feat(models): add Qwen3.8-27B Megatron support

[zpqiu]

What does this PR do?

Adds initial functional support for Qwen/Qwen3.8-27B on the Megatron/MCore backend.

• Adds a 2-node, 8-GPU-per-node GRPO recipe using Megatron TP4 + PP2 + CP2 and vLLM TP4.
• Inherits the Qwen3.5-9B Megatron recipe and overrides only the settings needed for Qwen3.8-27B. The effective rollout batch is 32 prompts x 8 generations = 256 samples, matching the global batch size of 256.
• Adds a 10-step nightly driver with checkpointing enabled and the same metric checks as the Qwen3.5-9B Megatron nightly.
• Documents the Qwen3.8 support status and registers the new model guide.
• AutoModel backend support will be validated in a follow-up PR.

Issues

Progress toward #3675. This PR covers the Megatron/MCore backend; AutoModel backend validation will follow in a separate PR.

Usage

uv run examples/run_grpo.py \
  --config examples/configs/recipes/llm/grpo-qwen3.8-27b-2n8g-megatron-tp4pp2cp2.yaml

Validation

• bash -n tests/test_suites/llm/grpo-qwen3.8-27b-2n8g-megatron-tp4pp2cp2.sh
• TEST_DRYRUN=1 bash tests/test_suites/llm/grpo-qwen3.8-27b-2n8g-megatron-tp4pp2cp2.sh
• python3 tools/config_cli.py minimize-check examples/configs/recipes/llm/grpo-qwen3.8-27b-2n8g-megatron-tp4pp2cp2.yaml
• git diff --check

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 relevant 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

Wandb Link: https://wandb.ai/nvidia/nemorl-pr3678 (including 100-step run and nightly test run).

[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.

[zpqiu]

Review of the Qwen3.8-27B Megatron recipe. Nice addition - the recipe itself is clean, it inherits rather than duplicating the 9B settings, and the guide/model-support.md/toctree wiring is all consistent.

One blocking item: the new nightly puts the suite over the GPU-hour cap, so a unit test that runs in L0 CI fails. Details inline, with the fix that matches how previous nightly-adding PRs handled it.

Verified correct, no action needed:

• Both nightly metric thresholds hold on real 10-step runs (numbers inline).
• NUM_MINUTES=60 matches fleet convention for 2n8g/10-step Megatron nightlies and has ~43% headroom over measured runtime.
• The guide's "context parallelism requires sequence packing" claim is backed by a hard assert at nemo_rl/models/megatron/setup.py#L862. (docs/model-quirks.md#L29 says the opposite, but it's scoped to the FSDP2 section - no conflict.)
• TP4 x PP2 x CP2 = 16 = 2 nodes x 8 GPUs; make_sequence_length_divisible_by: 16 is consistent with TP x CP x 2.
• Recipe naming, CONFIG_PATH derivation via common.env, and nightly.txt accounting all check out - the other 11 tests in test_recipes_and_test_suites.py pass.
• Docs: > [!NOTE] matches the dominant convention (28 files, incl. qwen3-5.md); all added links resolve; the PNG is 344K, small relative to sibling assets.
• No missing model registration - the qwen3_5 special-cases in policy/utils.py and vllm_worker.py are VLM-only, and Qwen3.8-27B is dense text-only.

Generated by Claude Code

[sharonyu-115]

/ok to test 38a9f71