[CI] Update vllm version to 20250922(5aeb925)

.github/workflows/format_pr_body.yaml

@@ -36,7 +36,7 @@ jobs:
 
       - name: Get vLLM version
         run: |
-          VLLM_COMMIT=c60e6137f0bf2034853919b3a9d705d7e06b93cf
+          VLLM_COMMIT=5aeb9254521023f97aca292b3478aa7ff485ffb2
           echo "VLLM_COMMIT=https://github.com/vllm-project/vllm/commit/$VLLM_COMMIT" >> $GITHUB_ENV
 
       - name: Checkout repository

.github/workflows/vllm_ascend_test.yaml

@@ -81,7 +81,7 @@ jobs:
         VLLM_USE_MODELSCOPE: True
     strategy:
       matrix:
-        vllm_version: [c60e6137f0bf2034853919b3a9d705d7e06b93cf, v0.10.2]
+        vllm_version: [5aeb9254521023f97aca292b3478aa7ff485ffb2, v0.10.2]
     steps:
       - name: Install packages
         run: |
@@ -136,7 +136,7 @@ jobs:
     name: e2e-light
     strategy:
       matrix:
-        vllm_version: [c60e6137f0bf2034853919b3a9d705d7e06b93cf, v0.10.2]
+        vllm_version: [5aeb9254521023f97aca292b3478aa7ff485ffb2, v0.10.2]
     # Note (yikun): If CI resource are limited we can split job into two chain jobs
     needs: [lint, changes]
     # only trigger e2e test after lint passed and the change is e2e related with pull request.

.github/workflows/vllm_ascend_test_full.yaml

@@ -68,7 +68,7 @@ jobs:
     name: e2e-full
     strategy:
       matrix:
-        vllm_version: [c60e6137f0bf2034853919b3a9d705d7e06b93cf, v0.10.2]
+        vllm_version: [5aeb9254521023f97aca292b3478aa7ff485ffb2, v0.10.2]
     needs: [changes]
     if: ${{ needs.changes.outputs.e2e_tracker == 'true' }}
     uses: ./.github/workflows/_e2e_test.yaml

docs/source/developer_guide/modeling/adding_a_new_model.md

@@ -61,7 +61,6 @@ from torch import nn
 from vllm.attention import Attention
 from vllm.config import VllmConfig
 from vllm.sequence import IntermediateTensors
-from vllm.model_executor.sampling_metadata import SamplingMetadata
 
 class CustomAttention(nn.Module):
     def __init__(self, vllm_config: VllmConfig, prefix: str):
@@ -123,8 +122,7 @@ class CustomModelForCausalLM(nn.Module):
         ...
 
     def compute_logits(self,
-                      hidden_states: torch.Tensor,
-                      sampling_metadata: SamplingMetadata) -> torch.Tensor:
+                      hidden_states: torch.Tensor) -> torch.Tensor:
         ...
 
     def load_weights(self, 

vllm_ascend/models/deepseek_mtp.py

@@ -36,7 +36,6 @@
     SharedHead)
 from vllm.model_executor.models.deepseek_v2 import DeepseekV2DecoderLayer
 from vllm.model_executor.models.utils import maybe_prefix
-from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.sequence import IntermediateTensors
 
 
@@ -168,14 +167,12 @@ def forward(
     def compute_logits(
         self,
         hidden_states: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
         spec_step_idx: int = 0,
     ) -> torch.Tensor:
         current_step_idx = (spec_step_idx % self.num_mtp_layers)
         mtp_layer = self.layers_list[current_step_idx]
         logits = self.logits_processor(mtp_layer.shared_head.head,
-                                       mtp_layer.shared_head(hidden_states),
-                                       sampling_metadata)
+                                       mtp_layer.shared_head(hidden_states))
         return logits
 
 

vllm_ascend/models/qwen3_next.py

@@ -54,7 +54,6 @@
     AutoWeightsLoader, PPMissingLayer, extract_layer_index,
     is_pp_missing_parameter, make_empty_intermediate_tensors_factory,
     make_layers, maybe_prefix)
-from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors
@@ -1308,10 +1307,8 @@ def get_mamba_state_shape_from_config(
     def compute_logits(
         self,
         hidden_states: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
     ) -> Optional[torch.Tensor]:
-        return self.logits_processor(self.lm_head, hidden_states,
-                                     sampling_metadata)
+        return self.logits_processor(self.lm_head, hidden_states)
 
     def load_weights(self, weights: Iterable[tuple[str,
                                                    torch.Tensor]]) -> set[str]:

vllm_ascend/spec_decode/eagle_proposer.py

@@ -471,7 +471,7 @@ def _propose(
                 hidden_states=self.hidden_states[:num_input_tokens],
             )
         sample_hidden_states = last_hidden_states[last_token_indices]
-        logits = self.model.compute_logits(sample_hidden_states, None)
+        logits = self.model.compute_logits(sample_hidden_states)
         draft_token_ids = logits.argmax(dim=-1)
 
         # Early exit if there is only one draft token to be generated.

vllm_ascend/torchair/models/qwen2.py

@@ -40,7 +40,6 @@
 from vllm.model_executor.models.qwen2 import Qwen2MLP, Qwen2Model
 from vllm.model_executor.models.utils import (AutoWeightsLoader,
                                               PPMissingLayer, maybe_prefix)
-from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.sequence import IntermediateTensors
 
 from vllm_ascend.ascend_config import get_ascend_config
@@ -345,10 +344,8 @@ def forward(
     def compute_logits(
         self,
         hidden_states: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
     ) -> Optional[torch.Tensor]:
-        logits = self.logits_processor(self.lm_head, hidden_states,
-                                       sampling_metadata)
+        logits = self.logits_processor(self.lm_head, hidden_states)
         return logits
 
     def load_weights(self, weights: Iterable[tuple[str,

vllm_ascend/torchair/models/torchair_deepseek_mtp.py

@@ -34,7 +34,6 @@
     DeepSeekMTP, DeepSeekMultiTokenPredictor, DeepSeekMultiTokenPredictorLayer,
     SharedHead)
 from vllm.model_executor.models.utils import maybe_prefix
-from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.sequence import IntermediateTensors
 
 from vllm_ascend.torchair.models.torchair_deepseek_v2 import \
@@ -172,14 +171,12 @@ def forward(
     def compute_logits(
         self,
         hidden_states: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
         spec_step_idx: int = 0,
     ) -> torch.Tensor:
         current_step_idx = (spec_step_idx % self.num_mtp_layers)
         mtp_layer = self.layers_list[current_step_idx]
         logits = self.logits_processor(mtp_layer.shared_head.head,
-                                       mtp_layer.shared_head(hidden_states),
-                                       sampling_metadata)
+                                       mtp_layer.shared_head(hidden_states))
         return logits
 
 

vllm_ascend/torchair/models/torchair_pangu_moe.py

@@ -53,7 +53,6 @@
 from vllm.model_executor.models.utils import (
     extract_layer_index, is_pp_missing_parameter,
     make_empty_intermediate_tensors_factory, make_layers, maybe_prefix)
-from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.model_executor.utils import set_weight_attrs
 from vllm.sequence import IntermediateTensors
 
@@ -937,18 +936,20 @@ def forward(
     def compute_logits(
         self,
         hidden_states: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
+        # For version compatibility, we still keep this argument here.
+        sampling_metadata  # type: ignore
     ) -> Optional[torch.Tensor]:
-        logits = self.logits_processor(self.lm_head, hidden_states,
-                                       sampling_metadata)
-        return logits
+        from vllm_ascend.utils import vllm_version_is
+        if vllm_version_is("0.10.2"):
+            return self.logits_processor(self.lm_head, hidden_states,
+                                         sampling_metadata)
+        return self.logits_processor(self.lm_head, hidden_states)
 
     def sample(
         self,
         logits: Optional[torch.Tensor],
-        sampling_metadata: SamplingMetadata,
     ) -> Optional[SamplerOutput]:
-        next_tokens = self.sampler(logits, sampling_metadata)
+        next_tokens = self.sampler(logits)
         return next_tokens
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):

vllm_ascend/worker/model_runner_v1.py

@@ -1942,7 +1942,7 @@ def execute_model(
                         num_scheduled_tokens_np, finished_sending,
                         finished_recving, kv_connector_output)
                 sample_hidden_states = hidden_states[logits_indices]
-                logits = self.model.compute_logits(sample_hidden_states, None)
+                logits = self.model.compute_logits(sample_hidden_states)
             if broadcast_pp_output:
                 model_output_broadcast_data = {
                     "logits": logits.contiguous(),
@@ -2349,7 +2349,7 @@ def _dummy_run(
 
                 def dummy_compute_logits(hidden_states):
                     return self.model.compute_logits(
-                        hidden_states[dummy_indices], None)
+                        hidden_states[dummy_indices])
 
             with set_ascend_forward_context(
                     attn_metadata,
@@ -2418,7 +2418,7 @@ def profile_run(self) -> None:
                 logit_indices = np.cumsum(num_scheduled_tokens) - 1
                 # TODO: need to rum a dummy sampler for generate task
                 hidden_states = hidden_states[logit_indices]
-                output = self.model.compute_logits(hidden_states, None)
+                output = self.model.compute_logits(hidden_states)
 
         NPUPlatform.synchronize()
         del hidden_states, output
@@ -3241,7 +3241,7 @@ def _get_prompt_logprobs_dict(
             req_idx = self.input_batch.req_id_to_index[req_id]
             offset = self.query_start_loc_np[req_idx].item()
             prompt_hidden_states = hidden_states[offset:offset + num_logits]
-            logits = self.model.compute_logits(prompt_hidden_states, None)
+            logits = self.model.compute_logits(prompt_hidden_states)
 
             # Get the "target" tokens for each index. For prompt at index i,
             # the token at prompt index i+1 is the "sampled" token we want