[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=9607d5eb449711b349d4c2bee0a9c94afcc7ed14
+          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

@@ -42,7 +42,7 @@ jobs:
   lint:
     uses: ./.github/workflows/pre-commit.yml
     with:
-      vllm: 9607d5eb449711b349d4c2bee0a9c94afcc7ed14
+      vllm: 5aeb9254521023f97aca292b3478aa7ff485ffb2
 
   changes:
     runs-on: ubuntu-latest
@@ -83,7 +83,7 @@ jobs:
         VLLM_USE_MODELSCOPE: True
     strategy:
       matrix:
-        vllm_version: [9607d5eb449711b349d4c2bee0a9c94afcc7ed14, v0.10.2]
+        vllm_version: [5aeb9254521023f97aca292b3478aa7ff485ffb2, v0.10.2]
     steps:
       - name: Install packages
         run: |
@@ -138,7 +138,7 @@ jobs:
     name: e2e-light
     strategy:
       matrix:
-        vllm_version: [9607d5eb449711b349d4c2bee0a9c94afcc7ed14, 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: [9607d5eb449711b349d4c2bee0a9c94afcc7ed14, 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):

tests/e2e/model_utils.py

@@ -19,7 +19,12 @@
 
 from typing import Dict, List, Optional, Sequence, Tuple, Union
 
-from vllm.sequence import PromptLogprobs, SampleLogprobs
+from vllm_ascend.utils import vllm_version_is
+
+if vllm_version_is("0.10.2"):
+    from vllm.sequence import PromptLogprobs, SampleLogprobs
+else:
+    from vllm.logprobs import PromptLogprobs, SampleLogprobs
 
 TokensText = Tuple[List[int], str]
 

tests/ut/core/test_schedule_config.py

@@ -27,7 +27,6 @@ def setUp(self):
             max_model_len=8192,
             is_multimodal_model=False,
             send_delta_data=False,
-            scheduler_delay_factor=0,
         )
 
     def test_initialize_from_config_with_default(self):
@@ -90,21 +89,6 @@ def test_not_implemented_send_delta_data(self):
             str(context.exception),
         )
 
-    def test_not_implemented_delay_factor(self):
-        with self.assertRaises(NotImplementedError) as context:
-            AscendSchedulerConfig.initialize_from_config(
-                self.basic_scheduler_config,
-                AscendSchedulerConfig(
-                    delay_factor=1,
-                    max_num_batched_tokens=2048,
-                    max_model_len=2048,
-                ),
-            )
-        self.assertIn(
-            "currently AscendScheduler doesn't support scheduler_delay_factor",
-            str(context.exception),
-        )
-
     def test_no_override(self):
         ascend_config = AscendSchedulerConfig.initialize_from_config(
             self.basic_scheduler_config, {})

tests/ut/models/test_deepseek_mtp.py

@@ -168,8 +168,6 @@ def setup_mtp(self, mocker: MockerFixture):
         mocker.patch(
             "vllm_ascend.models.deepseek_mtp.CustomDeepSeekMultiTokenPredictorLayer.__call__",
             return_value=None)
-        mocker.patch("vllm.model_executor.layers.sampler.get_sampler",
-                     return_value=None)
         mocker.patch(
             "vllm_ascend.ops.vocab_parallel_embedding.AscendVocabParallelEmbedding.__init__",
             return_value=None)

tests/ut/torchair/models/test_torchair_deepseek_mtp.py

@@ -165,8 +165,6 @@ def setup_mtp(self, mocker: MockerFixture):
         mocker.patch(
             "vllm_ascend.torchair.models.torchair_deepseek_mtp.TorchairDeepSeekMultiTokenPredictorLayer.__call__",
             return_value=None)
-        mocker.patch("vllm.model_executor.layers.sampler.get_sampler",
-                     return_value=None)
         mocker.patch(
             "vllm_ascend.ops.vocab_parallel_embedding.AscendVocabParallelEmbedding.__init__",
             return_value=None)

vllm_ascend/core/schedule_config.py

@@ -74,7 +74,7 @@ def __post_init__(self) -> None:
         if self.send_delta_data:
             raise NotImplementedError(
                 "currently AscendScheduler doesn't support send_delta_data.")
-        if self.delay_factor > 0:
+        if getattr(self, "scheduler_delay_factor", 0) > 0:
             raise NotImplementedError(
                 "currently AscendScheduler doesn't support scheduler_delay_factor."
             )

vllm_ascend/models/__init__.py

@@ -25,7 +25,7 @@ def register_model():
 
     ModelRegistry.register_model(
         "DeepseekV3ForCausalLM",
-        "vllm_ascend.models.deepseek_v3:CustomDeepseekV3ForCausalLM")
+        "vllm_ascend.models.deepseek_v2:CustomDeepseekV3ForCausalLM")
 
     ModelRegistry.register_model(
         "DeepSeekMTPModel",

vllm_ascend/models/deepseek_mtp.py

@@ -28,15 +28,13 @@
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization import QuantizationConfig
-from vllm.model_executor.layers.sampler import get_sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.models.deepseek_mtp import (
     DeepSeekMTP, DeepSeekMultiTokenPredictor, DeepSeekMultiTokenPredictorLayer,
     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
 
 
@@ -168,7 +166,7 @@ def forward(
     def compute_logits(
         self,
         hidden_states: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
+        sampling_metadata,  # type: ignore
         spec_step_idx: int = 0,
     ) -> torch.Tensor:
         current_step_idx = (spec_step_idx % self.num_mtp_layers)
@@ -188,8 +186,6 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
                                                        prefix=maybe_prefix(
                                                            prefix, "model"))
 
-        self.sampler = get_sampler()
-
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -204,4 +200,4 @@ def forward(
         hidden_states = self.model(input_ids, positions, kv_caches,
                                    attn_metadata, previous_hidden_states,
                                    inputs_embeds, spec_step_idx)
-        return hidden_states
+        return hidden_states

vllm_ascend/models/deepseek_v2.py

@@ -479,4 +479,8 @@ def load_weights(self, weights: Iterable[tuple[str,
         return loaded_params
 
 
+class CustomDeepseekV3ForCausalLM(CustomDeepseekV2ForCausalLM):
+    pass
+
+
 DeepseekV2DecoderLayer.__init__ = CustomDeepseekV2DecoderLayer.__init__

vllm_ascend/models/deepseek_v3.py

@@ -1,27 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved.
-# Copyright 2023 The vLLM team.
-# Copyright 2023 DeepSeek-AI and the HuggingFace Inc. team. All rights reserved.
-#
-# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
-# and OPT implementations in this library. It has been modified from its
-# original forms to accommodate minor architectural differences compared
-# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from vllm_ascend.models.deepseek_v2 import CustomDeepseekV2ForCausalLM
-
-
-class CustomDeepseekV3ForCausalLM(CustomDeepseekV2ForCausalLM):
-    pass

vllm_ascend/models/qwen3_next.py

@@ -50,7 +50,6 @@
     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
@@ -1079,9 +1078,9 @@ def get_mamba_state_shape_from_config(
             use_v1=True)
 
     def compute_logits(
-        self,
-        hidden_states: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
+            self,
+            hidden_states: torch.Tensor,
+            sampling_metadata,  # type: ignore
     ) -> Optional[torch.Tensor]:
         return self.logits_processor(self.lm_head, hidden_states,
                                      sampling_metadata)

vllm_ascend/ops/vocab_parallel_embedding.py

@@ -253,3 +253,15 @@ def _get_logits_normal(
             logits = logits[..., :self.org_vocab_size]
 
         return logits
+
+    def forward(
+        self,
+        lm_head: VocabParallelEmbedding,
+        hidden_states: torch.Tensor,
+        # keep this for version compatibility
+        sampling_metadata=None,  # type: ignore
+        embedding_bias: Optional[torch.Tensor] = None,
+    ) -> Optional[torch.Tensor]:
+        return super().forward(lm_head,
+                               hidden_states,
+                               embedding_bias=embedding_bias)

vllm_ascend/platform.py

@@ -142,7 +142,7 @@ def check_and_update_config(cls, vllm_config: VllmConfig) -> None:
                 "functionality is currently suboptimal.")
             if not model_config.is_multimodal_model and \
                 structured_outputs_config.backend == "auto" and \
-                not scheduler_config.delay_factor > 0 and \
+                not getattr(scheduler_config, "scheduler_delay_factor", 0) > 0 and \
                 not scheduler_config.send_delta_data and \
                 scheduler_config.policy == "fcfs":
                 ascend_scheduler_config.enabled = True

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
@@ -343,9 +342,9 @@ def forward(
         return hidden_states
 
     def compute_logits(
-        self,
-        hidden_states: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
+            self,
+            hidden_states: torch.Tensor,
+            sampling_metadata,  # type: ignore
     ) -> Optional[torch.Tensor]:
         logits = self.logits_processor(self.lm_head, hidden_states,
                                        sampling_metadata)

vllm_ascend/torchair/models/torchair_deepseek_mtp.py

@@ -27,14 +27,12 @@
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization import QuantizationConfig
-from vllm.model_executor.layers.sampler import get_sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.models.deepseek_mtp import (
     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 \
@@ -172,7 +170,7 @@ def forward(
     def compute_logits(
         self,
         hidden_states: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
+        sampling_metadata,  # type: ignore
         spec_step_idx: int = 0,
     ) -> torch.Tensor:
         current_step_idx = (spec_step_idx % self.num_mtp_layers)
@@ -199,8 +197,6 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         self.model = TorchairDeepSeekMultiTokenPredictor(
             vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model"))
 
-        self.sampler = get_sampler()
-
     def forward(
         self,
         input_ids: torch.Tensor,

vllm_ascend/torchair/models/torchair_deepseek_v2.py

@@ -52,7 +52,6 @@
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.model_executor.layers.sampler import get_sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import (
@@ -945,7 +944,6 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         else:
             self.lm_head = PPMissingLayer()
         self.logits_processor = LogitsProcessor(config.vocab_size)
-        self.sampler = get_sampler()
         self.make_empty_intermediate_tensors = (
             self.model.make_empty_intermediate_tensors)
 

vllm_ascend/torchair/models/torchair_pangu_moe.py

@@ -45,17 +45,16 @@
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.layers.rotary_embedding import get_rope
-from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     ParallelLMHead, VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.models.interfaces import SupportsPP
 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
+from vllm.v1.sample.sampler import Sampler
 
 from vllm_ascend.ascend_config import get_ascend_config
 from vllm_ascend.utils import ACL_FORMAT_FRACTAL_NZ, is_310p
@@ -913,7 +912,7 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         if self.config.tie_word_embeddings:
             self.lm_head.weight = self.model.embed_tokens.weight
         self.logits_processor = LogitsProcessor(config.vocab_size)
-        self.sampler = get_sampler()
+        self.sampler = Sampler()
         self.make_empty_intermediate_tensors = (
             self.model.make_empty_intermediate_tensors)
 
@@ -935,19 +934,19 @@ def forward(
         return hidden_states
 
     def compute_logits(
-        self,
-        hidden_states: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
+            self,
+            hidden_states: torch.Tensor,
+            sampling_metadata,  # type: ignore
     ) -> Optional[torch.Tensor]:
         logits = self.logits_processor(self.lm_head, hidden_states,
                                        sampling_metadata)
         return logits
 
     def sample(
-        self,
-        logits: Optional[torch.Tensor],
-        sampling_metadata: SamplingMetadata,
-    ) -> Optional[SamplerOutput]:
+            self,
+            logits: Optional[torch.Tensor],
+            sampling_metadata,  # type: ignore
+    ):
         next_tokens = self.sampler(logits, sampling_metadata)
         return next_tokens