Hey Alex, thanks so much for putting this together! Blackwell/DGX Spark support is exactly the kind of thing I've been wanting to add — the unified memory quirks and the NVRTC SM_121 compatibility issues are real pain points that I haven't had hardware to tackle myself.

A quick heads-up on timing: we're currently deep into the 0.4.0 release cycle on our feat/search-and-rag branch, which has diverged significantly from master (new ASR provider architecture, PyAnnote GPU optimizations, search/RAG features, etc.). That branch is where we're landing before the next release.

Once we cut 0.4.0 and merge to master, integrating this would be a great follow-up. In the meantime, I'd love it if you could keep an eye on how the 0.4.0 changes interact with your Blackwell setup — there are some GPU-side changes that may actually help with your use case, and it'd be great to have you test once we're closer to merging.

Really appreciate the contribution — and the clever workaround for the jiterator/fbank crash is something I hadn't seen documented anywhere else.

I've gone through the diff in detail and put together an integration plan for merging this into feat/search-and-rag: https://gist.github.com/davidamacey/d661109796840fc00187d165c52835dd

The short version: the additive pieces (Dockerfile.blackwell, docker-compose.blackwell.yml, the GPU stats fallback logic) are solid and I want to include all of them. There are a few things that'll need adjusting before they can land on the 0.4.0 branch:

• docker-compose.yml container_name removals — the branch still has those and various scripts depend on them by name, so I'll need to drop that part of the diff
• utility.py — the branch has a significantly rewritten multi-GPU version of update_gpu_stats() (with utilization %, temperature, per-device tracking). The Blackwell cudaMemGetInfo fallback is great — I'll graft it into the branch's version rather than replace it
• opentranscribe.sh detection — the file-existence check for docker-compose.blackwell.yml would activate the Blackwell overlay for all NVIDIA users once the file is in the repo. Planning to replace it with an actual nvidia-smi --query-gpu=compute_cap check for SM_12.x
• huggingface_hub==0.23.5 pin — hard-pinning that globally would conflict with other deps; I'll move it into Dockerfile.blackwell only where it's actually needed
• Dependency versions — the branch is on whisperx==3.8.1 and a custom pyannote fork with GPU optimisations; Dockerfile.blackwell will need to track those

Where your help would be most valuable:

1. whisperx==3.8.1 on DGX Spark — can you verify it installs and runs correctly with the NVIDIA PyTorch base? Same --no-deps trick you're already using should work
2. Compute cap detection — does nvidia-smi --query-gpu=compute_cap return 12.1 on your GB10? (Needed for the improved auto-detection)
3. pyannote fork — does pyannote.audio @ git+https://github.com/davidamacey/pyannote-audio.git@gpu-optimizations install cleanly on top of the NVIDIA container? The fork has significant GPU speedups that would benefit your setup

Planning to integrate this into 0.4.1 right after the 0.4.0 release. Happy to loop you in early if you want to test against the branch.