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

[AMD] remove redundancy H2D op in aiter attention backend (#19416)

Browse Source 
Co-authored-by: root <root@mia1-p01-g28.mia.tensorwave.lan>
This commit is contained in:
AMD-yanfeiwang
2026-02-27 02:15:55 -08:00
committed by GitHub
co-authored by root
parent 07da4bed7b
commit f69ca93d49
1 changed files with 1 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/aiter_backend.py
+1 -3
View File
@@ -1487,9 +1487,7 @@ class AiterAttnBackend(AttentionBackend):
k = k.to(fp8_dtype) k = k.to(fp8_dtype)
if v.dtype != fp8_dtype: if v.dtype != fp8_dtype:
v = v.to(fp8_dtype) v = v.to(fp8_dtype)
one_scale = torch.tensor( one_scale = torch.ones((), dtype=torch.float32, device=q.device)
1.0, dtype=torch.float32, device=q.device
)
kv_indptr_asm = qo_indptr kv_indptr_asm = qo_indptr
kv_indices_asm = self.forward_metadata.fp8_prefill_kv_indices kv_indices_asm = self.forward_metadata.fp8_prefill_kv_indices