[KVCache][Bugfix] Fix kv cache initialization error of attention layer (vllm-project#3113)

### What this PR does / why we need it?
Fixes vllm-project#3096 
1. Fix kv cache initialization error of attention layer. There are some
models with layer name like `attn.attn`, instead of `self_attn`, but the
initialization of kv cache tensors only check for `self_attn` and
`attn.attn`, which leding to the error `AssertionError: Some layers are
not correctly initialized`
2. Set the default value of input arg `sampling_metadata` in
`compute_logits` for the modeling files in vllm-ascend. Thus fixing the
error `Qwen3NextForCausalLM.compute_logits() missing 1 required
positional argument: 'sampling_metadata'`

### Does this PR introduce _any_ user-facing change?
N/A

### How was this patch tested?
test locally with internlm


- vLLM version: v0.10.2
- vLLM main:
vllm-project/vllm@5aeb925

---------

Signed-off-by: MengqingCao <cmq0113@163.com>

Author: MengqingCao

vllm_ascend/models/deepseek_mtp.py

@@ -166,7 +166,7 @@ def forward(
     def compute_logits(
         self,
         hidden_states: torch.Tensor,
-        sampling_metadata,  # type: ignore
+        sampling_metadata=None,  # type: ignore
         spec_step_idx: int = 0,
     ) -> torch.Tensor:
         current_step_idx = (spec_step_idx % self.num_mtp_layers)

vllm_ascend/models/qwen3_next.py

@@ -986,7 +986,7 @@ def get_mamba_state_shape_from_config(
     def compute_logits(
             self,
             hidden_states: torch.Tensor,
-            sampling_metadata,  # type: ignore
+            sampling_metadata=None,  # type: ignore
     ) -> Optional[torch.Tensor]:
         return self.logits_processor(self.lm_head, hidden_states,
                                      sampling_metadata)

vllm_ascend/torchair/models/qwen2.py

@@ -344,7 +344,7 @@ def forward(
     def compute_logits(
             self,
             hidden_states: torch.Tensor,
-            sampling_metadata,  # type: ignore
+            sampling_metadata=None,  # type: ignore
     ) -> Optional[torch.Tensor]:
         logits = self.logits_processor(self.lm_head, hidden_states,
                                        sampling_metadata)

vllm_ascend/torchair/models/torchair_deepseek_mtp.py

@@ -170,7 +170,7 @@ def forward(
     def compute_logits(
         self,
         hidden_states: torch.Tensor,
-        sampling_metadata,  # type: ignore
+        sampling_metadata=None,  # type: ignore
         spec_step_idx: int = 0,
     ) -> torch.Tensor:
         current_step_idx = (spec_step_idx % self.num_mtp_layers)

vllm_ascend/torchair/models/torchair_pangu_moe.py

@@ -936,7 +936,7 @@ def forward(
     def compute_logits(
             self,
             hidden_states: torch.Tensor,
-            sampling_metadata,  # type: ignore
+            sampling_metadata=None,  # type: ignore
     ) -> Optional[torch.Tensor]:
         logits = self.logits_processor(self.lm_head, hidden_states,
                                        sampling_metadata)

vllm_ascend/worker/model_runner_v1.py

@@ -2784,9 +2784,10 @@ def initialize_kv_cache_tensors(
             for idx in range(len(kv_cache_tensor.shared_by)):
                 layer_name = kv_cache_tensor.shared_by[idx]
                 if "linear_attn" in layer_name:
+                    # for mamba linear attention
                     for layer_name_inner in kv_cache_tensor.shared_by:
-                        if "self_attn" in layer_name_inner or layer_name_inner in kv_cache_raw_tensors.keys(
-                        ):
+                        if ("attn" in layer_name_inner and "linear_attn" not in layer_name_inner) or \
+                            layer_name_inner in kv_cache_raw_tensors.keys():
                             continue
                         if self.vllm_config.kv_transfer_config is None:
                             tensor = torch.zeros(kv_cache_tensor.size,
@@ -2800,7 +2801,8 @@ def initialize_kv_cache_tensors(
                             tensor = self._align_memory(
                                 tensor, alignment)[:kv_cache_tensor.size]
                         kv_cache_raw_tensors[layer_name_inner] = tensor
-                elif "self_attn" in layer_name:
+                elif "attn" in layer_name:
+                    # for other attentions, e.g., self_attn, sliding window attn
                     if self.vllm_config.kv_transfer_config is None:
                         k_tensor = torch.zeros(kv_cache_tensor.size // 2,
                                                dtype=torch.int8,