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

[Fix] Update to v0.1.5.post4 and refine HIP attention backend selection (#11161)

Browse Source
This commit is contained in:
jacky.cheng
2025-10-03 12:19:30 +08:00
committed by GitHub
parent a2faf8940c
commit b00a0c786f
2 changed files with 2 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
docker/Dockerfile.rocm
View File

@@ -42,7 +42,7 @@ ENV BUILD_TRITON="0"
ENV BUILD_LLVM="0"
ENV BUILD_AITER_ALL="1"
ENV BUILD_MOONCAKE="1"
ENV AITER_COMMIT="v0.1.5.post3"
ENV AITER_COMMIT="v0.1.5.post4"
ENV NO_DEPS_FLAG="--no-deps"
# ===============================

4
python/sglang/srt/model_executor/model_runner.py
View File

@@ -536,9 +536,7 @@ class ModelRunner:
elif _is_hip:
head_num = self.model_config.get_num_kv_heads(self.tp_size)
# TODO current aiter only support head number 16 or 128 head number
if (
head_num == 128 or head_num == 16
) and self.spec_algorithm.is_none():
if head_num == 128 or head_num == 16:
server_args.attention_backend = "aiter"
else:
server_args.attention_backend = "triton"