adapt llava pipeline to latest Transformers

- change TF pipeline name from image-to-text to image-text-to-text huggingface/transformers#34769
- simplify the preprocessing logics, keep the handling when image is empty string, and remove the multi-image inference in one run
- fix few README errors

comps/lvms/src/integrations/dependency/llava/Dockerfile.intel_hpu

@@ -22,7 +22,8 @@ COPY comps /home/user/comps
 # Install requirements and optimum habana
 RUN pip install --no-cache-dir --upgrade pip && \
     pip install --no-cache-dir -r /home/user/comps/lvms/src/integrations/dependency/llava/requirements.txt && \
-    pip install --no-cache-dir optimum[habana]
+    pip install git+/~https://github.com/huggingface/optimum-habana.git@transformers_future && \
+    pip install --no-cache-dir --upgrade Jinja2
 
 ENV PYTHONPATH=$PYTHONPATH:/home/user
 USER user

comps/lvms/src/integrations/dependency/llava/README.md

@@ -82,13 +82,10 @@ docker run -p 8399:8399 --runtime=habana -e HABANA_VISIBLE_DEVICES=all -e OMPI_M
 # Use curl/python
 
 # curl with an image and a prompt
-http_proxy="" curl http://localhost:9399/v1/lvm -XPOST -d '{"image": "iVBORw0KGgoAAAANSUhEUgAAAAoAAAAKCAYAAACNMs+9AAAAFUlEQVR42mP8/5+hnoEIwDiqkL4KAcT9GO0U4BxoAAAAAElFTkSuQmCC", "prompt":"What is this?"}' -H 'Content-Type: application/json'
-
-# curl with multiple images and a prompt (Note that depending on your MAX_IMAGES value, both images may not be sent to the LLaVA model)
-http_proxy="" curl http://localhost:9399/v1/lvm -XPOST -d '{"image": ["iVBORw0KGgoAAAANSUhEUgAAAAoAAAAKCAYAAACNMs+9AAAAFUlEQVR42mNkYPhfz0AEYBxVSF+FAP5FDvcfRYWgAAAAAElFTkSuQmCC", "iVBORw0KGgoAAAANSUhEUgAAAAoAAAAKCAYAAACNMs+9AAAAFUlEQVR42mNk+M9Qz0AEYBxVSF+FAAhKDveksOjmAAAAAElFTkSuQmCC"], "prompt":"What is in these images?"}' -H 'Content-Type: application/json'
+http_proxy="" curl http://localhost:8399/generate -XPOST -d '{"img_b64_str": "iVBORw0KGgoAAAANSUhEUgAAAAoAAAAKCAYAAACNMs+9AAAAFUlEQVR42mP8/5+hnoEIwDiqkL4KAcT9GO0U4BxoAAAAAElFTkSuQmCC", "prompt":"What is this?"}' -H 'Content-Type: application/json'
 
 # curl with a prompt only (no image)
-http_proxy="" curl http://localhost:9399/v1/lvm -XPOST -d '{"image": "", "prompt":"What is deep learning?"}' -H 'Content-Type: application/json'
+http_proxy="" curl http://localhost:8399/generate -XPOST -d '{"img_b64_str": "", "prompt":"What is deep learning?"}' -H 'Content-Type: application/json'
 
 # Test
 python check_llava_server.py

comps/lvms/src/integrations/dependency/llava/llava_server.py

@@ -15,7 +15,6 @@
 from fastapi import FastAPI, Request
 from fastapi.responses import JSONResponse, Response
 from transformers import AutoProcessor, pipeline
-from transformers.image_utils import load_image
 
 model_name_or_path = None
 model_dtype = None
@@ -27,79 +26,6 @@
 app = FastAPI()
 
 
-def pipeline_preprocess(self, image, prompt=None, timeout=None):
-    """
-    This replaces the preprocess function used by the image-to-text pipeline
-    (/~https://github.com/huggingface/transformers/blob/main/src/transformers/pipelines/image_to_text.py).
-    The original transformers image-to-text pipeline preprocess function requires that an image is passed in, and will
-    fail if the image parameter is null/empty. In order to support multimodal use cases with the same pipeline, this
-    preprocess function handles the case where there is no image with the prompt.
-    Also, the image-to-text pipeline typically treats multiple images passed in as a list as a batch (where it iterates
-    over the image inputs for generation). For that reason, the original pipeline_preprocess code would only get a
-    single image at a time. To support multiple images, the pipeline call is updated to send a list of lists for the
-    images (so that when iterated, we still get multiple images) and this pipeline_preprocess function has been updated
-    to handle a list of images in addition to single images.
-    """
-
-    if isinstance(image, list):
-        image = [load_image(i, timeout=timeout) for i in image]
-    elif image:
-        image = load_image(image, timeout=timeout)
-
-    if prompt is not None:
-        if not isinstance(prompt, str):
-            raise ValueError(
-                f"Received an invalid text input, got - {type(prompt)} - but expected a single string. "
-                "Note also that one single text can be provided for conditional image to text generation."
-            )
-
-        model_type = self.model.config.model_type
-
-        if model_type == "git":
-            if image:
-                model_inputs = self.image_processor(images=image, return_tensors=self.framework)
-                if self.framework == "pt":
-                    model_inputs = model_inputs.to(self.torch_dtype)
-            else:
-                model_inputs = {}
-            input_ids = self.tokenizer(text=prompt, add_special_tokens=False).input_ids
-            input_ids = [self.tokenizer.cls_token_id] + input_ids
-            input_ids = torch.tensor(input_ids).unsqueeze(0)
-            model_inputs.update({"input_ids": input_ids})
-        elif model_type == "pix2struct":
-            model_inputs = self.image_processor(images=image, header_text=prompt, return_tensors=self.framework)
-            if self.framework == "pt":
-                model_inputs = model_inputs.to(self.torch_dtype)
-
-        elif model_type != "vision-encoder-decoder":
-            if image:
-                # vision-encoder-decoder does not support conditional generation
-                model_inputs = self.image_processor(images=image, return_tensors=self.framework)
-
-                if self.framework == "pt":
-                    model_inputs = model_inputs.to(self.torch_dtype)
-            else:
-                model_inputs = {}
-
-            text_inputs = self.tokenizer(prompt, return_tensors=self.framework)
-            model_inputs.update(text_inputs)
-
-        else:
-            raise ValueError(f"Model type {model_type} does not support conditional text generation")
-
-    elif image:
-        model_inputs = self.image_processor(images=image, return_tensors=self.framework)
-        if self.framework == "pt":
-            model_inputs = model_inputs.to(self.torch_dtype)
-    else:
-        raise ValueError("Both image and prompt cannot be empty.")
-
-    if self.model.config.model_type == "git" and prompt is None:
-        model_inputs["input_ids"] = None
-
-    return model_inputs
-
-
 def process_image(image, max_len=1344, min_len=672):
     if max(image.size) > max_len:
         max_hw, min_hw = max(image.size), min(image.size)
@@ -122,11 +48,11 @@ async def health() -> Response:
 
 
 @app.post("/generate")
-async def generate(request: Request) -> Response:  # FIXME batch_size=1 for now
+async def generate(request: Request) -> Response:
     print("LLaVA generation begin.")
     request_dict = await request.json()
     prompt = request_dict.pop("prompt")
-    img_b64_str = request_dict.pop("img_b64_str")  # String or list of strings
+    img_b64_str = request_dict.pop("img_b64_str")   # Only accept string
     max_new_tokens = request_dict.pop("max_new_tokens", 100)
 
     # Determine the format of the role labels based on the model name
@@ -183,12 +109,9 @@ async def generate(request: Request) -> Response:  # FIXME batch_size=1 for now
 
     start = time.time()
 
-    # Override the pipeline preprocessing
-    generator.preprocess = pipeline_preprocess.__get__(generator, type(generator))
-
-    result = generator([images], prompt=prompt, batch_size=1, generate_kwargs=generate_kwargs)
+    result = generator(images, text=prompt, batch_size=1, generate_kwargs=generate_kwargs)
     end = time.time()
-    result = result[0][0]["generated_text"].split(output_assistant_label.strip())[-1].strip()
+    result = result[0]["generated_text"].split(output_assistant_label.strip())[-1].strip()
     print(f"LLaVA result = {result}, time = {(end-start) * 1000 }ms")
     if images:
         for i in images:
@@ -223,7 +146,7 @@ async def generate(request: Request) -> Response:  # FIXME batch_size=1 for now
     model_name_or_path = args.model_name_or_path
 
     generator = pipeline(
-        "image-to-text",
+        "image-text-to-text",
         model=args.model_name_or_path,
         torch_dtype=model_dtype,
         device=args.device,
@@ -266,11 +189,10 @@ async def generate(request: Request) -> Response:  # FIXME batch_size=1 for now
         },
     ]
     text_prompt = processor.apply_chat_template(conversation)
-
     for i in range(args.warmup):
-        generator(
+        res = generator(
             images,
-            prompt=text_prompt,
+            text=text_prompt,
             batch_size=1,
             generate_kwargs=generate_kwargs,
         )

comps/lvms/src/integrations/dependency/llava/requirements.txt

@@ -6,7 +6,7 @@ langchain-core
 opentelemetry-api
 opentelemetry-exporter-otlp
 opentelemetry-sdk
-optimum[habana]
+transformers
 prometheus-fastapi-instrumentator
 pydantic==2.7.2
 pydub