Update requirements.txt

* Fix batch count

* Add button to unload models manually.

* Add compiled pipeline option

* Add brevitas to requirements

* Tweaks to chatbot

* Change script loading trigger

.github/workflows/nightly.yml

@@ -53,6 +53,7 @@ jobs:
         python process_skipfiles.py
         $env:SHARK_PACKAGE_VERSION=${{ env.package_version }}
         pip install -e .
+        pip freeze -l
         pyinstaller .\apps\shark_studio\shark_studio.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

.github/workflows/test-studio.yml

@@ -81,6 +81,5 @@ jobs:
         source shark.venv/bin/activate
         pip install -r requirements.txt --no-cache-dir
         pip install -e .
-        pip uninstall -y torch
-        pip install torch==2.1.0+cpu -f https://download.pytorch.org/whl/torch_stable.html
-        python apps/shark_studio/tests/api_test.py
+        # Disabled due to hang when exporting test llama2
+        # python apps/shark_studio/tests/api_test.py

.gitignore

@@ -164,7 +164,7 @@ cython_debug/
 # vscode related
 .vscode
 
-# Shark related artefacts
+# Shark related artifacts
 *venv/
 shark_tmp/
 *.vmfb
@@ -172,6 +172,7 @@ shark_tmp/
 tank/dict_configs.py
 *.csv
 reproducers/
+apps/shark_studio/web/configs
 
 # ORT related artefacts
 cache_models/
@@ -188,6 +189,11 @@ variants.json
 # models folder
 apps/stable_diffusion/web/models/
 
+# model artifacts (SHARK)
+*.tempfile
+*.mlir
+*.vmfb
+
 # Stencil annotators.
 stencil_annotator/
 

apps/shark_studio/api/llm.py

@@ -3,8 +3,13 @@ from turbine_models.model_runner import vmfbRunner
 from turbine_models.gen_external_params.gen_external_params import gen_external_params
 import time
 from shark.iree_utils.compile_utils import compile_module_to_flatbuffer
-from apps.shark_studio.web.utils.file_utils import get_resource_path
+from apps.shark_studio.web.utils.file_utils import (
+    get_resource_path,
+    get_checkpoints_path,
+)
 from apps.shark_studio.modules.shared_cmd_opts import cmd_opts
+from apps.shark_studio.api.utils import parse_device
+from urllib.request import urlopen
 import iree.runtime as ireert
 from itertools import chain
 import gc
@@ -65,6 +70,7 @@ class LanguageModel:
         use_system_prompt=True,
         streaming_llm=False,
     ):
+        _, _, self.triple = parse_device(device)
         self.hf_model_name = llm_model_map[model_name]["hf_model_name"]
         self.device = device.split("=>")[-1].strip()
         self.backend = self.device.split("://")[0]
@@ -155,7 +161,9 @@ class LanguageModel:
                 use_auth_token=hf_auth_token,
             )
         elif not os.path.exists(self.tempfile_name):
-            self.torch_ir, self.tokenizer = llm_model_map[model_name]["initializer"](
+            self.torch_ir, self.tokenizer = llm_model_map[self.hf_model_name][
+                "initializer"
+            ](
                 self.hf_model_name,
                 hf_auth_token,
                 compile_to="torch",
@@ -163,6 +171,7 @@ class LanguageModel:
                 precision=self.precision,
                 quantization=self.quantization,
                 streaming_llm=self.streaming_llm,
+                decomp_attn=True,
             )
             with open(self.tempfile_name, "w+") as f:
                 f.write(self.torch_ir)
@@ -192,11 +201,27 @@ class LanguageModel:
             )
         elif self.backend == "vulkan":
             flags.extend(["--iree-stream-resource-max-allocation-size=4294967296"])
+        elif self.backend == "rocm":
+            flags.extend(
+                [
+                    "--iree-codegen-llvmgpu-enable-transform-dialect-jit=false",
+                    "--iree-llvmgpu-enable-prefetch=true",
+                    "--iree-opt-outer-dim-concat=true",
+                    "--iree-flow-enable-aggressive-fusion",
+                ]
+            )
+            if "gfx9" in self.triple:
+                flags.extend(
+                    [
+                        f"--iree-codegen-transform-dialect-library={get_mfma_spec_path(self.triple, get_checkpoints_path())}",
+                        "--iree-codegen-llvmgpu-use-vector-distribution=true",
+                    ]
+                )
         flags.extend(llm_model_map[self.hf_model_name]["compile_flags"])
         flatbuffer_blob = compile_module_to_flatbuffer(
             self.tempfile_name,
             device=self.device,
-            frontend="torch",
+            frontend="auto",
             model_config_path=None,
             extra_args=flags,
             write_to=self.vmfb_name,
@@ -258,8 +283,7 @@ class LanguageModel:
 
             history.append(format_out(token))
             while (
-                format_out(token)
-                != llm_model_map["meta-llama/Llama-2-7b-chat-hf"]["stop_token"]
+                format_out(token) != llm_model_map[self.hf_model_name]["stop_token"]
                 and len(history) < self.max_tokens
             ):
                 dec_time = time.time()
@@ -273,10 +297,7 @@ class LanguageModel:
 
             self.prev_token_len = token_len + len(history)
 
-            if (
-                format_out(token)
-                == llm_model_map["meta-llama/Llama-2-7b-chat-hf"]["stop_token"]
-            ):
+            if format_out(token) == llm_model_map[self.hf_model_name]["stop_token"]:
                 break
 
         for i in range(len(history)):
@@ -310,7 +331,7 @@ class LanguageModel:
                 self.first_input = False
 
             history.append(int(token))
-            while token != llm_model_map["meta-llama/Llama-2-7b-chat-hf"]["stop_token"]:
+            while token != llm_model_map[self.hf_model_name]["stop_token"]:
                 dec_time = time.time()
                 result = self.hf_mod(token.reshape([1, 1]), past_key_values=pkv)
                 history.append(int(token))
@@ -321,7 +342,7 @@ class LanguageModel:
 
             self.prev_token_len = token_len + len(history)
 
-            if token == llm_model_map["meta-llama/Llama-2-7b-chat-hf"]["stop_token"]:
+            if token == llm_model_map[self.hf_model_name]["stop_token"]:
                 break
         for i in range(len(history)):
             if type(history[i]) != int:
@@ -331,6 +352,17 @@ class LanguageModel:
         return result_output, total_time
 
 
+def get_mfma_spec_path(target_chip, save_dir):
+    url = "https://raw.githubusercontent.com/iree-org/iree/main/build_tools/pkgci/external_test_suite/attention_and_matmul_spec.mlir"
+    attn_spec = urlopen(url).read().decode("utf-8")
+    spec_path = os.path.join(save_dir, "attention_and_matmul_spec_mfma.mlir")
+    if os.path.exists(spec_path):
+        return spec_path
+    with open(spec_path, "w") as f:
+        f.write(attn_spec)
+    return spec_path
+
+
 def llm_chat_api(InputData: dict):
     from datetime import datetime as dt
 

apps/shark_studio/api/sd.py

@@ -1,54 +1,82 @@
 import gc
 import torch
+import gradio as gr
 import time
 import os
 import json
 import numpy as np
+import copy
+import importlib.util
+import sys
 from tqdm.auto import tqdm
 
 from pathlib import Path
 from random import randint
-from turbine_models.custom_models.sd_inference import clip, unet, vae
+from turbine_models.custom_models.sd_inference.sd_pipeline import SharkSDPipeline
+from turbine_models.custom_models.sdxl_inference.sdxl_compiled_pipeline import (
+    SharkSDXLPipeline,
+)
+
+
 from apps.shark_studio.api.controlnet import control_adapter_map
+from apps.shark_studio.api.utils import parse_device
 from apps.shark_studio.web.utils.state import status_label
 from apps.shark_studio.web.utils.file_utils import (
     safe_name,
     get_resource_path,
     get_checkpoints_path,
 )
-from apps.shark_studio.modules.pipeline import SharkPipelineBase
-from apps.shark_studio.modules.schedulers import get_schedulers
-from apps.shark_studio.modules.prompt_encoding import (
-    get_weighted_text_embeddings,
-)
+
 from apps.shark_studio.modules.img_processing import (
-    resize_stencil,
     save_output_img,
-    resamplers,
-    resampler_list,
 )
 
 from apps.shark_studio.modules.ckpt_processing import (
     preprocessCKPT,
-    process_custom_pipe_weights,
+    save_irpa,
 )
-from transformers import CLIPTokenizer
-from diffusers.image_processor import VaeImageProcessor
 
-sd_model_map = {
-    "clip": {
-        "initializer": clip.export_clip_model,
-    },
-    "unet": {
-        "initializer": unet.export_unet_model,
-    },
-    "vae_decode": {
-        "initializer": vae.export_vae_model,
-    },
+EMPTY_SD_MAP = {
+    "clip": None,
+    "scheduler": None,
+    "unet": None,
+    "vae_decode": None,
+}
+
+EMPTY_SDXL_MAP = {
+    "prompt_encoder": None,
+    "scheduled_unet": None,
+    "vae_decode": None,
+    "pipeline": None,
+    "full_pipeline": None,
+}
+
+EMPTY_FLAGS = {
+    "clip": None,
+    "unet": None,
+    "vae": None,
+    "pipeline": None,
 }
 
 
-class StableDiffusion(SharkPipelineBase):
+def load_script(source, module_name):
+    """
+    reads file source and loads it as a module
+
+    :param source: file to load
+    :param module_name: name of module to register in sys.modules
+    :return: loaded module
+    """
+
+    spec = importlib.util.spec_from_file_location(module_name, source)
+    module = importlib.util.module_from_spec(spec)
+    sys.modules[module_name] = module
+    spec.loader.exec_module(module)
+
+    return module
+
+
+class StableDiffusion:
     # This class is responsible for executing image generation and creating
     # /managing a set of compiled modules to run Stable Diffusion. The init
     # aims to be as general as possible, and the class will infer and compile
@@ -61,66 +89,45 @@ class StableDiffusion(SharkPipelineBase):
         height: int,
         width: int,
         batch_size: int,
+        steps: int,
+        scheduler: str,
         precision: str,
         device: str,
+        target_triple: str = None,
         custom_vae: str = None,
         num_loras: int = 0,
         import_ir: bool = True,
         is_controlled: bool = False,
-        hf_auth_token=None,
+        external_weights: str = "safetensors",
     ):
-        self.model_max_length = 77
-        self.batch_size = batch_size
         self.precision = precision
-        self.dtype = torch.float16 if precision == "fp16" else torch.float32
-        self.height = height
-        self.width = width
-        self.scheduler_obj = {}
-        static_kwargs = {
-            "pipe": {
-                "external_weights": "safetensors",
-            },
-            "clip": {"hf_model_name": base_model_id},
-            "unet": {
-                "hf_model_name": base_model_id,
-                "unet_model": unet.UnetModel(hf_model_name=base_model_id),
-                "batch_size": batch_size,
-                # "is_controlled": is_controlled,
-                # "num_loras": num_loras,
-                "height": height,
-                "width": width,
-                "precision": precision,
-                "max_length": self.model_max_length,
-            },
-            "vae_encode": {
-                "hf_model_name": base_model_id,
-                "vae_model": vae.VaeModel(
-                    hf_model_name=custom_vae if custom_vae else base_model_id,
-                ),
-                "batch_size": batch_size,
-                "height": height,
-                "width": width,
-                "precision": precision,
-            },
-            "vae_decode": {
-                "hf_model_name": base_model_id,
-                "vae_model": vae.VaeModel(
-                    hf_model_name=custom_vae if custom_vae else base_model_id,
-                ),
-                "batch_size": batch_size,
-                "height": height,
-                "width": width,
-                "precision": precision,
-            },
-        }
-        super().__init__(sd_model_map, base_model_id, static_kwargs, device, import_ir)
+        self.compiled_pipeline = False
+        self.base_model_id = base_model_id
+        self.custom_vae = custom_vae
+        self.is_sdxl = "xl" in self.base_model_id.lower()
+        self.is_custom = ".py" in self.base_model_id.lower()
+        if self.is_custom:
+            custom_module = load_script(
+                os.path.join(get_checkpoints_path("scripts"), self.base_model_id),
+                "custom_pipeline",
+            )
+            self.turbine_pipe = custom_module.StudioPipeline
+            self.model_map = custom_module.MODEL_MAP
+        elif self.is_sdxl:
+            self.turbine_pipe = SharkSDXLPipeline
+            self.model_map = EMPTY_SDXL_MAP
+        else:
+            self.turbine_pipe = SharkSDPipeline
+            self.model_map = EMPTY_SD_MAP
+        max_length = 64
+        target_backend, self.rt_device, triple = parse_device(device, target_triple)
         pipe_id_list = [
             safe_name(base_model_id),
             str(batch_size),
-            str(self.model_max_length),
+            str(max_length),
             f"{str(height)}x{str(width)}",
             precision,
-            self.device,
+            triple,
         ]
         if num_loras > 0:
             pipe_id_list.append(str(num_loras) + "lora")
@@ -129,305 +136,147 @@ class StableDiffusion(SharkPipelineBase):
         if custom_vae:
             pipe_id_list.append(custom_vae)
         self.pipe_id = "_".join(pipe_id_list)
-        print(f"\n[LOG] Pipeline initialized with pipe_id: {self.pipe_id}.")
-        del static_kwargs
-        gc.collect()
-
-    def prepare_pipe(self, custom_weights, adapters, embeddings, is_img2img):
-        print(f"\n[LOG] Preparing pipeline...")
-        self.is_img2img = is_img2img
-        self.schedulers = get_schedulers(self.base_model_id)
-
-        self.weights_path = os.path.join(
-            get_checkpoints_path(), self.safe_name(self.base_model_id)
+        self.pipeline_dir = Path(os.path.join(get_checkpoints_path(), self.pipe_id))
+        self.weights_path = Path(
+            os.path.join(
+                get_checkpoints_path(), safe_name(self.base_model_id + "_" + precision)
+            )
         )
         if not os.path.exists(self.weights_path):
             os.mkdir(self.weights_path)
 
-        for model in adapters:
-            self.model_map[model] = adapters[model]
+        decomp_attn = True
+        attn_spec = None
+        if triple in ["gfx940", "gfx942", "gfx90a"]:
+            decomp_attn = False
+            attn_spec = "mfma"
+        elif triple in ["gfx1100", "gfx1103", "gfx1150"]:
+            decomp_attn = False
+            attn_spec = "wmma"
+            if triple in ["gfx1103", "gfx1150"]:
+                # external weights have issues on igpu
+                external_weights = None
+        elif target_backend == "llvm-cpu":
+            decomp_attn = False
 
-        for submodel in self.static_kwargs:
-            if custom_weights:
-                custom_weights_params, _ = process_custom_pipe_weights(custom_weights)
-                if submodel not in ["clip", "clip2"]:
-                    self.static_kwargs[submodel][
-                        "external_weights"
-                    ] = custom_weights_params
-                else:
-                    self.static_kwargs[submodel]["external_weight_path"] = os.path.join(
-                        self.weights_path, submodel + ".safetensors"
+        self.sd_pipe = self.turbine_pipe(
+            hf_model_name=base_model_id,
+            scheduler_id=scheduler,
+            height=height,
+            width=width,
+            precision=precision,
+            max_length=max_length,
+            batch_size=batch_size,
+            num_inference_steps=steps,
+            device=target_backend,
+            iree_target_triple=triple,
+            ireec_flags=EMPTY_FLAGS,
+            attn_spec=attn_spec,
+            decomp_attn=decomp_attn,
+            pipeline_dir=self.pipeline_dir,
+            external_weights_dir=self.weights_path,
+            external_weights=external_weights,
+            custom_vae=custom_vae,
+        )
+        print(f"\n[LOG] Pipeline initialized with pipe_id: {self.pipe_id}.")
+        gc.collect()
+
+    def prepare_pipe(
+        self, custom_weights, adapters, embeddings, is_img2img, compiled_pipeline
+    ):
+        print(f"\n[LOG] Preparing pipeline...")
+        self.is_img2img = False
+        mlirs = copy.deepcopy(self.model_map)
+        vmfbs = copy.deepcopy(self.model_map)
+        weights = copy.deepcopy(self.model_map)
+        if not self.is_sdxl:
+            compiled_pipeline = False
+        self.compiled_pipeline = compiled_pipeline
+
+        if custom_weights:
+            custom_weights = os.path.join(
+                get_checkpoints_path("checkpoints"),
+                safe_name(self.base_model_id.split("/")[-1]),
+                custom_weights,
+            )
+            diffusers_weights_path = preprocessCKPT(custom_weights, self.precision)
+            for key in weights:
+                if key in ["scheduled_unet", "unet"]:
+                    unet_weights_path = os.path.join(
+                        diffusers_weights_path,
+                        "unet",
+                        "diffusion_pytorch_model.safetensors",
                     )
-            else:
-                self.static_kwargs[submodel]["external_weight_path"] = os.path.join(
-                    self.weights_path, submodel + ".safetensors"
-                )
+                    weights[key] = save_irpa(unet_weights_path, "unet.")
 
-        self.get_compiled_map(pipe_id=self.pipe_id)
-        print("\n[LOG] Pipeline successfully prepared for runtime.")
+                elif key in ["clip", "prompt_encoder"]:
+                    if not self.is_sdxl:
+                        sd1_path = os.path.join(
+                            diffusers_weights_path, "text_encoder", "model.safetensors"
+                        )
+                        weights[key] = save_irpa(sd1_path, "text_encoder_model.")
+                    else:
+                        clip_1_path = os.path.join(
+                            diffusers_weights_path, "text_encoder", "model.safetensors"
+                        )
+                        clip_2_path = os.path.join(
+                            diffusers_weights_path,
+                            "text_encoder_2",
+                            "model.safetensors",
+                        )
+                        weights[key] = [
+                            save_irpa(clip_1_path, "text_encoder_model_1."),
+                            save_irpa(clip_2_path, "text_encoder_model_2."),
+                        ]
+
+                elif key in ["vae_decode"] and weights[key] is None:
+                    vae_weights_path = os.path.join(
+                        diffusers_weights_path,
+                        "vae",
+                        "diffusion_pytorch_model.safetensors",
+                    )
+                    weights[key] = save_irpa(vae_weights_path, "vae.")
+
+        vmfbs, weights = self.sd_pipe.check_prepared(
+            mlirs, vmfbs, weights, interactive=False
+        )
+        print(f"\n[LOG] Loading pipeline to device {self.rt_device}.")
+        self.sd_pipe.load_pipeline(
+            vmfbs, weights, self.rt_device, self.compiled_pipeline
+        )
+        print(
+            "\n[LOG] Pipeline successfully prepared for runtime. Generating images..."
+        )
         return
 
-    def encode_prompts_weight(
-        self,
-        prompt,
-        negative_prompt,
-        do_classifier_free_guidance=True,
-    ):
-        # Encodes the prompt into text encoder hidden states.
-        self.load_submodels(["clip"])
-        self.tokenizer = CLIPTokenizer.from_pretrained(
-            self.base_model_id,
-            subfolder="tokenizer",
-        )
-        clip_inf_start = time.time()
-
-        text_embeddings, uncond_embeddings = get_weighted_text_embeddings(
-            pipe=self,
-            prompt=prompt,
-            uncond_prompt=negative_prompt if do_classifier_free_guidance else None,
-        )
-
-        if do_classifier_free_guidance:
-            text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
-
-        pad = (0, 0) * (len(text_embeddings.shape) - 2)
-        pad = pad + (
-            0,
-            self.static_kwargs["unet"]["max_length"] - text_embeddings.shape[1],
-        )
-        text_embeddings = torch.nn.functional.pad(text_embeddings, pad)
-
-        # SHARK: Report clip inference time
-        clip_inf_time = (time.time() - clip_inf_start) * 1000
-        if self.ondemand:
-            self.unload_submodels(["clip"])
-            gc.collect()
-        print(f"\n[LOG] Clip Inference time (ms) = {clip_inf_time:.3f}")
-
-        return text_embeddings.numpy().astype(np.float16)
-
-    def prepare_latents(
-        self,
-        generator,
-        num_inference_steps,
-        image,
-        strength,
-    ):
-        noise = torch.randn(
-            (
-                self.batch_size,
-                4,
-                self.height // 8,
-                self.width // 8,
-            ),
-            generator=generator,
-            dtype=self.dtype,
-        ).to("cpu")
-
-        self.scheduler.set_timesteps(num_inference_steps)
-        if self.is_img2img:
-            init_timestep = min(
-                int(num_inference_steps * strength), num_inference_steps
-            )
-            t_start = max(num_inference_steps - init_timestep, 0)
-            timesteps = self.scheduler.timesteps[t_start:]
-            latents = self.encode_image(image)
-            latents = self.scheduler.add_noise(latents, noise, timesteps[0].repeat(1))
-            return latents, [timesteps]
-        else:
-            self.scheduler.is_scale_input_called = True
-            latents = noise * self.scheduler.init_noise_sigma
-            return latents, self.scheduler.timesteps
-
-    def encode_image(self, input_image):
-        self.load_submodels(["vae_encode"])
-        vae_encode_start = time.time()
-        latents = self.run("vae_encode", input_image)
-        vae_inf_time = (time.time() - vae_encode_start) * 1000
-        if self.ondemand:
-            self.unload_submodels(["vae_encode"])
-        print(f"\n[LOG] VAE Encode Inference time (ms): {vae_inf_time:.3f}")
-
-        return latents
-
-    def produce_img_latents(
-        self,
-        latents,
-        text_embeddings,
-        guidance_scale,
-        total_timesteps,
-        cpu_scheduling,
-        mask=None,
-        masked_image_latents=None,
-        return_all_latents=False,
-    ):
-        # self.status = SD_STATE_IDLE
-        step_time_sum = 0
-        latent_history = [latents]
-        text_embeddings = torch.from_numpy(text_embeddings).to(self.dtype)
-        text_embeddings_numpy = text_embeddings.detach().numpy()
-        guidance_scale = torch.Tensor([guidance_scale]).to(self.dtype)
-        self.load_submodels(["unet"])
-        for i, t in tqdm(enumerate(total_timesteps)):
-            step_start_time = time.time()
-            timestep = torch.tensor([t]).to(self.dtype).detach().numpy()
-            latent_model_input = self.scheduler.scale_model_input(latents, t).to(
-                self.dtype
-            )
-            if mask is not None and masked_image_latents is not None:
-                latent_model_input = torch.cat(
-                    [
-                        torch.from_numpy(np.asarray(latent_model_input)).to(self.dtype),
-                        mask,
-                        masked_image_latents,
-                    ],
-                    dim=1,
-                ).to(self.dtype)
-            if cpu_scheduling:
-                latent_model_input = latent_model_input.detach().numpy()
-
-            # Profiling Unet.
-            # profile_device = start_profiling(file_path="unet.rdc")
-            noise_pred = self.run(
-                "unet",
-                [
-                    latent_model_input,
-                    timestep,
-                    text_embeddings_numpy,
-                    guidance_scale,
-                ],
-            )
-            # end_profiling(profile_device)
-
-            if cpu_scheduling:
-                noise_pred = torch.from_numpy(noise_pred.to_host())
-                latents = self.scheduler.step(noise_pred, t, latents).prev_sample
-            else:
-                latents = self.run("scheduler_step", (noise_pred, t, latents))
-
-            latent_history.append(latents)
-            step_time = (time.time() - step_start_time) * 1000
-            # print(
-            #     f"\n [LOG] step = {i} | timestep = {t} | time = {step_time:.2f}ms"
-            # )
-            step_time_sum += step_time
-
-            # if self.status == SD_STATE_CANCEL:
-            #    break
-
-        if self.ondemand:
-            self.unload_submodels(["unet"])
-            gc.collect()
-
-        avg_step_time = step_time_sum / len(total_timesteps)
-        print(f"\n[LOG] Average step time: {avg_step_time}ms/it")
-
-        if not return_all_latents:
-            return latents
-        all_latents = torch.cat(latent_history, dim=0)
-        return all_latents
-
-    def decode_latents(self, latents, cpu_scheduling=True):
-        latents_numpy = latents.to(self.dtype)
-        if cpu_scheduling:
-            latents_numpy = latents.detach().numpy()
-
-        # profile_device = start_profiling(file_path="vae.rdc")
-        vae_start = time.time()
-        images = self.run("vae_decode", latents_numpy).to_host()
-        vae_inf_time = (time.time() - vae_start) * 1000
-        # end_profiling(profile_device)
-        print(f"\n[LOG] VAE Inference time (ms): {vae_inf_time:.3f}")
-
-        images = torch.from_numpy(images).permute(0, 2, 3, 1).float().numpy()
-        pil_images = self.image_processor.numpy_to_pil(images)
-        return pil_images
-
     def generate_images(
         self,
         prompt,
         negative_prompt,
         image,
-        scheduler,
-        steps,
         strength,
         guidance_scale,
         seed,
         ondemand,
-        repeatable_seeds,
         resample_type,
         control_mode,
         hints,
     ):
-        # TODO: Batched args
-        self.image_processor = VaeImageProcessor(do_convert_rgb=True)
-        self.scheduler = self.schedulers[scheduler]
-        self.ondemand = ondemand
-        if self.is_img2img:
-            image, _ = self.image_processor.preprocess(image, resample_type)
-        else:
-            image = None
-
-        print("\n[LOG] Generating images...")
-        batched_args = [
-            prompt,
-            negative_prompt,
-            image,
-        ]
-        for arg in batched_args:
-            if not isinstance(arg, list):
-                arg = [arg] * self.batch_size
-            if len(arg) < self.batch_size:
-                arg = arg * self.batch_size
-            else:
-                arg = [arg[i] for i in range(self.batch_size)]
-
-        text_embeddings = self.encode_prompts_weight(
+        img = self.sd_pipe.generate_images(
             prompt,
             negative_prompt,
+            1,
+            guidance_scale,
+            seed,
+            return_imgs=True,
         )
-
-        uint32_info = np.iinfo(np.uint32)
-        uint32_min, uint32_max = uint32_info.min, uint32_info.max
-        if seed < uint32_min or seed >= uint32_max:
-            seed = randint(uint32_min, uint32_max)
-
-        generator = torch.manual_seed(seed)
-
-        init_latents, final_timesteps = self.prepare_latents(
-            generator=generator,
-            num_inference_steps=steps,
-            image=image,
-            strength=strength,
-        )
-
-        latents = self.produce_img_latents(
-            latents=init_latents,
-            text_embeddings=text_embeddings,
-            guidance_scale=guidance_scale,
-            total_timesteps=final_timesteps,
-            cpu_scheduling=True,  # until we have schedulers through Turbine
-        )
-
-        # Img latents -> PIL images
-        all_imgs = []
-        self.load_submodels(["vae_decode"])
-        for i in tqdm(range(0, latents.shape[0], self.batch_size)):
-            imgs = self.decode_latents(
-                latents=latents[i : i + self.batch_size],
-                cpu_scheduling=True,
-            )
-            all_imgs.extend(imgs)
-        if self.ondemand:
-            self.unload_submodels(["vae_decode"])
-
-        return all_imgs
+        return img
 
 
 def shark_sd_fn_dict_input(
     sd_kwargs: dict,
 ):
-    print("[LOG] Submitting Request...")
+    print("\n[LOG] Submitting Request...")
 
     for key in sd_kwargs:
         if sd_kwargs[key] in [None, []]:
@@ -437,9 +286,34 @@ def shark_sd_fn_dict_input(
         if key == "seed":
             sd_kwargs[key] = int(sd_kwargs[key])
 
-    for i in range(1):
-        generated_imgs = yield from shark_sd_fn(**sd_kwargs)
-        yield generated_imgs
+    # TODO: move these checks into the UI code so we don't have gradio warnings in a generalized dict input function.
+    if not sd_kwargs["device"]:
+        gr.Warning("No device specified. Please specify a device.")
+        return None, ""
+    if sd_kwargs["height"] not in [512, 1024]:
+        gr.Warning("Height must be 512 or 1024. This is a temporary limitation.")
+        return None, ""
+    if sd_kwargs["height"] != sd_kwargs["width"]:
+        gr.Warning("Height and width must be the same. This is a temporary limitation.")
+        return None, ""
+    if sd_kwargs["base_model_id"] == "stabilityai/sdxl-turbo":
+        if sd_kwargs["steps"] > 10:
+            gr.Warning("Max steps for sdxl-turbo is 10. 1 to 4 steps are recommended.")
+            return None, ""
+        if sd_kwargs["guidance_scale"] > 3:
+            gr.Warning(
+                "sdxl-turbo CFG scale should be less than 2.0 if using negative prompt, 0 otherwise."
+            )
+            return None, ""
+    if sd_kwargs["target_triple"] == "":
+        if parse_device(sd_kwargs["device"], sd_kwargs["target_triple"])[2] == "":
+            gr.Warning(
+                "Target device architecture could not be inferred. Please specify a target triple, e.g. 'gfx1100' for a Radeon 7900xtx."
+            )
+            return None, ""
+
+    generated_imgs = yield from shark_sd_fn(**sd_kwargs)
+    return generated_imgs
 
 
 def shark_sd_fn(
@@ -460,8 +334,9 @@ def shark_sd_fn(
     custom_vae: str,
     precision: str,
     device: str,
+    target_triple: str,
     ondemand: bool,
-    repeatable_seeds: bool,
+    compiled_pipeline: bool,
     resample_type: str,
     controlnets: dict,
     embeddings: dict,
@@ -471,8 +346,6 @@ def shark_sd_fn(
         sd_init_image = [sd_init_image]
     is_img2img = True if sd_init_image[0] is not None else False
 
-    print("\n[LOG] Performing Stable Diffusion Pipeline setup...")
-
     from apps.shark_studio.modules.shared_cmd_opts import cmd_opts
     import apps.shark_studio.web.utils.globals as global_obj
 
@@ -481,6 +354,7 @@ def shark_sd_fn(
     control_mode = None
     hints = []
     num_loras = 0
+    import_ir = True
     for i in embeddings:
         num_loras += 1 if embeddings[i] else 0
     if "model" in controlnets:
@@ -512,28 +386,29 @@ def shark_sd_fn(
         "batch_size": batch_size,
         "precision": precision,
         "device": device,
+        "target_triple": target_triple,
         "custom_vae": custom_vae,
         "num_loras": num_loras,
-        "import_ir": cmd_opts.import_mlir,
+        "import_ir": import_ir,
         "is_controlled": is_controlled,
+        "steps": steps,
+        "scheduler": scheduler,
     }
     submit_prep_kwargs = {
         "custom_weights": custom_weights,
         "adapters": adapters,
         "embeddings": embeddings,
         "is_img2img": is_img2img,
+        "compiled_pipeline": compiled_pipeline,
     }
     submit_run_kwargs = {
         "prompt": prompt,
         "negative_prompt": negative_prompt,
         "image": sd_init_image,
-        "steps": steps,
-        "scheduler": scheduler,
         "strength": strength,
         "guidance_scale": guidance_scale,
         "seed": seed,
         "ondemand": ondemand,
-        "repeatable_seeds": repeatable_seeds,
         "resample_type": resample_type,
         "control_mode": control_mode,
         "hints": hints,
@@ -566,22 +441,35 @@ def shark_sd_fn(
     for current_batch in range(batch_count):
         start_time = time.time()
         out_imgs = global_obj.get_sd_obj().generate_images(**submit_run_kwargs)
-        total_time = time.time() - start_time
-        text_output = f"Total image(s) generation time: {total_time:.4f}sec"
-        print(f"\n[LOG] {text_output}")
+        if not isinstance(out_imgs, list):
+            out_imgs = [out_imgs]
+        # total_time = time.time() - start_time
+        # text_output = f"Total image(s) generation time: {total_time:.4f}sec"
+        # print(f"\n[LOG] {text_output}")
         # if global_obj.get_sd_status() == SD_STATE_CANCEL:
         #     break
         # else:
-        save_output_img(
-            out_imgs[current_batch],
-            seed,
-            sd_kwargs,
-        )
+        for batch in range(batch_size):
+            save_output_img(
+                out_imgs[batch],
+                seed,
+                sd_kwargs,
+            )
         generated_imgs.extend(out_imgs)
+        # TODO: make seed changes over batch counts more configurable.
+        submit_run_kwargs["seed"] = submit_run_kwargs["seed"] + 1
         yield generated_imgs, status_label(
             "Stable Diffusion", current_batch + 1, batch_count, batch_size
         )
-    return generated_imgs, ""
+    return (generated_imgs, "")
+
+
+def unload_sd():
+    print("Unloading models.")
+    import apps.shark_studio.web.utils.globals as global_obj
+
+    global_obj.clear_cache()
+    gc.collect()
 
 
 def cancel_sd():
@@ -596,6 +484,10 @@ def view_json_file(file_path):
     return content
 
 
+def safe_name(name):
+    return name.replace("/", "_").replace("\\", "_").replace(".", "_")
+
+
 if __name__ == "__main__":
     from apps.shark_studio.modules.shared_cmd_opts import cmd_opts
     import apps.shark_studio.web.utils.globals as global_obj

apps/shark_studio/api/utils.py

@@ -52,6 +52,13 @@ def get_available_devices():
     set_iree_runtime_flags()
 
     available_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")
+    available_devices.extend(cpu_device)
+
     from shark.iree_utils.vulkan_utils import (
         get_all_vulkan_devices,
     )
@@ -64,17 +71,28 @@ def get_available_devices():
         id += 1
     if id != 0:
         print(f"vulkan devices are available.")
+
     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")
-    available_devices.extend(cpu_device)
+    hip_devices = get_devices_by_name("hip")
+    available_devices.extend(hip_devices)
+
+    for idx, device_str in enumerate(available_devices):
+        if "AMD Radeon(TM) Graphics =>" in device_str:
+            igpu_id_candidates = [
+                x.split("w/")[-1].split("=>")[0]
+                for x in available_devices
+                if "M Graphics" in x
+            ]
+            for igpu_name in igpu_id_candidates:
+                if igpu_name:
+                    available_devices[idx] = device_str.replace(
+                        "AMD Radeon(TM) Graphics", igpu_name
+                    )
+                break
     return available_devices
 
 
@@ -127,6 +145,57 @@ def set_iree_runtime_flags():
     set_iree_vulkan_runtime_flags(flags=vulkan_runtime_flags)
 
 
+def parse_device(device_str, target_override=""):
+    from shark.iree_utils.compile_utils import (
+        clean_device_info,
+        get_iree_target_triple,
+        iree_target_map,
+    )
+
+    rt_driver, device_id = clean_device_info(device_str)
+    target_backend = iree_target_map(rt_driver)
+    if device_id:
+        rt_device = f"{rt_driver}://{device_id}"
+    else:
+        rt_device = rt_driver
+
+    if target_override:
+        return target_backend, rt_device, target_override
+    match target_backend:
+        case "vulkan-spirv":
+            triple = get_iree_target_triple(device_str)
+            return target_backend, rt_device, triple
+        case "rocm":
+            triple = get_rocm_target_chip(device_str)
+            return target_backend, rt_device, triple
+        case "llvm-cpu":
+            return "llvm-cpu", "local-task", "x86_64-linux-gnu"
+
+
+def get_rocm_target_chip(device_str):
+    # TODO: Use a data file to map device_str to target chip.
+    rocm_chip_map = {
+        "6700": "gfx1031",
+        "6800": "gfx1030",
+        "6900": "gfx1030",
+        "7900": "gfx1100",
+        "MI300X": "gfx942",
+        "MI300A": "gfx940",
+        "MI210": "gfx90a",
+        "MI250": "gfx90a",
+        "MI100": "gfx908",
+        "MI50": "gfx906",
+        "MI60": "gfx906",
+        "780M": "gfx1103",
+    }
+    for key in rocm_chip_map:
+        if key in device_str:
+            return rocm_chip_map[key]
+    raise AssertionError(
+        f"Device {device_str} not recognized. Please file an issue at https://github.com/nod-ai/SHARK/issues."
+    )
+
+
 def get_all_devices(driver_name):
     """
     Inputs: driver_name

apps/shark_studio/modules/ckpt_processing.py

@@ -2,6 +2,11 @@ import os
 import json
 import re
 import requests
+import torch
+import safetensors
+from shark_turbine.aot.params import (
+    ParameterArchiveBuilder,
+)
 from io import BytesIO
 from pathlib import Path
 from tqdm import tqdm
@@ -15,21 +20,21 @@ from diffusers.pipelines.stable_diffusion.convert_from_ckpt import (
 )
 
 
-def get_path_to_diffusers_checkpoint(custom_weights):
+def get_path_to_diffusers_checkpoint(custom_weights, precision="fp16"):
     path = Path(custom_weights)
     diffusers_path = path.parent.absolute()
-    diffusers_directory_name = os.path.join("diffusers", path.stem)
+    diffusers_directory_name = os.path.join("diffusers", path.stem + f"_{precision}")
     complete_path_to_diffusers = diffusers_path / diffusers_directory_name
     complete_path_to_diffusers.mkdir(parents=True, exist_ok=True)
     path_to_diffusers = complete_path_to_diffusers.as_posix()
     return path_to_diffusers
 
 
-def preprocessCKPT(custom_weights, is_inpaint=False):
-    path_to_diffusers = get_path_to_diffusers_checkpoint(custom_weights)
+def preprocessCKPT(custom_weights, precision="fp16", is_inpaint=False):
+    path_to_diffusers = get_path_to_diffusers_checkpoint(custom_weights, precision)
     if next(Path(path_to_diffusers).iterdir(), None):
         print("Checkpoint already loaded at : ", path_to_diffusers)
-        return
+        return path_to_diffusers
     else:
         print(
             "Diffusers' checkpoint will be identified here : ",
@@ -51,8 +56,24 @@ def preprocessCKPT(custom_weights, is_inpaint=False):
         from_safetensors=from_safetensors,
         num_in_channels=num_in_channels,
     )
+    if precision == "fp16":
+        pipe.to(dtype=torch.float16)
     pipe.save_pretrained(path_to_diffusers)
+    del pipe
     print("Loading complete")
+    return path_to_diffusers
+
+
+def save_irpa(weights_path, prepend_str):
+    weights = safetensors.torch.load_file(weights_path)
+    archive = ParameterArchiveBuilder()
+    for key in weights.keys():
+        new_key = prepend_str + key
+        archive.add_tensor(new_key, weights[key])
+
+    irpa_file = weights_path.replace(".safetensors", ".irpa")
+    archive.save(irpa_file)
+    return irpa_file
 
 
 def convert_original_vae(vae_checkpoint):

apps/shark_studio/modules/schedulers.py

@@ -24,47 +24,47 @@ def get_schedulers(model_id):
         model_id,
         subfolder="scheduler",
     )
-    schedulers["DDPM"] = DDPMScheduler.from_pretrained(
-        model_id,
-        subfolder="scheduler",
-    )
-    schedulers["KDPM2Discrete"] = KDPM2DiscreteScheduler.from_pretrained(
-        model_id,
-        subfolder="scheduler",
-    )
-    schedulers["LMSDiscrete"] = LMSDiscreteScheduler.from_pretrained(
-        model_id,
-        subfolder="scheduler",
-    )
-    schedulers["DDIM"] = DDIMScheduler.from_pretrained(
-        model_id,
-        subfolder="scheduler",
-    )
-    schedulers["LCMScheduler"] = LCMScheduler.from_pretrained(
-        model_id,
-        subfolder="scheduler",
-    )
-    schedulers["DPMSolverMultistep"] = DPMSolverMultistepScheduler.from_pretrained(
-        model_id, subfolder="scheduler", algorithm_type="dpmsolver"
-    )
-    schedulers["DPMSolverMultistep++"] = DPMSolverMultistepScheduler.from_pretrained(
-        model_id, subfolder="scheduler", algorithm_type="dpmsolver++"
-    )
-    schedulers["DPMSolverMultistepKarras"] = (
-        DPMSolverMultistepScheduler.from_pretrained(
-            model_id,
-            subfolder="scheduler",
-            use_karras_sigmas=True,
-        )
-    )
-    schedulers["DPMSolverMultistepKarras++"] = (
-        DPMSolverMultistepScheduler.from_pretrained(
-            model_id,
-            subfolder="scheduler",
-            algorithm_type="dpmsolver++",
-            use_karras_sigmas=True,
-        )
-    )
+    # schedulers["DDPM"] = DDPMScheduler.from_pretrained(
+    #     model_id,
+    #     subfolder="scheduler",
+    # )
+    # schedulers["KDPM2Discrete"] = KDPM2DiscreteScheduler.from_pretrained(
+    #     model_id,
+    #     subfolder="scheduler",
+    # )
+    # schedulers["LMSDiscrete"] = LMSDiscreteScheduler.from_pretrained(
+    #     model_id,
+    #     subfolder="scheduler",
+    # )
+    # schedulers["DDIM"] = DDIMScheduler.from_pretrained(
+    #     model_id,
+    #     subfolder="scheduler",
+    # )
+    # schedulers["LCMScheduler"] = LCMScheduler.from_pretrained(
+    #     model_id,
+    #     subfolder="scheduler",
+    # )
+    # schedulers["DPMSolverMultistep"] = DPMSolverMultistepScheduler.from_pretrained(
+    #     model_id, subfolder="scheduler", algorithm_type="dpmsolver"
+    # )
+    # schedulers["DPMSolverMultistep++"] = DPMSolverMultistepScheduler.from_pretrained(
+    #     model_id, subfolder="scheduler", algorithm_type="dpmsolver++"
+    # )
+    # schedulers["DPMSolverMultistepKarras"] = (
+    #     DPMSolverMultistepScheduler.from_pretrained(
+    #         model_id,
+    #         subfolder="scheduler",
+    #         use_karras_sigmas=True,
+    #     )
+    # )
+    # schedulers["DPMSolverMultistepKarras++"] = (
+    #     DPMSolverMultistepScheduler.from_pretrained(
+    #         model_id,
+    #         subfolder="scheduler",
+    #         algorithm_type="dpmsolver++",
+    #         use_karras_sigmas=True,
+    #     )
+    # )
     schedulers["EulerDiscrete"] = EulerDiscreteScheduler.from_pretrained(
         model_id,
         subfolder="scheduler",
@@ -75,24 +75,24 @@ def get_schedulers(model_id):
             subfolder="scheduler",
         )
     )
-    schedulers["DEISMultistep"] = DEISMultistepScheduler.from_pretrained(
-        model_id,
-        subfolder="scheduler",
-    )
-    schedulers["DPMSolverSinglestep"] = DPMSolverSinglestepScheduler.from_pretrained(
-        model_id,
-        subfolder="scheduler",
-    )
-    schedulers["KDPM2AncestralDiscrete"] = (
-        KDPM2AncestralDiscreteScheduler.from_pretrained(
-            model_id,
-            subfolder="scheduler",
-        )
-    )
-    schedulers["HeunDiscrete"] = HeunDiscreteScheduler.from_pretrained(
-        model_id,
-        subfolder="scheduler",
-    )
+    # schedulers["DEISMultistep"] = DEISMultistepScheduler.from_pretrained(
+    #     model_id,
+    #     subfolder="scheduler",
+    # )
+    # schedulers["DPMSolverSinglestep"] = DPMSolverSinglestepScheduler.from_pretrained(
+    #     model_id,
+    #     subfolder="scheduler",
+    # )
+    # schedulers["KDPM2AncestralDiscrete"] = (
+    #     KDPM2AncestralDiscreteScheduler.from_pretrained(
+    #         model_id,
+    #         subfolder="scheduler",
+    #     )
+    # )
+    # schedulers["HeunDiscrete"] = HeunDiscreteScheduler.from_pretrained(
+    #     model_id,
+    #     subfolder="scheduler",
+    # )
     return schedulers
 
 
@@ -101,17 +101,18 @@ def export_scheduler_model(model):
 
 
 scheduler_model_map = {
+    "PNDM": export_scheduler_model("PNDMScheduler"),
+    # "DPMSolverSDE": export_scheduler_model("DpmSolverSDEScheduler"),
     "EulerDiscrete": export_scheduler_model("EulerDiscreteScheduler"),
     "EulerAncestralDiscrete": export_scheduler_model("EulerAncestralDiscreteScheduler"),
-    "LCM": export_scheduler_model("LCMScheduler"),
-    "LMSDiscrete": export_scheduler_model("LMSDiscreteScheduler"),
-    "PNDM": export_scheduler_model("PNDMScheduler"),
-    "DDPM": export_scheduler_model("DDPMScheduler"),
-    "DDIM": export_scheduler_model("DDIMScheduler"),
-    "DPMSolverMultistep": export_scheduler_model("DPMSolverMultistepScheduler"),
-    "KDPM2Discrete": export_scheduler_model("KDPM2DiscreteScheduler"),
-    "DEISMultistep": export_scheduler_model("DEISMultistepScheduler"),
-    "DPMSolverSinglestep": export_scheduler_model("DPMSolverSingleStepScheduler"),
-    "KDPM2AncestralDiscrete": export_scheduler_model("KDPM2AncestralDiscreteScheduler"),
-    "HeunDiscrete": export_scheduler_model("HeunDiscreteScheduler"),
+    # "LCM": export_scheduler_model("LCMScheduler"),
+    # "LMSDiscrete": export_scheduler_model("LMSDiscreteScheduler"),
+    # "DDPM": export_scheduler_model("DDPMScheduler"),
+    # "DDIM": export_scheduler_model("DDIMScheduler"),
+    # "DPMSolverMultistep": export_scheduler_model("DPMSolverMultistepScheduler"),
+    # "KDPM2Discrete": export_scheduler_model("KDPM2DiscreteScheduler"),
+    # "DEISMultistep": export_scheduler_model("DEISMultistepScheduler"),
+    # "DPMSolverSinglestep": export_scheduler_model("DPMSolverSingleStepScheduler"),
+    # "KDPM2AncestralDiscrete": export_scheduler_model("KDPM2AncestralDiscreteScheduler"),
+    # "HeunDiscrete": export_scheduler_model("HeunDiscreteScheduler"),
 }

apps/shark_studio/tests/api_test.py

@@ -36,6 +36,7 @@ class LLMAPITest(unittest.TestCase):
             device="cpu",
             precision="fp32",
             quantization="None",
+            streaming_llm=True,
         )
         count = 0
         label = "Turkishoure Turkish"

apps/shark_studio/web/configs/default_sd_config.json

@@ -1,28 +0,0 @@
-{
-  "prompt": [
-    "a photo taken of the front of a super-car drifting on a road near mountains at high speeds with smoke coming off the tires, front angle, front point of view, trees in the mountains of the background, ((sharp focus))"
-  ],
-  "negative_prompt": [
-    "watermark, signature, logo, text, lowres, ((monochrome, grayscale)), blurry, ugly, blur, oversaturated, cropped"
-  ],
-  "sd_init_image": [null],
-  "height": 512,
-  "width": 512,
-  "steps": 50,
-  "strength": 0.8,
-  "guidance_scale": 7.5,
-  "seed": "-1",
-  "batch_count": 1,
-  "batch_size": 1,
-  "scheduler": "EulerDiscrete",
-  "base_model_id": "stabilityai/stable-diffusion-2-1-base",
-  "custom_weights": null,
-  "custom_vae": null,
-  "precision": "fp16",
-  "device": "AMD Radeon RX 7900 XTX => vulkan://0",
-  "ondemand": false,
-  "repeatable_seeds": false,
-  "resample_type": "Nearest Neighbor",
-  "controlnets": {},
-  "embeddings": {}
-}

apps/shark_studio/web/index.py

@@ -76,8 +76,8 @@ def launch_webui(address):
 def webui():
     from apps.shark_studio.modules.shared_cmd_opts import cmd_opts
     from apps.shark_studio.web.ui.utils import (
-        nodicon_loc,
-        nodlogo_loc,
+        amdicon_loc,
+        amdlogo_loc,
     )
 
     launch_api = cmd_opts.api
@@ -172,9 +172,9 @@ def webui():
         analytics_enabled=False,
         title="Shark Studio 2.0 Beta",
     ) as studio_web:
-        nod_logo = Image.open(nodlogo_loc)
+        amd_logo = Image.open(amdlogo_loc)
         gr.Image(
-            value=nod_logo,
+            value=amd_logo,
             show_label=False,
             interactive=False,
             elem_id="tab_bar_logo",
@@ -209,7 +209,7 @@ def webui():
         inbrowser=True,
         server_name="0.0.0.0",
         server_port=cmd_opts.server_port,
-        favicon_path=nodicon_loc,
+        favicon_path=amdicon_loc,
     )
 
 

apps/shark_studio/web/ui/chat.py

@@ -137,7 +137,8 @@ with gr.Blocks(title="Chat") as chat_element:
             streaming_llm = gr.Checkbox(
                 label="Run in streaming mode (requires recompilation)",
                 value=True,
-                interactive=True,
+                interactive=False,
+                visible=False,
             )
             prompt_prefix = gr.Checkbox(
                 label="Add System Prompt",

apps/shark_studio/web/ui/css/sd_dark_theme.css

@@ -367,7 +367,7 @@ footer {
 #tab_bar_logo .image-container {
     object-fit: scale-down;
     position: absolute !important;
-    top: 14px;
+    top: 10px;
     right: 0px;
     height: 36px;
-}
+}

apps/shark_studio/web/ui/outputgallery.py

@@ -10,7 +10,7 @@ from apps.shark_studio.web.utils.file_utils import (
     get_generated_imgs_path,
     get_generated_imgs_todays_subdir,
 )
-from apps.shark_studio.web.ui.utils import nodlogo_loc
+from apps.shark_studio.web.ui.utils import amdlogo_loc
 from apps.shark_studio.web.utils.metadata import displayable_metadata
 
 # -- Functions for file, directory and image info querying
@@ -60,7 +60,7 @@ def output_subdirs() -> list[str]:
 # --- Define UI layout for Gradio
 
 with gr.Blocks() as outputgallery_element:
-    nod_logo = Image.open(nodlogo_loc)
+    amd_logo = Image.open(amdlogo_loc)
 
     with gr.Row(elem_id="outputgallery_gallery"):
         # needed to workaround gradio issue:
@@ -73,7 +73,7 @@ with gr.Blocks() as outputgallery_element:
         with gr.Column(scale=6):
             logo = gr.Image(
                 label="Getting subdirectories...",
-                value=nod_logo,
+                value=amd_logo,
                 interactive=False,
                 visible=True,
                 show_label=True,

apps/shark_studio/web/ui/sd.py

@@ -14,12 +14,12 @@ from apps.shark_studio.web.utils.file_utils import (
     get_checkpoints_path,
     get_checkpoints,
     get_configs_path,
-    write_default_sd_config,
+    write_default_sd_configs,
 )
 from apps.shark_studio.api.sd import (
-    sd_model_map,
     shark_sd_fn_dict_input,
     cancel_sd,
+    unload_sd,
 )
 from apps.shark_studio.api.controlnet import (
     cnet_preview,
@@ -33,7 +33,7 @@ from apps.shark_studio.modules.img_processing import (
 )
 from apps.shark_studio.modules.shared_cmd_opts import cmd_opts
 from apps.shark_studio.web.ui.utils import (
-    nodlogo_loc,
+    amdlogo_loc,
     none_to_str_none,
     str_none_to_none,
 )
@@ -45,11 +45,10 @@ from apps.shark_studio.modules import logger
 import apps.shark_studio.web.utils.globals as global_obj
 
 sd_default_models = [
-    "CompVis/stable-diffusion-v1-4",
     "runwayml/stable-diffusion-v1-5",
     "stabilityai/stable-diffusion-2-1-base",
     "stabilityai/stable-diffusion-2-1",
-    "stabilityai/stable-diffusion-xl-1.0",
+    "stabilityai/stable-diffusion-xl-base-1.0",
     "stabilityai/sdxl-turbo",
 ]
 
@@ -119,8 +118,9 @@ def pull_sd_configs(
     custom_vae,
     precision,
     device,
+    target_triple,
     ondemand,
-    repeatable_seeds,
+    compiled_pipeline,
     resample_type,
     controlnets,
     embeddings,
@@ -177,8 +177,9 @@ def load_sd_cfg(sd_json: dict, load_sd_config: str):
         sd_json["custom_vae"],
         sd_json["precision"],
         sd_json["device"],
+        sd_json["target_triple"],
         sd_json["ondemand"],
-        sd_json["repeatable_seeds"],
+        sd_json["compiled_pipeline"],
         sd_json["resample_type"],
         sd_json["controlnets"],
         sd_json["embeddings"],
@@ -231,14 +232,9 @@ def import_original(original_img, width, height):
 
 
 def base_model_changed(base_model_id):
-    ckpt_path = Path(
-        os.path.join(
-            cmd_opts.model_dir, "checkpoints", os.path.basename(str(base_model_id))
-        )
-    )
-    ckpt_path.mkdir(parents=True, exist_ok=True)
-
-    new_choices = get_checkpoints(ckpt_path) + get_checkpoints(model_type="checkpoints")
+    new_choices = get_checkpoints(
+        os.path.join("checkpoints", os.path.basename(str(base_model_id)))
+    ) + get_checkpoints(model_type="checkpoints")
 
     return gr.Dropdown(
         value=new_choices[0] if len(new_choices) > 0 else "None",
@@ -260,6 +256,11 @@ with gr.Blocks(title="Stable Diffusion") as sd_element:
                         choices=global_obj.get_device_list(),
                         allow_custom_value=False,
                     )
+                    target_triple = gr.Textbox(
+                        elem_id="target_triple",
+                        label="Architecture",
+                        value="",
+                    )
                     with gr.Row():
                         ondemand = gr.Checkbox(
                             value=cmd_opts.lowvram,
@@ -282,18 +283,19 @@ with gr.Blocks(title="Stable Diffusion") as sd_element:
                     elem_id="custom_model",
                     value="stabilityai/stable-diffusion-2-1-base",
                     choices=sd_default_models,
+                    allow_custom_value=True,
                 )  # base_model_id
                 with gr.Row():
                     height = gr.Slider(
                         384,
-                        768,
+                        1024,
                         value=cmd_opts.height,
                         step=8,
                         label="\U00002195\U0000FE0F Height",
                     )
                     width = gr.Slider(
                         384,
-                        768,
+                        1024,
                         value=cmd_opts.width,
                         step=8,
                         label="\U00002194\U0000FE0F Width",
@@ -604,35 +606,34 @@ with gr.Blocks(title="Stable Diffusion") as sd_element:
                                     interactive=True,
                                     visible=True,
                                 )
-                                repeatable_seeds = gr.Checkbox(
-                                    cmd_opts.repeatable_seeds,
-                                    label="Use Repeatable Seeds for Batches",
+                                compiled_pipeline = gr.Checkbox(
+                                    False,
+                                    label="Faster txt2img (SDXL only)",
                                 )
                             with gr.Row():
                                 stable_diffusion = gr.Button("Start")
-                                random_seed = gr.Button("Randomize Seed")
-                                random_seed.click(
-                                    lambda: -1,
-                                    inputs=[],
-                                    outputs=[seed],
+                                unload = gr.Button("Unload Models")
+                                unload.click(
+                                    fn=unload_sd,
                                     queue=False,
                                     show_progress=False,
                                 )
                                 stop_batch = gr.Button("Stop")
                     with gr.Tab(label="Config", id=102) as sd_tab_config:
                         with gr.Column(elem_classes=["sd-right-panel"]):
-                            Path(get_configs_path()).mkdir(parents=True, exist_ok=True)
-                            default_config_file = os.path.join(
-                                get_configs_path(),
-                                "default_sd_config.json",
-                            )
-                            write_default_sd_config(default_config_file)
-                            sd_json = gr.JSON(
-                                label="SD Config",
-                                elem_classes=["fill"],
-                                value=view_json_file(default_config_file),
-                                render=False,
-                            )
+                            with gr.Row(elem_classes=["fill"]):
+                                Path(get_configs_path()).mkdir(
+                                    parents=True, exist_ok=True
+                                )
+                                default_config_file = os.path.join(
+                                    get_configs_path(),
+                                    "default_sd_config.json",
+                                )
+                                write_default_sd_configs(get_configs_path())
+                                sd_json = gr.JSON(
+                                    elem_classes=["fill"],
+                                    value=view_json_file(default_config_file),
+                                )
                             with gr.Row():
                                 with gr.Column(scale=3):
                                     load_sd_config = gr.FileExplorer(
@@ -682,8 +683,9 @@ with gr.Blocks(title="Stable Diffusion") as sd_element:
                                         custom_vae,
                                         precision,
                                         device,
+                                        target_triple,
                                         ondemand,
-                                        repeatable_seeds,
+                                        compiled_pipeline,
                                         resample_type,
                                         cnet_config,
                                         embeddings_config,
@@ -695,8 +697,6 @@ with gr.Blocks(title="Stable Diffusion") as sd_element:
                                     inputs=[sd_json, sd_config_name],
                                     outputs=[sd_config_name],
                                 )
-                            with gr.Row(elem_classes=["fill"]):
-                                sd_json.render()
                         save_sd_config.click(
                             fn=save_sd_cfg,
                             inputs=[sd_json, sd_config_name],
@@ -708,6 +708,7 @@ with gr.Blocks(title="Stable Diffusion") as sd_element:
                                 value=f"{sd_model_info}\n"
                                 f"Images will be saved at "
                                 f"{get_generated_imgs_path()}",
+                                lines=2,
                                 elem_id="std_output",
                                 show_label=True,
                                 label="Log",
@@ -717,8 +718,6 @@ with gr.Blocks(title="Stable Diffusion") as sd_element:
                                 logger.read_sd_logs, None, std_output, every=1
                             )
                             sd_status = gr.Textbox(visible=False)
-                    with gr.Tab(label="Automation", id=104) as sd_tab_automation:
-                        pass
 
     pull_kwargs = dict(
         fn=pull_sd_configs,
@@ -740,8 +739,9 @@ with gr.Blocks(title="Stable Diffusion") as sd_element:
             custom_vae,
             precision,
             device,
+            target_triple,
             ondemand,
-            repeatable_seeds,
+            compiled_pipeline,
             resample_type,
             cnet_config,
             embeddings_config,

apps/shark_studio/web/ui/utils.py

@@ -10,8 +10,8 @@ def resource_path(relative_path):
     return os.path.join(base_path, relative_path)
 
 
-nodlogo_loc = resource_path("logos/nod-logo.png")
-nodicon_loc = resource_path("logos/nod-icon.png")
+amdlogo_loc = resource_path("logos/amd-logo.jpg")
+amdicon_loc = resource_path("logos/amd-icon.jpg")
 
 
 class HSLHue(IntEnum):

apps/shark_studio/web/utils/default_configs.py

@@ -0,0 +1,95 @@
+default_sd_config = r"""{
+  "prompt": [
+    "a photo taken of the front of a super-car drifting on a road near mountains at high speeds with smoke coming off the tires, front angle, front point of view, trees in the mountains of the background, ((sharp focus))"
+  ],
+  "negative_prompt": [
+    "watermark, signature, logo, text, lowres, ((monochrome, grayscale)), blurry, ugly, blur, oversaturated, cropped"
+  ],
+  "sd_init_image": [null],
+  "height": 512,
+  "width": 512,
+  "steps": 50,
+  "strength": 0.8,
+  "guidance_scale": 7.5,
+  "seed": "-1",
+  "batch_count": 1,
+  "batch_size": 1,
+  "scheduler": "EulerDiscrete",
+  "base_model_id": "stabilityai/stable-diffusion-2-1-base",
+  "custom_weights": null,
+  "custom_vae": null,
+  "precision": "fp16",
+  "device": "",
+  "target_triple": "",
+  "ondemand": false,
+  "compiled_pipeline": false,
+  "resample_type": "Nearest Neighbor",
+  "controlnets": {},
+  "embeddings": {}
+}"""
+
+sdxl_30steps = r"""{
+  "prompt": [
+    "a cat under the snow with blue eyes, covered by snow, cinematic style, medium shot, professional photo, animal"
+  ],
+  "negative_prompt": [
+    "watermark, signature, logo, text, lowres, ((monochrome, grayscale)), blurry, ugly, blur, oversaturated, cropped"
+  ],
+  "sd_init_image": [null],
+  "height": 1024,
+  "width": 1024,
+  "steps": 30,
+  "strength": 0.8,
+  "guidance_scale": 7.5,
+  "seed": "-1",
+  "batch_count": 1,
+  "batch_size": 1,
+  "scheduler": "EulerDiscrete",
+  "base_model_id": "stabilityai/stable-diffusion-xl-base-1.0",
+  "custom_weights": null,
+  "custom_vae": null,
+  "precision": "fp16",
+  "device": "",
+  "target_triple": "",
+  "ondemand": false,
+  "compiled_pipeline": true,
+  "resample_type": "Nearest Neighbor",
+  "controlnets": {},
+  "embeddings": {}
+}"""
+
+sdxl_turbo = r"""{
+  "prompt": [
+    "A cat wearing a hat that says 'TURBO' on it. The cat is sitting on a skateboard."
+  ],
+  "negative_prompt": [
+    ""
+  ],
+  "sd_init_image": [null],
+  "height": 512,
+  "width": 512,
+  "steps": 2,
+  "strength": 0.8,
+  "guidance_scale": 0,
+  "seed": "-1",
+  "batch_count": 1,
+  "batch_size": 1,
+  "scheduler": "EulerAncestralDiscrete",
+  "base_model_id": "stabilityai/sdxl-turbo",
+  "custom_weights": null,
+  "custom_vae": null,
+  "precision": "fp16",
+  "device": "",
+  "target_triple": "",
+  "ondemand": false,
+  "compiled_pipeline": true,
+  "resample_type": "Nearest Neighbor",
+  "controlnets": {},
+  "embeddings": {}
+}"""
+
+default_sd_configs = {
+    "default_sd_config.json": default_sd_config,
+    "sdxl-30steps.json": sdxl_30steps,
+    "sdxl-turbo.json": sdxl_turbo,
+}

apps/shark_studio/web/utils/file_utils.py

@@ -11,43 +11,18 @@ checkpoints_filetypes = (
     "*.safetensors",
 )
 
-default_sd_config = r"""{
-  "prompt": [
-    "a photo taken of the front of a super-car drifting on a road near mountains at high speeds with smoke coming off the tires, front angle, front point of view, trees in the mountains of the background, ((sharp focus))"
-  ],
-  "negative_prompt": [
-    "watermark, signature, logo, text, lowres, ((monochrome, grayscale)), blurry, ugly, blur, oversaturated, cropped"
-  ],
-  "sd_init_image": [null],
-  "height": 512,
-  "width": 512,
-  "steps": 50,
-  "strength": 0.8,
-  "guidance_scale": 7.5,
-  "seed": "-1",
-  "batch_count": 1,
-  "batch_size": 1,
-  "scheduler": "EulerDiscrete",
-  "base_model_id": "stabilityai/stable-diffusion-2-1-base",
-  "custom_weights": null,
-  "custom_vae": null,
-  "precision": "fp16",
-  "device": "AMD Radeon RX 7900 XTX => vulkan://0",
-  "ondemand": false,
-  "repeatable_seeds": false,
-  "resample_type": "Nearest Neighbor",
-  "controlnets": {},
-  "embeddings": {}
-}"""
+from apps.shark_studio.web.utils.default_configs import default_sd_configs
 
 
-def write_default_sd_config(path):
-    with open(path, "w") as f:
-        f.write(default_sd_config)
+def write_default_sd_configs(path):
+    for key in default_sd_configs.keys():
+        config_fpath = os.path.join(path, key)
+        with open(config_fpath, "w") as f:
+            f.write(default_sd_configs[key])
 
 
 def safe_name(name):
-    return name.replace("/", "_").replace("-", "_")
+    return name.split("/")[-1].replace("-", "_")
 
 
 def get_path_stem(path):

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

@@ -3,9 +3,8 @@ from pathlib import Path
 from apps.shark_studio.web.utils.file_utils import (
     get_checkpoint_pathfile,
 )
-from apps.shark_studio.api.sd import (
-    sd_model_map,
-)
+from apps.shark_studio.api.sd import EMPTY_SD_MAP as sd_model_map
+
 from apps.shark_studio.modules.schedulers import (
     scheduler_model_map,
 )

apps/shark_studio/web/utils/tmp_configs.py

@@ -17,7 +17,7 @@ def clear_tmp_mlir():
         and filename.endswith(".mlir")
     ]
     for filename in mlir_files:
-        os.remove(shark_tmp + filename)
+        os.remove(os.path.join(shark_tmp, filename))
     print(f"Clearing .mlir temporary files took {time() - cleanup_start:.4f} seconds.")
 
 

dataset/annotation_tool.py

@@ -10,7 +10,7 @@ from utils import get_datasets
 
 shark_root = Path(__file__).parent.parent
 demo_css = shark_root.joinpath("web/demo.css").resolve()
-nodlogo_loc = shark_root.joinpath("web/models/stable_diffusion/logos/nod-logo.png")
+nodlogo_loc = shark_root.joinpath("web/models/stable_diffusion/logos/amd-logo.jpg")
 
 
 with gr.Blocks(title="Dataset Annotation Tool", css=demo_css) as shark_web:

requirements.txt

@@ -1,13 +1,16 @@
--f https://download.pytorch.org/whl/nightly/cpu/torch_nightly.html
--f https://openxla.github.io/iree/pip-release-links.html
+-f https://download.pytorch.org/whl/nightly/cpu
+-f https://iree.dev/pip-release-links.html
 --pre
 
 setuptools
 wheel
 
+
 torch==2.3.0
 shark-turbine @ git+https://github.com/iree-org/iree-turbine.git@main
-turbine-models @ git+https://github.com/nod-ai/SHARK-Turbine.git@main#subdirectory=models
+turbine-models @ git+https://github.com/nod-ai/SHARK-Turbine.git@deprecated-constraints#subdirectory=models
+diffusers @ git+https://github.com/nod-ai/diffusers@0.29.0.dev0-shark
+brevitas @ git+https://github.com/Xilinx/brevitas.git@6695e8df7f6a2c7715b9ed69c4b78157376bb60b
 
 # SHARK Runner
 tqdm
@@ -17,8 +20,6 @@ google-cloud-storage
 
 # Testing
 pytest
-pytest-xdist
-pytest-forked
 Pillow
 parameterized
 
@@ -26,8 +27,10 @@ parameterized
 #accelerate is now required for diffusers import from ckpt.
 accelerate
 scipy
+transformers==4.37.1
+torchsde # Required for Stable Diffusion SDE schedulers.
 ftfy
-gradio==4.19.2
+gradio==4.29.0
 altair
 omegaconf
 # 0.3.2 doesn't have binaries for arm64

setup_venv.ps1

@@ -88,8 +88,8 @@ else {python -m venv .\shark.venv\}
 .\shark.venv\Scripts\activate
 python -m pip install --upgrade pip
 pip install wheel
-pip install -r requirements.txt
-# remove this when windows DLL issues are fixed from LLVM changes
-pip install --force-reinstall https://github.com/openxla/iree/releases/download/candidate-20240326.843/iree_compiler-20240326.843-cp311-cp311-win_amd64.whl https://github.com/openxla/iree/releases/download/candidate-20240326.843/iree_runtime-20240326.843-cp311-cp311-win_amd64.whl
+pip install --pre -r requirements.txt
+pip install --force-reinstall https://github.com/nod-ai/SRT/releases/download/candidate-20240528.279/iree_compiler-20240528.279-cp311-cp311-win_amd64.whl https://github.com/nod-ai/SRT/releases/download/candidate-20240528.279/iree_runtime-20240528.279-cp311-cp311-win_amd64.whl 
+pip install -e .
 
 Write-Host "Source your venv with ./shark.venv/Scripts/activate"

setup_venv.sh

@@ -84,21 +84,7 @@ 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 ${PYTORCH_URL}
-
-if [[ $(uname -s) = 'Linux' && ! -z "${IMPORTER}" ]]; then
-  T_VER=$($PYTHON -m pip show torch | grep Version)
-  T_VER_MIN=${T_VER:14:12}
-  TV_VER=$($PYTHON -m pip show torchvision | grep Version)
-  TV_VER_MAJ=${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/
-  if [ $? -eq 0 ];then
-    echo "Successfully Installed torch + cu118."
-  else
-    echo "Could not install torch + cu118." >&2
-  fi
-fi
+$PYTHON -m pip install --no-warn-conflicts -e . -f ${RUNTIME} -f ${PYTORCH_URL}
 
 if [[ -z "${NO_BREVITAS}" ]]; then
   $PYTHON -m pip install git+https://github.com/Xilinx/brevitas.git@dev

shark/iree_utils/_common.py

@@ -76,6 +76,7 @@ _IREE_DEVICE_MAP = {
     "vulkan": "vulkan",
     "metal": "metal",
     "rocm": "rocm",
+    "hip": "hip",
     "intel-gpu": "level_zero",
 }
 
@@ -94,6 +95,7 @@ _IREE_TARGET_MAP = {
     "vulkan": "vulkan-spirv",
     "metal": "metal",
     "rocm": "rocm",
+    "hip": "rocm",
     "intel-gpu": "opencl-spirv",
 }
 

shark/iree_utils/compile_utils.py

@@ -62,13 +62,16 @@ def get_iree_device_args(device, extra_args=[]):
         from shark.iree_utils.gpu_utils import get_iree_rocm_args
 
         return get_iree_rocm_args(device_num=device_num, extra_args=extra_args)
+    if device == "hip":
+        from shark.iree_utils.gpu_utils import get_iree_rocm_args
+        return get_iree_rocm_args(device_num=device_num, extra_args=extra_args, hip_driver=True)
     return []
 
 def get_iree_target_triple(device):
     args = get_iree_device_args(device)
     for flag in args:
-        if "triple" in flag.split("-"):
-            triple = flag.split("=")
+        if "triple" in flag:
+            triple = flag.split("=")[-1]
             return triple
     return ""
 
@@ -89,9 +92,9 @@ def clean_device_info(raw_device):
         if len(device_id) <= 2:
             device_id = int(device_id)
 
-    if device not in ["rocm", "vulkan"]:
+    if device not in ["hip", "rocm", "vulkan"]:
         device_id = None
-    if device in ["rocm", "vulkan"] and device_id == None:
+    if device in ["hip", "rocm", "vulkan"] and device_id == None:
         device_id = 0
     return device, device_id
 

shark/iree_utils/gpu_utils.py

@@ -52,7 +52,7 @@ def check_rocm_device_arch_in_args(extra_args):
     return None
 
 
-def get_rocm_device_arch(device_num=0, extra_args=[]):
+def get_rocm_device_arch(device_num=0, extra_args=[], hip_driver=False):
     # ROCM Device Arch selection:
     # 1 : User given device arch using `--iree-rocm-target-chip` flag
     # 2 : Device arch from `iree-run-module --dump_devices=rocm` for device on index <device_num>
@@ -68,15 +68,23 @@ def get_rocm_device_arch(device_num=0, extra_args=[]):
     arch_in_device_dump = None
 
     # get rocm arch from iree dump devices
-    def get_devices_info_from_dump(dump):
+    def get_devices_info_from_dump(dump, driver):
         from os import linesep
-
-        dump_clean = list(
-            filter(
-                lambda s: "--device=rocm" in s or "gpu-arch-name:" in s,
-                dump.split(linesep),
+        
+        if driver == "hip":
+            dump_clean = list(
+                filter(
+                    lambda s: "AMD" in s,
+                    dump.split(linesep),
+                )
+            )
+        else:
+            dump_clean = list(
+                filter(
+                    lambda s: f"--device={driver}" in s or "gpu-arch-name:" in s,
+                    dump.split(linesep),
+                )
             )
-        )
         arch_pairs = [
             (
                 dump_clean[i].split("=")[1].strip(),
@@ -87,16 +95,17 @@ def get_rocm_device_arch(device_num=0, extra_args=[]):
         return arch_pairs
 
     dump_device_info = None
+    driver = "hip" if hip_driver else "rocm"
     try:
         dump_device_info = run_cmd(
-            "iree-run-module --dump_devices=rocm", raise_err=True
+            "iree-run-module --dump_devices=" + driver, raise_err=True
         )
     except Exception as e:
-        print("could not execute `iree-run-module --dump_devices=rocm`")
+        print("could not execute `iree-run-module --dump_devices=" + driver + "`")
 
     if dump_device_info is not None:
         device_num = 0 if device_num is None else device_num
-        device_arch_pairs = get_devices_info_from_dump(dump_device_info[0])
+        device_arch_pairs = get_devices_info_from_dump(dump_device_info[0], driver)
         if len(device_arch_pairs) > device_num:  # can find arch in the list
             arch_in_device_dump = device_arch_pairs[device_num][1]
 
@@ -107,24 +116,22 @@ def get_rocm_device_arch(device_num=0, extra_args=[]):
     default_rocm_arch = "gfx1100"
     print(
         "Did not find ROCm architecture from `--iree-rocm-target-chip` flag"
-        "\n or from `iree-run-module --dump_devices=rocm` command."
+        "\n or from `iree-run-module --dump_devices` command."
         f"\nUsing {default_rocm_arch} as ROCm arch for compilation."
     )
     return default_rocm_arch
 
 
 # Get the default gpu args given the architecture.
-def get_iree_rocm_args(device_num=0, extra_args=[]):
+def get_iree_rocm_args(device_num=0, extra_args=[], hip_driver=False):
     ireert.flags.FUNCTION_INPUT_VALIDATION = False
-    rocm_flags = ["--iree-rocm-link-bc=true"]
-
+    rocm_flags = []
     if check_rocm_device_arch_in_args(extra_args) is None:
-        rocm_arch = get_rocm_device_arch(device_num, extra_args)
+        rocm_arch = get_rocm_device_arch(device_num, extra_args, hip_driver=hip_driver)
         rocm_flags.append(f"--iree-rocm-target-chip={rocm_arch}")
 
     return rocm_flags
 
-
 # Some constants taken from cuda.h
 CUDA_SUCCESS = 0
 CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT = 16