Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 17:24:32 +02:00
Code Issues Packages Projects Releases Wiki Activity

SDNQ atten fix enable_gqa + is_causal

Browse Source
This commit is contained in:
Disty0
2026-06-28 22:07:37 +03:00
parent e89f1a1fe8
commit 693639a7d3
1 changed files with 2 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sdnq/kernels/triton_atten.py
+2 -2
View File
@@ -116,7 +116,7 @@ def sdnq_attn_kernel(
lo = 0
if is_causal:
hi = tl.minimum(KN, (start_m + 1) * BLOCK_SIZE_M + KN - QN)
hi = tl.minimum(KN, (start_m + 1) * BLOCK_SIZE_M)
else:
hi = KN
@@ -141,7 +141,7 @@ def sdnq_attn_kernel(
qk = tl.dot(q, k, out_dtype=tl.float32)
if is_causal:
causal_mask = (offs_m[:, None] + (KN - QN)) >= (start_n + offs_n[None, :])
causal_mask = offs_m[:, None] >= (start_n + offs_n[None, :])
qk = tl.where(causal_mask, qk, -float('inf'))
if mask_ptr is not None: