[inference] fix: Support MCore dev inference mode

[yaoyu-33]

Summary

Fixes the current MCore dev bump signal in #3871.

Target: dev
Failure classification: Bridge-side compatibility issue across MCore variants.

Root cause:

• main/mcore-dev no longer exports InferenceMode from megatron.core.inference.utils, while Bridge VLM inference imported it directly.
• The unit/import check also flagged missing cache_position entries in the Qwen3ASR thinker forward docstrings.

Fix:

• Add a narrow Bridge VLM compatibility shim that uses MCore InferenceMode when present and falls back to no-op methods when the dev commit does not expose it.
• Route VLM engine code and its unit test through that shim.
• Document cache_position in the two Qwen3ASR forward docstrings flagged by the import check.

Guards:

• Added one MCore compatibility guard in src/megatron/bridge/inference/vlm/_mcore_compat.py.
• Removal TODO: remove when Megatron-Core dev exposes InferenceMode from megatron.core.inference.utils.

Validation

2026-05-18 PDT:

• CW interactive job 11877695: uv run python -m pytest tests/unit_tests/inference/vlm/test_vlm_engine.py tests/unit_tests/inference/vlm/test_base.py -q -> 12 passed.
• CW interactive job 11878377: uv run pre-commit run --all-files -> passed.
• Local: uvx pre-commit run --files src/megatron/bridge/inference/vlm/_mcore_compat.py src/megatron/bridge/inference/vlm/vlm_engine.py tests/unit_tests/inference/vlm/test_vlm_engine.py src/megatron/bridge/models/qwen3_asr/hf_qwen3_asr/modeling_qwen3_asr.py -> passed.
• Local: uvx ruff check ..., uvx ruff format --check ..., and git diff --check -> passed.

Local uv run python -m pytest ... in this workstation checkout was blocked before pytest by causal-conv1d==1.6.2.post1 build isolation requiring torch during wheel build, so the authoritative pytest/pre-commit runs above were done on CW interactive with the CI container.

[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.

[yaoyu-33]

/ok to test c2a1858

[claude]

Light Code Review

LGTM — clean, narrowly scoped fix for the MCore dev InferenceMode import breakage.

A couple of minor observations (non-blocking):

1. _mcore_compat.py error guard: The InferenceMode not in str(exc) check on L20 correctly re-raises unrelated ImportErrors. However, if megatron.core.inference.utils itself fails to import due to a transitive dependency issue whose error message happens to contain the string InferenceMode, the exception would be silently swallowed. This is an unlikely edge case and the current approach is pragmatic, but worth noting for the TODO removal later.

2. Docstring additions in modeling_qwen3_asr.py: The cache_position docstrings at L996-998 and L1179-1180 are correct and match the HF convention.

Suggested test cases

No perf tests impacted.

[yaoyu-33]

/ok to test 4daaa99

[yaoyu-33]

/ok to test 22d897e

[yaoyu-33]

/ok to test 34ab3ea

[yaoyu-33]

CI triage: unrelated/main, not PR-caused.

Evidence:

• PR head 34ab3ea01628942e8c2dd321e3aeaf23187b4422 failed L1_Launch_recipes_qwen_vl and gb200_L1_Launch_recipes_qwen_vl in tests/functional_tests/test_groups/recipes/test_qwen35_vl_recipes_finetune.py::TestQwen35VLFinetuneRecipes::test_sft_nothing_frozen with ValueError: GDN: cu_seqlens_q[-1]=563 does not match total_sequence_length=564.
  • H100 job: https://github.com/NVIDIA-NeMo/Megatron-Bridge/actions/runs/26122507905/job/76838800634
  • GB200 job: https://github.com/NVIDIA-NeMo/Megatron-Bridge/actions/runs/26122507905/job/76838800322
• The same failure appears on main after an unrelated push, run 26124138716 at SHA 2e3c3409f11465f98143a2e6da54889f8193b096, job gb200_L1_Launch_recipes_qwen_vl: https://github.com/NVIDIA-NeMo/Megatron-Bridge/actions/runs/26124138716/job/76843210427. It fails the same Qwen3.5-VL test with the same one-token GDN cu_seqlens mismatch.
• The PR changed files do not include the failing recipe/model path: no changes to src/megatron/bridge/recipes/qwen_vl/qwen35_vl.py, src/megatron/bridge/training/vlm_step.py, Qwen-VL model code, or 3rdparty/Megatron-LM/megatron/core/ssm/gated_delta_net.py.
• The failing surface matches the Qwen3.5-VL THD packing default. This PR head still has cfg.dataset.pack_sequences_in_batch = True from the older main state, while current main reverted that default in f8c61132 / Revert "[recipe] feat: enable THD packing by default for Qwen3.5-VL finetune" #3894 to False.

No patch from this PR is needed for the failure. The branch likely needs to be updated onto current main to pick up #3894 before rerunning CI. I did log-only triage and did not run local unit tests.