Merge remote-tracking branch 'origin/v0.15.1-dev' into v0.15.1-dev

# Conflicts:
#	vllm/model_executor/layers/fused_moe/config.py
#	vllm/model_executor/layers/fused_moe/layer.py
#	vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_marlin.py

vllm/model_executor/models/baichuan.py

@@ -209,10 +209,10 @@ class BaiChuanAttention(nn.Module):
                 quant_config=quant_config,
                 prefix=f"{prefix}.attn",
             )
-            self.quant_method = None
-            if quant_config is not None:
-                self.quant_method=quant_config.get_name()
-                self.quant_config=quant_config
+        self.quant_method = None
+        if quant_config is not None:
+            self.quant_method=quant_config.get_name()
+            self.quant_config=quant_config
 
     def forward(
         self,
@@ -334,7 +334,7 @@ class BaiChuanModel(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,
@@ -534,7 +534,7 @@ class BaiChuanBaseForCausalLM(nn.Module, SupportsLoRA, SupportsPP, 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,

vllm/model_executor/models/bailing_moe.py

@@ -440,7 +440,7 @@ class BailingMoeModel(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,
@@ -611,7 +611,7 @@ class BailingMoeForCausalLM(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,
@@ -640,4 +640,4 @@ class BailingMoeForCausalLM(nn.Module, SupportsPP, SupportsLoRA):
 
 
 class BailingMoeV2ForCausalLM(BailingMoeForCausalLM):
-    pass
+    pass
\ No newline at end of file

vllm/model_executor/models/bamba.py

@@ -311,7 +311,7 @@ class BambaModel(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,
@@ -493,7 +493,7 @@ class BambaForCausalLM(
 
     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,
@@ -514,4 +514,4 @@ class BambaForCausalLM(
 
     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/bert_with_rope.py

@@ -475,7 +475,7 @@ class BertWithRope(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,
@@ -726,4 +726,4 @@ class GteNewForSequenceClassification(nn.Module, SupportsCrossEncoding):
             positions=positions,
             inputs_embeds=inputs_embeds,
             intermediate_tensors=intermediate_tensors,
-        )
+        )
\ No newline at end of file

vllm/model_executor/models/blip2.py

@@ -641,7 +641,7 @@ class Blip2ForConditionalGeneration(
 
     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,
@@ -727,4 +727,4 @@ class Blip2ForConditionalGeneration(
             "the number of tokens per image."
         )
         num_images = num_vision_tokens / self._vision_tokens_per_image
-        return num_images * self.config.num_query_tokens
+        return num_images * self.config.num_query_tokens
\ No newline at end of file

vllm/model_executor/models/bloom.py

@@ -294,7 +294,7 @@ class BloomModel(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,
@@ -412,7 +412,7 @@ class BloomForCausalLM(nn.Module, SupportsPP, 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,

vllm/model_executor/models/chameleon.py

@@ -994,7 +994,7 @@ class ChameleonForConditionalGeneration(
 
     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,
@@ -1100,4 +1100,4 @@ class ChameleonForConditionalGeneration(
                 weight_loader = getattr(param, "weight_loader", default_weight_loader)
                 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/chatglm.py

@@ -381,7 +381,7 @@ class ChatGLMModel(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,
@@ -554,7 +554,7 @@ class ChatGLMForCausalLM(ChatGLMBaseModel, SupportsLoRA, SupportsPP, SupportsQua
 
     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/cohere2_vision.py

@@ -446,7 +446,7 @@ class Cohere2VisionForConditionalGeneration(nn.Module, SupportsMultiModal, Suppo
 
     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,
@@ -467,4 +467,4 @@ class Cohere2VisionForConditionalGeneration(nn.Module, SupportsMultiModal, Suppo
         self,
         hidden_states: torch.Tensor,
     ) -> torch.Tensor | None:
-        return self.language_model.compute_logits(hidden_states)
+        return self.language_model.compute_logits(hidden_states)
\ No newline at end of file

vllm/model_executor/models/commandr.py

@@ -312,7 +312,7 @@ class CohereModel(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,
@@ -438,7 +438,7 @@ class CohereForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsQuant):
     @torch.no_grad()
     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,
@@ -466,4 +466,4 @@ class CohereForCausalLM(nn.Module, SupportsLoRA, SupportsPP, SupportsQuant):
         loader = AutoWeightsLoader(
             self, skip_prefixes=["lm_head", "rotary_emb.inv_freq"]
         )
-        return loader.load_weights(weights)
+        return loader.load_weights(weights)
\ No newline at end of file

vllm/model_executor/models/dbrx.py

@@ -361,7 +361,7 @@ class DbrxModel(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,
@@ -462,7 +462,7 @@ class DbrxForCausalLM(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,
@@ -481,4 +481,4 @@ class DbrxForCausalLM(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/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,
@@ -1210,7 +1205,6 @@ class DeepseekV2ForCausalLM(
             os.environ['LLAMA_NN'] = '0'
             os.environ['LM_NN'] = '0'
 
-        self.use_w4a16_moe_sz = os.environ.get('AWQ_MOE_SZ') == '1'
         self.config = config
         self.quant_config = quant_config
 
@@ -1289,7 +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