fix(vlm): propagate NeMo Gym dynamic image padding

[youngeunkwon0405]

What does this PR do?

Propagates the existing NeMo-Gym pad_dynamic_image_shapes option through the initial deduplicated multimodal attachment path.

When grpo.deduplicate_multimodal_data=true, GRPO attaches the initial image-model inputs before repeating prompts. That path previously called attach_image_model_inputs_to_message without forwarding the environment's dynamic-padding option. A heterogeneous multi-image example could therefore attempt to stack different image resolutions before rollout even when the NeMo-Gym environment requested padding.

Before this fix, propagation stopped at the algorithm-to-rollout-helper boundary:

master_config.env.nemo_gym.pad_dynamic_image_shapes
  -> grpo_train / validate / AsyncTrajectoryCollector._process_batch
  -> attach_initial_nemo_gym_image_payloads(batch, processor)
  -> attach_image_model_inputs_to_message(...)
       pad_dynamic_image_shapes=False  # implicit default

The synchronous training, validation, and asynchronous collection call sites invoked attach_initial_nemo_gym_image_payloads without the flag. That helper then invoked attach_image_model_inputs_to_message without the flag as well, so the target function silently used its existing False default. This early deduplicated attachment shortcut is separate from the normal NeMo-Gym environment preprocessing path, which already carried the option correctly.

After this fix, the complete chain is:

master_config.env.nemo_gym.pad_dynamic_image_shapes
  -> grpo_train / validate / AsyncTrajectoryCollector._process_batch
  -> attach_initial_nemo_gym_image_payloads(..., pad_dynamic_image_shapes=...)
  -> attach_image_model_inputs_to_message(..., pad_dynamic_image_shapes=...)

This change:

• adds a keyword-only pad_dynamic_image_shapes argument to attach_initial_nemo_gym_image_payloads;
• forwards env.nemo_gym.pad_dynamic_image_shapes from synchronous training, validation, and asynchronous trajectory collection; and
• preserves the existing False default for every caller that does not opt in.

This PR is independent of the 32-GPU recipe PR and independent of #3721.

Issues

None.

Validation

• uvx --from ruff==0.9.9 ruff check nemo_rl/algorithms/async_utils/trajectory_collector.py nemo_rl/algorithms/grpo.py nemo_rl/experience/rollouts.py tests/unit/experience/test_rollouts.py
• uvx --from ruff==0.9.9 ruff format --check nemo_rl/algorithms/async_utils/trajectory_collector.py nemo_rl/algorithms/grpo.py nemo_rl/experience/rollouts.py tests/unit/experience/test_rollouts.py
• PYTHONPATH=. uv run --extra nemo_gym --group test pytest -q --noconftest tests/unit/experience/test_rollouts.py -k attach_initial_nemo_gym_image_payloads — 3 passed
• GPU Slurm validation job 6370288 — COMPLETED 0:0
• Ten-step private NeMo-Gym sync and async runs completed without mixed-image padding failures: sync W&B, async W&B

Before your PR is "Ready for review"

• Read and followed the contributor guidelines
• Added focused unit coverage for flag propagation and the default behavior
• Ran focused Linux GPU-container tests
• No documentation update is needed because this fixes existing internal option propagation without adding a new public option

[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.

[youngeunkwon0405]

This PR was reviewed by a coordinated team of 4 specialized agents (RL-codebase expert, test reviewer, independent bug scanner, and a devil's-advocate verification pass). Summary:

• The fix is correct and needed. Pre-PR, the dedup attachment path always invoked the processor with stacking semantics, breaking heterogeneous multi-image gym prompts even when the environment requested padding; the gym-actor side already honored the flag. All three production call sites are updated and no others exist.
• No bugs found. Config key path, master_config.env typing at all three call sites, and the downstream attach_image_model_inputs_to_message keyword signature were all verified against the code. Ruff check/format pass. The updated unit test was run locally and passes.
• The 5 inline comments are all low-severity polish (config-read convention + shared helper, two optional verified test hardenings, a docstring nit) plus one pre-existing out-of-scope observation. Nothing blocks merge.

Validation evidence (W&B sync/async runs + Slurm job) is appropriate for a correctness fix — thanks for including it.

Generated by Claude Code

[youngeunkwon0405]

/ok to test 36ed044

[youngeunkwon0405]

/ok to test ac228c1