From ba7a7438debd9fc4df5b088f289be9d512682508 Mon Sep 17 00:00:00 2001
From: cjackal <44624812+cjackal@users.noreply.github.com>
Date: Wed, 23 Sep 2026 05:37:09 +0900
Subject: [PATCH 1/2] [Bugfix][Spec Decode] Defer disposable DeepseekV32 MTP
 head for modelopt NVFP4

Signed-off-by: cjackal <44624812+cjackal@users.noreply.github.com>
---
 vllm/models/deepseek_v32/amd/mtp.py    | 7 +++++--
 vllm/models/deepseek_v32/nvidia/mtp.py | 7 +++++--
 2 files changed, 10 insertions(+), 4 deletions(-)

diff --git a/vllm/models/deepseek_v32/amd/mtp.py b/vllm/models/deepseek_v32/amd/mtp.py
index 46c35dec98fa..fdab94b92ac1 100644
--- a/vllm/models/deepseek_v32/amd/mtp.py
+++ b/vllm/models/deepseek_v32/amd/mtp.py
@@ -51,7 +51,6 @@ def __init__(self, vllm_config: VllmConfig, prefix: str) -> None:
         assert vllm_config.speculative_config is not None
         config = vllm_config.speculative_config.draft_model_config.hf_config
         self.config = config
-        quant_config = vllm_config.quant_config
 
         self.enorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
         self.hnorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
@@ -64,7 +63,9 @@ def __init__(self, vllm_config: VllmConfig, prefix: str) -> None:
             device=current_platform.device_type,
         )
         self.shared_head = SharedHead(
-            config=config, prefix=prefix, quant_config=quant_config
+            config=config,
+            prefix=prefix,
+            defer_lm_head=True,
         )
         self.mtp_block = DeepseekV32DecoderLayer(
             vllm_config,
@@ -286,6 +287,8 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
             is_fusion_moe_shared_experts_layer = (
                 self.is_fused_shared_expert_enabled and ("mlp.shared_experts" in name)
             )
+            if ".shared_head.head." in name:
+                continue
             name = self._rewrite_spec_layer_name(spec_layer, name)
 
             if _try_load_fp8_indexer_wk(
diff --git a/vllm/models/deepseek_v32/nvidia/mtp.py b/vllm/models/deepseek_v32/nvidia/mtp.py
index bdc6b95c15e4..c2a6b2e035b3 100644
--- a/vllm/models/deepseek_v32/nvidia/mtp.py
+++ b/vllm/models/deepseek_v32/nvidia/mtp.py
@@ -67,7 +67,6 @@ def __init__(self, vllm_config: VllmConfig, prefix: str) -> None:
         assert vllm_config.speculative_config is not None
         config = vllm_config.speculative_config.draft_model_config.hf_config
         self.config = config
-        quant_config = vllm_config.quant_config
 
         self.enorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
         self.hnorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
@@ -89,7 +88,9 @@ def __init__(self, vllm_config: VllmConfig, prefix: str) -> None:
             get_sparse_mla_index_group_max_rows(vllm_config),
         )
         self.shared_head = SharedHead(
-            config=config, prefix=prefix, quant_config=quant_config
+            config=config,
+            prefix=prefix,
+            defer_lm_head=True,
         )
         self.mtp_block = DeepseekV32DecoderLayer(
             vllm_config,
@@ -392,6 +393,8 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
                     weight_loader(param, loaded_weight)
                     loaded_params.add(name)
                 continue
+            if ".shared_head.head." in name:
+                continue
             name = self._rewrite_spec_layer_name(spec_layer, name)
 
             if _try_load_fp8_indexer_wk(

From b9c44be75893b71aaef349c553524233c94d544a Mon Sep 17 00:00:00 2001
From: cjackal <44624812+cjackal@users.noreply.github.com>
Date: Thu, 1 Oct 2026 01:35:41 +0900
Subject: [PATCH 2/2] add unittest

Co-Authored-By: Mikhail Kostryukov <mike@triptrack.net>
Signed-off-by: cjackal <44624812+cjackal@users.noreply.github.com>
---
 tests/v1/spec_decode/test_mtp.py | 34 ++++++++++++++++++++++++++++++++
 1 file changed, 34 insertions(+)

diff --git a/tests/v1/spec_decode/test_mtp.py b/tests/v1/spec_decode/test_mtp.py
index e3803509e39d..4ed3e1f98f63 100644
--- a/tests/v1/spec_decode/test_mtp.py
+++ b/tests/v1/spec_decode/test_mtp.py
@@ -57,6 +57,40 @@ def test_glm_mtp_defers_lm_head(default_vllm_config):
     assert layer.shared_head.head is None
 
 
+@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):
+    import importlib
+
+    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
+
+
 def _create_mtp_proposer(num_speculative_tokens: int) -> EagleProposer:
     """Create an MTP proposer with unified model configuration."""
     model_config = ModelConfig(