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

[AMD] [Fix] Enable aiter hd256 FP8 prefill FMHA on gfx950 (#33399)

Browse Source
This commit is contained in:
jacky.cheng
2026-08-04 02:40:33 -07:00
committed by GitHub
parent 5e6c37f2b4
commit 723c277640
1 changed files with 4 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/aiter_backend.py
+4 -5
View File
@@ -2339,8 +2339,7 @@ class AiterAttnBackend(AttentionBackend):
window_size = (layer.sliding_window_size, -1) window_size = (layer.sliding_window_size, -1)
if ( if (
get_bool_env_var("SGLANG_AITER_FMHA_FP8_ASM", "False") is_gfx95_supported()
and is_gfx95_supported()
and forward_batch.forward_mode.is_extend() and forward_batch.forward_mode.is_extend()
and forward_batch.extend_prefix_lens_cpu is not None and forward_batch.extend_prefix_lens_cpu is not None
and not any(forward_batch.extend_prefix_lens_cpu) and not any(forward_batch.extend_prefix_lens_cpu)
@@ -2361,9 +2360,9 @@ class AiterAttnBackend(AttentionBackend):
q_c.to(fp8_dtype), q_c.to(fp8_dtype),
k_c.to(fp8_dtype), k_c.to(fp8_dtype),
v_c.to(fp8_dtype), v_c.to(fp8_dtype),
fp8_q_descale, fp8_q_descale.reshape(1),
k_descale, k_descale.reshape(1),
v_descale, v_descale.reshape(1),
self.qo_indptr[:bs0], self.qo_indptr[:bs0],
self.qo_indptr[:bs0], self.qo_indptr[:bs0],
self.forward_metadata.max_q_len, self.forward_metadata.max_q_len,