A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
fix cuda graph (#22721)
Signed-off-by: fsx950223 <fsx950223@outlook.com>
who who who committed
Aug 20, 2025 at 14:24 UTC
d983769c41db224e0897fac2e9aefc5f57ad1122
1 file changed
+4
-3
vllm/v1/attention/backends/rocm_aiter_fa.py
+4
-3
@@ -2,7 +2,7 @@
2
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
"""Attention layer with AiterFlashAttention."""
4
from dataclasses import dataclass
5
-from typing import ClassVar, Optional
5
+from typing import Optional
6
7
import torch
8
@@ -11,7 +11,8 @@ from vllm.attention.backends.abstract import (AttentionBackend, AttentionImpl,
11
from vllm.config import VllmConfig
12
from vllm.logger import init_logger
13
from vllm.platforms import current_platform
14
-from vllm.v1.attention.backends.utils import (AttentionMetadataBuilder,
14
+from vllm.v1.attention.backends.utils import (AttentionCGSupport,
15
+ AttentionMetadataBuilder,
16
CommonAttentionMetadata)
17
from vllm.v1.kv_cache_interface import AttentionSpec
18
@@ -231,7 +232,7 @@ class AiterFlashAttentionMetadata:
232
233
class AiterFlashAttentionMetadataBuilder(
234
AttentionMetadataBuilder[AiterFlashAttentionMetadata]):
234
- full_cudagraph_supported: ClassVar[bool] = True
235
+ cudagraph_support = AttentionCGSupport.ALWAYS
236
237
def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],
238
vllm_config: VllmConfig, device: torch.device):