Fix loading custom inpainting models (#1103)

* Add doc on profiling with Shark

* Rename doc

.github/workflows/nightly.yml

@@ -44,7 +44,7 @@ jobs:
         body: |
           Automatic snapshot release of nod.ai SHARK.
         draft: true
-        prerelease: false
+        prerelease: true
 
     - name: Build Package 
       shell: powershell
@@ -67,9 +67,9 @@ jobs:
     #    $env:SHARK_PACKAGE_VERSION=${{ env.package_version }}
     #    pip wheel -v -w dist . --pre -f https://download.pytorch.org/whl/nightly/torch -f https://llvm.github.io/torch-mlir/package-index/ -f https://nod-ai.github.io/SHARK-Runtime/pip-release-links.html
 
-    - uses: actions/upload-artifact@v2
-      with:
-        path: dist/*
+    #- uses: actions/upload-artifact@v2
+    #  with:
+    #    path: dist/*
     
     - name: Upload Release Assets
       id: upload-release-assets
@@ -79,6 +79,7 @@ jobs:
       with:
         release_id: ${{ steps.create_release.outputs.id }}
         assets_path: ./dist/*
+        #asset_content_type: application/vnd.microsoft.portable-executable 
 
     - name: Publish Release
       id: publish_release
@@ -133,7 +134,7 @@ jobs:
         source iree.venv/bin/activate
         package_version="$(printf '%(%Y%m%d)T.${{ github.run_number }}')"
         SHARK_PACKAGE_VERSION=${package_version} \
-        pip wheel -v -w wheelhouse . --pre -f https://download.pytorch.org/whl/nightly/torch -f https://llvm.github.io/torch-mlir/package-index/ -f https://iree-org.github.io/iree/pip-release-links.html
+        pip wheel -v -w wheelhouse . --pre -f https://download.pytorch.org/whl/nightly/torch -f https://llvm.github.io/torch-mlir/package-index/ -f https://openxla.github.io/iree/pip-release-links.html
         # Install the built wheel
         pip install ./wheelhouse/nodai*
         # Validate the Models

README.md

@@ -27,7 +27,7 @@ Other users please ensure you have your latest vendor drivers and Vulkan SDK fro
 
 Install the Driver from [Prerequisites](https://github.com/nod-ai/SHARK#install-your-hardware-drivers) above 
 
-Download the stable release [539](https://github.com/nod-ai/SHARK/releases/download/20230216.539/shark_sd_20230216_539.exe) or if you are adventurous the latest .exe from [releases page](https://github.com/nod-ai/SHARK/releases).
+Download the [stable release](https://github.com/nod-ai/shark/releases/latest)
 
 Double click the .exe and you should have the [UI](http://localhost:8080/) in the browser. 
 

apps/stable_diffusion/scripts/__init__.py

@@ -1,2 +1,4 @@
 from apps.stable_diffusion.scripts.txt2img import txt2img_inf
 from apps.stable_diffusion.scripts.img2img import img2img_inf
+from apps.stable_diffusion.scripts.inpaint import inpaint_inf
+from apps.stable_diffusion.scripts.outpaint import outpaint_inf

apps/stable_diffusion/scripts/img2img.py

@@ -35,7 +35,7 @@ schedulers = None
 def img2img_inf(
     prompt: str,
     negative_prompt: str,
-    init_image: str,
+    init_image: Image,
     height: int,
     width: int,
     steps: int,
@@ -64,8 +64,11 @@ def img2img_inf(
     args.steps = steps
     args.strength = strength
     args.scheduler = scheduler
-    args.img_path = init_image
-    image = Image.open(args.img_path).convert("RGB")
+    args.img_path = "not none"
+
+    if init_image is None:
+        return None, "An Initial Image is required"
+    image = init_image.convert("RGB")
 
     # set ckpt_loc and hf_model_id.
     types = (
@@ -86,9 +89,6 @@ def img2img_inf(
     else:
         args.hf_model_id = custom_model
 
-    if image is None:
-        return None, "An Initial Image is required"
-
     args.save_metadata_to_json = save_metadata_to_json
     args.write_metadata_to_png = save_metadata_to_png
 
@@ -119,7 +119,7 @@ def img2img_inf(
         model_id = (
             args.hf_model_id
             if args.hf_model_id
-            else "runwayml/stable-diffusion-inpainting"
+            else "stabilityai/stable-diffusion-2-1-base"
         )
         schedulers = get_schedulers(model_id)
         scheduler_obj = schedulers[scheduler]

apps/stable_diffusion/scripts/inpaint.py

@@ -35,8 +35,7 @@ schedulers = None
 def inpaint_inf(
     prompt: str,
     negative_prompt: str,
-    image: Image,
-    mask_image: Image,
+    image_dict,
     height: int,
     width: int,
     steps: int,
@@ -62,6 +61,8 @@ def inpaint_inf(
     args.guidance_scale = guidance_scale
     args.steps = steps
     args.scheduler = scheduler
+    args.img_path = "not none"
+    args.mask_path = "not none"
 
     # set ckpt_loc and hf_model_id.
     types = (
@@ -138,6 +139,8 @@ def inpaint_inf(
     generated_imgs = []
     seeds = []
     img_seed = utils.sanitize_seed(seed)
+    image = image_dict["image"]
+    mask_image = image_dict["mask"]
     for i in range(batch_count):
         if i > 0:
             img_seed = utils.sanitize_seed(-1)

apps/stable_diffusion/scripts/outpaint.py

@@ -0,0 +1,275 @@
+import sys
+import torch
+import time
+from PIL import Image
+from dataclasses import dataclass
+from apps.stable_diffusion.src import (
+    args,
+    OutpaintPipeline,
+    get_schedulers,
+    set_init_device_flags,
+    utils,
+    clear_all,
+    save_output_img,
+)
+
+
+@dataclass
+class Config:
+    model_id: str
+    ckpt_loc: str
+    precision: str
+    batch_size: int
+    max_length: int
+    height: int
+    width: int
+    device: str
+
+
+outpaint_obj = None
+config_obj = None
+schedulers = None
+
+
+# Exposed to UI.
+def outpaint_inf(
+    prompt: str,
+    negative_prompt: str,
+    init_image: str,
+    pixels: int,
+    mask_blur: int,
+    directions: list,
+    noise_q: float,
+    color_variation: float,
+    height: int,
+    width: int,
+    steps: int,
+    guidance_scale: float,
+    seed: int,
+    batch_count: int,
+    batch_size: int,
+    scheduler: str,
+    custom_model: str,
+    hf_model_id: str,
+    precision: str,
+    device: str,
+    max_length: int,
+    save_metadata_to_json: bool,
+    save_metadata_to_png: bool,
+):
+    global outpaint_obj
+    global config_obj
+    global schedulers
+
+    args.prompts = [prompt]
+    args.negative_prompts = [negative_prompt]
+    args.guidance_scale = guidance_scale
+    args.steps = steps
+    args.scheduler = scheduler
+    args.img_path = init_image
+
+    # set ckpt_loc and hf_model_id.
+    types = (
+        ".ckpt",
+        ".safetensors",
+    )  # the tuple of file types
+    args.ckpt_loc = ""
+    args.hf_model_id = ""
+    if custom_model == "None":
+        if not hf_model_id:
+            return (
+                None,
+                "Please provide either custom model or huggingface model ID, both must not be empty",
+            )
+        args.hf_model_id = hf_model_id
+    elif ".ckpt" in custom_model or ".safetensors" in custom_model:
+        args.ckpt_loc = custom_model
+    else:
+        args.hf_model_id = custom_model
+
+    args.save_metadata_to_json = save_metadata_to_json
+    args.write_metadata_to_png = save_metadata_to_png
+
+    dtype = torch.float32 if precision == "fp32" else torch.half
+    cpu_scheduling = not scheduler.startswith("Shark")
+    new_config_obj = Config(
+        args.hf_model_id,
+        args.ckpt_loc,
+        precision,
+        batch_size,
+        max_length,
+        height,
+        width,
+        device,
+    )
+    if not outpaint_obj or config_obj != new_config_obj:
+        config_obj = new_config_obj
+        args.precision = precision
+        args.batch_size = batch_size
+        args.max_length = max_length
+        args.height = height
+        args.width = width
+        args.device = device.split("=>", 1)[1].strip()
+        args.iree_vulkan_target_triple = ""
+        args.use_tuned = True
+        args.import_mlir = False
+        set_init_device_flags()
+        model_id = (
+            args.hf_model_id
+            if args.hf_model_id
+            else "stabilityai/stable-diffusion-2-inpainting"
+        )
+        schedulers = get_schedulers(model_id)
+        scheduler_obj = schedulers[scheduler]
+        outpaint_obj = OutpaintPipeline.from_pretrained(
+            scheduler_obj,
+            args.import_mlir,
+            args.hf_model_id,
+            args.ckpt_loc,
+            args.custom_vae,
+            args.precision,
+            args.max_length,
+            args.batch_size,
+            args.height,
+            args.width,
+            args.use_base_vae,
+            args.use_tuned,
+        )
+
+    outpaint_obj.scheduler = schedulers[scheduler]
+
+    start_time = time.time()
+    outpaint_obj.log = ""
+    generated_imgs = []
+    seeds = []
+    img_seed = utils.sanitize_seed(seed)
+    image = Image.open(args.img_path)
+
+    left = True if "left" in directions else False
+    right = True if "right" in directions else False
+    top = True if "up" in directions else False
+    bottom = True if "down" in directions else False
+
+    for i in range(batch_count):
+        if i > 0:
+            img_seed = utils.sanitize_seed(-1)
+        out_imgs = outpaint_obj.generate_images(
+            prompt,
+            negative_prompt,
+            image,
+            pixels,
+            mask_blur,
+            left,
+            right,
+            top,
+            bottom,
+            noise_q,
+            color_variation,
+            batch_size,
+            height,
+            width,
+            steps,
+            guidance_scale,
+            img_seed,
+            args.max_length,
+            dtype,
+            args.use_base_vae,
+            cpu_scheduling,
+        )
+        save_output_img(out_imgs[0], img_seed)
+        generated_imgs.extend(out_imgs)
+        seeds.append(img_seed)
+        outpaint_obj.log += "\n"
+
+    total_time = time.time() - start_time
+    text_output = f"prompt={args.prompts}"
+    text_output += f"\nnegative prompt={args.negative_prompts}"
+    text_output += f"\nmodel_id={args.hf_model_id}, ckpt_loc={args.ckpt_loc}"
+    text_output += f"\nscheduler={args.scheduler}, device={device}"
+    text_output += f"\nsteps={args.steps}, guidance_scale={args.guidance_scale}, seed={seeds}"
+    text_output += f"\nsize={args.height}x{args.width}, batch-count={batch_count}, batch-size={args.batch_size}, max_length={args.max_length}"
+    text_output += outpaint_obj.log
+    text_output += f"\nTotal image generation time: {total_time:.4f}sec"
+
+    return generated_imgs, text_output
+
+
+if __name__ == "__main__":
+    if args.clear_all:
+        clear_all()
+
+    if args.img_path is None:
+        print("Flag --img_path is required.")
+        exit()
+    if "inpaint" not in args.hf_model_id:
+        print("Please use inpainting model with --hf_model_id.")
+        exit()
+
+    dtype = torch.float32 if args.precision == "fp32" else torch.half
+    cpu_scheduling = not args.scheduler.startswith("Shark")
+    set_init_device_flags()
+    schedulers = get_schedulers(args.hf_model_id)
+    scheduler_obj = schedulers[args.scheduler]
+    seed = args.seed
+    image = Image.open(args.img_path)
+
+    outpaint_obj = OutpaintPipeline.from_pretrained(
+        scheduler_obj,
+        args.import_mlir,
+        args.hf_model_id,
+        args.ckpt_loc,
+        args.custom_vae,
+        args.precision,
+        args.max_length,
+        args.batch_size,
+        args.height,
+        args.width,
+        args.use_base_vae,
+        args.use_tuned,
+    )
+
+    for current_batch in range(args.batch_count):
+        if current_batch > 0:
+            seed = -1
+        seed = utils.sanitize_seed(seed)
+
+        start_time = time.time()
+        generated_imgs = outpaint_obj.generate_images(
+            args.prompts,
+            args.negative_prompts,
+            image,
+            args.pixels,
+            args.mask_blur,
+            args.left,
+            args.right,
+            args.top,
+            args.bottom,
+            args.noise_q,
+            args.color_variation,
+            args.batch_size,
+            args.height,
+            args.width,
+            args.steps,
+            args.guidance_scale,
+            seed,
+            args.max_length,
+            dtype,
+            args.use_base_vae,
+            cpu_scheduling,
+        )
+        total_time = time.time() - start_time
+        text_output = f"prompt={args.prompts}"
+        text_output += f"\nnegative prompt={args.negative_prompts}"
+        text_output += (
+            f"\nmodel_id={args.hf_model_id}, ckpt_loc={args.ckpt_loc}"
+        )
+        text_output += f"\nscheduler={args.scheduler}, device={args.device}"
+        text_output += f"\nsteps={args.steps}, guidance_scale={args.guidance_scale}, seed={seed}, size={args.height}x{args.width}"
+        text_output += (
+            f", batch size={args.batch_size}, max_length={args.max_length}"
+        )
+        text_output += outpaint_obj.log
+        text_output += f"\nTotal image generation time: {total_time:.4f}sec"
+
+        save_output_img(generated_imgs[0], seed)
+        print(text_output)

apps/stable_diffusion/shark_sd.spec

@@ -15,12 +15,11 @@ datas += copy_metadata('filelock')
 datas += copy_metadata('numpy')
 datas += copy_metadata('tokenizers')
 datas += copy_metadata('importlib_metadata')
-datas += copy_metadata('torchvision')
 datas += copy_metadata('torch-mlir')
-datas += copy_metadata('diffusers')
-datas += copy_metadata('transformers')
 datas += copy_metadata('omegaconf')
 datas += copy_metadata('safetensors')
+datas += collect_data_files('diffusers')
+datas += collect_data_files('transformers')
 datas += collect_data_files('gradio')
 datas += collect_data_files('iree')
 datas += collect_data_files('google-cloud-storage')
@@ -44,7 +43,7 @@ a = Analysis(
     pathex=['.'],
     binaries=binaries,
     datas=datas,
-    hiddenimports=['shark', 'shark.*', 'shark.shark_inference', 'shark_inference', 'iree.tools.core', 'gradio', 'apps'],
+    hiddenimports=['shark', 'shark.shark_inference', 'apps'],
     hookspath=[],
     hooksconfig={},
     runtime_hooks=[],

apps/stable_diffusion/shark_sd_cli.spec

@@ -15,12 +15,11 @@ datas += copy_metadata('filelock')
 datas += copy_metadata('numpy')
 datas += copy_metadata('tokenizers')
 datas += copy_metadata('importlib_metadata')
-datas += copy_metadata('torchvision')
 datas += copy_metadata('torch-mlir')
-datas += copy_metadata('diffusers')
-datas += copy_metadata('transformers')
 datas += copy_metadata('omegaconf')
 datas += copy_metadata('safetensors')
+datas += collect_data_files('diffusers')
+datas += collect_data_files('transformers')
 datas += collect_data_files('gradio')
 datas += collect_data_files('iree')
 datas += collect_data_files('google-cloud-storage')
@@ -42,7 +41,7 @@ a = Analysis(
     pathex=['.'],
     binaries=binaries,
     datas=datas,
-    hiddenimports=['shark', 'shark.*', 'shark.shark_inference', 'shark_inference', 'iree.tools.core', 'gradio', 'apps'],
+    hiddenimports=['shark', 'shark.shark_inference', 'apps'],
     hookspath=[],
     hooksconfig={},
     runtime_hooks=[],

apps/stable_diffusion/src/__init__.py

@@ -8,7 +8,8 @@ from apps.stable_diffusion.src.utils import (
 )
 from apps.stable_diffusion.src.pipelines import (
     Text2ImagePipeline,
-    InpaintPipeline,
     Image2ImagePipeline,
+    InpaintPipeline,
+    OutpaintPipeline,
 )
 from apps.stable_diffusion.src.schedulers import get_schedulers

apps/stable_diffusion/src/models/model_wrappers.py

@@ -226,6 +226,11 @@ class SharkifyStableDiffusionModel:
                 )
                 self.in_channels = self.unet.in_channels
                 self.train(False)
+                if(args.attention_slicing is not None and args.attention_slicing != "none"):
+                    if(args.attention_slicing.isdigit()):
+                        self.unet.set_attention_slice(int(args.attention_slicing))
+                    else:
+                        self.unet.set_attention_slice(args.attention_slicing)
 
             def forward(
                 self, latent, timestep, text_embedding, guidance_scale

apps/stable_diffusion/src/pipelines/__init__.py

@@ -1,9 +1,12 @@
 from apps.stable_diffusion.src.pipelines.pipeline_shark_stable_diffusion_txt2img import (
     Text2ImagePipeline,
 )
-from apps.stable_diffusion.src.pipelines.pipeline_shark_stable_diffusion_inpaint import (
-    InpaintPipeline,
-)
 from apps.stable_diffusion.src.pipelines.pipeline_shark_stable_diffusion_img2img import (
     Image2ImagePipeline,
 )
+from apps.stable_diffusion.src.pipelines.pipeline_shark_stable_diffusion_inpaint import (
+    InpaintPipeline,
+)
+from apps.stable_diffusion.src.pipelines.pipeline_shark_stable_diffusion_outpaint import (
+    OutpaintPipeline,
+)

apps/stable_diffusion/src/pipelines/pipeline_shark_stable_diffusion_img2img.py

@@ -14,6 +14,7 @@ from diffusers import (
     EulerDiscreteScheduler,
     EulerAncestralDiscreteScheduler,
     DPMSolverMultistepScheduler,
+    DEISMultistepScheduler,
 )
 from apps.stable_diffusion.src.schedulers import SharkEulerDiscreteScheduler
 from apps.stable_diffusion.src.pipelines.pipeline_shark_stable_diffusion_utils import (
@@ -37,6 +38,7 @@ class Image2ImagePipeline(StableDiffusionPipeline):
             EulerAncestralDiscreteScheduler,
             DPMSolverMultistepScheduler,
             SharkEulerDiscreteScheduler,
+            DEISMultistepScheduler,
         ],
     ):
         super().__init__(vae, text_encoder, tokenizer, unet, scheduler)

apps/stable_diffusion/src/pipelines/pipeline_shark_stable_diffusion_inpaint.py

@@ -13,6 +13,7 @@ from diffusers import (
     EulerDiscreteScheduler,
     EulerAncestralDiscreteScheduler,
     DPMSolverMultistepScheduler,
+    DEISMultistepScheduler,
 )
 from apps.stable_diffusion.src.schedulers import SharkEulerDiscreteScheduler
 from apps.stable_diffusion.src.pipelines.pipeline_shark_stable_diffusion_utils import (
@@ -36,13 +37,16 @@ class InpaintPipeline(StableDiffusionPipeline):
             EulerAncestralDiscreteScheduler,
             DPMSolverMultistepScheduler,
             SharkEulerDiscreteScheduler,
+            DEISMultistepScheduler,
         ],
     ):
         super().__init__(vae, text_encoder, tokenizer, unet, scheduler)
         self.vae_encode = vae_encode
 
-    def prepare_mask_and_masked_image(self, image, mask):
+    def prepare_mask_and_masked_image(self, image, mask, height, width):
         # preprocess image
+        image = image.resize((width, height))
+        mask = mask.resize((width, height))
         if isinstance(image, (Image.Image, np.ndarray)):
             image = [image]
 
@@ -191,7 +195,7 @@ class InpaintPipeline(StableDiffusionPipeline):
 
         # Preprocess mask and image
         mask, masked_image = self.prepare_mask_and_masked_image(
-            image, mask_image
+            image, mask_image, height, width
         )
 
         # Prepare mask latent variables

apps/stable_diffusion/src/pipelines/pipeline_shark_stable_diffusion_outpaint.py

@@ -0,0 +1,542 @@
+import torch
+from tqdm.auto import tqdm
+import numpy as np
+from random import randint
+from PIL import Image, ImageDraw, ImageFilter
+from transformers import CLIPTokenizer
+from typing import Union
+from shark.shark_inference import SharkInference
+from diffusers import (
+    DDIMScheduler,
+    PNDMScheduler,
+    LMSDiscreteScheduler,
+    EulerDiscreteScheduler,
+    EulerAncestralDiscreteScheduler,
+    DPMSolverMultistepScheduler,
+    DEISMultistepScheduler,
+)
+from apps.stable_diffusion.src.schedulers import SharkEulerDiscreteScheduler
+from apps.stable_diffusion.src.pipelines.pipeline_shark_stable_diffusion_utils import (
+    StableDiffusionPipeline,
+)
+import math
+
+
+class OutpaintPipeline(StableDiffusionPipeline):
+    def __init__(
+        self,
+        vae_encode: SharkInference,
+        vae: SharkInference,
+        text_encoder: SharkInference,
+        tokenizer: CLIPTokenizer,
+        unet: SharkInference,
+        scheduler: Union[
+            DDIMScheduler,
+            PNDMScheduler,
+            LMSDiscreteScheduler,
+            EulerDiscreteScheduler,
+            EulerAncestralDiscreteScheduler,
+            DPMSolverMultistepScheduler,
+            SharkEulerDiscreteScheduler,
+            DEISMultistepScheduler,
+        ],
+    ):
+        super().__init__(vae, text_encoder, tokenizer, unet, scheduler)
+        self.vae_encode = vae_encode
+
+    def prepare_latents(
+        self,
+        batch_size,
+        height,
+        width,
+        generator,
+        num_inference_steps,
+        dtype,
+    ):
+        latents = torch.randn(
+            (
+                batch_size,
+                4,
+                height // 8,
+                width // 8,
+            ),
+            generator=generator,
+            dtype=torch.float32,
+        ).to(dtype)
+
+        self.scheduler.set_timesteps(num_inference_steps)
+        self.scheduler.is_scale_input_called = True
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    def prepare_mask_and_masked_image(
+        self, image, mask, mask_blur, width, height
+    ):
+        if mask_blur > 0:
+            mask = mask.filter(ImageFilter.GaussianBlur(mask_blur))
+        image = image.resize((width, height))
+        mask = mask.resize((width, height))
+
+        # preprocess image
+        if isinstance(image, (Image.Image, np.ndarray)):
+            image = [image]
+
+        if isinstance(image, list) and isinstance(image[0], Image.Image):
+            image = [np.array(i.convert("RGB"))[None, :] for i in image]
+            image = np.concatenate(image, axis=0)
+        elif isinstance(image, list) and isinstance(image[0], np.ndarray):
+            image = np.concatenate([i[None, :] for i in image], axis=0)
+
+        image = image.transpose(0, 3, 1, 2)
+        image = torch.from_numpy(image).to(dtype=torch.float32) / 127.5 - 1.0
+
+        # preprocess mask
+        if isinstance(mask, (Image.Image, np.ndarray)):
+            mask = [mask]
+
+        if isinstance(mask, list) and isinstance(mask[0], Image.Image):
+            mask = np.concatenate(
+                [np.array(m.convert("L"))[None, None, :] for m in mask], axis=0
+            )
+            mask = mask.astype(np.float32) / 255.0
+        elif isinstance(mask, list) and isinstance(mask[0], np.ndarray):
+            mask = np.concatenate([m[None, None, :] for m in mask], axis=0)
+
+        mask[mask < 0.5] = 0
+        mask[mask >= 0.5] = 1
+        mask = torch.from_numpy(mask)
+
+        masked_image = image * (mask < 0.5)
+
+        return mask, masked_image
+
+    def prepare_mask_latents(
+        self,
+        mask,
+        masked_image,
+        batch_size,
+        height,
+        width,
+        dtype,
+    ):
+        mask = torch.nn.functional.interpolate(
+            mask, size=(height // 8, width // 8)
+        )
+        mask = mask.to(dtype)
+
+        masked_image = masked_image.to(dtype)
+        masked_image_latents = self.vae_encode("forward", (masked_image,))
+        masked_image_latents = torch.from_numpy(masked_image_latents)
+
+        # duplicate mask and masked_image_latents for each generation per prompt, using mps friendly method
+        if mask.shape[0] < batch_size:
+            if not batch_size % mask.shape[0] == 0:
+                raise ValueError(
+                    "The passed mask and the required batch size don't match. Masks are supposed to be duplicated to"
+                    f" a total batch size of {batch_size}, but {mask.shape[0]} masks were passed. Make sure the number"
+                    " of masks that you pass is divisible by the total requested batch size."
+                )
+            mask = mask.repeat(batch_size // mask.shape[0], 1, 1, 1)
+        if masked_image_latents.shape[0] < batch_size:
+            if not batch_size % masked_image_latents.shape[0] == 0:
+                raise ValueError(
+                    "The passed images and the required batch size don't match. Images are supposed to be duplicated"
+                    f" to a total batch size of {batch_size}, but {masked_image_latents.shape[0]} images were passed."
+                    " Make sure the number of images that you pass is divisible by the total requested batch size."
+                )
+            masked_image_latents = masked_image_latents.repeat(
+                batch_size // masked_image_latents.shape[0], 1, 1, 1
+            )
+        return mask, masked_image_latents
+
+    def get_matched_noise(
+        self, _np_src_image, np_mask_rgb, noise_q=1, color_variation=0.05
+    ):
+        # helper fft routines that keep ortho normalization and auto-shift before and after fft
+        def _fft2(data):
+            if data.ndim > 2:  # has channels
+                out_fft = np.zeros(
+                    (data.shape[0], data.shape[1], data.shape[2]),
+                    dtype=np.complex128,
+                )
+                for c in range(data.shape[2]):
+                    c_data = data[:, :, c]
+                    out_fft[:, :, c] = np.fft.fft2(
+                        np.fft.fftshift(c_data), norm="ortho"
+                    )
+                    out_fft[:, :, c] = np.fft.ifftshift(out_fft[:, :, c])
+            else:  # one channel
+                out_fft = np.zeros(
+                    (data.shape[0], data.shape[1]), dtype=np.complex128
+                )
+                out_fft[:, :] = np.fft.fft2(
+                    np.fft.fftshift(data), norm="ortho"
+                )
+                out_fft[:, :] = np.fft.ifftshift(out_fft[:, :])
+
+            return out_fft
+
+        def _ifft2(data):
+            if data.ndim > 2:  # has channels
+                out_ifft = np.zeros(
+                    (data.shape[0], data.shape[1], data.shape[2]),
+                    dtype=np.complex128,
+                )
+                for c in range(data.shape[2]):
+                    c_data = data[:, :, c]
+                    out_ifft[:, :, c] = np.fft.ifft2(
+                        np.fft.fftshift(c_data), norm="ortho"
+                    )
+                    out_ifft[:, :, c] = np.fft.ifftshift(out_ifft[:, :, c])
+            else:  # one channel
+                out_ifft = np.zeros(
+                    (data.shape[0], data.shape[1]), dtype=np.complex128
+                )
+                out_ifft[:, :] = np.fft.ifft2(
+                    np.fft.fftshift(data), norm="ortho"
+                )
+                out_ifft[:, :] = np.fft.ifftshift(out_ifft[:, :])
+
+            return out_ifft
+
+        def _get_gaussian_window(width, height, std=3.14, mode=0):
+            window_scale_x = float(width / min(width, height))
+            window_scale_y = float(height / min(width, height))
+
+            window = np.zeros((width, height))
+            x = (np.arange(width) / width * 2.0 - 1.0) * window_scale_x
+            for y in range(height):
+                fy = (y / height * 2.0 - 1.0) * window_scale_y
+                if mode == 0:
+                    window[:, y] = np.exp(-(x**2 + fy**2) * std)
+                else:
+                    window[:, y] = (
+                        1 / ((x**2 + 1.0) * (fy**2 + 1.0))
+                    ) ** (std / 3.14)
+
+            return window
+
+        def _get_masked_window_rgb(np_mask_grey, hardness=1.0):
+            np_mask_rgb = np.zeros(
+                (np_mask_grey.shape[0], np_mask_grey.shape[1], 3)
+            )
+            if hardness != 1.0:
+                hardened = np_mask_grey[:] ** hardness
+            else:
+                hardened = np_mask_grey[:]
+            for c in range(3):
+                np_mask_rgb[:, :, c] = hardened[:]
+            return np_mask_rgb
+
+        def _match_cumulative_cdf(source, template):
+            src_values, src_unique_indices, src_counts = np.unique(
+                source.ravel(), return_inverse=True, return_counts=True
+            )
+            tmpl_values, tmpl_counts = np.unique(
+                template.ravel(), return_counts=True
+            )
+
+            # calculate normalized quantiles for each array
+            src_quantiles = np.cumsum(src_counts) / source.size
+            tmpl_quantiles = np.cumsum(tmpl_counts) / template.size
+
+            interp_a_values = np.interp(
+                src_quantiles, tmpl_quantiles, tmpl_values
+            )
+            return interp_a_values[src_unique_indices].reshape(source.shape)
+
+        def _match_histograms(image, reference):
+            if image.ndim != reference.ndim:
+                raise ValueError(
+                    "Image and reference must have the same number of channels."
+                )
+
+            if image.shape[-1] != reference.shape[-1]:
+                raise ValueError(
+                    "Number of channels in the input image and reference image must match!"
+                )
+
+            matched = np.empty(image.shape, dtype=image.dtype)
+            for channel in range(image.shape[-1]):
+                matched_channel = _match_cumulative_cdf(
+                    image[..., channel], reference[..., channel]
+                )
+                matched[..., channel] = matched_channel
+
+            matched = matched.astype(np.float64, copy=False)
+            return matched
+
+        width = _np_src_image.shape[0]
+        height = _np_src_image.shape[1]
+        num_channels = _np_src_image.shape[2]
+
+        np_src_image = _np_src_image[:] * (1.0 - np_mask_rgb)
+        np_mask_grey = np.sum(np_mask_rgb, axis=2) / 3.0
+        img_mask = np_mask_grey > 1e-6
+        ref_mask = np_mask_grey < 1e-3
+
+        # rather than leave the masked area black, we get better results from fft by filling the average unmasked color
+        windowed_image = _np_src_image * (
+            1.0 - _get_masked_window_rgb(np_mask_grey)
+        )
+        windowed_image /= np.max(windowed_image)
+        windowed_image += np.average(_np_src_image) * np_mask_rgb
+
+        src_fft = _fft2(
+            windowed_image
+        )  # get feature statistics from masked src img
+        src_dist = np.absolute(src_fft)
+        src_phase = src_fft / src_dist
+
+        # create a generator with a static seed to make outpainting deterministic / only follow global seed
+        rng = np.random.default_rng(0)
+
+        noise_window = _get_gaussian_window(
+            width, height, mode=1
+        )  # start with simple gaussian noise
+        noise_rgb = rng.random((width, height, num_channels))
+        noise_grey = np.sum(noise_rgb, axis=2) / 3.0
+        # the colorfulness of the starting noise is blended to greyscale with a parameter
+        noise_rgb *= color_variation
+        for c in range(num_channels):
+            noise_rgb[:, :, c] += (1.0 - color_variation) * noise_grey
+
+        noise_fft = _fft2(noise_rgb)
+        for c in range(num_channels):
+            noise_fft[:, :, c] *= noise_window
+        noise_rgb = np.real(_ifft2(noise_fft))
+        shaped_noise_fft = _fft2(noise_rgb)
+        shaped_noise_fft[:, :, :] = (
+            np.absolute(shaped_noise_fft[:, :, :]) ** 2
+            * (src_dist**noise_q)
+            * src_phase
+        )  # perform the actual shaping
+
+        # color_variation
+        brightness_variation = 0.0
+        contrast_adjusted_np_src = (
+            _np_src_image[:] * (brightness_variation + 1.0)
+            - brightness_variation * 2.0
+        )
+
+        shaped_noise = np.real(_ifft2(shaped_noise_fft))
+        shaped_noise -= np.min(shaped_noise)
+        shaped_noise /= np.max(shaped_noise)
+        shaped_noise[img_mask, :] = _match_histograms(
+            shaped_noise[img_mask, :] ** 1.0,
+            contrast_adjusted_np_src[ref_mask, :],
+        )
+        shaped_noise = (
+            _np_src_image[:] * (1.0 - np_mask_rgb) + shaped_noise * np_mask_rgb
+        )
+
+        matched_noise = shaped_noise[:]
+
+        return np.clip(matched_noise, 0.0, 1.0)
+
+    def generate_images(
+        self,
+        prompts,
+        neg_prompts,
+        image,
+        pixels,
+        mask_blur,
+        is_left,
+        is_right,
+        is_top,
+        is_bottom,
+        noise_q,
+        color_variation,
+        batch_size,
+        height,
+        width,
+        num_inference_steps,
+        guidance_scale,
+        seed,
+        max_length,
+        dtype,
+        use_base_vae,
+        cpu_scheduling,
+    ):
+        # prompts and negative prompts must be a list.
+        if isinstance(prompts, str):
+            prompts = [prompts]
+
+        if isinstance(neg_prompts, str):
+            neg_prompts = [neg_prompts]
+
+        prompts = prompts * batch_size
+        neg_prompts = neg_prompts * batch_size
+
+        # seed generator to create the inital latent noise. Also handle out of range seeds.
+        uint32_info = np.iinfo(np.uint32)
+        uint32_min, uint32_max = uint32_info.min, uint32_info.max
+        if seed < uint32_min or seed >= uint32_max:
+            seed = randint(uint32_min, uint32_max)
+        generator = torch.manual_seed(seed)
+
+        # Get initial latents
+        init_latents = self.prepare_latents(
+            batch_size=batch_size,
+            height=height,
+            width=width,
+            generator=generator,
+            num_inference_steps=num_inference_steps,
+            dtype=dtype,
+        )
+
+        # Get text embeddings from prompts
+        text_embeddings = self.encode_prompts(prompts, neg_prompts, max_length)
+
+        # guidance scale as a float32 tensor.
+        guidance_scale = torch.tensor(guidance_scale).to(torch.float32)
+
+        process_width = width
+        process_height = height
+        left = pixels if is_left else 0
+        right = pixels if is_right else 0
+        up = pixels if is_top else 0
+        down = pixels if is_bottom else 0
+        target_w = math.ceil((image.width + left + right) / 64) * 64
+        target_h = math.ceil((image.height + up + down) / 64) * 64
+
+        if left > 0:
+            left = left * (target_w - image.width) // (left + right)
+        if right > 0:
+            right = target_w - image.width - left
+        if up > 0:
+            up = up * (target_h - image.height) // (up + down)
+        if down > 0:
+            down = target_h - image.height - up
+
+        def expand(
+            init_img,
+            expand_pixels,
+            is_left=False,
+            is_right=False,
+            is_top=False,
+            is_bottom=False,
+        ):
+            is_horiz = is_left or is_right
+            is_vert = is_top or is_bottom
+            pixels_horiz = expand_pixels if is_horiz else 0
+            pixels_vert = expand_pixels if is_vert else 0
+
+            res_w = init_img.width + pixels_horiz
+            res_h = init_img.height + pixels_vert
+            process_res_w = math.ceil(res_w / 64) * 64
+            process_res_h = math.ceil(res_h / 64) * 64
+
+            img = Image.new("RGB", (process_res_w, process_res_h))
+            img.paste(
+                init_img,
+                (pixels_horiz if is_left else 0, pixels_vert if is_top else 0),
+            )
+
+            msk = Image.new("RGB", (process_res_w, process_res_h), "white")
+            draw = ImageDraw.Draw(msk)
+            draw.rectangle(
+                (
+                    expand_pixels + mask_blur if is_left else 0,
+                    expand_pixels + mask_blur if is_top else 0,
+                    msk.width - expand_pixels - mask_blur
+                    if is_right
+                    else res_w,
+                    msk.height - expand_pixels - mask_blur
+                    if is_bottom
+                    else res_h,
+                ),
+                fill="black",
+            )
+
+            np_image = (np.asarray(img) / 255.0).astype(np.float64)
+            np_mask = (np.asarray(msk) / 255.0).astype(np.float64)
+            noised = self.get_matched_noise(
+                np_image, np_mask, noise_q, color_variation
+            )
+            output_image = Image.fromarray(
+                np.clip(noised * 255.0, 0.0, 255.0).astype(np.uint8),
+                mode="RGB",
+            )
+
+            target_width = (
+                min(width, init_img.width + pixels_horiz)
+                if is_horiz
+                else img.width
+            )
+            target_height = (
+                min(height, init_img.height + pixels_vert)
+                if is_vert
+                else img.height
+            )
+            crop_region = (
+                0 if is_left else output_image.width - target_width,
+                0 if is_top else output_image.height - target_height,
+                target_width if is_left else output_image.width,
+                target_height if is_top else output_image.height,
+            )
+            mask_to_process = msk.crop(crop_region)
+            image_to_process = output_image.crop(crop_region)
+
+            # Preprocess mask and image
+            mask, masked_image = self.prepare_mask_and_masked_image(
+                image_to_process, mask_to_process, mask_blur, width, height
+            )
+
+            # Prepare mask latent variables
+            mask, masked_image_latents = self.prepare_mask_latents(
+                mask=mask,
+                masked_image=masked_image,
+                batch_size=batch_size,
+                height=height,
+                width=width,
+                dtype=dtype,
+            )
+
+            # Get Image latents
+            latents = self.produce_img_latents(
+                latents=init_latents,
+                text_embeddings=text_embeddings,
+                guidance_scale=guidance_scale,
+                total_timesteps=self.scheduler.timesteps,
+                dtype=dtype,
+                cpu_scheduling=cpu_scheduling,
+                mask=mask,
+                masked_image_latents=masked_image_latents,
+            )
+
+            # Img latents -> PIL images
+            all_imgs = []
+            for i in tqdm(range(0, latents.shape[0], batch_size)):
+                imgs = self.decode_latents(
+                    latents=latents[i : i + batch_size],
+                    use_base_vae=use_base_vae,
+                    cpu_scheduling=cpu_scheduling,
+                )
+                all_imgs.extend(imgs)
+
+            res_img = all_imgs[0].resize(
+                (image_to_process.width, image_to_process.height)
+            )
+            output_image.paste(
+                res_img,
+                (
+                    0 if is_left else output_image.width - res_img.width,
+                    0 if is_top else output_image.height - res_img.height,
+                ),
+            )
+            output_image = output_image.crop((0, 0, res_w, res_h))
+
+            return output_image
+
+        img = image.resize((width, height))
+        if left > 0:
+            img = expand(img, left, is_left=True)
+        if right > 0:
+            img = expand(img, right, is_right=True)
+        if up > 0:
+            img = expand(img, up, is_top=True)
+        if down > 0:
+            img = expand(img, down, is_bottom=True)
+
+        return [img]

apps/stable_diffusion/src/pipelines/pipeline_shark_stable_diffusion_txt2img.py

@@ -13,6 +13,7 @@ from diffusers import (
     EulerDiscreteScheduler,
     EulerAncestralDiscreteScheduler,
     DPMSolverMultistepScheduler,
+    DEISMultistepScheduler,
 )
 from apps.stable_diffusion.src.schedulers import SharkEulerDiscreteScheduler
 from apps.stable_diffusion.src.pipelines.pipeline_shark_stable_diffusion_utils import (
@@ -36,6 +37,7 @@ class Text2ImagePipeline(StableDiffusionPipeline):
             EulerAncestralDiscreteScheduler,
             DPMSolverMultistepScheduler,
             SharkEulerDiscreteScheduler,
+            DEISMultistepScheduler,
         ],
     ):
         super().__init__(vae, text_encoder, tokenizer, unet, scheduler)

apps/stable_diffusion/src/pipelines/pipeline_shark_stable_diffusion_utils.py

@@ -13,6 +13,7 @@ from diffusers import (
     EulerDiscreteScheduler,
     EulerAncestralDiscreteScheduler,
     DPMSolverMultistepScheduler,
+    DEISMultistepScheduler,
 )
 from shark.shark_inference import SharkInference
 from apps.stable_diffusion.src.schedulers import SharkEulerDiscreteScheduler
@@ -46,6 +47,7 @@ class StableDiffusionPipeline:
             EulerAncestralDiscreteScheduler,
             DPMSolverMultistepScheduler,
             SharkEulerDiscreteScheduler,
+            DEISMultistepScheduler,
         ],
     ):
         self.vae = vae
@@ -189,6 +191,7 @@ class StableDiffusionPipeline:
             EulerAncestralDiscreteScheduler,
             DPMSolverMultistepScheduler,
             SharkEulerDiscreteScheduler,
+            DEISMultistepScheduler,
         ],
         import_mlir: bool,
         model_id: str,
@@ -217,7 +220,11 @@ class StableDiffusionPipeline:
                 use_tuned=use_tuned,
                 low_cpu_mem_usage=low_cpu_mem_usage,
             )
-            if cls.__name__ in ["Image2ImagePipeline", "InpaintPipeline"]:
+            if cls.__name__ in [
+                "Image2ImagePipeline",
+                "InpaintPipeline",
+                "OutpaintPipeline",
+            ]:
                 clip, unet, vae, vae_encode = mlir_import()
                 return cls(
                     vae_encode, vae, clip, get_tokenizer(), unet, scheduler
@@ -225,7 +232,11 @@ class StableDiffusionPipeline:
             clip, unet, vae = mlir_import()
             return cls(vae, clip, get_tokenizer(), unet, scheduler)
         try:
-            if cls.__name__ in ["Image2ImagePipeline", "InpaintPipeline"]:
+            if cls.__name__ in [
+                "Image2ImagePipeline",
+                "InpaintPipeline",
+                "OutpaintPipeline",
+            ]:
                 return cls(
                     get_vae_encode(),
                     get_vae(),
@@ -252,7 +263,11 @@ class StableDiffusionPipeline:
                 use_tuned=use_tuned,
                 low_cpu_mem_usage=low_cpu_mem_usage,
             )
-            if cls.__name__ in ["Image2ImagePipeline", "InpaintPipeline"]:
+            if cls.__name__ in [
+                "Image2ImagePipeline",
+                "InpaintPipeline",
+                "OutpaintPipeline",
+            ]:
                 clip, unet, vae, vae_encode = mlir_import()
                 return cls(
                     vae_encode, vae, clip, get_tokenizer(), unet, scheduler

apps/stable_diffusion/src/schedulers/sd_schedulers.py

@@ -6,6 +6,7 @@ from diffusers import (
     KDPM2DiscreteScheduler,
     EulerDiscreteScheduler,
     EulerAncestralDiscreteScheduler,
+    DEISMultistepScheduler,
 )
 from apps.stable_diffusion.src.schedulers.shark_eulerdiscrete import (
     SharkEulerDiscreteScheduler,
@@ -46,6 +47,10 @@ def get_schedulers(model_id):
         model_id,
         subfolder="scheduler",
     )
+    schedulers["DEISMultistep"] = DEISMultistepScheduler.from_pretrained(
+        model_id,
+        subfolder="scheduler",
+    )
     schedulers[
         "SharkEulerDiscrete"
     ] = SharkEulerDiscreteScheduler.from_pretrained(

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

@@ -111,62 +111,6 @@
             }
         }
     },
-    "runwayml/stable-diffusion-inpainting": {
-        "unet": {
-            "latents": {
-                "shape": [
-                    "1*batch_size",
-                    9,
-                    "height",
-                    "width"
-                ],
-                "dtype": "f32"
-            },
-            "timesteps": {
-                "shape": [
-                    1
-                ],
-                "dtype": "f32"
-            },
-            "embedding": {
-                "shape": [
-                    "2*batch_size",
-                    "max_len",
-                    768
-                ],
-                "dtype": "f32"
-            },
-            "guidance_scale": {
-                "shape": 2,
-                "dtype": "f32"
-            }
-        },
-        "vae_encode": {
-            "image" : {
-                "shape" : [
-                    "1*batch_size",3,"8*height","8*width"
-                ],
-                "dtype":"f32"
-            }
-        },
-        "vae": {
-            "latents" : {
-                "shape" : [
-                    "1*batch_size",4,"height","width"
-                ],
-                "dtype":"f32"
-            }
-        },
-        "clip": {
-            "token" : {
-                "shape" : [
-                    "2*batch_size",
-                    "max_len"
-                ],
-                "dtype":"i64"
-            }
-        }
-    },
     "stabilityai/stable-diffusion-2-inpainting": {
         "unet": {
             "latents": {
@@ -222,5 +166,61 @@
                 "dtype":"i64"
             }
         }
+    },
+    "runwayml/stable-diffusion-inpainting": {
+        "unet": {
+            "latents": {
+                "shape": [
+                    "1*batch_size",
+                    9,
+                    "height",
+                    "width"
+                ],
+                "dtype": "f32"
+            },
+            "timesteps": {
+                "shape": [
+                    1
+                ],
+                "dtype": "f32"
+            },
+            "embedding": {
+                "shape": [
+                    "2*batch_size",
+                    "max_len",
+                    768
+                ],
+                "dtype": "f32"
+            },
+            "guidance_scale": {
+                "shape": 2,
+                "dtype": "f32"
+            }
+        },
+        "vae_encode": {
+            "image" : {
+                "shape" : [
+                    "1*batch_size",3,"8*height","8*width"
+                ],
+                "dtype":"f32"
+            }
+        },
+        "vae": {
+            "latents" : {
+                "shape" : [
+                    "1*batch_size",4,"height","width"
+                ],
+                "dtype":"f32"
+            }
+        },
+        "clip": {
+            "token" : {
+                "shape" : [
+                    "2*batch_size",
+                    "max_len"
+                ],
+                "dtype":"i64"
+            }
+        }
     }
 }

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

@@ -40,17 +40,6 @@
     "stablediffusion/v2_1/vae/fp16/length_77/untuned":"vae77_512_512_fp16_stabilityai_stable_diffusion_2_1_base",
     "stablediffusion/v2_1/vae/fp16/length_77/untuned/base":"vae2_8dec_fp16",
     "stablediffusion/v2_1/clip/fp32/length_77/untuned":"clip77_512_512_fp16_stabilityai_stable_diffusion_2_1_base",
-    "stablediffusion/inpaint_v1/unet/fp16/length_77/untuned":"unet_inpaint_fp16",
-    "stablediffusion/inpaint_v1/unet/fp32/length_77/untuned":"unet_inpaint_fp32",
-    "stablediffusion/inpaint_v1/vae_encode/fp16/length_77/untuned":"vae_encode_inpaint_fp16",
-    "stablediffusion/inpaint_v1/vae_encode/fp32/length_77/untuned":"vae_encode_inpaint_fp32",
-    "stablediffusion/inpaint_v1/vae/fp16/length_77/untuned":"vae_inpaint_fp16",
-    "stablediffusion/inpaint_v1/vae/fp32/length_77/untuned":"vae_inpaint_fp32",
-    "stablediffusion/inpaint_v1/clip/fp32/length_77/untuned":"clip_inpaint_fp32",
-    "stablediffusion/inpaint_v2/unet/fp16/length_77/untuned":"unet_inpaint_fp16",
-    "stablediffusion/inpaint_v2/vae_encode/fp16/length_77/untuned":"vae_encode_inpaint_fp16",
-    "stablediffusion/inpaint_v2/vae/fp16/length_77/untuned":"vae_inpaint_fp16",
-    "stablediffusion/inpaint_v2/clip/fp32/length_77/untuned":"clip_inpaint_fp32",
     "anythingv3/v1_4/unet/fp16/length_77/untuned":"av3_unet_19dec_fp16",
     "anythingv3/v1_4/unet/fp16/length_77/tuned":"av3_unet_19dec_fp16_tuned",
     "anythingv3/v1_4/unet/fp16/length_77/tuned/cuda":"av3_unet_19dec_fp16_cuda_tuned",

apps/stable_diffusion/src/utils/sd_annotation.py

@@ -82,14 +82,20 @@ def load_lower_configs():
         fetch_and_update_base_model_id,
     )
 
-    base_model_id = args.hf_model_id
     if args.ckpt_loc != "":
         base_model_id = fetch_and_update_base_model_id(args.ckpt_loc)
-    if base_model_id == "runwayml/stable-diffusion-v1-5":
-        base_model_id = "CompVis/stable-diffusion-v1-4"
+    else:
+        base_model_id = fetch_and_update_base_model_id(args.hf_model_id)
+        if base_model_id == "":
+            base_model_id = args.hf_model_id
 
     variant, version = get_variant_version(base_model_id)
 
+    if version == "inpaint_v1":
+        version = "v1_4"
+    elif version == "inpaint_v2":
+        version = "v2_1base"
+
     config_bucket = "gs://shark_tank/sd_tuned_configs/"
 
     device, device_spec_args = get_device_args()
@@ -136,11 +142,13 @@ def annotate_with_winograd(input_mlir, winograd_config_dir, model_name):
 
     if args.save_annotation:
         if model_name.split("_")[-1] != "tuned":
-            out_file_path = (
-                f"{args.annotation_output}/{model_name}_tuned_torch.mlir"
+            out_file_path = os.path.join(
+                args.annotation_output, model_name + "_tuned_torch.mlir"
             )
         else:
-            out_file_path = f"{args.annotation_output}/{model_name}_torch.mlir"
+            out_file_path = os.path.join(
+                args.annotation_output, model_name + "_torch.mlir"
+            )
         with open(out_file_path, "w") as f:
             f.write(str(winograd_model))
             f.close()

apps/stable_diffusion/src/utils/stable_args.py

@@ -35,12 +35,6 @@ p.add_argument(
     help="Path to the image input for img2img/inpainting",
 )
 
-p.add_argument(
-    "--mask_path",
-    type=str,
-    help="Path to the mask image input for inpainting",
-)
-
 p.add_argument(
     "--steps",
     type=int,
@@ -67,6 +61,7 @@ p.add_argument(
     "--height",
     type=int,
     default=512,
+    choices=range(384, 768, 8),
     help="the height of the output image.",
 )
 
@@ -74,6 +69,7 @@ p.add_argument(
     "--width",
     type=int,
     default=512,
+    choices=range(384, 768, 8),
     help="the width of the output image.",
 )
 
@@ -97,6 +93,75 @@ p.add_argument(
     default=0.8,
     help="the strength of change applied on the given input image for img2img",
 )
+
+##############################################################################
+### Inpainting and Outpainting Params
+##############################################################################
+
+p.add_argument(
+    "--mask_path",
+    type=str,
+    help="Path to the mask image input for inpainting",
+)
+
+p.add_argument(
+    "--pixels",
+    type=int,
+    default=128,
+    choices=range(8, 256, 8),
+    help="Number of expended pixels for one direction for outpainting",
+)
+
+p.add_argument(
+    "--mask_blur",
+    type=int,
+    default=8,
+    choices=range(0, 64),
+    help="Number of blur pixels for outpainting",
+)
+
+p.add_argument(
+    "--left",
+    default=False,
+    action=argparse.BooleanOptionalAction,
+    help="If expend left for outpainting",
+)
+
+p.add_argument(
+    "--right",
+    default=False,
+    action=argparse.BooleanOptionalAction,
+    help="If expend right for outpainting",
+)
+
+p.add_argument(
+    "--top",
+    default=False,
+    action=argparse.BooleanOptionalAction,
+    help="If expend top for outpainting",
+)
+
+p.add_argument(
+    "--bottom",
+    default=False,
+    action=argparse.BooleanOptionalAction,
+    help="If expend bottom for outpainting",
+)
+
+p.add_argument(
+    "--noise_q",
+    type=float,
+    default=1.0,
+    help="Fall-off exponent for outpainting (lower=higher detail) (min=0.0, max=4.0)",
+)
+
+p.add_argument(
+    "--color_variation",
+    type=float,
+    default=0.05,
+    help="Color variation for outpainting (min=0.0, max=1.0)",
+)
+
 ##############################################################################
 ### Model Config and Usage Params
 ##############################################################################
@@ -200,6 +265,13 @@ p.add_argument(
     help="Use the accelerate package to reduce cpu memory consumption",
 )
 
+p.add_argument(
+    "--attention_slicing",
+    type=str,
+    default="none",
+    help="Amount of attention slicing to use (one of 'max', 'auto', 'none', or an integer)",
+)
+
 ##############################################################################
 ### IREE - Vulkan supported flags
 ##############################################################################

apps/stable_diffusion/src/utils/utils.py

@@ -239,17 +239,15 @@ def set_init_device_flags():
         args.max_length = 64
 
     # Use tuned models in the case of fp16, vulkan rdna3 or cuda sm devices.
-    base_model_id = args.hf_model_id
     if args.ckpt_loc != "":
         base_model_id = fetch_and_update_base_model_id(args.ckpt_loc)
+    else:
+        base_model_id = fetch_and_update_base_model_id(args.hf_model_id)
+        if base_model_id == "":
+            base_model_id = args.hf_model_id
 
     if (
-        args.hf_model_id
-        in [
-            "runwayml/stable-diffusion-inpainting",
-            "stabilityai/stable-diffusion-2-inpainting",
-        ]
-        or args.precision != "fp16"
+        args.precision != "fp16"
         or args.height != 512
         or args.width != 512
         or args.batch_size != 1
@@ -257,7 +255,7 @@ def set_init_device_flags():
     ):
         args.use_tuned = False
 
-    elif args.ckpt_loc != "" and base_model_id not in [
+    elif base_model_id not in [
         "Linaqruf/anything-v3.0",
         "dreamlike-art/dreamlike-diffusion-1.0",
         "prompthero/openjourney",
@@ -266,6 +264,8 @@ def set_init_device_flags():
         "stabilityai/stable-diffusion-2-1-base",
         "CompVis/stable-diffusion-v1-4",
         "runwayml/stable-diffusion-v1-5",
+        "runwayml/stable-diffusion-inpainting",
+        "stabilityai/stable-diffusion-2-inpainting",
     ]:
         args.use_tuned = False
 
@@ -300,8 +300,6 @@ def set_init_device_flags():
         "stabilityai/stable-diffusion-2-1",
         "stabilityai/stable-diffusion-2-1-base",
         "CompVis/stable-diffusion-v1-4",
-        "runwayml/stable-diffusion-inpainting",
-        "stabilityai/stable-diffusion-2-inpainting",
     ]:
         args.import_mlir = True
 
@@ -439,10 +437,12 @@ def preprocessCKPT(custom_weights):
     print(
         "Loading diffusers' pipeline from original stable diffusion checkpoint"
     )
+    num_in_channels = 9 if "inpainting" in custom_weights else 4
     pipe = load_pipeline_from_original_stable_diffusion_ckpt(
         checkpoint_path=custom_weights,
         extract_ema=extract_ema,
         from_safetensors=from_safetensors,
+        num_in_channels=num_in_channels,
     )
     pipe.save_pretrained(path_to_diffusers)
     print("Loading complete")

apps/stable_diffusion/web/index.py

@@ -27,11 +27,16 @@ def resource_path(relative_path):
 
 dark_theme = resource_path("ui/css/sd_dark_theme.css")
 
-from apps.stable_diffusion.web.ui import txt2img_web, img2img_web
+from apps.stable_diffusion.web.ui import (
+    txt2img_web,
+    img2img_web,
+    inpaint_web,
+    outpaint_web,
+)
 
 sd_web = gr.TabbedInterface(
-    [txt2img_web, img2img_web],
-    ["Text-to-Image", "Image-to-Image"],
+    [txt2img_web, img2img_web, inpaint_web, outpaint_web],
+    ["Text-to-Image", "Image-to-Image", "Inpainting", "Outpainting"],
     css=dark_theme,
 )
 

apps/stable_diffusion/web/ui/__init__.py

@@ -1,2 +1,4 @@
 from apps.stable_diffusion.web.ui.txt2img_ui import txt2img_web
 from apps.stable_diffusion.web.ui.img2img_ui import img2img_web
+from apps.stable_diffusion.web.ui.inpaint_ui import inpaint_web
+from apps.stable_diffusion.web.ui.outpaint_ui import outpaint_web

apps/stable_diffusion/web/ui/img2img_ui.py

@@ -75,7 +75,7 @@ with gr.Blocks(title="Image-to-Image") as img2img_web:
                         elem_id="negative_prompt_box",
                     )
 
-                init_image = gr.Image(label="Input Image", type="filepath")
+                init_image = gr.Image(label="Input Image", type="pil")
 
                 with gr.Accordion(label="Advanced Options", open=False):
                     with gr.Row():
@@ -133,7 +133,7 @@ with gr.Blocks(title="Image-to-Image") as img2img_web:
                             0,
                             1,
                             value=args.strength,
-                            step=0.1,
+                            step=0.01,
                             label="Strength",
                         )
                     with gr.Row():

apps/stable_diffusion/web/ui/inpaint_ui.py

@@ -0,0 +1,224 @@
+import os
+import sys
+import glob
+from pathlib import Path
+import gradio as gr
+from PIL import Image
+from apps.stable_diffusion.scripts import inpaint_inf
+from apps.stable_diffusion.src import args
+from apps.stable_diffusion.web.ui.utils import (
+    available_devices,
+    nodlogo_loc,
+)
+
+
+with gr.Blocks(title="Inpainting") as inpaint_web:
+    with gr.Row(elem_id="ui_title"):
+        nod_logo = Image.open(nodlogo_loc)
+        with gr.Row():
+            with gr.Column(scale=1, elem_id="demo_title_outer"):
+                gr.Image(
+                    value=nod_logo,
+                    show_label=False,
+                    interactive=False,
+                    elem_id="top_logo",
+                ).style(width=150, height=50)
+    with gr.Row(elem_id="ui_body"):
+        with gr.Row():
+            with gr.Column(scale=1, min_width=600):
+                with gr.Row():
+                    ckpt_path = (
+                        Path(args.ckpt_dir)
+                        if args.ckpt_dir
+                        else Path(Path.cwd(), "models")
+                    )
+                    ckpt_path.mkdir(parents=True, exist_ok=True)
+                    types = (
+                        "*.ckpt",
+                        "*.safetensors",
+                    )  # the tuple of file types
+                    ckpt_files = ["None"]
+                    for extn in types:
+                        files = glob.glob(os.path.join(ckpt_path, extn))
+                        ckpt_files.extend(files)
+                    custom_model = gr.Dropdown(
+                        label=f"Models (Custom Model path: {ckpt_path})",
+                        value=args.ckpt_loc if args.ckpt_loc else "None",
+                        choices=ckpt_files
+                        + [
+                            "runwayml/stable-diffusion-inpainting",
+                            "stabilityai/stable-diffusion-2-inpainting",
+                        ],
+                    )
+                    hf_model_id = gr.Textbox(
+                        placeholder="Select 'None' in the Models dropdown on the left and enter model ID here e.g: SG161222/Realistic_Vision_V1.3",
+                        value="",
+                        label="HuggingFace Model ID",
+                        lines=3,
+                    )
+
+                with gr.Group(elem_id="prompt_box_outer"):
+                    prompt = gr.Textbox(
+                        label="Prompt",
+                        value=args.prompts[0],
+                        lines=1,
+                        elem_id="prompt_box",
+                    )
+                    negative_prompt = gr.Textbox(
+                        label="Negative Prompt",
+                        value=args.negative_prompts[0],
+                        lines=1,
+                        elem_id="negative_prompt_box",
+                    )
+
+                init_image = gr.Image(
+                    label="Masked Image",
+                    source="upload",
+                    tool="sketch",
+                    type="pil",
+                )
+
+                with gr.Accordion(label="Advanced Options", open=False):
+                    with gr.Row():
+                        scheduler = gr.Dropdown(
+                            label="Scheduler",
+                            value="PNDM",
+                            choices=[
+                                "DDIM",
+                                "PNDM",
+                                "DPMSolverMultistep",
+                                "EulerAncestralDiscrete",
+                            ],
+                        )
+                        with gr.Group():
+                            save_metadata_to_png = gr.Checkbox(
+                                label="Save prompt information to PNG",
+                                value=args.write_metadata_to_png,
+                                interactive=True,
+                            )
+                            save_metadata_to_json = gr.Checkbox(
+                                label="Save prompt information to JSON file",
+                                value=args.save_metadata_to_json,
+                                interactive=True,
+                            )
+                    with gr.Row():
+                        height = gr.Slider(
+                            384, 786, value=args.height, step=8, label="Height"
+                        )
+                        width = gr.Slider(
+                            384, 786, value=args.width, step=8, label="Width"
+                        )
+                        precision = gr.Radio(
+                            label="Precision",
+                            value=args.precision,
+                            choices=[
+                                "fp16",
+                                "fp32",
+                            ],
+                            visible=False,
+                        )
+                        max_length = gr.Radio(
+                            label="Max Length",
+                            value=args.max_length,
+                            choices=[
+                                64,
+                                77,
+                            ],
+                            visible=False,
+                        )
+                    with gr.Row():
+                        steps = gr.Slider(
+                            1, 100, value=args.steps, step=1, label="Steps"
+                        )
+                    with gr.Row():
+                        guidance_scale = gr.Slider(
+                            0,
+                            50,
+                            value=args.guidance_scale,
+                            step=0.1,
+                            label="CFG Scale",
+                        )
+                        batch_count = gr.Slider(
+                            1,
+                            100,
+                            value=args.batch_count,
+                            step=1,
+                            label="Batch Count",
+                            interactive=True,
+                        )
+                        batch_size = gr.Slider(
+                            1,
+                            4,
+                            value=args.batch_size,
+                            step=1,
+                            label="Batch Size",
+                            interactive=False,
+                            visible=False,
+                        )
+                with gr.Row():
+                    seed = gr.Number(
+                        value=args.seed, precision=0, label="Seed"
+                    )
+                    device = gr.Dropdown(
+                        label="Device",
+                        value=available_devices[0],
+                        choices=available_devices,
+                    )
+                with gr.Row():
+                    random_seed = gr.Button("Randomize Seed")
+                    random_seed.click(
+                        None,
+                        inputs=[],
+                        outputs=[seed],
+                        _js="() => Math.floor(Math.random() * 4294967295)",
+                    )
+                    stable_diffusion = gr.Button("Generate Image(s)")
+
+            with gr.Column(scale=1, min_width=600):
+                with gr.Group():
+                    gallery = gr.Gallery(
+                        label="Generated images",
+                        show_label=False,
+                        elem_id="gallery",
+                    ).style(grid=[2])
+                    std_output = gr.Textbox(
+                        value="Nothing to show.",
+                        lines=1,
+                        show_label=False,
+                    )
+                output_dir = args.output_dir if args.output_dir else Path.cwd()
+                output_dir = Path(output_dir, "generated_imgs")
+                output_loc = gr.Textbox(
+                    label="Saving Images at",
+                    value=output_dir,
+                    interactive=False,
+                )
+        kwargs = dict(
+            fn=inpaint_inf,
+            inputs=[
+                prompt,
+                negative_prompt,
+                init_image,
+                height,
+                width,
+                steps,
+                guidance_scale,
+                seed,
+                batch_count,
+                batch_size,
+                scheduler,
+                custom_model,
+                hf_model_id,
+                precision,
+                device,
+                max_length,
+                save_metadata_to_json,
+                save_metadata_to_png,
+            ],
+            outputs=[gallery, std_output],
+            show_progress=args.progress_bar,
+        )
+
+        prompt.submit(**kwargs)
+        negative_prompt.submit(**kwargs)
+        stable_diffusion.click(**kwargs)

apps/stable_diffusion/web/ui/outpaint_ui.py

@@ -0,0 +1,260 @@
+import os
+import sys
+import glob
+from pathlib import Path
+import gradio as gr
+from PIL import Image
+from apps.stable_diffusion.scripts import outpaint_inf
+from apps.stable_diffusion.src import args
+from apps.stable_diffusion.web.ui.utils import (
+    available_devices,
+    nodlogo_loc,
+)
+
+
+with gr.Blocks(title="Outpainting") as outpaint_web:
+    with gr.Row(elem_id="ui_title"):
+        nod_logo = Image.open(nodlogo_loc)
+        with gr.Row():
+            with gr.Column(scale=1, elem_id="demo_title_outer"):
+                gr.Image(
+                    value=nod_logo,
+                    show_label=False,
+                    interactive=False,
+                    elem_id="top_logo",
+                ).style(width=150, height=50)
+    with gr.Row(elem_id="ui_body"):
+        with gr.Row():
+            with gr.Column(scale=1, min_width=600):
+                with gr.Row():
+                    ckpt_path = (
+                        Path(args.ckpt_dir)
+                        if args.ckpt_dir
+                        else Path(Path.cwd(), "models")
+                    )
+                    ckpt_path.mkdir(parents=True, exist_ok=True)
+                    types = (
+                        "*.ckpt",
+                        "*.safetensors",
+                    )  # the tuple of file types
+                    ckpt_files = ["None"]
+                    for extn in types:
+                        files = glob.glob(os.path.join(ckpt_path, extn))
+                        ckpt_files.extend(files)
+                    custom_model = gr.Dropdown(
+                        label=f"Models (Custom Model path: {ckpt_path})",
+                        value=args.ckpt_loc if args.ckpt_loc else "None",
+                        choices=ckpt_files
+                        + [
+                            "runwayml/stable-diffusion-inpainting",
+                            "stabilityai/stable-diffusion-2-inpainting",
+                        ],
+                    )
+                    hf_model_id = gr.Textbox(
+                        placeholder="Select 'None' in the Models dropdown on the left and enter model ID here e.g: SG161222/Realistic_Vision_V1.3",
+                        value="",
+                        label="HuggingFace Model ID",
+                        lines=3,
+                    )
+
+                with gr.Group(elem_id="prompt_box_outer"):
+                    prompt = gr.Textbox(
+                        label="Prompt",
+                        value=args.prompts[0],
+                        lines=1,
+                        elem_id="prompt_box",
+                    )
+                    negative_prompt = gr.Textbox(
+                        label="Negative Prompt",
+                        value=args.negative_prompts[0],
+                        lines=1,
+                        elem_id="negative_prompt_box",
+                    )
+
+                init_image = gr.Image(label="Input Image", type="filepath")
+
+                with gr.Accordion(label="Advanced Options", open=False):
+                    with gr.Row():
+                        scheduler = gr.Dropdown(
+                            label="Scheduler",
+                            value="PNDM",
+                            choices=[
+                                "DDIM",
+                                "PNDM",
+                                "DPMSolverMultistep",
+                                "EulerAncestralDiscrete",
+                            ],
+                        )
+                        with gr.Group():
+                            save_metadata_to_png = gr.Checkbox(
+                                label="Save prompt information to PNG",
+                                value=args.write_metadata_to_png,
+                                interactive=True,
+                            )
+                            save_metadata_to_json = gr.Checkbox(
+                                label="Save prompt information to JSON file",
+                                value=args.save_metadata_to_json,
+                                interactive=True,
+                            )
+                    with gr.Row():
+                        pixels = gr.Slider(
+                            8,
+                            256,
+                            value=args.pixels,
+                            step=8,
+                            label="Pixels to expand",
+                        )
+                        mask_blur = gr.Slider(
+                            0,
+                            64,
+                            value=args.mask_blur,
+                            step=1,
+                            label="Mask blur",
+                        )
+                    with gr.Row():
+                        directions = gr.CheckboxGroup(
+                            label="Outpainting direction",
+                            choices=["left", "right", "up", "down"],
+                            value=["left", "right", "up", "down"],
+                        )
+                    with gr.Row():
+                        noise_q = gr.Slider(
+                            0.0,
+                            4.0,
+                            value=1.0,
+                            step=0.01,
+                            label="Fall-off exponent (lower=higher detail)",
+                        )
+                        color_variation = gr.Slider(
+                            0.0,
+                            1.0,
+                            value=0.05,
+                            step=0.01,
+                            label="Color variation",
+                        )
+                    with gr.Row():
+                        height = gr.Slider(
+                            384, 786, value=args.height, step=8, label="Height"
+                        )
+                        width = gr.Slider(
+                            384, 786, value=args.width, step=8, label="Width"
+                        )
+                        precision = gr.Radio(
+                            label="Precision",
+                            value=args.precision,
+                            choices=[
+                                "fp16",
+                                "fp32",
+                            ],
+                            visible=False,
+                        )
+                        max_length = gr.Radio(
+                            label="Max Length",
+                            value=args.max_length,
+                            choices=[
+                                64,
+                                77,
+                            ],
+                            visible=False,
+                        )
+                    with gr.Row():
+                        steps = gr.Slider(
+                            1, 100, value=20, step=1, label="Steps"
+                        )
+                    with gr.Row():
+                        guidance_scale = gr.Slider(
+                            0,
+                            50,
+                            value=args.guidance_scale,
+                            step=0.1,
+                            label="CFG Scale",
+                        )
+                        batch_count = gr.Slider(
+                            1,
+                            100,
+                            value=args.batch_count,
+                            step=1,
+                            label="Batch Count",
+                            interactive=True,
+                        )
+                        batch_size = gr.Slider(
+                            1,
+                            4,
+                            value=args.batch_size,
+                            step=1,
+                            label="Batch Size",
+                            interactive=False,
+                            visible=False,
+                        )
+                with gr.Row():
+                    seed = gr.Number(
+                        value=args.seed, precision=0, label="Seed"
+                    )
+                    device = gr.Dropdown(
+                        label="Device",
+                        value=available_devices[0],
+                        choices=available_devices,
+                    )
+                with gr.Row():
+                    random_seed = gr.Button("Randomize Seed")
+                    random_seed.click(
+                        None,
+                        inputs=[],
+                        outputs=[seed],
+                        _js="() => Math.floor(Math.random() * 4294967295)",
+                    )
+                    stable_diffusion = gr.Button("Generate Image(s)")
+
+            with gr.Column(scale=1, min_width=600):
+                with gr.Group():
+                    gallery = gr.Gallery(
+                        label="Generated images",
+                        show_label=False,
+                        elem_id="gallery",
+                    ).style(grid=[2])
+                    std_output = gr.Textbox(
+                        value="Nothing to show.",
+                        lines=1,
+                        show_label=False,
+                    )
+                output_dir = args.output_dir if args.output_dir else Path.cwd()
+                output_dir = Path(output_dir, "generated_imgs")
+                output_loc = gr.Textbox(
+                    label="Saving Images at",
+                    value=output_dir,
+                    interactive=False,
+                )
+        kwargs = dict(
+            fn=outpaint_inf,
+            inputs=[
+                prompt,
+                negative_prompt,
+                init_image,
+                pixels,
+                mask_blur,
+                directions,
+                noise_q,
+                color_variation,
+                height,
+                width,
+                steps,
+                guidance_scale,
+                seed,
+                batch_count,
+                batch_size,
+                scheduler,
+                custom_model,
+                hf_model_id,
+                precision,
+                device,
+                max_length,
+                save_metadata_to_json,
+                save_metadata_to_png,
+            ],
+            outputs=[gallery, std_output],
+            show_progress=args.progress_bar,
+        )
+
+        prompt.submit(**kwargs)
+        negative_prompt.submit(**kwargs)
+        stable_diffusion.click(**kwargs)

build_tools/stable_diffusion_testing.py

@@ -20,6 +20,33 @@ model_config_dicts = get_json_file(
 )
 
 
+def parse_sd_out(filename, command, device, use_tune, model_name, import_mlir):
+    with open(filename, "r+") as f:
+        lines = f.readlines()
+    metrics = {}
+    vals_to_read = [
+        "Clip Inference time",
+        "Average step",
+        "VAE Inference time",
+        "Total image generation",
+    ]
+    for line in lines:
+        for val in vals_to_read:
+            if val in line:
+                metrics[val] = line.split(" ")[-1].strip("\n")
+
+    metrics["Average step"] = metrics["Average step"].strip("ms/it")
+    metrics["Total image generation"] = metrics[
+        "Total image generation"
+    ].strip("sec")
+    metrics["device"] = device
+    metrics["use_tune"] = use_tune
+    metrics["model_name"] = model_name
+    metrics["import_mlir"] = import_mlir
+    metrics["command"] = command
+    return metrics
+
+
 def get_inpaint_inputs():
     os.mkdir("./test_images/inputs")
     img_url = (
@@ -39,6 +66,7 @@ def get_inpaint_inputs():
 def test_loop(device="vulkan", beta=False, extra_flags=[]):
     # Get golden values from tank
     shutil.rmtree("./test_images", ignore_errors=True)
+    model_metrics = []
     os.mkdir("./test_images")
     os.mkdir("./test_images/golden")
     get_inpaint_inputs()
@@ -52,9 +80,16 @@ def test_loop(device="vulkan", beta=False, extra_flags=[]):
         inpaint_prompt_text = '--prompt="Face of a yellow cat, high resolution, sitting on a park bench"'
     if beta:
         extra_flags.append("--beta_models=True")
+    extra_flags.append("--no-progress_bar")
+    to_skip = [
+        "Linaqruf/anything-v3.0",
+        "prompthero/openjourney",
+        "wavymulder/Analog-Diffusion",
+        "dreamlike-art/dreamlike-diffusion-1.0",
+    ]
     for import_opt in import_options:
         for model_name in hf_model_names:
-            if model_name == "Linaqruf/anything-v3.0":
+            if model_name in to_skip:
                 continue
             for use_tune in tuned_options:
                 command = (
@@ -73,7 +108,7 @@ def test_loop(device="vulkan", beta=False, extra_flags=[]):
                     ]
                     if "inpainting" not in model_name
                     else [
-                        "python",
+                        executable,
                         "apps/stable_diffusion/scripts/inpaint.py",
                         "--device=" + device,
                         inpaint_prompt_text,
@@ -91,12 +126,27 @@ def test_loop(device="vulkan", beta=False, extra_flags=[]):
                 command += extra_flags
                 if os.name == "nt":
                     command = " ".join(command)
-                generated_image = not subprocess.call(
-                    command, stdout=subprocess.DEVNULL
-                )
+                dumpfile_name = "_".join(model_name.split("/")) + ".txt"
+                dumpfile_name = os.path.join(os.getcwd(), dumpfile_name)
+                with open(dumpfile_name, "w+") as f:
+                    generated_image = not subprocess.call(
+                        command,
+                        stdout=f,
+                        stderr=f,
+                    )
                 if os.name != "nt":
                     command = " ".join(command)
                 if generated_image:
+                    model_metrics.append(
+                        parse_sd_out(
+                            dumpfile_name,
+                            command,
+                            device,
+                            use_tune,
+                            model_name,
+                            import_opt,
+                        )
+                    )
                     print(command)
                     print("Successfully generated image")
                     os.makedirs(
@@ -127,6 +177,22 @@ def test_loop(device="vulkan", beta=False, extra_flags=[]):
                     if "2_1_base" in model_name:
                         print("failed a known successful model.")
                         exit(1)
+    with open(os.path.join(os.getcwd(), "sd_testing_metrics.csv"), "w+") as f:
+        header = "model_name;device;use_tune;import_opt;Clip Inference time(ms);Average Step (ms/it);VAE Inference time(ms);total image generation(s);command\n"
+        f.write(header)
+        for metric in model_metrics:
+            output = [
+                metric["model_name"],
+                metric["device"],
+                metric["use_tune"],
+                metric["import_mlir"],
+                metric["Clip Inference time"],
+                metric["Average step"],
+                metric["VAE Inference time"],
+                metric["Total image generation"],
+                metric["command"],
+            ]
+            f.write(";".join(output) + "\n")
 
 
 parser = argparse.ArgumentParser()

docs/shark_iree_profiling.md

@@ -0,0 +1,118 @@
+# Overview
+
+This document is intended to provide a starting point for profiling with SHARK/IREE. At it's core
+[SHARK](https://github.com/nod-ai/SHARK/tree/main/tank) is a python API that links the MLIR lowerings from various
+frameworks + frontends (e.g. PyTorch -> Torch-MLIR) with the compiler + runtime offered by IREE. More information
+on model coverage and framework support can be found [here](https://github.com/nod-ai/SHARK/tree/main/tank). The intended
+use case for SHARK is for compilation and deployment of performant state of the art AI models.
+
+![image](https://user-images.githubusercontent.com/22101546/217151219-9bb184a3-cfb9-4788-bb7e-5b502953525c.png)
+
+## Benchmarking with SHARK
+
+TODO: Expand this section.
+
+SHARK offers native benchmarking support, although because it is model focused, fine grain profiling is
+hidden when compared against the common "model benchmarking suite" use case SHARK is good at.
+
+### SharkBenchmarkRunner
+
+SharkBenchmarkRunner is a class designed for benchmarking models against other runtimes.
+TODO: List supported runtimes for comparison + example on how to benchmark with it.
+
+## Directly profiling IREE
+
+A number of excellent developer resources on profiling with IREE can be
+found [here](https://github.com/iree-org/iree/tree/main/docs/developers/developing_iree). As a result this section will
+focus on the bridging the gap between the two.
+ - https://github.com/iree-org/iree/blob/main/docs/developers/developing_iree/profiling.md
+ - https://github.com/iree-org/iree/blob/main/docs/developers/developing_iree/profiling_with_tracy.md
+ - https://github.com/iree-org/iree/blob/main/docs/developers/developing_iree/profiling_vulkan_gpu.md
+ - https://github.com/iree-org/iree/blob/main/docs/developers/developing_iree/profiling_cpu_events.md
+
+Internally, SHARK builds a pair of IREE commands to compile + run a model. At a high level the flow starts with the
+model represented with a high level dialect (commonly Linalg) and is compiled to a flatbuffer (.vmfb) that
+the runtime is capable of ingesting. At this point (with potentially a few runtime flags) the compiled model is then run
+through the IREE runtime. This is all facilitated with the IREE python bindings, which offers a convenient method
+to capture the compile command SHARK comes up with. This is done by setting the environment variable
+`IREE_SAVE_TEMPS` to point to a directory of choice, e.g. for stable diffusion
+```
+# Linux
+$ export IREE_SAVE_TEMPS=/path/to/some/directory
+# Windows
+$ $env:IREE_SAVE_TEMPS="C:\path\to\some\directory"
+$ python apps/stable_diffusion/scripts/txt2img.py -p "a photograph of an astronaut riding a horse" --save_vmfb
+```
+NOTE: Currently this will only save the compile command + input MLIR for a single model if run in a pipeline.
+In the case of stable diffusion this (should) be UNet so to get examples for other models in the pipeline they
+need to be extracted and tested individually.
+
+The save temps directory should contain three files: `core-command-line.txt`, `core-input.mlir`, and `core-output.bin`.
+The command line for compilation will start something like this, where the `-` needs to be replaced with the path to `core-input.mlir`.
+```
+/home/quinn/nod/iree-build/compiler/bindings/python/iree/compiler/tools/../_mlir_libs/iree-compile - --iree-input-type=none ...
+```
+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-benchmark-module --module=benchmarks/module_forward_dispatch_${NUM}_vulkan_spirv_fb.vmfb --function=forward --device=vulkan
+```
+Where `${NUM}` is the dispatch number that you want to benchmark/profile in isolation.
+
+### Enabling Tracy for Vulkan profiling
+
+To begin profiling with Tracy, a build of IREE runtime with tracing enabled is needed. SHARK-Runtime builds an
+instrumented version alongside the normal version nightly (.whls typically found [here](https://github.com/nod-ai/SHARK-Runtime/releases)), however this is only available for Linux. For Windows, tracing can be enabled by enabling a CMake flag.
+```
+$env:IREE_ENABLE_RUNTIME_TRACING="ON"
+```
+Getting a trace can then be done by setting environment variable `TRACY_NO_EXIT=1` and running the program that is to be
+traced. Then, to actually capture the trace, use the `iree-tracy-capture` tool in a different terminal. Note that to get
+the capture and profiler tools the `IREE_BUILD_TRACY=ON` CMake flag needs to be set.
+```
+TRACY_NO_EXIT=1 python apps/stable_diffusion/scripts/txt2img.py -p "a photograph of an astronaut riding a horse"
+
+# (in another terminal, either on the same machine or through ssh with a tunnel through port 8086)
+iree-tracy-capture -o trace_filename.tracy
+```
+To do it over ssh, the flow looks like this
+```
+# From terminal 1 on local machine
+ssh -L 8086:localhost:8086 <remote_server_name>
+TRACY_NO_EXIT=1 python apps/stable_diffusion/scripts/txt2img.py -p "a photograph of an astronaut riding a horse"
+
+# From terminal 2 on local machine. Requires having built IREE with the CMake flag `IREE_BUILD_TRACY=ON` to build the required tooling.
+iree-tracy-capture -o /path/to/trace.tracy
+```
+
+The trace can then be viewed with
+```
+iree-tracy-profiler /path/to/trace.tracy
+```
+Capturing a runtime trace will work with any IREE tooling that uses the runtime. For example, `iree-benchmark-module`
+can be used for benchmarking an individual module. Importantly this means that any SHARK script can be profiled with tracy.
+
+NOTE: Not all backends have the same tracy support. This writeup is focused on CPU/Vulkan backends but there is recently added support for tracing on CUDA (requires the `--cuda_tracing` flag).
+
+## Experimental RGP support
+
+TODO: This section is temporary until proper RGP support is added.
+
+Currently, for stable diffusion there is a flag for enabling UNet to be visible to RGP with `--enable_rgp`. To get a proper capture though, the `DevModeSqttPrepareFrameCount=1` flag needs to be set for the driver (done with `VkPanel` on Windows).
+With these two settings, a single iteration of UNet can be captured.
+
+(AMD only) To get a dump of the pipelines (result of compiled SPIR-V) the `EnablePipelineDump=1` driver flag can be set. The
+files will typically be dumped to a directory called `spvPipeline` (on Linux `/var/tmp/spvPipeline`. The dumped files will
+include header information that can be used to map back to the source dispatch/SPIR-V, e.g.
+```
+[Version]
+version = 57 
+
+[CsSpvFile]
+fileName = Shader_0x946C08DFD0C10D9A.spv
+
+[CsInfo]
+entryPoint = forward_dispatch_193_matmul_256x65536x2304
+```

process_skipfiles.py

@@ -6,6 +6,16 @@ from distutils.sysconfig import get_python_lib
 import fileinput
 from pathlib import Path
 
+# Diffusers 0.13.1 fails with transformers __init.py errros in BLIP. So remove it for now until we fork it
+pix2pix_file = Path(
+    get_python_lib()
+    + "/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion_pix2pix_zero.py"
+)
+if pix2pix_file.exists():
+    print("Removing..%s", pix2pix_file)
+    pix2pix_file.unlink()
+
+
 path_to_skipfiles = Path(get_python_lib() + "/torch/_dynamo/skipfiles.py")
 
 modules_to_comment = ["abc,", "os,", "posixpath,", "_collections_abc,"]

requirements-importer.txt

@@ -15,7 +15,7 @@ iree-tools-tf
 
 # TensorFlow and JAX.
 gin-config
-tensorflow>=2.10.1
+tf-nightly
 keras>=2.10
 #tf-models-nightly
 #tensorflow-text-nightly

requirements.txt

@@ -16,7 +16,7 @@ parameterized
 
 # Add transformers, diffusers and scipy since it most commonly used
 transformers
-diffusers
+diffusers @ git+https://github.com/nod-ai/diffusers@e5810e686ea4ac499e325c2961808c8972dee039
 scipy
 ftfy
 gradio

setup_venv.ps1

@@ -61,6 +61,7 @@ $version = if($p -is [System.Management.Automation.ErrorRecord])
 else
 {
     # otherwise return complete Python list
+    $ErrorActionPreference = 'SilentlyContinue'
     $PyVer = py --list
 }
 
@@ -73,7 +74,8 @@ if ($PyVer -like "*venv*")
 
 Write-Host "Python versions found are"
 Write-Host ($PyVer | Out-String) # formatted output with line breaks
-if (!($PyVer -like "*3.11*")) # if 3.11 is not in list
+if (!($PyVer.length -ne 0)) {$p} # return Python --version String if py.exe is unavailable
+if (!($PyVer -like "*3.11*") -and !($p -like "*3.11*")) # if 3.11 is not in any list
 {
     Write-Host "Please install Python 3.11 and try again"
     break
@@ -81,12 +83,13 @@ if (!($PyVer -like "*3.11*")) # if 3.11 is not in list
 
 Write-Host "Installing Build Dependencies"
 # make sure we really use 3.11 from list, even if it's not the default.
-py -3.11 -m venv .\shark.venv\
+if (!($PyVer.length -ne 0)) {py -3.11 -m venv .\shark.venv\}
+else {python -m venv .\shark.venv\}
 .\shark.venv\Scripts\activate
 python -m pip install --upgrade pip
 pip install wheel
 pip install -r requirements.txt
-pip install --pre torch-mlir torch torchvision --extra-index-url https://download.pytorch.org/whl/nightly/cpu -f https://llvm.github.io/torch-mlir/package-index/
+pip install --pre torch-mlir torch --extra-index-url https://download.pytorch.org/whl/nightly/cpu -f https://llvm.github.io/torch-mlir/package-index/
 pip install --upgrade -f https://nod-ai.github.io/SHARK-Runtime/pip-release-links.html iree-compiler iree-runtime
 Write-Host "Building SHARK..."
 pip install -e . -f https://llvm.github.io/torch-mlir/package-index/ -f https://nod-ai.github.io/SHARK-Runtime/pip-release-links.html

shark/examples/shark_inference/sharded_bloom.py

@@ -9,9 +9,11 @@
 # -d, --download: set to true if you want to redownload the mlir files
 # -t --token_count: the number of tokens you want to generate
 # -pr --prompt: the prompt you want to feed to the model
+# -m --model_namme: the name of the model, e.g. bloom-560m
 #####################################################################################
 
 import os
+import io
 import torch
 import torch.nn as nn
 from collections import OrderedDict
@@ -22,14 +24,18 @@ from transformers.models.bloom.configuration_bloom import BloomConfig
 import json
 import sys
 import argparse
-from cuda.cudart import cudaSetDevice
 import json
+import urllib.request
 
 from torch.fx.experimental.proxy_tensor import make_fx
 from torch._decomp import get_decompositions
 from shark.shark_inference import SharkInference
 from shark.shark_downloader import download_public_file
-
+from transformers import (
+    BloomTokenizerFast,
+    BloomForSequenceClassification,
+    BloomForCausalLM,
+)
 from transformers.models.bloom.modeling_bloom import (
     BloomBlock,
     build_alibi_tensor,
@@ -47,16 +53,22 @@ class ShardedBloom:
         self.layers_initialized = False
 
         self.src_folder = src_folder
-        self.n_embed = config["n_embed"]
+        try:
+            self.n_embed = config["n_embed"]
+        except KeyError:
+            self.n_embed = config["hidden_size"]
         self.vocab_size = config["vocab_size"]
         self.n_layer = config["n_layer"]
-        self.n_head = config["num_attention_heads"]
+        try:
+            self.n_head = config["num_attention_heads"]
+        except KeyError:
+            self.n_head = config["n_head"]
 
     def _init_layer(self, layer_name, device, replace, device_idx):
         if replace or not os.path.exists(
             f"{self.src_folder}/{layer_name}.vmfb"
         ):
-            f_ = open(f"{self.src_folder}/{layer_name}.mlir")
+            f_ = open(f"{self.src_folder}/{layer_name}.mlir", encoding="utf-8")
             module = f_.read()
             f_.close()
             module = bytes(module, "utf-8")
@@ -292,90 +304,352 @@ def _prepare_attn_mask(
     return combined_attention_mask
 
 
-def download_560m(destination_folder):
+def download_model(destination_folder, model_name):
     download_public_file(
-        "https://bloom-560m/bloom_block_0.mlir", destination_folder
+        f"https://{model_name}/config.json", destination_folder
+    )
+    f = open(f"{destination_folder}/config.json")
+    config = json.load(f)
+    f.close()
+    n_blocks = config["n_layer"]
+    download_public_file(
+        f"https://{model_name}/lm_head.mlir", destination_folder
+    )
+    download_public_file(f"https://{model_name}/ln_f.mlir", destination_folder)
+    download_public_file(
+        f"https://{model_name}/word_embeddings.mlir", destination_folder
     )
     download_public_file(
-        "https://bloom-560m/bloom_block_1.mlir", destination_folder
+        f"https://{model_name}/word_embeddings_layernorm.mlir",
+        destination_folder,
     )
     download_public_file(
-        "https://bloom-560m/bloom_block_2.mlir", destination_folder
+        f"https://{model_name}/tokenizer.json", destination_folder
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_3.mlir", destination_folder
+
+    for i in range(n_blocks):
+        download_public_file(
+            f"https://{model_name}/bloom_block_{i}.mlir", destination_folder
+        )
+
+
+def compile_embeddings(embeddings_layer, input_ids, path):
+    input_ids_placeholder = torch_mlir.TensorPlaceholder.like(
+        input_ids, dynamic_axes=[1]
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_4.mlir", destination_folder
+    module = torch_mlir.compile(
+        embeddings_layer,
+        (input_ids_placeholder),
+        torch_mlir.OutputType.LINALG_ON_TENSORS,
+        use_tracing=False,
+        verbose=False,
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_5.mlir", destination_folder
+
+    bytecode_stream = io.BytesIO()
+    module.operation.write_bytecode(bytecode_stream)
+    bytecode = bytecode_stream.getvalue()
+
+    f_ = open(path, "w+")
+    f_.write(str(module))
+    f_.close()
+    return
+
+
+def compile_word_embeddings_layernorm(
+    embeddings_layer_layernorm, embeds, path
+):
+    embeds_placeholder = torch_mlir.TensorPlaceholder.like(
+        embeds, dynamic_axes=[1]
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_6.mlir", destination_folder
+    module = torch_mlir.compile(
+        embeddings_layer_layernorm,
+        (embeds_placeholder),
+        torch_mlir.OutputType.LINALG_ON_TENSORS,
+        use_tracing=False,
+        verbose=False,
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_7.mlir", destination_folder
+
+    bytecode_stream = io.BytesIO()
+    module.operation.write_bytecode(bytecode_stream)
+    bytecode = bytecode_stream.getvalue()
+
+    f_ = open(path, "w+")
+    f_.write(str(module))
+    f_.close()
+    return
+
+
+def strip_overloads(gm):
+    """
+    Modifies the target of graph nodes in :attr:`gm` to strip overloads.
+    Args:
+        gm(fx.GraphModule): The input Fx graph module to be modified
+    """
+    for node in gm.graph.nodes:
+        if isinstance(node.target, torch._ops.OpOverload):
+            node.target = node.target.overloadpacket
+    gm.recompile()
+
+
+def compile_to_mlir(
+    bblock,
+    hidden_states,
+    layer_past=None,
+    attention_mask=None,
+    head_mask=None,
+    use_cache=None,
+    output_attentions=False,
+    alibi=None,
+    block_index=0,
+    path=".",
+):
+    fx_g = make_fx(
+        bblock,
+        decomposition_table=get_decompositions(
+            [
+                torch.ops.aten.split.Tensor,
+                torch.ops.aten.split_with_sizes,
+            ]
+        ),
+        tracing_mode="real",
+        _allow_non_fake_inputs=False,
+    )(hidden_states, alibi, attention_mask)
+
+    fx_g.graph.set_codegen(torch.fx.graph.CodeGen())
+    fx_g.recompile()
+
+    strip_overloads(fx_g)
+
+    hidden_states_placeholder = TensorPlaceholder.like(
+        hidden_states, dynamic_axes=[1]
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_8.mlir", destination_folder
+    attention_mask_placeholder = TensorPlaceholder.like(
+        attention_mask, dynamic_axes=[2, 3]
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_9.mlir", destination_folder
+    alibi_placeholder = TensorPlaceholder.like(alibi, dynamic_axes=[2])
+
+    ts_g = torch.jit.script(fx_g)
+
+    module = torch_mlir.compile(
+        ts_g,
+        (
+            hidden_states_placeholder,
+            alibi_placeholder,
+            attention_mask_placeholder,
+        ),
+        torch_mlir.OutputType.LINALG_ON_TENSORS,
+        use_tracing=False,
+        verbose=False,
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_10.mlir", destination_folder
+
+    module_placeholder = module
+    module_context = module_placeholder.context
+
+    def check_valid_line(line, line_n, mlir_file_len):
+        if "private" in line:
+            return False
+        if "attributes" in line:
+            return False
+        if mlir_file_len - line_n == 2:
+            return False
+
+        return True
+
+    mlir_file_len = len(str(module).split("\n"))
+
+    def remove_constant_dim(line):
+        if "17x" in line:
+            line = re.sub("17x", "?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 eq" in line:
+            line = re.sub("c17", "dim", line)
+        if " 17," in line:
+            line = re.sub(" 17,", " %dim,", line)
+        return line
+
+    module = "\n".join(
+        [
+            remove_constant_dim(line)
+            for line, line_n in zip(
+                str(module).split("\n"), range(mlir_file_len)
+            )
+            if check_valid_line(line, line_n, mlir_file_len)
+        ]
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_11.mlir", destination_folder
+
+    module = module_placeholder.parse(module, context=module_context)
+    bytecode_stream = io.BytesIO()
+    module.operation.write_bytecode(bytecode_stream)
+    bytecode = bytecode_stream.getvalue()
+
+    f_ = open(path, "w+")
+    f_.write(str(module))
+    f_.close()
+    return
+
+
+def compile_ln_f(ln_f, hidden_layers, path):
+    hidden_layers_placeholder = torch_mlir.TensorPlaceholder.like(
+        hidden_layers, dynamic_axes=[1]
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_12.mlir", destination_folder
+    module = torch_mlir.compile(
+        ln_f,
+        (hidden_layers_placeholder),
+        torch_mlir.OutputType.LINALG_ON_TENSORS,
+        use_tracing=False,
+        verbose=False,
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_13.mlir", destination_folder
+
+    bytecode_stream = io.BytesIO()
+    module.operation.write_bytecode(bytecode_stream)
+    bytecode = bytecode_stream.getvalue()
+
+    f_ = open(path, "w+")
+    f_.write(str(module))
+    f_.close()
+    return
+
+
+def compile_lm_head(lm_head, hidden_layers, path):
+    hidden_layers_placeholder = torch_mlir.TensorPlaceholder.like(
+        hidden_layers, dynamic_axes=[1]
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_14.mlir", destination_folder
+    module = torch_mlir.compile(
+        lm_head,
+        (hidden_layers_placeholder),
+        torch_mlir.OutputType.LINALG_ON_TENSORS,
+        use_tracing=False,
+        verbose=False,
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_15.mlir", destination_folder
+
+    bytecode_stream = io.BytesIO()
+    module.operation.write_bytecode(bytecode_stream)
+    bytecode = bytecode_stream.getvalue()
+
+    f_ = open(path, "w+")
+    f_.write(str(module))
+    f_.close()
+    return
+
+
+def create_mlirs(destination_folder, model_name):
+    model_config = "bigscience/" + model_name
+    sample_input_ids = torch.ones([1, 17], dtype=torch.int64)
+
+    urllib.request.urlretrieve(
+        f"https://huggingface.co/bigscience/{model_name}/resolve/main/config.json",
+        filename=f"{destination_folder}/config.json",
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_16.mlir", destination_folder
+    urllib.request.urlretrieve(
+        f"https://huggingface.co/bigscience/bloom/resolve/main/tokenizer.json",
+        filename=f"{destination_folder}/tokenizer.json",
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_17.mlir", destination_folder
+
+    class HuggingFaceLanguage(torch.nn.Module):
+        def __init__(self):
+            super().__init__()
+            self.model = BloomForCausalLM.from_pretrained(model_config)
+
+        def forward(self, tokens):
+            return self.model.forward(tokens)[0]
+
+    class HuggingFaceBlock(torch.nn.Module):
+        def __init__(self, block):
+            super().__init__()
+            self.model = block
+
+        def forward(self, tokens, alibi, attention_mask):
+            output = self.model(
+                hidden_states=tokens,
+                alibi=alibi,
+                attention_mask=attention_mask,
+                use_cache=True,
+                output_attentions=False,
+            )
+            return (output[0], output[1][0], output[1][1])
+
+    model = HuggingFaceLanguage()
+
+    compile_embeddings(
+        model.model.transformer.word_embeddings,
+        sample_input_ids,
+        f"{destination_folder}/word_embeddings.mlir",
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_18.mlir", destination_folder
+
+    inputs_embeds = model.model.transformer.word_embeddings(sample_input_ids)
+
+    compile_word_embeddings_layernorm(
+        model.model.transformer.word_embeddings_layernorm,
+        inputs_embeds,
+        f"{destination_folder}/word_embeddings_layernorm.mlir",
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_19.mlir", destination_folder
+
+    hidden_states = model.model.transformer.word_embeddings_layernorm(
+        inputs_embeds
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_20.mlir", destination_folder
+
+    input_shape = sample_input_ids.size()
+
+    current_sequence_length = hidden_states.shape[1]
+    past_key_values_length = 0
+    past_key_values = tuple([None] * len(model.model.transformer.h))
+
+    attention_mask = torch.ones(
+        (hidden_states.shape[0], current_sequence_length), device="cpu"
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_21.mlir", destination_folder
+
+    alibi = build_alibi_tensor(
+        attention_mask,
+        model.model.transformer.n_head,
+        hidden_states.dtype,
+        "cpu",
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_22.mlir", destination_folder
+
+    causal_mask = _prepare_attn_mask(
+        attention_mask, input_shape, inputs_embeds, past_key_values_length
     )
-    download_public_file(
-        "https://bloom-560m/bloom_block_23.mlir", destination_folder
+
+    head_mask = model.model.transformer.get_head_mask(
+        None, model.model.transformer.config.n_layer
     )
-    download_public_file("https://bloom-560m/config.json", destination_folder)
-    download_public_file("https://bloom-560m/lm_head.mlir", destination_folder)
-    download_public_file("https://bloom-560m/ln_f.mlir", destination_folder)
-    download_public_file(
-        "https://bloom-560m/word_embeddings.mlir", destination_folder
+    output_attentions = model.model.transformer.config.output_attentions
+
+    all_hidden_states = ()
+
+    for i, (block, layer_past) in enumerate(
+        zip(model.model.transformer.h, past_key_values)
+    ):
+        all_hidden_states = all_hidden_states + (hidden_states,)
+
+        proxy_model = HuggingFaceBlock(block)
+
+        compile_to_mlir(
+            proxy_model,
+            hidden_states,
+            layer_past=layer_past,
+            attention_mask=causal_mask,
+            head_mask=head_mask[i],
+            use_cache=True,
+            output_attentions=output_attentions,
+            alibi=alibi,
+            block_index=i,
+            path=f"{destination_folder}/bloom_block_{i}.mlir",
+        )
+
+    compile_ln_f(
+        model.model.transformer.ln_f,
+        hidden_states,
+        f"{destination_folder}/ln_f.mlir",
     )
-    download_public_file(
-        "https://bloom-560m/word_embeddings_layernorm.mlir", destination_folder
-    )
-    download_public_file(
-        "https://bloom-560m/tokenizer.json", destination_folder
+    hidden_states = model.model.transformer.ln_f(hidden_states)
+    compile_lm_head(
+        model.model.lm_head,
+        hidden_states,
+        f"{destination_folder}/lm_head.mlir",
     )
 
 
@@ -387,6 +661,7 @@ if __name__ == "__main__":
     parser.add_argument("-c", "--recompile", default=False, type=bool)
     parser.add_argument("-d", "--download", default=False, type=bool)
     parser.add_argument("-t", "--token_count", default=10, type=int)
+    parser.add_argument("-m", "--model_name", default="bloom-560m")
     parser.add_argument(
         "-pr",
         "--prompt",
@@ -399,8 +674,10 @@ if __name__ == "__main__":
 
     if args.device == "cuda" and args.device_list is not None:
         IS_CUDA = True
+        from cuda.cudart import cudaSetDevice
     if args.download:
-        download_560m(args.model_path)
+        # download_model(args.model_path, args.model_name)
+        create_mlirs(args.model_path, args.model_name)
     from transformers import AutoTokenizer, AutoModelForCausalLM, BloomConfig
 
     tokenizer = AutoTokenizer.from_pretrained(args.model_path)