add deepseekv3 and llama4

vllm-v0.6.2/vllm/model_executor/models/deepseek_mtp.py

@@ -1,5 +1,4 @@
 """Inference-only DeepSeek V3 Multi-Token Prediction (MTP) model."""
-import re
 from typing import Iterable, List, Optional, Tuple
 
 import torch
@@ -18,7 +17,6 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.sequence import IntermediateTensors
 
 from .deepseek_v2 import DeepseekV2DecoderLayer
-from .utils import maybe_prefix
 
 
 class SharedHead(nn.Module):
@@ -240,6 +238,11 @@ class DeepSeekMTP(nn.Module):
                     and ".layers" not in name):
                 continue
 
+            # Strip "model." prefix since DeepSeekMTP holds
+            # embed_tokens and layers directly (no .model wrapper)
+            if name.startswith("model."):
+                name = name[len("model."):]
+
             self._load_single_weight(
                 name, loaded_weight, stacked_params_mapping,
                 params_dict)

vllm-v0.6.2/vllm/spec_decode/mlu_spec_decode_worker.py

@@ -159,9 +159,11 @@ class MLUSpecDecodeWorker(LoraNotSupportedWorkerBase):
                     draft_worker_kwargs[
                         "model_runner_cls"] = MLUTP1DraftModelRunner
                 else:
-                    if draft_model_config.hf_config.model_type == "eagle":
+                    if draft_model_config.hf_config.model_type in (
+                            "eagle", "deepseek_mtp"):
                         raise NotImplementedError(
-                            "EAGLE does not support TP > 1 yet")
+                            f"{draft_model_config.hf_config.model_type} "
+                            "does not support TP > 1 yet")
 
                     allow_zero_draft_token_step = False
                 proposer_worker = MLUMultiStepWorker(**draft_worker_kwargs)