fix(gym): make the gym venv prefetch dry run pass again

[yfw]

What does this PR do ?

Gym d980441d7 ("fix(cli): fail a dry run when a server's venv setup exits non-zero", NVIDIA-NeMo/Gym#2476) turned previously-swallowed dry-run failures into hard errors. Six servers in the prefetch configs had been failing all along; once that commit arrived via a Gym submodule bump, the rl-gym image build started failing with:

RuntimeError: Dry run failed to set up 6 servers.
  `prefetch_local_vllm_model` / `policy_model_reasoning_off` / `judge_model`
  / `safety_judge_model` / `swe_agents_val` / `swe_agents_train` exited with 1

Two independent causes, both ours:

1. Missing required config fields. Gym's VLLMModelConfig declares return_token_id_information and uses_reasoning_parser without defaults, and LocalVLLMModelConfig additionally requires vllm_serve_env_vars. The thin judge/reasoning-off clients and prefetch_local_vllm_model in prefetch_{super,ultra}_all_envs.yaml omitted them, so model_validate raised. Gym builds each server as setup_env_command(...) && ... python app.py (nemo_gym/cli/env.py:417-420), so the venv was built successfully and only the exit code was wrong -- this cause never damaged a venv. Cause 2 below is the one that leaves an image needing a rebake, since r2egym really was left uninstalled. Other RL configs (e.g. nemotron-3-super/ stage1_rlvr.yaml) already set these; the prefetch configs had simply drifted.

2. UV_LINK_MODE leaking into Gym's setup scripts. docker/Dockerfile exports UV_LINK_MODE=symlink for the gym prefetch RUN, and prefetch_venvs.py forwards the whole os.environ into the Gym Ray actor, so it reached swe_agents' r2e_gym.sh, which runs uv pip install --no-cache. uv rejects that pairing:

   Symlink-based installation is not supported with --no-cache.

   leaving r2egym uninstalled and SWEBenchWrapper asserting. Pin UV_LINK_MODE to hardlink for the Gym actor only, rather than changing the Dockerfile, so the outer prefetch keeps symlink semantics. hardlink is what the Dockerfile already uses for comparable nested builds and, unlike symlink, survives uv cache pruning. It is as space-efficient as copy but not as symlink: across Docker layers, linking out of the lower-layer uv cache makes overlayfs copy the bytes up. There is no cheaper mode available -- symlink is what breaks under --no-cache, and clone still copies up from a lower layer. Note docs/guides/models/nemotron/nemotron-3-super.md invokes this script the same way, which is why the fix belongs in the script rather than the Dockerfile. Upstream, r2e_gym.sh is the only uv --no-cache in Gym while its siblings swebench.sh / swebench_multilingual.sh omit it, so an upstream fix there would retire this pin.

prefetch_omni_envs.yaml needs no change; it defines no thin vllm clients.

Verified the YAML parses and every previously-missing field is present, and added
tests/unit/environments/test_nemo_gym_prefetch_configs.py, which runs Gym's own
model_validate over every standalone server block in examples/nemo_gym/prefetch_*.yaml
so this class of drift fails at PR time rather than in a manual image build. The test
deliberately does not restate the field list, so it tracks the Gym submodule pin.

Limits, kept honest: the test does not cover the UV_LINK_MODE/r2e_gym.sh half at all,
and passing it does not imply the image build succeeds. The build itself is the real test
for the link-mode half.

Issues

List issues that this PR closes (syntax):

Usage

• You can potentially add a usage example below

# Add a code snippet demonstrating how to use this

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]

The fix is correct, minimally scoped, and in the right place. Two inline comments below: one test gap (worth acting on), one comment-wording fix. Nothing blocking.

Both causes verified end-to-end:

• Cause 1 — 3-way check against the real gate (VLLMModelConfig.model_validate): PR head ACCEPTED, fields-deleted REJECTED on exactly ['return_token_id_information', 'uses_reasoning_parser']. Note Gym's parse_no_environment alone does not reject, so the parser isn't the gate — only model_validate is.
• Cause 2 — reproduced on uv 0.11.19: UV_LINK_MODE=symlink + --no-cache → error: Symlink-based installation is not supported with --no-cache.; hardlink installs fine. r2e_gym.sh is genuinely reached in a dry run: SWEBenchWrapper.model_post_init calls all four dataset processors' setup() unconditionally, so dataset_path: /dev/null doesn't matter. The "SWEBenchWrapper asserting" wording is literally accurate — app.py:447 is assert return_code == 0.
• Coverage is complete. Walking Gym's config MRO, the only other default-less fields are host/port/name (launcher-injected) and entrypoint (already present), so these three were exactly the full set — the fix isn't one field short of a second failure. These two prefetch files are also the only configs in the repo that define these servers from scratch; every other apparent gap inherits via _copy: or defaults:. prefetch_omni_envs.yaml correctly needs no change.
• The values can't bake a wrong artifact. They match Gym's own bundled local_vllm_model.yaml exactly (false/false/{}), and the venv is provably independent of them: the generated build command is byte-identical whichever way they're set, and all three thin clients share one venv keyed by server directory.

One wrong clause in the commit message. "so model_validate raised before the venv was built" is reversed — Gym builds the subprocess as setup_env_command(...) && ... python app.py (cli/env.py:417-420), so the venv was built successfully and only the exit code was wrong. The consequence matters more than the wording: cause 2 is why existing prefetched images need rebaking (r2egym really was left uninstalled); cause 1 never damaged a venv. The rest of the message holds, including "other RL configs already set these" — nemotron-3-super/stage1_rlvr.yaml sets all three.

Also checked clean: pre-commit (all hooks), PR title, sign-off, copyright, no paired tests/unit/reference_configs/ twin, and the two YAMLs' added blocks are byte-identical to each other. No performance evidence needed — build fix, no perf claim made.

Generated by Claude Code

[yfw]

/ok to test 0473d40

[yuki-97]

LGTM, just some nits.

[yfw]

/ok to test b212df0