[Main2Main] Upgrade vllm commit to 0120

.github/workflows/bot_pr_create.yaml

@@ -37,7 +37,7 @@ jobs:
     steps:
       - name: Get vLLM version
         run: |
-          VLLM_COMMIT=d68209402ddab3f54a09bc1f4de9a9495a283b60
+          VLLM_COMMIT=11b556878b958043e9c026919d4cba0236c85143
           echo "VLLM_COMMIT=https://github.com/vllm-project/vllm/commit/$VLLM_COMMIT" >> $GITHUB_ENV
 
       - name: Checkout repository

.github/workflows/dockerfiles/Dockerfile.lint

@@ -27,7 +27,7 @@ RUN apt-get update -y && \
 
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
 # For lint purpose, actually we need make a main2main matching.
-ARG VLLM_COMMIT=d68209402ddab3f54a09bc1f4de9a9495a283b60
+ARG VLLM_COMMIT=11b556878b958043e9c026919d4cba0236c85143
 RUN git clone $VLLM_REPO /vllm-workspace/vllm && \
     cd /vllm-workspace/vllm && \
     git checkout $VLLM_COMMIT

.github/workflows/pr_test_full.yaml

@@ -75,7 +75,7 @@ jobs:
     name: e2e-full
     strategy:
       matrix:
-        vllm_version: [d68209402ddab3f54a09bc1f4de9a9495a283b60, v0.14.1]
+        vllm_version: [11b556878b958043e9c026919d4cba0236c85143, v0.14.1]
     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: d68209402ddab3f54a09bc1f4de9a9495a283b60
+      vllm: 11b556878b958043e9c026919d4cba0236c85143
   changes:
     runs-on: linux-aarch64-a2-0
     outputs:
@@ -85,7 +85,7 @@ jobs:
     if: ${{ needs.lint.result == 'success' && (needs.changes.outputs.e2e_tracker == 'true' || needs.changes.outputs.ut_tracker == 'true') }}
     strategy:
       matrix:
-        vllm_version: [d68209402ddab3f54a09bc1f4de9a9495a283b60, v0.14.1]
+        vllm_version: [11b556878b958043e9c026919d4cba0236c85143, v0.14.1]
     uses: ./.github/workflows/_unit_test.yaml
     with:
       vllm: ${{ matrix.vllm_version }}
@@ -97,7 +97,7 @@ jobs:
     name: e2e-light
     strategy:
       matrix:
-        vllm_version: [d68209402ddab3f54a09bc1f4de9a9495a283b60, v0.14.1]
+        vllm_version: [11b556878b958043e9c026919d4cba0236c85143, v0.14.1]
     # 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: [d68209402ddab3f54a09bc1f4de9a9495a283b60]
+        vllm_version: [11b556878b958043e9c026919d4cba0236c85143]
     uses: ./.github/workflows/_unit_test.yaml
     with:
       vllm: ${{ matrix.vllm_version }}

docs/source/community/versioning_policy.md

@@ -55,7 +55,7 @@ For main branch of vLLM Ascend, we usually make it compatible with the latest vL
 
 | vLLM Ascend | vLLM         | Python           | Stable CANN | PyTorch/torch_npu  |
 |-------------|--------------|------------------|-------------|--------------------|
-|     main    | d68209402ddab3f54a09bc1f4de9a9495a283b60, v0.14.1 tag | >= 3.10, < 3.12   | 8.5.0 | 2.9.0 / 2.9.0 |
+|     main    | 11b556878b958043e9c026919d4cba0236c85143, v0.14.1 tag | >= 3.10, < 3.12   | 8.5.0 | 2.9.0 / 2.9.0 |
 
 ## Release cadence
 

tests/e2e/multicard/2-cards/test_aclgraph_capture_replay.py

@@ -109,8 +109,6 @@ def _run_worker_process(
             quantization="ascend" if "W8A8" in model_path else None,
             enable_expert_parallel=True if "DeepSeek" in model_path else False,
             trust_remote_code=True,
-            # vllm enables async scheduling by default, remove below when vllm >= 0.14.0
-            async_scheduling=False,
         )
 
         # Expose model config to the main test process
@@ -206,7 +204,8 @@ def test_models_aclgraph_capture_replay_metrics_dp2(
     # 2. Generation steps (max_tokens)
     # 3. Final step (likely EOS/idle step), no replay here
     total_steps = max_tokens + 1  # this includes the 1 and 2 above
-    expected_exec_model = (total_steps + 1) * dp_size
+    # vllm default enables Async scheduler, this will take 1 more steps
+    expected_exec_model = (total_steps + 1 + 1) * dp_size
 
     assert (
         num_execute_model == expected_exec_model

tests/e2e/multicard/2-cards/test_qwen3_performance.py

@@ -42,6 +42,7 @@
     "random_input_len": 128,
     "max_concurrency": 40,
     "random_output_len": 100,
+    "temperature": 0.0,
 }
 
 # NOTE: Any changes for the baseline throughput should be approved by team members.

tests/ut/attention/test_mla_v1.py

@@ -17,6 +17,7 @@
                                           AscendMLAPrefillMetadata,
                                           ChunkedContextMetadata)
 from vllm_ascend.attention.utils import AscendCommonAttentionMetadata
+from vllm_ascend.utils import vllm_version_is
 
 
 class TestAscendMLABackend(TestBase):
@@ -223,7 +224,9 @@ def mock_parent_init(self, kv_cache_spec, layer_names, vllm_config,
             )
 
         self.parent_init_patcher = patch(
-            "vllm.v1.attention.backends.mla.common.MLACommonMetadataBuilder.__init__",
+            ("vllm.v1.attention.backends.mla.common.MLACommonMetadataBuilder.__init__"
+ if vllm_version_is('0.14.1') else
+ "vllm.model_executor.layers.attention.mla_attention.MLACommonMetadataBuilder.__init__"),
             mock_parent_init)
         self.parent_init_patcher.start()
 
@@ -449,7 +452,9 @@ def mock_parent_init(self, kv_cache_spec, layer_names, vllm_config,
             )
 
         self.parent_init_patcher = patch(
-            "vllm.v1.attention.backends.mla.common.MLACommonMetadataBuilder.__init__",
+            ("vllm.v1.attention.backends.mla.common.MLACommonMetadataBuilder.__init__"
+ if vllm_version_is('0.14.1') else
+ "vllm.model_executor.layers.attention.mla_attention.MLACommonMetadataBuilder.__init__"),
             mock_parent_init)
         self.parent_init_patcher.start()
 

tests/ut/attention/test_sfa_v1.py

@@ -13,7 +13,7 @@
 from vllm_ascend.attention.sfa_v1 import (AscendSFABackend, AscendSFAImpl,
                                           AscendSFAMetadata,
                                           AscendSFAMetadataBuilder)
-from vllm_ascend.utils import enable_dsa_cp
+from vllm_ascend.utils import enable_dsa_cp, vllm_version_is
 
 
 class TestAscendSFABackend(TestBase):
@@ -124,7 +124,9 @@ def mock_parent_init(self, kv_cache_spec, layer_names, vllm_config,
             )
 
         self.parent_init_patcher = patch(
-            "vllm.v1.attention.backends.mla.common.MLACommonMetadataBuilder.__init__",
+            ("vllm.v1.attention.backends.mla.common.MLACommonMetadataBuilder.__init__"
+ if vllm_version_is('0.14.1') else
+ "vllm.model_executor.layers.attention.mla_attention.MLACommonMetadataBuilder.__init__"),
             mock_parent_init)
         self.parent_init_patcher.start()
 

tests/ut/eplb/core/test_eplb_utils.py

@@ -9,6 +9,7 @@
 
 from vllm_ascend.ascend_config import init_ascend_config
 from vllm_ascend.eplb.core.eplb_utils import init_eplb_config
+from vllm_ascend.utils import vllm_version_is
 # isort: on
 
 
@@ -20,8 +21,24 @@ def setUp(self, mock_fix_incompatible_config):
             "refresh": True,
             "eplb_config": {"dynamic_eplb": True, "num_redundant_experts": 2},
         }
-        moe_parallel_config = FusedMoEParallelConfig(2, 0, 1, 2, 1, 1, 1, 1, True, "hccl")
-        moe_config = FusedMoEConfig(8, 8, 8192, 5, moe_parallel_config, torch.float16)
+        if vllm_version_is('0.14.1'):
+            moe_parallel_config = FusedMoEParallelConfig(2, 0, 1, 2, 1, 1, 1, 1, True, "hccl")
+            moe_config = FusedMoEConfig(8, 8, 8192, 5, moe_parallel_config, torch.float16)
+        else:
+            from vllm.model_executor.layers.fused_moe.config import RoutingMethodType
+            moe_parallel_config = FusedMoEParallelConfig(2, 0, 1, 2, 1, 1, 1, 1, True, "hccl", enable_eplb=True)
+            moe_config = FusedMoEConfig(
+                num_experts=8,
+                experts_per_token=8,
+                hidden_dim=8192,
+                intermediate_size_per_partition=5,
+                num_local_experts=8,
+                activation="silu",
+                device="npu",
+                routing_method=RoutingMethodType.Simulated,
+                moe_parallel_config=moe_parallel_config,
+                in_dtype=torch.float16,
+            )
         moe_config.supports_eplb = True
         self.vllm_config = vllm_config
         self.moe_config = moe_config

tests/ut/spec_decode/test_eagle_proposer.py

@@ -51,6 +51,7 @@ def tearDown(self):
     def test_initialization_eagle_graph(self):
         self.vllm_config.speculative_config.method = "eagle"
         self.vllm_config.speculative_config.draft_model_config.get_hidden_size.return_value = 4096
+        self.vllm_config.speculative_config.draft_model_config.uses_mrope = False
         self.vllm_config.compilation_config.mode = CompilationMode.VLLM_COMPILE
         self.vllm_config.model_config.enforce_eager = False
         self.vllm_config.model_config.uses_mrope = False
@@ -65,10 +66,11 @@ def test_initialization_eagle_graph(self):
         self.assertEqual(proposer.hidden_size, 4096)
         self.assertTrue(proposer.use_cuda_graph)
 
-        self.assertEqual(proposer.input_ids.shape, (1024, ))
-        self.assertEqual(proposer.positions.shape, (1024, ))
-        self.assertEqual(proposer.hidden_states.shape, (1024, 4096))
-        self.assertEqual(proposer.arange.shape, (1024, ))
+        expected_max_num_tokens = proposer.max_num_tokens
+        self.assertEqual(proposer.input_ids.shape, (expected_max_num_tokens, ))
+        self.assertEqual(proposer.positions.shape, (expected_max_num_tokens, ))
+        self.assertEqual(proposer.hidden_states.shape, (expected_max_num_tokens, 4096))
+        self.assertEqual(proposer.arange.shape, (expected_max_num_tokens, ))
 
     def test_initialization_eagle3_enforce_eager(self):
         self.vllm_config.speculative_config.method = "eagle3"
@@ -83,7 +85,8 @@ def test_initialization_eagle3_enforce_eager(self):
 
         self.assertEqual(proposer.hidden_size, 2048)
         self.assertFalse(proposer.use_cuda_graph)
-        self.assertEqual(proposer.hidden_states.shape, (1024, 2048))
+        expected_max_num_tokens = proposer.max_num_tokens
+        self.assertEqual(proposer.hidden_states.shape, (expected_max_num_tokens, 2048))
 
     def test_initialization_eagle3_full_graph_async(self):
         self.vllm_config.speculative_config.method = "eagle3"
@@ -100,7 +103,8 @@ def test_initialization_eagle3_full_graph_async(self):
 
         self.assertEqual(proposer.hidden_size, 2048)
         self.assertTrue(proposer.use_cuda_graph)
-        self.assertEqual(proposer.hidden_states.shape, (1024, 2048))
+        expected_max_num_tokens = proposer.max_num_tokens
+        self.assertEqual(proposer.hidden_states.shape, (expected_max_num_tokens, 2048))
 
     def test_initialization_mtp_full_graph_async(self):
         self.vllm_config.speculative_config.method = "mtp"
@@ -117,7 +121,8 @@ def test_initialization_mtp_full_graph_async(self):
 
         self.assertEqual(proposer.hidden_size, 2048)
         self.assertFalse(proposer.use_cuda_graph)
-        self.assertEqual(proposer.hidden_states.shape, (1024, 2048))
+        expected_max_num_tokens = proposer.max_num_tokens
+        self.assertEqual(proposer.hidden_states.shape, (expected_max_num_tokens, 2048))
 
 
 class TestEagleProposerLoadModel(TestBase):

tests/ut/spec_decode/test_mtp_proposer.py

@@ -33,6 +33,7 @@ def vllm_config(self):
         config.speculative_config.method = "mtp"
         config.speculative_config.draft_model_config = MagicMock()
         config.speculative_config.draft_model_config.get_hidden_size.return_value = 4096
+        config.speculative_config.draft_model_config.uses_mrope = False
         config.speculative_config.speculative_token_tree = str([
             (i + 1) * (0, ) for i in range(2)
         ])

vllm_ascend/ascend_forward_context.py

@@ -19,6 +19,7 @@
     is_drafter_moe_model,
     is_moe_model,
     speculative_enable_dispatch_gmm_combine_decode,
+    vllm_version_is,
 )
 
 
@@ -42,20 +43,26 @@ def set_ascend_forward_context(
     batch_descriptor: BatchDescriptor | None = None,
     model_instance: torch.nn.Module = None,
     is_draft_model=False,
+    skip_compiled: bool = False,
 ):
     """A context manager that stores the current forward context,
     can be attention metadata, etc.
     We add some additional param into forward_context.
     """
-    with set_forward_context(
-        attn_metadata,
-        vllm_config,
-        virtual_engine=virtual_engine,
-        num_tokens=num_tokens,
-        num_tokens_across_dp=num_tokens_across_dp,
-        cudagraph_runtime_mode=aclgraph_runtime_mode,
-        batch_descriptor=batch_descriptor,
-    ):
+    forward_context_kwargs = {
+        "attn_metadata": attn_metadata,
+        "vllm_config": vllm_config,
+        "virtual_engine": virtual_engine,
+        "num_tokens": num_tokens,
+        "num_tokens_across_dp": num_tokens_across_dp,
+        "cudagraph_runtime_mode": aclgraph_runtime_mode,
+        "batch_descriptor": batch_descriptor,
+    }
+
+    if not vllm_version_is("0.14.1"):
+        forward_context_kwargs["skip_compiled"] = skip_compiled
+
+    with set_forward_context(**forward_context_kwargs):
         forward_context = get_forward_context()
 
         from vllm_ascend.ops.fused_moe.moe_comm_method import get_moe_comm_method

vllm_ascend/attention/mla_v1.py

@@ -11,7 +11,6 @@
 from vllm.model_executor.layers.linear import UnquantizedLinearMethod
 from vllm.utils.math_utils import cdiv, round_down
 from vllm.v1.attention.backend import AttentionBackend, AttentionCGSupport, MLAAttentionImpl  # type: ignore
-from vllm.v1.attention.backends.mla.common import MLACommonMetadataBuilder
 from vllm.v1.attention.backends.utils import PAD_SLOT_ID  # type: ignore
 from vllm.v1.kv_cache_interface import AttentionSpec, MLAAttentionSpec
 
@@ -45,12 +44,18 @@
 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.methods import AscendW8A8LinearMethod
-from vllm_ascend.utils import ACL_FORMAT_FRACTAL_ND, maybe_trans_nz, weak_ref_tensors
+from vllm_ascend.utils import ACL_FORMAT_FRACTAL_ND, maybe_trans_nz, 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
 
+# isort: off
+if vllm_version_is("0.14.1"):
+    from vllm.v1.attention.backends.mla.common import MLACommonMetadataBuilder  # type: ignore
+else:
+    from vllm.model_executor.layers.attention.mla_attention import MLACommonMetadataBuilder
+# isort: on
 
 MAX_O_PROJ_PREFETCH_SIZE = 16 * 1024 * 1024
 BUILD_METADATA_STEP_PREFILL = 0

vllm_ascend/attention/sfa_v1.py

@@ -12,7 +12,6 @@
 from vllm.model_executor.layers.linear import UnquantizedLinearMethod
 from vllm.triton_utils import HAS_TRITON
 from vllm.v1.attention.backend import AttentionBackend, AttentionCGSupport, MLAAttentionImpl  # type: ignore
-from vllm.v1.attention.backends.mla.common import MLACommonMetadataBuilder
 from vllm.v1.kv_cache_interface import AttentionSpec
 
 from vllm_ascend import envs
@@ -46,11 +45,17 @@
     enable_dsa_cp,
     enable_dsa_cp_with_layer_shard,
     maybe_trans_nz,
+    vllm_version_is,
 )
 from vllm_ascend.worker.npu_input_batch import NPUInputBatch
 
 if TYPE_CHECKING:
     from vllm.v1.core.sched.output import SchedulerOutput
+if vllm_version_is("0.14.1"):
+    from vllm.v1.attention.backends.mla.common import MLACommonMetadataBuilder  # type: ignore
+else:
+    from vllm.model_executor.layers.attention.mla_attention import MLACommonMetadataBuilder
+# isort: on
 
 # token count limits within bmm_transpose operator
 BMM_TRANS_MAX_SUPPORTED_TOKENS = 1024