sync v0.15.1 (models)

vllm/model_executor/models/deepseek_mtp.py

@@ -357,13 +357,14 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
                     weight_to_load = loaded_weight
 
                     if is_fusion_moe_shared_experts_layer:
-                        chunk_slice = slice(j * chunk_size, (j + 1) * chunk_size)
-                        if loaded_weight.ndim == 1:
-                            weight_to_load = loaded_weight[chunk_slice]
-                        elif split_dim == 0:
-                            weight_to_load = loaded_weight[chunk_slice, :]
+                        if split_dim == 0:
+                            weight_to_load = loaded_weight[
+                                j * chunk_size : (j + 1) * chunk_size, :
+                            ]
                         else:
-                            weight_to_load = loaded_weight[:, chunk_slice]
+                            weight_to_load = loaded_weight[
+                                :, j * chunk_size : (j + 1) * chunk_size
+                            ]
                         # Synthesize an expert-style name so expert mapping
                         # can route it
                         chunk_name = name.replace(

vllm/model_executor/models/deepseek_ocr.py

@@ -562,7 +562,7 @@ class DeepseekOCRForCausalLM(nn.Module, SupportsMultiModal, SupportsPP, Supports
 
     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,
@@ -596,4 +596,4 @@ class DeepseekOCRForCausalLM(nn.Module, SupportsMultiModal, SupportsPP, Supports
             language_model="language_model",
             connector="projector",
             tower_model=["sam_model", "vision_model"],
-        )
+        )
\ No newline at end of file

vllm/model_executor/models/deepseek_v2.py

@@ -77,7 +77,6 @@ from vllm.model_executor.model_loader.weight_utils import (
 from vllm.model_executor.models.utils import sequence_parallel_chunk
 from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors
-
 from vllm.v1.attention.backend import AttentionBackend
 from vllm.v1.attention.backends.mla.indexer import (
     DeepseekV32IndexerBackend,
@@ -92,7 +91,6 @@ from .utils import (
     make_layers,
     maybe_prefix,
 )
-
 from vllm import _custom_ops as ops
 from vllm.utils import W8a8GetCacheJSON
 
@@ -336,10 +334,7 @@ class DeepseekV2MoE(nn.Module):
             else None,
         )
 
-    def forward(self, hidden_states: torch.Tensor,
-                rms_weight: torch.Tensor | None = None,
-                residual: torch.Tensor | None = None
-                ) -> torch.Tensor:
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
         num_tokens, hidden_dim = hidden_states.shape
         hidden_states = hidden_states.view(-1, hidden_dim)
 
@@ -1106,7 +1101,7 @@ class DeepseekV2Model(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,
@@ -1288,7 +1283,7 @@ class DeepseekV2ForCausalLM(
 
     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/deepseek_vl2.py

@@ -614,7 +614,7 @@ class DeepseekVLV2ForCausalLM(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,
@@ -638,4 +638,4 @@ class DeepseekVLV2ForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         loader = AutoWeightsLoader(self)
         autoloaded_weights = loader.load_weights(weights, mapper=self.hf_to_vllm_mapper)
-        return autoloaded_weights
+        return autoloaded_weights
\ No newline at end of file

vllm/model_executor/models/dots1.py

@@ -394,7 +394,7 @@ class Dots1Model(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,
@@ -538,7 +538,7 @@ class Dots1ForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
 
     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,
@@ -563,4 +563,4 @@ class Dots1ForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
         return loader.load_weights(weights)
 
     def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
-        return self.model.get_expert_mapping()
+        return self.model.get_expert_mapping()
\ No newline at end of file

vllm/model_executor/models/dots_ocr.py

@@ -754,7 +754,7 @@ class DotsOCRForCausalLM(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA
 
     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,
@@ -790,4 +790,4 @@ class DotsOCRForCausalLM(nn.Module, SupportsMultiModal, SupportsPP, SupportsLoRA
             language_model="language_model",
             connector="vision_tower.merger",
             tower_model="vision_tower.",
-        )
+        )
\ No newline at end of file

vllm/model_executor/models/eagle2_5_vl.py

@@ -432,7 +432,7 @@ class Eagle2_5_VLForConditionalGeneration(
 
     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,
@@ -440,6 +440,7 @@ class Eagle2_5_VLForConditionalGeneration(
     ) -> IntermediateTensors:
         """Forward pass through the model."""
         if intermediate_tensors is not None:
+            input_ids = None
             inputs_embeds = None
 
         forward_kwargs = {
@@ -470,4 +471,4 @@ class Eagle2_5_VLForConditionalGeneration(
             language_model="language_model",
             connector="mlp1",
             tower_model="vision_model",
-        )
+        )
\ No newline at end of file

vllm/model_executor/models/ernie45_moe.py

@@ -466,7 +466,7 @@ class Ernie4_5_MoeModel(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,
@@ -546,6 +546,7 @@ class Ernie4_5_MoeModel(nn.Module):
                 # Skip layers on other devices.
                 if is_pp_missing_parameter(name, self):
                     continue
+
                 param = params_dict[name]
                 weight_loader = param.weight_loader
                 weight_loader(param, loaded_weight, shard_id)
@@ -727,7 +728,7 @@ class Ernie4_5_MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA, MixtureOfExpe
 
     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,
@@ -752,4 +753,4 @@ class Ernie4_5_MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA, MixtureOfExpe
         return loader.load_weights(weights)
 
     def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
-        return self.model.get_expert_mapping()
+        return self.model.get_expert_mapping()
\ No newline at end of file

vllm/model_executor/models/ernie45_vl.py

@@ -1650,7 +1650,7 @@ class Ernie4_5_VLMoeForConditionalGeneration(
 
     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,
@@ -1686,4 +1686,4 @@ class Ernie4_5_VLMoeForConditionalGeneration(
 
     def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         loader = AutoWeightsLoader(self)
-        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/ernie45_vl_moe.py

@@ -565,7 +565,7 @@ class Ernie4_5_VLMoeModel(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,
@@ -646,7 +646,7 @@ class Ernie4_5_VLMoeForCausalLM(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,
@@ -800,4 +800,4 @@ class Ernie4_5_VLMoeForCausalLM(nn.Module, SupportsPP):
                     )
                     weight_loader(param, loaded_weight)
             loaded_params.add(name)
-        return loaded_params
+        return loaded_params
\ No newline at end of file

vllm/model_executor/models/ernie_mtp.py

@@ -164,7 +164,7 @@ class ErnieMTP(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor | None,
+        input_ids: torch.Tensor,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
@@ -275,4 +275,4 @@ class ErnieMTP(nn.Module):
         name = name.replace(
             "model.mtp_block.0.", f"model.layers.{layer_idx}.mtp_block."
         )
-        return name
+        return name
\ No newline at end of file

vllm/model_executor/models/exaone.py

@@ -496,7 +496,7 @@ class ExaoneForCausalLM(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,
@@ -521,4 +521,4 @@ class ExaoneForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
             # processed with quantization, LoRA, fine-tuning, etc.
             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/exaone4.py

@@ -490,7 +490,7 @@ class Exaone4ForCausalLM(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,
@@ -515,4 +515,4 @@ class Exaone4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
             # processed with quantization, LoRA, fine-tuning, etc.
             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/exaone_moe.py

@@ -549,7 +549,7 @@ class ExaoneMoeForCausalLM(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,
@@ -576,4 +576,4 @@ class ExaoneMoeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
                 ["lm_head.", "mtp."] if self.config.tie_word_embeddings else ["mtp."]
             ),
         )
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/falcon.py

@@ -423,7 +423,7 @@ class FalconModel(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,

vllm/model_executor/models/falcon_h1.py

@@ -459,7 +459,7 @@ class FalconH1Model(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,
@@ -602,7 +602,7 @@ class FalconH1ForCausalLM(
 
     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,
@@ -678,4 +678,4 @@ class FalconH1ForCausalLM(
 
         if self.tie_word_embeddings:
             loaded_params.add("lm_head.weight")
-        return loaded_params
+        return loaded_params
\ No newline at end of file

vllm/model_executor/models/fuyu.py

@@ -340,7 +340,7 @@ class FuyuForCausalLM(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,
@@ -365,4 +365,4 @@ class FuyuForCausalLM(nn.Module, SupportsMultiModal, 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/gemma.py

@@ -297,7 +297,7 @@ class GemmaModel(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,
@@ -400,7 +400,7 @@ class GemmaForCausalLM(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,
@@ -422,4 +422,4 @@ class GemmaForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
             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/gemma2.py

@@ -410,7 +410,7 @@ class Gemma2ForCausalLM(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,
@@ -432,4 +432,4 @@ class Gemma2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
             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/gemma3.py

@@ -494,7 +494,7 @@ class Gemma3ForCausalLM(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,
@@ -517,4 +517,4 @@ class Gemma3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
             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