[Bug-fix] Fix MoE fused sum row offsets (#50220)
Signed-off-by: yaozhiyuan.666 <yaozhiyuan.666@bytedance.com> Co-authored-by: yaozhiyuan.666 <yaozhiyuan.666@bytedance.com>
yzyyzyhhh committed
Sep 5, 2026 at 00:39 UTC
1ff5edb02327583cd9c88f6bda3c485dbd504938
1 file changed
+1
-1
vllm/model_executor/layers/fused_moe/moe_fused_mul_sum.py
+1
-1
@@ -58,7 +58,7 @@ def moe_fused_mul_sum_kernel(
58
weights += (tl.load(w_row + n).to(tl.float32),) # type: ignore[assignment]
59
60
n_tiles: tl.constexpr = (hidden_size + BLOCK_K - 1) // BLOCK_K
61
- a_row = inputs_ptr + pid_m * stride_m
61
+ a_row = inputs_ptr + pid_m.to(tl.int64) * stride_m
62
out_row = outputs_ptr + pid_m * hidden_size
63
64
for t in tl.range(0, n_tiles):