A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 4, 01:15 UTC
[Bugfix] Fix GDN KKT precision loss on Hopper GPUs by aligning tl.dot operand layout with WGMMA (#42076)
Signed-off-by: kermit <ckeming@outlook.com>
Kermit committed
May 9, 2026 at 21:08 UTC
adb6d96516cb0cd6d1dec985cb5e5cdc20ed418b
1 file changed
+1
-1
vllm/model_executor/layers/fla/ops/chunk_scaled_dot_kkt.py
+1
-1
@@ -83,7 +83,7 @@ def chunk_scaled_dot_kkt_fwd_kernel(
83
)
84
b_k = tl.load(p_k, boundary_check=(0, 1))
85
b_kb = b_k * b_beta[:, None]
86
- b_A += tl.dot(b_kb.to(b_k.dtype), tl.trans(b_k))
86
+ b_A += tl.dot(b_kb, tl.trans(b_k).to(b_kb.dtype))
87
88
if USE_G:
89
p_g = tl.make_block_ptr(g + bos * H + i_h, (T,), (H,), (i_t * BT,), (BT,), (0,))