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

Drop vLLM 0.13.0 support (#6069)

Browse Source 
### 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>
This commit is contained in:
zhangxinyuehfad
2026-01-23 09:45:08 +08:00
committed by GitHub
parent 27a513b672
commit 819a4459ce
39 changed files with 86 additions and 272 deletions
Download Patch File Download Diff File Expand all files Collapse all files

41
vllm_ascend/attention/attention_v1.py
View File

@@ -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
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,
)
from vllm_ascend.utils import weak_ref_tensors
# default max value of sliding window size
SWA_INT_MAX = 2147483647

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

@@ -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
if vllm_version_is("0.13.0"):
from vllm.v1.attention.backends.utils import AttentionCGSupport
else:
from vllm.v1.attention.backend import AttentionCGSupport
from vllm_ascend.utils import cp_chunkedprefill_comm_stream, weak_ref_tensors
class AscendAttentionCPMetadataBuilder(AscendAttentionMetadataBuilder):

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

@@ -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
if vllm_version_is("0.13.0"):
from vllm.v1.attention.backends.utils import AttentionCGSupport
else:
from vllm.v1.attention.backend import AttentionCGSupport
from vllm_ascend.utils import weak_ref_tensors
MAX_O_PROJ_PREFETCH_SIZE = 16 * 1024 * 1024

16
vllm_ascend/attention/mla_v1.py
View File

@@ -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

13
vllm_ascend/attention/sfa_v1.py
View File

@@ -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

7
vllm_ascend/distributed/kv_transfer/kv_p2p/mooncake_connector.py
View File

@@ -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.attention.backends.abstract import AttentionMetadata # type: ignore
else:
from vllm.attention.backends import AttentionMetadata # type: ignore
from vllm.v1.attention.backend 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

7
vllm_ascend/distributed/kv_transfer/kv_p2p/mooncake_layerwise_connector.py
View File

@@ -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.attention.backends.abstract import AttentionMetadata # type: ignore
else:
from vllm.attention.backends import AttentionMetadata # type: ignore
from vllm.v1.attention.backend 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

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

@@ -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):

13
vllm_ascend/distributed/kv_transfer/kv_pool/cpu_offload/cpu_offload_connector.py
View File

@@ -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.attention.backends.abstract import \
AttentionMetadata # type: ignore
else:
from vllm.v1.attention.backend import AttentionType #type: ignore
from vllm.v1.attention.backend import AttentionMetadata #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

6
vllm_ascend/distributed/kv_transfer/kv_pool/ucm_connector.py
View File

@@ -9,16 +9,12 @@ 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.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

10
vllm_ascend/kv_offload/cpu_npu.py
View File

@@ -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__)

14
vllm_ascend/kv_offload/npu.py
View File

@@ -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):
@@ -27,10 +20,7 @@ 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)
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/fused_moe/fused_moe.py
View File

@@ -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:
@@ -451,12 +451,7 @@ 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
gate_out, _ = self._shared_experts.expert_gate(hidden_states) # type: ignore
shared_out = F.sigmoid(gate_out) * shared_out
return shared_out

9
vllm_ascend/ops/mla.py
View File

@@ -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):

8
vllm_ascend/ops/mm_encoder_attention.py
View File

@@ -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

25
vllm_ascend/ops/rotary_embedding.py
View File

@@ -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,
vllm_version_is)
get_ascend_device_type, has_rope, is_vl_model)
# Currently, rope ops used on npu requires detached cos && sin as inputs.
# However, RotaryEmbedding in vllm use cos_sin_cache as a whole variable.
@@ -637,18 +636,8 @@ 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)
x, cos, sin, origin_shape, origin_dtype = self._pre_process(
x, cos, sin)
head_dim = x.shape[-1]
# cos, sin: [seq_len, head_dim // 2]
@@ -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)
output = self._post_process(output, origin_shape, origin_dtype)
return output

9
vllm_ascend/ops/triton/mamba/causal_conv1d.py
View File

@@ -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
from vllm.v1.attention.backends.utils import PAD_SLOT_ID # type: ignore
def causal_conv1d_ref(

2
vllm_ascend/patch/platform/__init__.py
View File

@@ -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

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

@@ -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):

55
vllm_ascend/worker/model_runner_v1.py
View File

@@ -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)
@@ -746,10 +739,7 @@ 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()
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"]
self._prepare_multimodal_fields()
@@ -1575,16 +1565,10 @@ 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,
num_scheduled_tokens_np, kv_connector_output)
pool_output = self._pool(
hidden_states,
scheduler_output.total_num_scheduled_tokens,
num_scheduled_tokens_np, kv_connector_output)
if self.debugger is not None:
self.debugger.stop()
self.debugger.step()
@@ -1675,8 +1659,7 @@ 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)
self._copy_draft_token_ids_to_cpu(scheduler_output)
(
logprobs_lists,
@@ -1826,20 +1809,12 @@ 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,
discard_sampled_tokens_req_indices,
logprobs_tensors=logprobs_tensors,
)
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()

14
vllm_ascend/worker/worker.py
View File

@@ -58,16 +58,13 @@ 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
from vllm.utils.torch_utils import set_random_seed
torch_non_c_binding_in_graph_functions_npu = dict.fromkeys(
["torch.npu.current_stream"],
@@ -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