[Main2Main] Upgrade vllm commit to 0113 (#5839)

### What this PR does / why we need it? Upgrade vllm commit to 0113 (11b6af5280d6d6dfb8953af16e67b25f819b3be9) - Modify import paths due to the refactors https://github.com/vllm-project/vllm/pull/31916 https://github.com/vllm-project/vllm/pull/32054 - Fix `TypeError: NPUOffloadingSpec.__init__() takes 2 positional arguments but 3 were given` due to https://github.com/vllm-project/vllm/pull/24498 - Skip the async-scheduling tests in `tests/e2e/multicard/4-cards/long_sequence/test_mtp.py`, which are never verified https://github.com/vllm-project/vllm/pull/31998 - Skip some pooling tests, which are caused by https://github.com/vllm-project/vllm/pull/32148 where vllm is also failed https://buildkite.com/vllm/ci/builds/46705/steps/canvas?jid=019bb329-3834-4685-862b-1613b8e0f5d4 We will reopen those tests when main2main reachs https://github.com/vllm-project/vllm/pull/32243 - Skip some cases in `tests/e2e/multicard/4-cards/long_sequence/test_mtp.py`, which are broken by https://github.com/vllm-project/vllm/pull/32118 ### Does this PR introduce _any_ user-facing change? ### How was this patch tested? - vLLM version: v0.13.0 - vLLM main: 2f4e6548ef Signed-off-by: wjunLu <wjunlu217@gmail.com> Signed-off-by: hfadzxy <starmoon_zhang@163.com> Co-authored-by: hfadzxy <starmoon_zhang@163.com>
2026-01-15 09:48:53 +08:00
parent e67608041d
commit c11a05c4e1
29 changed files with 229 additions and 54 deletions
--- a/vllm_ascend/attention/attention_v1.py
+++ b/vllm_ascend/attention/attention_v1.py
@@ -22,15 +22,9 @@ from typing import ClassVar, List, Optional, Tuple, Type
 import torch
 import torch_npu
 import vllm.envs as envs_vllm
-from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
-                                              AttentionLayer, AttentionType)
-from vllm.attention.backends.registry import (AttentionBackendEnum,
-                                              register_backend)
 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.backends.utils import (AttentionCGSupport,
-                                              AttentionMetadataBuilder)
 from vllm.v1.core.sched.output import SchedulerOutput
 from vllm.v1.kv_cache_interface import AttentionSpec, CrossAttentionSpec

@@ -45,7 +39,23 @@ from vllm_ascend.compilation.acl_graph import (
    update_draft_graph_params_workspaces, update_graph_params_workspaces)
 from vllm_ascend.device.device_op import DeviceOperator
 from vllm_ascend.ops.flashcomm2_oshard_manager import flashcomm2_oshard_manager
-from vllm_ascend.utils import weak_ref_tensors
+from vllm_ascend.utils import vllm_version_is, weak_ref_tensors
+
+# isort: off
+if vllm_version_is('0.13.0'):
+    from vllm.v1.attention.backends.utils import (AttentionCGSupport,
+                                                  AttentionMetadataBuilder)
+    from vllm.attention.backends.abstract import (  # type: ignore
+        AttentionBackend, AttentionImpl, AttentionLayer, AttentionType)
+    from vllm.attention.backends.registry import (  # type: ignore
+        AttentionBackendEnum, register_backend)
+else:
+    from vllm.v1.attention.backend import (  # type: ignore
+        AttentionBackend, AttentionCGSupport, AttentionImpl, AttentionLayer,
+        AttentionType, AttentionMetadataBuilder)
+    from vllm.v1.attention.backends.registry import (  # type: ignore
+        AttentionBackendEnum, register_backend)
+# isort: on

 # default max value of sliding window size
 SWA_INT_MAX = 2147483647
--- a/vllm_ascend/attention/context_parallel/attention_cp.py
+++ b/vllm_ascend/attention/context_parallel/attention_cp.py
@@ -27,7 +27,6 @@ from vllm.distributed import (get_dcp_group,
                              get_decode_context_model_parallel_world_size,
                              get_pcp_group)
 from vllm.forward_context import ForwardContext, get_forward_context
-from vllm.v1.attention.backends.utils import AttentionCGSupport
 from vllm.v1.kv_cache_interface import AttentionSpec

 from vllm_ascend.attention.attention_v1 import (AscendAttentionBackendImpl,
@@ -41,7 +40,13 @@ from vllm_ascend.attention.utils import (AscendCommonAttentionMetadata,
                                         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, weak_ref_tensors
+from vllm_ascend.utils import (cp_chunkedprefill_comm_stream, vllm_version_is,
+                               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):
--- a/vllm_ascend/attention/context_parallel/mla_cp.py
+++ b/vllm_ascend/attention/context_parallel/mla_cp.py
@@ -10,7 +10,6 @@ from vllm.distributed import (get_dcp_group,
                              get_pcp_group)
 from vllm.forward_context import ForwardContext, get_forward_context
 from vllm.utils.math_utils import cdiv
-from vllm.v1.attention.backends.utils import AttentionCGSupport
 from vllm.v1.kv_cache_interface import AttentionSpec, MLAAttentionSpec

 # isort: off
@@ -28,7 +27,12 @@ from vllm_ascend.attention.context_parallel.common_cp import (
 from vllm_ascend.compilation.acl_graph import (get_draft_graph_params,
                                               get_graph_params,
                                               update_graph_params_workspaces)
-from vllm_ascend.utils import weak_ref_tensors
+from vllm_ascend.utils import weak_ref_tensors, vllm_version_is
+
+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

--- a/vllm_ascend/attention/mla_v1.py
+++ b/vllm_ascend/attention/mla_v1.py
@@ -5,14 +5,12 @@ import numpy as np
 import torch
 import torch_npu
 import vllm.envs as envs_vllm
-from vllm.attention.backends.abstract import AttentionBackend, MLAAttentionImpl
 from vllm.config import VllmConfig, get_current_vllm_config
 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.backends.mla.common import MLACommonMetadataBuilder
-from vllm.v1.attention.backends.utils import AttentionCGSupport
 from vllm.v1.kv_cache_interface import AttentionSpec, MLAAttentionSpec

 from vllm_ascend import envs
@@ -44,10 +42,17 @@ 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
--- a/vllm_ascend/attention/sfa_v1.py
+++ b/vllm_ascend/attention/sfa_v1.py
@@ -5,7 +5,6 @@ import torch
 import torch_npu
 import vllm.envs as envs_vllm
 from torch import nn
-from vllm.attention.backends.abstract import AttentionBackend, MLAAttentionImpl
 from vllm.config import CUDAGraphMode, VllmConfig, get_current_vllm_config
 from vllm.distributed import get_tensor_model_parallel_world_size, get_tp_group
 from vllm.forward_context import get_forward_context
@@ -13,7 +12,6 @@ from vllm.logger import logger
 from vllm.model_executor.layers.linear import UnquantizedLinearMethod
 from vllm.triton_utils import HAS_TRITON
 from vllm.v1.attention.backends.mla.common import MLACommonMetadataBuilder
-from vllm.v1.attention.backends.utils import AttentionCGSupport
 from vllm.v1.kv_cache_interface import AttentionSpec

 from vllm_ascend import envs
@@ -34,11 +32,20 @@ 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, maybe_trans_nz)
+                               enable_dsa_cp, maybe_trans_nz, vllm_version_is)
 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


 class AscendSFABackend(AttentionBackend):