EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

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

Browse Source 
### 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>
This commit is contained in:
wjunLu
2026-01-15 09:48:53 +08:00
committed by GitHub
parent e67608041d
commit c11a05c4e1
29 changed files with 229 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
vllm_ascend/attention/attention_v1.py
View File

@@ -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

9
vllm_ascend/attention/context_parallel/attention_cp.py
View File

@@ -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):

8
vllm_ascend/attention/context_parallel/mla_cp.py
View File

@@ -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

9
vllm_ascend/attention/mla_v1.py
View File

@@ -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

13
vllm_ascend/attention/sfa_v1.py
View File

@@ -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):

9
vllm_ascend/distributed/kv_transfer/kv_p2p/mooncake_connector.py
View File

@@ -43,13 +43,18 @@ 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
from vllm_ascend.utils import is_vl_model, vllm_version_is
# isort: off
if TYPE_CHECKING:
from vllm.attention.backends.abstract import AttentionMetadata
if vllm_version_is('0.13.0'):
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
# isort: on
GET_META_MSG = b"get_meta_msg"
DONE_RECVING_MSG = b"done_recving_msg"

9
vllm_ascend/distributed/kv_transfer/kv_p2p/mooncake_layerwise_connector.py
View File

@@ -38,13 +38,18 @@ 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
from vllm_ascend.utils import npu_stream_switch, vllm_version_is
# isort: off
if TYPE_CHECKING:
from vllm.attention.backends.abstract import AttentionMetadata
if vllm_version_is('0.13.0'):
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
# isort: on
DONE_SENDING_MSG = b"done_sending_msg"

9
vllm_ascend/distributed/kv_transfer/kv_pool/ascend_store/ascend_store_connector.py
View File

@@ -3,7 +3,6 @@ from typing import Any, Optional
import torch
import zmq
from vllm.attention.backends.abstract import AttentionMetadata
from vllm.config import VllmConfig
from vllm.distributed.kv_transfer.kv_connector.v1.base import (
KVConnectorBase_V1, KVConnectorMetadata, KVConnectorRole)
@@ -20,6 +19,14 @@ 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):

15
vllm_ascend/distributed/kv_transfer/kv_pool/cpu_offload/cpu_offload_connector.py
View File

@@ -9,7 +9,6 @@ from dataclasses import dataclass
from typing import TYPE_CHECKING, Any, Optional, Sequence
import torch
from vllm.attention.backends.abstract import AttentionType
from vllm.attention.layer import Attention, MLAAttention
from vllm.config import VllmConfig, get_layers_from_vllm_config
from vllm.distributed.ec_transfer import get_ec_transfer, has_ec_transfer
@@ -26,13 +25,25 @@ from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheSpec,
from vllm_ascend.distributed.kv_transfer.kv_pool.ascend_store.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:
from vllm.attention.backends.abstract import AttentionMetadata
if vllm_version_is('0.13.0'):
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

9
vllm_ascend/distributed/kv_transfer/ucm_connector.py
View File

@@ -9,16 +9,23 @@ 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:
from vllm.attention.backends.abstract import AttentionMetadata
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)
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
class UCMConnectorV1(KVConnectorBase_V1):

20
vllm_ascend/kv_offload/cpu_npu.py
View File

@@ -1,12 +1,20 @@
import numpy as np
import torch
from vllm.attention.backends.abstract import AttentionBackend
from vllm.logger import init_logger
from vllm.utils.platform_utils import is_pin_memory_available
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__)
@@ -166,3 +174,13 @@ class CpuNpuOffloadingHandler(OffloadingHandler):
for job_id, _ in results:
del self.transfer_events[job_id]
return results
def wait(self, job_ids: set[int]) -> None:
"""
Wait (block) until all specified transfer jobs are completed.
"""
for job_id in job_ids:
event = self.transfer_events.get(job_id)
if event is not None:
# This will block until the NPU event is complete
event.synchronize()

19
vllm_ascend/kv_offload/npu.py
View File

@@ -2,7 +2,6 @@ from collections.abc import Iterator
from typing import Optional
import torch
from vllm.attention.backends.abstract import AttentionBackend
from vllm.config import VllmConfig
from vllm.v1.kv_offload.abstract import LoadStoreSpec, OffloadingManager
from vllm.v1.kv_offload.backends.cpu import CPUBackend
@@ -10,14 +9,28 @@ from vllm.v1.kv_offload.lru_manager import LRUOffloadingManager
from vllm.v1.kv_offload.mediums import CPULoadStoreSpec, GPULoadStoreSpec
from vllm.v1.kv_offload.spec import OffloadingSpec
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):
def __init__(self, vllm_config: VllmConfig):
super().__init__(vllm_config)
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")
if not num_cpu_blocks:

9
vllm_ascend/ops/mla.py
View File

@@ -23,7 +23,6 @@ from typing import Optional
import torch
from torch import nn
from vllm.attention.backends.abstract import AttentionMetadata
from vllm.attention.layer import MLAAttention
from vllm.config import CacheConfig, get_current_vllm_config
from vllm.distributed import get_tensor_model_parallel_world_size
@@ -34,6 +33,14 @@ from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.utils.torch_utils import direct_register_custom_op
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):

9
vllm_ascend/ops/mm_encoder_attention.py
View File

@@ -19,10 +19,17 @@ import einops
import torch
import torch.nn.functional as F
import torch_npu
from vllm.attention.layers.mm_encoder_attention import MMEncoderAttention
from vllm.config import MultiModalConfig
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

2
vllm_ascend/ops/triton/mamba/causal_conv1d.py
View File

@@ -16,10 +16,12 @@ 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(

9
vllm_ascend/patch/worker/patch_qwen3_next.py
View File

@@ -18,7 +18,6 @@
import torch
from einops import rearrange
from torch import nn
from vllm.attention.backends.abstract import AttentionMetadata
from vllm.config import CUDAGraphMode
from vllm.forward_context import get_forward_context
from vllm.model_executor.layers.fla.ops import (
@@ -36,6 +35,14 @@ 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):

31
vllm_ascend/worker/model_runner_v1.py
View File

@@ -30,9 +30,7 @@ import numpy as np
import torch
import torch.distributed as dist
import torch.nn as nn
from vllm.attention.backends.abstract import AttentionBackend, AttentionType
from vllm.attention.layer import Attention, MLAAttention
from vllm.attention.selector import get_attn_backend
from vllm.config import (CompilationMode, CUDAGraphMode, VllmConfig,
get_layers_from_vllm_config)
from vllm.distributed import (get_tensor_model_parallel_world_size,
@@ -119,6 +117,15 @@ 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)
@@ -1817,12 +1824,20 @@ class NPUModelRunner(GPUModelRunner):
valid_sampled_token_ids[int(i)].clear()
else:
# Includes spec decode tokens.
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,
)
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,
discard_sampled_tokens_req_indices,
logprobs_tensors=logprobs_tensors,
)
else:
valid_sampled_token_ids = []
invalid_req_indices = discard_sampled_tokens_req_indices.tolist()