[Misc] rename torch_dtype to dtype (#26695)

Signed-off-by: wangxiyuan <wangxiyuan1007@gmail.com>

vllm/model_executor/models/chameleon.py

@@ -999,7 +999,7 @@ class ChameleonForConditionalGeneration(
             return []
         assert self.model.vqmodel is not None
         image_tokens = self.model.get_image_tokens(
-            image_input["data"].to(self.config.torch_dtype)
+            image_input["data"].to(self.config.dtype)
         )
         vision_embeddings = self.model.get_input_embeddings(image_tokens)
         return vision_embeddings

vllm/model_executor/models/ernie45_vl.py

@@ -1089,7 +1089,7 @@ class Ernie4_5VLMultiModalProcessor(BaseMultiModalProcessor[Ernie4_5_VLProcessin
         pixel_values = (
             rescale_factor * pixel_values.to(torch.float32) - image_mean_tensor
         ) / image_std_tensor
-        pixel_values = pixel_values.to(hf_config.torch_dtype)
+        pixel_values = pixel_values.to(hf_config.dtype)
         return pixel_values
 
     def _call_hf_processor(

vllm/model_executor/models/glm4v.py

@@ -615,7 +615,7 @@ class GLM4VForCausalLM(
         return None
 
     def _process_image_input(self, image_input: GLMVImagePixelInputs) -> torch.Tensor:
-        pixel_values = image_input["data"].to(dtype=self.config.torch_dtype)
+        pixel_values = image_input["data"].to(dtype=self.config.dtype)
 
         return self.transformer.vision(pixel_values)
 

vllm/model_executor/models/longcat_flash.py

@@ -114,7 +114,7 @@ class FlashConfig(PretrainedConfig):
         attention_dropout=0.0,
         mla_scale_q_lora=False,
         mla_scale_kv_lora=False,
-        torch_dtype="bfloat16",
+        dtype="bfloat16",
         params_dtype="bfloat16",
         router_dtype="float32",
         router_bias=False,
@@ -130,7 +130,7 @@ class FlashConfig(PretrainedConfig):
             bos_token_id=bos_token_id,
             eos_token_id=eos_token_id,
             tie_word_embeddings=tie_word_embeddings,
-            torch_dtype=torch_dtype,
+            dtype=dtype,
             params_dtype=params_dtype,
             router_dtype=router_dtype,
             topk_method=topk_method,

vllm/model_executor/models/nano_nemotron_vl.py

@@ -987,7 +987,7 @@ class NemotronH_Nano_VL_V2(
             prefix=maybe_prefix(prefix, "language_model"),
         )
         self.vision_model = self.get_vit_model_from_radio_config(config).to(
-            self.language_model.config.torch_dtype
+            self.language_model.config.dtype
         )
 
         # Construct the vision projection.
@@ -1008,7 +1008,7 @@ class NemotronH_Nano_VL_V2(
             ReLUSquaredActivation(),
             nn.Linear(vision_projection_hidden_size, llm_hidden_size, bias=False),
         )
-        self.mlp1 = self.mlp1.to(self.language_model.config.torch_dtype)
+        self.mlp1 = self.mlp1.to(self.language_model.config.dtype)
 
         self.config = config
         self.model_config = vllm_config.model_config

vllm/model_executor/models/qwen3_next.py

@@ -338,7 +338,7 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
             group_size=None,
             norm_before_gate=True,
             device=current_platform.current_device(),
-            dtype=config.torch_dtype,
+            dtype=config.dtype,
         )
 
         self.out_proj = RowParallelLinear(
@@ -847,7 +847,7 @@ class Qwen3NextDecoderLayer(nn.Module):
                     1,
                     1,
                     config.hidden_size,
-                    dtype=config.torch_dtype,
+                    dtype=config.dtype,
                 ),
             )
             self.ffn_layer_scale = torch.nn.Parameter(
@@ -855,7 +855,7 @@ class Qwen3NextDecoderLayer(nn.Module):
                     1,
                     1,
                     config.hidden_size,
-                    dtype=config.torch_dtype,
+                    dtype=config.dtype,
                 ),
             )
 

vllm/model_executor/models/transformers.py

@@ -530,7 +530,7 @@ class TransformersBase(nn.Module, SupportsQuant, SupportsLoRA, SupportsPP):
         with init_on_device_without_buffers("meta"):
             self.model: PreTrainedModel = AutoModel.from_config(
                 self.config,
-                torch_dtype=self.model_config.dtype,
+                dtype=self.model_config.dtype,
                 trust_remote_code=self.model_config.trust_remote_code,
             )
 

vllm/model_executor/models/transformers_pooling.py

@@ -157,7 +157,7 @@ class TransformersForSequenceClassification(TransformersPoolingBase):
         with torch.device("meta"):
             seq_cls_model = AutoModelForSequenceClassification.from_config(
                 self.config,
-                torch_dtype=self.model_config.dtype,
+                dtype=self.model_config.dtype,
                 trust_remote_code=self.model_config.trust_remote_code,
             )