switch to qwen2.5 vl (#2)

- switch to qwen2.5 vl (4480433cdbbbd3f77e86e967a2fa97e95417b18b)

app.py

@@ -1,14 +1,15 @@
-# import subprocess  # 🥲
-
-# subprocess.run(
-#     "pip install flash-attn --no-build-isolation",
-#     env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
-#     shell=True,
-# )
+import subprocess  # 🥲
 
+subprocess.run(
+    "pip install flash-attn --no-build-isolation",
+    env={"FLASH_ATTENTION_SKIP_CUDA_BUILD": "TRUE"},
+    shell=True,
+)
 import spaces
 import gradio as gr
-from transformers import AutoModelForCausalLM, AutoProcessor, GenerationConfig
+
+from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
+from qwen_vl_utils import process_vision_info
 import torch
 import os
 import json
@@ -17,19 +18,15 @@ from typing import Tuple
 
 os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
 
-# Load Molmo model
-model = AutoModelForCausalLM.from_pretrained(
-    'allenai/Molmo-7B-D-0924',
-    trust_remote_code=True,
-    torch_dtype='auto',
-    device_map='auto'
-)
-processor = AutoProcessor.from_pretrained(
-    'allenai/Molmo-7B-D-0924',
-    trust_remote_code=True,
-    torch_dtype='auto',
-    device_map='auto'
+
+model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+    "Qwen/Qwen2.5-VL-7B-Instruct",
+    torch_dtype=torch.bfloat16,
+    attn_implementation="flash_attention_2",
+    device_map="auto",
 )
+processor = AutoProcessor.from_pretrained("Qwen/Qwen2.5-VL-7B-Instruct")
+
 
 class GeneralRetrievalQuery(BaseModel):
     broad_topical_query: str
@@ -39,6 +36,7 @@ class GeneralRetrievalQuery(BaseModel):
     visual_element_query: str
     visual_element_explanation: str
 
+
 def get_retrieval_prompt(prompt_name: str) -> Tuple[str, GeneralRetrievalQuery]:
     if prompt_name != "general":
         raise ValueError("Only 'general' prompt is available in this version")
@@ -72,46 +70,77 @@ Format your response as a JSON object with the following structure:
 If there are no relevant visual elements, replace the third query with another specific detail query.
 
 Here is the document image to analyze:
+<image>
 
-Generate the queries based on this image and provide the response in the specified JSON format.
-Only return JSON. Don't return any extra explanation text. """
+Generate the queries based on this image and provide the response in the specified JSON format."""
 
     return prompt, GeneralRetrievalQuery
 
+
+# defined like this so we can later add more prompting options
 prompt, pydantic_model = get_retrieval_prompt("general")
 
+
 def _prep_data_for_input(image):
-    return processor.process(
-        images=[image],
-        text=prompt
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "image",
+                    "image": image,
+                },
+                {"type": "text", "text": prompt},
+            ],
+        }
+    ]
+
+    text = processor.apply_chat_template(
+        messages, tokenize=False, add_generation_prompt=True
+    )
+
+    image_inputs, video_inputs = process_vision_info(messages)
+
+    return processor(
+        text=[text],
+        images=image_inputs,
+        videos=video_inputs,
+        padding=True,
+        return_tensors="pt",
     )
 
-@spaces.GPU(duration=120)
+
+@spaces.GPU
 def generate_response(image):
     inputs = _prep_data_for_input(image)
-    inputs = {k: v.to(model.device).unsqueeze(0) for k, v in inputs.items()}
-    output = model.generate_from_batch(
-        inputs,
-        GenerationConfig(max_new_tokens=800, stop_token="<|endoftext|>"),
-        tokenizer=processor.tokenizer
+    inputs = inputs.to("cuda")
+
+    generated_ids = model.generate(**inputs, max_new_tokens=200)
+    generated_ids_trimmed = [
+        out_ids[len(in_ids) :]
+        for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
+    ]
+
+    output_text = processor.batch_decode(
+        generated_ids_trimmed,
+        skip_special_tokens=True,
+        clean_up_tokenization_spaces=False,
     )
-    generated_tokens = output[0, inputs['input_ids'].size(1):]
-    output_text = processor.tokenizer.decode(generated_tokens, skip_special_tokens=True)
-
     try:
-        return str(json.loads(output_text))
+        return json.loads(output_text[0])
     except Exception:
         gr.Warning("Failed to parse JSON from output")
-        return output_text
+        return {}
+
 
-title = "ColPali fine-tuning Query Generator"
+title = "ColPali Query Generator using Qwen2.5-VL"
 description = """[ColPali](https://huggingface.co/papers/2407.01449) is a very exciting new approach to multimodal document retrieval which aims to replace existing document retrievers which often rely on an OCR step with an end-to-end multimodal approach. 
 
 To train or fine-tune a ColPali model, we need a dataset of image-text pairs which represent the document images and the relevant text queries which those documents should match. 
 To make the ColPali models work even better we might want a dataset of query/image document pairs related to our domain or task. 
 
 One way in which we might go about generating such a dataset is to use a VLM to generate synthetic queries for us. 
-This space uses the [allenai/Molmo-7B-D-0924](https://huggingface.co/allenai/Molmo-7B-D-0924) model to generate queries for a document, based on an input document image. 
+This space uses the [Qwen/Qwen2.5-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-VL-7B-Instruct) VLM model to generate queries for a document, based on an input document image. 
 
 **Note** there is a lot of scope for improving to prompts and the quality of the generated queries! If you have any suggestions for improvements please [open a Discussion](https://huggingface.co/spaces/davanstrien/ColPali-Query-Generator/discussions/new)!
 
@@ -128,7 +157,7 @@ examples = [
 demo = gr.Interface(
     fn=generate_response,
     inputs=gr.Image(type="pil"),
-    outputs=gr.Text(),
+    outputs=gr.Json(),
     title=title,
     description=description,
     examples=examples,