sync v0.15.1 (models)

vllm/model_executor/models/mimo_v2_flash.py

@@ -478,7 +478,7 @@ class MiMoV2Model(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -693,7 +693,7 @@ class MiMoV2FlashForCausalLM(nn.Module, SupportsPP, MixtureOfExperts):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -715,4 +715,4 @@ class MiMoV2FlashForCausalLM(nn.Module, SupportsPP, MixtureOfExperts):
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         loader = AutoWeightsLoader(self)
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/minicpm.py

@@ -440,7 +440,7 @@ class MiniCPMModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -620,7 +620,7 @@ class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -654,4 +654,4 @@ class MiniCPMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsEagle3):
             self,
             skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
         )
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/minicpmv.py

@@ -1147,7 +1147,7 @@ class MiniCPMVBaseModel(nn.Module, SupportsMultiModal, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -1740,4 +1740,4 @@ class MiniCPMV(MiniCPMVBaseModel, SupportsMultiModal, SupportsLoRA):
         # so update values before init is called
         cls.packed_modules_mapping.update(instance_cls.packed_modules_mapping)
         cls.embedding_modules.update(instance_cls.embedding_modules)
-        return instance_cls(vllm_config=vllm_config, prefix=prefix)
+        return instance_cls(vllm_config=vllm_config, prefix=prefix)
\ No newline at end of file

vllm/model_executor/models/minimax_m2.py

@@ -362,7 +362,7 @@ class MiniMaxM2Model(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None,
         inputs_embeds: torch.Tensor | None = None,
@@ -521,7 +521,7 @@ class MiniMaxM2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -555,4 +555,4 @@ def get_spec_layer_idx_from_weight_name(
         for i in range(config.num_mtp_modules):
             if weight_name.startswith(f"model.layers.{layer_idx + i}."):
                 return layer_idx + i
-    return None
+    return None
\ No newline at end of file

vllm/model_executor/models/minimax_text_01.py

@@ -712,7 +712,7 @@ class MiniMaxText01ForCausalLM(nn.Module, HasInnerState, IsHybrid):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -1011,4 +1011,4 @@ class MiniMaxText01ForCausalLM(nn.Module, HasInnerState, IsHybrid):
 
     @classmethod
     def get_mamba_state_copy_func(cls) -> tuple[MambaStateCopyFunc]:
-        return MambaStateCopyFuncCalculator.linear_attention_state_copy_func()
+        return MambaStateCopyFuncCalculator.linear_attention_state_copy_func()
\ No newline at end of file

vllm/model_executor/models/minimax_vl_01.py

@@ -359,7 +359,7 @@ class MiniMaxVL01ForConditionalGeneration(nn.Module, SupportsMultiModal, Support
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -382,4 +382,4 @@ class MiniMaxVL01ForConditionalGeneration(nn.Module, SupportsMultiModal, Support
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         loader = AutoWeightsLoader(self)
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/mistral.py

@@ -156,8 +156,16 @@ class MistralDecoderLayer(LlamaDecoderLayer):
         )
 
         self.layer_idx = int(prefix.split(sep=".")[-1])
+        quant_config = self.get_quant_config(vllm_config)
         config = config or vllm_config.model_config.hf_config
 
+        do_fusion = getattr(
+            quant_config, "enable_quantization_scaling_fusion", False
+        ) and vllm_config.cache_config.cache_dtype.startswith("fp8")
+        if do_fusion:
+            self.input_layernorm.quant_scaling_from = self.self_attn.qkv_proj
+            self.post_attention_layernorm.quant_scaling_from = self.mlp.gate_up_proj
+
         if getattr(config, "ada_rms_norm_t_cond", False):
             self.ada_rms_norm_t_cond = nn.Sequential(
                 ColumnParallelLinear(
@@ -339,4 +347,4 @@ class MistralForCausalLM(LlamaForCausalLM):
             elif item in mapping and mapping[item] not in name:
                 name = name.replace(item, mapping[item])
 
-        return name, loaded_weight
+        return name, loaded_weight
\ No newline at end of file

vllm/model_executor/models/mistral3.py

@@ -539,7 +539,7 @@ class Mistral3ForConditionalGeneration(
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -609,4 +609,4 @@ class Mistral3ForConditionalGeneration(
             language_model="language_model",
             connector="multi_modal_projector",
             tower_model="vision_tower",
-        )
+        )
\ No newline at end of file

vllm/model_executor/models/mixtral.py

@@ -347,7 +347,7 @@ class MixtralModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None,
         inputs_embeds: torch.Tensor | None = None,
@@ -608,7 +608,7 @@ class MixtralForCausalLM(nn.Module, SupportsLoRA, SupportsPP, MixtureOfExperts):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/mllama4.py

@@ -901,7 +901,7 @@ class Llama4ForConditionalGeneration(
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -1161,4 +1161,4 @@ class Llama4ForConditionalGeneration(
             language_model="language_model",
             connector="multi_modal_projector.",
             tower_model="vision_model.",
-        )
+        )
\ No newline at end of file

vllm/model_executor/models/mlp_speculator.py

@@ -3,7 +3,6 @@
 
 import os
 import math
-from typing import Iterable, List, Set, Tuple, Optional
 from collections.abc import Iterable
 
 import torch
@@ -17,7 +16,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
 )
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm import _custom_ops as ops
-from vllm.distributed import tensor_model_parallel_all_gather, tensor_model_parallel_gather
 from vllm import envs
 
 from .utils import maybe_prefix

vllm/model_executor/models/modernbert.py

@@ -54,11 +54,12 @@ class ModernBertEmbeddings(nn.Module):
         input_ids: torch.Tensor,
         inputs_embeds: torch.Tensor | None = None,
     ) -> torch.Tensor:
-        if inputs_embeds is None:
+        if inputs_embeds is not None:
+            return self.norm(inputs_embeds)
+        else:
             inputs_embeds = self.tok_embeddings(input_ids)
-
-        embeddings = self.norm(inputs_embeds)
-        return embeddings
+            embeddings = self.norm(inputs_embeds)
+            return embeddings
 
 
 class ModernBertAttention(nn.Module):
@@ -454,4 +455,4 @@ class ModernBertForTokenClassification(nn.Module):
         )
         hidden_states = self.head(hidden_states)
         hidden_states = hidden_states.to(self.head_dtype)
-        return self.classifier(hidden_states)
+        return self.classifier(hidden_states)
\ No newline at end of file

vllm/model_executor/models/molmo.py

@@ -871,7 +871,7 @@ class MolmoModel(nn.Module, SupportsQuant):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -1591,4 +1591,4 @@ def _get_weights_with_merged_embedding(
         [embedding_weights["embedding"], embedding_weights["new_embedding"]],
         dim=0,
     )
-    yield ("model.embed_tokens.weight", embedding_weights)
+    yield ("model.embed_tokens.weight", embedding_weights)
\ No newline at end of file

vllm/model_executor/models/molmo2.py

@@ -1217,7 +1217,7 @@ class Molmo2TextModel(nn.Module, SupportsQuant):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -2805,4 +2805,4 @@ def _get_weights_with_merged_embedding(
         [embedding_weights["embedding"], embedding_weights["new_embedding"]],
         dim=0,
     )
-    yield ("model.embed_tokens.weight", embedding_weights)
+    yield ("model.embed_tokens.weight", embedding_weights)
\ No newline at end of file

vllm/model_executor/models/mpt.py

@@ -253,7 +253,7 @@ class MPTModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         position_ids: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None,
         inputs_embeds: torch.Tensor | None = None,
@@ -313,7 +313,7 @@ class MPTForCausalLM(nn.Module, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -332,4 +332,4 @@ class MPTForCausalLM(nn.Module, SupportsPP):
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         loader = AutoWeightsLoader(self)
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/nano_nemotron_vl.py

@@ -1917,7 +1917,7 @@ class NemotronH_Nano_VL_V2(
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/nemotron.py

@@ -477,7 +477,7 @@ class NemotronForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -496,4 +496,4 @@ class NemotronForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         loader = AutoWeightsLoader(self)
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/nemotron_h.py

@@ -601,7 +601,7 @@ class NemotronHModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -887,7 +887,7 @@ class NemotronHForCausalLM(
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -908,4 +908,4 @@ class NemotronHForCausalLM(
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         loader = AutoWeightsLoader(self, skip_prefixes=["mtp"])
-        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
+        return loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
\ No newline at end of file

vllm/model_executor/models/nemotron_nas.py

@@ -449,7 +449,7 @@ class DeciLMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, HasNoOps):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -471,4 +471,4 @@ class DeciLMForCausalLM(nn.Module, SupportsLoRA, SupportsPP, HasNoOps):
             self,
             skip_prefixes=(["lm_head."] if self.config.tie_word_embeddings else None),
         )
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/nemotron_parse.py

@@ -289,7 +289,7 @@ class MBartDecoderNoPos(nn.Module):
 
     def forward(
         self,
-        decoder_input_ids: torch.Tensor | None,
+        decoder_input_ids: torch.Tensor,
         *,
         encoder_hidden_states: torch.Tensor | None,
         inputs_embeds: torch.Tensor | None = None,
@@ -897,7 +897,7 @@ class NemotronParseForConditionalGeneration(nn.Module, SupportsMultiModal):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         encoder_outputs: list[torch.Tensor] | None = None,
         **kwargs,
@@ -957,4 +957,4 @@ class NemotronParseForConditionalGeneration(nn.Module, SupportsMultiModal):
         # Load encoder weights
         self.encoder.load_weights(encoder_weights)
         # Load decoder weights
-        self.decoder.load_weights(decoder_weights)
+        self.decoder.load_weights(decoder_weights)
\ No newline at end of file