[Main2Main] Upgrade vllm commit to 0109 (#5752)

### What this PR does / why we need it?
Upgrade vllm commit to 0109 (bde38c11df0ea066a740efe9b77fff5418be45df)

1. remove `init_cached_hf_modules ` due to
https://github.com/vllm-project/vllm/pull/31786
2. fix spec_decode e2e test due to
https://github.com/vllm-project/vllm/pull/29821 break
3. fix `vllm.v1.attention.backends.utils` duo to
https://github.com/vllm-project/vllm/pull/31891
4. fix `self.seq_lens - query_lens` on same device due to
https://github.com/vllm-project/vllm/pull/31773
5. skip model_runner_v2 e2e test due to `'_OpNamespace' '_C' object has
no attribute 'get_cuda_view_from_cpu_tensor'`

- vLLM version: v0.13.0
- vLLM main:
2f4e6548ef

Signed-off-by: hfadzxy <starmoon_zhang@163.com>

.github/workflows/_e2e_test.yaml

@@ -119,7 +119,7 @@ jobs:
           pytest -sv --durations=0 tests/e2e/singlecard/compile/test_norm_quant_fusion.py
   
           # model_runner_v2
-          pytest -sv --durations=0 tests/e2e/singlecard/model_runner_v2/test_basic.py
+          # pytest -sv --durations=0 tests/e2e/singlecard/model_runner_v2/test_basic.py
 
           # pooling
           pytest -sv --durations=0 tests/e2e/singlecard/pooling/test_classification.py
@@ -309,7 +309,7 @@ jobs:
         run: |
           pytest -sv --durations=0 tests/e2e/multicard/4-cards/test_data_parallel_tp2.py
           pytest -sv --durations=0 tests/e2e/multicard/4-cards/test_kimi_k2.py
-          pytest -sv --durations=0 tests/e2e/multicard/4-cards/test_qwen3_next.py
+          pytest -sv --durations=0 tests/e2e/multicard/4-cards/test_qwen3_next.py 
 
           # long_sequence
           pytest -sv --durations=0 tests/e2e/multicard/4-cards/long_sequence/test_accuracy.py

.github/workflows/bot_pr_create.yaml

@@ -37,7 +37,7 @@ jobs:
     steps:
       - name: Get vLLM version
         run: |
-          VLLM_COMMIT=2f4e6548efec402b913ffddc8726230d9311948d
+          VLLM_COMMIT=bde38c11df0ea066a740efe9b77fff5418be45df
           echo "VLLM_COMMIT=https://github.com/vllm-project/vllm/commit/$VLLM_COMMIT" >> $GITHUB_ENV
 
       - name: Checkout repository

.github/workflows/pr_test_full.yaml

@@ -75,7 +75,7 @@ jobs:
     name: e2e-full
     strategy:
       matrix:
-        vllm_version: [2f4e6548efec402b913ffddc8726230d9311948d, v0.13.0]
+        vllm_version: [bde38c11df0ea066a740efe9b77fff5418be45df, v0.13.0]
     needs: [changes]
     if: ${{ needs.changes.outputs.e2e_tracker == 'true' }}
     uses: ./.github/workflows/_e2e_test.yaml

.github/workflows/pr_test_light.yaml

@@ -41,7 +41,7 @@ jobs:
   lint:
     uses: ./.github/workflows/_pre_commit.yml
     with:
-      vllm: 2f4e6548efec402b913ffddc8726230d9311948d
+      vllm: bde38c11df0ea066a740efe9b77fff5418be45df
   changes:
     runs-on: linux-aarch64-a2-0
     outputs:
@@ -81,7 +81,7 @@ jobs:
     if: ${{ needs.lint.result == 'success' && (needs.changes.outputs.e2e_tracker == 'true' || needs.changes.outputs.ut_tracker == 'true') }}
     strategy:
       matrix:
-        vllm_version: [2f4e6548efec402b913ffddc8726230d9311948d, v0.13.0]
+        vllm_version: [bde38c11df0ea066a740efe9b77fff5418be45df, v0.13.0]
     uses: ./.github/workflows/_unit_test.yaml
     with:
       vllm: ${{ matrix.vllm_version }}
@@ -93,7 +93,7 @@ jobs:
     name: e2e-light
     strategy:
       matrix:
-        vllm_version: [2f4e6548efec402b913ffddc8726230d9311948d, v0.13.0]
+        vllm_version: [bde38c11df0ea066a740efe9b77fff5418be45df, v0.13.0]
     # 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/schedule_codecov_refresh.yaml

@@ -33,7 +33,7 @@ jobs:
     name: refresh codecov
     strategy:
       matrix:
-        vllm_version: [2f4e6548efec402b913ffddc8726230d9311948d]
+        vllm_version: [bde38c11df0ea066a740efe9b77fff5418be45df]
     uses: ./.github/workflows/_unit_test.yaml
     with:
       vllm: ${{ matrix.vllm_version }}

docs/source/community/versioning_policy.md

@@ -51,7 +51,7 @@ If you're using v0.7.3, don't forget to install [mindie-turbo](https://pypi.org/
 For main branch of vLLM Ascend, we usually make it compatible with the latest vLLM release and a newer commit hash of vLLM. Please note that this table is usually updated. Please check it regularly.
 | vLLM Ascend | vLLM         | Python           | Stable CANN | PyTorch/torch_npu  |
 |-------------|--------------|------------------|-------------|--------------------|
-|     main    | 2f4e6548efec402b913ffddc8726230d9311948d, v0.13.0 tag | >= 3.10, < 3.12   | 8.3.RC2 | 2.8.0 / 2.8.0 |
+|     main    | bde38c11df0ea066a740efe9b77fff5418be45df, v0.13.0 tag | >= 3.10, < 3.12   | 8.3.RC2 | 2.8.0 / 2.8.0 |
 
 ## Release cadence
 

tests/e2e/singlecard/compile/test_norm_quant_fusion.py

@@ -305,15 +305,16 @@ def test_rmsnorm_quant_fusion(
 
     vllm_config = VllmConfig(model_config=ModelConfig(dtype=dtype))
 
-    update_environment_variables({
+    with vllm.config.set_current_vllm_config(vllm_config):
-        "RANK": "0",
+        update_environment_variables({
-        "LOCAL_RANK": "0",
+            "RANK": "0",
-        "WORLD_SIZE": "1",
+            "LOCAL_RANK": "0",
-        "MASTER_ADDR": "localhost",
+            "WORLD_SIZE": "1",
-        "MASTER_PORT": "12345",
+            "MASTER_ADDR": "localhost",
-    })
+            "MASTER_PORT": "12345",
-    init_distributed_environment()
+        })
-    ensure_model_parallel_initialized(1, 1)
+        init_distributed_environment()
+        ensure_model_parallel_initialized(1, 1)
 
     with vllm.config.set_current_vllm_config(vllm_config):
         with set_ascend_forward_context(None, vllm_config):

tests/ut/attention/test_attention_cp.py

@@ -33,6 +33,11 @@ class TestAscendAttentionCPImpl(TestBase):
         self.layer_no_quant.layer_name = "test_layer"
         self.layer_no_quant._k_scale_float = 1.0
         self.layer_no_quant._v_scale_float = 1.0
+        self.mock_vllm_config = MagicMock()
+        self.config_patcher = patch(
+            'vllm_ascend.attention.attention_v1.get_current_vllm_config',
+            return_value=self.mock_vllm_config)
+        self.config_patcher.start()
 
         self.impl = AscendAttentionCPImpl(
             num_heads=8,

tests/ut/attention/test_attention_v1.py

@@ -13,6 +13,23 @@ from vllm_ascend.utils import AscendDeviceType
 
 class TestAscendAttentionBackend(TestBase):
 
+    def setUp(self):
+        self.mock_config = MagicMock()
+
+        mock_parallel_config = MagicMock()
+        mock_parallel_config.prefill_context_parallel_size = 1
+        mock_parallel_config.decode_context_parallel_size = 1
+
+        self.mock_config.parallel_config = mock_parallel_config
+
+        self.utils_patcher = patch(
+            'vllm_ascend.attention.utils.get_current_vllm_config',
+            return_value=self.mock_config)
+        self.utils_patcher.start()
+
+        from vllm_ascend.attention.utils import enable_cp
+        enable_cp.cache_clear()
+
     def test_get_name(self):
         self.assertEqual(AscendAttentionBackend.get_name(), "CUSTOM")
 
@@ -102,6 +119,19 @@ class TestAscendAttentionMetadataBuilder(TestBase):
 class TestAscendAttentionBackendImpl(TestBase):
 
     def setUp(self):
+        self.mock_event = MagicMock()
+        self.mock_event.record.return_value = None
+        self.mock_event.wait.return_value = None
+
+        self.mock_stream = MagicMock()
+        self.event_patcher = patch('torch_npu.npu.Event',
+                                   return_value=self.mock_event)
+        self.stream_patcher = patch('torch_npu.npu.current_stream',
+                                    return_value=self.mock_stream)
+
+        self.event_patcher.start()
+        self.stream_patcher.start()
+
         self.layer = MagicMock()
         self.layer.layer_name = "test_layer"
         self.layer._k_scale_float = 1.0
@@ -119,6 +149,11 @@ class TestAscendAttentionBackendImpl(TestBase):
         self.layer_no_quant.layer_name = "test_layer"
         self.layer_no_quant._k_scale_float = 1.0
         self.layer_no_quant._v_scale_float = 1.0
+        self.mock_vllm_config = MagicMock()
+        self.config_patcher = patch(
+            'vllm_ascend.attention.attention_v1.get_current_vllm_config',
+            return_value=self.mock_vllm_config)
+        self.config_patcher.start()
 
         self.impl = AscendAttentionBackendImpl(
             num_heads=8,

tests/ut/attention/test_mla_v1.py

@@ -22,6 +22,23 @@ from vllm_ascend.utils import vllm_version_is
 
 class TestAscendMLABackend(TestBase):
 
+    def setUp(self):
+        self.mock_config = MagicMock()
+
+        mock_parallel_config = MagicMock()
+        mock_parallel_config.prefill_context_parallel_size = 1
+        mock_parallel_config.decode_context_parallel_size = 1
+
+        self.mock_config.parallel_config = mock_parallel_config
+
+        self.utils_patcher = patch(
+            'vllm_ascend.attention.utils.get_current_vllm_config',
+            return_value=self.mock_config)
+        self.utils_patcher.start()
+
+        from vllm_ascend.attention.utils import enable_cp
+        enable_cp.cache_clear()
+
     def test_get_name(self):
         self.assertEqual(AscendMLABackend.get_name(), "ASCEND_MLA")
 

tests/ut/attention/test_sfa_v1.py

@@ -12,6 +12,7 @@ if 'torch_npu._inductor' not in sys.modules:
 from vllm_ascend.attention.sfa_v1 import (AscendSFABackend, AscendSFAImpl,
                                           AscendSFAMetadata,
                                           AscendSFAMetadataBuilder)
+from vllm_ascend.utils import enable_dsa_cp
 
 
 class TestAscendSFABackend(TestBase):
@@ -83,6 +84,27 @@ class TestAscendSFAMetadata(TestBase):
 
 class TestAscendSFAMetadataBuilder(TestBase):
 
+    def setUp(self):
+        self.mock_cfg = MagicMock()
+
+        self.mock_cfg.parallel_config = MagicMock()
+        self.mock_cfg.parallel_config.tensor_parallel_size = 1
+        self.mock_cfg.parallel_config.prefill_context_parallel_size = 1
+        self.mock_cfg.parallel_config.decode_context_parallel_size = 1
+
+        self.mock_cfg.compilation_config = MagicMock()
+        self.mock_cfg.compilation_config.pass_config = MagicMock()
+        self.mock_cfg.compilation_config.pass_config.enable_sp = False
+
+        self.mock_cfg.speculative_config.num_speculative_tokens = 0
+
+        self.patcher = patch("vllm.config.get_current_vllm_config",
+                             return_value=self.mock_cfg)
+        self.patcher.start()
+
+        if hasattr(enable_dsa_cp, "cache_clear"):
+            enable_dsa_cp.cache_clear()
+
     def test_ascend_sfa_metadata_builder_default(self):
         kv_cache_spec = MagicMock()
         layer_names = ["layer1", "layer2"]

tests/ut/ops/test_activation.py

@@ -13,10 +13,11 @@
 # This file is a part of the vllm-ascend project.
 #
 
-from unittest.mock import patch
+from unittest.mock import MagicMock, patch
 
 import pytest
 import torch
+from vllm.config import set_current_vllm_config
 from vllm.model_executor.layers.activation import QuickGELU, SiluAndMul
 
 from vllm_ascend.utils import AscendDeviceType
@@ -27,8 +28,20 @@ def dummy_tensor():
     return torch.randn(4, 8, dtype=torch.float16)
 
 
+@pytest.fixture
+def default_vllm_config():
+    mock_config = MagicMock()
+
+    mock_config.compilation_config.dispatch_forward_backend = "eager"
+
+    mock_config.compilation_config.custom_ops = ["all"]
+
+    with set_current_vllm_config(mock_config):
+        yield mock_config
+
+
 @patch("torch_npu.npu_fast_gelu", side_effect=lambda x: x + 1)
-def test_QuickGELU_forward(mock_gelu, dummy_tensor):
+def test_QuickGELU_forward(mock_gelu, dummy_tensor, default_vllm_config):
     layer = QuickGELU()
     out = layer.forward(dummy_tensor)
 
@@ -45,7 +58,7 @@ def test_QuickGELU_forward(mock_gelu, dummy_tensor):
        side_effect=lambda x: None)
 def test_SiluAndMul_forward(mock_maybe_prefetch_mlp_down_proj,
                             mock_maybe_wait_prefetch_done, mock_swiglu,
-                            is_310p, dummy_tensor):
+                            is_310p, dummy_tensor, default_vllm_config):
 
     with patch("vllm_ascend.utils.get_ascend_device_type",
                return_value=AscendDeviceType._310P

tests/ut/ops/test_layernorm.py

@@ -1,7 +1,8 @@
-from unittest.mock import patch
+from unittest.mock import MagicMock, patch
 
 import pytest
 import torch
+from vllm.config import set_current_vllm_config
 from vllm.model_executor.layers.layernorm import RMSNorm
 
 from vllm_ascend.utils import AscendDeviceType
@@ -20,13 +21,22 @@ def mock_add_rms_norm(x, residual, weight, eps):
     return 2 * x, None, 2 * residual
 
 
+@pytest.fixture(autouse=True)
+def default_vllm_config():
+    mock_config = MagicMock()
+    mock_config.compilation_config.custom_ops = ["all"]
+
+    with set_current_vllm_config(mock_config):
+        yield mock_config
+
+
 @pytest.mark.parametrize("is_310p", [True, False])
 @pytest.mark.parametrize("residual",
                          [None, torch.randn(4, 8, dtype=torch.float32)])
 @patch("torch_npu.npu_rms_norm", side_effect=mock_rms_norm)
 @patch("torch_npu.npu_add_rms_norm", side_effect=mock_add_rms_norm)
 def test_RMSNorm_forward(mock_add_rmsnorm, mock_rmsnorm, is_310p, residual,
-                         dummy_tensor):
+                         dummy_tensor, default_vllm_config):
 
     with patch("vllm_ascend.utils.get_ascend_device_type",
                return_value=AscendDeviceType._310P

tests/ut/ops/test_rotary_embedding.py

@@ -78,6 +78,12 @@ class TestAscendRotaryEmbedding(unittest.TestCase):
 
     def setUp(self):
         # Common setup for tests
+        self.config_patcher = patch('vllm.config.vllm.get_current_vllm_config')
+        self.mock_get_config = self.config_patcher.start()
+        mock_config = MagicMock()
+        mock_config.compilation_config.custom_ops = ["all"]
+
+        self.mock_get_config.return_value = mock_config
         self.positions = torch.tensor([1, 2, 3])
         self.query = torch.randn(3, 1, 32, dtype=torch.float16)
         self.key = torch.randn(3, 1, 32, dtype=torch.float16)
@@ -242,6 +248,12 @@ class TestAscendDeepseekScalingRotaryEmbedding(TestBase):
 
     def setUp(self):
         # Common setup for tests
+        self.config_patcher = patch('vllm.config.vllm.get_current_vllm_config')
+        self.mock_get_config = self.config_patcher.start()
+        mock_config = MagicMock()
+        mock_config.compilation_config.custom_ops = ["all"]
+
+        self.mock_get_config.return_value = mock_config
         self.positions = torch.tensor([1, 2, 3])
         self.query = torch.randn(3, 1, 32, dtype=torch.float16)
         self.key = torch.randn(3, 1, 32, dtype=torch.float16)
@@ -368,7 +380,11 @@ class TestAscendDeepseekScalingRotaryEmbedding(TestBase):
 class TestAscendMRotaryEmbedding(unittest.TestCase):
 
     def setUp(self):
-        # Common setup for tests
+        self.config_patcher = patch('vllm.config.vllm.get_current_vllm_config')
+        self.mock_get_config = self.config_patcher.start()
+        mock_config = MagicMock()
+        mock_config.compilation_config.custom_ops = ["all"]
+        self.mock_get_config.return_value = mock_config
         self.number_tokens = 3
         self.num_head = 8
         self.num_kvhead = 8

tests/ut/ops/test_token_dispatcher.py

@@ -29,6 +29,23 @@ from vllm_ascend.ops.fused_moe.token_dispatcher import (  # isort: skip
 class TestTokenDispatcherWithMC2(TestBase):
 
     def setUp(self):
+        self.config_patcher = patch(
+            'vllm_ascend.ops.fused_moe.token_dispatcher.get_current_vllm_config'
+        )
+        self.mock_get_config = self.config_patcher.start()
+
+        mock_config = MagicMock()
+
+        mock_config.scheduler_config.max_num_seqs = 256
+        mock_config.scheduler_config.decode_max_num_seqs = 256
+
+        mock_config.compilation_config.custom_ops = ["all"]
+
+        mock_config.speculative_config = None
+
+        mock_config.parallel_config.tensor_parallel_size = 1
+
+        self.mock_get_config.return_value = mock_config
         self.mc2_group = MagicMock()
         self.mc2_group.device_group.return_value._get_backend.return_value.get_hccl_comm_name.return_value = "hccl_123"
         self.mc2_group.rank_in_group = 0

tests/ut/ops/test_vocab_parallel_embedding.py

@@ -208,6 +208,15 @@ class TestCustomVocabParallelEmbedding(unittest.TestCase):
 class TestAscendLogitsProcessor(unittest.TestCase):
 
     def setUp(self):
+        self.mock_vllm_config = MagicMock()
+        self.mock_vllm_config.compilation_config.custom_ops = ["all"]
+
+        from vllm.config.vllm import set_current_vllm_config
+        set_current_vllm_config(self.mock_vllm_config)
+
+        self.config_patch = patch("vllm.config.vllm.get_current_vllm_config",
+                                  return_value=self.mock_vllm_config)
+        self.config_patch.start()
         self.vocab_size = 50
         self.num_embeddings = 50
         self.embedding_dim = 10

tests/ut/worker/test_worker_v1.py

@@ -5,6 +5,7 @@ import torch
 from vllm.config import CacheConfig, ModelConfig, ParallelConfig, VllmConfig
 
 from tests.ut.base import TestBase
+from vllm_ascend.utils import vllm_version_is
 
 init_cached_hf_modules_path = "vllm.utils.import_utils.init_cached_hf_modules"
 
@@ -52,7 +53,7 @@ class TestNPUWorker(TestBase):
     @patch("vllm_ascend.worker.worker.get_ascend_config")
     @patch("vllm_ascend.worker.worker.init_ascend_config")
     @patch("vllm_ascend.worker.worker.check_ascend_device_type")
-    @patch(init_cached_hf_modules_path)
+    @patch(init_cached_hf_modules_path, create=True)
     @patch("vllm_ascend.worker.worker.NPUWorker._init_profiler")
     def test_init_npu_worker_normal_case(
         self,
@@ -106,7 +107,7 @@ class TestNPUWorker(TestBase):
     @patch("vllm_ascend.worker.worker.get_ascend_config")
     @patch("vllm_ascend.worker.worker.init_ascend_config")
     @patch("vllm_ascend.worker.worker.check_ascend_device_type")
-    @patch(init_cached_hf_modules_path)
+    @patch(init_cached_hf_modules_path, create=True)
     @patch("vllm_ascend.worker.worker.NPUWorker._init_profiler")
     def test_init_npu_worker_with_trust_remote_code(
         self,
@@ -140,7 +141,10 @@ class TestNPUWorker(TestBase):
         )
 
         # Verify init_cached_hf_modules is called (trust_remote_code=True)
-        mock_init_cached_hf_modules.assert_called_once()
+        if vllm_version_is('0.13.0'):
+            mock_init_cached_hf_modules.assert_called_once()
+        else:
+            mock_init_cached_hf_modules.assert_not_called()
 
     @patch("vllm_ascend.utils.adapt_patch")
     @patch("vllm_ascend.ops")
@@ -149,7 +153,7 @@ class TestNPUWorker(TestBase):
     @patch("vllm_ascend.worker.worker.get_ascend_config")
     @patch("vllm_ascend.worker.worker.init_ascend_config")
     @patch("vllm_ascend.worker.worker.check_ascend_device_type")
-    @patch(init_cached_hf_modules_path)
+    @patch(init_cached_hf_modules_path, create=True)
     @patch("vllm_ascend.worker.worker.NPUWorker._init_profiler")
     def test_init_npu_worker_with_custom_cache_dtype(
         self,
@@ -813,10 +817,11 @@ class TestNPUWorker(TestBase):
                 mock_scheduler_output, None)
             self.assertEqual(result, mock_model_output)
 
+    @patch("vllm_ascend.worker.worker.enable_sp", return_value=False)
     @patch("vllm_ascend.worker.worker.get_pp_group")
     @patch("vllm_ascend.worker.worker.get_tp_group")
     def test_execute_model_middle_rank(self, mock_get_tp_group,
-                                       mock_get_pp_group):
+                                       mock_get_pp_group, mock_enable_sp):
         """Test execute_model method - middle rank case"""
         from vllm.sequence import IntermediateTensors
 
@@ -1113,12 +1118,14 @@ class TestNPUWorker(TestBase):
             worker.model_runner.initialize_kv_cache.assert_called_once_with(
                 mock_kv_cache_config)
 
+    @patch("vllm_ascend.worker.worker.enable_sp", return_value=False)
     @patch("vllm_ascend.worker.worker.get_pp_group")
     @patch("vllm_ascend.worker.worker.get_tp_group")
     @patch("vllm_ascend.worker.worker.EMPTY_MODEL_RUNNER_OUTPUT")
     def test_execute_model_kv_connector_not_finished(self, mock_empty_output,
                                                      mock_get_tp_group,
-                                                     mock_get_pp_group):
+                                                     mock_get_pp_group,
+                                                     mock_enable_sp):
         """Test execute_model method - kv_connector_output not finished sending/recving case"""
         from vllm.sequence import IntermediateTensors
 

vllm_ascend/attention/mla_v1.py

@@ -6,7 +6,6 @@ import torch
 import torch_npu
 import vllm.envs as envs_vllm
 from vllm.attention.backends.abstract import AttentionBackend, MLAAttentionImpl
-from vllm.attention.backends.utils import PAD_SLOT_ID
 from vllm.config import VllmConfig, get_current_vllm_config
 from vllm.forward_context import ForwardContext, get_forward_context
 from vllm.logger import logger
@@ -39,12 +38,17 @@ from vllm_ascend.ops.rotary_embedding import get_cos_and_sin_mla
 from vllm_ascend.ops.weight_prefetch import maybe_npu_prefetch
 from vllm_ascend.quantization.w8a8 import AscendW8A8LinearMethod
 from vllm_ascend.utils import (ACL_FORMAT_FRACTAL_ND, maybe_trans_nz,
-                               weak_ref_tensors)
+                               vllm_version_is, weak_ref_tensors)
 from vllm_ascend.worker.npu_input_batch import NPUInputBatch
 
 if TYPE_CHECKING:
     from vllm.v1.core.sched.output import SchedulerOutput
 
+if vllm_version_is('0.13.0'):
+    from vllm.attention.backends.utils import PAD_SLOT_ID  # type: ignore
+else:
+    from vllm.v1.attention.backends.utils import PAD_SLOT_ID  # type: ignore
+
 MAX_O_PROJ_PREFETCH_SIZE = 16 * 1024 * 1024
 BUILD_METADATA_STEP_PREFILL = 0
 BUILD_METADATA_STEP_DECODE = 1

vllm_ascend/ops/triton/mamba/causal_conv1d.py

@@ -13,7 +13,13 @@ import torch
 import torch.nn.functional as F
 import triton
 import triton.language as tl
-from vllm.attention.backends.utils import PAD_SLOT_ID
+
+from vllm_ascend.utils import vllm_version_is
+
+if vllm_version_is('0.13.0'):
+    from vllm.attention.backends.utils import PAD_SLOT_ID  # type: ignore
+else:
+    from vllm.v1.attention.backends.utils import PAD_SLOT_ID  # type: ignore
 
 
 def causal_conv1d_ref(

vllm_ascend/worker/model_runner_v1.py

@@ -1670,6 +1670,8 @@ class NPUModelRunner(GPUModelRunner):
                 attn_metadata,
                 aux_hidden_states,
             )
+            if not vllm_version_is('0.13.0'):
+                self._copy_draft_token_ids_to_cpu(scheduler_output)
 
         (
             logprobs_lists,
@@ -1983,7 +1985,7 @@ class NPUModelRunner(GPUModelRunner):
                     query_start_loc_cpu=self.query_start_loc.cpu[:num_reqs +
                                                                  1],
                     _seq_lens_cpu=self.seq_lens.cpu[:num_reqs],
-                    seq_lens=self.seq_lens.cpu[:num_reqs],
+                    seq_lens=self.seq_lens.gpu[:num_reqs],
                     num_reqs=num_reqs,
                     num_actual_tokens=num_tokens,
                     block_table_tensor=block_table_tensor[:num_reqs],

vllm_ascend/worker/worker.py

@@ -121,11 +121,12 @@ class NPUWorker(WorkerBase):
             self.cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[
                 self.cache_config.cache_dtype]
 
-        if self.model_config.trust_remote_code:
+        if vllm_version_is('0.13.0'):
-            # note: lazy import to avoid importing torch before initializing
+            if self.model_config.trust_remote_code:
-            from vllm.utils.import_utils import init_cached_hf_modules
+                # note: lazy import to avoid importing torch before initializing
+                from vllm.utils.import_utils import init_cached_hf_modules
 
-            init_cached_hf_modules()
+                init_cached_hf_modules()
 
         self.profiler = self._init_profiler()
         if vllm_config.model_config and vllm_config.model_config.enable_sleep_mode: