EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[Feature] Support FA3 backend for MLA (#4831)

Browse Source
This commit is contained in:
Baizhou Zhang
2025-03-28 18:30:14 -07:00
committed by GitHub
parent ec3ee0289d
commit 20c90be23d
3 changed files with 180 additions and 74 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
python/sglang/srt/model_executor/model_runner.py
View File

@@ -230,6 +230,10 @@ class ModelRunner:
elif server_args.enable_flashmla:
logger.info("MLA optimization is turned on. Use flashmla decode.")
server_args.attention_backend = "flashmla"
elif server_args.attention_backend == "fa3":
logger.info(
f"MLA optimization is turned on. Use flash attention 3 backend."
)
else:
logger.info("MLA optimization is turned on. Use triton backend.")
server_args.attention_backend = "triton"
@@ -879,7 +883,7 @@ class ModelRunner:
"Please use `--attention-backend flashinfer`."
)
logger.warning(
"FlashAttention v3 Backend is in Beta. Multimodal, Page > 1, FP8, MLA and Speculative Decoding are not supported."
"FlashAttention v3 Backend is in Beta. Multimodal, FP8, and Speculative Decoding are not supported."
)
from sglang.srt.layers.attention.flashattention_backend import (
FlashAttentionBackend,