Triton matmul_ogs Kernel missing multiplying expert weight

[zyongye]

Describe the bug

Hi Triton team,
I have been trying to integrate matmul_ogs kernel into vLLM. But I notice that in the reference implementation matmul_ogs,

triton/bench/triton_bench/matmul_ogs.py

Lines 615 to 618 in e341f7f

	for i, (lo, hi) in enumerate(offs):
	dst_idx = scatter_indx.dst_indx[lo:hi] // n_expts_act
	msk = dst_idx != -1
	out[dst_idx[msk], :] += y[lo:hi, :][msk, :].float()

after matmul between activation and expert weight, all expert output for a certain token is just added together instead of applying the expert weight in rdata.gate_scal. Since matmul_ogs and matmul_ogs_torchare equivalent, I think the same also applies to matmul_ogs? Is that normal, or am I missing something?

Thanks in advance

Environment details

Triton 3.3
CPU: Intel Xeon Gold 6126 CPU @ 2.60GHz
GPU: RTX 8xA6000, driver version 560.35.05, CUDA 12.6

[zyongye]

Ok I just need to pass routing_data.gate_scal into gamma field into matmul_ogs. Solved.