minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[AMD][DeepSeek V4] Set SGLANG_OPT_FLASHMLA_SPARSE_PREFILL to false on hip code path (#30237)

Browse Source
This commit is contained in:
Thomas Wang
2026-07-06 01:25:55 -07:00
committed by GitHub
parent b3ab56545b
commit 80decc78ec
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/arg_groups/deepseek_v4_hook.py
+2 -2
View File
@@ -28,8 +28,8 @@ def apply_deepseek_v4_defaults(server_args: ServerArgs, model_arch: str) -> None
# currently returns incorrect output for DeepSeek-V4-Flash on ROCm/HIP # currently returns incorrect output for DeepSeek-V4-Flash on ROCm/HIP
# (MI355X), which breaks the disaggregation nightly. Keep the previous # (MI355X), which breaks the disaggregation nightly. Keep the previous
# (dense prefill) behavior on ROCm until the sparse kernel is validated # (dense prefill) behavior on ROCm until the sparse kernel is validated
# there; an explicit env var still overrides this. # there;
if is_hip() and not envs.SGLANG_OPT_FLASHMLA_SPARSE_PREFILL.is_set(): if is_hip():
logger.warning( logger.warning(
"Disabling SGLANG_OPT_FLASHMLA_SPARSE_PREFILL by default on ROCm/HIP " "Disabling SGLANG_OPT_FLASHMLA_SPARSE_PREFILL by default on ROCm/HIP "
f"for {model_arch}; set it explicitly to override." f"for {model_arch}; set it explicitly to override."