Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-19 01:04:32 +02:00
Code Issues Packages Projects Releases Wiki Activity

sdnq atten fuse alpha and pv

Browse Source
This commit is contained in:
Disty0
2026-07-24 21:41:20 +03:00
parent d11a619a64
commit 9b296cbc1e
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
@@ -164,10 +164,10 @@ def sdnq_attn_kernel(
qk = qk - m_ij[:, None]
p = tl.exp2(qk)
l_i = tl.fma(l_i, alpha, tl.sum(p, 1))
acc *= alpha[:, None]
v = v_desc.load([start_n, 0])
if pv_is_quantized:
acc *= alpha[:, None]
v_scale = v_scale_desc.load([start_n])[None, :]
p *= v_scale
if v.dtype == tl.int8:
@@ -182,7 +182,7 @@ def sdnq_attn_kernel(
acc = tl.fma(tl.dot(p, v, out_dtype=tl.float32), p_scale, acc)
else:
p = p.to(v.dtype)
acc = tl.dot(p, v, acc, out_dtype=tl.float32)
acc = tl.fma(acc, alpha[:, None], tl.dot(p, v, out_dtype=tl.float32))
m_i = m_ij
l_i = 1 / l_i[:, None]