fix: Add vLLM Dockerfile with DeepSeek-OCR dependencies

[elasticdotventures]

Summary

Adds custom Dockerfile for vLLM to include missing Python dependencies required by the DeepSeek-OCR model.

Changes

• Dockerfile.vllm: New custom Dockerfile based on vllm/vllm-openai:v0.8.5.post1
  • Installs missing packages: addict, easydict, matplotlib
• docker-compose.yml: Updated vllm service to build from custom Dockerfile instead of using pre-built image
  • Changed from using pre-built vllm/vllm-openai:v0.8.5.post1 to building blender-mcp-vllm:local

Problem Solved

The vLLM container was failing to start with error:

ImportError: This modeling file requires the following packages that were not found in your environment: addict, matplotlib

Then after adding those dependencies:

ImportError: This modeling file requires the following packages that were not found in your environment: easydict

Testing

• ✅ Docker image builds successfully with all dependencies
• ✅ All three required packages (addict, easydict, matplotlib) are installed
• ⚠️ Container starts but encounters a secondary compatibility issue (see Known Issues)

Known Issues

⚠️ Container still fails with transformers compatibility issue:

ImportError: cannot import name 'LlamaFlashAttention2' from 'transformers.models.llama.modeling_llama'

This appears to be a deeper compatibility issue between DeepSeek-OCR model code and the transformers version in vllm v0.8.5.post1. The model's custom code expects LlamaFlashAttention2 which is not available in the installed transformers version.

Next Steps

• Investigate transformers version compatibility with DeepSeek-OCR
• Consider pinning transformers to a specific version
• Explore alternative vllm versions
• Test with a compatible model or wait for model/vllm updates

Related

• Related PR in blender-awesome repo: feat: Add vLLM Docker stack integration and improve justfile blender-agent-tools#2

🤖 Generated with Claude Code