fix(embed): pin stage3 eval torch to CUDA 12.9 build#164
Open
shan-nvidia wants to merge 1 commit intomainfrom
Open
fix(embed): pin stage3 eval torch to CUDA 12.9 build#164shan-nvidia wants to merge 1 commit intomainfrom
shan-nvidia wants to merge 1 commit intomainfrom
Conversation
Without an explicit pin, `uv` resolves `torch==2.11.0` from PyPI for the stage3 eval recipe, which requires CUDA 13 (`cuda-toolkit==13.0.2`). On systems with NVIDIA drivers that only support CUDA 12.x, `torch.cuda.is_available()` returns False and BEIR silently falls back to CPU, making eval effectively unusable on multi-GPU hosts. This mirrors the pattern nemo-automodel uses for other stages: pin `torch<=2.10.0` and point `[tool.uv.sources]` at the `pytorch-cu129` index so Linux gets `torch==2.10.0+cu129`. Stages 1 and 2 already get this via their nemo-automodel dependency; stage3 doesn't use nemo-automodel, so the configuration is added explicitly here. Signed-off-by: Steve Han <sthan@nvidia.com> Made-with: Cursor
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Without an explicit pin,
uvresolvestorch==2.11.0from PyPI for thestage3_evalrecipe, which requires CUDA 13 (cuda-toolkit==13.0.2). On systems whose NVIDIA driver only supports CUDA 12.x,torch.cuda.is_available()returns False and BEIR silently falls back to CPU, making eval effectively unusable on multi-GPU hosts — you just see theNVIDIA driver ... is too old (found version 12090)warning and the 0/4 progress bar crawls.Stages 1 and 2 avoid this because their
nemo-automodeldependency transitively pinstorch==2.10.0+cu129via[tool.uv.sources]+ thepytorch-cu129index. Stage 3 doesn't depend onnemo-automodel, so this PR adds the same configuration directly:torch<=2.10.0todependenciesnemo-automodel's[tool.uv.sources]torch selector and thepytorch-cpu/pytorch-cu129/pypiindex blocksAlso includes the regenerated
uv.lockso Linux resolves totorch 2.10.0+cu129.Test plan
rm -rf src/nemotron/recipes/embed/stage3_eval/.venvnemotron embed eval -c default <finetuned/eval/output paths>runs without theNVIDIA driver ... is too oldwarningnvidia-smiduring eval shows GPU memory and utilization across visible devices