Switch most compile flows to use ireec.compile_file. (#1863)

* Switch most compile flows to use ireec.compile_file.

* re-add input type to compile_str path.

* Check if mlir_module exists before checking if it's a path or pyobject.

* Fix some save_dir cases

.gitignore

@@ -193,3 +193,9 @@ 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/

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 [here](https://www.amd.com/en/support/kb/release-notes/rn-rad-win-23-2-1).
+* [AMD RDNA Users] Download the latest driver (23.2.1 is the oldest supported) [here](https://www.amd.com/en/support).
 * [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)
   

apps/language_models/langchain/h2oai_pipeline.py

@@ -20,7 +20,7 @@ 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
+from shark.shark_importer import import_with_fx, save_mlir
 from apps.stable_diffusion.src import args
 
 # Brevitas
@@ -237,7 +237,7 @@ class H2OGPTSHARKModel(torch.nn.Module):
             print(f"[DEBUG] converting torch to linalg")
             run_pipeline_with_repro_report(
                 module,
-                "builtin.module(func.func(torch-unpack-torch-tensor),torch-backend-to-linalg-on-tensors-backend-pipeline)",
+                "builtin.module(func.func(torch-unpack-quant-tensor),func.func(torch-convert-custom-quant-op),torch-backend-to-linalg-on-tensors-backend-pipeline)",
                 description="Lowering Torch Backend IR -> Linalg-on-Tensors Backend IR",
             )
         else:
@@ -256,6 +256,11 @@ 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):

apps/language_models/scripts/llama_ir_conversion_utils.py

@@ -0,0 +1,442 @@
+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

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

apps/language_models/shark_llama_cli.spec

@@ -0,0 +1,94 @@
+# -*- 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/vicuna4.py

@@ -47,18 +47,15 @@ from apps.language_models.src.model_wrappers.vicuna_sharded_model import (
 )
 from apps.language_models.src.model_wrappers.vicuna_model import (
     FirstVicuna,
-    SecondVicuna,
+    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_importer import import_with_fx
 from shark.shark_inference import SharkInference
 
-from brevitas_examples.llm.llm_quant.quantize import quantize_model
-from brevitas_examples.llm.llm_quant.run_utils import get_model_impl
 from transformers.models.llama.configuration_llama import LlamaConfig
 from transformers.models.llama.modeling_llama import (
     LlamaDecoderLayer,

apps/language_models/src/model_wrappers/vicuna_model.py

@@ -1,15 +1,13 @@
 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,
@@ -18,15 +16,24 @@ 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=torch.float32,
+                dtype=self.accumulates,
                 weight_bit_width=weight_bit_width,
                 weight_param_method="stats",
                 weight_scale_precision="float",
@@ -40,7 +47,9 @@ class FirstVicuna(torch.nn.Module):
     def forward(self, input_ids):
         op = self.model(input_ids=input_ids, use_cache=True)
         return_vals = []
-        return_vals.append(op.logits)
+        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])
@@ -48,11 +57,12 @@ class FirstVicuna(torch.nn.Module):
         return tuple(return_vals)
 
 
-class SecondVicuna(torch.nn.Module):
+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,
@@ -64,12 +74,21 @@ class SecondVicuna(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=torch.float32,
+                dtype=self.accumulates,
                 weight_bit_width=weight_bit_width,
                 weight_param_method="stats",
                 weight_scale_precision="float",
@@ -148,8 +167,6 @@ class SecondVicuna(torch.nn.Module):
         i63,
         i64,
     ):
-        # input_ids = input_tuple[0]
-        # input_tuple = torch.unbind(pkv, dim=0)
         token = i0
         past_key_values = (
             (i1, i2),
@@ -282,6 +299,842 @@ class SecondVicuna(torch.nn.Module):
             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:
+            return_vals.append(item[0])
+            return_vals.append(item[1])
+        return tuple(return_vals)
+
+
+class SecondVicuna70B(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,
+        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,
+    ):
+        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,
+            ),
+            (
+                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
+        )
+        return_vals = []
         return_vals.append(op.logits)
         temp_past_key_values = op.past_key_values
         for item in temp_past_key_values:
@@ -298,7 +1151,8 @@ class CombinedModel(torch.nn.Module):
     ):
         super().__init__()
         self.first_vicuna = FirstVicuna(first_vicuna_model_path)
-        self.second_vicuna = SecondVicuna(second_vicuna_model_path)
+        # NOT using this path for 13B currently, hence using `SecondVicuna7B`.
+        self.second_vicuna = SecondVicuna7B(second_vicuna_model_path)
 
     def forward(self, input_ids):
         first_output = self.first_vicuna(input_ids=input_ids)

apps/language_models/src/pipelines/SharkLLMBase.py

@@ -3,7 +3,10 @@ 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

@@ -7,7 +7,7 @@ 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
+from shark.shark_importer import import_with_fx, save_mlir
 from shark.shark_inference import SharkInference
 from transformers import AutoTokenizer, AutoModelForCausalLM
 from transformers.generation import (
@@ -28,7 +28,9 @@ parser = argparse.ArgumentParser(
     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(
     "--precision", "-p", default="fp16", help="fp32, fp16, int8, int4"
 )
@@ -49,7 +51,7 @@ parser.add_argument(
 )
 parser.add_argument(
     "--load_mlir_from_shark_tank",
-    default=False,
+    default=True,
     action=argparse.BooleanOptionalAction,
     help="download precompile mlir from shark tank",
 )
@@ -59,32 +61,52 @@ 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.",
+)
 
 
 class Falcon(SharkLLMBase):
     def __init__(
         self,
         model_name,
-        hf_model_path,
+        hf_model_path="tiiuae/falcon-7b-instruct",
+        hf_auth_token: str = None,
         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.shark_model = self.compile()
         self.src_model = self.get_src_model()
+        self.shark_model = self.compile()
 
     def get_tokenizer(self):
         tokenizer = AutoTokenizer.from_pretrained(
-            self.hf_model_path, trust_remote_code=True
+            self.hf_model_path,
+            trust_remote_code=True,
+            token=self.hf_auth_token,
         )
         tokenizer.padding_side = "left"
         tokenizer.pad_token_id = 11
@@ -92,13 +114,18 @@ class Falcon(SharkLLMBase):
 
     def get_src_model(self):
         print("Loading src model: ", self.model_name)
-        kwargs = {"torch_dtype": torch.float, "trust_remote_code": True}
+        kwargs = {
+            "torch_dtype": torch.float,
+            "trust_remote_code": True,
+            "token": self.hf_auth_token,
+            "device_map": "cpu" if args.device == "cpu" else "cuda:0",
+        }
         falcon_model = AutoModelForCausalLM.from_pretrained(
             self.hf_model_path, **kwargs
         )
         return falcon_model
 
-    def compile_falcon(self):
+    def compile(self):
         if args.use_precompiled_model:
             if not self.falcon_vmfb_path.exists():
                 # Downloading VMFB from shark_tank
@@ -120,37 +147,37 @@ class Falcon(SharkLLMBase):
             if vmfb is not None:
                 return vmfb
 
-        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'}"
-        )
+        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
-            # 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,
+            print(
+                f"[DEBUG] mlir not found at {self.falcon_mlir_path.absolute()}"
             )
-            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 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():
+                    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)
                 )
@@ -170,6 +197,7 @@ class Falcon(SharkLLMBase):
                     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")
@@ -189,35 +217,32 @@ class Falcon(SharkLLMBase):
                 bytecode = bytecode_stream.getvalue()
                 del module
 
-                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_ = 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=bytecode, device=self.device, mlir_dialect="linalg"
+            mlir_module=self.falcon_mlir_path,
+            device=self.device,
+            mlir_dialect="linalg",
         )
         path = shark_module.save_module(
-            self.falcon_vmfb_path.parent.absolute(),
+            self.falcon_vmfb_path,
             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-spirv-index-bits=64",
             ],
+            debug=self.debug,
         )
         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,
@@ -466,11 +491,26 @@ 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",
+        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,
@@ -497,7 +537,11 @@ if __name__ == "__main__":
             prompt = input("Please enter the prompt text: ")
         print("\nPrompt Text: ", prompt)
 
-        res_str = falcon.generate(prompt)
+        prompt_template = f"""A helpful assistant who helps the user with any questions asked.
+        User: {prompt}
+        Assistant:"""
+
+        res_str = falcon.generate(prompt_template)
         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
+from shark.shark_importer import import_with_fx, save_mlir
 import torch
 import torch_mlir
 from torch_mlir.compiler_utils import run_pipeline_with_repro_report
@@ -178,7 +178,7 @@ def load_vmfb(extended_model_name, device, mlir_dialect, extra_args=[]):
 
 
 def compile_module(
-    shark_module, extended_model_name, generate_vmfb, extra_args=[]
+    shark_module, extended_model_name, generate_vmfb, extra_args=[], debug=False,
 ):
     if generate_vmfb:
         vmfb_path = os.path.join(os.getcwd(), extended_model_name + ".vmfb")
@@ -190,7 +190,7 @@ 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
+                os.getcwd(), extended_model_name, extra_args, debug=debug
             )
             shark_module.load_module(path, extra_args=extra_args)
     else:
@@ -199,7 +199,7 @@ def compile_module(
 
 
 def compile_int_precision(
-    model, inputs, precision, device, generate_vmfb, extended_model_name
+    model, inputs, precision, device, generate_vmfb, extended_model_name, debug=False
 ):
     torchscript_module = import_with_fx(
         model,
@@ -219,7 +219,7 @@ def compile_int_precision(
     print(f"[DEBUG] converting torch to linalg")
     run_pipeline_with_repro_report(
         mlir_module,
-        "builtin.module(func.func(torch-unpack-torch-tensor),torch-backend-to-linalg-on-tensors-backend-pipeline)",
+        "builtin.module(func.func(torch-unpack-quant-tensor),func.func(torch-convert-custom-quant-op),torch-backend-to-linalg-on-tensors-backend-pipeline)",
         description="Lowering Torch Backend IR -> Linalg-on-Tensors Backend IR",
     )
     from contextlib import redirect_stdout
@@ -235,6 +235,12 @@ 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"
@@ -251,6 +257,7 @@ def compile_int_precision(
             extended_model_name=extended_model_name,
             generate_vmfb=generate_vmfb,
             extra_args=extra_args,
+            debug=debug,
         ),
         bytecode,
     )
@@ -294,6 +301,7 @@ 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

@@ -32,11 +32,13 @@ 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()
 
@@ -111,7 +113,7 @@ class SharkStableLM(SharkLLMBase):
         shark_module.compile()
 
         path = shark_module.save_module(
-            vmfb_path.parent.absolute(), vmfb_path.stem
+            vmfb_path.parent.absolute(), vmfb_path.stem, debug=self.debug
         )
         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):
+def get_vmfb_from_path(vmfb_path, device, mlir_dialect, device_id=None):
     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):
     print("Loading vmfb from: ", vmfb_path)
     print("Device from get_vmfb_from_path - ", device)
     shark_module = SharkInference(
-        None, device=device, mlir_dialect=mlir_dialect
+        None, device=device, mlir_dialect=mlir_dialect, device_idx=device_id
     )
     shark_module.load_module(vmfb_path)
     print("Successfully loaded vmfb")
@@ -28,7 +28,13 @@ def get_vmfb_from_path(vmfb_path, device, mlir_dialect):
 
 
 def get_vmfb_from_config(
-    shark_container, model, precision, device, vmfb_path, padding=None
+    shark_container,
+    model,
+    precision,
+    device,
+    vmfb_path,
+    padding=None,
+    device_id=None,
 ):
     vmfb_url = (
         f"gs://shark_tank/{shark_container}/{model}_{precision}_{device}"
@@ -37,4 +43,6 @@ 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")
+    return get_vmfb_from_path(
+        vmfb_path, device, "tm_tensor", device_id=device_id
+    )

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 --iree-stream-resource-index-bits=64 --iree-vm-target-index-bits=64 /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 /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 --iree-stream-resource-index-bits=64 --iree-vm-target-index-bits=64 /path/to/input/mlir -o /path/to/output/vmfb
+iree-compile --iree-input-type=none --iree-hal-target-backends=cuda /path/to/input/mlir -o /path/to/output/vmfb
 
 CPU:
-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
+iree-compile --iree-input-type=none --iree-hal-target-backends=llvm-cpu /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.cross_attention import LoRACrossAttnProcessor
+from diffusers.models.attention_processor import LoRAXFormersAttnProcessor
 
 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] = LoRACrossAttnProcessor(
+            lora_attn_procs[name] = LoRAXFormersAttnProcessor(
                 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")
@@ -31,20 +31,20 @@ 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("google_cloud_storage")
 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")
@@ -52,6 +52,7 @@ 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 += [
     ("src/utils/resources/prompts.json", "resources"),
     ("src/utils/resources/model_db.json", "resources"),
@@ -74,7 +75,13 @@ 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("transformers") if "tests" not in x
+    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)
 ]
 hiddenimports += [x for x in collect_submodules("iree") if "tests" not in x]
-hiddenimports += ["iree._runtime", "iree._runtime_libs"]
+hiddenimports += ["iree._runtime", "iree.compiler._mlir_libs._mlir.ir"]

apps/stable_diffusion/src/models/model_wrappers.py

@@ -177,9 +177,11 @@ 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:
@@ -339,7 +341,7 @@ class SharkifyStableDiffusionModel:
         )
         return shark_vae, vae_mlir
 
-    def get_controlled_unet(self):
+    def get_controlled_unet(self, use_large=False):
         class ControlledUnetModel(torch.nn.Module):
             def __init__(
                 self,
@@ -415,6 +417,16 @@ 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,
@@ -437,19 +449,19 @@ class SharkifyStableDiffusionModel:
         shark_controlled_unet, controlled_unet_mlir = compile_through_fx(
             unet,
             inputs,
-            extended_model_name=self.model_name["stencil_unet"],
+            extended_model_name=self.model_name[model_name],
             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="stencil_unet",
+            model_name=model_name,
             precision=self.precision,
             return_mlir=self.return_mlir,
         )
         return shark_controlled_unet, controlled_unet_mlir
 
-    def get_control_net(self):
+    def get_control_net(self, use_large=False):
         class StencilControlNetModel(torch.nn.Module):
             def __init__(
                 self, model_id=self.use_stencil, low_cpu_mem_usage=False
@@ -497,17 +509,34 @@ 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["stencil_adaptor"],
+            extended_model_name=self.model_name[model_name],
             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="stencil_adaptor",
+            model_name=model_name,
             precision=self.precision,
             return_mlir=self.return_mlir,
         )
@@ -681,8 +710,11 @@ class SharkifyStableDiffusionModel:
                 return self.text_encoder(input)[0]
 
         clip_model = CLIPText(low_cpu_mem_usage=self.low_cpu_mem_usage)
-        save_dir = os.path.join(self.sharktank_dir, self.model_name["clip"])
+        save_dir = ""
         if self.debug:
+            save_dir = os.path.join(
+                self.sharktank_dir, self.model_name["clip"]
+            )
             os.makedirs(
                 save_dir,
                 exist_ok=True,
@@ -748,7 +780,7 @@ class SharkifyStableDiffusionModel:
             else:
                 return self.get_unet(use_large=use_large)
         else:
-            return self.get_controlled_unet()
+            return self.get_controlled_unet(use_large=use_large)
 
     def vae_encode(self):
         try:
@@ -847,12 +879,14 @@ class SharkifyStableDiffusionModel:
         except Exception as e:
             sys.exit(e)
 
-    def controlnet(self):
+    def controlnet(self, use_large=False):
         try:
             self.inputs["stencil_adaptor"] = self.get_input_info_for(
                 base_models["stencil_adaptor"]
             )
-            compiled_stencil_adaptor, controlnet_mlir = self.get_control_net()
+            compiled_stencil_adaptor, controlnet_mlir = self.get_control_net(
+                use_large=use_large
+            )
 
             check_compilation(compiled_stencil_adaptor, "Stencil")
             if self.return_mlir:

apps/stable_diffusion/src/pipelines/pipeline_shark_stable_diffusion_img2img.py

@@ -84,13 +84,35 @@ 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
-        image = image.resize((width, height))
+        # Pre-process image
+        if resample_type == "Lanczos":
+            resample_type = Image.LANCZOS
+        elif resample_type == "Nearest Neighbor":
+            resample_type = Image.NEAREST
+        elif resample_type == "Bilinear":
+            resample_type = Image.BILINEAR
+        elif resample_type == "Bicubic":
+            resample_type = Image.BICUBIC
+        elif resample_type == "Adaptive":
+            resample_type = Image.ADAPTIVE
+        elif resample_type == "Antialias":
+            resample_type = Image.ANTIALIAS
+        elif resample_type == "Box":
+            resample_type = Image.BOX
+        elif resample_type == "Affine":
+            resample_type = Image.AFFINE
+        elif resample_type == "Cubic":
+            resample_type = Image.CUBIC
+        else:  # Fallback to Lanczos
+            resample_type = Image.LANCZOS
+
+        image = image.resize((width, height), resample=resample_type)
         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)
@@ -147,6 +169,7 @@ class Image2ImagePipeline(StableDiffusionPipeline):
         cpu_scheduling,
         max_embeddings_multiples,
         use_stencil,
+        resample_type,
     ):
         # prompts and negative prompts must be a list.
         if isinstance(prompts, str):
@@ -186,6 +209,7 @@ 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

@@ -58,6 +58,7 @@ 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:
@@ -68,6 +69,15 @@ 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,
@@ -111,8 +121,12 @@ class StencilPipeline(StableDiffusionPipeline):
         latent_history = [latents]
         text_embeddings = torch.from_numpy(text_embeddings).to(dtype)
         text_embeddings_numpy = text_embeddings.detach().numpy()
-        self.load_unet()
-        self.load_controlnet()
+        if text_embeddings.shape[1] <= self.model_max_length:
+            self.load_unet()
+            self.load_controlnet()
+        else:
+            self.load_unet_512()
+            self.load_controlnet_512()
         for i, t in tqdm(enumerate(total_timesteps)):
             step_start_time = time.time()
             timestep = torch.tensor([t]).to(dtype)
@@ -135,43 +149,82 @@ class StencilPipeline(StableDiffusionPipeline):
                 ).to(dtype)
             else:
                 latent_model_input_1 = latent_model_input
-            control = self.controlnet(
-                "forward",
-                (
-                    latent_model_input_1,
-                    timestep,
-                    text_embeddings,
-                    controlnet_hint,
-                ),
-                send_to_host=False,
-            )
+            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,
+                )
             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.
-            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,
-            )
+
+            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,
+                )
             end_profiling(profile_device)
 
             if cpu_scheduling:
@@ -191,7 +244,9 @@ 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"
 
@@ -218,6 +273,7 @@ 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/utils/sd_annotation.py

@@ -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
 
@@ -281,13 +281,9 @@ def sd_model_annotation(mlir_model, model_name, base_model_id=None):
         if "rdna2" not in args.iree_vulkan_target_triple.split("-")[0]:
             use_winograd = True
             winograd_config_dir = load_winograd_configs()
-            winograd_model = annotate_with_winograd(
+            tuned_model = annotate_with_winograd(
                 mlir_model, winograd_config_dir, model_name
             )
-            lowering_config_dir = load_lower_configs(base_model_id)
-            tuned_model = annotate_with_lower_configs(
-                winograd_model, lowering_config_dir, model_name, use_winograd
-            )
         else:
             tuned_model = mlir_model
     else:

apps/stable_diffusion/src/utils/stable_args.py

@@ -132,6 +132,57 @@ 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=[
+        "Lanczos",
+        "Nearest Neighbor",
+        "Bilinear",
+        "Bicubic",
+        "Adaptive",
+        "Antialias",
+        "Box",
+        "Affine",
+        "Cubic",
+    ],
+    help="The resample type to use when resizing an image before being run "
+    "through stable diffusion.",
+)
+
 ##############################################################################
 # Stable Diffusion Training Params
 ##############################################################################
@@ -407,6 +458,14 @@ 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
 ##############################################################################
@@ -519,6 +578,14 @@ 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,
@@ -574,6 +641,13 @@ p.add_argument(
     help="Flag for enabling rest API.",
 )
 
+p.add_argument(
+    "--debug",
+    default=False,
+    action=argparse.BooleanOptionalAction,
+    help="Flag for enabling debugging log in WebUI.",
+)
+
 p.add_argument(
     "--output_gallery",
     default=True,

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
+from shark.shark_importer import import_with_fx, save_mlir
 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
+from shark.iree_utils.gpu_utils import get_cuda_sm_cc, get_iree_rocm_args
 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
+                os.getcwd(), model_name, extra_args, debug=args.compile_debug
             )
             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"
@@ -168,6 +168,14 @@ 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,
@@ -179,17 +187,22 @@ 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)
 
 
@@ -470,12 +483,25 @@ def get_available_devices():
     set_iree_runtime_flags()
 
     available_devices = []
-    vulkan_devices = get_devices_by_name("vulkan")
+    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.")
     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")
@@ -499,7 +525,10 @@ 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(
@@ -572,7 +601,7 @@ 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=custom_weights,
+        checkpoint_path_or_dict=custom_weights,
         extract_ema=extract_ema,
         from_safetensors=from_safetensors,
         num_in_channels=num_in_channels,
@@ -822,6 +851,8 @@ 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:

apps/stable_diffusion/web/index.py

@@ -1,6 +1,7 @@
 from multiprocessing import Process, freeze_support
 import os
 import sys
+import logging
 
 if sys.platform == "darwin":
     # import before IREE to avoid torch-MLIR library issues
@@ -41,6 +42,8 @@ def launch_app(address):
 
 
 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(","):
@@ -153,9 +156,9 @@ if __name__ == "__main__":
         upscaler_sendto_img2img,
         upscaler_sendto_inpaint,
         upscaler_sendto_outpaint,
-        lora_train_web,
-        model_web,
-        model_config_web,
+        #  lora_train_web,
+        #  model_web,
+        #  model_config_web,
         hf_models,
         modelmanager_sendto_txt2img,
         modelmanager_sendto_img2img,
@@ -247,16 +250,16 @@ 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 (Experimental)", id=8):
+            #  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):
                 stablelm_chat.render()
-            with gr.TabItem(
-                label="Generate Sharding Config (Experimental)", id=9
-            ):
-                model_config_web.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):

apps/stable_diffusion/web/ui/img2img_ui.py

@@ -3,6 +3,7 @@ import torch
 import time
 import gradio as gr
 import PIL
+from math import ceil
 from PIL import Image
 import base64
 from io import BytesIO
@@ -67,6 +68,7 @@ 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,
@@ -245,7 +247,7 @@ def img2img_inf(
             batch_size,
             height,
             width,
-            steps,
+            ceil(steps / strength),
             strength,
             guidance_scale,
             seeds[current_batch],
@@ -255,6 +257,7 @@ 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(
@@ -348,6 +351,7 @@ def img2img_api(
         lora_hf_id="",
         ondemand=False,
         repeatable_seeds=False,
+        resample_type="Lanczos",
     )
 
     # Converts generator type to subscriptable
@@ -432,7 +436,7 @@ 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",
@@ -550,15 +554,6 @@ 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,
@@ -581,11 +576,35 @@ 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=[
+                                    "Lanczos",
+                                    "Nearest Neighbor",
+                                    "Bilinear",
+                                    "Bicubic",
+                                    "Adaptive",
+                                    "Antialias",
+                                    "Box",
+                                    "Affine",
+                                    "Cubic",
+                                ],
+                                label="Resample Type",
+                            )
                         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(
@@ -695,6 +714,7 @@ 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/minigpt4_ui.py

@@ -109,7 +109,7 @@ with gr.Blocks() as minigpt4_web:
     gr.Markdown(description)
 
     with gr.Row():
-        with gr.Column(scale=0.5):
+        with gr.Column():
             image = gr.Image(type="pil")
             upload_button = gr.Button(
                 value="Upload & Start Chat",

apps/stable_diffusion/web/ui/stablelm_ui.py

@@ -8,7 +8,7 @@ from transformers import (
 from apps.stable_diffusion.web.ui.utils import available_devices
 from datetime import datetime as dt
 import json
-import time
+import sys
 
 
 def user(message, history):
@@ -24,12 +24,9 @@ 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",
-    "vicuna4": "TheBloke/vicuna-7B-1.1-HF",
-    "StableLM": "stabilityai/stablelm-tuned-alpha-3b",
 }
 
 # NOTE: Each `model_name` should have its own start message
@@ -43,6 +40,15 @@ start_message = {
         "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": (
+        "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": (
         "System: You are a helpful, respectful and honest assistant. Always answer "
         "as helpfully as possible, while being safe.  Your answers should not "
@@ -52,60 +58,39 @@ start_message = {
         "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"
     ),
-    "vicuna4": (
-        "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):
     system_message = start_message[model_name]
 
-    if model_name in [
-        "StableLM",
-        "vicuna",
-        "vicuna4",
-        "vicuna1p3",
-        "llama2_7b",
-        "llama2_70b",
-    ]:
+    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:]]
+        )
+        msg = f"{B_INST} {B_SYS} {system_message} {E_SYS} {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()
+        msg = system_message + conversation
+        msg = msg.strip()
     return msg
 
 
@@ -149,14 +134,16 @@ def chat(
     model,
     device,
     precision,
+    download_vmfb,
     config_file,
     cli=False,
     progress=gr.Progress(),
 ):
     global past_key_values
     global model_vmfb_key
-
     global vicuna_model
+
+    device_id = None
     model_name, model_path = list(map(str.strip, model.split("=>")))
     if "cuda" in device:
         device = "cuda"
@@ -165,124 +152,125 @@ def chat(
     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")
 
-    new_model_vmfb_key = f"{model_name}#{model_path}#{device}#{precision}"
-    if model_name in [
-        "vicuna",
-        "vicuna4",
-        "vicuna1p3",
-        "codegen",
-        "llama2_7b",
-        "llama2_70b",
-    ]:
-        from apps.language_models.scripts.vicuna import ShardedVicuna
-        from apps.language_models.scripts.vicuna import UnshardedVicuna
-        from apps.stable_diffusion.src import args
+    from apps.language_models.scripts.vicuna import ShardedVicuna
+    from apps.language_models.scripts.vicuna import UnshardedVicuna
+    from apps.stable_diffusion.src import args
 
-        if 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
-            if args.iree_vulkan_target_triple != "":
-                _extra_args.append(
-                    f"-iree-vulkan-target-triple={args.iree_vulkan_target_triple}"
-                )
-
-            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,
-                    precision=precision,
-                    max_num_tokens=max_toks,
-                    extra_args_cmd=_extra_args,
-                )
-                #  else:
-                #      if config_file is not None:
-                #          config_file = open(config_file)
-                #          config_json = json.load(config_file)
-                #          config_file.close()
-                #      else:
-                #          config_json = get_default_config()
-                #      vicuna_model = ShardedVicuna(
-                #          model_name,
-                #          device=device,
-                #          precision=precision,
-                #          config_json=config_json,
-                #      )
-
-        prompt = create_prompt(model_name, history)
-
-        partial_text = ""
-        count = 0
-        start_time = time.time()
-        for text, msg in progress.tqdm(
-            vicuna_model.generate(prompt, cli=cli),
-            desc="generating response",
-        ):
-            count += 1
-            if "formatted" in msg:
-                history[-1][1] = text
-                end_time = time.time()
-                tokens_per_sec = count / (end_time - start_time)
-                yield history, str(
-                    format(tokens_per_sec, ".2f")
-                ) + " tokens/sec"
-            else:
-                partial_text += text + " "
-                history[-1][1] = partial_text
-                yield history, ""
-
-        return history, ""
-
-    # else Model is StableLM
-    global sharkModel
-    from apps.language_models.src.pipelines.stablelm_pipeline import (
-        SharkStableLM,
-    )
-
-    if new_model_vmfb_key != model_vmfb_key:
+    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_new_tokens=512
-        shark_slm = SharkStableLM(
-            model_name
-        )  # pass elements from UI as required
+        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 target triple : {vulkan_target_triple}")
+
+        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.")
 
-    # 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 = ""
-    for new_text in words_list:
-        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
+    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):
@@ -318,6 +306,7 @@ def llm_chat_api(InputData: dict):
         UnshardedVicuna,
     )
 
+    device_id = None
     if vicuna_model == 0:
         if "cuda" in device:
             device = "cuda"
@@ -326,6 +315,7 @@ def llm_chat_api(InputData: dict):
         elif "task" in device:
             device = "cpu-task"
         elif "vulkan" in device:
+            device_id = int(device.split("://")[1])
             device = "vulkan"
         else:
             print("unrecognized device")
@@ -336,6 +326,9 @@ 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
@@ -398,7 +391,7 @@ with gr.Blocks(title="Chatbot") as stablelm_chat:
         )
         model = gr.Dropdown(
             label="Select Model",
-            value=model_choices[4],
+            value=model_choices[0],
             choices=model_choices,
         )
         supported_devices = available_devices
@@ -406,27 +399,31 @@ with gr.Blocks(title="Chatbot") as stablelm_chat:
         # 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)
-        devices = gr.Dropdown(
+        device = gr.Dropdown(
             label="Device",
             value=supported_devices[0]
             if enabled
             else "Only CUDA Supported for now",
             choices=supported_devices,
             interactive=enabled,
-            #  multiselect=True,
+            # multiselect=True,
         )
         precision = gr.Radio(
             label="Precision",
-            value="int8",
+            value="int4",
             choices=[
                 "int4",
                 "int8",
                 "fp16",
             ],
-            visible=True,
+            visible=False,
         )
         tokens_time = gr.Textbox(label="Tokens generated per second")
+        download_vmfb = gr.Checkbox(
+            label="Download vmfb from Shark tank if available",
+            value=True,
+            interactive=True,
+        )
 
     with gr.Row(visible=False):
         with gr.Group():
@@ -458,19 +455,45 @@ with gr.Blocks(title="Chatbot") as stablelm_chat:
     )
 
     submit_event = msg.submit(
-        fn=user, inputs=[msg, chatbot], outputs=[msg, chatbot], queue=False
+        fn=user,
+        inputs=[msg, chatbot],
+        outputs=[msg, chatbot],
+        show_progress=False,
+        queue=False,
     ).then(
         fn=chat,
-        inputs=[system_msg, chatbot, model, devices, precision, config_file],
+        inputs=[
+            system_msg,
+            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], queue=False
+        fn=user,
+        inputs=[msg, chatbot],
+        outputs=[msg, chatbot],
+        show_progress=False,
+        queue=False,
     ).then(
         fn=chat,
-        inputs=[system_msg, chatbot, model, devices, precision, config_file],
+        inputs=[
+            system_msg,
+            chatbot,
+            model,
+            device,
+            precision,
+            download_vmfb,
+            config_file,
+        ],
         outputs=[chatbot, tokens_time],
+        show_progress=False,
         queue=True,
     )
     stop.click(

apps/stable_diffusion/web/ui/txt2img_ui.py

@@ -4,6 +4,7 @@ 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
@@ -26,6 +27,7 @@ from apps.stable_diffusion.src import (
     utils,
     save_output_img,
     prompt_examples,
+    Image2ImagePipeline,
 )
 from apps.stable_diffusion.src.utils import (
     get_generated_imgs_path,
@@ -62,6 +64,11 @@ 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,
@@ -200,6 +207,81 @@ 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
@@ -271,6 +353,11 @@ def txt2img_api(
         lora_hf_id="",
         ondemand=False,
         repeatable_seeds=False,
+        use_hiresfix=False,
+        hiresfix_height=512,
+        hiresfix_width=512,
+        hiresfix_strength=0.6,
+        resample_type="Nearest Neighbor",
     )
 
     # Convert Generator to Subscriptable
@@ -460,6 +547,49 @@ with gr.Blocks(title="Text-to-Image") as txt2img_web:
                             label="Low VRAM",
                             interactive=True,
                         )
+                    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=[
+                                    "Lanczos",
+                                    "Nearest Neighbor",
+                                    "Bilinear",
+                                    "Bicubic",
+                                    "Adaptive",
+                                    "Antialias",
+                                    "Box",
+                                    "Affine",
+                                    "Cubic",
+                                ],
+                                label="Resample Type",
+                            )
+                        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():
                         with gr.Column(scale=3):
                             batch_count = gr.Slider(
@@ -495,16 +625,6 @@ with gr.Blocks(title="Text-to-Image") as txt2img_web:
                         value=available_devices[0],
                         choices=available_devices,
                     )
-                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,
@@ -530,6 +650,18 @@ with gr.Blocks(title="Text-to-Image") as txt2img_web:
                         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")
@@ -565,6 +697,11 @@ 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",

apps/stable_diffusion/web/ui/utils.py

@@ -25,7 +25,7 @@ class Config:
     device: str
     use_lora: str
     use_stencil: str
-    ondemand: str
+    ondemand: str  # should this be expecting a bool instead?
 
 
 custom_model_filetypes = (

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 --iree-stream-resource-index-bits=64 --iree-vm-target-index-bits=64 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  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 --iree-stream-resource-index-bits=64 --iree-vm-target-index-bits=64 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  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 --iree-stream-resource-index-bits=64 --iree-vm-target-index-bits=64 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  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 --iree-stream-resource-index-bits=64 --iree-vm-target-index-bits=64 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  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 --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-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-benchmark-module --module=benchmarks/module_forward_dispatch_${NUM}_vulkan_spirv_fb.vmfb --function=forward --device=vulkan
 ```

inference/CMakeLists.txt

@@ -1,192 +0,0 @@
-# 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/srt 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

@@ -1,100 +0,0 @@
-# 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/srt
-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

@@ -1,39 +0,0 @@
-# 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

@@ -1,30 +0,0 @@
-# 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
 
-# Temorary workaround for transformers/__init__.py.
-path_to_tranformers_hook = Path(
+# Temporary workaround for transformers/__init__.py.
+path_to_transformers_hook = Path(
     get_python_lib()
     + "/_pyinstaller_hooks_contrib/hooks/stdhooks/hook-transformers.py"
 )
-if path_to_tranformers_hook.is_file():
+if path_to_transformers_hook.is_file():
     pass
 else:
-    with open(path_to_tranformers_hook, "w") as f:
+    with open(path_to_transformers_hook, "w") as f:
         f.write("module_collection_mode = 'pyz+py'")
 
 path_to_skipfiles = Path(get_python_lib() + "/torch/_dynamo/skipfiles.py")

requirements-importer-macos.txt

@@ -8,19 +8,8 @@ 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

@@ -9,23 +9,13 @@ tabulate
 tqdm
 
 #iree-compiler  | iree-runtime should already be installed
-iree-tools-tflite
 iree-tools-xla
-iree-tools-tf
 
-# TensorFlow and JAX.
+# Modelling and JAX.
 gin-config
-tf-nightly
-keras
-#tf-models-nightly
-#tensorflow-text-nightly
 transformers
 diffusers
-#tensorflow-probability
 #jax[cpu]
-
-
-# tflitehub dependencies.
 Pillow
 
 # Testing and support.

requirements.txt

@@ -18,13 +18,14 @@ Pillow
 parameterized
 
 # Add transformers, diffusers and scipy since it most commonly used
+tokenizers==0.13.3
 transformers
-diffusers==0.19.3
+diffusers
 #accelerate is now required for diffusers import from ckpt.
 accelerate
 scipy
 ftfy
-gradio
+gradio==3.44.3
 altair
 omegaconf
 # 0.3.2 doesn't have binaries for arm64
@@ -46,4 +47,4 @@ pefile
 pyinstaller
 
 # vicuna quantization
-brevitas @ git+https://github.com/Xilinx/brevitas.git@dev
+brevitas @ git+https://github.com/Xilinx/brevitas.git@56edf56a3115d5ac04f19837b388fd7d3b1ff7ea

setup_venv.sh

@@ -130,14 +130,13 @@ 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/cpu/
 
-if [[ $(uname -s) = 'Linux' && ! -z "${BENCHMARK}" ]]; then
+if [[ $(uname -s) = 'Linux' && ! -z "${IMPORTER}" ]]; then
   T_VER=$($PYTHON -m pip show torch | grep Version)
-  TORCH_VERSION=${T_VER:9:17}
+  T_VER_MIN=${T_VER:14:12}
   TV_VER=$($PYTHON -m pip show torchvision | grep Version)
-  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-${TV_VERSION}%2Bcu118-cp311-cp311-linux_x86_64.whl
+  TV_VER_MAJ=${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
@@ -146,7 +145,7 @@ if [[ $(uname -s) = 'Linux' && ! -z "${BENCHMARK}" ]]; then
 fi
 
 if [[ -z "${NO_BREVITAS}" ]]; then
-  $PYTHON -m pip install git+https://github.com/Xilinx/brevitas.git@llm
+  $PYTHON -m pip install git+https://github.com/Xilinx/brevitas.git@dev
 fi
 
 if [[ -z "${CONDA_PREFIX}" && "$SKIP_VENV" != "1" ]]; then

shark/iree_utils/_common.py

@@ -52,6 +52,8 @@ def iree_device_map(device):
     )
     if len(uri_parts) == 1:
         return iree_driver
+    elif "rocm" in uri_parts:
+        return "rocm"
     else:
         return f"{iree_driver}://{uri_parts[1]}"
 
@@ -63,7 +65,6 @@ def get_supported_device_list():
 _IREE_DEVICE_MAP = {
     "cpu": "local-task",
     "cpu-task": "local-task",
-    "AMD-AIE": "local-task",
     "cpu-sync": "local-sync",
     "cuda": "cuda",
     "vulkan": "vulkan",
@@ -82,7 +83,6 @@ def iree_target_map(device):
 _IREE_TARGET_MAP = {
     "cpu": "llvm-cpu",
     "cpu-task": "llvm-cpu",
-    "AMD-AIE": "llvm-cpu",
     "cpu-sync": "llvm-cpu",
     "cuda": "cuda",
     "vulkan": "vulkan",
@@ -121,7 +121,10 @@ def check_device_drivers(device):
         return False
     elif device == "rocm":
         try:
-            subprocess.check_output("rocminfo")
+            if sys.platform == "win32":
+                subprocess.check_output("hipinfo")
+            else:
+                subprocess.check_output("rocminfo")
         except Exception:
             return True
 

shark/iree_utils/benchmark_utils.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import iree._runtime.scripts.iree_benchmark_module as benchmark_module
 from shark.iree_utils._common import run_cmd, iree_device_map
 from shark.iree_utils.cpu_utils import get_cpu_count
 import numpy as np
@@ -102,15 +101,13 @@ def build_benchmark_args_non_tensor_input(
     and whether it is training or not.
     Outputs: string that execute benchmark-module on target model.
     """
-    path = benchmark_module.__path__[0]
+    path = os.path.join(os.environ["VIRTUAL_ENV"], "bin")
     if platform.system() == "Windows":
-        benchmarker_path = os.path.join(
-            path, "..", "..", "iree-benchmark-module.exe"
-        )
+        benchmarker_path = os.path.join(path, "iree-benchmark-module.exe")
+        time_extractor = None
     else:
-        benchmarker_path = os.path.join(
-            path, "..", "..", "iree-benchmark-module"
-        )
+        benchmarker_path = os.path.join(path, "iree-benchmark-module")
+        time_extractor = "| awk 'END{{print $2 $3}}'"
     benchmark_cl = [benchmarker_path, f"--module={input_file}"]
     # TODO: The function named can be passed as one of the args.
     if function_name:
@@ -135,7 +132,7 @@ def run_benchmark_module(benchmark_cl):
     benchmark_path = benchmark_cl[0]
     assert os.path.exists(
         benchmark_path
-    ), "Cannot find benchmark_module, Please contact SHARK maintainer on discord."
+    ), "Cannot find iree_benchmark_module, Please contact SHARK maintainer on discord."
     bench_stdout, bench_stderr = run_cmd(" ".join(benchmark_cl))
     try:
         regex_split = re.compile("(\d+[.]*\d*)(  *)([a-zA-Z]+)")

shark/iree_utils/compile_utils.py

@@ -46,7 +46,7 @@ def get_iree_device_args(device, extra_args=[]):
     if device_uri[0] == "cpu":
         from shark.iree_utils.cpu_utils import get_iree_cpu_args
 
-        data_tiling_flag = ["--iree-flow-enable-data-tiling"]
+        data_tiling_flag = ["--iree-opt-data-tiling"]
         u_kernel_flag = ["--iree-llvmcpu-enable-microkernels"]
         stack_size_flag = ["--iree-llvmcpu-stack-allocation-limit=256000"]
 
@@ -84,7 +84,7 @@ def get_iree_frontend_args(frontend):
     elif frontend in ["tensorflow", "tf", "mhlo", "stablehlo"]:
         return [
             "--iree-llvmcpu-target-cpu-features=host",
-            "--iree-flow-demote-i64-to-i32",
+            "--iree-input-demote-i64-to-i32",
         ]
     else:
         # Frontend not found.
@@ -92,14 +92,27 @@ def get_iree_frontend_args(frontend):
 
 
 # Common args to be used given any frontend or device.
-def get_iree_common_args():
-    return [
-        "--iree-stream-resource-index-bits=64",
+def get_iree_common_args(debug=False):
+    common_args = [
         "--iree-stream-resource-max-allocation-size=4294967295",
-        "--iree-vm-target-index-bits=64",
         "--iree-vm-bytecode-module-strip-source-map=true",
         "--iree-util-zero-fill-elided-attrs",
     ]
+    if debug == True:
+        common_args.extend(
+            [
+                "--iree-opt-strip-assertions=false",
+                "--verify=true",
+            ]
+        )
+    else:
+        common_args.extend(
+            [
+                "--iree-opt-strip-assertions=true",
+                "--verify=false",
+            ]
+        )
+    return common_args
 
 
 # Args that are suitable only for certain models or groups of models.
@@ -278,14 +291,17 @@ def compile_module_to_flatbuffer(
     model_config_path,
     extra_args,
     model_name="None",
+    debug=False,
+    compile_str=False,
 ):
     # Setup Compile arguments wrt to frontends.
-    input_type = ""
+    input_type = "auto"
     args = get_iree_frontend_args(frontend)
     args += get_iree_device_args(device, extra_args)
-    args += get_iree_common_args()
+    args += get_iree_common_args(debug=debug)
     args += get_model_specific_args()
     args += extra_args
+    args += shark_args.additional_compile_args
 
     if frontend in ["tensorflow", "tf"]:
         input_type = "auto"
@@ -296,10 +312,7 @@ def compile_module_to_flatbuffer(
     elif frontend in ["tm_tensor"]:
         input_type = ireec.InputType.TM_TENSOR
 
-    # TODO: make it simpler.
-    # Compile according to the input type, else just try compiling.
-    if input_type != "":
-        # Currently for MHLO/TOSA.
+    if compile_str:
         flatbuffer_blob = ireec.compile_str(
             module,
             target_backends=[iree_target_map(device)],
@@ -307,9 +320,10 @@ def compile_module_to_flatbuffer(
             input_type=input_type,
         )
     else:
-        # Currently for Torch.
-        flatbuffer_blob = ireec.compile_str(
+        assert os.path.isfile(module)
+        flatbuffer_blob = ireec.compile_file(
             module,
+            input_type=input_type,
             target_backends=[iree_target_map(device)],
             extra_args=args,
         )
@@ -343,7 +357,8 @@ def load_vmfb_using_mmap(
     flatbuffer_blob_or_path, device: str, device_idx: int = None
 ):
     print(f"Loading module {flatbuffer_blob_or_path}...")
-
+    if "rocm" in device:
+        device = "rocm"
     with DetailLogger(timeout=2.5) as dl:
         # First get configs.
         if device_idx is not None:
@@ -388,6 +403,11 @@ def load_vmfb_using_mmap(
             dl.log(f"mmap {flatbuffer_blob_or_path}")
             ctx = ireert.SystemContext(config=config)
             dl.log(f"ireert.SystemContext created")
+            if "vulkan" in device:
+                # Vulkan pipeline creation consumes significant amount of time.
+                print(
+                    "\tCompiling Vulkan shaders. This may take a few minutes."
+                )
             ctx.add_vm_module(mmaped_vmfb)
             dl.log(f"module initialized")
             mmaped_vmfb = getattr(ctx.modules, mmaped_vmfb.name)
@@ -410,10 +430,18 @@ def get_iree_compiled_module(
     extra_args: list = [],
     device_idx: int = None,
     mmap: bool = False,
+    debug: bool = False,
+    compile_str: bool = False,
 ):
     """Given a module returns the compiled .vmfb and configs"""
     flatbuffer_blob = compile_module_to_flatbuffer(
-        module, device, frontend, model_config_path, extra_args
+        module,
+        device,
+        frontend,
+        model_config_path,
+        extra_args,
+        debug,
+        compile_str,
     )
     temp_file_to_unlink = None
     # TODO: Currently mmap=True control flow path has been switched off for mmap.
@@ -469,10 +497,18 @@ def export_iree_module_to_vmfb(
     model_config_path: str = None,
     module_name: str = None,
     extra_args: list = [],
+    debug: bool = False,
+    compile_str: bool = False,
 ):
     # Compiles the module given specs and saves it as .vmfb file.
     flatbuffer_blob = compile_module_to_flatbuffer(
-        module, device, mlir_dialect, model_config_path, extra_args
+        module,
+        device,
+        mlir_dialect,
+        model_config_path,
+        extra_args,
+        debug,
+        compile_str,
     )
     if module_name is None:
         device_name = (
@@ -480,9 +516,9 @@ def export_iree_module_to_vmfb(
         )
         module_name = f"{mlir_dialect}_{device_name}"
     filename = os.path.join(directory, module_name + ".vmfb")
-    print(f"Saved vmfb in {filename}.")
     with open(filename, "wb") as f:
         f.write(flatbuffer_blob)
+    print(f"Saved vmfb in {filename}.")
     return filename
 
 

shark/iree_utils/gpu_utils.py

@@ -17,6 +17,7 @@
 import functools
 import iree.runtime as ireert
 import ctypes
+import sys
 from shark.parser import shark_args
 
 
@@ -42,21 +43,51 @@ def get_iree_gpu_args():
 @functools.cache
 def get_iree_rocm_args():
     ireert.flags.FUNCTION_INPUT_VALIDATION = False
-    # get arch from rocminfo.
+    # get arch from hipinfo.
+    import os
     import re
     import subprocess
 
-    rocm_arch = re.match(
-        r".*(gfx\w+)",
-        subprocess.check_output(
-            "rocminfo | grep -i 'gfx'", shell=True, text=True
-        ),
-    ).group(1)
-    print(f"Found rocm arch {rocm_arch}...")
+    if sys.platform == "win32":
+        if "HIP_PATH" in os.environ:
+            rocm_path = os.environ["HIP_PATH"]
+            print(f"Found a ROCm installation at {rocm_path}.")
+        else:
+            print("Failed to find ROCM_PATH. Defaulting to C:\\AMD\\ROCM\\5.5")
+            rocm_path = "C:\\AMD\\ROCM\\5.5"
+    else:
+        if "ROCM_PATH" in os.environ:
+            rocm_path = os.environ["ROCM_PATH"]
+            print(f"Found a ROCm installation at {rocm_path}.")
+        else:
+            print("Failed to find ROCM_PATH. Defaulting to /opt/rocm")
+            rocm_path = "/opt/rocm/"
+
+    try:
+        if sys.platform == "win32":
+            rocm_arch = re.search(
+                r"gfx\d{3,}",
+                subprocess.check_output("hipinfo", shell=True, text=True),
+            ).group(0)
+        else:
+            rocm_arch = re.match(
+                r".*(gfx\w+)",
+                subprocess.check_output(
+                    "rocminfo | grep -i 'gfx'", shell=True, text=True
+                ),
+            ).group(1)
+        print(f"Found rocm arch {rocm_arch}...")
+    except:
+        print(
+            "Failed to find ROCm architecture from hipinfo / rocminfo. Defaulting to gfx1100."
+        )
+        rocm_arch = "gfx1100"
+
+    bc_path = os.path.join(rocm_path, "amdgcn", "bitcode")
     return [
         f"--iree-rocm-target-chip={rocm_arch}",
         "--iree-rocm-link-bc=true",
-        "--iree-rocm-bc-dir=/opt/rocm/amdgcn/bitcode",
+        f"--iree-rocm-bc-dir={bc_path}",
     ]
 
 

shark/iree_utils/vulkan_target_env_utils.py

@@ -57,11 +57,8 @@ def get_version(triple):
 @functools.cache
 def get_extensions(triple):
     def make_ext_list(ext_list):
-        res = ""
-        for e in ext_list:
-            res += e + ", "
-        res = f"[{res[:-2]}]"
-        return res
+        res = ", ".join(ext_list)
+        return f"[{res}]"
 
     arch, product, os = triple
     if arch == "m1":
@@ -119,7 +116,7 @@ def get_extensions(triple):
     ]
 
     if get_vendor(triple) == "NVIDIA" or arch == "rdna3":
-        ext.append("VK_NV_cooperative_matrix")
+        ext.append("VK_KHR_cooperative_matrix")
     if get_vendor(triple) == ["NVIDIA", "AMD", "Intel"]:
         ext.append("VK_KHR_shader_integer_dot_product")
     return make_ext_list(ext_list=ext)
@@ -247,7 +244,7 @@ def get_vulkan_target_capabilities(triple):
         if arch == "rdna3":
             # TODO: Get scope value
             cap["coopmatCases"] = [
-                "mSize = 16, nSize = 16, kSize = 16, aType = f16, bType = f16, cType = f16, resultType = f16, scope = #vk.scope<Subgroup>"
+                "mSize = 16, nSize = 16, kSize = 16, aType = f16, bType = f16, cType = f16, resultType = f16, accSat = false, scope = #vk.scope<Subgroup>"
             ]
 
         if product == "rx5700xt":
@@ -468,9 +465,9 @@ def get_vulkan_target_capabilities(triple):
         cap["variablePointersStorageBuffer"] = True
 
         cap["coopmatCases"] = [
-            "mSize = 8, nSize = 8, kSize = 32, aType = i8, bType = i8, cType = i32, resultType = i32, scope = #vk.scope<Subgroup>",
-            "mSize = 16, nSize = 16, kSize = 16, aType = f16, bType = f16, cType = f16, resultType = f16, scope = #vk.scope<Subgroup>",
-            "mSize = 16, nSize = 16, kSize = 16, aType = f16, bType = f16, cType = f32, resultType = f32, scope = #vk.scope<Subgroup>",
+            "mSize = 8, nSize = 8, kSize = 32, aType = i8, bType = i8, cType = i32, resultType = i32, accSat = false, scope = #vk.scope<Subgroup>",
+            "mSize = 16, nSize = 16, kSize = 16, aType = f16, bType = f16, cType = f16, resultType = f16, accSat = false, scope = #vk.scope<Subgroup>",
+            "mSize = 16, nSize = 16, kSize = 16, aType = f16, bType = f16, cType = f32, resultType = f32, accSat = false, scope = #vk.scope<Subgroup>",
         ]
 
     elif arch == "adreno":
@@ -531,7 +528,7 @@ def get_vulkan_target_capabilities(triple):
                 cmc = ""
                 for case in v:
                     cmc += f"#vk.coop_matrix_props<{case}>, "
-                res += f"cooperativeMatrixPropertiesNV = [{cmc[:-2]}], "
+                res += f"cooperativeMatrixPropertiesKHR = [{cmc[:-2]}], "
             else:
                 res += f"{k} = {get_comma_sep_str(v)}, "
         else:

shark/iree_utils/vulkan_utils.py

@@ -23,11 +23,19 @@ from shark.iree_utils.vulkan_target_env_utils import get_vulkan_target_env_flag
 from shark.parser import shark_args
 
 
+@functools.cache
+def get_all_vulkan_devices():
+    from iree.runtime import get_driver
+
+    driver = get_driver("vulkan")
+    device_list_src = driver.query_available_devices()
+    device_list_src.sort(key=lambda d: d["path"])
+    return [d["name"] for d in device_list_src]
+
+
 @functools.cache
 def get_vulkan_device_name(device_num=0):
-    vulkaninfo_dump, _ = run_cmd("vulkaninfo")
-    vulkaninfo_dump = vulkaninfo_dump.split(linesep)
-    vulkaninfo_list = [s.strip() for s in vulkaninfo_dump if "deviceName" in s]
+    vulkaninfo_list = get_all_vulkan_devices()
     if len(vulkaninfo_list) == 0:
         raise ValueError("No device name found in VulkanInfo!")
     if len(vulkaninfo_list) > 1:
@@ -111,6 +119,8 @@ def get_vulkan_target_triple(device_name):
     # Windows: AMD Radeon RX 7900 XTX
     elif all(x in device_name for x in ("RX", "7900")):
         triple = f"rdna3-7900-{system_os}"
+    elif all(x in device_name for x in ("Radeon", "780M")):
+        triple = f"rdna3-780m-{system_os}"
     elif all(x in device_name for x in ("AMD", "PRO", "W7900")):
         triple = f"rdna3-w7900-{system_os}"
     elif any(x in device_name for x in ("AMD", "Radeon")):
@@ -178,9 +188,7 @@ def get_iree_vulkan_args(device_num=0, extra_args=[]):
 @functools.cache
 def get_iree_vulkan_runtime_flags():
     vulkan_runtime_flags = [
-        f"--vulkan_large_heap_block_size={shark_args.vulkan_large_heap_block_size}",
         f"--vulkan_validation_layers={'true' if shark_args.vulkan_validation_layers else 'false'}",
-        f"--vulkan_vma_allocator={'true' if shark_args.vulkan_vma_allocator else 'false'}",
     ]
     return vulkan_runtime_flags
 

shark/parser.py

@@ -14,8 +14,21 @@
 
 import argparse
 import os
+import shlex
 import subprocess
 
+
+class SplitStrToListAction(argparse.Action):
+    def __init__(self, option_strings, dest, *args, **kwargs):
+        super(SplitStrToListAction, self).__init__(
+            option_strings=option_strings, dest=dest, *args, **kwargs
+        )
+
+    def __call__(self, parser, namespace, values, option_string=None):
+        del parser, option_string
+        setattr(namespace, self.dest, shlex.split(values[0]))
+
+
 parser = argparse.ArgumentParser(description="SHARK runner.")
 
 parser.add_argument(
@@ -24,6 +37,13 @@ parser.add_argument(
     default="cpu",
     help="Device on which shark_runner runs. options are cpu, cuda, and vulkan",
 )
+parser.add_argument(
+    "--additional_compile_args",
+    default=list(),
+    nargs=1,
+    action=SplitStrToListAction,
+    help="Additional arguments to pass to the compiler. These are appended as the last arguments.",
+)
 parser.add_argument(
     "--enable_tf32",
     type=bool,
@@ -133,13 +153,6 @@ parser.add_argument(
     help="Profiles vulkan device and collects the .rdc info.",
 )
 
-parser.add_argument(
-    "--vulkan_large_heap_block_size",
-    default="2073741824",
-    help="Flag for setting VMA preferredLargeHeapBlockSize for "
-    "vulkan device, default is 4G.",
-)
-
 parser.add_argument(
     "--vulkan_validation_layers",
     default=False,
@@ -147,11 +160,4 @@ parser.add_argument(
     help="Flag for disabling vulkan validation layers when benchmarking.",
 )
 
-parser.add_argument(
-    "--vulkan_vma_allocator",
-    default=False,
-    action=argparse.BooleanOptionalAction,
-    help="Flag for enabling / disabling Vulkan VMA Allocator.",
-)
-
 shark_args, unknown = parser.parse_known_args()

shark/shark_benchmark_runner.py

@@ -84,6 +84,13 @@ class SharkBenchmarkRunner(SharkRunner):
         self.extra_args = extra_args
         self.import_args = {}
         self.temp_file_to_unlink = None
+        if not os.path.isfile(mlir_module):
+            print(
+                "Warning: Initializing SharkRunner with a mlir string/bytecode object will duplicate the model in RAM at compile time. To avoid this, initialize SharkInference with a path to a MLIR module on your hard disk instead."
+            )
+            self.compile_str = True
+        else:
+            self.compile_str = False
         SharkRunner.__init__(
             self,
             mlir_module,
@@ -98,6 +105,7 @@ class SharkBenchmarkRunner(SharkRunner):
             ".",
             self.mlir_dialect,
             extra_args=self.extra_args,
+            compile_str=self.compile_str,
         )
         params = load_flatbuffer(
             self.vmfb_file,

shark/shark_compile.py

@@ -1,7 +1,7 @@
 import os
 import tempfile
 from shark.shark_inference import SharkInference
-from shark.shark_importer import import_with_fx
+from shark.shark_importer import import_with_fx, save_mlir
 import torch
 import torch_mlir
 from torch_mlir.compiler_utils import run_pipeline_with_repro_report
@@ -115,7 +115,7 @@ def compile_int_precision(
     print(f"[DEBUG] converting torch to linalg")
     run_pipeline_with_repro_report(
         mlir_module,
-        "builtin.module(func.func(torch-unpack-torch-tensor),torch-backend-to-linalg-on-tensors-backend-pipeline)",
+        "builtin.module(func.func(torch-unpack-quant-tensor),func.func(torch-convert-custom-quant-op),torch-backend-to-linalg-on-tensors-backend-pipeline)",
         description="Lowering Torch Backend IR -> Linalg-on-Tensors Backend IR",
     )
     from contextlib import redirect_stdout
@@ -130,10 +130,17 @@ def compile_int_precision(
     mlir_module = mlir_module.encode("UTF-8")
     mlir_module = BytesIO(mlir_module)
     bytecode = mlir_module.read()
+    bytecode_path = os.path.join(
+        os.getcwd(), f"{extended_model_name}_linalg.mlirbc"
+    )
+    with open(bytecode_path, "wb") as f:
+        f.write(bytecode)
+    del bytecode
+    del mlir_module
     print(f"Elided IR written for {extended_model_name}")
-    return bytecode
+    return bytecode_path
     shark_module = SharkInference(
-        mlir_module=bytecode, device=device, mlir_dialect="tm_tensor"
+        mlir_module=bytecode_path, device=device, mlir_dialect="tm_tensor"
     )
     extra_args = [
         "--iree-hal-dump-executable-sources-to=ies",
@@ -148,7 +155,7 @@ def compile_int_precision(
             generate_vmfb=generate_vmfb,
             extra_args=extra_args,
         ),
-        bytecode,
+        bytecode_path,
     )
 
 
@@ -201,7 +208,7 @@ def shark_compile_through_fx(
         ]
     else:
         (
-            mlir_module,
+            bytecode,
             _,
         ) = import_with_fx(
             model=model,
@@ -212,6 +219,11 @@ def shark_compile_through_fx(
             model_name=extended_model_name,
             save_dir=save_dir,
         )
+        mlir_module = save_mlir(
+            mlir_module=bytecode,
+            model_name=extended_model_name,
+            mlir_dialect=mlir_dialect,
+        )
 
     shark_module = SharkInference(
         mlir_module,

shark/shark_downloader.py

@@ -275,11 +275,11 @@ def download_model(
     model_dir = os.path.join(WORKDIR, model_dir_name)
     tuned_str = "" if tuned is None else "_" + tuned
     suffix = f"{dyn_str}_{frontend}{tuned_str}.mlir"
-    filename = os.path.join(model_dir, model_name + suffix)
+    mlir_filename = os.path.join(model_dir, model_name + suffix)
     print(
-        f"Verifying that model artifacts were downloaded successfully to {filename}..."
+        f"Verifying that model artifacts were downloaded successfully to {mlir_filename}..."
     )
-    if not os.path.exists(filename):
+    if not os.path.exists(mlir_filename):
         from tank.generate_sharktank import gen_shark_files
 
         print(
@@ -287,13 +287,11 @@ def download_model(
         )
         gen_shark_files(model_name, frontend, WORKDIR, import_args)
 
-    assert os.path.exists(filename), f"MLIR not found at {filename}"
-    with open(filename, mode="rb") as f:
-        mlir_file = f.read()
+    assert os.path.exists(mlir_filename), f"MLIR not found at {mlir_filename}"
     function_name = str(np.load(os.path.join(model_dir, "function_name.npy")))
     inputs = np.load(os.path.join(model_dir, "inputs.npz"))
     golden_out = np.load(os.path.join(model_dir, "golden_out.npz"))
 
     inputs_tuple = tuple([inputs[key] for key in inputs])
     golden_out_tuple = tuple([golden_out[key] for key in golden_out])
-    return mlir_file, function_name, inputs_tuple, golden_out_tuple
+    return mlir_filename, function_name, inputs_tuple, golden_out_tuple

shark/shark_eager/shark_eager.py

@@ -1,6 +1,6 @@
 from typing import Any, Dict, List, Tuple
 from collections import defaultdict
-from shark.shark_importer import import_with_fx
+from shark.shark_importer import import_with_fx, save_mlir
 import torchvision.models as models
 import copy
 import io
@@ -20,10 +20,16 @@ def shark_backend(fx_g: torch.fx.GraphModule, inputs, device: str = "cpu"):
     bytecode_stream = io.BytesIO()
     mlir_module.operation.write_bytecode(bytecode_stream)
     bytecode = bytecode_stream.getvalue()
+    bytecode_path = save_mlir(
+        bytecode,
+        model_name="shark_eager_module",
+        frontend="torch",
+        mlir_dialect="tm_tensor",
+    )
     from shark.shark_inference import SharkInference
 
     shark_module = SharkInference(
-        mlir_module=bytecode,
+        mlir_module=bytecode_path,
         device=device,
         mlir_dialect="tm_tensor",
     )

shark/shark_generate_model_config.py

@@ -3,8 +3,8 @@ import json
 import numpy as np
 
 import torch_mlir
-from iree.compiler import compile_str
-from shark.shark_importer import import_with_fx, get_f16_inputs
+from iree.compiler import compile_file
+from shark.shark_importer import import_with_fx, get_f16_inputs, save_mlir
 
 
 class GenerateConfigFile:
@@ -54,9 +54,15 @@ class GenerateConfigFile:
             verbose=False,
         )
         module = module.operation.get_asm(large_elements_limit=4)
+        module_file = save_mlir(
+            module,
+            model_name="module_pre_split",
+            frontend="torch",
+            mlir_dialect="linalg",
+        )
         compiled_module_str = str(
-            compile_str(
-                str(module),
+            compile_file(
+                module_file,
                 target_backends=[backend],
                 extra_args=[
                     "--compile-to=flow",
@@ -138,7 +144,7 @@ if __name__ == "__main__":
     firstVicunaCompileInput = (compilation_input_ids,)
     from apps.language_models.src.model_wrappers.vicuna_model import (
         FirstVicuna,
-        SecondVicuna,
+        SecondVicuna7B,
         CombinedModel,
     )
 

shark/shark_importer.py

@@ -451,6 +451,65 @@ def transform_fx(fx_g, quantized=False):
     fx_g.graph.lint()
 
 
+def gptq_transforms(fx_g):
+    import torch
+
+    for node in fx_g.graph.nodes:
+        if node.op == "call_function":
+            if node.target in [
+                torch.ops.aten.arange,
+                torch.ops.aten.empty,
+                torch.ops.aten.ones,
+                torch.ops.aten._to_copy,
+            ]:
+                if node.kwargs.get("device") == torch.device(device="cuda:0"):
+                    updated_kwargs = node.kwargs.copy()
+                    updated_kwargs["device"] = torch.device(device="cpu")
+                    node.kwargs = updated_kwargs
+
+            if node.target in [
+                torch.ops.aten._to_copy,
+            ]:
+                if node.kwargs.get("dtype") == torch.bfloat16:
+                    updated_kwargs = node.kwargs.copy()
+                    updated_kwargs["dtype"] = torch.float16
+                    node.kwargs = updated_kwargs
+
+            # Inputs of aten.native_layer_norm should be upcasted to fp32.
+            if node.target in [torch.ops.aten.native_layer_norm]:
+                with fx_g.graph.inserting_before(node):
+                    new_node_arg0 = fx_g.graph.call_function(
+                        torch.ops.prims.convert_element_type,
+                        args=(node.args[0], torch.float32),
+                        kwargs={},
+                    )
+                    node.args = (
+                        new_node_arg0,
+                        node.args[1],
+                        node.args[2],
+                        node.args[3],
+                        node.args[4],
+                    )
+
+            # Downcasting the result of native_layer_norm back to fp16.
+            if node.name.startswith("getitem"):
+                with fx_g.graph.inserting_before(node):
+                    if node.args[0].target in [
+                        torch.ops.aten.native_layer_norm
+                    ]:
+                        new_node = fx_g.graph.call_function(
+                            torch.ops.aten._to_copy,
+                            args=(node,),
+                            kwargs={"dtype": torch.float32},
+                        )
+                        node.append(new_node)
+                        node.replace_all_uses_with(new_node)
+                        new_node.args = (node,)
+                        new_node.kwargs = {"dtype": torch.float32}
+
+    fx_g.graph.lint()
+
+
 # Doesn't replace the None type.
 def change_fx_graph_return_to_tuple(fx_g):
     for node in fx_g.graph.nodes:
@@ -504,27 +563,12 @@ def import_with_fx(
     is_dynamic=False,
     tracing_required=False,
     precision="fp32",
+    is_gptq=False,
 ):
     import torch
     from torch.fx.experimental.proxy_tensor import make_fx
     from torch._decomp import get_decompositions
     from typing import List
-    from brevitas_examples.llm.llm_quant.export import (
-        block_quant_layer_level_manager,
-    )
-    from brevitas_examples.llm.llm_quant.export import (
-        brevitas_layer_export_mode,
-    )
-    from brevitas_examples.llm.llm_quant.sharded_mlir_group_export import (
-        LinearWeightBlockQuantHandlerFwd,
-    )
-    from brevitas_examples.llm.llm_quant.export import replace_call_fn_target
-    from brevitas_examples.llm.llm_quant.sharded_mlir_group_export import (
-        matmul_rhs_group_quant_placeholder,
-    )
-    from brevitas.backport.fx.experimental.proxy_tensor import (
-        make_fx as brevitas_make_fx,
-    )
 
     golden_values = None
     if debug:
@@ -596,8 +640,30 @@ def import_with_fx(
         torch.ops.aten.native_layer_norm,
         torch.ops.aten.masked_fill.Tensor,
         torch.ops.aten.masked_fill.Scalar,
+        torch.ops.aten._scaled_dot_product_flash_attention.default,
+        torch.ops.aten.index_add,
+        torch.ops.aten.index_add_,
     ]
-    if precision in ["int4", "int8"]:
+    if precision in ["int4", "int8"] and not is_gptq:
+        from brevitas_examples.llm.llm_quant.export import (
+            block_quant_layer_level_manager,
+        )
+        from brevitas_examples.llm.llm_quant.export import (
+            brevitas_layer_export_mode,
+        )
+        from brevitas_examples.llm.llm_quant.sharded_mlir_group_export import (
+            LinearWeightBlockQuantHandlerFwd,
+        )
+        from brevitas_examples.llm.llm_quant.export import (
+            replace_call_fn_target,
+        )
+        from brevitas_examples.llm.llm_quant.sharded_mlir_group_export import (
+            matmul_rhs_group_quant_placeholder,
+        )
+        from brevitas.backport.fx.experimental.proxy_tensor import (
+            make_fx as brevitas_make_fx,
+        )
+
         export_context_manager = brevitas_layer_export_mode
         export_class = block_quant_layer_level_manager(
             export_handlers=[LinearWeightBlockQuantHandlerFwd]
@@ -647,6 +713,10 @@ def import_with_fx(
         add_upcast(fx_g)
         fx_g.recompile()
 
+    if is_gptq:
+        gptq_transforms(fx_g)
+        fx_g.recompile()
+
     if mlir_type == "fx":
         return fx_g
 
@@ -677,5 +747,27 @@ def import_with_fx(
         )
         return mlir_module, func_name
 
-    mlir_module, func_name = mlir_importer.import_mlir()
+    mlir_module, func_name = mlir_importer.import_mlir(mlir_type=mlir_type)
     return mlir_module, func_name
+
+
+# Saves a .mlir module python object to the directory 'dir' with 'model_name' and returns a path to the saved file.
+def save_mlir(
+    mlir_module,
+    model_name,
+    mlir_dialect="linalg",
+    frontend="torch",
+    dir=tempfile.gettempdir(),
+):
+    model_name_mlir = (
+        model_name + "_" + frontend + "_" + mlir_dialect + ".mlir"
+    )
+    if dir == "":
+        dir = tempfile.gettempdir()
+    mlir_path = os.path.join(dir, model_name_mlir)
+    print(f"saving {model_name_mlir} to {dir}")
+    if frontend == "torch":
+        with open(mlir_path, "wb") as mlir_file:
+            mlir_file.write(mlir_module)
+
+    return mlir_path

shark/shark_inference.py

@@ -39,7 +39,7 @@ class SharkInference:
     Attributes
     ----------
     mlir_module : str
-        mlir_module represented in string; modules from torch-mlir are serialized in bytecode format.
+        mlir_module or path represented in string; modules from torch-mlir are serialized in bytecode format.
     device : str
         device to execute the mlir_module on.
         currently supports cpu, cuda, vulkan, and metal backends.
@@ -65,7 +65,7 @@ class SharkInference:
 
     def __init__(
         self,
-        mlir_module: bytes,
+        mlir_module,
         device: str = "none",
         mlir_dialect: str = "linalg",
         is_benchmark: bool = False,
@@ -75,6 +75,14 @@ class SharkInference:
         mmap: bool = True,
     ):
         self.mlir_module = mlir_module
+        if mlir_module is not None:
+            if mlir_module and not os.path.isfile(mlir_module):
+                print(
+                    "Warning: Initializing SharkInference with a mlir string/bytecode object will duplicate the model in RAM at compile time. To avoid this, initialize SharkInference with a path to a MLIR module on your hard disk instead."
+                )
+                self.compile_str = True
+            else:
+                self.compile_str = False
         self.device = shark_args.device if device == "none" else device
         self.mlir_dialect = mlir_dialect
         self.is_benchmark = is_benchmark
@@ -192,7 +200,9 @@ class SharkInference:
 
     # TODO: Instead of passing directory and having names decided by the module
     # , user may want to save the module with manual names.
-    def save_module(self, dir=os.getcwd(), module_name=None, extra_args=[]):
+    def save_module(
+        self, dir=os.getcwd(), module_name=None, extra_args=[], debug=False
+    ):
         return export_iree_module_to_vmfb(
             self.mlir_module,
             self.device,
@@ -200,6 +210,8 @@ class SharkInference:
             self.mlir_dialect,
             module_name=module_name,
             extra_args=extra_args,
+            debug=debug,
+            compile_str=self.compile_str,
         )
 
     # load and return the module.

shark/shark_runner.py

@@ -45,7 +45,7 @@ class SharkRunner:
     Attributes
     ----------
     mlir_module : str
-        mlir_module represented in string.
+        mlir_module path, string, or bytecode.
     device : str
         device to execute the mlir_module on.
         currently supports cpu, cuda, vulkan, and metal backends.
@@ -74,6 +74,14 @@ class SharkRunner:
         device_idx: int = None,
     ):
         self.mlir_module = mlir_module
+        if self.mlir_module is not None:
+            if not os.path.isfile(mlir_module):
+                print(
+                    "Warning: Initializing SharkRunner with a mlir string/bytecode object will duplicate the model in RAM at compile time. To avoid this, initialize SharkInference with a path to a MLIR module on your hard disk instead."
+                )
+                self.compile_str = True
+            else:
+                self.compile_str = False
         self.device = shark_args.device if device == "none" else device
         self.mlir_dialect = mlir_dialect
         self.extra_args = extra_args
@@ -91,6 +99,7 @@ class SharkRunner:
                 self.mlir_dialect,
                 extra_args=self.extra_args,
                 device_idx=self.device_idx,
+                compile_str=self.compile_str,
             )
             self.iree_compilation_module = params["vmfb"]
             self.iree_config = params["config"]

shark/shark_trainer.py

@@ -15,7 +15,7 @@
 from shark.parser import shark_args
 from shark.shark_runner import SharkRunner
 from shark.backward_makefx import MakeFxModule
-from shark.shark_importer import import_with_fx
+from shark.shark_importer import import_with_fx, save_mlir
 import numpy as np
 from tqdm import tqdm
 import sys
@@ -69,7 +69,7 @@ class SharkTrainer:
             self.frontend = frontend
 
     # Training function is needed in the case of torch_fn.
-    def compile(self, training_fn=None, extra_args=[]):
+    def compile(self, training_fn=None, mlir_type="linalg", extra_args=[]):
         if self.frontend in ["torch", "pytorch"]:
             packed_inputs = (
                 dict(self.model.named_parameters()),
@@ -77,7 +77,18 @@ class SharkTrainer:
                 tuple(self.input),
             )
             mlir_module, func_name = import_with_fx(
-                training_fn, packed_inputs, False, [], training=True
+                training_fn,
+                packed_inputs,
+                False,
+                [],
+                training=True,
+                mlir_type=mlir_type,
+            )
+            mlir_module = save_mlir(
+                mlir_module,
+                model_name="shark_model",
+                frontend="torch",
+                mlir_dialect=mlir_type,
             )
             self.shark_runner = SharkRunner(
                 mlir_module,

tank/all_models.csv

@@ -1,24 +1,6 @@
-resnet50,stablehlo,tf,1e-2,1e-3,default,nhcw-nhwc,False,False,False,"","macos"
-albert-base-v2,stablehlo,tf,1e-2,1e-2,default,None,False,False,False,"",""
-roberta-base,stablehlo,tf,1e-02,1e-3,default,nhcw-nhwc,True,True,True,"","macos"
-bert-base-uncased,stablehlo,tf,1e-2,1e-3,default,None,False,False,False,"","enabled_windows"
-camembert-base,stablehlo,tf,1e-2,1e-3,default,None,True,True,True,"",""
-dbmdz/convbert-base-turkish-cased,stablehlo,tf,1e-2,1e-3,default,nhcw-nhwc,True,True,False,"https://github.com/iree-org/iree/issues/9971",""
-distilbert-base-uncased,stablehlo,tf,1e-2,1e-3,default,None,False,False,False,"",""
-facebook/convnext-tiny-224,stablehlo,tf,1e-2,1e-3,tf_vit,nhcw-nhwc,True,True,False,"https://github.com/nod-ai/SHARK/issues/311 & https://github.com/nod-ai/SHARK/issues/342","macos"
-funnel-transformer/small,stablehlo,tf,1e-2,1e-3,default,None,True,True,False,"https://github.com/nod-ai/SHARK/issues/201",""
-google/electra-small-discriminator,stablehlo,tf,1e-2,1e-3,default,None,False,False,False,"",""
-google/mobilebert-uncased,stablehlo,tf,1e-2,1e-3,default,None,True,False,False,"Fails during iree-compile","macos"
-google/vit-base-patch16-224,stablehlo,tf,1e-2,1e-3,tf_vit,nhcw-nhwc,False,False,False,"",""
-microsoft/MiniLM-L12-H384-uncased,stablehlo,tf,1e-2,1e-3,tf_hf,None,True,False,False,"Fails during iree-compile.",""
-microsoft/layoutlm-base-uncased,stablehlo,tf,1e-2,1e-3,default,None,False,False,False,"",""
-microsoft/mpnet-base,stablehlo,tf,1e-2,1e-2,default,None,True,True,True,"",""
-alexnet,linalg,torch,1e-2,1e-3,default,None,True,True,False,"https://github.com/nod-ai/SHARK/issues/879",""
-bert-base-cased,linalg,torch,1e-2,1e-3,default,None,False,True,False,"",""
 bert-base-uncased,linalg,torch,1e-2,1e-3,default,None,False,True,False,"",""
 bert-base-uncased_fp16,linalg,torch,1e-1,1e-1,default,None,True,True,True,"",""
 bert-large-uncased,linalg,torch,1e-2,1e-3,default,None,False,True,False,"",""
-bert-large-uncased,stablehlo,tf,1e-2,1e-3,default,None,False,False,False,"",""
 facebook/deit-small-distilled-patch16-224,linalg,torch,1e-2,1e-3,default,nhcw-nhwc,False,True,False,"Fails during iree-compile.",""
 google/vit-base-patch16-224,linalg,torch,1e-2,1e-3,default,nhcw-nhwc,False,True,False,"https://github.com/nod-ai/SHARK/issues/311",""
 microsoft/beit-base-patch16-224-pt22k-ft22k,linalg,torch,1e-2,1e-3,default,nhcw-nhwc,False,True,False,"https://github.com/nod-ai/SHARK/issues/390","macos"
@@ -32,14 +14,8 @@ resnet50,linalg,torch,1e-2,1e-3,default,nhcw-nhwc,False,False,False,"","macos"
 resnet50_fp16,linalg,torch,1e-2,1e-2,default,nhcw-nhwc/img2col,True,True,True,"Numerics issues, awaiting cuda-independent fp16 integration",""
 squeezenet1_0,linalg,torch,1e-2,1e-3,default,nhcw-nhwc,False,False,False,"","macos"
 wide_resnet50_2,linalg,torch,1e-2,1e-3,default,nhcw-nhwc/img2col,True,False,False,"","macos"
-efficientnet-v2-s,stablehlo,tf,1e-02,1e-3,default,nhcw-nhwc,False,False,False,"","macos"
 mnasnet1_0,linalg,torch,1e-2,1e-3,default,nhcw-nhwc,True,True,True,"","macos"
 efficientnet_b0,linalg,torch,1e-2,1e-3,default,nhcw-nhwc,True,True,True,"https://github.com/nod-ai/SHARK/issues/1487","macos"
 efficientnet_b7,linalg,torch,1e-2,1e-3,default,nhcw-nhwc,True,True,True,"https://github.com/nod-ai/SHARK/issues/1487","macos"
-efficientnet_b0,stablehlo,tf,1e-2,1e-3,default,nhcw-nhwc,False,False,False,"",""
-efficientnet_b7,stablehlo,tf,1e-2,1e-3,default,nhcw-nhwc,False,False,False,"Fails on MacOS builder, VK device lost","macos"
-gpt2,stablehlo,tf,1e-2,1e-3,default,None,True,False,False,"","macos"
 t5-base,linalg,torch,1e-2,1e-3,default,None,True,True,True,"Inputs for seq2seq models in torch currently unsupported.","macos"
-t5-base,stablehlo,tf,1e-2,1e-3,default,None,False,False,False,"","macos"
 t5-large,linalg,torch,1e-2,1e-3,default,None,True,True,True,"Inputs for seq2seq models in torch currently unsupported","macos"
-t5-large,stablehlo,tf,1e-2,1e-3,default,None,False,False,False,"","macos"

tank/examples/bert_tf/bert_large_run.py

@@ -85,8 +85,6 @@ if __name__ == "__main__":
     args = [
         "--iree-llvmcpu-target-cpu-features=host",
         "--iree-mhlo-demote-i64-to-i32=false",
-        "--iree-stream-resource-index-bits=64",
-        "--iree-vm-target-index-bits=64",
     ]
     backend_config = "dylib"
     # backend = "cuda"

tank/examples/opt/README.md

@@ -1,3 +1,26 @@
-# Running Different OPT Variants
+# Run OPT for sentence completion through SHARK
 
-To run different sizes of OPT, change the string `OPT_MODEL` string in `opt_torch_test.py`. The default is 350m parameters. 66b cases also exist in the file, simply uncomment the test cases.
+From base SHARK directory, follow instructions to set up a virtual environment with SHARK. (`./setup_venv.sh` or `./setup_venv.ps1`)
+Then, you may run opt_causallm.py to get a very simple sentence completion application running through SHARK
+```
+python opt_causallm.py
+```
+
+# Run OPT performance comparison on SHARK vs. PyTorch
+
+```
+python opt_perf_comparison.py --max-seq-len=512 --model-name=facebook/opt-1.3b \
+        --platform=shark
+```
+Any OPT model from huggingface should work with this script, and you can choose between `--platform=shark` or `--platform=huggingface` to generate benchmarks of OPT inference on SHARK / PyTorch. 
+
+# Run a small suite of OPT models through the benchmark script
+
+```
+python opt_perf_comparison_batch.py
+```
+This script will run benchmarks from a suite of OPT configurations:
+- Sequence Lengths: 32, 128, 256, 512
+- Parameter Counts: 125m, 350m, 1.3b
+
+note: Most of these scripts are written for use on CPU, as perf comparisons against pytorch can be problematic across platforms otherwise.

tank/examples/opt/opt_causallm.py

@@ -36,9 +36,7 @@ def create_module(model_name, tokenizer, device):
 
     mlir_path = f"./{OPT_FS_NAME}_causallm_{MAX_SEQUENCE_LENGTH}_torch.mlir"
     if os.path.isfile(mlir_path):
-        with open(mlir_path, "r") as f:
-            model_mlir = f.read()
-        print(f"Loaded .mlir from {mlir_path}")
+        print(f"Found .mlir from {mlir_path}")
     else:
         (model_mlir, func_name) = import_with_fx(
             model=opt_model,
@@ -50,16 +48,17 @@ def create_module(model_name, tokenizer, device):
         with open(mlir_path, "w") as f:
             f.write(model_mlir)
         print(f"Saved mlir at {mlir_path}")
+        del model_mlir
 
     shark_module = SharkInference(
-        model_mlir,
+        mlir_path,
         device=device,
         mlir_dialect="tm_tensor",
         is_benchmark=False,
     )
 
     vmfb_name = f"{OPT_FS_NAME}_causallm_{MAX_SEQUENCE_LENGTH}_torch_{device}"
-    shark_module.save_module(module_name=vmfb_name)
+    shark_module.save_module(module_name=vmfb_name, debug=False)
     vmfb_path = vmfb_name + ".vmfb"
     return vmfb_path
 

tank/examples/opt/opt_causallm_torch_test.py

@@ -6,7 +6,7 @@ import numpy as np
 from shark_opt_wrapper import OPTForCausalLMModel
 from shark.iree_utils._common import check_device_drivers, device_driver_info
 from shark.shark_inference import SharkInference
-from shark.shark_importer import import_with_fx
+from shark.shark_importer import import_with_fx, save_mlir
 from transformers import AutoTokenizer, OPTForCausalLM
 
 OPT_MODEL = "facebook/opt-1.3b"
@@ -57,9 +57,10 @@ class OPTModuleTester:
         with open(mlir_path, "w") as f:
             f.write(mlir_module)
         print(f"Saved mlir at {mlir_path}")
+        del mlir_module
 
         shark_module = SharkInference(
-            mlir_module,
+            mlir_path,
             device=device,
             mlir_dialect="tm_tensor",
             is_benchmark=self.benchmark,

tank/examples/opt/opt_perf_comparison.py

@@ -1,18 +1,45 @@
+"""
+Script for comparing OPT model performance between SHARK and Huggingface
+PyTorch.
+
+Usage Example:
+
+python opt_perf_comparison.py --max-seq-len=32 --model-name=facebook/opt-125m \
+        --platform=shark
+
+python opt_perf_comparison.py --max-seq-len=512 --model-name=facebook/opt-1.3b \
+        --platform=shark
+
+See parse_args() below for command line argument usage.
+"""
+
+import argparse
 import collections
 import json
-import time
 import os
+import psutil
+import time
+from typing import Tuple
 
 from shark.shark_inference import SharkInference
 from shark.shark_importer import import_with_fx
 from transformers import AutoTokenizer, OPTForCausalLM
 from shark_opt_wrapper import OPTForCausalLMModel
 
-MODEL_NAME = "facebook/opt-1.3b"
-OPT_MODELNAME = "opt-1.3b"
-OPT_FS_NAME = "opt_1-3b"
-MAX_SEQUENCE_LENGTH = 512
 DEVICE = "cpu"
+PLATFORM_SHARK = "shark"
+PLATFORM_HUGGINGFACE = "huggingface"
+
+# Dict keys for reports.
+REPORT_PLATFORM = "platform"
+REPORT_MODEL_NAME = "model"
+REPORT_MAX_SEQ_LEN = "max_seq_len"
+REPORT_LOAD_TIME = "load_time_sec"
+REPORT_RUN_TIME = "run_time_sec"
+REPORT_LOAD_PHYSICAL_MEMORY_MB = "load_physical_MB"
+REPORT_LOAD_VIRTUAL_MEMORY_MB = "load_virtual_MB"
+REPORT_RUN_PHYSICAL_MEMORY_MB = "run_physical_MB"
+REPORT_RUN_VIRTUAL_MEMORY_MB = "run_virtual_MB"
 
 PROMPTS = [
     "What is the meaning of life?",
@@ -30,15 +57,27 @@ PROMPTS = [
 ModelWrapper = collections.namedtuple("ModelWrapper", ["model", "tokenizer"])
 
 
-def create_vmfb_module(model_name, tokenizer, device):
-    opt_base_model = OPTForCausalLM.from_pretrained("facebook/" + model_name)
+def get_memory_info():
+    pid = os.getpid()
+    process = psutil.Process(pid)
+    return process.memory_info()
+
+
+def create_vmfb_module(
+    model_name: str,
+    tokenizer,
+    device: str,
+    max_seq_len: int,
+    recompile_shark: bool,
+):
+    opt_base_model = OPTForCausalLM.from_pretrained(model_name)
     opt_base_model.eval()
     opt_model = OPTForCausalLMModel(opt_base_model)
     encoded_inputs = tokenizer(
-        "What is the meaning of life?",
+        PROMPTS[0],
         padding="max_length",
         truncation=True,
-        max_length=MAX_SEQUENCE_LENGTH,
+        max_length=max_seq_len,
         return_tensors="pt",
     )
     inputs = (
@@ -48,8 +87,16 @@ def create_vmfb_module(model_name, tokenizer, device):
     # np.save("model_inputs_0.npy", inputs[0])
     # np.save("model_inputs_1.npy", inputs[1])
 
-    mlir_path = f"./{OPT_FS_NAME}_causallm_{MAX_SEQUENCE_LENGTH}_torch.mlir"
-    if os.path.isfile(mlir_path):
+    opt_fs_name = get_opt_fs_name(model_name)
+    mlir_path = f"./{opt_fs_name}_causallm_{max_seq_len}_torch.mlir"
+    # If MLIR has already been loaded and recompilation is not requested, use
+    # the loaded MLIR file.
+    has_mlir = os.path.isfile(mlir_path)
+    # The purpose of recompile_shark is to measure compilation time; the
+    # compilation time can be correctly measured only when MLIR has already been
+    # loaded.
+    assert not recompile_shark or has_mlir
+    if has_mlir:
         with open(mlir_path, "r") as f:
             model_mlir = f.read()
         print(f"Loaded .mlir from {mlir_path}")
@@ -58,7 +105,7 @@ def create_vmfb_module(model_name, tokenizer, device):
             model=opt_model,
             inputs=inputs,
             is_f16=False,
-            model_name=OPT_FS_NAME,
+            model_name=opt_fs_name,
             return_str=True,
         )
         with open(mlir_path, "w") as f:
@@ -72,18 +119,25 @@ def create_vmfb_module(model_name, tokenizer, device):
         is_benchmark=False,
     )
 
-    vmfb_name = f"{OPT_FS_NAME}_causallm_{MAX_SEQUENCE_LENGTH}_torch_{DEVICE}_tiled_ukernels"
+    vmfb_name = (
+        f"{opt_fs_name}_causallm_{max_seq_len}_torch_{DEVICE}_tiled_ukernels"
+    )
     shark_module.save_module(module_name=vmfb_name)
     vmfb_path = vmfb_name + ".vmfb"
     return vmfb_path
 
 
-def load_shark_model() -> ModelWrapper:
-    vmfb_name = f"{OPT_FS_NAME}_causallm_{MAX_SEQUENCE_LENGTH}_torch_{DEVICE}_tiled_ukernels.vmfb"
-    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, use_fast=False)
-    if not os.path.isfile(vmfb_name):
+def load_shark_model(
+    model_name: str, max_seq_len: int, recompile_shark: bool
+) -> ModelWrapper:
+    opt_fs_name = get_opt_fs_name(model_name)
+    vmfb_name = f"{opt_fs_name}_causallm_{max_seq_len}_torch_{DEVICE}_tiled_ukernels.vmfb"
+    tokenizer = AutoTokenizer.from_pretrained(model_name, use_fast=False)
+    if recompile_shark or not os.path.isfile(vmfb_name):
         print(f"vmfb not found. compiling and saving to {vmfb_name}")
-        create_vmfb_module(OPT_MODELNAME, tokenizer, DEVICE)
+        create_vmfb_module(
+            model_name, tokenizer, DEVICE, max_seq_len, recompile_shark
+        )
     shark_module = SharkInference(mlir_module=None, device="cpu-task")
     shark_module.load_module(vmfb_name)
     return ModelWrapper(model=shark_module, tokenizer=tokenizer)
@@ -94,20 +148,10 @@ def run_shark_model(model_wrapper: ModelWrapper, tokens):
     return model_wrapper.model("forward", tokens)
 
 
-def run_shark():
-    model_wrapper = load_shark_model()
-
-    prompt = "What is the meaning of life?"
-    logits = run_shark_model(model_wrapper, prompt)
-
-    # Print output logits to validate vs. pytorch + base transformers
-    print(logits[0])
-
-
-def load_huggingface_model() -> ModelWrapper:
+def load_huggingface_model(model_name: str) -> ModelWrapper:
     return ModelWrapper(
-        model=OPTForCausalLM.from_pretrained(MODEL_NAME),
-        tokenizer=AutoTokenizer.from_pretrained(MODEL_NAME),
+        model=OPTForCausalLM.from_pretrained(model_name),
+        tokenizer=AutoTokenizer.from_pretrained(model_name),
     )
 
 
@@ -117,47 +161,71 @@ def run_huggingface_model(model_wrapper: ModelWrapper, tokens):
     )
 
 
-def run_huggingface():
-    model_wrapper = load_huggingface_model()
-    prompt = "What is the meaning of life?"
-    logits = run_huggingface_model(model_wrapper, prompt)
-
-    print(logits[0])
-
-
 def save_json(data, filename):
     with open(filename, "w") as file:
         json.dump(data, file)
 
 
-def collect_huggingface_logits():
+def collect_huggingface_logits(
+    model_name: str, max_seq_len: int, to_save_json: bool
+) -> Tuple[float, float]:
+    # Load
     t0 = time.time()
-    model_wrapper = load_huggingface_model()
-    print("--- Took {} seconds to load Huggingface.".format(time.time() - t0))
+    model_wrapper = load_huggingface_model(model_name)
+    load_time = time.time() - t0
+    print("--- Took {} seconds to load Huggingface.".format(load_time))
+    load_memory_info = get_memory_info()
+
     results = []
     tokenized_prompts = []
     for prompt in PROMPTS:
         tokens = model_wrapper.tokenizer(
             prompt,
             padding="max_length",
-            max_length=MAX_SEQUENCE_LENGTH,
+            max_length=max_seq_len,
             truncation=True,
             return_tensors="pt",
         )
         tokenized_prompts.append(tokens)
+
+    # Run
     t0 = time.time()
     for idx, tokens in enumerate(tokenized_prompts):
         print("prompt: {}".format(PROMPTS[idx]))
         logits = run_huggingface_model(model_wrapper, tokens)
-        results.append([PROMPTS[idx], logits[0].tolist()])
-    print("--- Took {} seconds to run Huggingface.".format(time.time() - t0))
-    save_json(results, "/tmp/huggingface.json")
+        if to_save_json:
+            results.append([PROMPTS[idx], logits[0].tolist()])
+    run_time = time.time() - t0
+    print("--- Took {} seconds to run Huggingface.".format(run_time))
+    if to_save_json:
+        save_json(results, "/tmp/huggingface.json")
+    run_memory_info = get_memory_info()
+    return {
+        REPORT_PLATFORM: PLATFORM_HUGGINGFACE,
+        REPORT_MODEL_NAME: model_name,
+        REPORT_MAX_SEQ_LEN: max_seq_len,
+        REPORT_LOAD_TIME: load_time,
+        REPORT_RUN_TIME: run_time / len(PROMPTS),
+        REPORT_LOAD_PHYSICAL_MEMORY_MB: load_memory_info.rss >> 20,
+        REPORT_LOAD_VIRTUAL_MEMORY_MB: load_memory_info.vms >> 20,
+        REPORT_RUN_PHYSICAL_MEMORY_MB: run_memory_info.rss >> 20,
+        REPORT_RUN_VIRTUAL_MEMORY_MB: run_memory_info.vms >> 20,
+    }
 
 
-def collect_shark_logits():
+def collect_shark_logits(
+    model_name: str,
+    max_seq_len: int,
+    recompile_shark: bool,
+    to_save_json: bool,
+) -> Tuple[float, float]:
+    # Load
     t0 = time.time()
-    model_wrapper = load_shark_model()
-    print("--- Took {} seconds to load Shark.".format(time.time() - t0))
+    model_wrapper = load_shark_model(model_name, max_seq_len, recompile_shark)
+    load_time = time.time() - t0
+    print("--- Took {} seconds to load Shark.".format(load_time))
+    load_memory_info = get_memory_info()
+
     results = []
     tokenized_prompts = []
     for prompt in PROMPTS:
@@ -165,7 +233,7 @@ def collect_shark_logits():
             prompt,
             padding="max_length",
             truncation=True,
-            max_length=MAX_SEQUENCE_LENGTH,
+            max_length=max_seq_len,
             return_tensors="pt",
         )
         inputs = (
@@ -173,16 +241,100 @@ def collect_shark_logits():
             tokens["attention_mask"],
         )
         tokenized_prompts.append(inputs)
+
+    # Run
     t0 = time.time()
     for idx, tokens in enumerate(tokenized_prompts):
         print("prompt: {}".format(PROMPTS[idx]))
         logits = run_shark_model(model_wrapper, tokens)
         lst = [e.tolist() for e in logits]
-        results.append([PROMPTS[idx], lst])
-    print("--- Took {} seconds to run Shark.".format(time.time() - t0))
-    save_json(results, "/tmp/shark.json")
+        if to_save_json:
+            results.append([PROMPTS[idx], lst])
+    run_time = time.time() - t0
+    print("--- Took {} seconds to run Shark.".format(run_time))
+    if to_save_json:
+        save_json(results, "/tmp/shark.json")
+    platform_postfix = "-compile" if recompile_shark else "-precompiled"
+    run_memory_info = get_memory_info()
+    return {
+        REPORT_PLATFORM: PLATFORM_SHARK + platform_postfix,
+        REPORT_MODEL_NAME: model_name,
+        REPORT_MAX_SEQ_LEN: max_seq_len,
+        REPORT_LOAD_TIME: load_time,
+        REPORT_RUN_TIME: run_time / len(PROMPTS),
+        REPORT_LOAD_PHYSICAL_MEMORY_MB: load_memory_info.rss >> 20,
+        REPORT_LOAD_VIRTUAL_MEMORY_MB: load_memory_info.vms >> 20,
+        REPORT_RUN_PHYSICAL_MEMORY_MB: run_memory_info.rss >> 20,
+        REPORT_RUN_VIRTUAL_MEMORY_MB: run_memory_info.vms >> 20,
+    }
+
+
+def get_opt_fs_name(model_name: str) -> str:
+    """Cleanses the model name ino a file system-friendly name.
+
+    Example: get_opt_fs_name('facebook/opt-1.3b') == 'opt_1-3b'
+    """
+    slash_split = model_name.split("/")
+    assert 1 <= len(slash_split) <= 2, "There should be at most one slash."
+    model_name = slash_split[-1]
+    for src_pattern, dest_pattern in (("-", "_"), (".", "-")):
+        model_name = model_name.replace(src_pattern, dest_pattern)
+    return model_name
+
+
+def parse_args():
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--save-json",
+        help="If set, saves output JSON.",
+        action=argparse.BooleanOptionalAction,
+        default=False,
+    )
+    parser.add_argument(
+        "--max-seq-len", help="Max sequence length", type=int, default=32
+    )
+    parser.add_argument(
+        "--model-name",
+        help="Model name",
+        type=str,
+        choices=[
+            "facebook/opt-125m",
+            "facebook/opt-350m",
+            "facebook/opt-1.3b",
+            "facebook/opt-6.7b",
+        ],
+        default="facebook/opt-1.3b",
+    )
+    parser.add_argument(
+        "--recompile-shark",
+        help="If set, recompiles MLIR",
+        action=argparse.BooleanOptionalAction,
+        default=False,
+    )
+    parser.add_argument(
+        "--platform",
+        help="Either shark or huggingface",
+        type=str,
+        choices=[PLATFORM_SHARK, PLATFORM_HUGGINGFACE],
+        default=PLATFORM_SHARK,
+    )
+    args = parser.parse_args()
+    print("args={}".format(args))
+    return args
 
 
 if __name__ == "__main__":
-    collect_shark_logits()
-    collect_huggingface_logits()
+    args = parse_args()
+    if args.platform == PLATFORM_SHARK:
+        shark_report = collect_shark_logits(
+            args.model_name,
+            args.max_seq_len,
+            args.recompile_shark,
+            args.save_json,
+        )
+        print("# Summary: {}".format(json.dumps(shark_report)))
+    else:
+        huggingface_report = collect_huggingface_logits(
+            args.model_name, args.max_seq_len, args.save_json
+        )
+        print("# Summary: {}".format(json.dumps(huggingface_report)))

tank/examples/opt/opt_perf_comparison_batch.py

@@ -0,0 +1,30 @@
+"""
+Script for running opt_perf_comparison.py in batch with a series of arguments.
+
+Usage: python opt_perf_comparison_batch.py
+"""
+
+from typing import Iterable, List
+import shlex
+import subprocess
+
+
+def make_commands() -> Iterable[List[str]]:
+    command = shlex.split("python opt_perf_comparison.py --no-save-json")
+    max_seq_lens = [32, 128, 256, 512]
+    model_names = ["facebook/opt-" + e for e in ["125m", "350m", "1.3b"]]
+    for max_seq_len in max_seq_lens:
+        for model_name in model_names:
+            yield command + [
+                f"--max-seq-len={max_seq_len}",
+                f"--model-name={model_name}",
+            ]
+
+
+def main():
+    for command in make_commands():
+        result = subprocess.run(command, check=True)
+
+
+if __name__ == "__main__":
+    main()

tank/examples/opt/shark_hf_base_opt.py

@@ -2,7 +2,7 @@ import os
 import torch
 from transformers import AutoTokenizer, OPTForCausalLM
 from shark.shark_inference import SharkInference
-from shark.shark_importer import import_with_fx
+from shark.shark_importer import import_with_fx, save_mlir
 from shark_opt_wrapper import OPTForCausalLMModel
 
 model_name = "facebook/opt-1.3b"
@@ -25,11 +25,13 @@ inputs = (
     model=model,
     inputs=inputs,
     is_f16=False,
-    debug=True,
-    model_name=model_name.split("/")[1],
-    save_dir=".",
 )
-
+mlir_module = save_mlir(
+    mlir_module,
+    model_name=model_name.split("/")[1],
+    frontend="torch",
+    mlir_dialect="linalg",
+)
 shark_module = SharkInference(
     mlir_module,
     device="cpu-sync",

tank/generate_sharktank.py

@@ -36,7 +36,7 @@ def save_torch_model(torch_model_list, local_tank_cache, import_args):
         get_hf_img_cls_model,
         get_fp16_model,
     )
-    from shark.shark_importer import import_with_fx
+    from shark.shark_importer import import_with_fx, save_mlir
 
     with open(torch_model_list) as csvfile:
         torch_reader = csv.reader(csvfile, delimiter=",")
@@ -130,133 +130,6 @@ def save_torch_model(torch_model_list, local_tank_cache, import_args):
                     )
 
 
-def save_tf_model(tf_model_list, local_tank_cache, import_args):
-    from tank.model_utils_tf import (
-        get_causal_image_model,
-        get_masked_lm_model,
-        get_causal_lm_model,
-        get_keras_model,
-        get_TFhf_model,
-        get_tfhf_seq2seq_model,
-    )
-    import os
-
-    os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2"
-    import tensorflow as tf
-
-    visible_default = tf.config.list_physical_devices("GPU")
-    try:
-        tf.config.set_visible_devices([], "GPU")
-        visible_devices = tf.config.get_visible_devices()
-        for device in visible_devices:
-            assert device.device_type != "GPU"
-    except:
-        # Invalid device or cannot modify virtual devices once initialized.
-        pass
-
-    with open(tf_model_list) as csvfile:
-        tf_reader = csv.reader(csvfile, delimiter=",")
-        fields = next(tf_reader)
-        for row in tf_reader:
-            tf_model_name = row[0]
-            model_type = row[1]
-
-            model = None
-            input = None
-            print(f"Generating artifacts for model {tf_model_name}")
-            if model_type == "hf":
-                model, input, _ = get_masked_lm_model(
-                    tf_model_name, import_args
-                )
-            elif model_type == "img":
-                model, input, _ = get_causal_image_model(
-                    tf_model_name, import_args
-                )
-            elif model_type == "keras":
-                model, input, _ = get_keras_model(tf_model_name, import_args)
-            elif model_type == "TFhf":
-                model, input, _ = get_TFhf_model(tf_model_name, import_args)
-            elif model_type == "tfhf_seq2seq":
-                model, input, _ = get_tfhf_seq2seq_model(
-                    tf_model_name, import_args
-                )
-            elif model_type == "hf_causallm":
-                model, input, _ = get_causal_lm_model(
-                    tf_model_name, import_args
-                )
-
-            tf_model_name = tf_model_name.replace("/", "_")
-            if import_args["batch_size"] != 1:
-                tf_model_dir = os.path.join(
-                    local_tank_cache,
-                    str(tf_model_name)
-                    + "_tf"
-                    + f"_BS{str(import_args['batch_size'])}",
-                )
-            else:
-                tf_model_dir = os.path.join(
-                    local_tank_cache, str(tf_model_name) + "_tf"
-                )
-            os.makedirs(tf_model_dir, exist_ok=True)
-            mlir_importer = SharkImporter(
-                model,
-                inputs=input,
-                frontend="tf",
-            )
-            mlir_importer.import_debug(
-                is_dynamic=False,
-                dir=tf_model_dir,
-                model_name=tf_model_name,
-            )
-
-
-def save_tflite_model(tflite_model_list, local_tank_cache, import_args):
-    from shark.tflite_utils import TFLitePreprocessor
-
-    with open(tflite_model_list) as csvfile:
-        tflite_reader = csv.reader(csvfile, delimiter=",")
-        for row in tflite_reader:
-            print("\n")
-            tflite_model_name = row[0]
-            tflite_model_link = row[1]
-            print("tflite_model_name", tflite_model_name)
-            print("tflite_model_link", tflite_model_link)
-            tflite_model_name_dir = os.path.join(
-                local_tank_cache, str(tflite_model_name) + "_tflite"
-            )
-            os.makedirs(tflite_model_name_dir, exist_ok=True)
-            print(f"TMP_TFLITE_MODELNAME_DIR = {tflite_model_name_dir}")
-
-            # Preprocess to get SharkImporter input import_args
-            tflite_preprocessor = TFLitePreprocessor(str(tflite_model_name))
-            raw_model_file_path = tflite_preprocessor.get_raw_model_file()
-            inputs = tflite_preprocessor.get_inputs()
-            tflite_interpreter = tflite_preprocessor.get_interpreter()
-
-            # Use SharkImporter to get SharkInference input import_args
-            my_shark_importer = SharkImporter(
-                module=tflite_interpreter,
-                inputs=inputs,
-                frontend="tflite",
-                raw_model_file=raw_model_file_path,
-            )
-            my_shark_importer.import_debug(
-                dir=tflite_model_name_dir,
-                model_name=tflite_model_name,
-                func_name="main",
-            )
-            mlir_hash = create_hash(
-                os.path.join(
-                    tflite_model_name_dir,
-                    tflite_model_name + "_tflite" + ".mlir",
-                )
-            )
-            np.save(
-                os.path.join(tflite_model_name_dir, "hash"),
-                np.array(mlir_hash),
-            )
-
-
 def check_requirements(frontend):
     import importlib
 
@@ -265,10 +138,6 @@ def check_requirements(frontend):
         tv_spec = importlib.util.find_spec("torchvision")
         has_pkgs = tv_spec is not None
 
-    elif frontend in ["tensorflow", "tf"]:
-        tf_spec = importlib.util.find_spec("tensorflow")
-        has_pkgs = tf_spec is not None
-
     return has_pkgs
 
 
@@ -287,27 +156,11 @@ def gen_shark_files(modelname, frontend, tank_dir, importer_args):
         torch_model_csv = os.path.join(
             os.path.dirname(__file__), "torch_model_list.csv"
         )
-        tf_model_csv = os.path.join(
-            os.path.dirname(__file__), "tf_model_list.csv"
-        )
         custom_model_csv = tempfile.NamedTemporaryFile(
             dir=os.path.dirname(__file__),
             delete=True,
         )
-        # Create a temporary .csv with only the desired entry.
-        if frontend == "tf":
-            with open(tf_model_csv, mode="r") as src:
-                reader = csv.reader(src)
-                for row in reader:
-                    if row[0] == modelname:
-                        target = row
-            with open(custom_model_csv.name, mode="w") as trg:
-                writer = csv.writer(trg)
-                writer.writerow(["modelname", "src"])
-                writer.writerow(target)
-            save_tf_model(custom_model_csv.name, tank_dir, import_args)
-
-        elif frontend == "torch":
+        if frontend == "torch":
             with open(torch_model_csv, mode="r") as src:
                 reader = csv.reader(src)
                 for row in reader:
@@ -341,18 +194,6 @@ if __name__ == "__main__":
     #         Please see: https://github.com/nod-ai/SHARK/blob/main/tank/torch_model_list.csv""",
     # )
     # parser.add_argument(
-    #    "--tf_model_csv",
-    #    type=lambda x: is_valid_file(x),
-    #    default="./tank/tf_model_list.csv",
-    #    help="Contains the file with tf model name and args.",
-    # )
-    # parser.add_argument(
-    #    "--tflite_model_csv",
-    #    type=lambda x: is_valid_file(x),
-    #    default="./tank/tflite/tflite_model_list.csv",
-    #    help="Contains the file with tf model name and args.",
-    # )
-    # parser.add_argument(
     #    "--ci_tank_dir",
     #    type=bool,
     #    default=False,
@@ -369,11 +210,5 @@ if __name__ == "__main__":
     torch_model_csv = os.path.join(
         os.path.dirname(__file__), "torch_model_list.csv"
     )
-    tf_model_csv = os.path.join(os.path.dirname(__file__), "tf_model_list.csv")
-    tflite_model_csv = os.path.join(
-        os.path.dirname(__file__), "tflite", "tflite_model_list.csv"
-    )
 
     save_torch_model(torch_model_csv, WORKDIR, import_args)
-    # save_tf_model(tf_model_csv, WORKDIR, import_args)
-    # save_tflite_model(tflite_model_csv, WORKDIR, import_args)

tank/test_models.py

@@ -287,6 +287,9 @@ class SharkModuleTester:
         repro_path = os.path.join("reproducers", self.tmp_prefix, "*")
 
         bashCommand = f"gsutil cp -r {repro_path} gs://shark-public/builder/repro_artifacts/{self.ci_sha}/{self.tmp_prefix}/"
+        print(
+            f"Uploading reproducer {repro_path} to gs://shark-public/builder/repro_artifacts/{self.ci_sha}/{self.tmp_prefix}/"
+        )
         process = subprocess.run(bashCommand.split())
 
     def postprocess_outputs(self, golden_out, result):

tank/tf_model_list.csv

@@ -1,28 +0,0 @@
-model_name, model_type
-albert-base-v2,hf
-bert-base-uncased,hf
-camembert-base,hf
-dbmdz/convbert-base-turkish-cased,hf
-distilbert-base-uncased,hf
-google/electra-small-discriminator,hf
-funnel-transformer/small,hf
-microsoft/layoutlm-base-uncased,hf
-google/mobilebert-uncased,hf
-microsoft/mpnet-base,hf
-roberta-base,hf
-resnet50,keras
-xlm-roberta-base,hf
-microsoft/MiniLM-L12-H384-uncased,TFhf
-funnel-transformer/small,hf
-microsoft/mpnet-base,hf
-facebook/convnext-tiny-224,img
-google/vit-base-patch16-224,img
-efficientnet-v2-s,keras
-bert-large-uncased,hf
-t5-base,tfhf_seq2seq
-t5-large,tfhf_seq2seq
-efficientnet_b0,keras
-efficientnet_b7,keras
-gpt2,hf_causallm
-t5-base,tfhf_seq2seq
-t5-large,tfhf_seq2seq