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

[Refactor] Rename cudagraph_support to aclgraph_support (#3104)

Browse Source 
### What this PR does / why we need it?
Updates the `cudagraph_support` attribute to `aclgraph_support` to use
terminology appropriate for the Ascend platform (ACL graphs instead of
CUDA graphs).

This change also explicitly disables graph support for the MLA attention
backend.

### Does this PR introduce _any_ user-facing change?
None.

### How was this patch tested?
None needed.

- vLLM version: v0.10.2
- vLLM main:
5aeb925452

Signed-off-by: Yizhou Liu <liu_yizhou@outlook.com>
This commit is contained in:
Yizhou
2025-09-23 11:30:31 +08:00
committed by GitHub
parent d2399ab97b
commit 39a85c49fa
3 changed files with 10 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm_ascend/attention/attention_v1.py
View File

@@ -199,8 +199,8 @@ class AscendMetadata:
class AscendAttentionMetadataBuilder: class AscendAttentionMetadataBuilder:
# Does this backend/builder support CUDA Graphs for attention (default: no). # Does this backend/builder support ACL Graphs for attention (default: no).
cudagraph_support: ClassVar[AttentionCGSupport] = \ aclgraph_support: ClassVar[AttentionCGSupport] = \
AttentionCGSupport.UNIFORM_SINGLE_TOKEN_DECODE AttentionCGSupport.UNIFORM_SINGLE_TOKEN_DECODE
# Does this backend/builder reorder the batch? # Does this backend/builder reorder the batch?
# If not, set this to None. Otherwise set it to the query # If not, set this to None. Otherwise set it to the query

7
vllm_ascend/attention/mla_v1.py
View File

@@ -1,5 +1,6 @@
from dataclasses import dataclass from dataclasses import dataclass
from typing import TYPE_CHECKING, NamedTuple, Optional, Tuple, Type, TypeVar from typing import (TYPE_CHECKING, ClassVar, NamedTuple, Optional, Tuple, Type,
TypeVar)
import torch import torch
import torch_npu import torch_npu
@@ -12,6 +13,7 @@ from vllm.distributed import get_tensor_model_parallel_world_size, get_tp_group
from vllm.model_executor.layers.linear import (LinearBase, from vllm.model_executor.layers.linear import (LinearBase,
UnquantizedLinearMethod) UnquantizedLinearMethod)
from vllm.utils import cdiv, round_down from vllm.utils import cdiv, round_down
from vllm.v1.attention.backends.utils import AttentionCGSupport
from vllm_ascend.ascend_config import get_ascend_config from vllm_ascend.ascend_config import get_ascend_config
from vllm_ascend.attention.attention_v1 import AscendAttentionState from vllm_ascend.attention.attention_v1 import AscendAttentionState
@@ -163,6 +165,9 @@ M = TypeVar("M", bound=AscendMLAMetadata)
class AscendMLAMetadataBuilder: class AscendMLAMetadataBuilder:
# Does this backend/builder support ACL Graphs for attention (default: no).
aclgraph_support: ClassVar[AttentionCGSupport] = \
AttentionCGSupport.NEVER
""" """
NOTE: Please read the comment at the top of the file before trying to NOTE: Please read the comment at the top of the file before trying to
understand this class understand this class

4
vllm_ascend/worker/model_runner_v1.py
View File

@@ -3259,8 +3259,8 @@ class NPUModelRunner(LoRAModelRunnerMixin):
builder = attn_group.metadata_builder builder = attn_group.metadata_builder
else: else:
builder = attn_group.get_metadata_builder() builder = attn_group.get_metadata_builder()
if builder.cudagraph_support.value < min_ag_support.value: if builder.aclgraph_support.value < min_ag_support.value:
min_ag_support = builder.cudagraph_support min_ag_support = builder.aclgraph_support
min_ag_builder_name = builder.__class__.__name__ min_ag_builder_name = builder.__class__.__name__
# This is an imitation of compilation_config.splitting_ops_contain_attention() # This is an imitation of compilation_config.splitting_ops_contain_attention()