EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Fix cuda graph mode in flashinfer attn backend (#10056)

Browse Source
This commit is contained in:
Ben Barsdell
2025-09-07 15:59:48 +10:00
committed by GitHub
parent 85ed8e0a5e
commit a12061df4c
1 changed files with 6 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

10
python/sglang/srt/layers/attention/flashinfer_backend.py
View File

@@ -501,8 +501,9 @@ class FlashInferAttnBackend(AttentionBackend):
sm_scale=layer.scaling,
window_left=layer.sliding_window_size,
logits_soft_cap=logits_soft_cap,
k_scale=layer.k_scale,
v_scale=layer.v_scale,
# Must use _float to avoid device-to-host copy that breaks cuda graph capture.
k_scale=layer.k_scale_float,
v_scale=layer.v_scale_float,
)
else:
causal = True
@@ -580,8 +581,9 @@ class FlashInferAttnBackend(AttentionBackend):
forward_batch.token_to_kv_pool.get_kv_buffer(layer.layer_id),
sm_scale=layer.scaling,
logits_soft_cap=layer.logit_cap,
k_scale=layer.k_scale,
v_scale=layer.v_scale,
# Must use _float to avoid device-to-host copy that breaks cuda graph capture.
k_scale=layer.k_scale_float,
v_scale=layer.v_scale_float,
)
return o.view(-1, layer.tp_q_head_num * layer.head_dim)