fix(runtime): honor system Python for all actor environments

[jepio]

What does this PR do ?

Makes NEMO_RL_PY_EXECUTABLES_SYSTEM=1 apply to the FSDP, Automodel, NeMo Gym, trajectory collector, replay buffer, and synchronous rollout actor environments. This prevents those actors from creating specialized uv environments when the runtime uses system Python.

The default mappings remain unchanged when the flag is disabled. ModelOpt retains its existing independent handling of the system-Python flag.

Issues

N/A

Usage

Set NEMO_RL_PY_EXECUTABLES_SYSTEM=1 when all actor dependencies are available in the system Python environment.

Before your PR is "Ready for review"

Pre checks:

• Read and followed the contributor guidelines.
• Added focused tests for both flag states.
• Ran the focused unit tests locally in Lima.
• No documentation change is needed.

Additional Information

Validation in Lima:

• Focused unit tests: 2 passed.
• End-of-file, trailing-whitespace, Ruff lint, Ruff format, and Pyrefly checks passed for the changed Python files.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[jepio]

/ok to test b869b51

[jepio]

/ok to test 799ff4f

[jepio]

Agent-team review of PR #3982. Clean, well-scoped bug fix — 3 actors (AsyncTrajectoryCollector, ReplayBuffer, SyncRolloutActor) were hardcoded to raw PY_EXECUTABLES.VLLM instead of the flag-aware VLLM_EXECUTABLE, so they silently ignored NEMO_RL_PY_EXECUTABLES_SYSTEM=1. No bugs found; registry completeness and ModelOpt interaction independently verified. Two low-severity, non-blocking suggestions below (doc + test coverage).

Generated by Claude Code

[jepio]

nemo_rl/distributed/ray_actor_environment_registry.py:33

1 action item.

NEMO_RL_PY_EXECUTABLES_SYSTEM isn't documented anywhere in the repo outside this PR's description (only referenced in the two registry .py files). This PR expands its effect from generation-only backends to the full training path (FSDP/Automodel/NeMo Gym) plus the async rollout actors, so it's worth a sentence in docs/design-docs/uv.md for future readers.

AI-1

Add a sentence to docs/design-docs/uv.md (around line 43, end of the "Worker Configuration" paragraph) documenting the flag, e.g.:

Setting the NEMO_RL_PY_EXECUTABLES_SYSTEM=1 environment variable overrides most registry entries to PY_EXECUTABLES.SYSTEM, skipping per-actor uv venv creation when all dependencies are already available in the system Python environment.

[jepio]

tests/unit/distributed/test_ray_actor_environment_registry.py:66

1 action item, optional.

Pre-existing coverage gap (not introduced by this PR): no test guards that the always-PY_EXECUTABLES.SYSTEM actors (DynamoVllmWorker, MathEnvironment, etc.) stay pinned to SYSTEM regardless of the flag. Verified locally in lima (4 passed, 1 warning in 33.90s).

AI-1

Add a parametrized regression test (also needs import pytest added near the top, after import importlib):

Suggested change

	_restore_actor_environment_registry(original_registry)
	finally:
	_restore_actor_environment_registry(original_registry)
	ALWAYS_SYSTEM_EXECUTABLES = [
	"nemo_rl.models.generation.dynamo.dynamo_worker.DynamoVllmWorker",
	"nemo_rl.environments.math_environment.MathEnvironment",
	"nemo_rl.environments.math_environment.MathMultiRewardEnvironment",
	"nemo_rl.environments.vlm_environment.VLMEnvironment",
	"nemo_rl.environments.code_environment.CodeEnvironment",
	"nemo_rl.environments.reward_model_environment.RewardModelEnvironment",
	"nemo_rl.environments.code_jaccard_environment.CodeJaccardEnvironment",
	"nemo_rl.environments.games.sliding_puzzle.SlidingPuzzleEnv",
	"nemo_rl.environments.tools.retriever.RAGEnvironment",
	]
	@pytest.mark.parametrize("system_flag", ["0", "1"])
	def test_always_system_actor_environments_are_unaffected_by_flag(
	monkeypatch, system_flag
	):
	original_registry = registry.ACTOR_ENVIRONMENT_REGISTRY
	try:
	with monkeypatch.context() as context:
	context.setenv("NEMO_RL_PY_EXECUTABLES_SYSTEM", system_flag)
	importlib.reload(registry)
	assert {
	actor: registry.get_actor_python_env(actor)
	for actor in ALWAYS_SYSTEM_EXECUTABLES
	} == dict.fromkeys(
	ALWAYS_SYSTEM_EXECUTABLES, registry.PY_EXECUTABLES.SYSTEM
	)
	finally:
	_restore_actor_environment_registry(original_registry)

[yuki-97]

closing in favor of #4020.