[Bugfix] Guard EPLB VLM unwrap for models without language_model

[esmeetu]

Summary

Forward-fix for the regression reported in #42636 (open as a draft revert of #39805).

In #39805 the EPLB initialization path was extended to unwrap VLM wrappers via moe_candidate.get_language_model() so wrapper models like KimiK25ForConditionalGeneration would register their inner MoE language model with EPLB. That call was made unconditionally for any SupportsMultiModal model, including non-EPLB loads.

For NemotronParseForConditionalGeneration, although it does call _mark_language_model (vllm/model_executor/models/nemotron_parse.py:593), its marked module MBartDecoderNoPos does not expose embed_input_ids. SupportsMultiModal.get_language_model() therefore falls through both the marked-name lookup and the children-fallback in vllm/model_executor/models/interfaces.py:176-211 and raises NotImplementedError, breaking engine-core init for any user loading NemotronParse, even without EPLB.

This PR:

• Gates the MoE-resolution block on self.parallel_config.enable_eplb, so non-EPLB loads never call get_language_model() — restoring pre-[Bugfix] Fix EPLB initialization for VLM wrapper models #39805 behavior for all non-EPLB users.
• Defensively wraps the get_language_model() call in try/except NotImplementedError so EPLB-on configs with a VLM that cannot resolve its language model fall through cleanly instead of crashing during load.

Behavior matrix:

Scenario	Before	After
EPLB off, NemotronParse (regression in #42636)	NotImplementedError	block skipped
EPLB on, KimiK25 (VLM+MoE) — original #39805 target	unwrap → register	unwrap → register
EPLB on, DeepSeek-style (plain MoE)	_moe_model = self.model	unchanged
EPLB on, VLM without resolvable LM	NotImplementedError	caught, _moe_model=None, skipped

Test plan

• Verified the two failing tests cited in Revert "[Bugfix] Fix EPLB initialization for VLM wrapper models" (#39805) #42636 do not use EPLB, so they exercise scenario A (EPLB off → block skipped, no get_language_model() call):
  • tests/models/multimodal/generation/test_nemotron_parse.py::test_models[nvidia/NVIDIA-Nemotron-Parse-v1.2]
  • tests/models/test_initialization.py::test_can_initialize_large_subset[NemotronParseForConditionalGeneration]
• ast.parse syntax check passes; pre-commit hooks pass locally.
• CI re-run on these tests (please add ready label).

AI assistance disclosure

Code change drafted with assistance from Claude (Anthropic). The submitter reviewed every changed line and verified correctness against the four scenarios above.

[claude]

Claude Code Review

This repository is configured for manual code reviews. Comment @claude review to trigger a review and subscribe this PR to future pushes, or @claude review once for a one-time review.

_{Tip: disable this comment in your organization's Code Review settings.}

[gemini-code-assist]

Code Review

This pull request updates the load_model method in GPUModelRunner to conditionally resolve MoE models only when EPLB is enabled. It also introduces error handling for NotImplementedError when calling get_language_model() on multi-modal models, which prevents potential crashes for models like NemotronParse that do not expose a language module. I have no feedback to provide as there were no review comments.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c323462725

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Disable EPLB after MoE unwrap failure

Catching NotImplementedError here sets moe_candidate to None and continues loading, but EPLB remains enabled. In the V1 runner, execute_model always calls eplb_step(), and eplb_step asserts self._moe_model is not None whenever enable_eplb is true, so this path deterministically crashes on the first step for multimodal models whose get_language_model() cannot resolve. Before this change, that configuration failed fast during load; now it fails later with an assertion, so the intended “skip cleanly” behavior is not achieved.

Useful? React with 👍 / 👎.