Skip to content

fix(workflow): size auto tiles for the segment GPU's VRAM - #27

Merged
lguerard merged 1 commit into
mainfrom
fix/workflow-gpu-mem
Jun 24, 2026
Merged

fix(workflow): size auto tiles for the segment GPU's VRAM#27
lguerard merged 1 commit into
mainfrom
fix/workflow-gpu-mem

Conversation

@lguerard

Copy link
Copy Markdown
Contributor

Summary

On SLURM the prepare (tile planning) job runs on a CPU node, so it can't see the segment GPU. With tile_shape: "auto" it logged:

GPU memory query failed (nvidia-ml-py not installed?); using 8 GiB default.

…and sized every tile for 8 GiB — harmless but under-utilises bigger GPUs.

Add a gpu_memory_gb config field forwarded to auto_tile_shape_cellpose(gpu_memory=…) so tiles are sized for the real GPU (e.g. 24, 40, 80) with no query and no warning. null keeps the old 8 GiB-guess behaviour. Documented in the config + guide.

Test

auto_tile_shape_cellpose((64,4096,4096), do_3D=True, use_gpu=True, gpu_memory=24G)(64,1753,1753) and emits no warning; gpu_memory=None(64,915,915) with the fallback warning. Full local workflow run still 12/12.

🤖 Generated with Claude Code

Tile planning (prepare) runs on a CPU node, so it cannot query the
segment GPU and falls back to an 8 GiB guess, logging "GPU memory query
failed". Add a gpu_memory_gb config option forwarded to
auto_tile_shape_cellpose so tiles are sized for the real GPU (e.g. 24,
40, 80) without a query or warning. null keeps the previous behaviour.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
@lguerard
lguerard merged commit 2efe635 into main Jun 24, 2026
2 checks passed
@lguerard
lguerard deleted the fix/workflow-gpu-mem branch June 24, 2026 13:39
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant