[Model] Remove incorrect SupportsPP from MTP models (#32150)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/model_executor/models/deepseek_mtp.py

@@ -32,7 +32,6 @@ from .deepseek_v2 import (
     DeepseekV2MoE,
     get_spec_layer_idx_from_weight_name,
 )
-from .interfaces import SupportsPP
 from .utils import maybe_prefix
 
 logger = init_logger(__name__)
@@ -181,7 +180,7 @@ class DeepSeekMultiTokenPredictor(nn.Module):
 
 
 @support_torch_compile
-class DeepSeekMTP(nn.Module, SupportsPP, DeepseekV2MixtureOfExperts):
+class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
         self.config = vllm_config.model_config.hf_config

vllm/model_executor/models/ernie_mtp.py

@@ -39,7 +39,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.sequence import IntermediateTensors
 
-from .interfaces import SupportsPP
 from .llama import LlamaDecoderLayer
 from .utils import is_pp_missing_parameter, maybe_prefix
 
@@ -143,7 +142,7 @@ class ErnieMultiTokenPredictor(nn.Module):
         return logits
 
 
-class ErnieMTP(nn.Module, SupportsPP):
+class ErnieMTP(nn.Module):
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
 

vllm/model_executor/models/glm4_moe_mtp.py

@@ -47,7 +47,6 @@ from .glm4_moe import (
     Glm4MoeDecoderLayer,
     get_spec_layer_idx_from_weight_name,
 )
-from .interfaces import SupportsPP
 from .utils import maybe_prefix
 
 
@@ -184,7 +183,7 @@ class Glm4MoeMultiTokenPredictor(nn.Module):
         return logits
 
 
-class Glm4MoeMTP(nn.Module, SupportsPP, Glm4MixtureOfExperts):
+class Glm4MoeMTP(nn.Module, Glm4MixtureOfExperts):
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
         self.config = vllm_config.model_config.hf_config

vllm/model_executor/models/longcat_flash_mtp.py

@@ -24,7 +24,6 @@ from vllm.model_executor.models.longcat_flash import FlashConfig
 from vllm.sequence import IntermediateTensors
 
 from .deepseek_v2 import DeepseekV2DecoderLayer
-from .interfaces import SupportsPP
 from .utils import maybe_prefix
 
 
@@ -124,7 +123,7 @@ class LongCatMultiTokenPredictor(nn.Module):
         )
 
 
-class LongCatFlashMTP(nn.Module, SupportsPP):
+class LongCatFlashMTP(nn.Module):
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
         # LongCat MTP without MoE layers

vllm/model_executor/models/openpangu_mtp.py

@@ -43,7 +43,6 @@ from vllm.model_executor.models.deepseek_mtp import (
 from vllm.model_executor.models.utils import maybe_prefix
 from vllm.sequence import IntermediateTensors
 
-from .interfaces import SupportsPP
 from .openpangu import OpenPanguDecoderLayer
 
 
@@ -92,7 +91,7 @@ class OpenPanguMultiTokenPredictor(DeepSeekMultiTokenPredictor):
 
 
 @support_torch_compile
-class OpenPanguMTP(nn.Module, SupportsPP):
+class OpenPanguMTP(nn.Module):
     def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
         self.config = vllm_config.model_config.hf_config

vllm/model_executor/models/qwen3_next_mtp.py

@@ -27,7 +27,6 @@ from vllm.model_executor.models.qwen3_next import (
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.configs import Qwen3NextConfig
 
-from .interfaces import SupportsPP
 from .utils import (
     AutoWeightsLoader,
     is_pp_missing_parameter,
@@ -221,7 +220,7 @@ class Qwen3NextMultiTokenPredictor(nn.Module):
 
 
 @support_torch_compile
-class Qwen3NextMTP(nn.Module, SupportsPP, QwenNextMixtureOfExperts):
+class Qwen3NextMTP(nn.Module, QwenNextMixtureOfExperts):
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",
@@ -253,9 +252,6 @@ class Qwen3NextMTP(nn.Module, SupportsPP, QwenNextMixtureOfExperts):
             prefix=maybe_prefix(prefix, "lm_head"),
         )
         self.logits_processor = LogitsProcessor(config.vocab_size)
-        self.make_empty_intermediate_tensors = (
-            self.model.make_empty_intermediate_tensors
-        )
         self.set_moe_parameters()
 
     def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: