fix(recipe): pin expert tensor parallelism on Qwen3.5-VL MoE pretrain recipes

[pruprakash]

What does this PR do ?

Fixes three Qwen3.5-VL MoE pretrain recipes raising RuntimeError at initialize_model_parallel, by pinning the expert tensor parallel size they left unset.

Changelog

• src/megatron/bridge/recipes/qwen_vl/h100/qwen35_vl.py: set expert_tensor_parallel_size = 1 on the 35B-A3B/8gpu, 122B-A10B/128gpu and 397B-A17B/512gpu pretrain mock recipes.
• tests/unit_tests/recipes/qwen_vl/test_qwen35_vl_recipes.py: 1 parametrized test (3 cases) asserting the dense and expert grids divide each recipe's named world size, built through a provider fake that leaves ETP unset.

GitHub Actions CI

See the CI section in the Contributing doc for how to trigger the CI. A Nvidia developer will need to approve and trigger the CI for external contributors.

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 add or update any necessary documentation?
• Does the PR affect components that are optional to install? (Ex: Numba, Pynini, Apex etc)
  • Reviewer: Does the PR have correct import guards for all optional libraries?

Additional Information

• Root cause: ETP unset, so Megatron-Core resolves it to the dense TP (model_parallel_config.py:506, parallel_state.py:871) and the expert grid needs 4x the ranks each recipe name advertises.
• Blast radius: tutorials/data/datacomp/train_qwen36_example.sh and both qwen3.6-35b-a3b card leaves launch the 35B-A3B recipe on 8 GPUs with no overrides; all fail at init.
• Regression? No, latent since 4d7ce790 (refactor(recipes): add h100 recipe namespace #4622). Masked in unit tests because the shared _FakeModelCfg pre-sets ETP=1.
• Verification: 8xH100, nvcr.io/nvidian/nemo:nightly. initialize_model_parallel at the recipe topology: ETP unset gives world_size (8) is not divisible by expert_tensor_model_pipeline_parallel size (32); ETP=1 gives OK (dp=1 etp=1 ep=4). Unit red-green 3 failed to 3 passed; pytest tests/unit_tests/recipes/qwen_vl/ 225 passed.
• Not included: the card leaves' GBS 512 vs recipe GBS 32 mismatch is a separate card-side defect.
• Related to # (issue)

[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.

[claude]

Light review — LGTM

The fix pins expert_tensor_parallel_size = 1 on the three Qwen3.5-VL MoE pretrain mock recipes. Verified that with the fix every expert grid (ETPEPPP) divides the named world size:

• 35b_a3b_8gpu: dense 412=8 (ok), expert 142=8 (ok)
• 122b_a10b_128gpu: dense 428=64 -> 128 (ok), expert 188=64 -> 128 (ok)
• 397b_a17b_512gpu: dense 4216=128 -> 512 (ok), expert 11616=256 -> 512 (ok)

Without the fix an unset value resolves to TP=4, yielding expert grids of 32 / 256 / 1024 that do not divide the world size, so the change is both correct and necessary. The new parametrized unit test (test_qwen35_vl_moe_pretrain_mock_grids_divide_their_named_world_size) is the right GPU-free coverage and correctly simulates the unset case via _FakeModelCfgWithoutExpertTensorParallel.

Related (out of scope, follow-up): the sibling qwen3_vl MoE mock configs — qwen3_vl_30b_a3b_pretrain_8gpu_h100_bf16_mock_config and qwen3_vl_235b_a22b_pretrain_256gpu_h100_bf16_mock_config — set expert_model_parallel_size but omit expert_tensor_parallel_size, the identical latent issue. Consider extending the same fix and test to them.

Suggested test cases

• No perf tests impacted. (No scripts/performance/configs/ files changed.)
• Relevant unit coverage is included: tests/unit_tests/recipes/qwen_vl/test_qwen35_vl_recipes.py::test_qwen35_vl_moe_pretrain_mock_grids_divide_their_named_world_size (parametrized over the 8/128/512-GPU recipes).

[pruprakash]

/ok to test 1403572

[qiaochuz-nv]

Retrospective QA-fix tracking issue: #6394. It records the fix authored by @pruprakash and delivered by this merged PR.