feat(launcher): expose GPUs to eval container for compute-eval#912
Open
feat(launcher): expose GPUs to eval container for compute-eval#912
Conversation
…tadata init The bare `except ModuleNotFoundError: pass` silently swallowed errors from transitive dependencies (pydantic, structlog, etc.) when nemo_evaluator was on PYTHONPATH as source but its deps were not installed. This made extract_framework_yml silently disappear from the namespace. Now only suppresses the error when nemo_evaluator itself is the missing module; any other missing dependency is re-raised. Co-Authored-By: Claude Opus 4.6 (1M context) <[email protected]> Signed-off-by: Wojciech Prazuch <[email protected]>
The top-level `from nemo_evaluator.core.input import ...` in loading.py triggered the full nemo_evaluator import chain (including pydantic) at module import time. In docs environments where nemo_evaluator source is on PYTHONPATH but pydantic is not installed, this caused a ModuleNotFoundError for pydantic that was silently swallowed by the try/except in __init__.py, making extract_framework_yml and parse_framework_to_irs unavailable. Move the import to the only call site (parse_framework_to_irs) so loading.py can be imported without the full nemo_evaluator dep tree. Signed-off-by: Wojciech Prazuch <[email protected]>
Add `execution.evaluation_gpu_visible` config option (default: false).
When true, exports `NVIDIA_VISIBLE_DEVICES=all` and passes it to the
eval container, enabling GPU access for benchmarks that compile and
execute CUDA code (e.g. compute-eval).
Usage in config:
execution:
evaluation_gpu_visible: true
Without this, the eval srun runs without GPU visibility and CUDA
binaries fail with cudaErrorInsufficientDriver.
Validated with compute-eval on HSG: pass@1 = 51.25% (41/80).
Signed-off-by: Wojciech Prazuch <[email protected]>
wprazuch
force-pushed
the
wprazuch/eval-gpu-visibility
branch
from
463f8f0 to
5ccb8f9
Compare
chtruong814
added
docs-only
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
2 participants
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
NVIDIA_VISIBLE_DEVICES=allbefore the eval srun command--container-envWhy
Benchmarks like compute-eval compile and execute CUDA code inside the eval container. Without GPU access,
nvcccan't detect the target architecture (-arch=nativefalls back to default) and compiled binaries fail withcudaErrorInsufficientDriver.Testing
Validated with compute-eval on HSG (driver 580/CUDA 13.0):
cudaErrorInsufficientDriver)Test plan
🤖 Generated with Claude Code