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

[AMD] Fix DeepSeek-R1-MXFP4 accuracy with AITER FP8 (#37762)

Browse Source
This commit is contained in:
Bingxu Chen
2026-09-21 10:51:37 -07:00
committed by GitHub
parent e6931ca889
commit 632919e498
1 changed files with 1 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mla_rocm.py
+1
View File
@@ -349,6 +349,7 @@ def _fused_rope_cat_and_cache(
q_nope_out.dtype q_nope_out.dtype
if attn.kv_cache_dtype == "fp8_e4m3" if attn.kv_cache_dtype == "fp8_e4m3"
and attn.current_attention_backend == "aiter" and attn.current_attention_backend == "aiter"
and q_nope_out.shape[-2] == 12
else kv_cache_dtype else kv_cache_dtype
) )
kv_pool = get_token_to_kv_pool() kv_pool = get_token_to_kv_pool()