DRAFT fix: Nano-v3 recipe run fix.

[snowmanwwg]

1. The PR fixes the old NeMo-Gym PARENT_DIR crash in NeMo-RL.

When NeMo-RL starts a NeMo-Gym actor, Python could accidentally grab the wrong folder named nemo_gym:

• Wrong one: examples/nemo_gym
• Correct one: 3rdparty/Gym-workspace/Gym/nemo_gym

The wrong one is just an examples folder, so it does not contain PARENT_DIR. That caused:

ImportError: cannot import name 'PARENT_DIR' from 'nemo_gym'

The PR makes NeMo-RL do this before starting NeMo-Gym:

1. Put the real NeMo-Gym package path first.
2. If Python already loaded the wrong examples/nemo_gym, clear it out.
3. Then import NeMo-Gym normally.

It also fixes a smaller environment bug: NeMo-RL was setting VIRTUAL_ENV to the Python executable path like:

.../bin/python

but it should point to the venv folder:

.../

So the PR makes that correct too.

2. Slurm Ray job cleanup
   After the driver command finished, ray.sub did not signal the Ray sidecars to stop, so the Slurm allocation could remain alive even after training completed.
   The PR touches LOG_DIR/ENDED after a non-empty COMMAND exits and preserves the driver exit code.

3. Context-length overflow handling
   In nemo_rl/models/generation/vllm/vllm_worker_async.py:716, I changed the async chat endpoint so our local ValueError for overlong prompts is converted to HTTP 400, same as vLLM
   validation errors. Also a related Gym PR that fix it in the corresponding gym side fix: handle vllm context length errors in nano v3 recipe Gym#1752

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests?
• Did you run the unit tests and functional tests locally? Visit our Testing Guide for how to run tests
• Did you add or update any necessary documentation? Visit our Document Development Guide for how to write, build and test the docs.

Additional Information

• ...

[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.

[yfw]

vllm_worker_async.py:733

VLLMValidationError is a subclass of ValueError in vLLM 0.20 (vllm/exceptions.py: class VLLMValidationError(ValueError)). Because of that, isinstance(e, ValueError) here is True for every VLLMValidationError, so the guard re-raises any VLLMValidationError whose message lacks "max_model_len"/"maximum context length" — e.g. the sampling-param validation errors raised throughout vllm/sampling_params.py. Before this change (except VLLMValidationError) all of those returned HTTP 400; now they propagate instead, narrowing behavior for the exact type the handler was built for.

Consider gating the substring filter so it only applies to plain ValueError, never to VLLMValidationError:

Suggested change

	raise
	if not isinstance(e, VLLMValidationError) and not (
	"max_model_len" in message or "maximum context length" in message
	):
	raise

[yfw]

Also, do you have examples for when the current behavior is a problem?

[yfw]

nemo_gym.py:42

These fixes currently ship without tests. Two are pure, CPU-only logic and would be cheap to cover:

1. _ensure_nemo_gym_package_precedence — monkeypatch sys.path/sys.modules and point at a temp gym_init; assert the namespace-package shadow is purged, a real package (with __file__/PARENT_DIR) is preserved, a missing submodule is a no-op, and the call is idempotent. Home: tests/unit/environments/test_nemo_gym.py.
2. The vLLM ValueError→400 filter — a context-length ValueError (message containing max_model_len) → 400, a generic ValueError → re-raise, and a VLLMValidationError without those substrings → 400 (this last case guards the regression flagged in the other comment). tests/unit/models/generation/test_vllm_generation.py:240 already stubs VLLMValidationError (note: the stub there subclasses Exception, not ValueError — it should subclass ValueError to mirror the real class).

Since the sys.path/sys.modules manipulation is the subtlest change here, a unit test for (1) would be especially valuable before merge.

[yfw]

Let's separate this one out to another PR. It seems pretty unrelated to the rest

[yfw]

The PR fixes the old NeMo-Gym PARENT_DIR crash in NeMo-RL.

When does this happen?