Getting DeepSeek-OCR Working on an Nvidia Spark via Brute Force with Claude Code (simonwillison.net)

DeepSeek released DeepSeek‑OCR, a 6.6GB PyTorch/CUDA‑fine‑tuned OCR model, and a user successfully ran it on an NVIDIA Spark GB10 (sm_121) by delegating the environment setup to Claude Code. Inside a Docker container (nvcr.io/nvidia/cuda:13.0.1-devel-ubuntu24.04) Claude Code was given broad permissions and scripted cloning the repo, enabling LFS, installing PyTorch ARM64 wheels, and running inference against a test image. The main technical snag was hardware/driver compatibility: PyTorch 2.5.1 failed for sm_121 (“no kernel image”), but a search found PyTorch 2.9.0 ARM64 wheels for CUDA 12.8/12.9/13.0 that allowed the model to load (warning about max supported CUDA) and run. The agent tried different prompt modes (grounding, free OCR, markdown) to balance text extraction vs. bounding‑box outputs; bounding boxes worked immediately while text output required switching prompts. This is significant because it demonstrates an effective agentic loop for real‑world ML ops: an LLM agent can diagnose wheel availability, install compatible binaries, and iterate on prompts to get usable outputs on new ARM CUDA hardware in under 40 minutes with minimal human time. Practical takeaways include using Docker sandboxes to safely permit autonomous agents, the value of checking upstream wheel indexes for platform‑specific binaries, and monitoring containers via VS Code Remote/Dev Containers. It’s a clear example of AI assisting deployment and debugging of research models — speeding experimentation while still needing human judgment for final validation and reproducibility.