Merge tag 'v0.15.0rc1' into v0.15.0rc1-ori

vllm/model_executor/models/deepseek_mtp.py

@@ -213,7 +213,7 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
@@ -316,7 +316,11 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
                     # Determine split axis based on op type
                     # gate/up: ColumnParallel → split along dim 0
                     # down: RowParallel → split along dim 1
-                    split_dim = 1 if "down_proj.weight" in name else 0
+                    split_dim = (
+                        1
+                        if ("down_proj.weight" in name and loaded_weight.ndim > 1)
+                        else 0
+                    )
                     total = loaded_weight.shape[split_dim]
                     assert total % num_chunks == 0, (
                         f"Shared expert weight dim {total} "
@@ -329,14 +333,13 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
                     weight_to_load = loaded_weight
 
                     if is_fusion_moe_shared_experts_layer:
-                        if split_dim == 0:
-                            weight_to_load = loaded_weight[
-                                j * chunk_size : (j + 1) * chunk_size, :
-                            ]
+                        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, :]
                         else:
-                            weight_to_load = loaded_weight[
-                                :, j * chunk_size : (j + 1) * chunk_size
-                            ]
+                            weight_to_load = loaded_weight[:, chunk_slice]
                         # 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,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/deepseek_v2.py

@@ -1087,7 +1087,7 @@ class DeepseekV2Model(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None,
         inputs_embeds: torch.Tensor | None = None,
@@ -1260,7 +1260,7 @@ class DeepseekV2ForCausalLM(
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         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,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/dots1.py

@@ -394,7 +394,7 @@ class Dots1Model(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         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,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

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,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/eagle2_5_vl.py

@@ -432,7 +432,7 @@ class Eagle2_5_VLForConditionalGeneration(
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -440,7 +440,6 @@ class Eagle2_5_VLForConditionalGeneration(
     ) -> IntermediateTensors:
         """Forward pass through the model."""
         if intermediate_tensors is not None:
-            input_ids = None
             inputs_embeds = None
 
         forward_kwargs = {

vllm/model_executor/models/ernie45_moe.py

@@ -466,7 +466,7 @@ class Ernie4_5_MoeModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,
@@ -728,7 +728,7 @@ class Ernie4_5_MoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA, MixtureOfExpe
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/ernie45_vl.py

@@ -1650,7 +1650,7 @@ class Ernie4_5_VLMoeForConditionalGeneration(
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

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,
+        input_ids: torch.Tensor | None,
         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,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/ernie_mtp.py

@@ -164,7 +164,7 @@ class ErnieMTP(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         hidden_states: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,

vllm/model_executor/models/exaone.py

@@ -496,7 +496,7 @@ class ExaoneForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/exaone4.py

@@ -490,7 +490,7 @@ class Exaone4ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/exaone_moe.py

@@ -549,7 +549,7 @@ class ExaoneMoeForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/falcon.py

@@ -402,7 +402,7 @@ class FalconModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         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,
+        input_ids: torch.Tensor | None,
         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,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/fuyu.py

@@ -340,7 +340,7 @@ class FuyuForCausalLM(nn.Module, SupportsMultiModal, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/gemma.py

@@ -297,7 +297,7 @@ class GemmaModel(nn.Module):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         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,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/gemma2.py

@@ -410,7 +410,7 @@ class Gemma2ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,

vllm/model_executor/models/gemma3.py

@@ -494,7 +494,7 @@ class Gemma3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP):
 
     def forward(
         self,
-        input_ids: torch.Tensor,
+        input_ids: torch.Tensor | None,
         positions: torch.Tensor,
         intermediate_tensors: IntermediateTensors | None = None,
         inputs_embeds: torch.Tensor | None = None,