[Quickfix] update CachedRequestState as NewRequestData changed (#2367)

MengqingCao · web-flow · commit 61866b8ac6e8 · 2025-08-15T07:35:27.000+08:00
### What this PR does / why we need it? 1. update `CachedRequestState` as `NewRequestData` changed in vllm-project/vllm#22570 2. drop maintenance of vllm v0.10.0 in the branch main ### Does this PR introduce _any_ user-facing change? N/A ### How was this patch tested? CI passed with existing test. - vLLM version: v0.10.0 - vLLM main: vllm-project/vllm@92ff41a --------- Signed-off-by: MengqingCao <cmq0113@163.com>
diff --git a/.github/workflows/vllm_ascend_test.yaml b/.github/workflows/vllm_ascend_test.yaml
@@ -81,7 +81,7 @@ jobs:
         VLLM_USE_MODELSCOPE: True
     strategy:
       matrix:
-        vllm_version: [main, v0.10.0]
+        vllm_version: [main]
     steps:
       - name: Install packages
         run: |
@@ -137,7 +137,7 @@ jobs:
       max-parallel: 2
       matrix:
         os: [linux-aarch64-a2-1]
-        vllm_version: [main, v0.10.0]
+        vllm_version: [main]
     name: singlecard e2e test
     runs-on: ${{ matrix.os }}
     container:
@@ -185,9 +185,6 @@ jobs:
         run: |
           pip install -r requirements-dev.txt
           pip install -v -e .
-          if [[ "${{ matrix.vllm_version }}" == "v0.10.0" ]]; then
-            pip install "transformers<4.54.0"
-          fi
 
       - name: Run e2e test
         env:
@@ -222,7 +219,7 @@ jobs:
       max-parallel: 2
       matrix:
         os: [linux-aarch64-a2-2]
-        vllm_version: [main, v0.10.0]
+        vllm_version: [main]
     name: multicard e2e test
     runs-on: ${{ matrix.os }}
     container:
@@ -270,9 +267,6 @@ jobs:
         run: |
           pip install -r requirements-dev.txt
           pip install -v -e .
-          if [[ "${{ matrix.vllm_version }}" == "v0.10.0" ]]; then
-            pip install "transformers<4.54.0"
-          fi
 
       - name: Run vllm-project/vllm-ascend test
         env:
diff --git a/.github/workflows/vllm_ascend_test_310p.yaml b/.github/workflows/vllm_ascend_test_310p.yaml
@@ -53,7 +53,7 @@ jobs:
       max-parallel: 2
       matrix:
         os: [linux-aarch64-310p-1, linux-aarch64-310p-4]
-        vllm_version: [main, v0.10.0]
+        vllm_version: [main]
     name: 310p e2e test
     runs-on: ${{ matrix.os }}
     container:
diff --git a/tests/ut/core/test_scheduler.py b/tests/ut/core/test_scheduler.py
@@ -50,7 +50,7 @@ def create_requests(
             request_id=f"{i}",
             prompt_token_ids=[i] * num_tokens,
             sampling_params=sampling_params,
-            multi_modal_inputs=mm_inputs,
+            multi_modal_kwargs=mm_inputs,
             multi_modal_placeholders=mm_position,
             multi_modal_hashes=None,
             eos_token_id=EOS_TOKEN_ID,
diff --git a/tests/ut/kv_connector/test_remote_decode_lifecycle.py b/tests/ut/kv_connector/test_remote_decode_lifecycle.py
@@ -25,7 +25,6 @@
                                          create_model_runner_output,
                                          create_request, create_scheduler,
                                          create_vllm_config)
-from vllm_ascend.utils import vllm_version_is
 
 
 def test_basic_lifecycle():
@@ -103,13 +102,10 @@ def test_basic_lifecycle():
 
     # (3b): execute_model()
     model_runner_output = copy.deepcopy(EMPTY_MODEL_RUNNER_OUTPUT)
-    if vllm_version_is("0.10.0"):
-        model_runner_output.finished_sending = [request_id]
-    else:
-        from vllm.v1.worker.kv_connector_model_runner_mixin import \
-            KVConnectorOutput  # type: ignore  # noqa
-        model_runner_output.kv_connector_output = KVConnectorOutput(
-            finished_sending=[request_id])
+    from vllm.v1.worker.kv_connector_model_runner_mixin import \
+        KVConnectorOutput  # type: ignore  # noqa
+    model_runner_output.kv_connector_output = KVConnectorOutput(
+        finished_sending=[request_id])
 
     # (3c): update_from_output()
     scheduler.update_from_output(scheduler_output, model_runner_output)
@@ -164,13 +160,10 @@ def test_prefix_cache_lifecycle():
     scheduler_output = scheduler.schedule()
     scheduler.schedule()
     model_runner_output = copy.deepcopy(EMPTY_MODEL_RUNNER_OUTPUT)
-    if vllm_version_is("0.10.0"):
-        model_runner_output.finished_sending = [request_remote.request_id]
-    else:
-        from vllm.v1.worker.kv_connector_model_runner_mixin import \
-            KVConnectorOutput  # noqa
-        model_runner_output.kv_connector_output = KVConnectorOutput(
-            finished_sending=[request_remote.request_id])
+    from vllm.v1.worker.kv_connector_model_runner_mixin import \
+        KVConnectorOutput  # noqa
+    model_runner_output.kv_connector_output = KVConnectorOutput(
+        finished_sending=[request_remote.request_id])
     scheduler.update_from_output(scheduler_output, model_runner_output)
     _ = scheduler.schedule()
     assert_scheduler_empty(scheduler)
diff --git a/tests/ut/kv_connector/test_remote_prefill_lifecycle.py b/tests/ut/kv_connector/test_remote_prefill_lifecycle.py
@@ -25,7 +25,6 @@
                                          create_model_runner_output,
                                          create_request, create_scheduler,
                                          create_vllm_config)
-from vllm_ascend.utils import vllm_version_is
 
 
 def test_basic_lifecycle():
@@ -91,13 +90,10 @@ def test_basic_lifecycle():
 
     # (2b): forward(): request finishes recv.
     model_runner_output = copy.deepcopy(EMPTY_MODEL_RUNNER_OUTPUT)
-    if vllm_version_is("0.10.0"):
-        model_runner_output.finished_recving = [request_id]
-    else:
-        from vllm.v1.worker.kv_connector_model_runner_mixin import \
-            KVConnectorOutput  # type: ignore  # noqa
-        model_runner_output.kv_connector_output = KVConnectorOutput(
-            finished_recving=[request_id])
+    from vllm.v1.worker.kv_connector_model_runner_mixin import \
+        KVConnectorOutput  # type: ignore  # noqa
+    model_runner_output.kv_connector_output = KVConnectorOutput(
+        finished_recving=[request_id])
 
     # (2c): update_from_output():
     engine_core_outputs = scheduler.update_from_output(scheduler_output,
@@ -211,13 +207,10 @@ def test_full_block_prompt():
     # # STEP (2): Recv.
     scheduler_output = scheduler.schedule()
     model_runner_output = copy.deepcopy(EMPTY_MODEL_RUNNER_OUTPUT)
-    if vllm_version_is("0.10.0"):
-        model_runner_output.finished_recving = [request_id]
-    else:
-        from vllm.v1.worker.kv_connector_model_runner_mixin import \
-            KVConnectorOutput  # type: ignore  # noqa
-        model_runner_output.kv_connector_output = KVConnectorOutput(
-            finished_recving=[request_id])
+    from vllm.v1.worker.kv_connector_model_runner_mixin import \
+        KVConnectorOutput  # type: ignore  # noqa
+    model_runner_output.kv_connector_output = KVConnectorOutput(
+        finished_recving=[request_id])
     scheduler.update_from_output(scheduler_output, model_runner_output)
     assert len(scheduler.waiting) == 1
     assert (request_id in scheduler.finished_recving_kv_req_ids)
diff --git a/tests/ut/kv_connector/utils.py b/tests/ut/kv_connector/utils.py
@@ -157,7 +157,7 @@ def create_request(
         request_id=f"id-{request_id}",
         prompt_token_ids=prompt_token_ids,
         sampling_params=sampling_params,
-        multi_modal_inputs=None,
+        multi_modal_kwargs=None,
         multi_modal_placeholders=None,
         multi_modal_hashes=None,
         **({
@@ -187,19 +187,11 @@ def create_model_runner_output(
 
     # Make output data structure.
     extra_args = {}
-    if not vllm_version_is("0.10.0"):
-        from vllm.v1.worker.kv_connector_model_runner_mixin import \
-            KVConnectorOutput  # type: ignore  # noqa
-        kv_connector_output = KVConnectorOutput(
-            finished_sending=finished_sending,
-            finished_recving=finished_recving)
-        extra_args = {"kv_connector_output": kv_connector_output}
-    else:
-        extra_args = {
-            "finished_sending": finished_sending,
-            "finished_recving": finished_recving,
-        }
-
+    from vllm.v1.worker.kv_connector_model_runner_mixin import \
+        KVConnectorOutput  # type: ignore  # noqa
+    kv_connector_output = KVConnectorOutput(finished_sending=finished_sending,
+                                            finished_recving=finished_recving)
+    extra_args = {"kv_connector_output": kv_connector_output}
     return ModelRunnerOutput(
         req_ids=req_ids,
         req_id_to_index=req_id_to_index,
diff --git a/tests/ut/worker/test_input_batch.py b/tests/ut/worker/test_input_batch.py
@@ -12,7 +12,7 @@ def mock_cached_request_state(req_id="1", prompt=[1, 2, 3], output=[4, 5, 6]):
     return CachedRequestState(
         req_id=req_id,
         prompt_token_ids=prompt,
-        mm_inputs=[],
+        mm_kwargs=[],
         mm_positions=[],
         sampling_params=SamplingParams(),
         pooling_params=None,
diff --git a/vllm_ascend/models/qwen2_5_vl.py b/vllm_ascend/models/qwen2_5_vl.py
@@ -30,8 +30,7 @@
 from vllm.config import VllmConfig
 from vllm.distributed import parallel_state
 from vllm.distributed import utils as dist_utils
-from vllm.model_executor.layers.activation import (_ACTIVATION_REGISTRY,
-                                                   get_act_and_mul_fn)
+from vllm.model_executor.layers.activation import get_act_and_mul_fn
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
@@ -43,8 +42,6 @@
 from vllm.model_executor.models.utils import maybe_prefix
 from vllm.multimodal import MULTIMODAL_REGISTRY
 
-from vllm_ascend.utils import vllm_version_is
-
 MIN_PAD_SIZE = 64  # min_size to pad weight
 MAX_PAD_SIZE = 128  # max_size to pad weight
 
@@ -202,8 +199,6 @@ def __init__(
         )
 
         act_fn = get_act_and_mul_fn(vision_config.hidden_act)
-        if vllm_version_is("0.10.0"):
-            act_fn = _ACTIVATION_REGISTRY[vision_config.hidden_act]
         self.blocks = nn.ModuleList([
             AscendQwen2_5_VisionBlock(
                 dim=self.hidden_size,
@@ -303,12 +298,9 @@ def load_weights(self, weights: Iterable[Tuple[str,
             ("qkv_proj", "q_proj", "q"),
             ("qkv_proj", "k_proj", "k"),
             ("qkv_proj", "v_proj", "v"),
+            ("mlp.gate_up_proj.", "mlp.gate_proj.", 0),
+            ("mlp.gate_up_proj.", "mlp.up_proj.", 1),
         ]
-        if not vllm_version_is("0.10.0"):
-            stacked_params_mapping.extend([
-                ("mlp.gate_up_proj.", "mlp.gate_proj.", 0),
-                ("mlp.gate_up_proj.", "mlp.up_proj.", 1),
-            ])
         params_dict = dict(self.named_parameters(remove_duplicate=False))
         loaded_params: Set[str] = set()
         for name, loaded_weight in weights:
diff --git a/vllm_ascend/models/qwen2_5_vl_without_padding.py b/vllm_ascend/models/qwen2_5_vl_without_padding.py
@@ -30,8 +30,7 @@
 from vllm.config import VllmConfig
 from vllm.distributed import parallel_state
 from vllm.distributed import utils as dist_utils
-from vllm.model_executor.layers.activation import (_ACTIVATION_REGISTRY,
-                                                   get_act_and_mul_fn)
+from vllm.model_executor.layers.activation import get_act_and_mul_fn
 from vllm.model_executor.layers.layernorm import RMSNorm
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.models.qwen2_5_vl import (
@@ -43,7 +42,6 @@
 from vllm.multimodal import MULTIMODAL_REGISTRY
 
 from vllm_ascend.models.qwen2_5_vl import AscendQwen2_5_VisionRotaryEmbedding
-from vllm_ascend.utils import vllm_version_is
 
 
 class AscendQwen2_5_VisionAttention_Without_Padding(Qwen2_5_VisionAttention):
@@ -175,8 +173,6 @@ def __init__(
         )
 
         act_fn = get_act_and_mul_fn(vision_config.hidden_act)
-        if vllm_version_is("0.10.0"):
-            act_fn = _ACTIVATION_REGISTRY[vision_config.hidden_act]
         self.blocks = nn.ModuleList([
             AscendQwen2_5_VisionBlock_Without_Padding(
                 dim=self.hidden_size,
diff --git a/vllm_ascend/multistream/ms_split.py b/vllm_ascend/multistream/ms_split.py
@@ -105,7 +105,7 @@ def model_input_split_v1_mla_attn(
     [block_table_pre,
      block_table_post] = split_attn_tensor_type(attn_metadata.block_tables,
                                                 seq_index)
-
+    assert attn_metadata.attn_mask is not None
     if attn_metadata.attn_state == AscendAttentionState.PrefillNoCache or attn_metadata.attn_state == AscendAttentionState.PrefillCacheHit:
         # the attn_mla kernel in torch npu only accept 128*128 attn mask
         attn_mask_pre = attn_mask_post = attn_metadata.attn_mask
diff --git a/vllm_ascend/patch/platform/__init__.py b/vllm_ascend/patch/platform/__init__.py
@@ -14,12 +14,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from vllm_ascend.utils import vllm_version_is
-
-# Import specific patches for different versions
-if vllm_version_is("0.10.0"):
-    from vllm_ascend.patch.platform import patch_0_10_0  # noqa: F401
-    from vllm_ascend.patch.platform import patch_common  # noqa: F401
-else:
-    from vllm_ascend.patch.platform import patch_common  # noqa: F401
-    from vllm_ascend.patch.platform import patch_main  # noqa: F401
+from vllm_ascend.patch.platform import patch_common  # noqa: F401
+from vllm_ascend.patch.platform import patch_main  # noqa: F401
diff --git a/vllm_ascend/patch/platform/patch_0_10_0/__init__.py b/vllm_ascend/patch/platform/patch_0_10_0/__init__.py
diff --git a/vllm_ascend/patch/worker/__init__.py b/vllm_ascend/patch/worker/__init__.py
@@ -15,12 +15,5 @@
 # limitations under the License.
 #
 
-from vllm_ascend.utils import vllm_version_is
-
-# Import specific patches for different versions
-if vllm_version_is("0.10.0"):
-    from vllm_ascend.patch.worker import patch_0_10_0  # noqa: F401
-    from vllm_ascend.patch.worker import patch_common  # noqa: F401
-else:
-    from vllm_ascend.patch.worker import patch_common  # noqa: F401
-    from vllm_ascend.patch.worker import patch_main  # noqa: F401
+from vllm_ascend.patch.worker import patch_common  # noqa: F401
+from vllm_ascend.patch.worker import patch_main  # noqa: F401
diff --git a/vllm_ascend/patch/worker/patch_0_10_0/__init__.py b/vllm_ascend/patch/worker/patch_0_10_0/__init__.py
diff --git a/vllm_ascend/patch/worker/patch_0_10_0/patch_sampler_gather_logprobs.py b/vllm_ascend/patch/worker/patch_0_10_0/patch_sampler_gather_logprobs.py
diff --git a/vllm_ascend/worker/model_runner_v1.py b/vllm_ascend/worker/model_runner_v1.py
diff --git a/vllm_ascend/worker/npu_input_batch.py b/vllm_ascend/worker/npu_input_batch.py
diff --git a/vllm_ascend/worker/worker_v1.py b/vllm_ascend/worker/worker_v1.py
