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

[AMD] Restore AITER verify runtime sizing reverted by #34647 (#38575)

Browse Source
This commit is contained in:
Chengze Fan
2026-09-10 05:57:47 -07:00
committed by GitHub
parent 613002281e
commit 12771786f2
1 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/layers/attention/aiter_backend.py
+4 -2
View File
@@ -1542,8 +1542,8 @@ class AiterAttnBackend(AttentionBackend):
run_graph=False, run_graph=False,
) )
else: else:
draft_num = forward_batch.input_ids.shape[0] // bs
bs = len(forward_batch.req_pool_indices) bs = len(forward_batch.req_pool_indices)
draft_num = spec_info.draft_token_num
if self._use_unified_verify: if self._use_unified_verify:
page_table, qo_indptr, max_q_len, swa_page_table = ( page_table, qo_indptr, max_q_len, swa_page_table = (
@@ -2793,7 +2793,9 @@ class AiterAttnBackend(AttentionBackend):
v=v_unified, v=v_unified,
out=o.view(-1, layer.tp_q_head_num, layer.v_head_dim), out=o.view(-1, layer.tp_q_head_num, layer.v_head_dim),
cu_seqlens_q=self.forward_metadata.qo_indptr, cu_seqlens_q=self.forward_metadata.qo_indptr,
seqused_k=forward_batch.seq_lens + self.num_draft_tokens, seqused_k=(
forward_batch.seq_lens + self.forward_metadata.max_q_len
),
max_seqlen_q=self.forward_metadata.max_q_len, max_seqlen_q=self.forward_metadata.max_q_len,
max_seqlen_k=max_kv_len, max_seqlen_k=max_kv_len,
softmax_scale=layer.scaling, softmax_scale=layer.scaling,