Merge pull request #45 from huggingface/sd-xl-fix

Fix Stable diffusion pipeline

Author: philschmid

dockerfiles/pytorch/cpu/environment.yaml

@@ -8,6 +8,6 @@ dependencies:
   - transformers[sklearn,sentencepiece,audio,vision]==4.31.0
   - sentence_transformers==2.2.2
   - torchvision==0.14.1
-  - diffusers==0.19.3
+  - diffusers==0.20.0
   - accelerate==0.21.0
   - safetensors

dockerfiles/pytorch/gpu/environment.yaml

@@ -9,6 +9,6 @@ dependencies:
   - transformers[sklearn,sentencepiece,audio,vision]==4.31.0
   - sentence_transformers==2.2.2
   - torchvision==0.14.1
-  - diffusers==0.19.3
+  - diffusers==0.20.0
   - accelerate==0.21.0
   - safetensors

src/huggingface_inference_toolkit/diffusers_utils.py

@@ -1,6 +1,8 @@
 import importlib.util
 import logging
 
+from transformers.utils.import_utils import is_torch_bf16_gpu_available
+
 logger = logging.getLogger(__name__)
 logging.basicConfig(format="%(asctime)s | %(levelname)s | %(message)s", level=logging.INFO)
 
@@ -20,7 +22,7 @@ class IEAutoPipelineForText2Image:
     def __init__(self, model_dir: str, device: str = None):  # needs "cuda" for GPU
         dtype = torch.float32
         if device == "cuda":
-            dtype = torch.float16
+            dtype = torch.bfloat16 if is_torch_bf16_gpu_available() else torch.float16
         device_map = "auto" if device == "cuda" else None
 
         self.pipeline = AutoPipelineForText2Image.from_pretrained(model_dir, torch_dtype=dtype, device_map=device_map)
@@ -43,11 +45,7 @@ def __call__(
             logger.warning("Sending num_images_per_prompt > 1 to pipeline is not supported. Using default value 1.")
 
         # Call pipeline with parameters
-        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)
+        out = self.pipeline(prompt, num_images_per_prompt=1, **kwargs)
         return out.images[0]
 
 

src/huggingface_inference_toolkit/webservice_starlette.py

@@ -75,7 +75,6 @@ async def predict(request):
         # check for query parameter and add them to the body
         if request.query_params and "parameters" not in deserialized_body:
             deserialized_body["parameters"] = convert_params_to_int_or_bool(dict(request.query_params))
-        print(deserialized_body)
 
         # tracks request time
         start_time = perf_counter()