chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[AMD] Update aiter to v0.1.10.post2 (#18423)

Browse Source 
Co-authored-by: kkHuang-amd <wunhuang@amd.com>
Co-authored-by: YC Tseng <yctseng@amd.com>
This commit is contained in:
Bingxu Chen
2026-02-09 14:08:24 +08:00
committed by GitHub
parent b564dcec61
commit 3f3c201243
15 changed files with 79 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
python/sglang/srt/layers/attention/aiter_backend.py
View File

@@ -268,6 +268,7 @@ class AiterAttnBackend(AttentionBackend):
self,
qo_indptr,
kv_indptr,
kv_last_page_len,
work_metadata,
work_info_set,
work_indptr,
@@ -287,6 +288,7 @@ class AiterAttnBackend(AttentionBackend):
meta = get_mla_metadata_v1(
qo_indptr,
kv_indptr,
kv_last_page_len,
self.num_head // nhead_kv,
nhead_kv,
True,
@@ -367,6 +369,7 @@ class AiterAttnBackend(AttentionBackend):
self.make_mla_meta_data(
qo_indptr,
kv_indptr,
kv_last_page_len,
work_metadata,
work_info_set,
work_indptr,
@@ -423,6 +426,7 @@ class AiterAttnBackend(AttentionBackend):
self.make_mla_meta_data(
qo_indptr,
kv_indptr,
self.kv_last_page_len[:bs],
work_metadata,
work_info_set,
work_indptr,
@@ -518,6 +522,7 @@ class AiterAttnBackend(AttentionBackend):
self.make_mla_meta_data(
qo_indptr,
kv_indptr,
self.kv_last_page_len[:bs],
work_metadata,
work_info_set,
work_indptr,
@@ -716,6 +721,7 @@ class AiterAttnBackend(AttentionBackend):
self.make_mla_meta_data(
qo_indptr,
kv_indptr,
kv_last_page_len,
self.work_metadata,
self.work_info_set,
self.work_indptr,
@@ -786,6 +792,7 @@ class AiterAttnBackend(AttentionBackend):
self.make_mla_meta_data(
qo_indptr,
kv_indptr,
kv_last_page_len,
self.work_metadata,
self.work_info_set,
self.work_indptr,
@@ -872,6 +879,7 @@ class AiterAttnBackend(AttentionBackend):
self.make_mla_meta_data(
qo_indptr,
kv_indptr,
kv_last_page_len,
self.work_metadata,
self.work_info_set,
self.work_indptr,
@@ -1144,6 +1152,7 @@ class AiterAttnBackend(AttentionBackend):
self.make_mla_meta_data(
self.forward_metadata.qo_indptr,
self.forward_metadata.kv_indptr,
self.forward_metadata.kv_last_page_len,
work_metadata,
work_info_set,
work_indptr,
@@ -1165,8 +1174,8 @@ class AiterAttnBackend(AttentionBackend):
self.forward_metadata.kv_indices,
self.forward_metadata.kv_last_page_len,
self.forward_metadata.max_q_len,
layer.scaling,
layer.logit_cap,
sm_scale=layer.scaling,
logit_cap=layer.logit_cap,
work_meta_data=work_metadata,
work_indptr=work_indptr,
work_info_set=work_info_set,
@@ -1195,6 +1204,7 @@ class AiterAttnBackend(AttentionBackend):
self.make_mla_meta_data(
self.forward_metadata.qo_indptr,
self.forward_metadata.kv_indptr,
self.forward_metadata.kv_last_page_len,
work_metadata,
work_info_set,
work_indptr,
@@ -1232,8 +1242,8 @@ class AiterAttnBackend(AttentionBackend):
self.forward_metadata.kv_indices,
self.forward_metadata.kv_last_page_len,
self.forward_metadata.max_q_len,
layer.scaling,
layer.logit_cap,
sm_scale=layer.scaling,
logit_cap=layer.logit_cap,
work_meta_data=work_metadata,
work_indptr=work_indptr,
work_info_set=work_info_set,
@@ -1262,8 +1272,8 @@ class AiterAttnBackend(AttentionBackend):
self.forward_metadata.kv_indices,
self.forward_metadata.kv_last_page_len,
self.forward_metadata.max_q_len,
layer.scaling,
layer.logit_cap,
sm_scale=layer.scaling,
logit_cap=layer.logit_cap,
work_meta_data=work_metadata,
work_indptr=work_indptr,
work_info_set=work_info_set,
@@ -1353,6 +1363,7 @@ class AiterAttnBackend(AttentionBackend):
self.make_mla_meta_data(
self.forward_metadata.qo_indptr,
self.forward_metadata.kv_indptr,
self.forward_metadata.kv_last_page_len,
work_metadata,
work_info_set,
work_indptr,
@@ -1374,8 +1385,8 @@ class AiterAttnBackend(AttentionBackend):
self.forward_metadata.kv_indices,
self.forward_metadata.kv_last_page_len,
self.forward_metadata.max_q_len,
layer.scaling,
layer.logit_cap,
sm_scale=layer.scaling,
logit_cap=layer.logit_cap,
work_meta_data=work_metadata,
work_indptr=work_indptr,
work_info_set=work_info_set,