Drop vLLM 0.13.0 support (#6069)

### What this PR does / why we need it?
Drop vLLM 0.13.0 support, upgrade to 0.14.0

- vLLM version: v0.13.0
- vLLM main:
d68209402d

---------

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

.github/workflows/_e2e_nightly_multi_node.yaml

@@ -32,7 +32,7 @@ on:
         description: how many pods will be pulled up via lws.yaml, indicates number of nodes we need
       vllm_version:
         required: false
-        default: "v0.13.0"
+        default: "v0.14.0"
         type: string
         description: vllm version to use
       vllm_ascend_remote_url:

.github/workflows/nightly_test_a2.yaml

@@ -66,7 +66,7 @@ jobs:
             tests: tests/e2e/nightly/single_node/ops/multicard_ops_a2/
     uses: ./.github/workflows/_e2e_nightly_single_node.yaml
     with:
-      vllm: v0.13.0
+      vllm: v0.14.0
       runner: ${{ matrix.test_config.os }}
       tests: ${{ matrix.test_config.tests }}
       name: ${{ matrix.test_config.name }}
@@ -137,7 +137,7 @@ jobs:
               - Qwen3-Omni-30B-A3B-Instruct
     uses: ./.github/workflows/_e2e_nightly_single_node_models.yaml
     with:
-      vllm: v0.13.0
+      vllm: v0.14.0
       runner: ${{ matrix.test_config.os }}
       model_list: ${{ toJson(matrix.test_config.model_list) }}
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-910b-ubuntu22.04-py3.11'

.github/workflows/nightly_test_a3.yaml

@@ -158,7 +158,7 @@ jobs:
             tests: tests/e2e/nightly/single_node/models/test_deepseek_v3_2_w8a8.py
     uses: ./.github/workflows/_e2e_nightly_single_node.yaml
     with:
-      vllm: v0.13.0
+      vllm: v0.14.0
       runner: ${{ matrix.test_config.os }}
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/vllm-ascend:nightly-a3'
       tests: ${{ matrix.test_config.tests }}
@@ -178,7 +178,7 @@ jobs:
     uses: ./.github/workflows/_e2e_nightly_single_node.yaml
     with:
       runner: ${{ matrix.test_config.os }}
-      vllm: v0.13.0
+      vllm: v0.14.0
       image: 'swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/vllm-ascend:nightly-a3'
       tests: ${{ matrix.test_config.tests }}
       name: ${{ matrix.test_config.name }}

.github/workflows/pr_test_full.yaml

@@ -75,7 +75,7 @@ jobs:
     name: e2e-full
     strategy:
       matrix:
-        vllm_version: [d68209402ddab3f54a09bc1f4de9a9495a283b60, v0.13.0]
+        vllm_version: [d68209402ddab3f54a09bc1f4de9a9495a283b60, v0.14.0]
     needs: [changes]
     if: ${{ needs.changes.outputs.e2e_tracker == 'true' }}
     uses: ./.github/workflows/_e2e_test.yaml

.github/workflows/pr_test_light.yaml

@@ -84,7 +84,7 @@ jobs:
     if: ${{ needs.lint.result == 'success' && (needs.changes.outputs.e2e_tracker == 'true' || needs.changes.outputs.ut_tracker == 'true') }}
     strategy:
       matrix:
-        vllm_version: [d68209402ddab3f54a09bc1f4de9a9495a283b60, v0.13.0]
+        vllm_version: [d68209402ddab3f54a09bc1f4de9a9495a283b60, v0.14.0]
     uses: ./.github/workflows/_unit_test.yaml
     with:
       vllm: ${{ matrix.vllm_version }}
@@ -96,7 +96,7 @@ jobs:
     name: e2e-light
     strategy:
       matrix:
-        vllm_version: [d68209402ddab3f54a09bc1f4de9a9495a283b60, v0.13.0]
+        vllm_version: [d68209402ddab3f54a09bc1f4de9a9495a283b60, v0.14.0]
     # 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_test_benchmarks.yaml

@@ -51,7 +51,7 @@ jobs:
     strategy:
       matrix:
         include:
-          - vllm_branch: v0.13.0
+          - vllm_branch: v0.14.0
             vllm_ascend_branch: main
       max-parallel: 1
     container:

Dockerfile

@@ -48,7 +48,7 @@ RUN pip config set global.index-url ${PIP_INDEX_URL}
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.13.0
+ARG VLLM_TAG=v0.14.0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -v -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \

Dockerfile.310p

@@ -40,7 +40,7 @@ RUN pip config set global.index-url ${PIP_INDEX_URL}
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.13.0
+ARG VLLM_TAG=v0.14.0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -v -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \

Dockerfile.310p.openEuler

@@ -36,7 +36,7 @@ COPY . /vllm-workspace/vllm-ascend/
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.13.0
+ARG VLLM_TAG=v0.14.0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \

Dockerfile.a3

@@ -47,7 +47,7 @@ RUN apt-get update -y && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.13.0
+ARG VLLM_TAG=v0.14.0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -v -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \

Dockerfile.a3.openEuler

@@ -50,7 +50,7 @@ RUN yum update -y && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.13.0
+ARG VLLM_TAG=v0.14.0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \

Dockerfile.openEuler

@@ -50,7 +50,7 @@ RUN yum update -y && \
 
 # Install vLLM
 ARG VLLM_REPO=https://github.com/vllm-project/vllm.git
-ARG VLLM_TAG=v0.13.0
+ARG VLLM_TAG=v0.14.0
 RUN git clone --depth 1 $VLLM_REPO --branch $VLLM_TAG /vllm-workspace/vllm
 # In x86, triton will be installed by vllm. But in Ascend, triton doesn't work correctly. we need to uninstall it.
 RUN VLLM_TARGET_DEVICE="empty" python3 -m pip install -e /vllm-workspace/vllm/[audio] --extra-index https://download.pytorch.org/whl/cpu/ && \

docs/source/community/versioning_policy.md

@@ -53,7 +53,7 @@ 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, v0.13.0 tag | >= 3.10, < 3.12   | 8.5.0 | 2.9.0 / 2.9.0 |
+|     main    | d68209402ddab3f54a09bc1f4de9a9495a283b60, v0.14.0 tag | >= 3.10, < 3.12   | 8.5.0 | 2.9.0 / 2.9.0 |
 
 ## Release cadence
 

docs/source/conf.py

@@ -77,7 +77,7 @@ myst_substitutions = {
     # CANN image tag
     "cann_image_tag": "8.5.0-910b-ubuntu22.04-py3.11",
     # vllm version in ci
-    "ci_vllm_version": "v0.13.0",
+    "ci_vllm_version": "v0.14.0",
 }
 
 # For cross-file header anchors

tests/e2e/multicard/4-cards/long_sequence/test_mtp.py

@@ -21,7 +21,6 @@ import os
 import pytest
 
 from tests.e2e.conftest import VllmRunner
-from vllm_ascend.utils import vllm_version_is
 
 os.environ["HCCL_BUFFSIZE"] = "512"
 
@@ -51,8 +50,7 @@ def test_pcp_dcp_mtp1_eager():
         runner.generate_greedy(prompts, 32)
 
 
-@pytest.mark.skipif(
+@pytest.mark.skip(
-    not vllm_version_is('0.13.0'),
     reason="vLLM PR-32118 break this",
 )
 def test_pcp_dcp_mtp3_eager():
@@ -80,8 +78,7 @@ def test_pcp_dcp_mtp3_eager():
         runner.generate_greedy(prompts, 32)
 
 
-@pytest.mark.skipif(
+@pytest.mark.skip(
-    not vllm_version_is('0.13.0'),
     reason="vLLM PR-32118 break this",
 )
 def test_pcp_dcp_mtp3_piecewise_graph():
@@ -112,8 +109,7 @@ def test_pcp_dcp_mtp3_piecewise_graph():
         runner.generate_greedy(prompts, 32)
 
 
-@pytest.mark.skipif(
+@pytest.mark.skip(
-    not vllm_version_is('0.13.0'),
     reason="vLLM PR-32118 break this",
 )
 def test_pcp_dcp_mtp3_full_graph():

tests/ut/attention/test_mla_v1.py

@@ -17,7 +17,6 @@ from vllm_ascend.attention.mla_v1 import (AscendMLABackend,
                                           AscendMLAPrefillMetadata,
                                           ChunkedContextMetadata)
 from vllm_ascend.attention.utils import AscendCommonAttentionMetadata
-from vllm_ascend.utils import vllm_version_is
 
 
 class TestAscendMLABackend(TestBase):
@@ -477,9 +476,6 @@ class TestAscendMLAMetadataBuilderBuild(TestBase):
         self.mock_vllm_config.model_config = model_config
         self.kv_cache_spec = MagicMock()
         self.kv_cache_spec.num_layers = 32
-        if vllm_version_is('0.13.0'):
-            self.kv_cache_spec.head_size = 128
-        else:
         self.kv_cache_spec.head_size = 64
         self.kv_cache_spec.num_heads = 32
 

tests/ut/test_platform.py

@@ -5,17 +5,11 @@ import pytest
 import torch
 from vllm.config.compilation import CompilationMode, CUDAGraphMode
 from vllm.platforms import PlatformEnum
+from vllm.v1.attention.selector import AttentionSelectorConfig  # type: ignore
 
 from tests.ut.base import TestBase
 from vllm_ascend.platform import NPUPlatform
-from vllm_ascend.utils import ASCEND_QUANTIZATION_METHOD, COMPRESSED_TENSORS_METHOD, AscendDeviceType, vllm_version_is
+from vllm_ascend.utils import ASCEND_QUANTIZATION_METHOD, COMPRESSED_TENSORS_METHOD, AscendDeviceType 
-
-# isort: off
-if vllm_version_is("0.13.0"):
-    from vllm.attention.selector import AttentionSelectorConfig  # type: ignore
-else:
-    from vllm.v1.attention.selector import AttentionSelectorConfig  # type: ignore
-# isort: on
 
 
 class TestNPUPlatform(TestBase):

tests/ut/worker/test_worker_v1.py

@@ -5,7 +5,6 @@ import torch
 from vllm.config import CacheConfig, ModelConfig, ParallelConfig, ProfilerConfig, 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"
 
@@ -141,9 +140,6 @@ class TestNPUWorker(TestBase):
         )
 
         # Verify init_cached_hf_modules is called (trust_remote_code=True)
-        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")

vllm_ascend/attention/attention_v1.py

@@ -25,6 +25,18 @@ import vllm.envs as envs_vllm
 from vllm.config import VllmConfig, get_current_vllm_config
 from vllm.forward_context import ForwardContext, get_forward_context
 from vllm.utils.math_utils import cdiv
+from vllm.v1.attention.backend import (  # type: ignore
+    AttentionBackend,
+    AttentionCGSupport,
+    AttentionImpl,
+    AttentionLayer,
+    AttentionMetadataBuilder,
+    AttentionType,
+)
+from vllm.v1.attention.backends.registry import (  # type: ignore
+    AttentionBackendEnum,
+    register_backend,
+)
 from vllm.v1.core.sched.output import SchedulerOutput
 from vllm.v1.kv_cache_interface import AttentionSpec, CrossAttentionSpec
 
@@ -44,34 +56,7 @@ from vllm_ascend.compilation.acl_graph import (
 )
 from vllm_ascend.device.device_op import DeviceOperator
 from vllm_ascend.ops.flashcomm2_oshard_manager import flashcomm2_oshard_manager
-from vllm_ascend.utils import vllm_version_is, weak_ref_tensors
+from vllm_ascend.utils import weak_ref_tensors
-
-if vllm_version_is("0.13.0"):
-    from vllm.attention.backends.abstract import (  # type: ignore
-        AttentionBackend,
-        AttentionImpl,
-        AttentionLayer,
-        AttentionType,
-    )
-    from vllm.attention.backends.registry import (  # type: ignore
-        AttentionBackendEnum,
-        register_backend,
-    )
-    from vllm.v1.attention.backends.utils import AttentionCGSupport, AttentionMetadataBuilder
-else:
-    from vllm.v1.attention.backend import (  # type: ignore
-        AttentionBackend,
-        AttentionCGSupport,
-        AttentionImpl,
-        AttentionLayer,
-        AttentionMetadataBuilder,
-        AttentionType,
-    )
-    from vllm.v1.attention.backends.registry import (  # type: ignore
-        AttentionBackendEnum,
-        register_backend,
-    )
-
 
 # default max value of sliding window size
 SWA_INT_MAX = 2147483647

vllm_ascend/attention/context_parallel/attention_cp.py

@@ -29,6 +29,7 @@ from vllm.distributed import (
     get_pcp_group,
 )
 from vllm.forward_context import ForwardContext, get_forward_context
+from vllm.v1.attention.backend import AttentionCGSupport
 from vllm.v1.kv_cache_interface import AttentionSpec
 
 from vllm_ascend.attention.attention_v1 import (
@@ -49,12 +50,7 @@ from vllm_ascend.attention.utils import (
     split_decodes_and_prefills,
 )
 from vllm_ascend.compilation.acl_graph import get_graph_params, update_graph_params_workspaces
-from vllm_ascend.utils import cp_chunkedprefill_comm_stream, vllm_version_is, weak_ref_tensors
+from vllm_ascend.utils import cp_chunkedprefill_comm_stream, weak_ref_tensors
-
-if vllm_version_is("0.13.0"):
-    from vllm.v1.attention.backends.utils import AttentionCGSupport
-else:
-    from vllm.v1.attention.backend import AttentionCGSupport
 
 
 class AscendAttentionCPMetadataBuilder(AscendAttentionMetadataBuilder):

vllm_ascend/attention/context_parallel/mla_cp.py

@@ -12,6 +12,7 @@ from vllm.distributed import (
 )
 from vllm.forward_context import ForwardContext, get_forward_context
 from vllm.utils.math_utils import cdiv
+from vllm.v1.attention.backend import AttentionCGSupport
 from vllm.v1.kv_cache_interface import AttentionSpec, MLAAttentionSpec
 
 from vllm_ascend.attention.attention_v1 import AscendAttentionState
@@ -37,12 +38,7 @@ from vllm_ascend.attention.context_parallel.common_cp import (
 )
 from vllm_ascend.attention.utils import AscendCommonAttentionMetadata
 from vllm_ascend.compilation.acl_graph import get_draft_graph_params, get_graph_params, update_graph_params_workspaces
-from vllm_ascend.utils import vllm_version_is, weak_ref_tensors
+from vllm_ascend.utils import weak_ref_tensors
-
-if vllm_version_is("0.13.0"):
-    from vllm.v1.attention.backends.utils import AttentionCGSupport
-else:
-    from vllm.v1.attention.backend import AttentionCGSupport
 
 MAX_O_PROJ_PREFETCH_SIZE = 16 * 1024 * 1024
 

vllm_ascend/attention/mla_v1.py

@@ -10,7 +10,10 @@ from vllm.forward_context import ForwardContext, get_forward_context
 from vllm.logger import logger
 from vllm.model_executor.layers.linear import UnquantizedLinearMethod
 from vllm.utils.math_utils import cdiv, round_down
+from vllm.v1.attention.backend import (  # type: ignore
+    AttentionBackend, AttentionCGSupport, MLAAttentionImpl)
 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
 
 from vllm_ascend import envs
@@ -35,23 +38,12 @@ 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,
-                               vllm_version_is, weak_ref_tensors)
+                               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('0.13.0'):
-    from vllm.v1.attention.backends.utils import AttentionCGSupport
-    from vllm.attention.backends.abstract import (  # type: ignore
-        AttentionBackend, MLAAttentionImpl)
-    from vllm.attention.backends.utils import PAD_SLOT_ID  # type: ignore
-else:
-    from vllm.v1.attention.backend import (  # type: ignore
-        AttentionBackend, AttentionCGSupport, MLAAttentionImpl)
-    from vllm.v1.attention.backends.utils import PAD_SLOT_ID  # type: ignore
-# isort: on
 
 MAX_O_PROJ_PREFETCH_SIZE = 16 * 1024 * 1024
 BUILD_METADATA_STEP_PREFILL = 0

vllm_ascend/attention/sfa_v1.py

@@ -12,6 +12,8 @@ from vllm.forward_context import get_forward_context
 from vllm.logger import logger
 from vllm.model_executor.layers.linear import UnquantizedLinearMethod
 from vllm.triton_utils import HAS_TRITON
+from vllm.v1.attention.backend import (  # type: ignore
+    AttentionBackend, AttentionCGSupport, MLAAttentionImpl)
 from vllm.v1.attention.backends.mla.common import MLACommonMetadataBuilder
 from vllm.v1.kv_cache_interface import AttentionSpec
 
@@ -35,20 +37,11 @@ from vllm_ascend.ops.triton.rope import rope_forward_triton
 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, _round_up, dispose_layer,
-                               enable_dsa_cp, enable_dsa_cp_with_layer_shard, maybe_trans_nz, vllm_version_is)
+                               enable_dsa_cp, enable_dsa_cp_with_layer_shard, maybe_trans_nz)
 from vllm_ascend.worker.npu_input_batch import NPUInputBatch
 
-# isort: off
 if TYPE_CHECKING:
     from vllm.v1.core.sched.output import SchedulerOutput
-if vllm_version_is('0.13.0'):
-    from vllm.v1.attention.backends.utils import AttentionCGSupport
-    from vllm.attention.backends.abstract import (  # type: ignore
-        AttentionBackend, MLAAttentionImpl)
-else:
-    from vllm.v1.attention.backend import (  # type: ignore
-        AttentionBackend, AttentionCGSupport, MLAAttentionImpl)
-# isort: on
 
 # token count limits within bmm_transpose operator
 BMM_TRANS_MAX_SUPPORTED_TOKENS = 1024

vllm_ascend/distributed/kv_transfer/kv_p2p/mooncake_connector.py

@@ -43,14 +43,11 @@ from vllm.v1.request import RequestStatus
 from vllm_ascend.ascend_config import get_ascend_config, init_ascend_config
 from vllm_ascend.distributed.kv_transfer.utils.mooncake_transfer_engine import global_te
 from vllm_ascend.distributed.kv_transfer.utils.utils import get_transfer_timeout_value
-from vllm_ascend.utils import is_vl_model, vllm_version_is
+from vllm_ascend.utils import is_vl_model
 
 # isort: off
 if TYPE_CHECKING:
-    if vllm_version_is('0.13.0'):
+    from vllm.v1.attention.backend import AttentionMetadata  # type: ignore
-        from vllm.attention.backends.abstract import AttentionMetadata  # type: ignore
-    else:
-        from vllm.attention.backends import AttentionMetadata  # type: ignore
     from vllm.forward_context import ForwardContext
     from vllm.v1.core.kv_cache_manager import KVCacheBlocks
     from vllm.v1.request import Request

vllm_ascend/distributed/kv_transfer/kv_p2p/mooncake_layerwise_connector.py

@@ -38,14 +38,11 @@ from vllm_ascend.distributed.kv_transfer.utils.mooncake_transfer_engine import \
     global_te
 from vllm_ascend.distributed.kv_transfer.utils.utils import (
     align_memory, get_transfer_timeout_value, kv_alltoall_and_rearrange)
-from vllm_ascend.utils import npu_stream_switch, vllm_version_is
+from vllm_ascend.utils import npu_stream_switch
 
 # isort: off
 if TYPE_CHECKING:
-    if vllm_version_is('0.13.0'):
+    from vllm.v1.attention.backend import AttentionMetadata  # type: ignore
-        from vllm.attention.backends.abstract import AttentionMetadata  # type: ignore
-    else:
-        from vllm.attention.backends import AttentionMetadata  # type: ignore
     from vllm.forward_context import ForwardContext
     from vllm.v1.core.kv_cache_manager import KVCacheBlocks
     from vllm.v1.request import Request

vllm_ascend/distributed/kv_transfer/kv_pool/ascend_store/ascend_store_connector.py

@@ -9,6 +9,7 @@ from vllm.distributed.kv_transfer.kv_connector.v1.base import (
 from vllm.forward_context import ForwardContext
 from vllm.logger import logger
 from vllm.utils.network_utils import make_zmq_socket
+from vllm.v1.attention.backend import AttentionMetadata  # type: ignore
 from vllm.v1.core.kv_cache_manager import KVCacheBlocks
 from vllm.v1.core.sched.output import SchedulerOutput
 from vllm.v1.kv_cache_interface import KVCacheConfig
@@ -19,14 +20,6 @@ from vllm_ascend.distributed.kv_transfer.kv_pool.ascend_store.pool_scheduler imp
     KVPoolScheduler, get_zmq_rpc_path_lookup)
 from vllm_ascend.distributed.kv_transfer.kv_pool.ascend_store.pool_worker import \
     KVPoolWorker
-from vllm_ascend.utils import vllm_version_is
-
-# isort: off
-if vllm_version_is('0.13.0'):
-    from vllm.attention.backends.abstract import AttentionMetadata  # type: ignore
-else:
-    from vllm.v1.attention.backend import AttentionMetadata  # type: ignore
-# isort: on
 
 
 class AscendStoreConnector(KVConnectorBase_V1):

vllm_ascend/distributed/kv_transfer/kv_pool/cpu_offload/cpu_offload_connector.py

@@ -24,25 +24,14 @@ from vllm.v1.kv_cache_interface import FullAttentionSpec, KVCacheSpec
 
 from vllm_ascend.distributed.kv_transfer.kv_pool.cpu_offload.metadata import (
     MetadataServer, MetadataServerProc, MLAConfig)
-from vllm_ascend.utils import vllm_version_is
 
-# isort: off
-if vllm_version_is('0.13.0'):
-    from vllm.attention.backends.abstract import AttentionType  # type: ignore
-else:
-    from vllm.v1.attention.backend import AttentionType  # type: ignore
 
 if TYPE_CHECKING:
-    if vllm_version_is('0.13.0'):
+    from vllm.v1.attention.backend import AttentionMetadata  #type: ignore
-        from vllm.attention.backends.abstract import \
-            AttentionMetadata  # type: ignore
-    else:
-        from vllm.v1.attention.backend import AttentionType  #type: ignore
     from vllm.forward_context import ForwardContext
     from vllm.v1.core.kv_cache_manager import KVCacheBlocks
     from vllm.v1.kv_cache_interface import KVCacheConfig
     from vllm.v1.request import Request
-# isort: on
 
 
 @dataclass

vllm_ascend/distributed/kv_transfer/kv_pool/ucm_connector.py

@@ -9,15 +9,11 @@ from vllm.distributed.kv_transfer.kv_connector.v1.base import (
 from vllm.logger import init_logger
 from vllm.v1.core.sched.output import SchedulerOutput
 
-from vllm_ascend.utils import vllm_version_is
 
 logger = init_logger(__name__)
 
 # isort: off
 if TYPE_CHECKING:
-    if vllm_version_is('0.13.0'):
-        from vllm.attention.backends.abstract import AttentionMetadata  # type: ignore
-    else:
     from vllm.v1.attention.backend import AttentionMetadata  # type: ignore
     from vllm.distributed.kv_transfer.kv_connector.v1.metrics import (
         KVConnectorPromMetrics, KVConnectorStats, PromMetric, PromMetricT)

vllm_ascend/kv_offload/cpu_npu.py

@@ -2,19 +2,11 @@ import numpy as np
 import torch
 from vllm.logger import init_logger
 from vllm.utils.platform_utils import is_pin_memory_available
+from vllm.v1.attention.backend import AttentionBackend  # type: ignore
 from vllm.v1.kv_offload.mediums import CPULoadStoreSpec, GPULoadStoreSpec
 from vllm.v1.kv_offload.worker.worker import (OffloadingHandler,
                                               TransferResult, TransferSpec)
 
-from vllm_ascend.utils import vllm_version_is
-
-# isort: off
-if vllm_version_is('0.13.0'):
-    from vllm.attention.backends.abstract import AttentionBackend  # type: ignore
-else:
-    from vllm.v1.attention.backend import AttentionBackend  # type: ignore
-# isort: on
-
 logger = init_logger(__name__)
 
 

vllm_ascend/kv_offload/npu.py

@@ -3,6 +3,7 @@ from typing import Optional
 
 import torch
 from vllm.config import VllmConfig
+from vllm.v1.attention.backend import AttentionBackend  # type: ignore
 from vllm.v1.kv_offload.abstract import LoadStoreSpec, OffloadingManager
 from vllm.v1.kv_offload.backends.cpu import CPUBackend
 from vllm.v1.kv_offload.lru_manager import LRUOffloadingManager
@@ -12,14 +13,6 @@ from vllm.v1.kv_offload.worker.worker import OffloadingHandler
 from vllm.v1.kv_cache_interface import KVCacheConfig
 
 from vllm_ascend.kv_offload.cpu_npu import CpuNpuOffloadingHandler
-from vllm_ascend.utils import vllm_version_is
-
-# isort: off
-if vllm_version_is('0.13.0'):
-    from vllm.attention.backends.abstract import AttentionBackend  # type: ignore
-else:
-    from vllm.v1.attention.backend import AttentionBackend  # type: ignore
-# isort: on
 
 
 class NPUOffloadingSpec(OffloadingSpec):
@@ -27,9 +20,6 @@ class NPUOffloadingSpec(OffloadingSpec):
     def __init__(self,
                  vllm_config: VllmConfig,
                  kv_cache_config: Optional[KVCacheConfig] = None):
-        if vllm_version_is('0.13.0'):
-            super().__init__(vllm_config)
-        else:
         super().__init__(vllm_config, kv_cache_config)
 
         num_cpu_blocks = self.extra_config.get("num_cpu_blocks")

vllm_ascend/ops/fused_moe/fused_moe.py

@@ -50,7 +50,7 @@ from vllm_ascend.quantization.w8a8_dynamic import \
 from vllm_ascend.utils import (AscendDeviceType, enable_sp,
                                get_ascend_device_type, maybe_trans_nz,
                                npu_stream_switch, shared_expert_dp_enabled,
-                               shared_experts_calculation_stream, vllm_version_is)
+                               shared_experts_calculation_stream)
 
 @dataclass
 class FusedMoEResult:
@@ -451,11 +451,6 @@ class AscendSharedFusedMoE(SharedFusedMoE, AscendFusedMoE):
         # Qwen3-Next specific gating mechanism
         if hasattr(self._shared_experts, "expert_gate") and \
             self._shared_experts.expert_gate is not None:
-            if vllm_version_is('0.13.0'):
-                # TODO(jianzs): remove this branch after vLLM new version is
-                # released
-                gate_out = self._shared_experts.expert_gate(hidden_states)  # type: ignore
-            else:
             gate_out, _ = self._shared_experts.expert_gate(hidden_states)  # type: ignore
             shared_out = F.sigmoid(gate_out) * shared_out
         return shared_out

vllm_ascend/ops/mla.py

@@ -31,16 +31,9 @@ from vllm.model_executor.layers.mla import (MLAModules,
                                             MultiHeadLatentAttentionWrapper)
 from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.utils.torch_utils import direct_register_custom_op
+from vllm.v1.attention.backend import AttentionMetadata  # type: ignore
 
 from vllm_ascend.ascend_config import get_ascend_config
-from vllm_ascend.utils import vllm_version_is
-
-# isort: off
-if vllm_version_is('0.13.0'):
-    from vllm.attention.backends.abstract import AttentionMetadata  # type: ignore
-else:
-    from vllm.v1.attention.backend import AttentionMetadata  # type: ignore
-# isort: on
 
 
 class IndexerWrapper(nn.Module):

vllm_ascend/ops/mm_encoder_attention.py

@@ -20,16 +20,10 @@ import torch
 import torch.nn.functional as F
 import torch_npu
 from vllm.config import MultiModalConfig
+from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention  # type: ignore
 
 import vllm_ascend.envs as envs_ascend
-from vllm_ascend.utils import vllm_version_is
 
-# isort: off
-if vllm_version_is('0.13.0'):
-    from vllm.attention.layers.mm_encoder_attention import MMEncoderAttention  # type: ignore
-else:
-    from vllm.model_executor.layers.attention.mm_encoder_attention import MMEncoderAttention  # type: ignore
-# isort: on
 
 MIN_PAD_SIZE = 64  # min_size to pad weight
 MAX_PAD_SIZE = 128  # max_size to pad weight

vllm_ascend/ops/rotary_embedding.py

@@ -31,8 +31,7 @@ if HAS_TRITON:
 
 from vllm_ascend.platform import NPUPlatform
 from vllm_ascend.utils import (AscendDeviceType, enable_custom_op,
-                               get_ascend_device_type, has_rope, is_vl_model,
+                               get_ascend_device_type, has_rope, is_vl_model)
-                               vllm_version_is)
 
 # Currently, rope ops used on npu requires detached cos && sin as inputs.
 # However, RotaryEmbedding in vllm use cos_sin_cache as a whole variable.
@@ -637,16 +636,6 @@ class AscendApplyRotaryEmb(ApplyRotaryEmb):
         cos: torch.Tensor,
         sin: torch.Tensor,
     ) -> torch.Tensor:
-        if vllm_version_is('0.13.0'):
-            origin_shape = x.shape
-            origin_dtype = x.dtype
-            if len(origin_shape) == 3:
-                x = x.unsqueeze(0)
-            if self.enable_fp32_compute:
-                x = x.float()
-                cos = cos.float()
-                sin = sin.float()
-        else:
         x, cos, sin, origin_shape, origin_dtype = self._pre_process(
             x, cos, sin)
 
@@ -660,12 +649,6 @@ class AscendApplyRotaryEmb(ApplyRotaryEmb):
 
         output = torch_npu.npu_rotary_mul(x, cos, sin)
 
-        if vllm_version_is('0.13.0'):
-            if len(origin_shape) == 3:
-                output = output.squeeze(0)
-            if self.enable_fp32_compute:
-                output = output.to(origin_dtype)
-        else:
         output = self._post_process(output, origin_shape, origin_dtype)
 
         return output

vllm_ascend/ops/triton/mamba/causal_conv1d.py

@@ -14,14 +14,7 @@ import torch.nn.functional as F
 import triton
 import triton.language as tl
 
-from vllm_ascend.utils import vllm_version_is
-
-# isort: off
-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
-# isort: on
 
 
 def causal_conv1d_ref(

vllm_ascend/patch/platform/__init__.py

@@ -27,5 +27,5 @@ if os.getenv("DYNAMIC_EPLB", "false").lower() in ("true", "1") or os.getenv(
         "EXPERT_MAP_RECORD", "false") == "true":
     import vllm_ascend.patch.platform.patch_multiproc_executor  # noqa
 
-if envs.VLLM_ASCEND_BALANCE_SCHEDULING and vllm_version_is('0.13.0'):
+if envs.VLLM_ASCEND_BALANCE_SCHEDULING and vllm_version_is('0.14.0'):
     import vllm_ascend.patch.platform.patch_balance_schedule  # noqa

vllm_ascend/patch/worker/patch_qwen3_next.py

@@ -28,6 +28,7 @@ from vllm.model_executor.layers.mamba.ops.causal_conv1d import (
 from vllm.model_executor.models.qwen3_next import (Qwen3NextGatedDeltaNet,
                                                    fused_gdn_gating)
 from vllm.triton_utils import triton
+from vllm.v1.attention.backend import AttentionMetadata  # type: ignore
 from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadata
 
 from vllm_ascend.ops.triton.fla.fused_qkvzba_split_reshape import \
@@ -35,14 +36,6 @@ from vllm_ascend.ops.triton.fla.fused_qkvzba_split_reshape import \
 from vllm_ascend.ops.triton.fla.sigmoid_gating import \
     fused_sigmoid_gating_delta_rule_update
 from vllm_ascend.ops.triton.fused_gdn_gating import fused_gdn_gating_patch
-from vllm_ascend.utils import vllm_version_is
-
-# isort: off
-if vllm_version_is('0.13.0'):
-    from vllm.attention.backends.abstract import AttentionMetadata  # type: ignore
-else:
-    from vllm.v1.attention.backend import AttentionMetadata  # type: ignore
-# isort: on
 
 
 class AscendQwen3Next_GatedDeltaNet(nn.Module, MambaBase):

vllm_ascend/worker/model_runner_v1.py

@@ -50,8 +50,10 @@ from vllm.sequence import IntermediateTensors
 from vllm.utils.import_utils import LazyLoader
 from vllm.utils.math_utils import cdiv
 from vllm.utils.mem_utils import DeviceMemoryProfiler
+from vllm.v1.attention.backend import AttentionBackend, AttentionType  # type: ignore
 from vllm.v1.attention.backends.gdn_attn import GDNAttentionMetadataBuilder
 from vllm.v1.attention.backends.utils import CommonAttentionMetadata
+from vllm.v1.attention.selector import get_attn_backend  # type: ignore
 from vllm.v1.core.sched.output import SchedulerOutput
 from vllm.v1.kv_cache_interface import (AttentionSpec,
                                         EncoderOnlyAttentionSpec,
@@ -102,7 +104,7 @@ from vllm_ascend.spec_decode.mtp_proposer import MtpProposer
 from vllm_ascend.utils import (AscendDeviceType, ProfileExecuteDuration,
                                enable_sp, get_ascend_device_type, is_moe_model,
                                lmhead_tp_enable, maybe_trans_nz,
-                               set_weight_prefetch_method, vllm_version_is)
+                               set_weight_prefetch_method)
 from vllm_ascend.worker.npu_input_batch import NPUInputBatch
 from vllm_ascend.worker.pcp_utils import PCPManager
 
@@ -115,15 +117,6 @@ if TYPE_CHECKING:
 else:
     xgr = LazyLoader("xgr", globals(), "xgrammar")
 
-# isort: off
-if vllm_version_is('0.13.0'):
-    from vllm.attention.backends.abstract import (  # type: ignore
-        AttentionBackend, AttentionType)
-    from vllm.attention.selector import get_attn_backend  # type: ignore
-else:
-    from vllm.v1.attention.selector import get_attn_backend  # type: ignore
-    from vllm.v1.attention.backend import AttentionBackend, AttentionType  # type: ignore
-# isort: on
 import torch_npu
 
 # if true, allow tensor initialization and casting with internal format (e.g., NZ)
@@ -746,9 +739,6 @@ class NPUModelRunner(GPUModelRunner):
 
         # _prepare_inputs may reorder the batch, so we must gather
         # multi-modal outputs after that to ensure the correct order
-        if vllm_version_is('0.13.0'):
-            model_kwargs = self._init_model_kwargs(num_input_tokens)
-        else:
         model_kwargs = self._init_model_kwargs()
         if self.is_multimodal_model and not self.model_config.is_encoder_decoder:
             self.multimodal_cpu_fields = ["grid_thw"]
@@ -1575,12 +1565,6 @@ class NPUModelRunner(GPUModelRunner):
                 logits = None
             else:
                 if self.input_batch.pooling_params:
-                    if vllm_version_is('0.13.0'):
-                        pool_output = self._pool(
-                            hidden_states,
-                            scheduler_output.total_num_scheduled_tokens,
-                            num_scheduled_tokens_np)
-                    else:
                     pool_output = self._pool(
                         hidden_states,
                         scheduler_output.total_num_scheduled_tokens,
@@ -1675,7 +1659,6 @@ 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)
 
         (
@@ -1826,14 +1809,6 @@ class NPUModelRunner(GPUModelRunner):
                     valid_sampled_token_ids[int(i)].clear()
             else:
                 # Includes spec decode tokens.
-                if vllm_version_is('0.13.0'):
-                    valid_sampled_token_ids, cu_num_tokens = RejectionSampler.parse_output(
-                        sampled_token_ids,
-                        self.input_batch.vocab_size,
-                        discard_sampled_tokens_req_indices,
-                        return_cu_num_tokens=logprobs_tensors is not None,
-                    )
-                else:
                 valid_sampled_token_ids, cu_num_tokens = RejectionSampler.parse_output(
                     sampled_token_ids,
                     self.input_batch.vocab_size,

vllm_ascend/worker/worker.py

@@ -58,15 +58,12 @@ from vllm_ascend.distributed.parallel_state import init_ascend_model_parallel
 from vllm_ascend.ops.triton.triton_utils import init_device_properties_triton
 from vllm_ascend.utils import (AscendDeviceType, check_ascend_device_type,
                                enable_sp, get_ascend_device_type,
-                               register_ascend_customop, vllm_version_is)
+                               register_ascend_customop)
 from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
 
 torch._dynamo.trace_rules.clear_lru_cache()  # noqa: E402
 from torch._dynamo.variables import TorchInGraphFunctionVariable  # noqa: E402
 
-if vllm_version_is("0.13.0"):
-    from vllm.model_executor.utils import set_random_seed
-else:
 from vllm.utils.torch_utils import set_random_seed
 
 torch_non_c_binding_in_graph_functions_npu = dict.fromkeys(
@@ -121,13 +118,6 @@ class NPUWorker(WorkerBase):
             self.cache_dtype = STR_DTYPE_TO_TORCH_DTYPE[
                 self.cache_config.cache_dtype]
 
-        if vllm_version_is('0.13.0'):
-            if self.model_config.trust_remote_code:
-                # note: lazy import to avoid importing torch before initializing
-                from vllm.utils.import_utils import 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:
             # Buffers saved before sleep