[Bugfix][Spec Decode] Defer disposable DeepseekV32 MTP head for modelopt nvfp4

[cjackal]

Purpose

#55442 deferred the disposable ParallelLMHead in Glm5Next MTP, but GLM-5.3(glm_moe_dsa) which needs the same fix does not take that path.

SharedHead appends head to the layer prefix (model.layers.N.head), while modelopt checkpoints name it model.layers.N.shared_head.
The ignore entry never matches in ModelOptQuantConfigBase.is_layer_excluded, the head is allocated as NVFP4, and the weight_scale NaN-sentinel check added in #52501 aborts get_model() before the proposer can swap in the target lm_head:

...
(Worker_TP7_DCP7_EP7 pid=1656) ERROR 09-23 00:30:38 [multiproc_executor.py:943] RuntimeError: NVFP4 weight_scale for layer 'parallel_lm_head' was never loaded (still NaN). The checkpoint likely stores this layer as BF16 (not FP4). Fix: pass quant_config=None when constructing this layer, or add it to the quantization ignore list.
(EngineCore pid=606) INFO 09-23 00:30:38 [multiproc_executor.py:472] [shutdown] Executor: waiting for worker exit count=8
...

Repro: Inferact/GLM-5.3-NVFP4 with --speculative-config.method mtp.

Test Plan

Run Inferact/GLM-5.3-NVFP4 with --speculative-config.method mtp and check that model weights are loaded successfully.

Test Result

...
(Worker_TP0_DCP0_EP0 pid=1697)
(Worker_TP0_DCP0_EP0 pid=1697) INFO 09-23 01:55:28 [default_loader.py:430] Loading weights took 232.85 seconds
(Worker_TP0_DCP0_EP0 pid=1697) INFO 09-23 01:55:31 [unquantized.py:498] Using MoEPrepareAndFinalizeNoDPEPMonolithic
...

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[drakosha]

Hit the same failure with Inferact/GLM-5.3-NVFP4 and MTP on 4x H200 (main af7f9488c),
tracked as part 1 of #59306. This is the right place for the fix: since #52861 the GLM DSA draft is
built by DeepseekV32MultiTokenPredictorLayer, and #55442 only covered the Glm5Next path.

Two notes from checking the diff:

1. Head deferral on GLM: validated on our side with an equivalent change (head skipped when the
   target is glm_moe_dsa); the model loads and serves with num_speculative_tokens=3. For GLM
   checkpoints the .shared_head.head. skip in load_weights is a no-op (zai-org/GLM-5.2 and
   the Inferact NVFP4 quant carry only model.layers.78.shared_head.norm.weight for the MTP
   layer), so that run covers this diff as far as GLM goes. deepseek-ai/DeepSeek-V3.2 does ship
   model.layers.61.shared_head.head.weight, so for V3.2 the skip is what keeps loading working;
   I have not run a V3.2 checkpoint against it. ROCm I cannot test.

2. [Bugfix][Model][Spec Decode] Defer disposable GLM MTP head #55442 came with a unit test for the Glm5Next layer. The same test for this layer fails on
   main and passes with this PR, in case you want to add it to tests/v1/spec_decode/test_mtp.py:

@pytest.mark.parametrize(
    "module_name",
    ["vllm.models.deepseek_v32.nvidia.mtp", "vllm.models.deepseek_v32.amd.mtp"],
)
def test_deepseek_v32_mtp_defers_lm_head(default_vllm_config, module_name):
    """The DSA draft (GLM-5.2/5.3 and DeepSeek-V3.2) must not build a head.

    The placeholder is replaced by the target ``lm_head`` after loading; on
    quantized GLM checkpoints, which ship no ``shared_head.head``, building it
    trips the NVFP4 unloaded-scale check first. The draft config reaches the
    layer with ``model_type`` already rewritten to ``deepseek_mtp``.
    """
    from vllm.model_executor.models import deepseek_mtp

    mtp = importlib.import_module(module_name)

    config = mock.MagicMock(
        hidden_size=16,
        rms_norm_eps=1e-5,
        index_topk=8,
        model_type="deepseek_mtp",
    )
    vllm_config = mock.MagicMock()
    vllm_config.speculative_config.draft_model_config.hf_config = config
    vllm_config.speculative_config.num_speculative_tokens = 1
    vllm_config.scheduler_config.max_num_batched_tokens = 4
    vllm_config.scheduler_config.max_num_seqs = 1

    with (
        mock.patch.object(mtp, "DeepseekV32DecoderLayer", return_value=nn.Identity()),
        mock.patch.object(deepseek_mtp, "ParallelLMHead") as parallel_lm_head,
        mock.patch.object(mtp.current_platform, "device_type", "cpu"),
    ):
        layer = mtp.DeepseekV32MultiTokenPredictorLayer(vllm_config, "model.layers.1")

    parallel_lm_head.assert_not_called()
    assert layer.shared_head.head is None

(import importlib at the top of the file.) On main: 2 failed, 1 passed, with ParallelLMHead
built under prefix model.layers.1.head, which is also why the modelopt ignore list never
matches. With this PR: 3 passed. Run inside vllm/vllm-openai:nightly on CPU with
VLLM_TARGET_DEVICE=cpu.

Separate from this PR: the self._eh_plan = build_glm52_plan(...) if config.model_type == "glm_moe_dsa" else None line a few lines up (#49791) tests the draft config, whose model_type
is already deepseek_mtp after SpeculativeConfig.hf_config_override, so the GLM low-latency
plan for eh_proj has never been built. Noted in #59306.

AI assistance was used for this check.