Merge pull request #43 from huggingface/diffusers-auto

Add Stable Diffusion XL and Diffusers autopipeline

Author: philschmid

README.md

@@ -40,6 +40,8 @@ docker build -t starlette-transformers:gpu -f dockerfiles/tensorflow/gpu/Dockerf
 ```bash
 docker run -ti -p 5000:5000 -e HF_MODEL_ID=distilbert-base-uncased-distilled-squad -e HF_TASK=question-answering starlette-transformers:cpu
 docker run -ti -p 5000:5000 --gpus all -e HF_MODEL_ID=nlpconnect/vit-gpt2-image-captioning -e HF_TASK=image-to-text starlette-transformers:gpu
+docker run -ti -p 5000:5000 --gpus all -e HF_MODEL_ID=echarlaix/tiny-random-stable-diffusion-xl -e HF_TASK=text-to-image starlette-transformers:gpu
+docker run -ti -p 5000:5000 --gpus all -e HF_MODEL_ID=stabilityai/stable-diffusion-xl-base-1.0 -e HF_TASK=text-to-image starlette-transformers:gpu
 docker run -ti -p 5000:5000 -e HF_MODEL_DIR=/repository -v $(pwd)/distilbert-base-uncased-emotion:/repository starlette-transformers:cpu
 ```
 

dockerfiles/pytorch/cpu/Dockerfile

@@ -23,7 +23,7 @@ RUN apt-get update \
 # install micromamba
 ENV MAMBA_ROOT_PREFIX=/opt/conda
 ENV PATH=/opt/conda/bin:$PATH
-RUN curl -L https://micromamba.snakepit.net/api/micromamba/linux-64/latest | tar -xj "bin/micromamba" \
+RUN curl -L https://micro.mamba.pm/api/micromamba/linux-64/latest | tar -xj "bin/micromamba" \
     && touch /root/.bashrc \
     && ./bin/micromamba shell init -s bash -p /opt/conda  \
     && grep -v '[ -z "\$PS1" ] && return' /root/.bashrc  > /opt/conda/bashrc

dockerfiles/pytorch/cpu/environment.yaml

@@ -5,8 +5,9 @@ dependencies:
 - python=3.9.13
 - pytorch::pytorch=1.13.1=py3.9_cpu_0
 - pip:
-  - transformers[sklearn,sentencepiece,audio,vision]==4.27.2
+  - transformers[sklearn,sentencepiece,audio,vision]==4.31.0
   - sentence_transformers==2.2.2
   - torchvision==0.14.1
-  - diffusers==0.14.0
-  - accelerate==0.17.1
+  - diffusers==0.19.3
+  - accelerate==0.21.0
+  - safetensors

dockerfiles/pytorch/gpu/Dockerfile

@@ -25,7 +25,7 @@ ENV MAMBA_ROOT_PREFIX=/opt/conda
 ENV PATH=/opt/conda/bin:$PATH
 ENV LD_LIBRARY_PATH="/opt/conda/lib:${LD_LIBRARY_PATH}"
 
-RUN curl -L https://micromamba.snakepit.net/api/micromamba/linux-64/latest | tar -xj "bin/micromamba" \
+RUN curl -L https://micro.mamba.pm/api/micromamba/linux-64/latest | tar -xj "bin/micromamba" \
     && touch /root/.bashrc \
     && ./bin/micromamba shell init -s bash -p /opt/conda  \
     && grep -v '[ -z "\$PS1" ] && return' /root/.bashrc  > /opt/conda/bashrc

dockerfiles/pytorch/gpu/environment.yaml

@@ -9,5 +9,6 @@ dependencies:
   - transformers[sklearn,sentencepiece,audio,vision]==4.31.0
   - sentence_transformers==2.2.2
   - torchvision==0.14.1
-  - diffusers==0.18.2
-  - accelerate==0.21.0
+  - diffusers==0.19.3
+  - accelerate==0.21.0
+  - safetensors

dockerfiles/tensorflow/cpu/Dockerfile

@@ -23,7 +23,7 @@ RUN apt-get update \
 # install micromamba
 ENV MAMBA_ROOT_PREFIX=/opt/conda
 ENV PATH=/opt/conda/bin:$PATH
-RUN curl -L https://micromamba.snakepit.net/api/micromamba/linux-64/latest | tar -xj "bin/micromamba" \
+RUN curl -L https://micro.mamba.pm/api/micromamba/linux-64/latest | tar -xj "bin/micromamba" \
     && touch /root/.bashrc \
     && ./bin/micromamba shell init -s bash -p /opt/conda  \
     && grep -v '[ -z "\$PS1" ] && return' /root/.bashrc  > /opt/conda/bashrc

dockerfiles/tensorflow/gpu/Dockerfile

@@ -26,7 +26,7 @@ ENV MAMBA_ROOT_PREFIX=/opt/conda
 ENV PATH=/opt/conda/bin:$PATH
 ENV LD_LIBRARY_PATH="/opt/conda/lib:${LD_LIBRARY_PATH}"
 
-RUN curl -L https://micromamba.snakepit.net/api/micromamba/linux-64/latest | tar -xj "bin/micromamba" \
+RUN curl -L https://micro.mamba.pm/api/micromamba/linux-64/latest | tar -xj "bin/micromamba" \
     && touch /root/.bashrc \
     && ./bin/micromamba shell init -s bash -p /opt/conda  \
     && grep -v '[ -z "\$PS1" ] && return' /root/.bashrc  > /opt/conda/bashrc

src/huggingface_inference_toolkit/diffusers_utils.py

@@ -1,6 +1,8 @@
 import importlib.util
-import json
-import os
+import logging
+
+logger = logging.getLogger(__name__)
+logging.basicConfig(format="%(asctime)s | %(levelname)s | %(message)s", level=logging.INFO)
 
 _diffusers = importlib.util.find_spec("diffusers") is not None
 
@@ -11,60 +13,46 @@ def is_diffusers_available():
 
 if is_diffusers_available():
     import torch
-    from diffusers import DPMSolverMultistepScheduler, StableDiffusionPipeline
-
-
-def check_supported_pipeline(model_dir):
-    try:
-        with open(os.path.join(model_dir, "model_index.json")) as json_file:
-            data = json.load(json_file)
-            if data["_class_name"] == "StableDiffusionPipeline":
-                return True
-            else:
-                return False
-    except Exception:
-        return False
+    from diffusers import AutoPipelineForText2Image, DPMSolverMultistepScheduler, StableDiffusionPipeline
 
 
-class DiffusersPipelineImageToText:
+class IEAutoPipelineForText2Image:
     def __init__(self, model_dir: str, device: str = None):  # needs "cuda" for GPU
-        self.pipeline = StableDiffusionPipeline.from_pretrained(model_dir, torch_dtype=torch.float16)
+        dtype = torch.float32
+        if device == "cuda":
+            dtype = torch.float16
+        device_map = "auto" if device == "cuda" else None
+
+        self.pipeline = AutoPipelineForText2Image.from_pretrained(model_dir, torch_dtype=dtype, device_map=device_map)
         # try to use DPMSolverMultistepScheduler
-        try:
-            self.pipeline.scheduler = DPMSolverMultistepScheduler.from_config(self.pipeline.scheduler.config)
-        except Exception:
-            pass
+        if isinstance(self.pipeline, StableDiffusionPipeline):
+            try:
+                self.pipeline.scheduler = DPMSolverMultistepScheduler.from_config(self.pipeline.scheduler.config)
+            except Exception:
+                pass
         self.pipeline.to(device)
 
     def __call__(
         self,
         prompt,
-        num_inference_steps=25,
-        guidance_scale=7.5,
-        num_images_per_prompt=1,
-        height=None,
-        width=None,
-        negative_prompt=None,
+        **kwargs,
     ):
         # TODO: add support for more images (Reason is correct output)
-        num_images_per_prompt = 1
+        if "num_images_per_prompt" in kwargs:
+            kwargs.pop("num_images_per_prompt")
+            logger.warning("Sending num_images_per_prompt > 1 to pipeline is not supported. Using default value 1.")
 
         # Call pipeline with parameters
-        out = self.pipeline(
-            prompt,
-            num_inference_steps=num_inference_steps,
-            guidance_scale=guidance_scale,
-            num_images_per_prompt=num_images_per_prompt,
-            negative_prompt=negative_prompt,
-            height=height,
-            width=width,
-        )
-
+        if self.pipeline.device.type == "cuda":
+            with torch.autocast("cuda"):
+                out = self.pipeline(prompt, num_images_per_prompt=1)
+        else:
+            out = self.pipeline(prompt, num_images_per_prompt=1)
         return out.images[0]
 
 
 DIFFUSERS_TASKS = {
-    "text-to-image": DiffusersPipelineImageToText,
+    "text-to-image": IEAutoPipelineForText2Image,
 }
 
 

src/huggingface_inference_toolkit/utils.py

@@ -4,14 +4,13 @@
 from pathlib import Path
 from typing import Optional, Union
 
-from huggingface_hub import login, snapshot_download
+from huggingface_hub import HfApi, login, snapshot_download
 from transformers import WhisperForConditionalGeneration, pipeline
 from transformers.file_utils import is_tf_available, is_torch_available
 from transformers.pipelines import Conversation, Pipeline
 
 from huggingface_inference_toolkit.const import HF_DEFAULT_PIPELINE_NAME, HF_MODULE_NAME
 from huggingface_inference_toolkit.diffusers_utils import (
-    check_supported_pipeline,
     get_diffusers_pipeline,
     is_diffusers_available,
 )
@@ -46,11 +45,12 @@ def is_optimum_available():
     "pt": "pytorch*",
     "flax": "flax*",
     "rust": "rust*",
-    "onnx": "*onnx",
+    "onnx": "*onnx*",
     "safetensors": "*safetensors",
     "coreml": "*mlmodel",
     "tflite": "*tflite",
     "savedmodel": "*tar.gz",
+    "openvino": "*openvino*",
     "ckpt": "*ckpt",
 }
 
@@ -59,18 +59,8 @@ def create_artifact_filter(framework):
     """
     Returns a list of regex pattern based on the DL Framework. which will be to used to ignore files when downloading
     """
-    ignore_regex_list = [
-        "pytorch*",
-        "tf*",
-        "flax*",
-        "rust*",
-        "*onnx",
-        "*safetensors",
-        "*mlmodel",
-        "*tflite",
-        "*tar.gz",
-        "*ckpt",
-    ]
+    ignore_regex_list = list(set(framework2weight.values()))
+
     pattern = framework2weight.get(framework, None)
     if pattern in ignore_regex_list:
         ignore_regex_list.remove(pattern)
@@ -157,6 +147,12 @@ def _load_repository_from_hf(
     if not target_dir.exists():
         target_dir.mkdir(parents=True)
 
+    # check if safetensors weights are available
+    if framework == "pytorch":
+        files = HfApi().model_info(repository_id).siblings
+        if any(f.rfilename.endswith("safetensors") for f in files):
+            framework = "safetensors"
+
     # create regex to only include the framework specific weights
     ignore_regex = create_artifact_filter(framework)
     logger.info(f"Ignore regex pattern for files, which are not downloaded: { ', '.join(ignore_regex) }")
@@ -259,7 +255,7 @@ def get_pipeline(task: str, model_dir: Path, **kwargs) -> Pipeline:
         "sentence-ranking",
     ]:
         hf_pipeline = get_sentence_transformers_pipeline(task=task, model_dir=model_dir, device=device, **kwargs)
-    elif is_diffusers_available() and check_supported_pipeline(model_dir) and task == "text-to-image":
+    elif is_diffusers_available() and task == "text-to-image":
         hf_pipeline = get_diffusers_pipeline(task=task, model_dir=model_dir, device=device, **kwargs)
     else:
         hf_pipeline = pipeline(task=task, model=model_dir, device=device, **kwargs)
@@ -287,8 +283,8 @@ def convert_params_to_int_or_bool(params):
     for k, v in params.items():
         if v.isnumeric():
             params[k] = int(v)
-        if v == 'false':
+        if v == "false":
             params[k] = False
-        if v == 'true':
+        if v == "true":
             params[k] = True
     return params

tests/unit/test_diffusers.py

@@ -3,8 +3,8 @@
 from PIL import Image
 from transformers.testing_utils import require_torch, slow
 
-from huggingface_inference_toolkit.handler import get_inference_handler_either_custom_or_default_handler
-from huggingface_inference_toolkit.diffusers_utils import get_diffusers_pipeline, DiffusersPipelineImageToText
+
+from huggingface_inference_toolkit.diffusers_utils import get_diffusers_pipeline, IEAutoPipelineForText2Image
 from huggingface_inference_toolkit.utils import _load_repository_from_hf, get_pipeline
 
 
@@ -15,7 +15,7 @@ def test_get_diffusers_pipeline():
             "hf-internal-testing/tiny-stable-diffusion-torch", tmpdirname, framework="pytorch"
         )
         pipe = get_pipeline("text-to-image", storage_dir.as_posix())
-        assert isinstance(pipe, DiffusersPipelineImageToText)
+        assert isinstance(pipe, IEAutoPipelineForText2Image)
 
 
 @slow