-        if len(supported_devices) == 0:
-            supported_devices = ["cpu-task"]
-        supported_devices = [x for x in supported_devices if "sync" not in x]
-        device = gr.Dropdown(
-            label="Device",
-            value=supported_devices[0],
-            choices=supported_devices,
-            interactive=enabled,
-            allow_custom_value=True,
-        )
-        precision = gr.Radio(
-            label="Precision",
-            value="int4",
-            choices=[
-                # "int4",
-                # "int8",
-                # "fp16",
-                "fp32",
-            ],
-            visible=False,
-        )
-        tokens_time = gr.Textbox(label="Tokens generated per second")
-        with gr.Column():
-            download_vmfb = gr.Checkbox(
-                label="Download vmfb from Shark tank if available",
-                value=True,
-                interactive=True,
-            )
-            prompt_prefix = gr.Checkbox(
-                label="Add System Prompt",
-                value=False,
-                interactive=True,
-            )
-
-    chatbot = gr.Chatbot(height=500)
-    with gr.Row():
-        with gr.Column():
-            msg = gr.Textbox(
-                label="Chat Message Box",
-                placeholder="Chat Message Box",
-                show_label=False,
-                interactive=enabled,
-                container=False,
-            )
-        with gr.Column():
-            with gr.Row():
-                submit = gr.Button("Submit", interactive=enabled)
-                stop = gr.Button("Stop", interactive=enabled)
-                clear = gr.Button("Clear", interactive=enabled)
-
-    with gr.Row(visible=False):
-        with gr.Group():
-            config_file = gr.File(
-                label="Upload sharding configuration", visible=False
-            )
-            json_view_button = gr.Button(label="View as JSON", visible=False)
-        json_view = gr.JSON(interactive=True, visible=False)
-        json_view_button.click(
-            fn=view_json_file, inputs=[config_file], outputs=[json_view]
-        )
-    submit_event = msg.submit(
-        fn=user,
-        inputs=[msg, chatbot],
-        outputs=[msg, chatbot],
-        show_progress=False,
-        queue=False,
-    ).then(
-        fn=chat_fn,
-        inputs=[
-            prompt_prefix,
-            chatbot,
-            model,
-            device,
-            precision,
-            download_vmfb,
-            config_file,
-        ],
-        outputs=[chatbot, tokens_time],
-        show_progress=False,
-        queue=True,
-    )
-    submit_click_event = submit.click(
-        fn=user,
-        inputs=[msg, chatbot],
-        outputs=[msg, chatbot],
-        show_progress=False,
-        queue=False,
-    ).then(
-        fn=chat_fn,
-        inputs=[
-            prompt_prefix,
-            chatbot,
-            model,
-            device,
-            precision,
-            download_vmfb,
-            config_file,
-        ],
-        outputs=[chatbot, tokens_time],
-        show_progress=False,
-        queue=True,
-    )
-    stop.click(
-        fn=None,
-        inputs=None,
-        outputs=None,
-        cancels=[submit_event, submit_click_event],
-        queue=False,
-    )
-    clear.click(lambda: None, None, [chatbot], queue=False)

apps/stable_diffusion/profiling_with_iree.md

@@ -7,16 +7,16 @@ Compile Commands FP32/FP16:
 
 ```shell
 Vulkan AMD: 
-iree-compile --iree-input-type=none --iree-hal-target-backends=vulkan --iree-vulkan-target-triple=rdna2-unknown-linux /path/to/input/mlir -o /path/to/output/vmfb
+iree-compile --iree-input-type=none --iree-hal-target-backends=vulkan --iree-vulkan-target-triple=rdna2-unknown-linux --iree-stream-resource-index-bits=64 --iree-vm-target-index-bits=64 /path/to/input/mlir -o /path/to/output/vmfb
 
 #  add --mlir-print-debuginfo --mlir-print-op-on-diagnostic=true for debug
 #  use –iree-input-type=auto or "mhlo_legacy" or "stablehlo" for TF models
 
 CUDA NVIDIA:
-iree-compile --iree-input-type=none --iree-hal-target-backends=cuda /path/to/input/mlir -o /path/to/output/vmfb
+iree-compile --iree-input-type=none --iree-hal-target-backends=cuda --iree-stream-resource-index-bits=64 --iree-vm-target-index-bits=64 /path/to/input/mlir -o /path/to/output/vmfb
 
 CPU:
-iree-compile --iree-input-type=none --iree-hal-target-backends=llvm-cpu /path/to/input/mlir -o /path/to/output/vmfb
+iree-compile --iree-input-type=none --iree-hal-target-backends=llvm-cpu  --iree-stream-resource-index-bits=64 --iree-vm-target-index-bits=64 /path/to/input/mlir -o /path/to/output/vmfb
 ```
 
 

apps/stable_diffusion/scripts/train_lora_word.py

@@ -34,7 +34,7 @@ from PIL import Image
 from tqdm.auto import tqdm
 from transformers import CLIPFeatureExtractor, CLIPTextModel, CLIPTokenizer
 from diffusers.loaders import AttnProcsLayers
-from diffusers.models.attention_processor import LoRAXFormersAttnProcessor
+from diffusers.models.cross_attention import LoRACrossAttnProcessor
 
 import torch_mlir
 from torch_mlir.dynamo import make_simple_dynamo_backend
@@ -287,7 +287,7 @@ def lora_train(
                 block_id = int(name[len("down_blocks.")])
                 hidden_size = unet.config.block_out_channels[block_id]
 
-            lora_attn_procs[name] = LoRAXFormersAttnProcessor(
+            lora_attn_procs[name] = LoRACrossAttnProcessor(
                 hidden_size=hidden_size,
                 cross_attention_dim=cross_attention_dim,
             )

apps/stable_diffusion/shark_studio_imports.py

@@ -15,8 +15,8 @@ pathex = [
 
 # datafiles for pyinstaller
 datas = []
+datas += collect_data_files("torch")
 datas += copy_metadata("torch")
-datas += copy_metadata("tokenizers")
 datas += copy_metadata("tqdm")
 datas += copy_metadata("regex")
 datas += copy_metadata("requests")
@@ -30,30 +30,26 @@ datas += copy_metadata("safetensors")
 datas += copy_metadata("Pillow")
 datas += copy_metadata("sentencepiece")
 datas += copy_metadata("pyyaml")
-datas += copy_metadata("huggingface-hub")
-datas += collect_data_files("torch")
 datas += collect_data_files("tokenizers")
 datas += collect_data_files("tiktoken")
 datas += collect_data_files("accelerate")
 datas += collect_data_files("diffusers")
 datas += collect_data_files("transformers")
 datas += collect_data_files("pytorch_lightning")
+datas += collect_data_files("opencv_python")
 datas += collect_data_files("skimage")
 datas += collect_data_files("gradio")
 datas += collect_data_files("gradio_client")
 datas += collect_data_files("iree")
-datas += collect_data_files("shark", include_py_files=True)
+datas += collect_data_files("google_cloud_storage")
+datas += collect_data_files("shark")
 datas += collect_data_files("timm", include_py_files=True)
-datas += collect_data_files("tqdm")
 datas += collect_data_files("tkinter")
 datas += collect_data_files("webview")
 datas += collect_data_files("sentencepiece")
 datas += collect_data_files("jsonschema")
 datas += collect_data_files("jsonschema_specifications")
 datas += collect_data_files("cpuinfo")
-datas += collect_data_files("langchain")
-datas += collect_data_files("cv2")
-datas += collect_data_files("einops")
 datas += [
     ("src/utils/resources/prompts.json", "resources"),
     ("src/utils/resources/model_db.json", "resources"),
@@ -76,13 +72,6 @@ datas += [
 hiddenimports = ["shark", "shark.shark_inference", "apps"]
 hiddenimports += [x for x in collect_submodules("skimage") if "tests" not in x]
 hiddenimports += [
-    x for x in collect_submodules("diffusers") if "tests" not in x
-]
-blacklist = ["tests", "convert"]
-hiddenimports += [
-    x
-    for x in collect_submodules("transformers")
-    if not any(kw in x for kw in blacklist)
+    x for x in collect_submodules("transformers") if "tests" not in x
 ]
 hiddenimports += [x for x in collect_submodules("iree") if "tests" not in x]
-hiddenimports += ["iree._runtime", "iree.compiler._mlir_libs._mlir.ir"]

apps/stable_diffusion/src/models/model_wrappers.py

@@ -8,7 +8,6 @@ import traceback
 import subprocess
 import sys
 import os
-import requests
 from apps.stable_diffusion.src.utils import (
     compile_through_fx,
     get_opt_flags,
@@ -17,7 +16,6 @@ from apps.stable_diffusion.src.utils import (
     preprocessCKPT,
     convert_original_vae,
     get_path_to_diffusers_checkpoint,
-    get_civitai_checkpoint,
     fetch_and_update_base_model_id,
     get_path_stem,
     get_extended_name,
@@ -96,19 +94,21 @@ class SharkifyStableDiffusionModel:
         self.height = height // 8
         self.width = width // 8
         self.batch_size = batch_size
-        self.custom_weights = custom_weights.strip()
+        self.custom_weights = custom_weights
         self.use_quantize = use_quantize
         if custom_weights != "":
-            if custom_weights.startswith("https://civitai.com/api/"):
-                # download the checkpoint from civitai if we don't already have it
-                weights_path = get_civitai_checkpoint(custom_weights)
-
-                # act as if we were given the local file as custom_weights originally
+            if "civitai" in custom_weights:
+                weights_id = custom_weights.split("/")[-1]
+                # TODO: use model name and identify file type by civitai rest api
+                weights_path = (
+                    str(Path.cwd()) + "/models/" + weights_id + ".safetensors"
+                )
+                if not os.path.isfile(weights_path):
+                    subprocess.run(
+                        ["wget", custom_weights, "-O", weights_path]
+                    )
                 custom_weights = get_path_to_diffusers_checkpoint(weights_path)
                 self.custom_weights = weights_path
-
-                # needed to ensure webui sets the correct model name metadata
-                args.ckpt_loc = weights_path
             else:
                 assert custom_weights.lower().endswith(
                     (".ckpt", ".safetensors")
@@ -116,7 +116,6 @@ class SharkifyStableDiffusionModel:
                 custom_weights = get_path_to_diffusers_checkpoint(
                     custom_weights
                 )
-
         self.model_id = model_id if custom_weights == "" else custom_weights
         # TODO: remove the following line when stable-diffusion-2-1 works
         if self.model_id == "stabilityai/stable-diffusion-2-1":
@@ -178,11 +177,9 @@ class SharkifyStableDiffusionModel:
             "unet",
             "unet512",
             "stencil_unet",
-            "stencil_unet_512",
             "vae",
             "vae_encode",
             "stencil_adaptor",
-            "stencil_adaptor_512",
         ]
         index = 0
         for model in sub_model_list:
@@ -342,7 +339,7 @@ class SharkifyStableDiffusionModel:
         )
         return shark_vae, vae_mlir
 
-    def get_controlled_unet(self, use_large=False):
+    def get_controlled_unet(self):
         class ControlledUnetModel(torch.nn.Module):
             def __init__(
                 self,
@@ -418,16 +415,6 @@ class SharkifyStableDiffusionModel:
         is_f16 = True if self.precision == "fp16" else False
 
         inputs = tuple(self.inputs["unet"])
-        model_name = "stencil_unet"
-        if use_large:
-            pad = (0, 0) * (len(inputs[2].shape) - 2)
-            pad = pad + (0, 512 - inputs[2].shape[1])
-            inputs = (
-                inputs[:2]
-                + (torch.nn.functional.pad(inputs[2], pad),)
-                + inputs[3:]
-            )
-            model_name = "stencil_unet_512"
         input_mask = [
             True,
             True,
@@ -450,19 +437,19 @@ class SharkifyStableDiffusionModel:
         shark_controlled_unet, controlled_unet_mlir = compile_through_fx(
             unet,
             inputs,
-            extended_model_name=self.model_name[model_name],
+            extended_model_name=self.model_name["stencil_unet"],
             is_f16=is_f16,
             f16_input_mask=input_mask,
             use_tuned=self.use_tuned,
             extra_args=get_opt_flags("unet", precision=self.precision),
             base_model_id=self.base_model_id,
-            model_name=model_name,
+            model_name="stencil_unet",
             precision=self.precision,
             return_mlir=self.return_mlir,
         )
         return shark_controlled_unet, controlled_unet_mlir
 
-    def get_control_net(self, use_large=False):
+    def get_control_net(self):
         class StencilControlNetModel(torch.nn.Module):
             def __init__(
                 self, model_id=self.use_stencil, low_cpu_mem_usage=False
@@ -510,34 +497,17 @@ class SharkifyStableDiffusionModel:
         is_f16 = True if self.precision == "fp16" else False
 
         inputs = tuple(self.inputs["stencil_adaptor"])
-        if use_large:
-            pad = (0, 0) * (len(inputs[2].shape) - 2)
-            pad = pad + (0, 512 - inputs[2].shape[1])
-            inputs = (
-                inputs[0],
-                inputs[1],
-                torch.nn.functional.pad(inputs[2], pad),
-                inputs[3],
-            )
-            save_dir = os.path.join(
-                self.sharktank_dir, self.model_name["stencil_adaptor_512"]
-            )
-        else:
-            save_dir = os.path.join(
-                self.sharktank_dir, self.model_name["stencil_adaptor"]
-            )
         input_mask = [True, True, True, True]
-        model_name = "stencil_adaptor" if use_large else "stencil_adaptor_512"
         shark_cnet, cnet_mlir = compile_through_fx(
             scnet,
             inputs,
-            extended_model_name=self.model_name[model_name],
+            extended_model_name=self.model_name["stencil_adaptor"],
             is_f16=is_f16,
             f16_input_mask=input_mask,
             use_tuned=self.use_tuned,
             extra_args=get_opt_flags("unet", precision=self.precision),
             base_model_id=self.base_model_id,
-            model_name=model_name,
+            model_name="stencil_adaptor",
             precision=self.precision,
             return_mlir=self.return_mlir,
         )
@@ -711,11 +681,8 @@ class SharkifyStableDiffusionModel:
                 return self.text_encoder(input)[0]
 
         clip_model = CLIPText(low_cpu_mem_usage=self.low_cpu_mem_usage)
-        save_dir = ""
+        save_dir = os.path.join(self.sharktank_dir, self.model_name["clip"])
         if self.debug:
-            save_dir = os.path.join(
-                self.sharktank_dir, self.model_name["clip"]
-            )
             os.makedirs(
                 save_dir,
                 exist_ok=True,
@@ -781,7 +748,7 @@ class SharkifyStableDiffusionModel:
             else:
                 return self.get_unet(use_large=use_large)
         else:
-            return self.get_controlled_unet(use_large=use_large)
+            return self.get_controlled_unet()
 
     def vae_encode(self):
         try:
@@ -880,14 +847,12 @@ class SharkifyStableDiffusionModel:
         except Exception as e:
             sys.exit(e)
 
-    def controlnet(self, use_large=False):
+    def controlnet(self):
         try:
             self.inputs["stencil_adaptor"] = self.get_input_info_for(
                 base_models["stencil_adaptor"]
             )
-            compiled_stencil_adaptor, controlnet_mlir = self.get_control_net(
-                use_large=use_large
-            )
+            compiled_stencil_adaptor, controlnet_mlir = self.get_control_net()
 
             check_compilation(compiled_stencil_adaptor, "Stencil")
             if self.return_mlir:

apps/stable_diffusion/src/pipelines/pipeline_shark_stable_diffusion_img2img.py

@@ -29,10 +29,6 @@ from apps.stable_diffusion.src.models import (
     SharkifyStableDiffusionModel,
     get_vae_encode,
 )
-from apps.stable_diffusion.src.utils import (
-    resamplers,
-    resampler_list,
-)
 
 
 class Image2ImagePipeline(StableDiffusionPipeline):
@@ -88,21 +84,13 @@ class Image2ImagePipeline(StableDiffusionPipeline):
         num_inference_steps,
         strength,
         dtype,
-        resample_type,
     ):
         # Pre process image -> get image encoded -> process latents
 
         # TODO: process with variable HxW combos
 
-        # Pre-process image
-        resample_type = (
-            resamplers[resample_type]
-            if resample_type in resampler_list
-            # Fallback to Lanczos
-            else Image.Resampling.LANCZOS
-        )
-
-        image = image.resize((width, height), resample=resample_type)
+        # Pre process image
+        image = image.resize((width, height))
         image_arr = np.stack([np.array(i) for i in (image,)], axis=0)
         image_arr = image_arr / 255.0
         image_arr = torch.from_numpy(image_arr).permute(0, 3, 1, 2).to(dtype)
@@ -159,7 +147,6 @@ class Image2ImagePipeline(StableDiffusionPipeline):
         cpu_scheduling,
         max_embeddings_multiples,
         use_stencil,
-        resample_type,
     ):
         # prompts and negative prompts must be a list.
         if isinstance(prompts, str):
@@ -199,7 +186,6 @@ class Image2ImagePipeline(StableDiffusionPipeline):
             num_inference_steps=num_inference_steps,
             strength=strength,
             dtype=dtype,
-            resample_type=resample_type,
         )
 
         # Get Image latents

apps/stable_diffusion/src/pipelines/pipeline_shark_stable_diffusion_stencil.py

@@ -58,7 +58,6 @@ class StencilPipeline(StableDiffusionPipeline):
     ):
         super().__init__(scheduler, sd_model, import_mlir, use_lora, ondemand)
         self.controlnet = None
-        self.controlnet_512 = None
 
     def load_controlnet(self):
         if self.controlnet is not None:
@@ -69,15 +68,6 @@ class StencilPipeline(StableDiffusionPipeline):
         del self.controlnet
         self.controlnet = None
 
-    def load_controlnet_512(self):
-        if self.controlnet_512 is not None:
-            return
-        self.controlnet_512 = self.sd_model.controlnet(use_large=True)
-
-    def unload_controlnet_512(self):
-        del self.controlnet_512
-        self.controlnet_512 = None
-
     def prepare_latents(
         self,
         batch_size,
@@ -121,12 +111,8 @@ class StencilPipeline(StableDiffusionPipeline):
         latent_history = [latents]
         text_embeddings = torch.from_numpy(text_embeddings).to(dtype)
         text_embeddings_numpy = text_embeddings.detach().numpy()
-        if text_embeddings.shape[1] <= self.model_max_length:
-            self.load_unet()
-            self.load_controlnet()
-        else:
-            self.load_unet_512()
-            self.load_controlnet_512()
+        self.load_unet()
+        self.load_controlnet()
         for i, t in tqdm(enumerate(total_timesteps)):
             step_start_time = time.time()
             timestep = torch.tensor([t]).to(dtype)
@@ -149,82 +135,43 @@ class StencilPipeline(StableDiffusionPipeline):
                 ).to(dtype)
             else:
                 latent_model_input_1 = latent_model_input
-            if text_embeddings.shape[1] <= self.model_max_length:
-                control = self.controlnet(
-                    "forward",
-                    (
-                        latent_model_input_1,
-                        timestep,
-                        text_embeddings,
-                        controlnet_hint,
-                    ),
-                    send_to_host=False,
-                )
-            else:
-                control = self.controlnet_512(
-                    "forward",
-                    (
-                        latent_model_input_1,
-                        timestep,
-                        text_embeddings,
-                        controlnet_hint,
-                    ),
-                    send_to_host=False,
-                )
+            control = self.controlnet(
+                "forward",
+                (
+                    latent_model_input_1,
+                    timestep,
+                    text_embeddings,
+                    controlnet_hint,
+                ),
+                send_to_host=False,
+            )
             timestep = timestep.detach().numpy()
             # Profiling Unet.
             profile_device = start_profiling(file_path="unet.rdc")
             # TODO: Pass `control` as it is to Unet. Same as TODO mentioned in model_wrappers.py.
-
-            if text_embeddings.shape[1] <= self.model_max_length:
-                noise_pred = self.unet(
-                    "forward",
-                    (
-                        latent_model_input,
-                        timestep,
-                        text_embeddings_numpy,
-                        guidance_scale,
-                        control[0],
-                        control[1],
-                        control[2],
-                        control[3],
-                        control[4],
-                        control[5],
-                        control[6],
-                        control[7],
-                        control[8],
-                        control[9],
-                        control[10],
-                        control[11],
-                        control[12],
-                    ),
-                    send_to_host=False,
-                )
-            else:
-                print(self.unet_512)
-                noise_pred = self.unet_512(
-                    "forward",
-                    (
-                        latent_model_input,
-                        timestep,
-                        text_embeddings_numpy,
-                        guidance_scale,
-                        control[0],
-                        control[1],
-                        control[2],
-                        control[3],
-                        control[4],
-                        control[5],
-                        control[6],
-                        control[7],
-                        control[8],
-                        control[9],
-                        control[10],
-                        control[11],
-                        control[12],
-                    ),
-                    send_to_host=False,
-                )
+            noise_pred = self.unet(
+                "forward",
+                (
+                    latent_model_input,
+                    timestep,
+                    text_embeddings_numpy,
+                    guidance_scale,
+                    control[0],
+                    control[1],
+                    control[2],
+                    control[3],
+                    control[4],
+                    control[5],
+                    control[6],
+                    control[7],
+                    control[8],
+                    control[9],
+                    control[10],
+                    control[11],
+                    control[12],
+                ),
+                send_to_host=False,
+            )
             end_profiling(profile_device)
 
             if cpu_scheduling:
@@ -244,9 +191,7 @@ class StencilPipeline(StableDiffusionPipeline):
 
         if self.ondemand:
             self.unload_unet()
-            self.unload_unet_512()
             self.unload_controlnet()
-            self.unload_controlnet_512()
         avg_step_time = step_time_sum / len(total_timesteps)
         self.log += f"\nAverage step time: {avg_step_time}ms/it"
 
@@ -273,7 +218,6 @@ class StencilPipeline(StableDiffusionPipeline):
         cpu_scheduling,
         max_embeddings_multiples,
         use_stencil,
-        resample_type,
     ):
         # Control Embedding check & conversion
         # TODO: 1. Change `num_images_per_prompt`.

apps/stable_diffusion/src/schedulers/shark_eulerdiscrete.py

@@ -84,6 +84,9 @@ class SharkEulerDiscreteScheduler(EulerDiscreteScheduler):
             iree_flags.append(
                 f"-iree-vulkan-target-triple={args.iree_vulkan_target_triple}"
             )
+        # Disable bindings fusion to work with moltenVK.
+        if sys.platform == "darwin":
+            iree_flags.append("-iree-stream-fuse-binding=false")
 
         def _import(self):
             scaling_model = ScalingModel()

apps/stable_diffusion/src/utils/__init__.py

@@ -41,8 +41,3 @@ from apps.stable_diffusion.src.utils.utils import (
     resize_stencil,
     _compile_module,
 )
-from apps.stable_diffusion.src.utils.civitai import get_civitai_checkpoint
-from apps.stable_diffusion.src.utils.resamplers import (
-    resamplers,
-    resampler_list,
-)

apps/stable_diffusion/src/utils/civitai.py

@@ -1,42 +0,0 @@
-import re
-import requests
-from apps.stable_diffusion.src.utils.stable_args import args
-
-from pathlib import Path
-from tqdm import tqdm
-
-
-def get_civitai_checkpoint(url: str):
-    with requests.get(url, allow_redirects=True, stream=True) as response:
-        response.raise_for_status()
-
-        # civitai api returns the filename in the content disposition
-        base_filename = re.findall(
-            '"([^"]*)"', response.headers["Content-Disposition"]
-        )[0]
-        destination_path = (
-            Path.cwd() / (args.ckpt_dir or "models") / base_filename
-        )
-
-        # we don't have this model downloaded yet
-        if not destination_path.is_file():
-            print(
-                f"downloading civitai model from {url} to {destination_path}"
-            )
-
-            size = int(response.headers["content-length"], 0)
-            progress_bar = tqdm(total=size, unit="iB", unit_scale=True)
-
-            with open(destination_path, "wb") as f:
-                for chunk in response.iter_content(chunk_size=65536):
-                    f.write(chunk)
-                    progress_bar.update(len(chunk))
-
-            progress_bar.close()
-
-        # we already have this model downloaded
-        else:
-            print(f"civitai model already downloaded to {destination_path}")
-
-        response.close()
-        return destination_path.as_posix()

apps/stable_diffusion/src/utils/resamplers.py

@@ -1,12 +0,0 @@
-import PIL.Image as Image
-
-resamplers = {
-    "Lanczos": Image.Resampling.LANCZOS,
-    "Nearest Neighbor": Image.Resampling.NEAREST,
-    "Bilinear": Image.Resampling.BILINEAR,
-    "Bicubic": Image.Resampling.BICUBIC,
-    "Hamming": Image.Resampling.HAMMING,
-    "Box": Image.Resampling.BOX,
-}
-
-resampler_list = resamplers.keys()

apps/stable_diffusion/src/utils/resources/opt_flags.json

@@ -11,12 +11,12 @@
     "untuned": {
       "fp16": {
         "default_compilation_flags": [
-          "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-global-opt-detach-elementwise-from-named-ops,iree-global-opt-convert-1x1-filter-conv2d-to-matmul,iree-preprocessing-convert-conv2d-to-img2col,iree-preprocessing-pad-linalg-ops{pad-size=32}))"
+          "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-flow-detach-elementwise-from-named-ops,iree-flow-convert-1x1-filter-conv2d-to-matmul,iree-preprocessing-convert-conv2d-to-img2col,iree-preprocessing-pad-linalg-ops{pad-size=32}))"
         ]
       },
       "fp32": {
         "default_compilation_flags": [
-          "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-global-opt-detach-elementwise-from-named-ops,iree-global-opt-convert-1x1-filter-conv2d-to-matmul,iree-preprocessing-convert-conv2d-to-img2col,iree-preprocessing-pad-linalg-ops{pad-size=16}))"
+          "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-flow-detach-elementwise-from-named-ops,iree-flow-convert-1x1-filter-conv2d-to-matmul,iree-preprocessing-convert-conv2d-to-img2col,iree-preprocessing-pad-linalg-ops{pad-size=16}))"
         ]
       }
     }
@@ -28,7 +28,7 @@
         "specified_compilation_flags": {
           "cuda": [],
           "default_device": [
-            "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-global-opt-detach-elementwise-from-named-ops,iree-global-opt-convert-1x1-filter-conv2d-to-matmul,iree-preprocessing-convert-conv2d-to-img2col,iree-preprocessing-pad-linalg-ops{pad-size=32},iree-linalg-ext-convert-conv2d-to-winograd))"
+            "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-flow-detach-elementwise-from-named-ops,iree-flow-convert-1x1-filter-conv2d-to-matmul,iree-preprocessing-convert-conv2d-to-img2col,iree-preprocessing-pad-linalg-ops{pad-size=32},iree-linalg-ext-convert-conv2d-to-winograd))"
           ]
         }
       },
@@ -37,7 +37,7 @@
         "specified_compilation_flags": {
           "cuda": [],
           "default_device": [
-            "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-global-opt-detach-elementwise-from-named-ops,iree-global-opt-convert-1x1-filter-conv2d-to-matmul,iree-preprocessing-convert-conv2d-to-img2col,iree-preprocessing-pad-linalg-ops{pad-size=16},iree-linalg-ext-convert-conv2d-to-winograd))"
+            "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-flow-detach-elementwise-from-named-ops,iree-flow-convert-1x1-filter-conv2d-to-matmul,iree-preprocessing-convert-conv2d-to-img2col,iree-preprocessing-pad-linalg-ops{pad-size=16},iree-linalg-ext-convert-conv2d-to-winograd))"
           ]
         }
       }
@@ -45,12 +45,12 @@
     "untuned": {
       "fp16": {
         "default_compilation_flags": [
-          "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-global-opt-detach-elementwise-from-named-ops,iree-global-opt-convert-1x1-filter-conv2d-to-matmul,iree-preprocessing-convert-conv2d-to-img2col,iree-preprocessing-pad-linalg-ops{pad-size=32},iree-linalg-ext-convert-conv2d-to-winograd))"
+          "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-flow-detach-elementwise-from-named-ops,iree-flow-convert-1x1-filter-conv2d-to-matmul,iree-preprocessing-convert-conv2d-to-img2col,iree-preprocessing-pad-linalg-ops{pad-size=32},iree-linalg-ext-convert-conv2d-to-winograd))"
         ]
       },
       "fp32": {
         "default_compilation_flags": [
-          "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-global-opt-detach-elementwise-from-named-ops,iree-global-opt-convert-1x1-filter-conv2d-to-matmul,iree-preprocessing-convert-conv2d-to-img2col,iree-preprocessing-pad-linalg-ops{pad-size=16},iree-linalg-ext-convert-conv2d-to-winograd))"
+          "--iree-preprocessing-pass-pipeline=builtin.module(func.func(iree-flow-detach-elementwise-from-named-ops,iree-flow-convert-1x1-filter-conv2d-to-matmul,iree-preprocessing-convert-conv2d-to-img2col,iree-preprocessing-pad-linalg-ops{pad-size=16},iree-linalg-ext-convert-conv2d-to-winograd))"
         ]
       }
     }

apps/stable_diffusion/src/utils/sd_annotation.py

@@ -109,7 +109,7 @@ def load_lower_configs(base_model_id=None):
             spec = spec.split("-")[0]
 
     if args.annotation_model == "vae":
-        if not spec or spec in ["sm_80"]:
+        if not spec or spec in ["rdna3", "sm_80"]:
             config_name = (
                 f"{args.annotation_model}_{args.precision}_{device}.json"
             )
@@ -158,9 +158,9 @@ def load_lower_configs(base_model_id=None):
                 f"{spec}.json"
             )
 
+    full_gs_url = config_bucket + config_name
     lowering_config_dir = os.path.join(WORKDIR, "configs", config_name)
     print("Loading lowering config file from ", lowering_config_dir)
-    full_gs_url = config_bucket + config_name
     download_public_file(full_gs_url, lowering_config_dir, True)
     return lowering_config_dir
 
@@ -203,8 +203,8 @@ def dump_after_mlir(input_mlir, use_winograd):
     if use_winograd:
         preprocess_flag = (
             "--iree-preprocessing-pass-pipeline=builtin.module"
-            "(func.func(iree-global-opt-detach-elementwise-from-named-ops,"
-            "iree-global-opt-convert-1x1-filter-conv2d-to-matmul,"
+            "(func.func(iree-flow-detach-elementwise-from-named-ops,"
+            "iree-flow-convert-1x1-filter-conv2d-to-matmul,"
             "iree-preprocessing-convert-conv2d-to-img2col,"
             "iree-preprocessing-pad-linalg-ops{pad-size=32},"
             "iree-linalg-ext-convert-conv2d-to-winograd))"
@@ -212,8 +212,8 @@ def dump_after_mlir(input_mlir, use_winograd):
     else:
         preprocess_flag = (
             "--iree-preprocessing-pass-pipeline=builtin.module"
-            "(func.func(iree-global-opt-detach-elementwise-from-named-ops,"
-            "iree-global-opt-convert-1x1-filter-conv2d-to-matmul,"
+            "(func.func(iree-flow-detach-elementwise-from-named-ops,"
+            "iree-flow-convert-1x1-filter-conv2d-to-matmul,"
             "iree-preprocessing-convert-conv2d-to-img2col,"
             "iree-preprocessing-pad-linalg-ops{pad-size=32}))"
         )

apps/stable_diffusion/src/utils/stable_args.py

@@ -2,8 +2,6 @@ import argparse
 import os
 from pathlib import Path
 
-from apps.stable_diffusion.src.utils.resamplers import resampler_list
-
 
 def path_expand(s):
     return Path(s).expanduser().resolve()
@@ -134,47 +132,6 @@ p.add_argument(
     "img2img.",
 )
 
-p.add_argument(
-    "--use_hiresfix",
-    type=bool,
-    default=False,
-    help="Use Hires Fix to do higher resolution images, while trying to "
-    "avoid the issues that come with it. This is accomplished by first "
-    "generating an image using txt2img, then running it through img2img.",
-)
-
-p.add_argument(
-    "--hiresfix_height",
-    type=int,
-    default=768,
-    choices=range(128, 769, 8),
-    help="The height of the Hires Fix image.",
-)
-
-p.add_argument(
-    "--hiresfix_width",
-    type=int,
-    default=768,
-    choices=range(128, 769, 8),
-    help="The width of the Hires Fix image.",
-)
-
-p.add_argument(
-    "--hiresfix_strength",
-    type=float,
-    default=0.6,
-    help="The denoising strength to apply for the Hires Fix.",
-)
-
-p.add_argument(
-    "--resample_type",
-    type=str,
-    default="Nearest Neighbor",
-    choices=resampler_list,
-    help="The resample type to use when resizing an image before being run "
-    "through stable diffusion.",
-)
-
 ##############################################################################
 # Stable Diffusion Training Params
 ##############################################################################
@@ -245,30 +202,28 @@ p.add_argument(
     "--left",
     default=False,
     action=argparse.BooleanOptionalAction,
-    help="If extend left for outpainting.",
+    help="If expend left for outpainting.",
 )
 
 p.add_argument(
     "--right",
     default=False,
     action=argparse.BooleanOptionalAction,
-    help="If extend right for outpainting.",
+    help="If expend right for outpainting.",
 )
 
 p.add_argument(
-    "--up",
     "--top",
     default=False,
     action=argparse.BooleanOptionalAction,
-    help="If extend top for outpainting.",
+    help="If expend top for outpainting.",
 )
 
 p.add_argument(
-    "--down",
     "--bottom",
     default=False,
     action=argparse.BooleanOptionalAction,
-    help="If extend bottom for outpainting.",
+    help="If expend bottom for outpainting.",
 )
 
 p.add_argument(
@@ -300,7 +255,7 @@ p.add_argument(
 
 p.add_argument(
     "--import_mlir",
-    default=True,
+    default=False,
     action=argparse.BooleanOptionalAction,
     help="Imports the model from torch module to shark_module otherwise "
     "downloads the model from shark_tank.",
@@ -323,7 +278,7 @@ p.add_argument(
 
 p.add_argument(
     "--use_tuned",
-    default=False,
+    default=True,
     action=argparse.BooleanOptionalAction,
     help="Download and use the tuned version of the model if available.",
 )
@@ -416,7 +371,7 @@ p.add_argument(
 
 p.add_argument(
     "--use_stencil",
-    choices=["canny", "openpose", "scribble", "zoedepth"],
+    choices=["canny", "openpose", "scribble"],
     help="Enable the stencil feature.",
 )
 
@@ -452,14 +407,6 @@ p.add_argument(
     help="Specify your own huggingface authentication tokens for models like Llama2.",
 )
 
-p.add_argument(
-    "--device_allocator_heap_key",
-    type=str,
-    default="",
-    help="Specify heap key for device caching allocator."
-    "Expected form: max_allocation_size;max_allocation_capacity;max_free_allocation_count"
-    "Example: --device_allocator_heap_key='*;1gib' (will limit caching on device to 1 gigabyte)",
-)
 ##############################################################################
 # IREE - Vulkan supported flags
 ##############################################################################
@@ -572,20 +519,6 @@ p.add_argument(
     "in shark importer. Does nothing if import_mlir is false (the default).",
 )
 
-p.add_argument(
-    "--compile_debug",
-    default=False,
-    action=argparse.BooleanOptionalAction,
-    help="Flag to toggle debug assert/verify flags for imported IR in the"
-    "iree-compiler. Default to false.",
-)
-
-p.add_argument(
-    "--iree_constant_folding",
-    default=True,
-    action=argparse.BooleanOptionalAction,
-    help="Controls constant folding in iree-compile for all SD models.",
-)
 
 ##############################################################################
 # Web UI flags
@@ -635,25 +568,6 @@ p.add_argument(
     help="Flag for enabling rest API.",
 )
 
-p.add_argument(
-    "--api_accept_origin",
-    action="append",
-    type=str,
-    help="An origin to be accepted by the REST api for Cross Origin"
-    "Resource Sharing (CORS). Use multiple times for multiple origins, "
-    'or use --api_accept_origin="*" to accept all origins. If no origins '
-    "are set no CORS headers will be returned by the api. Use, for "
-    "instance, if you need to access the REST api from Javascript running "
-    "in a web browser.",
-)
-
-p.add_argument(
-    "--debug",
-    default=False,
-    action=argparse.BooleanOptionalAction,
-    help="Flag for enabling debugging log in WebUI.",
-)
-
 p.add_argument(
     "--output_gallery",
     default=True,
@@ -731,18 +645,6 @@ p.add_argument(
     help="Specifies whether the docuchat's web version is running or not.",
 )
 
-##############################################################################
-# rocm Flags
-##############################################################################
-
-p.add_argument(
-    "--iree_rocm_target_chip",
-    type=str,
-    default="",
-    help="Add the rocm device architecture ex gfx1100, gfx90a, etc. Use `hipinfo` "
-    "or `iree-run-module --dump_devices=rocm` or `hipinfo` to get desired arch name",
-)
-
 args, unknown = p.parse_known_args()
 if args.import_debug:
     os.environ["IREE_SAVE_TEMPS"] = os.path.join(

apps/stable_diffusion/src/utils/stencils/__init__.py

@@ -1,3 +1,2 @@
 from apps.stable_diffusion.src.utils.stencils.canny import CannyDetector
 from apps.stable_diffusion.src.utils.stencils.openpose import OpenposeDetector
-from apps.stable_diffusion.src.utils.stencils.zoe import ZoeDetector

apps/stable_diffusion/src/utils/stencils/stencil_utils.py

@@ -4,7 +4,6 @@ import torch
 from apps.stable_diffusion.src.utils.stencils import (
     CannyDetector,
     OpenposeDetector,
-    ZoeDetector,
 )
 
 stencil = {}
@@ -118,9 +117,6 @@ def controlnet_hint_conversion(
         case "scribble":
             print("Working with scribble")
             controlnet_hint = hint_scribble(image)
-        case "zoedepth":
-            print("Working with ZoeDepth")
-            controlnet_hint = hint_zoedepth(image)
         case _:
             return None
     controlnet_hint = controlnet_hint_shaping(
@@ -131,7 +127,7 @@ def controlnet_hint_conversion(
 
 stencil_to_model_id_map = {
     "canny": "lllyasviel/control_v11p_sd15_canny",
-    "zoedepth": "lllyasviel/control_v11f1p_sd15_depth",
+    "depth": "lllyasviel/control_v11p_sd15_depth",
     "hed": "lllyasviel/sd-controlnet-hed",
     "mlsd": "lllyasviel/control_v11p_sd15_mlsd",
     "normal": "lllyasviel/control_v11p_sd15_normalbae",
@@ -188,16 +184,3 @@ def hint_scribble(image: Image.Image):
         detected_map = np.zeros_like(input_image, dtype=np.uint8)
         detected_map[np.min(input_image, axis=2) < 127] = 255
         return detected_map
-
-
-# Stencil 4. Depth (Only Zoe Preprocessing)
-def hint_zoedepth(image: Image.Image):
-    with torch.no_grad():
-        input_image = np.array(image)
-
-        if not "depth" in stencil:
-            stencil["depth"] = ZoeDetector()
-
-        detected_map = stencil["depth"](input_image)
-        detected_map = HWC3(detected_map)
-        return detected_map

apps/stable_diffusion/src/utils/stencils/zoe/__init__.py

@@ -1,58 +0,0 @@
-import numpy as np
-import torch
-from pathlib import Path
-import requests
-
-
-from einops import rearrange
-
-remote_model_path = (
-    "https://huggingface.co/lllyasviel/Annotators/resolve/main/ZoeD_M12_N.pt"
-)
-
-
-class ZoeDetector:
-    def __init__(self):
-        cwd = Path.cwd()
-        ckpt_path = Path(cwd, "stencil_annotator")
-        ckpt_path.mkdir(parents=True, exist_ok=True)
-        modelpath = ckpt_path / "ZoeD_M12_N.pt"
-
-        with requests.get(remote_model_path, stream=True) as r:
-            r.raise_for_status()
-            with open(modelpath, "wb") as f:
-                for chunk in r.iter_content(chunk_size=8192):
-                    f.write(chunk)
-
-        model = torch.hub.load(
-            "monorimet/ZoeDepth:torch_update",
-            "ZoeD_N",
-            pretrained=False,
-            force_reload=False,
-        )
-        model.load_state_dict(
-            torch.load(modelpath, map_location=model.device)["model"]
-        )
-        model.eval()
-        self.model = model
-
-    def __call__(self, input_image):
-        assert input_image.ndim == 3
-        image_depth = input_image
-        with torch.no_grad():
-            image_depth = torch.from_numpy(image_depth).float()
-            image_depth = image_depth / 255.0
-            image_depth = rearrange(image_depth, "h w c -> 1 c h w")
-            depth = self.model.infer(image_depth)
-
-            depth = depth[0, 0].cpu().numpy()
-
-            vmin = np.percentile(depth, 2)
-            vmax = np.percentile(depth, 85)
-
-            depth -= vmin
-            depth /= vmax - vmin
-            depth = 1.0 - depth
-            depth_image = (depth * 255.0).clip(0, 255).astype(np.uint8)
-
-            return depth_image

apps/stable_diffusion/src/utils/utils.py

@@ -18,14 +18,14 @@ import tempfile
 import torch
 from safetensors.torch import load_file
 from shark.shark_inference import SharkInference
-from shark.shark_importer import import_with_fx, save_mlir
+from shark.shark_importer import import_with_fx
 from shark.iree_utils.vulkan_utils import (
     set_iree_vulkan_runtime_flags,
     get_vulkan_target_triple,
     get_iree_vulkan_runtime_flags,
 )
 from shark.iree_utils.metal_utils import get_metal_target_triple
-from shark.iree_utils.gpu_utils import get_cuda_sm_cc, get_iree_rocm_args
+from shark.iree_utils.gpu_utils import get_cuda_sm_cc
 from apps.stable_diffusion.src.utils.stable_args import args
 from apps.stable_diffusion.src.utils.resources import opt_flags
 from apps.stable_diffusion.src.utils.sd_annotation import sd_model_annotation
@@ -78,7 +78,7 @@ def _compile_module(shark_module, model_name, extra_args=[]):
                     )
                 )
             path = shark_module.save_module(
-                os.getcwd(), model_name, extra_args, debug=args.compile_debug
+                os.getcwd(), model_name, extra_args
             )
             shark_module.load_module(path, extra_args=extra_args)
     else:
@@ -154,8 +154,8 @@ def compile_through_fx(
         f16_input_mask=f16_input_mask,
         debug=debug,
         model_name=extended_model_name,
+        save_dir=save_dir,
     )
-
     if use_tuned:
         if "vae" in extended_model_name.split("_")[0]:
             args.annotation_model = "vae"
@@ -168,14 +168,6 @@ def compile_through_fx(
             mlir_module, extended_model_name, base_model_id
         )
 
-    if not os.path.isdir(save_dir):
-        save_dir = ""
-
-    mlir_module = save_mlir(
-        mlir_module,
-        model_name=extended_model_name,
-        dir=save_dir,
-    )
     shark_module = SharkInference(
         mlir_module,
         device=args.device if device is None else device,
@@ -187,22 +179,17 @@ def compile_through_fx(
             mlir_module,
         )
 
+    del mlir_module
     gc.collect()
 
 
 def set_iree_runtime_flags():
-    # TODO: This function should be device-agnostic and piped properly
-    # to general runtime driver init.
     vulkan_runtime_flags = get_iree_vulkan_runtime_flags()
     if args.enable_rgp:
         vulkan_runtime_flags += [
             f"--enable_rgp=true",
             f"--vulkan_debug_utils=true",
         ]
-    if args.device_allocator_heap_key:
-        vulkan_runtime_flags += [
-            f"--device_allocator=caching:device_local={args.device_allocator_heap_key}",
-        ]
     set_iree_vulkan_runtime_flags(flags=vulkan_runtime_flags)
 
 
@@ -477,38 +464,18 @@ def get_available_devices():
                         f"{device_name} => {driver_name.replace('local', 'cpu')}"
                     )
                 else:
-                    # for drivers with single devices
-                    # let the default device be selected without any indexing
-                    if len(device_list_dict) == 1:
-                        device_list.append(f"{device_name} => {driver_name}")
-                    else:
-                        device_list.append(
-                            f"{device_name} => {driver_name}://{i}"
-                        )
+                    device_list.append(f"{device_name} => {driver_name}://{i}")
         return device_list
 
     set_iree_runtime_flags()
 
     available_devices = []
-    from shark.iree_utils.vulkan_utils import (
-        get_all_vulkan_devices,
-    )
-
-    vulkaninfo_list = get_all_vulkan_devices()
-    vulkan_devices = []
-    id = 0
-    for device in vulkaninfo_list:
-        vulkan_devices.append(f"{device.strip()} => vulkan://{id}")
-        id += 1
-    if id != 0:
-        print(f"vulkan devices are available.")
+    vulkan_devices = get_devices_by_name("vulkan")
     available_devices.extend(vulkan_devices)
     metal_devices = get_devices_by_name("metal")
     available_devices.extend(metal_devices)
     cuda_devices = get_devices_by_name("cuda")
     available_devices.extend(cuda_devices)
-    rocm_devices = get_devices_by_name("rocm")
-    available_devices.extend(rocm_devices)
     cpu_device = get_devices_by_name("cpu-sync")
     available_devices.extend(cpu_device)
     cpu_device = get_devices_by_name("cpu-task")
@@ -532,15 +499,10 @@ def get_opt_flags(model, precision="fp16"):
         iree_flags.append(
             f"-iree-vulkan-target-triple={args.iree_vulkan_target_triple}"
         )
-    if "rocm" in args.device:
-        rocm_args = get_iree_rocm_args()
-        iree_flags.extend(rocm_args)
-        print(iree_flags)
-    if args.iree_constant_folding == False:
-        iree_flags.append("--iree-opt-const-expr-hoisting=False")
-        iree_flags.append(
-            "--iree-codegen-linalg-max-constant-fold-elements=9223372036854775807"
-        )
+
+    # Disable bindings fusion to work with moltenVK.
+    if sys.platform == "darwin":
+        iree_flags.append("-iree-stream-fuse-binding=false")
 
     if "default_compilation_flags" in opt_flags[model][is_tuned][precision]:
         iree_flags += opt_flags[model][is_tuned][precision][
@@ -604,7 +566,7 @@ def preprocessCKPT(custom_weights, is_inpaint=False):
     )
     num_in_channels = 9 if is_inpaint else 4
     pipe = download_from_original_stable_diffusion_ckpt(
-        checkpoint_path_or_dict=custom_weights,
+        checkpoint_path=custom_weights,
         extract_ema=extract_ema,
         from_safetensors=from_safetensors,
         num_in_channels=num_in_channels,
@@ -811,12 +773,11 @@ def batch_seeds(
     seeds = seeds[:batch_count] + [-1] * (batch_count - len(seeds))
 
     if repeatable:
-        if all(seed < 0 for seed in seeds):
-            seeds[0] = sanitize_seed(seeds[0])
-
         # set seed for the rng based on what we have so far
         saved_random_state = random_getstate()
-        seed_random(str([n for n in seeds if n > -1]))
+        if all(seed < 0 for seed in seeds):
+            seeds[0] = sanitize_seed(seeds[0])
+        seed_random(str(seeds))
 
     # generate any seeds that are unspecified
     seeds = [sanitize_seed(seed) for seed in seeds]
@@ -855,8 +816,6 @@ def clear_all():
     elif os.name == "unix":
         shutil.rmtree(os.path.join(home, ".cache/AMD/VkCache"))
         shutil.rmtree(os.path.join(home, ".local/shark_tank"))
-    if args.local_tank_cache != "":
-        shutil.rmtree(args.local_tank_cache)
 
 
 def get_generated_imgs_path() -> Path:
@@ -902,13 +861,6 @@ def save_output_img(output_img, img_seed, extra_info=None):
         pngInfo = PngImagePlugin.PngInfo()
 
         if args.write_metadata_to_png:
-            # Using a conditional expression caused problems, so setting a new
-            # variable for now.
-            if args.use_hiresfix:
-                png_size_text = f"{args.hiresfix_width}x{args.hiresfix_height}"
-            else:
-                png_size_text = f"{args.width}x{args.height}"
-
             pngInfo.add_text(
                 "parameters",
                 f"{args.prompts[0]}"
@@ -917,7 +869,7 @@ def save_output_img(output_img, img_seed, extra_info=None):
                 f"Sampler: {args.scheduler}, "
                 f"CFG scale: {args.guidance_scale}, "
                 f"Seed: {img_seed},"
-                f"Size: {png_size_text}, "
+                f"Size: {args.width}x{args.height}, "
                 f"Model: {img_model}, "
                 f"VAE: {img_vae}, "
                 f"LoRA: {img_lora}",
@@ -944,10 +896,8 @@ def save_output_img(output_img, img_seed, extra_info=None):
         "CFG_SCALE": args.guidance_scale,
         "PRECISION": args.precision,
         "STEPS": args.steps,
-        "HEIGHT": args.height
-        if not args.use_hiresfix
-        else args.hiresfix_height,
-        "WIDTH": args.width if not args.use_hiresfix else args.hiresfix_width,
+        "HEIGHT": args.height,
+        "WIDTH": args.width,
         "MAX_LENGTH": args.max_length,
         "OUTPUT": out_img_path,
         "VAE": img_vae,
@@ -985,10 +935,6 @@ def get_generation_text_info(seeds, device):
     )
     text_output += (
         f"\nsize={args.height}x{args.width}, "
-        if not args.use_hiresfix
-        else f"\nsize={args.hiresfix_height}x{args.hiresfix_width}, "
-    )
-    text_output += (
         f"batch_count={args.batch_count}, "
         f"batch_size={args.batch_size}, "
         f"max_length={args.max_length}"

apps/stable_diffusion/web/api/__init__.py

@@ -1 +0,0 @@
-from apps.stable_diffusion.web.api.sdapi_v1 import sdapi

apps/stable_diffusion/web/api/sdapi_v1.py

@@ -1,579 +0,0 @@
-import os
-
-from collections import defaultdict
-from enum import Enum
-from fastapi import FastAPI
-from pydantic import BaseModel, Field, conlist, model_validator
-
-from apps.stable_diffusion.web.api.utils import (
-    frozen_args,
-    sampler_aliases,
-    encode_pil_to_base64,
-    decode_base64_to_image,
-    get_model_from_request,
-    get_scheduler_from_request,
-    get_lora_params,
-    get_device,
-    GenerationInputData,
-    GenerationResponseData,
-)
-
-from apps.stable_diffusion.web.ui.utils import (
-    get_custom_model_files,
-    get_custom_model_pathfile,
-    predefined_models,
-    predefined_paint_models,
-    predefined_upscaler_models,
-    scheduler_list,
-)
-from apps.stable_diffusion.web.ui.txt2img_ui import txt2img_inf
-from apps.stable_diffusion.web.ui.img2img_ui import img2img_inf
-from apps.stable_diffusion.web.ui.inpaint_ui import inpaint_inf
-from apps.stable_diffusion.web.ui.outpaint_ui import outpaint_inf
-from apps.stable_diffusion.web.ui.upscaler_ui import upscaler_inf
-
-sdapi = FastAPI()
-
-
-# Rest API: /sdapi/v1/sd-models (lists available models)
-class AppParam(str, Enum):
-    txt2img = "txt2img"
-    img2img = "img2img"
-    inpaint = "inpaint"
-    outpaint = "outpaint"
-    upscaler = "upscaler"
-
-
-@sdapi.get(
-    "/v1/sd-models",
-    summary="lists available models",
-    description=(
-        "This is all the models that this server currently knows about.\n "
-        "Models listed may still have a compilation and build pending that "
-        "will be triggered the first time they are used."
-    ),
-)
-def sd_models_api(app: AppParam = frozen_args.app):
-    match app:
-        case "inpaint" | "outpaint":
-            checkpoint_type = "inpainting"
-            predefined = predefined_paint_models
-        case "upscaler":
-            checkpoint_type = "upscaler"
-            predefined = predefined_upscaler_models
-        case _:
-            checkpoint_type = ""
-            predefined = predefined_models
-
-    return [
-        {
-            "title": model_file,
-            "model_name": model_file,
-            "hash": None,
-            "sha256": None,
-            "filename": get_custom_model_pathfile(model_file),
-            "config": None,
-        }
-        for model_file in get_custom_model_files(
-            custom_checkpoint_type=checkpoint_type
-        )
-    ] + [
-        {
-            "title": model,
-            "model_name": model,
-            "hash": None,
-            "sha256": None,
-            "filename": None,
-            "config": None,
-        }
-        for model in predefined
-    ]
-
-
-# Rest API: /sdapi/v1/samplers (lists schedulers)
-@sdapi.get(
-    "/v1/samplers",
-    summary="lists available schedulers/samplers",
-    description=(
-        "These are all the Schedulers defined and available. Not "
-        "every scheduler is compatible with all apis. Aliases are "
-        "equivalent samplers in A1111 if they are known."
-    ),
-)
-def sd_samplers_api():
-    reverse_sampler_aliases = defaultdict(list)
-    for key, value in sampler_aliases.items():
-        reverse_sampler_aliases[value].append(key)
-
-    return (
-        {
-            "name": scheduler,
-            "aliases": reverse_sampler_aliases.get(scheduler, []),
-            "options": {},
-        }
-        for scheduler in scheduler_list
-    )
-
-
-# Rest API: /sdapi/v1/options (lists application level options)
-@sdapi.get(
-    "/v1/options",
-    summary="lists current settings of application level options",
-    description=(
-        "A subset of the command line arguments set at startup renamed "
-        "to correspond to the A1111 naming. Only a small subset of A1111 "
-        "options are returned."
-    ),
-)
-def options_api():
-    # This is mostly just enough to support what Koboldcpp wants, with a
-    # few other things that seemed obvious
-    return {
-        "samples_save": True,
-        "samples_format": frozen_args.output_img_format,
-        "sd_model_checkpoint": os.path.basename(frozen_args.ckpt_loc)
-        if frozen_args.ckpt_loc
-        else frozen_args.hf_model_id,
-        "sd_lora": frozen_args.use_lora,
-        "sd_vae": frozen_args.custom_vae or "Automatic",
-        "enable_pnginfo": frozen_args.write_metadata_to_png,
-    }
-
-
-# Rest API: /sdapi/v1/cmd-flags (lists command line argument settings)
-@sdapi.get(
-    "/v1/cmd-flags",
-    summary="lists the command line arguments value that were set on startup.",
-)
-def cmd_flags_api():
-    return vars(frozen_args)
-
-
-# Rest API: /sdapi/v1/txt2img (Text to image)
-class ModelOverrideSettings(BaseModel):
-    sd_model_checkpoint: str = get_model_from_request(
-        fallback_model="stabilityai/stable-diffusion-2-1-base"
-    )
-
-
-class Txt2ImgInputData(GenerationInputData):
-    enable_hr: bool = frozen_args.use_hiresfix
-    hr_resize_y: int = Field(
-        default=frozen_args.hiresfix_height, ge=128, le=768, multiple_of=8
-    )
-    hr_resize_x: int = Field(
-        default=frozen_args.hiresfix_width, ge=128, le=768, multiple_of=8
-    )
-    override_settings: ModelOverrideSettings = None
-
-
-@sdapi.post(
-    "/v1/txt2img",
-    summary="Does text to image generation",
-    response_model=GenerationResponseData,
-)
-def txt2img_api(InputData: Txt2ImgInputData):
-    model_id = get_model_from_request(
-        InputData,
-        fallback_model="stabilityai/stable-diffusion-2-1-base",
-    )
-    scheduler = get_scheduler_from_request(
-        InputData, "txt2img_hires" if InputData.enable_hr else "txt2img"
-    )
-    (lora_weights, lora_hf_id) = get_lora_params(frozen_args.use_lora)
-
-    print(
-        f"Prompt: {InputData.prompt}, "
-        f"Negative Prompt: {InputData.negative_prompt}, "
-        f"Seed: {InputData.seed},"
-        f"Model: {model_id}, "
-        f"Scheduler: {scheduler}. "
-    )
-
-    res = txt2img_inf(
-        InputData.prompt,
-        InputData.negative_prompt,
-        InputData.height,
-        InputData.width,
-        InputData.steps,
-        InputData.cfg_scale,
-        InputData.seed,
-        batch_count=InputData.n_iter,
-        batch_size=1,
-        scheduler=scheduler,
-        model_id=model_id,
-        custom_vae=frozen_args.custom_vae or "None",
-        precision="fp16",
-        device=get_device(frozen_args.device),
-        max_length=frozen_args.max_length,
-        save_metadata_to_json=frozen_args.save_metadata_to_json,
-        save_metadata_to_png=frozen_args.write_metadata_to_png,
-        lora_weights=lora_weights,
-        lora_hf_id=lora_hf_id,
-        ondemand=frozen_args.ondemand,
-        repeatable_seeds=False,
-        use_hiresfix=InputData.enable_hr,
-        hiresfix_height=InputData.hr_resize_y,
-        hiresfix_width=InputData.hr_resize_x,
-        hiresfix_strength=frozen_args.hiresfix_strength,
-        resample_type=frozen_args.resample_type,
-    )
-
-    # Since we're not streaming we just want the last generator result
-    for items_so_far in res:
-        items = items_so_far
-
-    return {
-        "images": encode_pil_to_base64(items[0]),
-        "parameters": {},
-        "info": items[1],
-    }
-
-
-# Rest API: /sdapi/v1/img2img (Image to image)
-class StencilParam(str, Enum):
-    canny = "canny"
-    openpose = "openpose"
-    scribble = "scribble"
-    zoedepth = "zoedepth"
-
-
-class Img2ImgInputData(GenerationInputData):
-    init_images: conlist(str, min_length=1, max_length=2)
-    denoising_strength: float = frozen_args.strength
-    use_stencil: StencilParam = frozen_args.use_stencil
-    override_settings: ModelOverrideSettings = None
-
-    @model_validator(mode="after")
-    def check_image_supplied_for_scribble_stencil(self) -> "Img2ImgInputData":
-        if (
-            self.use_stencil == StencilParam.scribble
-            and len(self.init_images) < 2
-        ):
-            raise ValueError(
-                "a second image must be supplied for the controlnet:scribble stencil"
-            )
-
-        return self
-
-
-@sdapi.post(
-    "/v1/img2img",
-    summary="Does image to image generation",
-    response_model=GenerationResponseData,
-)
-def img2img_api(
-    InputData: Img2ImgInputData,
-):
-    model_id = get_model_from_request(
-        InputData,
-        fallback_model="stabilityai/stable-diffusion-2-1-base",
-    )
-    scheduler = get_scheduler_from_request(InputData, "img2img")
-    (lora_weights, lora_hf_id) = get_lora_params(frozen_args.use_lora)
-
-    init_image = decode_base64_to_image(InputData.init_images[0])
-    mask_image = (
-        decode_base64_to_image(InputData.init_images[1])
-        if len(InputData.init_images) > 1
-        else None
-    )
-
-    print(
-        f"Prompt: {InputData.prompt}, "
-        f"Negative Prompt: {InputData.negative_prompt}, "
-        f"Seed: {InputData.seed}, "
-        f"Model: {model_id}, "
-        f"Scheduler: {scheduler}."
-    )
-
-    res = img2img_inf(
-        InputData.prompt,
-        InputData.negative_prompt,
-        {"image": init_image, "mask": mask_image},
-        InputData.height,
-        InputData.width,
-        InputData.steps,
-        InputData.denoising_strength,
-        InputData.cfg_scale,
-        InputData.seed,
-        batch_count=InputData.n_iter,
-        batch_size=1,
-        scheduler=scheduler,
-        model_id=model_id,
-        custom_vae=frozen_args.custom_vae or "None",
-        precision="fp16",
-        device=get_device(frozen_args.device),
-        max_length=frozen_args.max_length,
-        use_stencil=InputData.use_stencil,
-        save_metadata_to_json=frozen_args.save_metadata_to_json,
-        save_metadata_to_png=frozen_args.write_metadata_to_png,
-        lora_weights=lora_weights,
-        lora_hf_id=lora_hf_id,
-        ondemand=frozen_args.ondemand,
-        repeatable_seeds=False,
-        resample_type=frozen_args.resample_type,
-    )
-
-    # Since we're not streaming we just want the last generator result
-    for items_so_far in res:
-        items = items_so_far
-
-    return {
-        "images": encode_pil_to_base64(items[0]),
-        "parameters": {},
-        "info": items[1],
-    }
-
-
-# Rest API: /sdapi/v1/inpaint (Inpainting)
-class PaintModelOverideSettings(BaseModel):
-    sd_model_checkpoint: str = get_model_from_request(
-        checkpoint_type="inpainting",
-        fallback_model="stabilityai/stable-diffusion-2-inpainting",
-    )
-
-
-class InpaintInputData(GenerationInputData):
-    image: str = Field(description="Base64 encoded input image")
-    mask: str = Field(description="Base64 encoded mask image")
-    is_full_res: bool = False  # Is this setting backwards in the UI?
-    full_res_padding: int = Field(default=32, ge=0, le=256, multiple_of=4)
-    denoising_strength: float = frozen_args.strength
-    use_stencil: StencilParam = frozen_args.use_stencil
-    override_settings: PaintModelOverideSettings = None
-
-
-@sdapi.post(
-    "/v1/inpaint",
-    summary="Does inpainting generation on an image",
-    response_model=GenerationResponseData,
-)
-def inpaint_api(
-    InputData: InpaintInputData,
-):
-    model_id = get_model_from_request(
-        InputData,
-        checkpoint_type="inpainting",
-        fallback_model="stabilityai/stable-diffusion-2-inpainting",
-    )
-    scheduler = get_scheduler_from_request(InputData, "inpaint")
-    (lora_weights, lora_hf_id) = get_lora_params(frozen_args.use_lora)
-
-    init_image = decode_base64_to_image(InputData.image)
-    mask = decode_base64_to_image(InputData.mask)
-
-    print(
-        f"Prompt: {InputData.prompt}, "
-        f'Negative Prompt: {InputData.negative_prompt}", '
-        f'Seed: {InputData.seed}", '
-        f"Model: {model_id}, "
-        f"Scheduler: {scheduler}."
-    )
-
-    res = inpaint_inf(
-        InputData.prompt,
-        InputData.negative_prompt,
-        {"image": init_image, "mask": mask},
-        InputData.height,
-        InputData.width,
-        InputData.is_full_res,
-        InputData.full_res_padding,
-        InputData.steps,
-        InputData.cfg_scale,
-        InputData.seed,
-        batch_count=InputData.n_iter,
-        batch_size=1,
-        scheduler=scheduler,
-        model_id=model_id,
-        custom_vae=frozen_args.custom_vae or "None",
-        precision="fp16",
-        device=get_device(frozen_args.device),
-        max_length=frozen_args.max_length,
-        save_metadata_to_json=frozen_args.save_metadata_to_json,
-        save_metadata_to_png=frozen_args.write_metadata_to_png,
-        lora_weights=lora_weights,
-        lora_hf_id=lora_hf_id,
-        ondemand=frozen_args.ondemand,
-        repeatable_seeds=False,
-    )
-
-    # Since we're not streaming we just want the last generator result
-    for items_so_far in res:
-        items = items_so_far
-
-    return {
-        "images": encode_pil_to_base64(items[0]),
-        "parameters": {},
-        "info": items[1],
-    }
-
-
-# Rest API: /sdapi/v1/outpaint (Outpainting)
-class DirectionParam(str, Enum):
-    left = "left"
-    right = "right"
-    up = "up"
-    down = "down"
-
-
-class OutpaintInputData(GenerationInputData):
-    init_images: list[str]
-    pixels: int = Field(
-        default=frozen_args.pixels, ge=8, le=256, multiple_of=8
-    )
-    mask_blur: int = Field(default=frozen_args.mask_blur, ge=0, le=64)
-    directions: set[DirectionParam] = [
-        direction
-        for direction in ["left", "right", "up", "down"]
-        if vars(frozen_args)[direction]
-    ]
-    noise_q: float = frozen_args.noise_q
-    color_variation: float = frozen_args.color_variation
-    override_settings: PaintModelOverideSettings = None
-
-
-@sdapi.post(
-    "/v1/outpaint",
-    summary="Does outpainting generation on an image",
-    response_model=GenerationResponseData,
-)
-def outpaint_api(
-    InputData: OutpaintInputData,
-):
-    model_id = get_model_from_request(
-        InputData,
-        checkpoint_type="inpainting",
-        fallback_model="stabilityai/stable-diffusion-2-inpainting",
-    )
-    scheduler = get_scheduler_from_request(InputData, "outpaint")
-    (lora_weights, lora_hf_id) = get_lora_params(frozen_args.use_lora)
-
-    init_image = decode_base64_to_image(InputData.init_images[0])
-
-    print(
-        f"Prompt: {InputData.prompt}, "
-        f"Negative Prompt: {InputData.negative_prompt}, "
-        f"Seed: {InputData.seed}, "
-        f"Model: {model_id}, "
-        f"Scheduler: {scheduler}."
-    )
-
-    res = outpaint_inf(
-        InputData.prompt,
-        InputData.negative_prompt,
-        init_image,
-        InputData.pixels,
-        InputData.mask_blur,
-        InputData.directions,
-        InputData.noise_q,
-        InputData.color_variation,
-        InputData.height,
-        InputData.width,
-        InputData.steps,
-        InputData.cfg_scale,
-        InputData.seed,
-        batch_count=InputData.n_iter,
-        batch_size=1,
-        scheduler=scheduler,
-        model_id=model_id,
-        custom_vae=frozen_args.custom_vae or "None",
-        precision="fp16",
-        device=get_device(frozen_args.device),
-        max_length=frozen_args.max_length,
-        save_metadata_to_json=frozen_args.save_metadata_to_json,
-        save_metadata_to_png=frozen_args.write_metadata_to_png,
-        lora_weights=lora_weights,
-        lora_hf_id=lora_hf_id,
-        ondemand=frozen_args.ondemand,
-        repeatable_seeds=False,
-    )
-
-    # Since we're not streaming we just want the last generator result
-    for items_so_far in res:
-        items = items_so_far
-
-    return {
-        "images": encode_pil_to_base64(items[0]),
-        "parameters": {},
-        "info": items[1],
-    }
-
-
-# Rest API: /sdapi/v1/upscaler (Upscaling)
-class UpscalerModelOverideSettings(BaseModel):
-    sd_model_checkpoint: str = get_model_from_request(
-        checkpoint_type="upscaler",
-        fallback_model="stabilityai/stable-diffusion-x4-upscaler",
-    )
-
-
-class UpscalerInputData(GenerationInputData):
-    init_images: list[str] = Field(
-        description="Base64 encoded image to upscale"
-    )
-    noise_level: int = frozen_args.noise_level
-    override_settings: UpscalerModelOverideSettings = None
-
-
-@sdapi.post(
-    "/v1/upscaler",
-    summary="Does image upscaling",
-    response_model=GenerationResponseData,
-)
-def upscaler_api(
-    InputData: UpscalerInputData,
-):
-    model_id = get_model_from_request(
-        InputData,
-        checkpoint_type="upscaler",
-        fallback_model="stabilityai/stable-diffusion-x4-upscaler",
-    )
-    scheduler = get_scheduler_from_request(InputData, "upscaler")
-    (lora_weights, lora_hf_id) = get_lora_params(frozen_args.use_lora)
-
-    init_image = decode_base64_to_image(InputData.init_images[0])
-
-    print(
-        f"Prompt: {InputData.prompt}, "
-        f"Negative Prompt: {InputData.negative_prompt}, "
-        f"Seed: {InputData.seed}, "
-        f"Model: {model_id}, "
-        f"Scheduler: {scheduler}."
-    )
-
-    res = upscaler_inf(
-        InputData.prompt,
-        InputData.negative_prompt,
-        init_image,
-        InputData.height,
-        InputData.width,
-        InputData.steps,
-        InputData.noise_level,
-        InputData.cfg_scale,
-        InputData.seed,
-        batch_count=InputData.n_iter,
-        batch_size=1,
-        scheduler=scheduler,
-        model_id=model_id,
-        custom_vae=frozen_args.custom_vae or "None",
-        precision="fp16",
-        device=get_device(frozen_args.device),
-        max_length=frozen_args.max_length,
-        save_metadata_to_json=frozen_args.save_metadata_to_json,
-        save_metadata_to_png=frozen_args.write_metadata_to_png,
-        lora_weights=lora_weights,
-        lora_hf_id=lora_hf_id,
-        ondemand=frozen_args.ondemand,
-        repeatable_seeds=False,
-    )
-
-    # Since we're not streaming we just want the last generator result
-    for items_so_far in res:
-        items = items_so_far
-
-    return {
-        "images": encode_pil_to_base64(items[0]),
-        "parameters": {},
-        "info": items[1],
-    }

apps/stable_diffusion/web/api/utils.py

@@ -1,211 +0,0 @@
-import base64
-import pickle
-
-from argparse import Namespace
-from fastapi.exceptions import HTTPException
-from io import BytesIO
-from PIL import Image
-from pydantic import BaseModel, Field
-
-from apps.stable_diffusion.src import args
-from apps.stable_diffusion.web.ui.utils import (
-    available_devices,
-    get_custom_model_files,
-    predefined_models,
-    predefined_paint_models,
-    predefined_upscaler_models,
-    scheduler_list,
-    scheduler_list_cpu_only,
-)
-
-
-# Probably overly cautious, but try to ensure we only use the starting
-# args in each api call, as the code does `args.<whatever> = <changed_value>`
-# in lots of places and in testing, it seemed to me, these changes leaked
-# into subsequent api calls.
-
-# Roundtripping through pickle for deepcopy, there is probably a better way
-frozen_args = Namespace(**(pickle.loads(pickle.dumps(vars(args)))))
-
-# an attempt to map some of the A1111 sampler names to scheduler names
-# https://github.com/huggingface/diffusers/issues/4167 is where the
-# (not so obvious) ones come from
-sampler_aliases = {
-    # a1111/onnx (these point to diffusers classes in A1111)
-    "pndm": "PNDM",
-    "heun": "HeunDiscrete",
-    "ddim": "DDIM",
-    "ddpm": "DDPM",
-    "euler": "EulerDiscrete",
-    "euler-ancestral": "EulerAncestralDiscrete",
-    "dpm": "DPMSolverMultistep",
-    # a1111/k_diffusion (the obvious ones)
-    "Euler a": "EulerAncestralDiscrete",
-    "Euler": "EulerDiscrete",
-    "LMS": "LMSDiscrete",
-    "Heun": "HeunDiscrete",
-    # a1111/k_diffusion (not so obvious)
-    "DPM++ 2M": "DPMSolverMultistep",
-    "DPM++ 2M Karras": "DPMSolverMultistepKarras",
-    "DPM++ 2M SDE": "DPMSolverMultistep++",
-    "DPM++ 2M SDE Karras": "DPMSolverMultistepKarras++",
-    "DPM2": "KDPM2Discrete",
-    "DPM2 a": "KDPM2AncestralDiscrete",
-}
-
-allowed_schedulers = {
-    "txt2img": {
-        "schedulers": scheduler_list,
-        "fallback": "SharkEulerDiscrete",
-    },
-    "txt2img_hires": {
-        "schedulers": scheduler_list_cpu_only,
-        "fallback": "DEISMultistep",
-    },
-    "img2img": {
-        "schedulers": scheduler_list_cpu_only,
-        "fallback": "EulerDiscrete",
-    },
-    "inpaint": {
-        "schedulers": scheduler_list_cpu_only,
-        "fallback": "DDIM",
-    },
-    "outpaint": {
-        "schedulers": scheduler_list_cpu_only,
-        "fallback": "DDIM",
-    },
-    "upscaler": {
-        "schedulers": scheduler_list_cpu_only,
-        "fallback": "DDIM",
-    },
-}
-
-# base pydantic model for sd generation apis
-
-
-class GenerationInputData(BaseModel):
-    prompt: str = ""
-    negative_prompt: str = ""
-    hf_model_id: str | None = None
-    height: int = Field(
-        default=frozen_args.height, ge=128, le=768, multiple_of=8
-    )
-    width: int = Field(
-        default=frozen_args.width, ge=128, le=768, multiple_of=8
-    )
-    sampler_name: str = frozen_args.scheduler
-    cfg_scale: float = Field(default=frozen_args.guidance_scale, ge=1)
-    steps: int = Field(default=frozen_args.steps, ge=1, le=100)
-    seed: int = frozen_args.seed
-    n_iter: int = Field(default=frozen_args.batch_count)
-
-
-class GenerationResponseData(BaseModel):
-    images: list[str] = Field(description="Generated images, Base64 encoded")
-    properties: dict = {}
-    info: str
-
-
-# image encoding/decoding
-
-
-def encode_pil_to_base64(images: list[Image.Image]):
-    encoded_imgs = []
-    for image in images:
-        with BytesIO() as output_bytes:
-            if frozen_args.output_img_format.lower() == "png":
-                image.save(output_bytes, format="PNG")
-
-            elif frozen_args.output_img_format.lower() in ("jpg", "jpeg"):
-                image.save(output_bytes, format="JPEG")
-            else:
-                raise HTTPException(
-                    status_code=500, detail="Invalid image format"
-                )
-            bytes_data = output_bytes.getvalue()
-            encoded_imgs.append(base64.b64encode(bytes_data))
-    return encoded_imgs
-
-
-def decode_base64_to_image(encoding: str):
-    if encoding.startswith("data:image/"):
-        encoding = encoding.split(";", 1)[1].split(",", 1)[1]
-    try:
-        image = Image.open(BytesIO(base64.b64decode(encoding)))
-        return image
-    except Exception as err:
-        print(err)
-        raise HTTPException(status_code=400, detail="Invalid encoded image")
-
-
-# get valid sd models/vaes/schedulers etc.
-
-
-def get_predefined_models(custom_checkpoint_type: str):
-    match custom_checkpoint_type:
-        case "inpainting":
-            return predefined_paint_models
-        case "upscaler":
-            return predefined_upscaler_models
-        case _:
-            return predefined_models
-
-
-def get_model_from_request(
-    request_data=None,
-    checkpoint_type: str = "",
-    fallback_model: str = "",
-):
-    model = None
-    if request_data:
-        if request_data.hf_model_id:
-            model = request_data.hf_model_id
-        elif request_data.override_settings:
-            model = request_data.override_settings.sd_model_checkpoint
-
-    # if the request didn't specify a model try the command line args
-    result = model or frozen_args.ckpt_loc or frozen_args.hf_model_id
-
-    # make sure whatever we have is a valid model for the checkpoint type
-    if result in get_custom_model_files(
-        custom_checkpoint_type=checkpoint_type
-    ) + get_predefined_models(checkpoint_type):
-        return result
-    # if not return what was specified as the fallback
-    else:
-        return fallback_model
-
-
-def get_scheduler_from_request(
-    request_data: GenerationInputData, operation: str
-):
-    allowed = allowed_schedulers[operation]
-
-    requested = request_data.sampler_name
-    requested = sampler_aliases.get(requested, requested)
-
-    return (
-        requested
-        if requested in allowed["schedulers"]
-        else allowed["fallback"]
-    )
-
-
-def get_lora_params(use_lora: str):
-    # TODO: since the inference functions in the webui, which we are
-    # still calling into for the api, jam these back together again before
-    # handing them off to the pipeline, we should remove this nonsense
-    # and unify their selection in the UI and command line args proper
-    if use_lora in get_custom_model_files("lora"):
-        return (use_lora, "")
-
-    return ("None", use_lora)
-
-
-def get_device(device_str: str):
-    # first substring match in the list available devices, with first
-    # device when none are matched
-    return next(
-        (device for device in available_devices if device_str in device),
-        available_devices[0],
-    )

apps/stable_diffusion/web/index.py

@@ -1,8 +1,6 @@
-from multiprocessing import freeze_support
+from multiprocessing import Process, freeze_support
 import os
 import sys
-import logging
-import apps.stable_diffusion.web.utils.app as app
 
 if sys.platform == "darwin":
     # import before IREE to avoid torch-MLIR library issues
@@ -22,54 +20,64 @@ if args.clear_all:
     clear_all()
 
 
+def launch_app(address):
+    from tkinter import Tk
+    import webview
+
+    window = Tk()
+
+    # get screen width and height of display and make it more reasonably
+    # sized as we aren't making it full-screen or maximized
+    width = int(window.winfo_screenwidth() * 0.81)
+    height = int(window.winfo_screenheight() * 0.91)
+    webview.create_window(
+        "SHARK AI Studio",
+        url=address,
+        width=width,
+        height=height,
+        text_select=True,
+    )
+    webview.start(private_mode=False)
+
+
 if __name__ == "__main__":
-    if args.debug:
-        logging.basicConfig(level=logging.DEBUG)
     # required to do multiprocessing in a pyinstaller freeze
     freeze_support()
     if args.api or "api" in args.ui.split(","):
         from apps.stable_diffusion.web.ui import (
+            txt2img_api,
+            img2img_api,
+            upscaler_api,
+            inpaint_api,
+            outpaint_api,
             llm_chat_api,
         )
-        from apps.stable_diffusion.web.api import sdapi
 
         from fastapi import FastAPI, APIRouter
-        from fastapi.middleware.cors import CORSMiddleware
         import uvicorn
 
         # init global sd pipeline and config
         global_obj._init()
 
-        api = FastAPI()
-        api.mount("/sdapi/", sdapi)
+        app = FastAPI()
+        app.add_api_route("/sdapi/v1/txt2img", txt2img_api, methods=["post"])
+        app.add_api_route("/sdapi/v1/img2img", img2img_api, methods=["post"])
+        app.add_api_route("/sdapi/v1/inpaint", inpaint_api, methods=["post"])
+        app.add_api_route("/sdapi/v1/outpaint", outpaint_api, methods=["post"])
+        app.add_api_route("/sdapi/v1/upscaler", upscaler_api, methods=["post"])
 
         # chat APIs needed for compatibility with multiple extensions using OpenAI API
-        api.add_api_route(
+        app.add_api_route(
             "/v1/chat/completions", llm_chat_api, methods=["post"]
         )
-        api.add_api_route("/v1/completions", llm_chat_api, methods=["post"])
-        api.add_api_route("/chat/completions", llm_chat_api, methods=["post"])
-        api.add_api_route("/completions", llm_chat_api, methods=["post"])
-        api.add_api_route(
+        app.add_api_route("/v1/completions", llm_chat_api, methods=["post"])
+        app.add_api_route("/chat/completions", llm_chat_api, methods=["post"])
+        app.add_api_route("/completions", llm_chat_api, methods=["post"])
+        app.add_api_route(
             "/v1/engines/codegen/completions", llm_chat_api, methods=["post"]
         )
-        api.include_router(APIRouter())
-
-        # deal with CORS requests if CORS accept origins are set
-        if args.api_accept_origin:
-            print(
-                f"API Configured for CORS. Accepting origins: { args.api_accept_origin }"
-            )
-            api.add_middleware(
-                CORSMiddleware,
-                allow_origins=args.api_accept_origin,
-                allow_methods=["GET", "POST"],
-                allow_headers=["*"],
-            )
-        else:
-            print("API not configured for CORS")
-
-        uvicorn.run(api, host="0.0.0.0", port=args.server_port)
+        app.include_router(APIRouter())
+        uvicorn.run(app, host="0.0.0.0", port=args.server_port)
         sys.exit(0)
 
     # Setup to use shark_tmp for gradio's temporary image files and clear any
@@ -83,10 +91,7 @@ if __name__ == "__main__":
     import gradio as gr
 
     # Create custom models folders if they don't exist
-    from apps.stable_diffusion.web.ui.utils import (
-        create_custom_models_folders,
-        nodicon_loc,
-    )
+    from apps.stable_diffusion.web.ui.utils import create_custom_models_folders
 
     create_custom_models_folders()
 
@@ -102,6 +107,7 @@ if __name__ == "__main__":
     from apps.stable_diffusion.web.ui import (
         txt2img_web,
         txt2img_custom_model,
+        txt2img_hf_model_id,
         txt2img_gallery,
         txt2img_png_info_img,
         txt2img_status,
@@ -109,10 +115,10 @@ if __name__ == "__main__":
         txt2img_sendto_inpaint,
         txt2img_sendto_outpaint,
         txt2img_sendto_upscaler,
-        # h2ogpt_upload,
-        # h2ogpt_web,
+        h2ogpt_web,
         img2img_web,
         img2img_custom_model,
+        img2img_hf_model_id,
         img2img_gallery,
         img2img_init_image,
         img2img_status,
@@ -121,6 +127,7 @@ if __name__ == "__main__":
         img2img_sendto_upscaler,
         inpaint_web,
         inpaint_custom_model,
+        inpaint_hf_model_id,
         inpaint_gallery,
         inpaint_init_image,
         inpaint_status,
@@ -129,6 +136,7 @@ if __name__ == "__main__":
         inpaint_sendto_upscaler,
         outpaint_web,
         outpaint_custom_model,
+        outpaint_hf_model_id,
         outpaint_gallery,
         outpaint_init_image,
         outpaint_status,
@@ -137,15 +145,15 @@ if __name__ == "__main__":
         outpaint_sendto_upscaler,
         upscaler_web,
         upscaler_custom_model,
+        upscaler_hf_model_id,
         upscaler_gallery,
         upscaler_init_image,
         upscaler_status,
         upscaler_sendto_img2img,
         upscaler_sendto_inpaint,
         upscaler_sendto_outpaint,
-        #  lora_train_web,
-        #  model_web,
-        #  model_config_web,
+        lora_train_web,
+        model_web,
         hf_models,
         modelmanager_sendto_txt2img,
         modelmanager_sendto_img2img,
@@ -200,18 +208,9 @@ if __name__ == "__main__":
         )
 
     with gr.Blocks(
-        css=dark_theme, analytics_enabled=False, title="SHARK AI Studio"
+        css=dark_theme, analytics_enabled=False, title="Stable Diffusion"
     ) as sd_web:
         with gr.Tabs() as tabs:
-            # NOTE: If adding, removing, or re-ordering tabs, make sure that they
-            # have a unique id that doesn't clash with any of the other tabs,
-            # and that the order in the code here is the order they should
-            # appear in the ui, as the id value doesn't determine the order.
-
-            # Where possible, avoid changing the id of any tab that is the
-            # destination of one of the 'send to' buttons. If you do have to change
-            # that id, make sure you update the relevant register_button_click calls
-            # further down with the new id.
             with gr.TabItem(label="Text-to-Image", id=0):
                 txt2img_web.render()
             with gr.TabItem(label="Image-to-Image", id=1):
@@ -237,31 +236,16 @@ if __name__ == "__main__":
                         upscaler_status,
                     ]
                 )
-            #  with gr.TabItem(label="Model Manager", id=6):
-            #      model_web.render()
-            #  with gr.TabItem(label="LoRA Training (Experimental)", id=7):
-            #      lora_train_web.render()
-            with gr.TabItem(label="Chat Bot", id=8):
+            with gr.TabItem(label="Model Manager", id=6):
+                model_web.render()
+            with gr.TabItem(label="LoRA Training (Experimental)", id=8):
+                lora_train_web.render()
+            with gr.TabItem(label="Chat Bot (Experimental)", id=7):
                 stablelm_chat.render()
-            #  with gr.TabItem(
-            #      label="Generate Sharding Config (Experimental)", id=9
-            #  ):
-            #      model_config_web.render()
-            with gr.TabItem(label="MultiModal (Experimental)", id=10):
+            with gr.TabItem(label="MultiModal (Experimental)", id=9):
                 minigpt4_web.render()
-            # with gr.TabItem(label="DocuChat Upload", id=11):
-            #     h2ogpt_upload.render()
-            # with gr.TabItem(label="DocuChat(Experimental)", id=12):
-            #     h2ogpt_web.render()
-
-            actual_port = app.usable_port()
-            if actual_port != args.server_port:
-                sd_web.load(
-                    fn=lambda: gr.Info(
-                        f"Port {args.server_port} is in use by another application. "
-                        f"Shark is running on port {actual_port} instead."
-                    )
-                )
+            with gr.TabItem(label="DocuChat(Experimental)", id=10):
+                h2ogpt_web.render()
 
         # send to buttons
         register_button_click(
@@ -395,38 +379,42 @@ if __name__ == "__main__":
             modelmanager_sendto_txt2img,
             0,
             [hf_models],
-            [txt2img_custom_model, tabs],
+            [txt2img_custom_model, txt2img_hf_model_id, tabs],
         )
         register_modelmanager_button(
             modelmanager_sendto_img2img,
             1,
             [hf_models],
-            [img2img_custom_model, tabs],
+            [img2img_custom_model, img2img_hf_model_id, tabs],
         )
         register_modelmanager_button(
             modelmanager_sendto_inpaint,
             2,
             [hf_models],
-            [inpaint_custom_model, tabs],
+            [inpaint_custom_model, inpaint_hf_model_id, tabs],
         )
         register_modelmanager_button(
             modelmanager_sendto_outpaint,
             3,
             [hf_models],
-            [outpaint_custom_model, tabs],
+            [outpaint_custom_model, outpaint_hf_model_id, tabs],
         )
         register_modelmanager_button(
             modelmanager_sendto_upscaler,
             4,
             [hf_models],
-            [upscaler_custom_model, tabs],
+            [upscaler_custom_model, upscaler_hf_model_id, tabs],
         )
 
     sd_web.queue()
+    if args.ui == "app":
+        t = Process(
+            target=launch_app, args=[f"http://localhost:{args.server_port}"]
+        )
+        t.start()
     sd_web.launch(
         share=args.share,
-        inbrowser=not app.launch(actual_port),
+        inbrowser=args.ui == "web",
         server_name="0.0.0.0",
-        server_port=actual_port,
-        favicon_path=nodicon_loc,
+        server_port=args.server_port,
     )

apps/stable_diffusion/web/ui/__init__.py

@@ -1,7 +1,9 @@
 from apps.stable_diffusion.web.ui.txt2img_ui import (
     txt2img_inf,
+    txt2img_api,
     txt2img_web,
     txt2img_custom_model,
+    txt2img_hf_model_id,
     txt2img_gallery,
     txt2img_png_info_img,
     txt2img_status,
@@ -12,8 +14,10 @@ from apps.stable_diffusion.web.ui.txt2img_ui import (
 )
 from apps.stable_diffusion.web.ui.img2img_ui import (
     img2img_inf,
+    img2img_api,
     img2img_web,
     img2img_custom_model,
+    img2img_hf_model_id,
     img2img_gallery,
     img2img_init_image,
     img2img_status,
@@ -23,8 +27,10 @@ from apps.stable_diffusion.web.ui.img2img_ui import (
 )
 from apps.stable_diffusion.web.ui.inpaint_ui import (
     inpaint_inf,
+    inpaint_api,
     inpaint_web,
     inpaint_custom_model,
+    inpaint_hf_model_id,
     inpaint_gallery,
     inpaint_init_image,
     inpaint_status,
@@ -34,8 +40,10 @@ from apps.stable_diffusion.web.ui.inpaint_ui import (
 )
 from apps.stable_diffusion.web.ui.outpaint_ui import (
     outpaint_inf,
+    outpaint_api,
     outpaint_web,
     outpaint_custom_model,
+    outpaint_hf_model_id,
     outpaint_gallery,
     outpaint_init_image,
     outpaint_status,
@@ -45,8 +53,10 @@ from apps.stable_diffusion.web.ui.outpaint_ui import (
 )
 from apps.stable_diffusion.web.ui.upscaler_ui import (
     upscaler_inf,
+    upscaler_api,
     upscaler_web,
     upscaler_custom_model,
+    upscaler_hf_model_id,
     upscaler_gallery,
     upscaler_init_image,
     upscaler_status,
@@ -68,7 +78,7 @@ from apps.stable_diffusion.web.ui.stablelm_ui import (
     stablelm_chat,
     llm_chat_api,
 )
-from apps.stable_diffusion.web.ui.generate_config import model_config_web
+from apps.stable_diffusion.web.ui.h2ogpt import h2ogpt_web
 from apps.stable_diffusion.web.ui.minigpt4_ui import minigpt4_web
 from apps.stable_diffusion.web.ui.outputgallery_ui import (
     outputgallery_web,

apps/stable_diffusion/web/ui/generate_config.py

@@ -1,41 +0,0 @@
-import gradio as gr
-import torch
-from transformers import AutoTokenizer
-from apps.language_models.src.model_wrappers.vicuna_model import CombinedModel
-from shark.shark_generate_model_config import GenerateConfigFile
-
-
-def get_model_config():
-    hf_model_path = "TheBloke/vicuna-7B-1.1-HF"
-    tokenizer = AutoTokenizer.from_pretrained(hf_model_path, use_fast=False)
-    compilation_prompt = "".join(["0" for _ in range(17)])
-    compilation_input_ids = tokenizer(
-        compilation_prompt,
-        return_tensors="pt",
-    ).input_ids
-    compilation_input_ids = torch.tensor(compilation_input_ids).reshape(
-        [1, 19]
-    )
-    firstVicunaCompileInput = (compilation_input_ids,)
-
-    model = CombinedModel()
-    c = GenerateConfigFile(model, 1, ["gpu_id"], firstVicunaCompileInput)
-    return c.split_into_layers()
-
-
-with gr.Blocks() as model_config_web:
-    with gr.Row():
-        hf_models = gr.Dropdown(
-            label="Model List",
-            choices=["Vicuna"],
-            value="Vicuna",
-            visible=True,
-        )
-        get_model_config_btn = gr.Button(value="Get Model Config")
-    json_view = gr.JSON()
-
-    get_model_config_btn.click(
-        fn=get_model_config,
-        inputs=[],
-        outputs=[json_view],
-    )

apps/stable_diffusion/web/ui/h2ogpt.py

@@ -12,10 +12,6 @@ from apps.language_models.langchain.enums import (
     LangChainAction,
 )
 import apps.language_models.langchain.gen as gen
-from gpt_langchain import (
-    path_to_docs,
-    create_or_update_db,
-)
 from apps.stable_diffusion.src import args
 
 
@@ -37,15 +33,8 @@ start_message = """
 
 def create_prompt(history):
     system_message = start_message
-    for item in history:
-        print("His item: ", item)
 
-    conversation = "<|endoftext|>".join(
-        [
-            "<|endoftext|><|answer|>".join([item[0], item[1]])
-            for item in history
-        ]
-    )
+    conversation = "".join(["".join([item[0], item[1]]) for item in history])
 
     msg = system_message + conversation
     msg = msg.strip()
@@ -55,12 +44,10 @@ def create_prompt(history):
 def chat(curr_system_message, history, device, precision):
     args.run_docuchat_web = True
     global h2ogpt_model
-    global sharkModel
     global h2ogpt_tokenizer
     global model_state
     global langchain
     global userpath_selector
-    from apps.language_models.langchain.h2oai_pipeline import generate_token
 
     if h2ogpt_model == 0:
         if "cuda" in device:
@@ -115,14 +102,9 @@ def chat(curr_system_message, history, device, precision):
             prompt_type=None,
             prompt_dict=None,
         )
-        from apps.language_models.langchain.h2oai_pipeline import (
-            H2OGPTSHARKModel,
-        )
-
-        sharkModel = H2OGPTSHARKModel()
 
     prompt = create_prompt(history)
-    output_dict = langchain.evaluate(
+    output = langchain.evaluate(
         model_state=model_state,
         my_db_state=None,
         instruction=prompt,
@@ -182,22 +164,14 @@ def chat(curr_system_message, history, device, precision):
         model_lock=True,
         user_path=userpath_selector.value,
     )
-
-    output = generate_token(sharkModel, **output_dict)
     for partial_text in output:
-        history[-1][1] = partial_text
+        history[-1][1] = partial_text["response"]
         yield history
+
     return history
 
 
-userpath_selector = gr.Textbox(
-    label="Document Directory",
-    value=str(os.path.abspath("apps/language_models/langchain/user_path/")),
-    interactive=True,
-    container=True,
-)
-
-with gr.Blocks(title="DocuChat") as h2ogpt_web:
+with gr.Blocks(title="H2OGPT") as h2ogpt_web:
     with gr.Row():
         supported_devices = available_devices
         enabled = len(supported_devices) > 0
@@ -212,7 +186,6 @@ with gr.Blocks(title="DocuChat") as h2ogpt_web:
             else "Only CUDA Supported for now",
             choices=supported_devices,
             interactive=enabled,
-            allow_custom_value=True,
         )
         precision = gr.Radio(
             label="Precision",
@@ -225,6 +198,14 @@ with gr.Blocks(title="DocuChat") as h2ogpt_web:
             ],
             visible=True,
         )
+        userpath_selector = gr.Textbox(
+            label="Document Directory",
+            value=str(
+                os.path.abspath("apps/language_models/langchain/user_path/")
+            ),
+            interactive=True,
+            container=True,
+        )
     chatbot = gr.Chatbot(height=500)
     with gr.Row():
         with gr.Column():
@@ -268,100 +249,3 @@ with gr.Blocks(title="DocuChat") as h2ogpt_web:
         queue=False,
     )
     clear.click(lambda: None, None, [chatbot], queue=False)
-
-
-with gr.Blocks(title="DocuChat Upload") as h2ogpt_upload:
-    import pathlib
-
-    upload_path = None
-    database = None
-    database_directory = os.path.abspath(
-        "apps/language_models/langchain/db_path/"
-    )
-
-    def read_path():
-        global upload_path
-        filenames = [
-            [f]
-            for f in os.listdir(upload_path)
-            if os.path.isfile(os.path.join(upload_path, f))
-        ]
-        filenames.sort()
-        return filenames
-
-    def upload_file(f):
-        names = []
-        for tmpfile in f:
-            name = tmpfile.name.split("/")[-1]
-            basename = os.path.join(upload_path, name)
-            with open(basename, "wb") as w:
-                with open(tmpfile.name, "rb") as r:
-                    w.write(r.read())
-        update_or_create_db()
-        return read_path()
-
-    def update_userpath(newpath):
-        global upload_path
-        upload_path = newpath
-        pathlib.Path(upload_path).mkdir(parents=True, exist_ok=True)
-        return read_path()
-
-    def update_or_create_db():
-        global database
-        global upload_path
-
-        sources = path_to_docs(
-            upload_path,
-            verbose=True,
-            fail_any_exception=False,
-            n_jobs=-1,
-            chunk=True,
-            chunk_size=512,
-            url=None,
-            enable_captions=False,
-            captions_model=None,
-            caption_loader=None,
-            enable_ocr=False,
-        )
-
-        pathlib.Path(database_directory).mkdir(parents=True, exist_ok=True)
-
-        database = create_or_update_db(
-            "chroma",
-            database_directory,
-            "UserData",
-            sources,
-            False,
-            True,
-            True,
-            "sentence-transformers/all-MiniLM-L6-v2",
-        )
-
-    def first_run():
-        global database
-        if database is None:
-            update_or_create_db()
-
-    update_userpath(
-        os.path.abspath("apps/language_models/langchain/user_path/")
-    )
-    h2ogpt_upload.load(fn=first_run)
-    h2ogpt_web.load(fn=first_run)
-
-    with gr.Column():
-        text = gr.DataFrame(
-            col_count=(1, "fixed"),
-            type="array",
-            label="Documents",
-            value=read_path(),
-        )
-        with gr.Row():
-            upload = gr.UploadButton(
-                label="Upload documents",
-                file_count="multiple",
-            )
-            upload.upload(fn=upload_file, inputs=upload, outputs=text)
-            userpath_selector.render()
-            userpath_selector.input(
-                fn=update_userpath, inputs=userpath_selector, outputs=text
-            ).then(fn=update_or_create_db)

apps/stable_diffusion/web/ui/img2img_ui.py

@@ -3,8 +3,10 @@ import torch
 import time
 import gradio as gr
 import PIL
-from math import ceil
 from PIL import Image
+import base64
+from io import BytesIO
+from fastapi.exceptions import HTTPException
 from apps.stable_diffusion.web.ui.utils import (
     available_devices,
     nodlogo_loc,
@@ -27,7 +29,6 @@ from apps.stable_diffusion.src import (
 from apps.stable_diffusion.src.utils import (
     get_generated_imgs_path,
     get_generation_text_info,
-    resampler_list,
 )
 from apps.stable_diffusion.web.utils.common_label_calc import status_label
 import numpy as np
@@ -53,7 +54,8 @@ def img2img_inf(
     batch_count: int,
     batch_size: int,
     scheduler: str,
-    model_id: str,
+    custom_model: str,
+    hf_model_id: str,
     custom_vae: str,
     precision: str,
     device: str,
@@ -65,7 +67,6 @@ def img2img_inf(
     lora_hf_id: str,
     ondemand: bool,
     repeatable_seeds: bool,
-    resample_type: str,
 ):
     from apps.stable_diffusion.web.ui.utils import (
         get_custom_model_pathfile,
@@ -100,17 +101,21 @@ def img2img_inf(
     args.ckpt_loc = ""
     args.hf_model_id = ""
     args.custom_vae = ""
-
-    # .safetensor or .chkpt on the custom model path
-    if model_id in get_custom_model_files():
-        args.ckpt_loc = get_custom_model_pathfile(model_id)
-    # civitai download
-    elif "civitai" in model_id:
-        args.ckpt_loc = model_id
-    # either predefined or huggingface
+    if custom_model == "None":
+        if not hf_model_id:
+            return (
+                None,
+                "Please provide either custom model or huggingface model ID, "
+                "both must not be empty.",
+            )
+        if "civitai" in hf_model_id:
+            args.ckpt_loc = hf_model_id
+        else:
+            args.hf_model_id = hf_model_id
+    elif ".ckpt" in custom_model or ".safetensors" in custom_model:
+        args.ckpt_loc = get_custom_model_pathfile(custom_model)
     else:
-        args.hf_model_id = model_id
-
+        args.hf_model_id = custom_model
     if custom_vae != "None":
         args.custom_vae = get_custom_model_pathfile(custom_vae, model="vae")
 
@@ -240,7 +245,7 @@ def img2img_inf(
             batch_size,
             height,
             width,
-            ceil(steps / strength),
+            steps,
             strength,
             guidance_scale,
             seeds[current_batch],
@@ -250,7 +255,6 @@ def img2img_inf(
             cpu_scheduling,
             args.max_embeddings_multiples,
             use_stencil=use_stencil,
-            resample_type=resample_type,
         )
         total_time = time.time() - start_time
         text_output = get_generation_text_info(
@@ -275,6 +279,87 @@ def img2img_inf(
     return generated_imgs, text_output, ""
 
 
+def decode_base64_to_image(encoding):
+    if encoding.startswith("data:image/"):
+        encoding = encoding.split(";", 1)[1].split(",", 1)[1]
+    try:
+        image = Image.open(BytesIO(base64.b64decode(encoding)))
+        return image
+    except Exception as err:
+        print(err)
+        raise HTTPException(status_code=500, detail="Invalid encoded image")
+
+
+def encode_pil_to_base64(images):
+    encoded_imgs = []
+    for image in images:
+        with BytesIO() as output_bytes:
+            if args.output_img_format.lower() == "png":
+                image.save(output_bytes, format="PNG")
+
+            elif args.output_img_format.lower() in ("jpg", "jpeg"):
+                image.save(output_bytes, format="JPEG")
+            else:
+                raise HTTPException(
+                    status_code=500, detail="Invalid image format"
+                )
+            bytes_data = output_bytes.getvalue()
+            encoded_imgs.append(base64.b64encode(bytes_data))
+    return encoded_imgs
+
+
+# Img2Img Rest API.
+def img2img_api(
+    InputData: dict,
+):
+    print(
+        f'Prompt: {InputData["prompt"]}, '
+        f'Negative Prompt: {InputData["negative_prompt"]}, '
+        f'Seed: {InputData["seed"]}.'
+    )
+    init_image = decode_base64_to_image(InputData["init_images"][0])
+    res = img2img_inf(
+        InputData["prompt"],
+        InputData["negative_prompt"],
+        init_image,
+        InputData["height"],
+        InputData["width"],
+        InputData["steps"],
+        InputData["denoising_strength"],
+        InputData["cfg_scale"],
+        InputData["seed"],
+        batch_count=1,
+        batch_size=1,
+        scheduler="EulerDiscrete",
+        custom_model="None",
+        hf_model_id=InputData["hf_model_id"]
+        if "hf_model_id" in InputData.keys()
+        else "stabilityai/stable-diffusion-2-1-base",
+        custom_vae="None",
+        precision="fp16",
+        device=available_devices[0],
+        max_length=64,
+        use_stencil=InputData["use_stencil"]
+        if "use_stencil" in InputData.keys()
+        else "None",
+        save_metadata_to_json=False,
+        save_metadata_to_png=False,
+        lora_weights="None",
+        lora_hf_id="",
+        ondemand=False,
+        repeatable_seeds=False,
+    )
+
+    # Converts generator type to subscriptable
+    res = next(res)
+
+    return {
+        "images": encode_pil_to_base64(res[0]),
+        "parameters": {},
+        "info": res[1],
+    }
+
+
 with gr.Blocks(title="Image-to-Image") as img2img_web:
     with gr.Row(elem_id="ui_title"):
         nod_logo = Image.open(nodlogo_loc)
@@ -293,19 +378,31 @@ with gr.Blocks(title="Image-to-Image") as img2img_web:
             with gr.Column(scale=1, min_width=600):
                 with gr.Row():
                     # janky fix for overflowing text
-                    i2i_model_info = (
-                        f"Custom Model Path: {str(get_custom_model_path())}"
+                    i2i_model_info = (str(get_custom_model_path())).replace(
+                        "\\", "\n\\"
                     )
+                    i2i_model_info = f"Custom Model Path: {i2i_model_info}"
                     img2img_custom_model = gr.Dropdown(
                         label=f"Models",
-                        info="Select, or enter HuggingFace Model ID or Civitai model download URL",
+                        info=i2i_model_info,
                         elem_id="custom_model",
                         value=os.path.basename(args.ckpt_loc)
                         if args.ckpt_loc
                         else "stabilityai/stable-diffusion-2-1-base",
-                        choices=get_custom_model_files() + predefined_models,
-                        allow_custom_value=True,
-                        scale=2,
+                        choices=["None"]
+                        + get_custom_model_files()
+                        + predefined_models,
+                    )
+                    img2img_hf_model_id = gr.Textbox(
+                        elem_id="hf_model_id",
+                        placeholder="Select 'None' in the Models dropdown "
+                        "on the left and enter model ID here "
+                        "e.g: SG161222/Realistic_Vision_V1.3, "
+                        "https://civitai.com/api/download/models/15236",
+                        value="",
+                        label="HuggingFace Model ID or Civitai model "
+                        "download URL",
+                        lines=3,
                     )
                     # janky fix for overflowing text
                     i2i_vae_info = (str(get_custom_model_path("vae"))).replace(
@@ -320,8 +417,6 @@ with gr.Blocks(title="Image-to-Image") as img2img_web:
                         if args.custom_vae
                         else "None",
                         choices=["None"] + get_custom_model_files("vae"),
-                        allow_custom_value=True,
-                        scale=1,
                     )
 
                 with gr.Group(elem_id="prompt_box_outer"):
@@ -337,7 +432,7 @@ with gr.Blocks(title="Image-to-Image") as img2img_web:
                         lines=2,
                         elem_id="negative_prompt_box",
                     )
-                # TODO: make this import image prompt info if it exists
+
                 img2img_init_image = gr.Image(
                     label="Input Image",
                     source="upload",
@@ -352,13 +447,7 @@ with gr.Blocks(title="Image-to-Image") as img2img_web:
                             elem_id="stencil_model",
                             label="Stencil model",
                             value="None",
-                            choices=[
-                                "None",
-                                "canny",
-                                "openpose",
-                                "scribble",
-                                "zoedepth",
-                            ],
+                            choices=["None", "canny", "openpose", "scribble"],
                         )
 
                     def show_canvas(choice):
@@ -419,7 +508,6 @@ with gr.Blocks(title="Image-to-Image") as img2img_web:
                         ).replace("\\", "\n\\")
                         i2i_lora_info = f"LoRA Path: {i2i_lora_info}"
                         lora_weights = gr.Dropdown(
-                            allow_custom_value=True,
                             label=f"Standalone LoRA Weights",
                             info=i2i_lora_info,
                             elem_id="lora_weights",
@@ -443,7 +531,6 @@ with gr.Blocks(title="Image-to-Image") as img2img_web:
                             label="Scheduler",
                             value="EulerDiscrete",
                             choices=scheduler_list_cpu_only,
-                            allow_custom_value=True,
                         )
                         with gr.Group():
                             save_metadata_to_png = gr.Checkbox(
@@ -463,6 +550,15 @@ with gr.Blocks(title="Image-to-Image") as img2img_web:
                         width = gr.Slider(
                             384, 768, value=args.width, step=8, label="Width"
                         )
+                        precision = gr.Radio(
+                            label="Precision",
+                            value=args.precision,
+                            choices=[
+                                "fp16",
+                                "fp32",
+                            ],
+                            visible=True,
+                        )
                         max_length = gr.Radio(
                             label="Max Length",
                             value=args.max_length,
@@ -485,26 +581,11 @@ with gr.Blocks(title="Image-to-Image") as img2img_web:
                                 step=0.01,
                                 label="Denoising Strength",
                             )
-                            resample_type = gr.Dropdown(
-                                value=args.resample_type,
-                                choices=resampler_list,
-                                label="Resample Type",
-                                allow_custom_value=True,
-                            )
                         ondemand = gr.Checkbox(
                             value=args.ondemand,
                             label="Low VRAM",
                             interactive=True,
                         )
-                        precision = gr.Radio(
-                            label="Precision",
-                            value=args.precision,
-                            choices=[
-                                "fp16",
-                                "fp32",
-                            ],
-                            visible=True,
-                        )
                     with gr.Row():
                         with gr.Column(scale=3):
                             guidance_scale = gr.Slider(
@@ -548,8 +629,17 @@ with gr.Blocks(title="Image-to-Image") as img2img_web:
                         label="Device",
                         value=available_devices[0],
                         choices=available_devices,
-                        allow_custom_value=True,
                     )
+                with gr.Row():
+                    random_seed = gr.Button("Randomize Seed")
+                    random_seed.click(
+                        lambda: -1,
+                        inputs=[],
+                        outputs=[seed],
+                        queue=False,
+                    )
+                    stop_batch = gr.Button("Stop Batch")
+                    stable_diffusion = gr.Button("Generate Image(s)")
 
             with gr.Column(scale=1, min_width=600):
                 with gr.Group():
@@ -561,26 +651,13 @@ with gr.Blocks(title="Image-to-Image") as img2img_web:
                         object_fit="contain",
                     )
                     std_output = gr.Textbox(
-                        value=f"{i2i_model_info}\n"
-                        f"Images will be saved at "
+                        value=f"Images will be saved at "
                         f"{get_generated_imgs_path()}",
-                        lines=2,
+                        lines=1,
                         elem_id="std_output",
                         show_label=False,
                     )
                     img2img_status = gr.Textbox(visible=False)
-                with gr.Row():
-                    stable_diffusion = gr.Button("Generate Image(s)")
-                    random_seed = gr.Button("Randomize Seed")
-                    random_seed.click(
-                        lambda: -1,
-                        inputs=[],
-                        outputs=[seed],
-                        queue=False,
-                    )
-                    stop_batch = gr.Button("Stop Batch")
-                with gr.Row():
-                    blank_thing_for_row = None
                 with gr.Row():
                     img2img_sendto_inpaint = gr.Button(value="SendTo Inpaint")
                     img2img_sendto_outpaint = gr.Button(
@@ -606,6 +683,7 @@ with gr.Blocks(title="Image-to-Image") as img2img_web:
                 batch_size,
                 scheduler,
                 img2img_custom_model,
+                img2img_hf_model_id,
                 custom_vae,
                 precision,
                 device,
@@ -617,7 +695,6 @@ with gr.Blocks(title="Image-to-Image") as img2img_web:
                 lora_hf_id,
                 ondemand,
                 repeatable_seeds,
-                resample_type,
             ],
             outputs=[img2img_gallery, std_output, img2img_status],
             show_progress="minimal" if args.progress_bar else "none",

apps/stable_diffusion/web/ui/inpaint_ui.py

@@ -4,6 +4,9 @@ import time
 import sys
 import gradio as gr
 from PIL import Image
+import base64
+from io import BytesIO
+from fastapi.exceptions import HTTPException
 from apps.stable_diffusion.web.ui.utils import (
     available_devices,
     nodlogo_loc,
@@ -50,7 +53,8 @@ def inpaint_inf(
     batch_count: int,
     batch_size: int,
     scheduler: str,
-    model_id: str,
+    custom_model: str,
+    hf_model_id: str,
     custom_vae: str,
     precision: str,
     device: str,
@@ -85,17 +89,21 @@ def inpaint_inf(
     args.ckpt_loc = ""
     args.hf_model_id = ""
     args.custom_vae = ""
-
-    # .safetensor or .chkpt on the custom model path
-    if model_id in get_custom_model_files(custom_checkpoint_type="inpainting"):
-        args.ckpt_loc = get_custom_model_pathfile(model_id)
-    # civitai download
-    elif "civitai" in model_id:
-        args.ckpt_loc = model_id
-    # either predefined or huggingface
+    if custom_model == "None":
+        if not hf_model_id:
+            return (
+                None,
+                "Please provide either custom model or huggingface model ID, "
+                "both must not be empty.",
+            )
+        if "civitai" in hf_model_id:
+            args.ckpt_loc = hf_model_id
+        else:
+            args.hf_model_id = hf_model_id
+    elif ".ckpt" in custom_model or ".safetensors" in custom_model:
+        args.ckpt_loc = get_custom_model_pathfile(custom_model)
     else:
-        args.hf_model_id = model_id
-
+        args.hf_model_id = custom_model
     if custom_vae != "None":
         args.custom_vae = get_custom_model_pathfile(custom_vae, model="vae")
 
@@ -220,6 +228,86 @@ def inpaint_inf(
     return generated_imgs, text_output
 
 
+def decode_base64_to_image(encoding):
+    if encoding.startswith("data:image/"):
+        encoding = encoding.split(";", 1)[1].split(",", 1)[1]
+    try:
+        image = Image.open(BytesIO(base64.b64decode(encoding)))
+        return image
+    except Exception as err:
+        print(err)
+        raise HTTPException(status_code=500, detail="Invalid encoded image")
+
+
+def encode_pil_to_base64(images):
+    encoded_imgs = []
+    for image in images:
+        with BytesIO() as output_bytes:
+            if args.output_img_format.lower() == "png":
+                image.save(output_bytes, format="PNG")
+
+            elif args.output_img_format.lower() in ("jpg", "jpeg"):
+                image.save(output_bytes, format="JPEG")
+            else:
+                raise HTTPException(
+                    status_code=500, detail="Invalid image format"
+                )
+            bytes_data = output_bytes.getvalue()
+            encoded_imgs.append(base64.b64encode(bytes_data))
+    return encoded_imgs
+
+
+# Inpaint Rest API.
+def inpaint_api(
+    InputData: dict,
+):
+    print(
+        f'Prompt: {InputData["prompt"]}, '
+        f'Negative Prompt: {InputData["negative_prompt"]}, '
+        f'Seed: {InputData["seed"]}.'
+    )
+    init_image = decode_base64_to_image(InputData["image"])
+    mask = decode_base64_to_image(InputData["mask"])
+    res = inpaint_inf(
+        InputData["prompt"],
+        InputData["negative_prompt"],
+        {"image": init_image, "mask": mask},
+        InputData["height"],
+        InputData["width"],
+        InputData["is_full_res"],
+        InputData["full_res_padding"],
+        InputData["steps"],
+        InputData["cfg_scale"],
+        InputData["seed"],
+        batch_count=1,
+        batch_size=1,
+        scheduler="EulerDiscrete",
+        custom_model="None",
+        hf_model_id=InputData["hf_model_id"]
+        if "hf_model_id" in InputData.keys()
+        else "stabilityai/stable-diffusion-2-inpainting",
+        custom_vae="None",
+        precision="fp16",
+        device=available_devices[0],
+        max_length=64,
+        save_metadata_to_json=False,
+        save_metadata_to_png=False,
+        lora_weights="None",
+        lora_hf_id="",
+        ondemand=False,
+        repeatable_seeds=False,
+    )
+
+    # Converts generator type to subscriptable
+    res = next(res)
+
+    return {
+        "images": encode_pil_to_base64(res[0]),
+        "parameters": {},
+        "info": res[1],
+    }
+
+
 with gr.Blocks(title="Inpainting") as inpaint_web:
     with gr.Row(elem_id="ui_title"):
         nod_logo = Image.open(nodlogo_loc)
@@ -239,21 +327,34 @@ with gr.Blocks(title="Inpainting") as inpaint_web:
                 with gr.Row():
                     # janky fix for overflowing text
                     inpaint_model_info = (
-                        f"Custom Model Path: {str(get_custom_model_path())}"
+                        str(get_custom_model_path())
+                    ).replace("\\", "\n\\")
+                    inpaint_model_info = (
+                        f"Custom Model Path: {inpaint_model_info}"
                     )
                     inpaint_custom_model = gr.Dropdown(
                         label=f"Models",
-                        info="Select, or enter HuggingFace Model ID or Civitai model download URL",
+                        info=inpaint_model_info,
                         elem_id="custom_model",
                         value=os.path.basename(args.ckpt_loc)
                         if args.ckpt_loc
                         else "stabilityai/stable-diffusion-2-inpainting",
-                        choices=get_custom_model_files(
+                        choices=["None"]
+                        + get_custom_model_files(
                             custom_checkpoint_type="inpainting"
                         )
                         + predefined_paint_models,
-                        allow_custom_value=True,
-                        scale=2,
+                    )
+                    inpaint_hf_model_id = gr.Textbox(
+                        elem_id="hf_model_id",
+                        placeholder="Select 'None' in the Models dropdown "
+                        "on the left and enter model ID here "
+                        "e.g: ghunkins/stable-diffusion-liberty-inpainting, "
+                        "https://civitai.com/api/download/models/3433",
+                        value="",
+                        label="HuggingFace Model ID or Civitai model "
+                        "download URL",
+                        lines=3,
                     )
                     # janky fix for overflowing text
                     inpaint_vae_info = (
@@ -268,8 +369,6 @@ with gr.Blocks(title="Inpainting") as inpaint_web:
                         if args.custom_vae
                         else "None",
                         choices=["None"] + get_custom_model_files("vae"),
-                        allow_custom_value=True,
-                        scale=1,
                     )
 
                 with gr.Group(elem_id="prompt_box_outer"):
@@ -307,7 +406,6 @@ with gr.Blocks(title="Inpainting") as inpaint_web:
                             elem_id="lora_weights",
                             value="None",
                             choices=["None"] + get_custom_model_files("lora"),
-                            allow_custom_value=True,
                         )
                         lora_hf_id = gr.Textbox(
                             elem_id="lora_hf_id",
@@ -326,7 +424,6 @@ with gr.Blocks(title="Inpainting") as inpaint_web:
                             label="Scheduler",
                             value="EulerDiscrete",
                             choices=scheduler_list_cpu_only,
-                            allow_custom_value=True,
                         )
                         with gr.Group():
                             save_metadata_to_png = gr.Checkbox(
@@ -430,8 +527,17 @@ with gr.Blocks(title="Inpainting") as inpaint_web:
                         label="Device",
                         value=available_devices[0],
                         choices=available_devices,
-                        allow_custom_value=True,
                     )
+                with gr.Row():
+                    random_seed = gr.Button("Randomize Seed")
+                    random_seed.click(
+                        lambda: -1,
+                        inputs=[],
+                        outputs=[seed],
+                        queue=False,
+                    )
+                    stop_batch = gr.Button("Stop Batch")
+                    stable_diffusion = gr.Button("Generate Image(s)")
 
             with gr.Column(scale=1, min_width=600):
                 with gr.Group():
@@ -443,26 +549,14 @@ with gr.Blocks(title="Inpainting") as inpaint_web:
                         object_fit="contain",
                     )
                     std_output = gr.Textbox(
-                        value=f"{inpaint_model_info}\n"
-                        "Images will be saved at "
+                        value=f"Images will be saved at "
                         f"{get_generated_imgs_path()}",
-                        lines=2,
+                        lines=1,
                         elem_id="std_output",
                         show_label=False,
                     )
                     inpaint_status = gr.Textbox(visible=False)
-                with gr.Row():
-                    stable_diffusion = gr.Button("Generate Image(s)")
-                    random_seed = gr.Button("Randomize Seed")
-                    random_seed.click(
-                        lambda: -1,
-                        inputs=[],
-                        outputs=[seed],
-                        queue=False,
-                    )
-                    stop_batch = gr.Button("Stop Batch")
-                with gr.Row():
-                    blank_thing_for_row = None
+
                 with gr.Row():
                     inpaint_sendto_img2img = gr.Button(value="SendTo Img2Img")
                     inpaint_sendto_outpaint = gr.Button(
@@ -489,6 +583,7 @@ with gr.Blocks(title="Inpainting") as inpaint_web:
                 batch_size,
                 scheduler,
                 inpaint_custom_model,
+                inpaint_hf_model_id,
                 custom_vae,
                 precision,
                 device,

apps/stable_diffusion/web/ui/lora_train_ui.py

@@ -50,7 +50,6 @@ with gr.Blocks(title="Lora Training") as lora_train_web:
                                 choices=["None"]
                                 + get_custom_model_files()
                                 + predefined_models,
-                                allow_custom_value=True,
                             )
                             hf_model_id = gr.Textbox(
                                 elem_id="hf_model_id",
@@ -74,7 +73,6 @@ with gr.Blocks(title="Lora Training") as lora_train_web:
                         elem_id="lora_weights",
                         value="None",
                         choices=["None"] + get_custom_model_files("lora"),
-                        allow_custom_value=True,
                     )
                     lora_hf_id = gr.Textbox(
                         elem_id="lora_hf_id",
@@ -107,7 +105,6 @@ with gr.Blocks(title="Lora Training") as lora_train_web:
                             label="Scheduler",
                             value=args.scheduler,
                             choices=scheduler_list,
-                            allow_custom_value=True,
                         )
                     with gr.Row():
                         height = gr.Slider(
@@ -180,7 +177,6 @@ with gr.Blocks(title="Lora Training") as lora_train_web:
                         label="Device",
                         value=available_devices[0],
                         choices=available_devices,
-                        allow_custom_value=True,
                     )
                 with gr.Row():
                     with gr.Column(scale=2):

apps/stable_diffusion/web/ui/minigpt4_ui.py

@@ -109,7 +109,7 @@ with gr.Blocks() as minigpt4_web:
     gr.Markdown(description)
 
     with gr.Row():
-        with gr.Column():
+        with gr.Column(scale=0.5):
             image = gr.Image(type="pil")
             upload_button = gr.Button(
                 value="Upload & Start Chat",
@@ -143,7 +143,6 @@ with gr.Blocks() as minigpt4_web:
                 # else "Only CUDA Supported for now",
                 choices=["cuda"],
                 interactive=False,
-                allow_custom_value=True,
             )
 
         with gr.Column():

apps/stable_diffusion/web/ui/model_manager.py

@@ -98,7 +98,6 @@ with gr.Blocks() as model_web:
         choices=None,
         value=None,
         visible=False,
-        allow_custom_value=True,
     )
     # TODO: select and SendTo
     civit_models = gr.Gallery(

apps/stable_diffusion/web/ui/outpaint_ui.py

@@ -53,7 +53,8 @@ def outpaint_inf(
     batch_count: int,
     batch_size: int,
     scheduler: str,
-    model_id: str,
+    custom_model: str,
+    hf_model_id: str,
     custom_vae: str,
     precision: str,
     device: str,
@@ -87,17 +88,21 @@ def outpaint_inf(
     args.ckpt_loc = ""
     args.hf_model_id = ""
     args.custom_vae = ""
-
-    # .safetensor or .chkpt on the custom model path
-    if model_id in get_custom_model_files(custom_checkpoint_type="inpainting"):
-        args.ckpt_loc = get_custom_model_pathfile(model_id)
-    # civitai download
-    elif "civitai" in model_id:
-        args.ckpt_loc = model_id
-    # either predefined or huggingface
+    if custom_model == "None":
+        if not hf_model_id:
+            return (
+                None,
+                "Please provide either custom model or huggingface model ID, "
+                "both must not be empty.",
+            )
+        if "civitai" in hf_model_id:
+            args.ckpt_loc = hf_model_id
+        else:
+            args.hf_model_id = hf_model_id
+    elif ".ckpt" in custom_model or ".safetensors" in custom_model:
+        args.ckpt_loc = get_custom_model_pathfile(custom_model)
     else:
-        args.hf_model_id = model_id
-
+        args.hf_model_id = custom_model
     if custom_vae != "None":
         args.custom_vae = get_custom_model_pathfile(custom_vae, model="vae")
 
@@ -228,6 +233,88 @@ def outpaint_inf(
     return generated_imgs, text_output, ""
 
 
+def decode_base64_to_image(encoding):
+    if encoding.startswith("data:image/"):
+        encoding = encoding.split(";", 1)[1].split(",", 1)[1]
+    try:
+        image = Image.open(BytesIO(base64.b64decode(encoding)))
+        return image
+    except Exception as err:
+        print(err)
+        raise HTTPException(status_code=500, detail="Invalid encoded image")
+
+
+def encode_pil_to_base64(images):
+    encoded_imgs = []
+    for image in images:
+        with BytesIO() as output_bytes:
+            if args.output_img_format.lower() == "png":
+                image.save(output_bytes, format="PNG")
+
+            elif args.output_img_format.lower() in ("jpg", "jpeg"):
+                image.save(output_bytes, format="JPEG")
+            else:
+                raise HTTPException(
+                    status_code=500, detail="Invalid image format"
+                )
+            bytes_data = output_bytes.getvalue()
+            encoded_imgs.append(base64.b64encode(bytes_data))
+    return encoded_imgs
+
+
+# Inpaint Rest API.
+def outpaint_api(
+    InputData: dict,
+):
+    print(
+        f'Prompt: {InputData["prompt"]}, '
+        f'Negative Prompt: {InputData["negative_prompt"]}, '
+        f'Seed: {InputData["seed"]}'
+    )
+    init_image = decode_base64_to_image(InputData["init_images"][0])
+    res = outpaint_inf(
+        InputData["prompt"],
+        InputData["negative_prompt"],
+        init_image,
+        InputData["pixels"],
+        InputData["mask_blur"],
+        InputData["directions"],
+        InputData["noise_q"],
+        InputData["color_variation"],
+        InputData["height"],
+        InputData["width"],
+        InputData["steps"],
+        InputData["cfg_scale"],
+        InputData["seed"],
+        batch_count=1,
+        batch_size=1,
+        scheduler="EulerDiscrete",
+        custom_model="None",
+        hf_model_id=InputData["hf_model_id"]
+        if "hf_model_id" in InputData.keys()
+        else "stabilityai/stable-diffusion-2-inpainting",
+        custom_vae="None",
+        precision="fp16",
+        device=available_devices[0],
+        max_length=64,
+        save_metadata_to_json=False,
+        save_metadata_to_png=False,
+        lora_weights="None",
+        lora_hf_id="",
+        ondemand=False,
+        repeatable_seeds=False,
+    )
+
+    # Convert Generator to Subscriptable
+    res = next(res)
+
+    return {
+        "images": encode_pil_to_base64(res[0]),
+        "parameters": {},
+        "info": res[1],
+    }
+
+
 with gr.Blocks(title="Outpainting") as outpaint_web:
     with gr.Row(elem_id="ui_title"):
         nod_logo = Image.open(nodlogo_loc)
@@ -245,22 +332,36 @@ with gr.Blocks(title="Outpainting") as outpaint_web:
         with gr.Row():
             with gr.Column(scale=1, min_width=600):
                 with gr.Row():
+                    # janky fix for overflowing text
                     outpaint_model_info = (
-                        f"Custom Model Path: {str(get_custom_model_path())}"
+                        str(get_custom_model_path())
+                    ).replace("\\", "\n\\")
+                    outpaint_model_info = (
+                        f"Custom Model Path: {outpaint_model_info}"
                     )
                     outpaint_custom_model = gr.Dropdown(
                         label=f"Models",
-                        info="Select, or enter HuggingFace Model ID or Civitai model download URL",
+                        info=outpaint_model_info,
                         elem_id="custom_model",
                         value=os.path.basename(args.ckpt_loc)
                         if args.ckpt_loc
                         else "stabilityai/stable-diffusion-2-inpainting",
-                        choices=get_custom_model_files(
+                        choices=["None"]
+                        + get_custom_model_files(
                             custom_checkpoint_type="inpainting"
                         )
                         + predefined_paint_models,
-                        allow_custom_value=True,
-                        scale=2,
+                    )
+                    outpaint_hf_model_id = gr.Textbox(
+                        elem_id="hf_model_id",
+                        placeholder="Select 'None' in the Models dropdown "
+                        "on the left and enter model ID here "
+                        "e.g: ghunkins/stable-diffusion-liberty-inpainting, "
+                        "https://civitai.com/api/download/models/3433",
+                        value="",
+                        label="HuggingFace Model ID or Civitai model "
+                        "download URL",
+                        lines=3,
                     )
                     # janky fix for overflowing text
                     outpaint_vae_info = (
@@ -275,9 +376,8 @@ with gr.Blocks(title="Outpainting") as outpaint_web:
                         if args.custom_vae
                         else "None",
                         choices=["None"] + get_custom_model_files("vae"),
-                        allow_custom_value=True,
-                        scale=1,
                     )
+
                 with gr.Group(elem_id="prompt_box_outer"):
                     prompt = gr.Textbox(
                         label="Prompt",
@@ -311,7 +411,6 @@ with gr.Blocks(title="Outpainting") as outpaint_web:
                             elem_id="lora_weights",
                             value="None",
                             choices=["None"] + get_custom_model_files("lora"),
-                            allow_custom_value=True,
                         )
                         lora_hf_id = gr.Textbox(
                             elem_id="lora_hf_id",
@@ -330,7 +429,6 @@ with gr.Blocks(title="Outpainting") as outpaint_web:
                             label="Scheduler",
                             value="EulerDiscrete",
                             choices=scheduler_list_cpu_only,
-                            allow_custom_value=True,
                         )
                         with gr.Group():
                             save_metadata_to_png = gr.Checkbox(
@@ -457,8 +555,17 @@ with gr.Blocks(title="Outpainting") as outpaint_web:
                         label="Device",
                         value=available_devices[0],
                         choices=available_devices,
-                        allow_custom_value=True,
                     )
+                with gr.Row():
+                    random_seed = gr.Button("Randomize Seed")
+                    random_seed.click(
+                        lambda: -1,
+                        inputs=[],
+                        outputs=[seed],
+                        queue=False,
+                    )
+                    stop_batch = gr.Button("Stop Batch")
+                    stable_diffusion = gr.Button("Generate Image(s)")
 
             with gr.Column(scale=1, min_width=600):
                 with gr.Group():
@@ -470,26 +577,13 @@ with gr.Blocks(title="Outpainting") as outpaint_web:
                         object_fit="contain",
                     )
                     std_output = gr.Textbox(
-                        value=f"{outpaint_model_info}\n"
-                        f"Images will be saved at "
+                        value=f"Images will be saved at "
                         f"{get_generated_imgs_path()}",
-                        lines=2,
+                        lines=1,
                         elem_id="std_output",
                         show_label=False,
                     )
                     outpaint_status = gr.Textbox(visible=False)
-                with gr.Row():
-                    stable_diffusion = gr.Button("Generate Image(s)")
-                    random_seed = gr.Button("Randomize Seed")
-                    random_seed.click(
-                        lambda: -1,
-                        inputs=[],
-                        outputs=[seed],
-                        queue=False,
-                    )
-                    stop_batch = gr.Button("Stop Batch")
-                with gr.Row():
-                    blank_thing_for_row = None
                 with gr.Row():
                     outpaint_sendto_img2img = gr.Button(value="SendTo Img2Img")
                     outpaint_sendto_inpaint = gr.Button(value="SendTo Inpaint")
@@ -517,6 +611,7 @@ with gr.Blocks(title="Outpainting") as outpaint_web:
                 batch_size,
                 scheduler,
                 outpaint_custom_model,
+                outpaint_hf_model_id,
                 custom_vae,
                 precision,
                 device,

apps/stable_diffusion/web/ui/outputgallery_ui.py

@@ -109,7 +109,6 @@ with gr.Blocks() as outputgallery_web:
                             value="",
                             interactive=True,
                             elem_classes="dropdown_no_container",
-                            allow_custom_value=True,
                         )
                     with gr.Column(
                         scale=1,

apps/stable_diffusion/web/ui/stablelm_ui.py

@@ -7,8 +7,6 @@ from transformers import (
 )
 from apps.stable_diffusion.web.ui.utils import available_devices
 from datetime import datetime as dt
-import json
-import sys
 
 
 def user(message, history):
@@ -24,81 +22,81 @@ past_key_values = None
 
 model_map = {
     "llama2_7b": "meta-llama/Llama-2-7b-chat-hf",
-    "llama2_13b": "meta-llama/Llama-2-13b-chat-hf",
     "llama2_70b": "meta-llama/Llama-2-70b-chat-hf",
+    "codegen": "Salesforce/codegen25-7b-multi",
+    "vicuna1p3": "lmsys/vicuna-7b-v1.3",
     "vicuna": "TheBloke/vicuna-7B-1.1-HF",
+    "StableLM": "stabilityai/stablelm-tuned-alpha-3b",
 }
 
 # NOTE: Each `model_name` should have its own start message
 start_message = {
     "llama2_7b": (
-        "You are a helpful, respectful and honest assistant. Always answer "
-        "as helpfully as possible, while being safe. Your answers should not "
-        "include any harmful, unethical, racist, sexist, toxic, dangerous, or "
-        "illegal content. Please ensure that your responses are socially "
-        "unbiased and positive in nature. If a question does not make any "
-        "sense, or is not factually coherent, explain why instead of "
-        "answering something not correct. If you don't know the answer "
-        "to a question, please don't share false information."
-    ),
-    "llama2_13b": (
-        "You are a helpful, respectful and honest assistant. Always answer "
-        "as helpfully as possible, while being safe. Your answers should not "
-        "include any harmful, unethical, racist, sexist, toxic, dangerous, or "
-        "illegal content. Please ensure that your responses are socially "
-        "unbiased and positive in nature. If a question does not make any "
-        "sense, or is not factually coherent, explain why instead of "
-        "answering something not correct. If you don't know the answer "
-        "to a question, please don't share false information."
+        "System: You are a helpful, respectful and honest assistant. Always answer "
+        "as helpfully as possible, while being safe.  Your answers should not "
+        "include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal "
+        "content. Please ensure that your responses are socially unbiased and positive "
+        "in nature. If a question does not make any sense, or is not factually coherent, "
+        "explain why instead of answering something not correct. If you don't know the "
+        "answer to a question, please don't share false information."
     ),
     "llama2_70b": (
-        "You are a helpful, respectful and honest assistant. Always answer "
-        "as helpfully as possible, while being safe. Your answers should not "
-        "include any harmful, unethical, racist, sexist, toxic, dangerous, or "
-        "illegal content. Please ensure that your responses are socially "
-        "unbiased and positive in nature. If a question does not make any "
-        "sense, or is not factually coherent, explain why instead of "
-        "answering something not correct. If you don't know the answer "
-        "to a question, please don't share false information."
+        "System: You are a helpful, respectful and honest assistant. Always answer "
+        "as helpfully as possible, while being safe.  Your answers should not "
+        "include any harmful, unethical, racist, sexist, toxic, dangerous, or illegal "
+        "content. Please ensure that your responses are socially unbiased and positive "
+        "in nature. If a question does not make any sense, or is not factually coherent, "
+        "explain why instead of answering something not correct. If you don't know the "
+        "answer to a question, please don't share false information."
+    ),
+    "StableLM": (
+        "<|SYSTEM|># StableLM Tuned (Alpha version)"
+        "\n- StableLM is a helpful and harmless open-source AI language model "
+        "developed by StabilityAI."
+        "\n- StableLM is excited to be able to help the user, but will refuse "
+        "to do anything that could be considered harmful to the user."
+        "\n- StableLM is more than just an information source, StableLM is also "
+        "able to write poetry, short stories, and make jokes."
+        "\n- StableLM will refuse to participate in anything that "
+        "could harm a human."
     ),
     "vicuna": (
-        "A chat between a curious user and an artificial intelligence "
-        "assistant. The assistant gives helpful, detailed, and "
-        "polite answers to the user's questions.\n"
+        "A chat between a curious user and an artificial intelligence assistant. "
+        "The assistant gives helpful, detailed, and polite answers to the user's "
+        "questions.\n"
     ),
+    "vicuna1p3": (
+        "A chat between a curious user and an artificial intelligence assistant. "
+        "The assistant gives helpful, detailed, and polite answers to the user's "
+        "questions.\n"
+    ),
+    "codegen": "",
 }
 
 
-def create_prompt(model_name, history, prompt_prefix):
-    system_message = ""
-    if prompt_prefix:
-        system_message = start_message[model_name]
+def create_prompt(model_name, history):
+    system_message = start_message[model_name]
 
-    if "llama2" in model_name:
-        B_INST, E_INST = "[INST]", "[/INST]"
-        B_SYS, E_SYS = "<<SYS>>\n", "\n<</SYS>>\n\n"
-        conversation = "".join(
-            [f"{B_INST} {item[0]} {E_INST} {item[1]} " for item in history[1:]]
-        )
-        if prompt_prefix:
-            msg = f"{B_INST} {B_SYS}{system_message}{E_SYS}{history[0][0]} {E_INST} {history[0][1]} {conversation}"
-        else:
-            msg = f"{B_INST} {history[0][0]} {E_INST} {history[0][1]} {conversation}"
-    elif model_name in ["vicuna"]:
+    if model_name in [
+        "StableLM",
+        "vicuna",
+        "vicuna1p3",
+        "llama2_7b",
+        "llama2_70b",
+    ]:
         conversation = "".join(
             [
                 "".join(["<|USER|>" + item[0], "<|ASSISTANT|>" + item[1]])
                 for item in history
             ]
         )
-        msg = system_message + conversation
-        msg = msg.strip()
     else:
         conversation = "".join(
             ["".join([item[0], item[1]]) for item in history]
         )
-        msg = system_message + conversation
-        msg = msg.strip()
+
+    msg = system_message + conversation
+    msg = msg.strip()
     return msg
 
 
@@ -107,193 +105,84 @@ def set_vicuna_model(model):
     vicuna_model = model
 
 
-def get_default_config():
-    import torch
-    from transformers import AutoTokenizer
-
-    hf_model_path = "TheBloke/vicuna-7B-1.1-HF"
-    tokenizer = AutoTokenizer.from_pretrained(hf_model_path, use_fast=False)
-    compilation_prompt = "".join(["0" for _ in range(17)])
-    compilation_input_ids = tokenizer(
-        compilation_prompt,
-        return_tensors="pt",
-    ).input_ids
-    compilation_input_ids = torch.tensor(compilation_input_ids).reshape(
-        [1, 19]
-    )
-    firstVicunaCompileInput = (compilation_input_ids,)
-    from apps.language_models.src.model_wrappers.vicuna_model import (
-        CombinedModel,
-    )
-    from shark.shark_generate_model_config import GenerateConfigFile
-
-    model = CombinedModel()
-    c = GenerateConfigFile(model, 1, ["gpu_id"], firstVicunaCompileInput)
-    c.split_into_layers()
-
-
-def clean_device_info(raw_device):
-    # return appropriate device and device_id for consumption by LLM pipeline
-    # Multiple devices only supported for vulkan and rocm (as of now).
-    # default device must be selected for all others
-
-    device_id = None
-    device = (
-        raw_device
-        if "=>" not in raw_device
-        else raw_device.split("=>")[1].strip()
-    )
-    if "://" in device:
-        device, device_id = device.split("://")
-        device_id = int(device_id)  # using device index in webui
-
-    if device not in ["rocm", "vulkan"]:
-        device_id = None
-
-    return device, device_id
-
-
-model_vmfb_key = ""
-
-
 # TODO: Make chat reusable for UI and API
-def chat(
-    prompt_prefix,
-    history,
-    model,
-    device,
-    precision,
-    download_vmfb,
-    config_file,
-    cli=False,
-    progress=gr.Progress(),
-):
+def chat(curr_system_message, history, model, device, precision, cli=True):
     global past_key_values
-    global model_vmfb_key
+
     global vicuna_model
-
     model_name, model_path = list(map(str.strip, model.split("=>")))
-    device, device_id = clean_device_info(device)
 
-    from apps.language_models.scripts.vicuna import ShardedVicuna
-    from apps.language_models.scripts.vicuna import UnshardedVicuna
-    from apps.stable_diffusion.src import args
-
-    new_model_vmfb_key = f"{model_name}#{model_path}#{device}#{device_id}#{precision}#{download_vmfb}"
-    if vicuna_model is None or new_model_vmfb_key != model_vmfb_key:
-        model_vmfb_key = new_model_vmfb_key
-        max_toks = 128 if model_name == "codegen" else 512
-
-        # get iree flags that need to be overridden, from commandline args
-        _extra_args = []
-        # vulkan target triple
-        vulkan_target_triple = args.iree_vulkan_target_triple
-        from shark.iree_utils.vulkan_utils import (
-            get_all_vulkan_devices,
-            get_vulkan_target_triple,
+    if model_name in [
+        "vicuna",
+        "vicuna1p3",
+        "codegen",
+        "llama2_7b",
+        "llama2_70b",
+    ]:
+        from apps.language_models.scripts.vicuna import (
+            UnshardedVicuna,
         )
+        from apps.stable_diffusion.src import args
 
-        if device == "vulkan":
-            vulkaninfo_list = get_all_vulkan_devices()
-            if vulkan_target_triple == "":
-                # We already have the device_id extracted via WebUI, so we directly use
-                # that to find the target triple.
-                vulkan_target_triple = get_vulkan_target_triple(
-                    vulkaninfo_list[device_id]
-                )
-            _extra_args.append(
-                f"-iree-vulkan-target-triple={vulkan_target_triple}"
-            )
-            if "rdna" in vulkan_target_triple:
-                flags_to_add = [
-                    "--iree-spirv-index-bits=64",
-                ]
-                _extra_args = _extra_args + flags_to_add
+        if vicuna_model == 0:
+            if "cuda" in device:
+                device = "cuda"
+            elif "sync" in device:
+                device = "cpu-sync"
+            elif "task" in device:
+                device = "cpu-task"
+            elif "vulkan" in device:
+                device = "vulkan"
+            else:
+                print("unrecognized device")
 
-            if device_id is None:
-                id = 0
-                for device in vulkaninfo_list:
-                    target_triple = get_vulkan_target_triple(
-                        vulkaninfo_list[id]
-                    )
-                    if target_triple == vulkan_target_triple:
-                        device_id = id
-                        break
-                    id += 1
-
-                assert (
-                    device_id
-                ), f"no vulkan hardware for target-triple '{vulkan_target_triple}' exists"
-            print(f"Will use vulkan target triple : {vulkan_target_triple}")
-
-        elif "rocm" in device:
-            # add iree rocm flags
-            if args.iree_rocm_target_chip != "":
-                _extra_args.append(
-                    f"--iree-rocm-target-chip={args.iree_rocm_target_chip}"
-                )
-                print(f"extra args = {_extra_args}")
-
-        if model_name == "vicuna4":
-            vicuna_model = ShardedVicuna(
-                model_name,
-                hf_model_path=model_path,
-                device=device,
-                precision=precision,
-                max_num_tokens=max_toks,
-                compressed=True,
-                extra_args_cmd=_extra_args,
-            )
-        else:
-            #  if config_file is None:
+            max_toks = 128 if model_name == "codegen" else 512
             vicuna_model = UnshardedVicuna(
                 model_name,
                 hf_model_path=model_path,
                 hf_auth_token=args.hf_auth_token,
                 device=device,
-                vulkan_target_triple=vulkan_target_triple,
                 precision=precision,
                 max_num_tokens=max_toks,
-                download_vmfb=download_vmfb,
-                load_mlir_from_shark_tank=True,
-                extra_args_cmd=_extra_args,
-                device_id=device_id,
             )
+        prompt = create_prompt(model_name, history)
 
-    if vicuna_model is None:
-        sys.exit("Unable to instantiate the model object, exiting.")
+        for partial_text in vicuna_model.generate(prompt, cli=cli):
+            history[-1][1] = partial_text
+            yield history
 
-    prompt = create_prompt(model_name, history, prompt_prefix)
+        return history
+
+    # else Model is StableLM
+    global sharkModel
+    from apps.language_models.src.pipelines.stablelm_pipeline import (
+        SharkStableLM,
+    )
+
+    if sharkModel == 0:
+        # max_new_tokens=512
+        shark_slm = SharkStableLM(
+            model_name
+        )  # pass elements from UI as required
+
+    # Construct the input message string for the model by concatenating the
+    # current system message and conversation history
+    if len(curr_system_message.split()) > 160:
+        print("clearing context")
+    prompt = create_prompt(model_name, history)
+    generate_kwargs = dict(prompt=prompt)
+
+    words_list = shark_slm.generate(**generate_kwargs)
 
     partial_text = ""
-    token_count = 0
-    total_time_ms = 0.001  # In order to avoid divide by zero error
-    prefill_time = 0
-    is_first = True
-    for text, msg, exec_time in progress.tqdm(
-        vicuna_model.generate(prompt, cli=cli),
-        desc="generating response",
-    ):
-        if msg is None:
-            if is_first:
-                prefill_time = exec_time
-                is_first = False
-            else:
-                total_time_ms += exec_time
-                token_count += 1
-            partial_text += text + " "
-            history[-1][1] = partial_text
-            yield history, f"Prefill: {prefill_time:.2f}"
-        elif "formatted" in msg:
-            history[-1][1] = text
-            tokens_per_sec = (token_count / total_time_ms) * 1000
-            yield history, f"Prefill: {prefill_time:.2f} seconds\n Decode: {tokens_per_sec:.2f} tokens/sec"
-        else:
-            sys.exit(
-                "unexpected message from the vicuna generate call, exiting."
-            )
-
-    return history, ""
+    for new_text in words_list:
+        print(new_text)
+        partial_text += new_text
+        history[-1][1] = partial_text
+        # Yield an empty string to clean up the message textbox and the updated
+        # conversation history
+        yield history
+    return words_list
 
 
 def llm_chat_api(InputData: dict):
@@ -329,9 +218,17 @@ def llm_chat_api(InputData: dict):
         UnshardedVicuna,
     )
 
-    device_id = None
     if vicuna_model == 0:
-        device, device_id = clean_device_info(device)
+        if "cuda" in device:
+            device = "cuda"
+        elif "sync" in device:
+            device = "cpu-sync"
+        elif "task" in device:
+            device = "cpu-task"
+        elif "vulkan" in device:
+            device = "vulkan"
+        else:
+            print("unrecognized device")
 
         vicuna_model = UnshardedVicuna(
             model_name,
@@ -339,9 +236,6 @@ def llm_chat_api(InputData: dict):
             device=device,
             precision=precision,
             max_num_tokens=max_toks,
-            download_vmfb=True,
-            load_mlir_from_shark_tank=True,
-            device_id=device_id,
         )
 
     # TODO: add role dict for different models
@@ -406,13 +300,13 @@ with gr.Blocks(title="Chatbot") as stablelm_chat:
             label="Select Model",
             value=model_choices[0],
             choices=model_choices,
-            allow_custom_value=True,
         )
         supported_devices = available_devices
         enabled = len(supported_devices) > 0
         # show cpu-task device first in list for chatbot
         supported_devices = supported_devices[-1:] + supported_devices[:-1]
         supported_devices = [x for x in supported_devices if "sync" not in x]
+        print(supported_devices)
         device = gr.Dropdown(
             label="Device",
             value=supported_devices[0]
@@ -420,39 +314,23 @@ with gr.Blocks(title="Chatbot") as stablelm_chat:
             else "Only CUDA Supported for now",
             choices=supported_devices,
             interactive=enabled,
-            allow_custom_value=True,
-            # multiselect=True,
         )
         precision = gr.Radio(
             label="Precision",
-            value="int4",
+            value="fp16",
             choices=[
                 "int4",
                 "int8",
                 "fp16",
+                "fp32",
             ],
-            visible=False,
+            visible=True,
         )
-        tokens_time = gr.Textbox(label="Tokens generated per second")
-        with gr.Column():
-            download_vmfb = gr.Checkbox(
-                label="Download vmfb from Shark tank if available",
-                value=True,
-                interactive=True,
-            )
-            prompt_prefix = gr.Checkbox(
-                label="Add System Prompt",
-                value=False,
-                interactive=True,
-            )
-
-    with gr.Row(visible=False):
+    with gr.Row():
         with gr.Group():
-            config_file = gr.File(
-                label="Upload sharding configuration", visible=False
-            )
-            json_view_button = gr.Button(label="View as JSON", visible=False)
-        json_view = gr.JSON(interactive=True, visible=False)
+            config_file = gr.File(label="Upload sharding configuration")
+            json_view_button = gr.Button("View as JSON")
+        json_view = gr.JSON()
         json_view_button.click(
             fn=view_json_file, inputs=[config_file], outputs=[json_view]
         )
@@ -471,47 +349,24 @@ with gr.Blocks(title="Chatbot") as stablelm_chat:
                 submit = gr.Button("Submit", interactive=enabled)
                 stop = gr.Button("Stop", interactive=enabled)
                 clear = gr.Button("Clear", interactive=enabled)
+    system_msg = gr.Textbox(
+        start_message, label="System Message", interactive=False, visible=False
+    )
 
     submit_event = msg.submit(
-        fn=user,
-        inputs=[msg, chatbot],
-        outputs=[msg, chatbot],
-        show_progress=False,
-        queue=False,
+        fn=user, inputs=[msg, chatbot], outputs=[msg, chatbot], queue=False
     ).then(
         fn=chat,
-        inputs=[
-            prompt_prefix,
-            chatbot,
-            model,
-            device,
-            precision,
-            download_vmfb,
-            config_file,
-        ],
-        outputs=[chatbot, tokens_time],
-        show_progress=False,
+        inputs=[system_msg, chatbot, model, device, precision],
+        outputs=[chatbot],
         queue=True,
     )
     submit_click_event = submit.click(
-        fn=user,
-        inputs=[msg, chatbot],
-        outputs=[msg, chatbot],
-        show_progress=False,
-        queue=False,
+        fn=user, inputs=[msg, chatbot], outputs=[msg, chatbot], queue=False
     ).then(
         fn=chat,
-        inputs=[
-            prompt_prefix,
-            chatbot,
-            model,
-            device,
-            precision,
-            download_vmfb,
-            config_file,
-        ],
-        outputs=[chatbot, tokens_time],
-        show_progress=False,
+        inputs=[system_msg, chatbot, model, device, precision],
+        outputs=[chatbot],
         queue=True,
     )
     stop.click(

apps/stable_diffusion/web/ui/txt2img_ui.py

@@ -4,14 +4,15 @@ import time
 import sys
 import gradio as gr
 from PIL import Image
-from math import ceil
+import base64
+from io import BytesIO
+from fastapi.exceptions import HTTPException
 from apps.stable_diffusion.web.ui.utils import (
     available_devices,
     nodlogo_loc,
     get_custom_model_path,
     get_custom_model_files,
     scheduler_list,
-    scheduler_list_cpu_only,
     predefined_models,
     cancel_sd,
 )
@@ -25,12 +26,10 @@ from apps.stable_diffusion.src import (
     utils,
     save_output_img,
     prompt_examples,
-    Image2ImagePipeline,
 )
 from apps.stable_diffusion.src.utils import (
     get_generated_imgs_path,
     get_generation_text_info,
-    resampler_list,
 )
 
 # set initial values of iree_vulkan_target_triple, use_tuned and import_mlir.
@@ -51,7 +50,8 @@ def txt2img_inf(
     batch_count: int,
     batch_size: int,
     scheduler: str,
-    model_id: str,
+    custom_model: str,
+    hf_model_id: str,
     custom_vae: str,
     precision: str,
     device: str,
@@ -62,11 +62,6 @@ def txt2img_inf(
     lora_hf_id: str,
     ondemand: bool,
     repeatable_seeds: bool,
-    use_hiresfix: bool,
-    hiresfix_height: int,
-    hiresfix_width: int,
-    hiresfix_strength: float,
-    resample_type: str,
 ):
     from apps.stable_diffusion.web.ui.utils import (
         get_custom_model_pathfile,
@@ -89,17 +84,21 @@ def txt2img_inf(
     args.ckpt_loc = ""
     args.hf_model_id = ""
     args.custom_vae = ""
-
-    # .safetensor or .chkpt on the custom model path
-    if model_id in get_custom_model_files():
-        args.ckpt_loc = get_custom_model_pathfile(model_id)
-    # civitai download
-    elif "civitai" in model_id:
-        args.ckpt_loc = model_id
-    # either predefined or huggingface
+    if custom_model == "None":
+        if not hf_model_id:
+            return (
+                None,
+                "Please provide either custom model or huggingface model ID, "
+                "both must not be empty",
+            )
+        if "civitai" in hf_model_id:
+            args.ckpt_loc = hf_model_id
+        else:
+            args.hf_model_id = hf_model_id
+    elif ".ckpt" in custom_model or ".safetensors" in custom_model:
+        args.ckpt_loc = get_custom_model_pathfile(custom_model)
     else:
-        args.hf_model_id = model_id
-
+        args.hf_model_id = custom_model
     if custom_vae != "None":
         args.custom_vae = get_custom_model_pathfile(custom_vae, model="vae")
 
@@ -139,11 +138,6 @@ def txt2img_inf(
         args.max_length = max_length
         args.height = height
         args.width = width
-        args.use_hiresfix = use_hiresfix
-        args.hiresfix_height = hiresfix_height
-        args.hiresfix_width = hiresfix_width
-        args.hiresfix_strength = hiresfix_strength
-        args.resample_type = resample_type
         args.device = device.split("=>", 1)[1].strip()
         args.iree_vulkan_target_triple = init_iree_vulkan_target_triple
         args.iree_metal_target_platform = init_iree_metal_target_platform
@@ -206,81 +200,6 @@ def txt2img_inf(
             cpu_scheduling,
             args.max_embeddings_multiples,
         )
-        # TODO: allow user to save original image
-        # TODO: add option to let user keep both pipelines loaded, and unload
-        #  either at will
-        # TODO: add custom step value slider
-        # TODO: add option to use secondary model for the img2img pass
-        if use_hiresfix is True:
-            new_config_obj = Config(
-                "img2img",
-                args.hf_model_id,
-                args.ckpt_loc,
-                args.custom_vae,
-                precision,
-                1,
-                max_length,
-                height,
-                width,
-                device,
-                use_lora=args.use_lora,
-                use_stencil="None",
-                ondemand=ondemand,
-            )
-
-            global_obj.clear_cache()
-            global_obj.set_cfg_obj(new_config_obj)
-            set_init_device_flags()
-            model_id = (
-                args.hf_model_id
-                if args.hf_model_id
-                else "stabilityai/stable-diffusion-2-1-base"
-            )
-            global_obj.set_schedulers(get_schedulers(model_id))
-            scheduler_obj = global_obj.get_scheduler(args.scheduler)
-
-            global_obj.set_sd_obj(
-                Image2ImagePipeline.from_pretrained(
-                    scheduler_obj,
-                    args.import_mlir,
-                    args.hf_model_id,
-                    args.ckpt_loc,
-                    args.custom_vae,
-                    args.precision,
-                    args.max_length,
-                    1,
-                    hiresfix_height,
-                    hiresfix_width,
-                    args.use_base_vae,
-                    args.use_tuned,
-                    low_cpu_mem_usage=args.low_cpu_mem_usage,
-                    debug=args.import_debug if args.import_mlir else False,
-                    use_lora=args.use_lora,
-                    ondemand=args.ondemand,
-                )
-            )
-
-            global_obj.set_sd_scheduler(args.scheduler)
-
-            out_imgs = global_obj.get_sd_obj().generate_images(
-                prompt,
-                negative_prompt,
-                out_imgs[0],
-                batch_size,
-                hiresfix_height,
-                hiresfix_width,
-                ceil(steps / hiresfix_strength),
-                hiresfix_strength,
-                guidance_scale,
-                seeds[current_batch],
-                args.max_length,
-                dtype,
-                args.use_base_vae,
-                cpu_scheduling,
-                args.max_embeddings_multiples,
-                use_stencil="None",
-                resample_type=resample_type,
-            )
         total_time = time.time() - start_time
         text_output = get_generation_text_info(
             seeds[: current_batch + 1], device
@@ -300,6 +219,70 @@ def txt2img_inf(
     return generated_imgs, text_output, ""
 
 
+def encode_pil_to_base64(images):
+    encoded_imgs = []
+    for image in images:
+        with BytesIO() as output_bytes:
+            if args.output_img_format.lower() == "png":
+                image.save(output_bytes, format="PNG")
+
+            elif args.output_img_format.lower() in ("jpg", "jpeg"):
+                image.save(output_bytes, format="JPEG")
+            else:
+                raise HTTPException(
+                    status_code=500, detail="Invalid image format"
+                )
+            bytes_data = output_bytes.getvalue()
+            encoded_imgs.append(base64.b64encode(bytes_data))
+    return encoded_imgs
+
+
+# Text2Img Rest API.
+def txt2img_api(
+    InputData: dict,
+):
+    print(
+        f'Prompt: {InputData["prompt"]}, '
+        f'Negative Prompt: {InputData["negative_prompt"]}, '
+        f'Seed: {InputData["seed"]}.'
+    )
+    res = txt2img_inf(
+        InputData["prompt"],
+        InputData["negative_prompt"],
+        InputData["height"],
+        InputData["width"],
+        InputData["steps"],
+        InputData["cfg_scale"],
+        InputData["seed"],
+        batch_count=1,
+        batch_size=1,
+        scheduler="EulerDiscrete",
+        custom_model="None",
+        hf_model_id=InputData["hf_model_id"]
+        if "hf_model_id" in InputData.keys()
+        else "stabilityai/stable-diffusion-2-1-base",
+        custom_vae="None",
+        precision="fp16",
+        device=available_devices[0],
+        max_length=64,
+        save_metadata_to_json=False,
+        save_metadata_to_png=False,
+        lora_weights="None",
+        lora_hf_id="",
+        ondemand=False,
+        repeatable_seeds=False,
+    )
+
+    # Convert Generator to Subscriptable
+    res = next(res)
+
+    return {
+        "images": encode_pil_to_base64(res[0]),
+        "parameters": {},
+        "info": res[1],
+    }
+
+
 with gr.Blocks(title="Text-to-Image") as txt2img_web:
     with gr.Row(elem_id="ui_title"):
         nod_logo = Image.open(nodlogo_loc)
@@ -319,18 +302,32 @@ with gr.Blocks(title="Text-to-Image") as txt2img_web:
                 with gr.Row():
                     with gr.Column(scale=10):
                         with gr.Row():
-                            t2i_model_info = f"Custom Model Path: {str(get_custom_model_path())}"
+                            # janky fix for overflowing text
+                            t2i_model_info = (
+                                str(get_custom_model_path())
+                            ).replace("\\", "\n\\")
+                            t2i_model_info = (
+                                f"Custom Model Path: {t2i_model_info}"
+                            )
                             txt2img_custom_model = gr.Dropdown(
                                 label=f"Models",
-                                info="Select, or enter HuggingFace Model ID or Civitai model download URL",
+                                info=t2i_model_info,
                                 elem_id="custom_model",
                                 value=os.path.basename(args.ckpt_loc)
                                 if args.ckpt_loc
                                 else "stabilityai/stable-diffusion-2-1-base",
-                                choices=get_custom_model_files()
+                                choices=["None"]
+                                + get_custom_model_files()
                                 + predefined_models,
-                                allow_custom_value=True,
-                                scale=2,
+                            )
+                            txt2img_hf_model_id = gr.Textbox(
+                                elem_id="hf_model_id",
+                                placeholder="Select 'None' in the dropdown "
+                                "on the left and enter model ID here.",
+                                value="",
+                                label="HuggingFace Model ID or Civitai model "
+                                "download URL.",
+                                lines=3,
                             )
                             # janky fix for overflowing text
                             t2i_vae_info = (
@@ -346,8 +343,6 @@ with gr.Blocks(title="Text-to-Image") as txt2img_web:
                                 else "None",
                                 choices=["None"]
                                 + get_custom_model_files("vae"),
-                                allow_custom_value=True,
-                                scale=1,
                             )
                     with gr.Column(scale=1, min_width=170):
                         txt2img_png_info_img = gr.Image(
@@ -384,7 +379,6 @@ with gr.Blocks(title="Text-to-Image") as txt2img_web:
                             elem_id="lora_weights",
                             value="None",
                             choices=["None"] + get_custom_model_files("lora"),
-                            allow_custom_value=True,
                         )
                         lora_hf_id = gr.Textbox(
                             elem_id="lora_hf_id",
@@ -403,7 +397,6 @@ with gr.Blocks(title="Text-to-Image") as txt2img_web:
                             label="Scheduler",
                             value=args.scheduler,
                             choices=scheduler_list,
-                            allow_custom_value=True,
                         )
                         with gr.Column():
                             save_metadata_to_png = gr.Checkbox(
@@ -490,41 +483,6 @@ with gr.Blocks(title="Text-to-Image") as txt2img_web:
                             args.repeatable_seeds,
                             label="Repeatable Seeds",
                         )
-                with gr.Accordion(label="Hires Fix Options", open=False):
-                    with gr.Group():
-                        with gr.Row():
-                            use_hiresfix = gr.Checkbox(
-                                value=args.use_hiresfix,
-                                label="Use Hires Fix",
-                                interactive=True,
-                            )
-                            resample_type = gr.Dropdown(
-                                value=args.resample_type,
-                                choices=resampler_list,
-                                label="Resample Type",
-                                allow_custom_value=False,
-                            )
-                        hiresfix_height = gr.Slider(
-                            384,
-                            768,
-                            value=args.hiresfix_height,
-                            step=8,
-                            label="Hires Fix Height",
-                        )
-                        hiresfix_width = gr.Slider(
-                            384,
-                            768,
-                            value=args.hiresfix_width,
-                            step=8,
-                            label="Hires Fix Width",
-                        )
-                        hiresfix_strength = gr.Slider(
-                            0,
-                            1,
-                            value=args.hiresfix_strength,
-                            step=0.01,
-                            label="Hires Fix Denoising Strength",
-                        )
                 with gr.Row():
                     seed = gr.Textbox(
                         value=args.seed,
@@ -536,8 +494,17 @@ with gr.Blocks(title="Text-to-Image") as txt2img_web:
                         label="Device",
                         value=available_devices[0],
                         choices=available_devices,
-                        allow_custom_value=True,
                     )
+                with gr.Row():
+                    random_seed = gr.Button("Randomize Seed")
+                    random_seed.click(
+                        lambda: -1,
+                        inputs=[],
+                        outputs=[seed],
+                        queue=False,
+                    )
+                    stop_batch = gr.Button("Stop Batch")
+                    stable_diffusion = gr.Button("Generate Image(s)")
                 with gr.Accordion(label="Prompt Examples!", open=False):
                     ex = gr.Examples(
                         examples=prompt_examples,
@@ -556,26 +523,13 @@ with gr.Blocks(title="Text-to-Image") as txt2img_web:
                         object_fit="contain",
                     )
                     std_output = gr.Textbox(
-                        value=f"{t2i_model_info}\n"
-                        f"Images will be saved at "
+                        value=f"Images will be saved at "
                         f"{get_generated_imgs_path()}",
                         lines=1,
                         elem_id="std_output",
                         show_label=False,
                     )
                     txt2img_status = gr.Textbox(visible=False)
-                with gr.Row():
-                    stable_diffusion = gr.Button("Generate Image(s)")
-                    random_seed = gr.Button("Randomize Seed")
-                    random_seed.click(
-                        lambda: -1,
-                        inputs=[],
-                        outputs=[seed],
-                        queue=False,
-                    )
-                    stop_batch = gr.Button("Stop Batch")
-                with gr.Row():
-                    blank_thing_for_row = None
                 with gr.Row():
                     txt2img_sendto_img2img = gr.Button(value="SendTo Img2Img")
                     txt2img_sendto_inpaint = gr.Button(value="SendTo Inpaint")
@@ -600,6 +554,7 @@ with gr.Blocks(title="Text-to-Image") as txt2img_web:
                 batch_size,
                 scheduler,
                 txt2img_custom_model,
+                txt2img_hf_model_id,
                 custom_vae,
                 precision,
                 device,
@@ -610,11 +565,6 @@ with gr.Blocks(title="Text-to-Image") as txt2img_web:
                 lora_hf_id,
                 ondemand,
                 repeatable_seeds,
-                use_hiresfix,
-                hiresfix_height,
-                hiresfix_width,
-                hiresfix_strength,
-                resample_type,
             ],
             outputs=[txt2img_gallery, std_output, txt2img_status],
             show_progress="minimal" if args.progress_bar else "none",
@@ -649,6 +599,7 @@ with gr.Blocks(title="Text-to-Image") as txt2img_web:
                 width,
                 height,
                 txt2img_custom_model,
+                txt2img_hf_model_id,
                 lora_weights,
                 lora_hf_id,
                 custom_vae,
@@ -664,28 +615,9 @@ with gr.Blocks(title="Text-to-Image") as txt2img_web:
                 width,
                 height,
                 txt2img_custom_model,
+                txt2img_hf_model_id,
                 lora_weights,
                 lora_hf_id,
                 custom_vae,
             ],
         )
-
-        # SharkEulerDiscrete doesn't work with img2img which hires_fix uses
-        def set_compatible_schedulers(hires_fix_selected):
-            if hires_fix_selected:
-                return gr.Dropdown.update(
-                    choices=scheduler_list_cpu_only,
-                    value="DEISMultistep",
-                )
-            else:
-                return gr.Dropdown.update(
-                    choices=scheduler_list,
-                    value="SharkEulerDiscrete",
-                )
-
-        use_hiresfix.change(
-            fn=set_compatible_schedulers,
-            inputs=[use_hiresfix],
-            outputs=[scheduler],
-            queue=False,
-        )

apps/stable_diffusion/web/ui/upscaler_ui.py

@@ -3,6 +3,9 @@ import torch
 import time
 import gradio as gr
 from PIL import Image
+import base64
+from io import BytesIO
+from fastapi.exceptions import HTTPException
 from apps.stable_diffusion.web.ui.utils import (
     available_devices,
     nodlogo_loc,
@@ -43,7 +46,8 @@ def upscaler_inf(
     batch_count: int,
     batch_size: int,
     scheduler: str,
-    model_id: str,
+    custom_model: str,
+    hf_model_id: str,
     custom_vae: str,
     precision: str,
     device: str,
@@ -81,17 +85,21 @@ def upscaler_inf(
     args.ckpt_loc = ""
     args.hf_model_id = ""
     args.custom_vae = ""
-
-    # .safetensor or .chkpt on the custom model path
-    if model_id in get_custom_model_files(custom_checkpoint_type="upscaler"):
-        args.ckpt_loc = get_custom_model_pathfile(model_id)
-    # civitai download
-    elif "civitai" in model_id:
-        args.ckpt_loc = model_id
-    # either predefined or huggingface
+    if custom_model == "None":
+        if not hf_model_id:
+            return (
+                None,
+                "Please provide either custom model or huggingface model ID, "
+                "both must not be empty.",
+            )
+        if "civitai" in hf_model_id:
+            args.ckpt_loc = hf_model_id
+        else:
+            args.hf_model_id = hf_model_id
+    elif ".ckpt" in custom_model or ".safetensors" in custom_model:
+        args.ckpt_loc = get_custom_model_pathfile(custom_model)
     else:
-        args.hf_model_id = model_id
-
+        args.hf_model_id = custom_model
     if custom_vae != "None":
         args.custom_vae = get_custom_model_pathfile(custom_vae, model="vae")
 
@@ -244,6 +252,83 @@ def upscaler_inf(
     yield generated_imgs, text_output, ""
 
 
+def decode_base64_to_image(encoding):
+    if encoding.startswith("data:image/"):
+        encoding = encoding.split(";", 1)[1].split(",", 1)[1]
+    try:
+        image = Image.open(BytesIO(base64.b64decode(encoding)))
+        return image
+    except Exception as err:
+        print(err)
+        raise HTTPException(status_code=500, detail="Invalid encoded image")
+
+
+def encode_pil_to_base64(images):
+    encoded_imgs = []
+    for image in images:
+        with BytesIO() as output_bytes:
+            if args.output_img_format.lower() == "png":
+                image.save(output_bytes, format="PNG")
+
+            elif args.output_img_format.lower() in ("jpg", "jpeg"):
+                image.save(output_bytes, format="JPEG")
+            else:
+                raise HTTPException(
+                    status_code=500, detail="Invalid image format"
+                )
+            bytes_data = output_bytes.getvalue()
+            encoded_imgs.append(base64.b64encode(bytes_data))
+    return encoded_imgs
+
+
+# Upscaler Rest API.
+def upscaler_api(
+    InputData: dict,
+):
+    print(
+        f'Prompt: {InputData["prompt"]}, '
+        f'Negative Prompt: {InputData["negative_prompt"]}, '
+        f'Seed: {InputData["seed"]}'
+    )
+    init_image = decode_base64_to_image(InputData["init_images"][0])
+    res = upscaler_inf(
+        InputData["prompt"],
+        InputData["negative_prompt"],
+        init_image,
+        InputData["height"],
+        InputData["width"],
+        InputData["steps"],
+        InputData["noise_level"],
+        InputData["cfg_scale"],
+        InputData["seed"],
+        batch_count=1,
+        batch_size=1,
+        scheduler="EulerDiscrete",
+        custom_model="None",
+        hf_model_id=InputData["hf_model_id"]
+        if "hf_model_id" in InputData.keys()
+        else "stabilityai/stable-diffusion-2-1-base",
+        custom_vae="None",
+        precision="fp16",
+        device=available_devices[0],
+        max_length=64,
+        save_metadata_to_json=False,
+        save_metadata_to_png=False,
+        lora_weights="None",
+        lora_hf_id="",
+        ondemand=False,
+        repeatable_seeds=False,
+    )
+    # Converts generator type to subscriptable
+    res = next(res)
+
+    return {
+        "images": encode_pil_to_base64(res[0]),
+        "parameters": {},
+        "info": res[1],
+    }
+
+
 with gr.Blocks(title="Upscaler") as upscaler_web:
     with gr.Row(elem_id="ui_title"):
         nod_logo = Image.open(nodlogo_loc)
@@ -261,22 +346,36 @@ with gr.Blocks(title="Upscaler") as upscaler_web:
         with gr.Row():
             with gr.Column(scale=1, min_width=600):
                 with gr.Row():
+                    # janky fix for overflowing text
                     upscaler_model_info = (
-                        f"Custom Model Path: {str(get_custom_model_path())}"
+                        str(get_custom_model_path())
+                    ).replace("\\", "\n\\")
+                    upscaler_model_info = (
+                        f"Custom Model Path: {upscaler_model_info}"
                     )
                     upscaler_custom_model = gr.Dropdown(
                         label=f"Models",
-                        info="Select, or enter HuggingFace Model ID or Civitai model download URL",
+                        info=upscaler_model_info,
                         elem_id="custom_model",
                         value=os.path.basename(args.ckpt_loc)
                         if args.ckpt_loc
                         else "stabilityai/stable-diffusion-x4-upscaler",
-                        choices=get_custom_model_files(
+                        choices=["None"]
+                        + get_custom_model_files(
                             custom_checkpoint_type="upscaler"
                         )
                         + predefined_upscaler_models,
-                        allow_custom_value=True,
-                        scale=2,
+                    )
+                    upscaler_hf_model_id = gr.Textbox(
+                        elem_id="hf_model_id",
+                        placeholder="Select 'None' in the Models dropdown "
+                        "on the left and enter model ID here "
+                        "e.g: SG161222/Realistic_Vision_V1.3, "
+                        "https://civitai.com/api/download/models/15236",
+                        value="",
+                        label="HuggingFace Model ID or Civitai model "
+                        "download URL",
+                        lines=3,
                     )
                     # janky fix for overflowing text
                     upscaler_vae_info = (
@@ -291,8 +390,6 @@ with gr.Blocks(title="Upscaler") as upscaler_web:
                         if args.custom_vae
                         else "None",
                         choices=["None"] + get_custom_model_files("vae"),
-                        allow_custom_value=True,
-                        scale=1,
                     )
 
                 with gr.Group(elem_id="prompt_box_outer"):
@@ -328,7 +425,6 @@ with gr.Blocks(title="Upscaler") as upscaler_web:
                             elem_id="lora_weights",
                             value="None",
                             choices=["None"] + get_custom_model_files("lora"),
-                            allow_custom_value=True,
                         )
                         lora_hf_id = gr.Textbox(
                             elem_id="lora_hf_id",
@@ -347,7 +443,6 @@ with gr.Blocks(title="Upscaler") as upscaler_web:
                             label="Scheduler",
                             value="DDIM",
                             choices=scheduler_list_cpu_only,
-                            allow_custom_value=True,
                         )
                         with gr.Group():
                             save_metadata_to_png = gr.Checkbox(
@@ -452,8 +547,17 @@ with gr.Blocks(title="Upscaler") as upscaler_web:
                         label="Device",
                         value=available_devices[0],
                         choices=available_devices,
-                        allow_custom_value=True,
                     )
+                with gr.Row():
+                    random_seed = gr.Button("Randomize Seed")
+                    random_seed.click(
+                        lambda: -1,
+                        inputs=[],
+                        outputs=[seed],
+                        queue=False,
+                    )
+                    stop_batch = gr.Button("Stop Batch")
+                    stable_diffusion = gr.Button("Generate Image(s)")
 
             with gr.Column(scale=1, min_width=600):
                 with gr.Group():
@@ -465,26 +569,14 @@ with gr.Blocks(title="Upscaler") as upscaler_web:
                         object_fit="contain",
                     )
                     std_output = gr.Textbox(
-                        value=f"{upscaler_model_info}\n"
-                        f"Images will be saved at "
+                        value=f"Images will be saved at "
                         f"{get_generated_imgs_path()}",
-                        lines=2,
+                        lines=1,
                         elem_id="std_output",
                         show_label=False,
                     )
                     upscaler_status = gr.Textbox(visible=False)
-                with gr.Row():
-                    stable_diffusion = gr.Button("Generate Image(s)")
-                    random_seed = gr.Button("Randomize Seed")
-                    random_seed.click(
-                        lambda: -1,
-                        inputs=[],
-                        outputs=[seed],
-                        queue=False,
-                    )
-                    stop_batch = gr.Button("Stop Batch")
-                with gr.Row():
-                    blank_thing_for_row = None
+
                 with gr.Row():
                     upscaler_sendto_img2img = gr.Button(value="SendTo Img2Img")
                     upscaler_sendto_inpaint = gr.Button(value="SendTo Inpaint")
@@ -508,6 +600,7 @@ with gr.Blocks(title="Upscaler") as upscaler_web:
                 batch_size,
                 scheduler,
                 upscaler_custom_model,
+                upscaler_hf_model_id,
                 custom_vae,
                 precision,
                 device,

apps/stable_diffusion/web/ui/utils.py

@@ -25,7 +25,7 @@ class Config:
     device: str
     use_lora: str
     use_stencil: str
-    ondemand: str  # should this be expecting a bool instead?
+    ondemand: str
 
 
 custom_model_filetypes = (
@@ -170,5 +170,4 @@ def cancel_sd():
 
 
 nodlogo_loc = resource_path("logos/nod-logo.png")
-nodicon_loc = resource_path("logos/nod-icon.png")
 available_devices = get_available_devices()

apps/stable_diffusion/web/utils/app.py

@@ -1,105 +0,0 @@
-import os
-import sys
-import webview
-import webview.util
-import socket
-
-from contextlib import closing
-from multiprocessing import Process
-
-from apps.stable_diffusion.src import args
-
-
-def webview2_installed():
-    if sys.platform != "win32":
-        return False
-
-    # On windows we want to ensure we have MS webview2 available so we don't fall back
-    # to MSHTML (aka ye olde Internet Explorer) which is deprecated by pywebview, and
-    # apparently causes SHARK not to load in properly.
-
-    # Checking these registry entries is how Microsoft says to detect a webview2 installation:
-    # https://learn.microsoft.com/en-us/microsoft-edge/webview2/concepts/distribution
-    import winreg
-
-    path = r"SOFTWARE\WOW6432Node\Microsoft\EdgeUpdate\Clients\{F3017226-FE2A-4295-8BDF-00C3A9A7E4C5}"
-
-    # only way can find if a registry entry even exists is to try and open it
-    try:
-        # check for an all user install
-        with winreg.OpenKey(
-            winreg.HKEY_LOCAL_MACHINE,
-            path,
-            0,
-            winreg.KEY_QUERY_VALUE | winreg.KEY_WOW64_64KEY,
-        ) as registry_key:
-            value, type = winreg.QueryValueEx(registry_key, "pv")
-
-    # if it didn't exist, we want to continue on...
-    except WindowsError:
-        try:
-            # ...to check for a current user install
-            with winreg.OpenKey(
-                winreg.HKEY_CURRENT_USER,
-                path,
-                0,
-                winreg.KEY_QUERY_VALUE | winreg.KEY_WOW64_64KEY,
-            ) as registry_key:
-                value, type = winreg.QueryValueEx(registry_key, "pv")
-        except WindowsError:
-            value = None
-    finally:
-        return (value is not None) and value != "" and value != "0.0.0.0"
-
-
-def window(address):
-    from tkinter import Tk
-
-    window = Tk()
-
-    # get screen width and height of display and make it more reasonably
-    # sized as we aren't making it full-screen or maximized
-    width = int(window.winfo_screenwidth() * 0.81)
-    height = int(window.winfo_screenheight() * 0.91)
-    webview.create_window(
-        "SHARK AI Studio",
-        url=address,
-        width=width,
-        height=height,
-        text_select=True,
-    )
-    webview.start(private_mode=False, storage_path=os.getcwd())
-
-
-def usable_port():
-    # Make sure we can actually use the port given in args.server_port. If
-    # not ask the OS for a port and return that as our port to use.
-
-    port = args.server_port
-
-    with closing(socket.socket(socket.AF_INET, socket.SOCK_STREAM)) as sock:
-        try:
-            sock.bind(("0.0.0.0", port))
-        except OSError:
-            with closing(
-                socket.socket(socket.AF_INET, socket.SOCK_STREAM)
-            ) as sock:
-                sock.bind(("0.0.0.0", 0))
-                sock.setsockopt(socket.SOL_SOCKET, socket.SO_REUSEADDR, 1)
-                return sock.getsockname()[1]
-
-    return port
-
-
-def launch(port):
-    # setup to launch as an app if app mode has been requested and we're able
-    # to do it, answering whether we succeeded.
-    if args.ui == "app" and (sys.platform != "win32" or webview2_installed()):
-        try:
-            t = Process(target=window, args=[f"http://localhost:{port}"])
-            t.start()
-            return True
-        except webview.util.WebViewException:
-            return False
-    else:
-        return False

apps/stable_diffusion/web/utils/metadata/png_metadata.py

@@ -149,6 +149,7 @@ def import_png_metadata(
     width,
     height,
     custom_model,
+    hf_model_id,
     custom_lora,
     hf_lora_id,
     custom_vae,
@@ -174,8 +175,10 @@ def import_png_metadata(
 
         if "Model" in metadata and png_custom_model:
             custom_model = png_custom_model
-        elif "Model" in metadata and png_hf_model_id:
-            custom_model = png_hf_model_id
+            hf_model_id = ""
+        if "Model" in metadata and png_hf_model_id:
+            custom_model = "None"
+            hf_model_id = png_hf_model_id
 
         if "LoRA" in metadata and lora_custom_model:
             custom_lora = lora_custom_model
@@ -214,6 +217,7 @@ def import_png_metadata(
         width,
         height,
         custom_model,
+        hf_model_id,
         custom_lora,
         hf_lora_id,
         custom_vae,

benchmarks/tests/test_benchmark.py

@@ -129,12 +129,12 @@ pytest_benchmark_param = pytest.mark.parametrize(
         pytest.param(True, "cpu", marks=pytest.mark.skip),
         pytest.param(
             False,
-            "cuda",
+            "gpu",
             marks=pytest.mark.skipif(
-                check_device_drivers("cuda"), reason="nvidia-smi not found"
+                check_device_drivers("gpu"), reason="nvidia-smi not found"
             ),
         ),
-        pytest.param(True, "cuda", marks=pytest.mark.skip),
+        pytest.param(True, "gpu", marks=pytest.mark.skip),
         pytest.param(
             False,
             "vulkan",

build_tools/image_comparison.py

@@ -24,13 +24,13 @@ def get_image(url, local_filename):
             shutil.copyfileobj(res.raw, f)
 
 
-def compare_images(new_filename, golden_filename, upload=False):
+def compare_images(new_filename, golden_filename):
     new = np.array(Image.open(new_filename)) / 255.0
     golden = np.array(Image.open(golden_filename)) / 255.0
     diff = np.abs(new - golden)
     mean = np.mean(diff)
     if mean > 0.1:
-        if os.name != "nt" and upload == True:
+        if os.name != "nt":
             subprocess.run(
                 [
                     "gsutil",
@@ -39,7 +39,7 @@ def compare_images(new_filename, golden_filename, upload=False):
                     "gs://shark_tank/testdata/builder/",
                 ]
             )
-        raise AssertionError("new and golden not close")
+        raise SystemExit("new and golden not close")
     else:
         print("SUCCESS")
 

build_tools/populate_sharktank_ci.sh

@@ -1,6 +1,5 @@
 #!/bin/bash
 
-IMPORTER=1 BENCHMARK=1 NO_BREVITAS=1 ./setup_venv.sh
+IMPORTER=1 BENCHMARK=1 ./setup_venv.sh
 source $GITHUB_WORKSPACE/shark.venv/bin/activate
-python build_tools/stable_diffusion_testing.py --gen
 python tank/generate_sharktank.py

build_tools/stable_diffusion_testing.py

@@ -63,14 +63,7 @@ def get_inpaint_inputs():
     open("./test_images/inputs/mask.png", "wb").write(mask.content)
 
 
-def test_loop(
-    device="vulkan",
-    beta=False,
-    extra_flags=[],
-    upload_bool=True,
-    exit_on_fail=True,
-    do_gen=False,
-):
+def test_loop(device="vulkan", beta=False, extra_flags=[]):
     # Get golden values from tank
     shutil.rmtree("./test_images", ignore_errors=True)
     model_metrics = []
@@ -88,8 +81,6 @@ def test_loop(
     if beta:
         extra_flags.append("--beta_models=True")
     extra_flags.append("--no-progress_bar")
-    if do_gen:
-        extra_flags.append("--import_debug")
     to_skip = [
         "Linaqruf/anything-v3.0",
         "prompthero/openjourney",
@@ -190,14 +181,7 @@ def test_loop(
                         "./test_images/golden/" + model_name + "/*.png"
                     )
                     golden_file = glob(golden_path)[0]
-                    try:
-                        compare_images(
-                            test_file, golden_file, upload=upload_bool
-                        )
-                    except AssertionError as e:
-                        print(e)
-                        if exit_on_fail == True:
-                            raise
+                    compare_images(test_file, golden_file)
                 else:
                     print(command)
                     print("failed to generate image for this configuration")
@@ -216,9 +200,6 @@ def test_loop(
                             extra_flags.remove(
                                 "--iree_vulkan_target_triple=rdna2-unknown-windows"
                             )
-            if do_gen:
-                prepare_artifacts()
-
     with open(os.path.join(os.getcwd(), "sd_testing_metrics.csv"), "w+") as f:
         header = "model_name;device;use_tune;import_opt;Clip Inference time(ms);Average Step (ms/it);VAE Inference time(ms);total image generation(s);command\n"
         f.write(header)
@@ -237,49 +218,15 @@ def test_loop(
             f.write(";".join(output) + "\n")
 
 
-def prepare_artifacts():
-    gen_path = os.path.join(os.getcwd(), "gen_shark_tank")
-    if not os.path.isdir(gen_path):
-        os.mkdir(gen_path)
-    for dirname in os.listdir(os.getcwd()):
-        for modelname in ["clip", "unet", "vae"]:
-            if modelname in dirname and "vmfb" not in dirname:
-                if not os.path.isdir(os.path.join(gen_path, dirname)):
-                    shutil.move(os.path.join(os.getcwd(), dirname), gen_path)
-                    print(f"Moved dir: {dirname} to {gen_path}.")
-
-
 parser = argparse.ArgumentParser()
 
 parser.add_argument("-d", "--device", default="vulkan")
 parser.add_argument(
     "-b", "--beta", action=argparse.BooleanOptionalAction, default=False
 )
-parser.add_argument("-e", "--extra_args", type=str, default=None)
-parser.add_argument(
-    "-u", "--upload", action=argparse.BooleanOptionalAction, default=True
-)
-parser.add_argument(
-    "-x", "--exit_on_fail", action=argparse.BooleanOptionalAction, default=True
-)
-parser.add_argument(
-    "-g", "--gen", action=argparse.BooleanOptionalAction, default=False
-)
+
 
 if __name__ == "__main__":
     args = parser.parse_args()
     print(args)
-    extra_args = []
-    if args.extra_args:
-        for arg in args.extra_args.split(","):
-            extra_args.append(arg)
-    test_loop(
-        args.device,
-        args.beta,
-        extra_args,
-        args.upload,
-        args.exit_on_fail,
-        args.gen,
-    )
-    if args.gen:
-        prepare_artifacts()
+    test_loop(args.device, args.beta, [])

cpp/CMakeLists.txt

@@ -27,7 +27,7 @@ include(FetchContent)
 
 FetchContent_Declare(
   iree
-  GIT_REPOSITORY https://github.com/nod-ai/srt.git
+  GIT_REPOSITORY https://github.com/nod-ai/shark-runtime.git
   GIT_TAG shark 
   GIT_SUBMODULES_RECURSE OFF
   GIT_SHALLOW OFF

cpp/README.md

@@ -40,7 +40,7 @@ cmake --build build/
 *Prepare the model*
 ```bash
 wget https://storage.googleapis.com/shark_tank/latest/resnet50_tf/resnet50_tf.mlir
-iree-compile --iree-input-type=auto --iree-vm-bytecode-module-output-format=flatbuffer-binary --iree-hal-target-backends=vulkan --iree-llvmcpu-embedded-linker-path=`python3 -c 'import sysconfig; print(sysconfig.get_paths()["purelib"])'`/iree/compiler/tools/../_mlir_libs/iree-lld --mlir-print-debuginfo --mlir-print-op-on-diagnostic=false --mlir-pass-pipeline-crash-reproducer=ist/core-reproducer.mlir --iree-llvmcpu-target-cpu-features=host -iree-vulkan-target-triple=rdna2-unknown-linux  resnet50_tf.mlir -o resnet50_tf.vmfb
+iree-compile --iree-input-type=auto --iree-vm-bytecode-module-output-format=flatbuffer-binary --iree-hal-target-backends=vulkan --iree-llvmcpu-embedded-linker-path=`python3 -c 'import sysconfig; print(sysconfig.get_paths()["purelib"])'`/iree/compiler/tools/../_mlir_libs/iree-lld --mlir-print-debuginfo --mlir-print-op-on-diagnostic=false --mlir-pass-pipeline-crash-reproducer=ist/core-reproducer.mlir --iree-llvmcpu-target-cpu-features=host -iree-vulkan-target-triple=rdna2-unknown-linux --iree-stream-resource-index-bits=64 --iree-vm-target-index-bits=64 resnet50_tf.mlir -o resnet50_tf.vmfb
 ```
 *Prepare the input*
 
@@ -65,18 +65,18 @@ A tool for benchmarking other models is built and can be invoked with a command
 see `./build/vulkan_gui/iree-vulkan-gui --help` for an explanation on the function input. For example, stable diffusion unet can be tested with the following commands:
 ```bash
 wget https://storage.googleapis.com/shark_tank/quinn/stable_diff_tf/stable_diff_tf.mlir
-iree-compile --iree-input-type=auto --iree-vm-bytecode-module-output-format=flatbuffer-binary --iree-hal-target-backends=vulkan --mlir-print-debuginfo --mlir-print-op-on-diagnostic=false --iree-llvmcpu-target-cpu-features=host -iree-vulkan-target-triple=rdna2-unknown-linux  stable_diff_tf.mlir -o stable_diff_tf.vmfb
+iree-compile --iree-input-type=auto --iree-vm-bytecode-module-output-format=flatbuffer-binary --iree-hal-target-backends=vulkan --mlir-print-debuginfo --mlir-print-op-on-diagnostic=false --iree-llvmcpu-target-cpu-features=host -iree-vulkan-target-triple=rdna2-unknown-linux --iree-stream-resource-index-bits=64 --iree-vm-target-index-bits=64 stable_diff_tf.mlir -o stable_diff_tf.vmfb
 ./build/vulkan_gui/iree-vulkan-gui --module-file=stable_diff_tf.vmfb --function_input=2x4x64x64xf32 --function_input=1xf32 --function_input=2x77x768xf32
 ```
 VAE and Autoencoder are also available
 ```bash
 # VAE
 wget https://storage.googleapis.com/shark_tank/quinn/stable_diff_tf/vae_tf/vae.mlir
-iree-compile --iree-input-type=auto --iree-vm-bytecode-module-output-format=flatbuffer-binary --iree-hal-target-backends=vulkan --mlir-print-debuginfo --mlir-print-op-on-diagnostic=false --iree-llvmcpu-target-cpu-features=host -iree-vulkan-target-triple=rdna2-unknown-linux  vae.mlir -o vae.vmfb
+iree-compile --iree-input-type=auto --iree-vm-bytecode-module-output-format=flatbuffer-binary --iree-hal-target-backends=vulkan --mlir-print-debuginfo --mlir-print-op-on-diagnostic=false --iree-llvmcpu-target-cpu-features=host -iree-vulkan-target-triple=rdna2-unknown-linux --iree-stream-resource-index-bits=64 --iree-vm-target-index-bits=64 vae.mlir -o vae.vmfb
 ./build/vulkan_gui/iree-vulkan-gui --module-file=stable_diff_tf.vmfb --function_input=1x4x64x64xf32
 
 # CLIP Autoencoder
 wget https://storage.googleapis.com/shark_tank/quinn/stable_diff_tf/clip_tf/clip_autoencoder.mlir
-iree-compile --iree-input-type=auto --iree-vm-bytecode-module-output-format=flatbuffer-binary --iree-hal-target-backends=vulkan --mlir-print-debuginfo --mlir-print-op-on-diagnostic=false --iree-llvmcpu-target-cpu-features=host -iree-vulkan-target-triple=rdna2-unknown-linux  clip_autoencoder.mlir -o clip_autoencoder.vmfb
+iree-compile --iree-input-type=auto --iree-vm-bytecode-module-output-format=flatbuffer-binary --iree-hal-target-backends=vulkan --mlir-print-debuginfo --mlir-print-op-on-diagnostic=false --iree-llvmcpu-target-cpu-features=host -iree-vulkan-target-triple=rdna2-unknown-linux --iree-stream-resource-index-bits=64 --iree-vm-target-index-bits=64 clip_autoencoder.mlir -o clip_autoencoder.vmfb
 ./build/vulkan_gui/iree-vulkan-gui --module-file=stable_diff_tf.vmfb --function_input=1x77xi32 --function_input=1x77xi32
 ```

docs/shark_iree_profiling.md

@@ -55,7 +55,7 @@ The command line for compilation will start something like this, where the `-` n
 The `-o output_filename.vmfb` flag can be used to specify the location to save the compiled vmfb. Note that a dump of the
 dispatches that can be compiled + run in isolation can be generated by adding `--iree-hal-dump-executable-benchmarks-to=/some/directory`. Say, if they are in the `benchmarks` directory, the following compile/run commands would work for Vulkan on RDNA3.
 ```
-iree-compile --iree-input-type=none --iree-hal-target-backends=vulkan --iree-vulkan-target-triple=rdna3-unknown-linux  benchmarks/module_forward_dispatch_${NUM}_vulkan_spirv_fb.mlir -o benchmarks/module_forward_dispatch_${NUM}_vulkan_spirv_fb.vmfb
+iree-compile --iree-input-type=none --iree-hal-target-backends=vulkan --iree-vulkan-target-triple=rdna3-unknown-linux --iree-stream-resource-index-bits=64 --iree-vm-target-index-bits=64 benchmarks/module_forward_dispatch_${NUM}_vulkan_spirv_fb.mlir -o benchmarks/module_forward_dispatch_${NUM}_vulkan_spirv_fb.vmfb
 
 iree-benchmark-module --module=benchmarks/module_forward_dispatch_${NUM}_vulkan_spirv_fb.vmfb --function=forward --device=vulkan
 ```
@@ -63,8 +63,8 @@ Where `${NUM}` is the dispatch number that you want to benchmark/profile in isol
 
 ### Enabling Tracy for Vulkan profiling
 
-To begin profiling with Tracy, a build of IREE runtime with tracing enabled is needed. SHARK-Runtime (SRT) builds an
-instrumented version alongside the normal version nightly (.whls typically found [here](https://github.com/nod-ai/SRT/releases)), however this is only available for Linux. For Windows, tracing can be enabled by enabling a CMake flag.
+To begin profiling with Tracy, a build of IREE runtime with tracing enabled is needed. SHARK-Runtime builds an
+instrumented version alongside the normal version nightly (.whls typically found [here](https://github.com/nod-ai/SHARK-Runtime/releases)), however this is only available for Linux. For Windows, tracing can be enabled by enabling a CMake flag.
 ```
 $env:IREE_ENABLE_RUNTIME_TRACING="ON"
 ```

docs/shark_sd_koboldcpp.md

@@ -1,140 +0,0 @@
-# Overview
-
-In [1.47.2](https://github.com/LostRuins/koboldcpp/releases/tag/v1.47.2) [Koboldcpp](https://github.com/LostRuins/koboldcpp) added AUTOMATIC1111 integration for image generation. Since SHARK implements a small subset of the A1111 REST api, you can also use SHARK for this. This document gives a starting point for how to get this working.
-
-## In Action
-
-![preview](https://user-images.githubusercontent.com/121311569/280557602-bb97bad0-fdf5-4922-a2cc-4f327f2760db.jpg)
-
-## Memory considerations
-
-Since both Koboldcpp and SHARK will use VRAM on your graphic card(s) running both at the same time using the same card will impose extra limitations on the model size you can fully offload to the video card in Koboldcpp. For me, on a RX 7900 XTX on Windows with 24 GiB of VRAM, the limit was about a 13 Billion parameter model with Q5_K_M quantisation.
-
-## Performance Considerations
-
-When using SHARK for image generation, especially with Koboldcpp, you need to be aware that it is currently designed to pay a large upfront cost in time compiling and tuning the model you select, to get an optimal individual image generation time. You need to be the judge as to whether this trade-off is going to be worth it for your OS and hardware combination.
-
-It means that the first time you run a particular Stable Diffusion model for a particular combination of image size, LoRA, and VAE, SHARK will spend *many minutes* - even on a beefy machaine with very fast graphics card with lots of memory - building that model combination just so it can save it to disk. It may even have to go away and download the model if it doesn't already have it locally. Once it has done its build of a model combination for your hardware once, it shouldn't need to do it again until you upgrade to a newer SHARK version, install different drivers or change your graphics hardware. It will just upload the files it generated the first time to your graphics card and proceed from there.
-
-This does mean however, that on a brand new fresh install of SHARK that has not generated any images on a model you haven't selected before, the first image Koboldcpp requests may look like it is *never* going finish and that the whole process has broken. Be forewarned, make yourself a cup of coffee, and expect a lot of messages about compilation and tuning from SHARK in the terminal you ran it from.
-
-## Setup SHARK and prerequisites:
-
- * Make sure you have suitable drivers for your graphics card installed. See the prerequisties section of the [README](https://github.com/nod-ai/SHARK#readme).
- * Download the latest SHARK studio .exe from [here](https://github.com/nod-ai/SHARK/releases) or follow the instructions in the [README](https://github.com/nod-ai/SHARK#readme) for an advanced, Linux or Mac install.
- * Run SHARK from terminal/PowerShell with the `--api` flag. Since koboldcpp also expects both CORS support and the image generator to be running on port `7860` rather than SHARK default of `8080`, also include both the `--api_cors_origin` flag with a suitable origin (use `="*"` to enable all origins) and `--server_port=7860` on the command line. (See the if you want to run SHARK on a different port)
-
-```powershell
-## Run the .exe in API mode, with CORS support, on the A1111 endpoint port:
-.\node_ai_shark_studio_<date>_<ver>.exe --api --api_cors_origin="*"  --server_port=7860
-
-## Run trom the base directory of a source clone of SHARK on Windows:
-.\setup_venv.ps1
-python .\apps\stable_diffusion\web\index.py --api --api_cors_origin="*"  --server_port=7860
-
-## Run a the base directory of a source clone of SHARK on Linux:
-./setup_venv.sh
-source shark.venv/bin/activate
-python ./apps/stable_diffusion/web/index.py --api --api_cors_origin="*"  --server_port=7860
-
-## An example giving improved performance on AMD cards using vulkan, that runs on the same port as A1111
-.\node_ai_shark_studio_20320901_2525.exe --api --api_cors_origin="*" --device_allocator="caching" --server_port=7860
-
-## Since the api respects most applicable SHARK command line arguments for options not specified,
-## or currently unimplemented by API, there might be some you want to set, as listed in `--help`
-.\node_ai_shark_studio_20320901_2525.exe --help
-
-## For instance, the example above, but with a a custom VAE specified
-.\node_ai_shark_studio_20320901_2525.exe --api --api_cors_origin="*" --device_allocator="caching" --server_port=7860 --custom_vae="clearvae_v23.safetensors"
-
-## An example with multiple specific CORS origins
-python apps/stable_diffusion/web/index.py --api --api_cors_origin="koboldcpp.example.com:7001" --api_cors_origin="koboldcpp.example.com:7002" --server_port=7860
-```
-
-SHARK should start in server mode, and you should see something like this:
-
-![SHARK API startup](https://user-images.githubusercontent.com/121311569/280556294-c3f7fc1a-c8e2-467d-afe6-365638d6823a.png)
-
-* Note: When running in api mode with `--api`, the .exe will not function as a webUI. Thus, the address or port shown in the terminal output will only be useful for API requests.
-
-
-## Configure Koboldcpp for local image generation:
-
-* Get the latest [Koboldcpp](https://github.com/LostRuins/koboldcpp/releases) if you don't already have it. If you have a recent AMD card that has ROCm HIP [support for Windows](https://rocmdocs.amd.com/en/latest/release/windows_support.html#windows-supported-gpus) or [support for Linux](https://rocmdocs.amd.com/en/latest/release/gpu_os_support.html#linux-supported-gpus), you'll likely prefer [YellowRosecx's ROCm fork](https://github.com/YellowRoseCx/koboldcpp-rocm).
-* Start Koboldcpp in another terminal/Powershell and setup your model configuration. Refer to the [Koboldcpp README](https://github.com/YellowRoseCx/koboldcpp-rocm) for more details on how to do this if this is your first time using Koboldcpp.
-* Once the main UI has loaded into your browser click the settings button, go to the advanced tab, and then choose *Local A1111* from the generate images dropdown:
-
-  ![Settings button location](https://user-images.githubusercontent.com/121311569/280556246-10692d79-e89f-4fdf-87ba-82f3d78ed49d.png)
-
-  ![Advanced Settings with 'Local A1111' location](https://user-images.githubusercontent.com/121311569/280556234-6ebc8ba7-1469-442a-93a7-5626a094ddf1.png)
-
-  *if you get an error here, see the next section [below](#connecting-to-shark-on-a-different-address-or-port)*
-
-* A list of Stable Diffusion models available to your SHARK instance should now be listed in the box below *generate images*. The default value will usually be set to `stabilityai/stable-diffusion-2-1-base`. Choose the model you want to use for image generation from the list (but see [performance considerations](#performance-considerations)).
-* You should now be ready to generate images, either by clicking the 'Add Img' button above the text entry box:
-
-  ![Add Image Button](https://user-images.githubusercontent.com/121311569/280556161-846c7883-4a83-4458-a56a-bd9f93ca354c.png)
-
-  ...or by selecting the 'Autogenerate' option in the settings:
-
-  ![Setting the autogenerate images option](https://user-images.githubusercontent.com/121311569/280556230-ae221a46-ba68-499b-a519-c8f290bbbeae.png)
-
-  *I often find that even if I have selected autogenerate I have to do an 'add img' to get things started off*
-
-* There is one final piece of image generation configuration within Koboldcpp you might want to do. This is also in the generate images section of advanced settings. Here there is, not very obviously, a 'style' button:
-
-  ![Selecting the 'styles' button](https://user-images.githubusercontent.com/121311569/280556694-55cd1c55-a059-4b54-9293-63d66a32368e.png)
-
-  This will bring up a dialog box where you can enter a short text that will sent as a prefix to the Prompt sent to SHARK:
-
-  ![Entering extra image styles](https://user-images.githubusercontent.com/121311569/280556172-4aab9794-7a77-46d7-bdda-43df570ad19a.png)
-
-
-## Connecting to SHARK on a different address or port
-
-If you didn't set the port to `--server_port=7860` when starting SHARK, or you are running it on different machine on your network than you are running Koboldcpp, or to where you are running the koboldcpp's kdlite client frontend, then you very likely got the following error:
-
-  ![Can't find the A1111 endpoint error](https://user-images.githubusercontent.com/121311569/280555857-601f53dc-35e9-4027-9180-baa61d2393ba.png)
-
-As long as SHARK is running correctly, this means you need to set the url and port to the correct values in Koboldcpp. For instance. to set the port that Koboldcpp looks for an image generator to SHARK's default port of 8080:
-
-* Select the cog icon the Generate Images section of Advanced settings:
-
-     ![Selecting the endpoint cog](https://user-images.githubusercontent.com/121311569/280555866-4287ecc5-f29f-4c03-8f5a-abeaf31b0442.png)
-
-* Then edit the port number at the end of the url in the 'A1111 Endpoint Selection' dialog box to read 8080:
-
-     ![Changing the endpoint port](https://user-images.githubusercontent.com/121311569/280556170-f8848b7b-6fc9-4cf7-80eb-5c312f332fd9.png)
-
-* Similarly, when running SHARK on a different machine you will need to change host part of the endpoint url to the hostname or ip address where SHARK is running, similarly:
-
-    ![Changing the endpoint hostname](https://user-images.githubusercontent.com/121311569/280556167-c6541dea-0f85-417a-b661-fdf4dc40d05f.png)
-
-## Examples
-
-Here's how Koboldcpp shows an image being requested:
-
-  ![An image being generated]((https://user-images.githubusercontent.com/121311569/280556210-bb1c9efd-79ac-478e-b726-b25b82ef2186.png)
-
-The generated image in context in story mode:
-
- ![A generated image](https://user-images.githubusercontent.com/121311569/280556179-4e9f3752-f349-4cba-bc6a-f85f8dc79b10.jpg)
-
-And the same image when clicked on:
-
- ![A selected image](https://user-images.githubusercontent.com/121311569/280556216-2ca4c0a4-3889-4ef5-8a09-30084fb34081.jpg)
-
-
-## Where to find the images in SHARK
-
-Even though Koboldcpp requests images at a size of 512x512, it resizes then to 256x256, converts them to `.jpeg`, and only shows them at 200x200 in the main text window. It does this so it can save them compactly embedded in your story as a `data://` uri.
-
-However the images at the original size are saved by SHARK in its `output_dir` which is usually a folder named for the current date. inside `generated_imgs` folder in the SHARK installation directory.
-
-You can browse these, either using the Output Gallery tab from within the SHARK web ui:
-
-  ![SHARK web ui output gallery tab](https://user-images.githubusercontent.com/121311569/280556582-9303ca85-2594-4a8c-97a2-fbd72337980b.jpg)
-
-...or by browsing to the `output_dir` in your operating system's file manager:
-
-  ![SHARK output directory subfolder in Windows File Explorer](https://user-images.githubusercontent.com/121311569/280556297-66173030-2324-415c-a236-ef3fcd73e6ed.jpg)

inference/CMakeLists.txt

@@ -0,0 +1,192 @@
+# Copyright 2021, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions
+# are met:
+#  * Redistributions of source code must retain the above copyright
+#    notice, this list of conditions and the following disclaimer.
+#  * Redistributions in binary form must reproduce the above copyright
+#    notice, this list of conditions and the following disclaimer in the
+#    documentation and/or other materials provided with the distribution.
+#  * Neither the name of NVIDIA CORPORATION nor the names of its
+#    contributors may be used to endorse or promote products derived
+#    from this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+# PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+cmake_minimum_required(VERSION 3.17)
+
+project(sharkbackend LANGUAGES C CXX)
+
+#
+# Options
+#
+
+option(TRITON_ENABLE_GPU "Enable GPU support in backend" ON)
+option(TRITON_ENABLE_STATS "Include statistics collections in backend" ON)
+
+set(TRITON_COMMON_REPO_TAG "main" CACHE STRING "Tag for triton-inference-server/common repo")
+set(TRITON_CORE_REPO_TAG "main" CACHE STRING "Tag for triton-inference-server/core repo")
+set(TRITON_BACKEND_REPO_TAG "main" CACHE STRING "Tag for triton-inference-server/backend repo")
+
+if(NOT CMAKE_BUILD_TYPE)
+  set(CMAKE_BUILD_TYPE Release)
+endif()
+
+#
+# Dependencies
+#
+# FetchContent requires us to include the transitive closure of all
+# repos that we depend on so that we can override the tags.
+#
+include(FetchContent)
+
+FetchContent_Declare(
+  repo-common
+  GIT_REPOSITORY https://github.com/triton-inference-server/common.git
+  GIT_TAG ${TRITON_COMMON_REPO_TAG}
+  GIT_SHALLOW ON
+)
+FetchContent_Declare(
+  repo-core
+  GIT_REPOSITORY https://github.com/triton-inference-server/core.git
+  GIT_TAG ${TRITON_CORE_REPO_TAG}
+  GIT_SHALLOW ON
+)
+FetchContent_Declare(
+  repo-backend
+  GIT_REPOSITORY https://github.com/triton-inference-server/backend.git
+  GIT_TAG ${TRITON_BACKEND_REPO_TAG}
+  GIT_SHALLOW ON
+)
+FetchContent_MakeAvailable(repo-common repo-core repo-backend)
+
+#
+# The backend must be built into a shared library. Use an ldscript to
+# hide all symbols except for the TRITONBACKEND API.
+#
+configure_file(src/libtriton_dshark.ldscript libtriton_dshark.ldscript COPYONLY)
+
+add_library(
+  triton-dshark-backend SHARED
+  src/dshark.cc
+  #src/dshark_driver_module.c
+)
+
+add_library(
+  SharkBackend::triton-dshark-backend ALIAS triton-dshark-backend
+)
+
+target_include_directories(
+  triton-dshark-backend
+  PRIVATE
+    ${CMAKE_CURRENT_SOURCE_DIR}/src
+)
+
+list(APPEND CMAKE_MODULE_PATH "${PROJECT_BINARY_DIR}/lib/cmake/mlir")
+
+add_subdirectory(thirdparty/shark-runtime EXCLUDE_FROM_ALL)
+
+target_link_libraries(triton-dshark-backend PRIVATE iree_base_base
+  iree_hal_hal
+  iree_hal_cuda_cuda
+  iree_hal_cuda_registration_registration
+  iree_hal_vmvx_registration_registration
+  iree_hal_dylib_registration_registration
+  iree_modules_hal_hal
+  iree_vm_vm
+  iree_vm_bytecode_module
+  iree_hal_local_loaders_system_library_loader
+  iree_hal_local_loaders_vmvx_module_loader
+  )
+
+target_compile_features(triton-dshark-backend PRIVATE cxx_std_11)
+
+
+target_link_libraries(
+  triton-dshark-backend
+  PRIVATE
+    triton-core-serverapi   # from repo-core
+    triton-core-backendapi  # from repo-core
+    triton-core-serverstub  # from repo-core
+    triton-backend-utils    # from repo-backend
+)
+
+if(WIN32)
+  set_target_properties(
+    triton-dshark-backend PROPERTIES
+    POSITION_INDEPENDENT_CODE ON
+    OUTPUT_NAME triton_dshark
+  )
+else()
+  set_target_properties(
+    triton-dshark-backend PROPERTIES
+    POSITION_INDEPENDENT_CODE ON
+    OUTPUT_NAME triton_dshark
+    LINK_DEPENDS ${CMAKE_CURRENT_BINARY_DIR}/libtriton_dshark.ldscript
+    LINK_FLAGS "-Wl,--version-script libtriton_dshark.ldscript"
+  )
+endif()
+
+
+
+#
+# Install
+#
+include(GNUInstallDirs)
+set(INSTALL_CONFIGDIR ${CMAKE_INSTALL_LIBDIR}/cmake/SharkBackend)
+
+install(
+  TARGETS
+    triton-dshark-backend
+  EXPORT
+    triton-dshark-backend-targets
+  LIBRARY DESTINATION ${CMAKE_INSTALL_PREFIX}/backends/dshark
+  RUNTIME DESTINATION ${CMAKE_INSTALL_PREFIX}/backends/dshark
+)
+
+install(
+  EXPORT
+    triton-dshark-backend-targets
+  FILE
+    SharkBackendTargets.cmake
+  NAMESPACE
+    SharkBackend::
+  DESTINATION
+    ${INSTALL_CONFIGDIR}
+)
+
+include(CMakePackageConfigHelpers)
+configure_package_config_file(
+  ${CMAKE_CURRENT_LIST_DIR}/cmake/SharkBackendConfig.cmake.in
+  ${CMAKE_CURRENT_BINARY_DIR}/SharkBackendConfig.cmake
+  INSTALL_DESTINATION ${INSTALL_CONFIGDIR}
+)
+
+install(
+  FILES
+  ${CMAKE_CURRENT_BINARY_DIR}/SharkBackendConfig.cmake
+  DESTINATION ${INSTALL_CONFIGDIR}
+)
+
+#
+# Export from build tree
+#
+export(
+  EXPORT triton-dshark-backend-targets
+  FILE ${CMAKE_CURRENT_BINARY_DIR}/SharkBackendTargets.cmake
+  NAMESPACE SharkBackend::
+)
+
+export(PACKAGE SharkBackend)
+

inference/README.md

@@ -0,0 +1,100 @@
+# SHARK Triton Backend
+
+The triton backend for shark.
+
+# Build
+
+Install SHARK
+
+```
+git clone https://github.com/nod-ai/SHARK.git
+# skip above step if dshark is already installed
+cd SHARK/inference
+```
+
+install dependancies
+
+```
+apt-get install patchelf rapidjson-dev python3-dev
+git submodule update --init
+```
+
+update the submodules of iree
+
+```
+cd thirdparty/shark-runtime
+git submodule update --init
+```
+
+Next, make the backend and install it
+
+```
+cd ../..
+mkdir build && cd build
+cmake -DTRITON_ENABLE_GPU=ON \
+-DIREE_HAL_DRIVER_CUDA=ON \
+-DIREE_TARGET_BACKEND_CUDA=ON \
+-DMLIR_ENABLE_CUDA_RUNNER=ON \
+-DCMAKE_INSTALL_PREFIX:PATH=`pwd`/install \
+-DTRITON_BACKEND_REPO_TAG=r22.02 \
+-DTRITON_CORE_REPO_TAG=r22.02 \
+-DTRITON_COMMON_REPO_TAG=r22.02 ..
+make install
+```
+
+# Incorporating into Triton
+
+There are much more in depth explenations for the following steps in triton's documentation:
+https://github.com/triton-inference-server/server/blob/main/docs/compose.md#triton-with-unsupported-and-custom-backends
+
+There should be a file at /build/install/backends/dshark/libtriton_dshark.so.  You will need to copy it into your triton server image.  
+More documentation is in the link above, but to create the docker image, you need to run the compose.py command in the triton-backend server repo
+
+
+To first build your image, clone the tritonserver repo.
+
+```
+git clone https://github.com/triton-inference-server/server.git
+```
+
+then run `compose.py` to build a docker compose file 
+```
+cd server
+python3 compose.py --repoagent checksum --dry-run
+```
+
+Because dshark is a third party backend, you will need to manually modify the `Dockerfile.compose` to include the dshark backend.  To do this, in the Dockerfile.compose file produced, copy this line.
+the dshark backend will be located in the build folder from earlier under `/build/install/backends`
+
+```
+COPY /path/to/build/install/backends/dshark /opt/tritonserver/backends/dshark
+```
+
+Next run 
+```
+docker build -t tritonserver_custom -f Dockerfile.compose .
+docker run -it --gpus=1 --net=host -v/path/to/model_repos:/models  tritonserver_custom:latest tritonserver --model-repository=/models
+```
+
+where `path/to/model_repos` is where you are storing the models you want to run
+
+if your not using gpus, omit `--gpus=1`
+
+```
+docker run -it  --net=host -v/path/to/model_repos:/models  tritonserver_custom:latest tritonserver --model-repository=/models
+```
+
+# Setting up a model
+
+to include a model in your backend, add a directory with your model name to your model repository directory.  examples of models can be seen here: https://github.com/triton-inference-server/backend/tree/main/examples/model_repos/minimal_models
+
+make sure to adjust the input correctly in the config.pbtxt file, and save a vmfb file under 1/model.vmfb
+
+# CUDA
+
+if you're having issues with cuda, make sure your correct drivers are installed, and that `nvidia-smi` works, and also make sure that the nvcc compiler is on the path.
+
+
+
+
+

inference/cmake/SharkBackendConfig.cmake.in

@@ -0,0 +1,39 @@
+# Copyright 2021, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions
+# are met:
+#  * Redistributions of source code must retain the above copyright
+#    notice, this list of conditions and the following disclaimer.
+#  * Redistributions in binary form must reproduce the above copyright
+#    notice, this list of conditions and the following disclaimer in the
+#    documentation and/or other materials provided with the distribution.
+#  * Neither the name of NVIDIA CORPORATION nor the names of its
+#    contributors may be used to endorse or promote products derived
+#    from this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+# PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+include(CMakeFindDependencyMacro)
+
+get_filename_component(
+  SHARKBACKEND_CMAKE_DIR "${CMAKE_CURRENT_LIST_FILE}" PATH
+)
+
+list(APPEND CMAKE_MODULE_PATH ${SHARKBACKEND_CMAKE_DIR})
+
+if(NOT TARGET SharkBackend::triton-dshark-backend)
+  include("${SHARKBACKEND_CMAKE_DIR}/SharkBackendTargets.cmake")
+endif()
+
+set(SHARKBACKEND_LIBRARIES SharkBackend::triton-dshark-backend)

inference/src/libtriton_dshark.ldscript

@@ -0,0 +1,30 @@
+# Copyright 2021, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions
+# are met:
+#  * Redistributions of source code must retain the above copyright
+#    notice, this list of conditions and the following disclaimer.
+#  * Redistributions in binary form must reproduce the above copyright
+#    notice, this list of conditions and the following disclaimer in the
+#    documentation and/or other materials provided with the distribution.
+#  * Neither the name of NVIDIA CORPORATION nor the names of its
+#    contributors may be used to endorse or promote products derived
+#    from this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
+# EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
+# PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE COPYRIGHT OWNER OR
+# CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
+# EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
+# PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
+# PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
+# OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+{
+  global:
+    TRITONBACKEND_*;
+  local: *;
+};

process_skipfiles.py

@@ -6,15 +6,15 @@ from distutils.sysconfig import get_python_lib
 import fileinput
 from pathlib import Path
 
-# Temporary workaround for transformers/__init__.py.
-path_to_transformers_hook = Path(
+# Temorary workaround for transformers/__init__.py.
+path_to_tranformers_hook = Path(
     get_python_lib()
     + "/_pyinstaller_hooks_contrib/hooks/stdhooks/hook-transformers.py"
 )
-if path_to_transformers_hook.is_file():
+if path_to_tranformers_hook.is_file():
     pass
 else:
-    with open(path_to_transformers_hook, "w") as f:
+    with open(path_to_tranformers_hook, "w") as f:
         f.write("module_collection_mode = 'pyz+py'")
 
 path_to_skipfiles = Path(get_python_lib() + "/torch/_dynamo/skipfiles.py")

pyproject.toml

@@ -5,7 +5,7 @@ requires = [
     "packaging",
 
     "numpy>=1.22.4",
-    "torch-mlir>=20230620.875",
+    "torch-mlir>=20221021.633",
     "iree-compiler>=20221022.190",
     "iree-runtime>=20221022.190",
 ]

requirements-importer-macos.txt

@@ -8,8 +8,19 @@ torchvision
 tqdm
 
 #iree-compiler  | iree-runtime should already be installed
+#these dont work ok osx
+#iree-tools-tflite
+#iree-tools-xla
+#iree-tools-tf
 
+# TensorFlow and JAX.
+gin-config
+tensorflow-macos
+tensorflow-metal
+#tf-models-nightly
+#tensorflow-text-nightly
 transformers
+tensorflow-probability
 #jax[cpu]
 
 # tflitehub dependencies.

requirements-importer.txt

@@ -3,19 +3,29 @@
 
 numpy>1.22.4
 pytorch-triton
-torchvision 
+torchvision==0.16.0.dev20230322 
 tabulate
 
 tqdm
 
 #iree-compiler  | iree-runtime should already be installed
+iree-tools-tflite
 iree-tools-xla
+iree-tools-tf
 
-# Modelling and JAX.
+# TensorFlow and JAX.
 gin-config
+tensorflow>2.11
+keras
+#tf-models-nightly
+#tensorflow-text-nightly
 transformers
 diffusers
+#tensorflow-probability
 #jax[cpu]
+
+
+# tflitehub dependencies.
 Pillow
 
 # Testing and support.

requirements.txt

@@ -1,6 +1,3 @@
--f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html
---pre
-
 setuptools
 wheel
 
@@ -17,20 +14,17 @@ pytest-forked
 Pillow
 parameterized
 
-#shark-turbine @ git+https://github.com/nod-ai/SHARK-Turbine.git@main
 # Add transformers, diffusers and scipy since it most commonly used
-tokenizers==0.13.3
 transformers
 diffusers
 #accelerate is now required for diffusers import from ckpt.
 accelerate
 scipy
 ftfy
-gradio==3.44.3
+gradio
 altair
 omegaconf
-# 0.3.2 doesn't have binaries for arm64
-safetensors==0.3.1
+safetensors
 opencv-python
 scikit-image
 pytorch_lightning # for runwayml models
@@ -41,13 +35,10 @@ py-cpuinfo
 tiktoken # for codegen
 joblib # for langchain
 timm # for MiniGPT4
-langchain
-einops # for zoedepth
-pydantic==2.4.1 # pin until pyinstaller-hooks-contrib works with beta versions
 
 # Keep PyInstaller at the end. Sometimes Windows Defender flags it but most folks can continue even if it errors
 pefile
 pyinstaller
 
 # vicuna quantization
-brevitas @ git+https://github.com/Xilinx/brevitas.git@56edf56a3115d5ac04f19837b388fd7d3b1ff7ea
+brevitas @ git+https://github.com/Xilinx/brevitas.git@dev

rest_api_tests/api_test.py

@@ -4,7 +4,7 @@ import base64
 from io import BytesIO
 
 
-def upscaler_test(verbose=False):
+def upscaler_test():
     # Define values here
     prompt = ""
     negative_prompt = ""
@@ -44,17 +44,10 @@ def upscaler_test(verbose=False):
 
     res = requests.post(url=url, json=data, headers=headers, timeout=1000)
 
-    print(
-        f"[upscaler] response from server was : {res.status_code} {res.reason}"
-    )
-
-    if verbose or res.status_code != 200:
-        print(
-            f"\n{res.json()['info'] if res.status_code == 200 else res.content}\n"
-        )
+    print(f"response from server was : {res.status_code}")
 
 
-def img2img_test(verbose=False):
+def img2img_test():
     # Define values here
     prompt = "Paint a rabbit riding on the dog"
     negative_prompt = "ugly, bad art, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, body out of frame, blurry, bad anatomy, blurred, watermark, grainy, tiling, signature, cut off, draft"
@@ -94,16 +87,7 @@ def img2img_test(verbose=False):
 
     res = requests.post(url=url, json=data, headers=headers, timeout=1000)
 
-    res = requests.post(url=url, json=data, headers=headers, timeout=1000)
-
-    print(
-        f"[img2img] response from server was : {res.status_code} {res.reason}"
-    )
-
-    if verbose or res.status_code != 200:
-        print(
-            f"\n{res.json()['info'] if res.status_code == 200 else res.content}\n"
-        )
+    print(f"response from server was : {res.status_code}")
 
     # NOTE Uncomment below to save the picture
 
@@ -119,7 +103,7 @@ def img2img_test(verbose=False):
     # response_img.save(r"rest_api_tests/response_img.png")
 
 
-def inpainting_test(verbose=False):
+def inpainting_test():
     prompt = "Paint a rabbit riding on the dog"
     negative_prompt = "ugly, bad art, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, body out of frame, blurry, bad anatomy, blurred, watermark, grainy, tiling, signature, cut off, draft"
     seed = 2121991605
@@ -166,17 +150,10 @@ def inpainting_test(verbose=False):
 
     res = requests.post(url=url, json=data, headers=headers, timeout=1000)
 
-    print(
-        f"[inpaint] response from server was : {res.status_code} {res.reason}"
-    )
-
-    if verbose or res.status_code != 200:
-        print(
-            f"\n{res.json()['info'] if res.status_code == 200 else res.content}\n"
-        )
+    print(f"[Inpainting] response from server was : {res.status_code}")
 
 
-def outpainting_test(verbose=False):
+def outpainting_test():
     prompt = "Paint a rabbit riding on the dog"
     negative_prompt = "ugly, bad art, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, body out of frame, blurry, bad anatomy, blurred, watermark, grainy, tiling, signature, cut off, draft"
     seed = 2121991605
@@ -223,17 +200,10 @@ def outpainting_test(verbose=False):
 
     res = requests.post(url=url, json=data, headers=headers, timeout=1000)
 
-    print(
-        f"[outpaint] response from server was : {res.status_code} {res.reason}"
-    )
-
-    if verbose or res.status_code != 200:
-        print(
-            f"\n{res.json()['info'] if res.status_code == 200 else res.content}\n"
-        )
+    print(f"[Outpaint] response from server was : {res.status_code}")
 
 
-def txt2img_test(verbose=False):
+def txt2img_test():
     prompt = "Paint a rabbit in a top hate"
     negative_prompt = "ugly, bad art, poorly drawn hands, poorly drawn feet, poorly drawn face, out of frame, extra limbs, disfigured, deformed, body out of frame, blurry, bad anatomy, blurred, watermark, grainy, tiling, signature, cut off, draft"
     seed = 2121991605
@@ -262,119 +232,12 @@ def txt2img_test(verbose=False):
 
     res = requests.post(url=url, json=data, headers=headers, timeout=1000)
 
-    print(
-        f"[txt2img] response from server was : {res.status_code} {res.reason}"
-    )
-
-    if verbose or res.status_code != 200:
-        print(
-            f"\n{res.json()['info'] if res.status_code == 200 else res.content}\n"
-        )
-
-
-def sd_models_test(verbose=False):
-    url = "http://127.0.0.1:8080/sdapi/v1/sd-models"
-
-    headers = {
-        "User-Agent": "PythonTest",
-        "Accept": "*/*",
-        "Accept-Encoding": "gzip, deflate, br",
-    }
-
-    res = requests.get(url=url, headers=headers, timeout=1000)
-
-    print(
-        f"[sd_models] response from server was : {res.status_code} {res.reason}"
-    )
-
-    if verbose or res.status_code != 200:
-        print(f"\n{res.json() if res.status_code == 200 else res.content}\n")
-
-
-def sd_samplers_test(verbose=False):
-    url = "http://127.0.0.1:8080/sdapi/v1/samplers"
-
-    headers = {
-        "User-Agent": "PythonTest",
-        "Accept": "*/*",
-        "Accept-Encoding": "gzip, deflate, br",
-    }
-
-    res = requests.get(url=url, headers=headers, timeout=1000)
-
-    print(
-        f"[sd_samplers] response from server was : {res.status_code} {res.reason}"
-    )
-
-    if verbose or res.status_code != 200:
-        print(f"\n{res.json() if res.status_code == 200 else res.content}\n")
-
-
-def options_test(verbose=False):
-    url = "http://127.0.0.1:8080/sdapi/v1/options"
-
-    headers = {
-        "User-Agent": "PythonTest",
-        "Accept": "*/*",
-        "Accept-Encoding": "gzip, deflate, br",
-    }
-
-    res = requests.get(url=url, headers=headers, timeout=1000)
-
-    print(
-        f"[options] response from server was : {res.status_code} {res.reason}"
-    )
-
-    if verbose or res.status_code != 200:
-        print(f"\n{res.json() if res.status_code == 200 else res.content}\n")
-
-
-def cmd_flags_test(verbose=False):
-    url = "http://127.0.0.1:8080/sdapi/v1/cmd-flags"
-
-    headers = {
-        "User-Agent": "PythonTest",
-        "Accept": "*/*",
-        "Accept-Encoding": "gzip, deflate, br",
-    }
-
-    res = requests.get(url=url, headers=headers, timeout=1000)
-
-    print(
-        f"[cmd-flags] response from server was : {res.status_code} {res.reason}"
-    )
-
-    if verbose or res.status_code != 200:
-        print(f"\n{res.json() if res.status_code == 200 else res.content}\n")
+    print(f"[txt2img] response from server was : {res.status_code}")
 
 
 if __name__ == "__main__":
-    import argparse
-
-    parser = argparse.ArgumentParser(
-        description=(
-            "Exercises the Stable Diffusion REST API of Shark. Make sure "
-            "Shark is running in API mode on 127.0.0.1:8080 before running"
-            "this script."
-        ),
-    )
-    parser.add_argument(
-        "-v",
-        "--verbose",
-        action="store_true",
-        help=(
-            "also display selected info from the JSON response for "
-            "successful requests"
-        ),
-    )
-    args = parser.parse_args()
-
-    sd_models_test(args.verbose)
-    sd_samplers_test(args.verbose)
-    options_test(args.verbose)
-    cmd_flags_test(args.verbose)
-    txt2img_test(args.verbose)
-    img2img_test(args.verbose)
-    upscaler_test(args.verbose)
-    inpainting_test(args.verbose)
-    outpainting_test(args.verbose)
+    txt2img_test()
+    img2img_test()
+    upscaler_test()
+    inpainting_test()
+    outpainting_test()

setup_venv.ps1

@@ -90,8 +90,8 @@ python -m pip install --upgrade pip
 pip install wheel
 pip install -r requirements.txt
 pip install --pre torch-mlir torch --extra-index-url https://download.pytorch.org/whl/nightly/cpu -f https://llvm.github.io/torch-mlir/package-index/
-pip install --upgrade -f https://nod-ai.github.io/SRT/pip-release-links.html iree-compiler iree-runtime
+pip install --upgrade -f https://nod-ai.github.io/SHARK-Runtime/pip-release-links.html iree-compiler iree-runtime
 Write-Host "Building SHARK..."
-pip install -e . -f https://llvm.github.io/torch-mlir/package-index/ -f https://nod-ai.github.io/SRT/pip-release-links.html
+pip install -e . -f https://llvm.github.io/torch-mlir/package-index/ -f https://nod-ai.github.io/SHARK-Runtime/pip-release-links.html
 Write-Host "Build and installation completed successfully"
 Write-Host "Source your venv with ./shark.venv/Scripts/activate"

setup_venv.sh

@@ -86,7 +86,6 @@ $PYTHON -m pip install --upgrade -r "$TD/requirements.txt"
 if [ "$torch_mlir_bin" = true ]; then
   if [[ $(uname -s) = 'Darwin' ]]; then
     echo "MacOS detected. Installing torch-mlir from .whl, to avoid dependency problems with torch."
-    $PYTHON -m pip uninstall -y timm #TEMP FIX FOR MAC
     $PYTHON -m pip install --pre --no-cache-dir torch-mlir -f https://llvm.github.io/torch-mlir/package-index/ -f https://download.pytorch.org/whl/nightly/torch/
   else
     $PYTHON -m pip install --pre torch-mlir -f https://llvm.github.io/torch-mlir/package-index/
@@ -104,7 +103,7 @@ else
 fi
 if [[ -z "${USE_IREE}" ]]; then
   rm .use-iree
-  RUNTIME="https://nod-ai.github.io/SRT/pip-release-links.html"
+  RUNTIME="https://nod-ai.github.io/SHARK-Runtime/pip-release-links.html"
 else
   touch ./.use-iree
   RUNTIME="https://openxla.github.io/iree/pip-release-links.html"
@@ -129,21 +128,16 @@ if [[ ! -z "${IMPORTER}" ]]; then
   fi
 fi
 
-if [[ $(uname -s) = 'Darwin' ]]; then
-  PYTORCH_URL=https://download.pytorch.org/whl/nightly/torch/
-else
-  PYTORCH_URL=https://download.pytorch.org/whl/nightly/cpu/
-fi
+$PYTHON -m pip install --no-warn-conflicts -e . -f https://llvm.github.io/torch-mlir/package-index/ -f ${RUNTIME} -f https://download.pytorch.org/whl/nightly/torch/
 
-$PYTHON -m pip install --no-warn-conflicts -e . -f https://llvm.github.io/torch-mlir/package-index/ -f ${RUNTIME} -f ${PYTORCH_URL}
-
-if [[ $(uname -s) = 'Linux' && ! -z "${IMPORTER}" ]]; then
+if [[ $(uname -s) = 'Linux' && ! -z "${BENCHMARK}" ]]; then
   T_VER=$($PYTHON -m pip show torch | grep Version)
-  T_VER_MIN=${T_VER:14:12}
+  TORCH_VERSION=${T_VER:9:17}
   TV_VER=$($PYTHON -m pip show torchvision | grep Version)
-  TV_VER_MAJ=${TV_VER:9:6}
-  $PYTHON -m pip uninstall -y torchvision
-  $PYTHON -m pip install torchvision==${TV_VER_MAJ}${T_VER_MIN} --no-deps -f https://download.pytorch.org/whl/nightly/cpu/torchvision/
+  TV_VERSION=${TV_VER:9:18}
+  $PYTHON -m pip uninstall -y torch torchvision
+  $PYTHON -m pip install -U --pre --no-warn-conflicts triton
+  $PYTHON -m pip install --no-deps https://download.pytorch.org/whl/nightly/cu118/torch-${TORCH_VERSION}%2Bcu118-cp311-cp311-linux_x86_64.whl https://download.pytorch.org/whl/nightly/cu118/torchvision-${TV_VERSION}%2Bcu118-cp311-cp311-linux_x86_64.whl
   if [ $? -eq 0 ];then
     echo "Successfully Installed torch + cu118."
   else
@@ -151,8 +145,14 @@ if [[ $(uname -s) = 'Linux' && ! -z "${IMPORTER}" ]]; then
   fi
 fi
 
-if [[ -z "${NO_BREVITAS}" ]]; then
-  $PYTHON -m pip install git+https://github.com/Xilinx/brevitas.git@dev
+if [[ ! -z "${ONNX}" ]]; then
+  echo "${Yellow}Installing ONNX and onnxruntime for benchmarks..."
+  $PYTHON -m pip install onnx onnxruntime psutil
+  if [ $? -eq 0 ];then
+    echo "Successfully installed ONNX and ONNX runtime."
+  else
+    echo "Could not install ONNX." >&2
+  fi
 fi
 
 if [[ -z "${CONDA_PREFIX}" && "$SKIP_VENV" != "1" ]]; then

shark/examples/shark_inference/ESRGAN/esrgan.py

@@ -177,7 +177,7 @@ def compile_through_fx(model, inputs, mlir_loc=None):
     mlir_model = str(module)
     func_name = "forward"
     shark_module = SharkInference(
-        mlir_model, device=args.device, mlir_dialect="linalg"
+        mlir_model, func_name, device=args.device, mlir_dialect="linalg"
     )
     shark_module.compile()
 

shark/examples/shark_inference/albert_maskfill_pt.py

@@ -43,7 +43,9 @@ if __name__ == "__main__":
     minilm_mlir, func_name = mlir_importer.import_mlir(
         is_dynamic=False, tracing_required=True
     )
-    shark_module = SharkInference(minilm_mlir)
+    shark_module = SharkInference(
+        minilm_mlir, func_name, mlir_dialect="linalg"
+    )
     shark_module.compile()
     token_logits = torch.tensor(shark_module.forward(inputs))
     mask_id = torch.where(

shark/examples/shark_inference/albert_maskfill_tf.py

@@ -54,7 +54,7 @@ if __name__ == "__main__":
     minilm_mlir, func_name = mlir_importer.import_mlir(
         is_dynamic=False, tracing_required=False
     )
-    shark_module = SharkInference(minilm_mlir, mlir_dialect="mhlo")
+    shark_module = SharkInference(minilm_mlir, func_name, mlir_dialect="mhlo")
     shark_module.compile()
     output_idx = 0
     data_idx = 1

shark/examples/shark_inference/bloom_tank.py

@@ -6,7 +6,7 @@ mlir_model, func_name, inputs, golden_out = download_model(
 )
 
 shark_module = SharkInference(
-    mlir_model, device="cpu", mlir_dialect="tm_tensor"
+    mlir_model, func_name, device="cpu", mlir_dialect="tm_tensor"
 )
 shark_module.compile()
 result = shark_module.forward(inputs)

shark/examples/shark_inference/mhlo_example.py

@@ -13,7 +13,9 @@ arg0 = np.ones((1, 4)).astype(np.float32)
 arg1 = np.ones((4, 1)).astype(np.float32)
 
 print("Running shark on cpu backend")
-shark_module = SharkInference(mhlo_ir, device="cpu", mlir_dialect="mhlo")
+shark_module = SharkInference(
+    mhlo_ir, function_name="forward", device="cpu", mlir_dialect="mhlo"
+)
 
 # Generate the random inputs and feed into the graph.
 x = shark_module.generate_random_inputs()
@@ -21,11 +23,15 @@ shark_module.compile()
 print(shark_module.forward(x))
 
 print("Running shark on cuda backend")
-shark_module = SharkInference(mhlo_ir, device="cuda", mlir_dialect="mhlo")
+shark_module = SharkInference(
+    mhlo_ir, function_name="forward", device="cuda", mlir_dialect="mhlo"
+)
 shark_module.compile()
 print(shark_module.forward(x))
 
 print("Running shark on vulkan backend")
-shark_module = SharkInference(mhlo_ir, device="vulkan", mlir_dialect="mhlo")
+shark_module = SharkInference(
+    mhlo_ir, function_name="forward", device="vulkan", mlir_dialect="mhlo"
+)
 shark_module.compile()
 print(shark_module.forward(x))

shark/examples/shark_inference/minilm_jit.py

@@ -8,7 +8,9 @@ mlir_model, func_name, inputs, golden_out = download_model(
 )
 
 
-shark_module = SharkInference(mlir_model, device="cpu", mlir_dialect="linalg")
+shark_module = SharkInference(
+    mlir_model, func_name, device="cpu", mlir_dialect="linalg"
+)
 shark_module.compile()
 result = shark_module.forward(inputs)
 print("The obtained result via shark is: ", result)

shark/examples/shark_inference/resnest.py

@@ -33,7 +33,7 @@ mlir_importer = SharkImporter(
 
 print(golden_out)
 
-shark_module = SharkInference(vision_mlir, mlir_dialect="linalg")
+shark_module = SharkInference(vision_mlir, func_name, mlir_dialect="linalg")
 shark_module.compile()
 result = shark_module.forward((input,))
 print("Obtained result", result)

shark/examples/shark_inference/resnet50_fp16.py

@@ -49,7 +49,9 @@ module = torch_mlir.compile(
 mlir_model = module
 func_name = "forward"
 
-shark_module = SharkInference(mlir_model, device="cuda", mlir_dialect="linalg")
+shark_module = SharkInference(
+    mlir_model, func_name, device="cuda", mlir_dialect="linalg"
+)
 shark_module.compile()
 
 

shark/examples/shark_inference/simple_dlrm.py

@@ -360,7 +360,7 @@ mlir_importer = SharkImporter(
 )
 
 shark_module = SharkInference(
-    dlrm_mlir, device="vulkan", mlir_dialect="linalg"
+    dlrm_mlir, func_name, device="vulkan", mlir_dialect="linalg"
 )
 shark_module.compile()
 result = shark_module.forward(input_dlrm)

shark/examples/shark_inference/sparse_arch.py

@@ -294,7 +294,7 @@ def test_dlrm() -> None:
     )
 
     shark_module = SharkInference(
-        dlrm_mlir, device="cpu", mlir_dialect="linalg"
+        dlrm_mlir, func_name, device="cpu", mlir_dialect="linalg"
     )
     shark_module.compile()
     result = shark_module.forward(inputs)