Seperate everything into files

Dockerfile

@@ -1,25 +1,37 @@
+# =============================================================================
+# vLLM Server with DotsOCR Model for Hugging Face Spaces
+# =============================================================================
+
 # Base: official vLLM OpenAI-compatible server (tested version family)
 FROM vllm/vllm-openai:v0.9.1
 
-# Install required packages as root
-RUN pip3 install flash_attn==2.8.0.post2
-RUN pip3 install transformers==4.51.3
+# Add metadata labels
+LABEL maintainer="kybtech-vllm-dotsocr"
+LABEL description="vLLM server with DotsOCR model for document OCR"
+LABEL version="1.0"
 
-# vLLM needs the custom model to be registered before main() runs.
-# The model authors recommend importing their vLLM adapter into the vllm CLI module.
-# Do this as root before switching to user (required for HF Spaces)
-RUN sed -i '/^from vllm\.entrypoints\.cli\.main import main$/a\
-from DotsOCR import modeling_dots_ocr_vllm' $(which vllm)
+# =============================================================================
+# SYSTEM SETUP (as root)
+# =============================================================================
 
-# Show the patched part of the vllm script for verification
-RUN grep -A 1 'from vllm.entrypoints.cli.main import main' $(which vllm)
+# Install additional Python packages
+COPY requirements.txt /tmp/requirements.txt
+RUN pip3 install -r /tmp/requirements.txt
 
-# Set up a new user named "user" with user ID 1000
-RUN useradd -m -u 1000 user
+# Patch vLLM entrypoint to register DotsOCR model
+# This must be done as root before switching to user (required for HF Spaces)
+COPY patch_vllm.sh /tmp/patch_vllm.sh
+RUN chmod +x /tmp/patch_vllm.sh && /tmp/patch_vllm.sh
 
-# Copy the startup script and make it executable (as root)
+# Create user and copy application files
+RUN useradd -m -u 1000 user
 COPY start_server.sh /home/user/app/start_server.sh
-RUN chmod +x /home/user/app/start_server.sh
+COPY download_model.py /home/user/app/download_model.py
+RUN chmod +x /home/user/app/start_server.sh /home/user/app/download_model.py
+
+# =============================================================================
+# APPLICATION SETUP (as user)
+# =============================================================================
 
 # Switch to the "user" user
 USER user
@@ -39,20 +51,7 @@ ENV HF_HUB_ENABLE_HF_TRANSFER=1 \
 # Pre-download the model repo using Hugging Face cache
 # Note: dots.ocr requires the directory name to avoid '.' (see model card).
 ARG MODEL_ID=rednote-hilab/dots.ocr
-RUN python3 - <<'PY'
-from huggingface_hub import snapshot_download
-import os
-# Download model to HF cache (default location: ~/.cache/huggingface/hub)
-# This automatically handles caching, deduplication, and proper directory structure
-model_path = snapshot_download(
-    repo_id=os.environ.get("MODEL_ID", "rednote-hilab/dots.ocr"),
-    allow_patterns=["*"]
-)
-print(f"Model downloaded to: {model_path}")
-# Write the model path to a file for later use
-with open("/home/user/app/model_path.txt", "w") as f:
-    f.write(model_path)
-PY
+RUN MODEL_ID=$MODEL_ID python3 /home/user/app/download_model.py
 
 # Set the model path from the downloaded location
 RUN HF_MODEL_PATH=$(cat /home/user/app/model_path.txt) && \
@@ -63,9 +62,17 @@ RUN HF_MODEL_PATH=$(cat /home/user/app/model_path.txt) && \
 ENV HF_MODEL_PATH=/home/user/.cache/huggingface/hub/models--rednote-hilab--dots.ocr
 ENV PYTHONPATH="/home/user/.cache/huggingface/hub/models--rednote-hilab--dots.ocr:${PYTHONPATH}"
 
+# =============================================================================
+# RUNTIME CONFIGURATION
+# =============================================================================
+
 # Expose the Space port
 EXPOSE 7860
 ENV PORT=7860
 
+# Add health check (using httpx instead of curl)
+HEALTHCHECK --interval=30s --timeout=10s --start-period=60s --retries=3 \
+    CMD python3 -c "import httpx; httpx.get(f'http://localhost:{PORT}/health', timeout=5)" || exit 1
+
 # Use the startup script as entrypoint
 ENTRYPOINT ["/home/user/app/start_server.sh"]

download_model.py

@@ -0,0 +1,36 @@
+#!/usr/bin/env python3
+"""
+Download the dots.ocr model from Hugging Face Hub.
+This script downloads the model to the HF cache and writes the path to a file.
+"""
+
+import os
+import sys
+from huggingface_hub import snapshot_download
+
+
+def main():
+    # Get model ID from environment variable or use default
+    model_id = os.environ.get("MODEL_ID", "rednote-hilab/dots.ocr")
+    
+    print(f"Downloading model: {model_id}")
+    
+    # Download model to HF cache (default location: ~/.cache/huggingface/hub)
+    # This automatically handles caching, deduplication, and proper directory structure
+    model_path = snapshot_download(
+        repo_id=model_id,
+        allow_patterns=["*"]
+    )
+    
+    print(f"Model downloaded to: {model_path}")
+    
+    # Write the model path to a file for later use
+    output_file = "/home/user/app/model_path.txt"
+    with open(output_file, "w") as f:
+        f.write(model_path)
+    
+    print(f"Model path written to: {output_file}")
+
+
+if __name__ == "__main__":
+    main()

patch_vllm.sh

@@ -0,0 +1,16 @@
+#!/bin/bash
+# patch_vllm.sh
+# Patches the vLLM entrypoint to register the DotsOCR model
+
+set -e
+
+echo "Patching vLLM entrypoint to register DotsOCR model..."
+
+# Patch the vLLM script to import the DotsOCR modeling code
+sed -i '/^from vllm\.entrypoints\.cli\.main import main$/a\
+from DotsOCR import modeling_dots_ocr_vllm' $(which vllm)
+
+echo "vLLM script after patch:"
+grep -A 1 'from vllm.entrypoints.cli.main import main' $(which vllm)
+
+echo "vLLM patching completed successfully!"

requirements.txt

@@ -0,0 +1,5 @@
+# Additional packages required for dots.ocr
+flash_attn==2.8.0.post2
+transformers==4.51.3
+huggingface_hub>=0.24.0
+httpx>=0.24.0