diff --git a/vllm/model_executor/layers/fused_moe/moe_fused_mul_sum.py b/vllm/model_executor/layers/fused_moe/moe_fused_mul_sum.py index f9854d1773..d8ad9b9455 100644 --- a/vllm/model_executor/layers/fused_moe/moe_fused_mul_sum.py +++ b/vllm/model_executor/layers/fused_moe/moe_fused_mul_sum.py @@ -58,7 +58,7 @@ def moe_fused_mul_sum_kernel( weights += (tl.load(w_row + n).to(tl.float32),) # type: ignore[assignment] n_tiles: tl.constexpr = (hidden_size + BLOCK_K - 1) // BLOCK_K - a_row = inputs_ptr + pid_m * stride_m + a_row = inputs_ptr + pid_m.to(tl.int64) * stride_m out_row = outputs_ptr + pid_m * hidden_size for t in tl.range(0, n_tiles):