Skip to content

feat(launcher): expose GPUs to eval container for compute-eval #912

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wprazuch wants to merge 3 commits into
base: main
Choose a base branch
from
Open

feat(launcher): expose GPUs to eval container for compute-eval #912

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
33debd3
fix(launcher): re-raise unrelated ModuleNotFoundError in container_me…
wprazuch Apr 14, 2026
4d51bee
fix(launcher): lazy-import nemo_evaluator in loading.py to fix docs env
wprazuch Apr 14, 2026
5ccb8f9
feat(launcher): configurable GPU access for eval container
wprazuch Apr 16, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 6 additions & 2 deletions ...nemo-evaluator-launcher/src/nemo_evaluator_launcher/common/container_metadata/__init__.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,6 +56,10 @@
"parse_framework_to_irs",
]
)
except ModuleNotFoundError:
except ModuleNotFoundError as exc:
# Allow importing this package for IR-only workflows (docs autogen, etc.)
pass
# Only suppress when nemo_evaluator itself is missing; re-raise if a
# transitive dependency (pydantic, structlog, …) is absent — that is a
# broken installation, not an IR-only workflow.
if exc.name is None or not exc.name.startswith("nemo_evaluator"):
raise
3 changes: 2 additions & 1 deletion .../nemo-evaluator-launcher/src/nemo_evaluator_launcher/common/container_metadata/loading.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -25,7 +25,6 @@
from typing import Optional

import yaml
from nemo_evaluator.core.input import get_framework_evaluations

from nemo_evaluator_launcher.common.container_metadata.intermediate_repr import (
HarnessIntermediateRepresentation,
Expand Down Expand Up @@ -943,6 +942,8 @@ def parse_framework_to_irs(
temp_file_path = temp_file.name

try:
from nemo_evaluator.core.input import get_framework_evaluations

include_internal = (
importlib.util.find_spec("nemo_evaluator_internal") is not None
or importlib.util.find_spec("nemo_evaluator_launcher_internal")
Expand Down
11 changes: 10 additions & 1 deletion packages/nemo-evaluator-launcher/src/nemo_evaluator_launcher/executors/slurm/executor.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1014,11 +1014,20 @@ def _create_slurm_sbatch_script(
aux_extra_env_names.extend(endpoint_vars)

s += "# evaluation client\n"
# When evaluation_gpu_visible is true, expose GPUs to the eval container.
# Required for benchmarks that compile/execute CUDA code (e.g. compute-eval).
eval_gpu_visible = cfg.execution.get("evaluation_gpu_visible", False)
extra_eval_env_names: list[str] = []
if eval_gpu_visible:
s += "export NVIDIA_VISIBLE_DEVICES=all\n"
extra_eval_env_names.append("NVIDIA_VISIBLE_DEVICES")
s += "srun --mpi pmix --overlap "
s += '--nodelist "${PRIMARY_NODE}" --nodes 1 --ntasks 1 '
s += "--container-image {} ".format(eval_image)
# Combine eval env vars with auxiliary endpoint env vars
all_eval_env_names = sorted(set(list(eval_env_vars.keys()) + aux_extra_env_names))
all_eval_env_names = sorted(
set(list(eval_env_vars.keys()) + aux_extra_env_names + extra_eval_env_names)
)
if all_eval_env_names:
s += "--container-env {} ".format(",".join(all_eval_env_names))
if not cfg.execution.get("mounts", {}).get("mount_home", True):
Expand Down
12 changes: 3 additions & 9 deletions packages/nemo-evaluator-launcher/tests/unit_tests/test_loading.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -208,9 +208,7 @@ def test_create_task_irs(self, mock_extract):
class TestParseFrameworkToIrs:
"""Test parsing framework.yml to IRs."""

@patch(
"nemo_evaluator_launcher.common.container_metadata.loading.get_framework_evaluations"
)
@patch("nemo_evaluator.core.input.get_framework_evaluations")
def test_parse_framework_to_irs_success(self, mock_get_evaluations):
"""Test successful parsing of framework.yml."""
framework_content = """
Expand All @@ -237,9 +235,7 @@ def test_parse_framework_to_irs_success(self, mock_get_evaluations):
assert harness_ir.name == "test-harness" # Original name preserved
assert len(task_irs) > 0

@patch(
"nemo_evaluator_launcher.common.container_metadata.loading.get_framework_evaluations"
)
@patch("nemo_evaluator.core.input.get_framework_evaluations")
def test_parse_framework_to_irs_preserves_case(self, mock_get_evaluations):
"""Test that framework name case is preserved."""
framework_content = """
Expand Down Expand Up @@ -274,9 +270,7 @@ def test_parse_framework_to_irs_preserves_name(self):
"""
# This should not raise an error - no validation in loading.py
# We'll mock get_framework_evaluations to avoid actual parsing
with patch(
"nemo_evaluator_launcher.common.container_metadata.loading.get_framework_evaluations"
) as mock_get:
with patch("nemo_evaluator.core.input.get_framework_evaluations") as mock_get:
mock_get.return_value = ("test-harness-123", {}, {})
harness_ir, task_irs = parse_framework_to_irs(
framework_content, "test:latest", None
Expand Down
Loading