diff --git a/tests/kernels/moe/test_ocp_mx_moe.py b/tests/kernels/moe/test_ocp_mx_moe.py index a96e47fe43..8c620afbc8 100644 --- a/tests/kernels/moe/test_ocp_mx_moe.py +++ b/tests/kernels/moe/test_ocp_mx_moe.py @@ -659,19 +659,6 @@ def test_trtllm_gen_mxfp4_fused_moe( check_accuracy(ref_result, tg_result, atol=0, rtol=0.3, percent=0.8) -def _interleave_scales_lastdim_by4(scales: torch.Tensor) -> torch.Tensor: - """Interleave scales on the last dimension by groups of 4, matching - the transformation in mxfp4.py's BF16 (Hopper) path.""" - s = scales.to(torch.uint8) - s_shape = s.shape - assert s_shape[-1] % 4 == 0 - s = s.reshape(*s_shape[:-1], s_shape[-1] // 4, 4) - # Move the 4-group dimension before the row dimension - permuted = s.permute(0, 2, 1, 3) - # Merge the row dim with the 4-group dim - return permuted.reshape(s_shape[0], s_shape[-1] // 4, s_shape[1] * 4) - - @pytest.mark.parametrize("topk", [1, 4]) @pytest.mark.parametrize("num_experts", [32]) @pytest.mark.parametrize("num_tokens", [1, 128]) @@ -771,13 +758,25 @@ def test_flashinfer_cutlass_mxfp4_fused_moe( w1_w, w3_w = torch.chunk(w13_q, 2, dim=1) w13_q_swapped = torch.cat([w3_w, w1_w], dim=1) + # SM90 mixed-input GEMM expects weights/scales in an interleaved layout; + # without it the FP4->BF16 LUT reads bytes from wrong positions for K>128. + from flashinfer.fused_moe import ( + interleave_moe_scales_for_sm90_mixed_gemm, + interleave_moe_weights_for_sm90_mixed_gemm, + ) + + w13_q_swapped = interleave_moe_weights_for_sm90_mixed_gemm( + w13_q_swapped, quant_type="fp4" + ) + w2_q = interleave_moe_weights_for_sm90_mixed_gemm(w2_q, quant_type="fp4") + b1, b3 = torch.chunk(bias13.to(torch.float32), 2, dim=-1) w13_b = torch.cat([b3, b1], dim=-1).to(torch.bfloat16) w1_s, w3_s = torch.chunk(w13_scale, 2, dim=1) w13_s = torch.cat([w3_s, w1_s], dim=1) - w13_s_inter = _interleave_scales_lastdim_by4(w13_s) - w2_s_inter = _interleave_scales_lastdim_by4(w2_scale) + w13_s_inter = interleave_moe_scales_for_sm90_mixed_gemm(w13_s) + w2_s_inter = interleave_moe_scales_for_sm90_mixed_gemm(w2_scale) routing_weights = torch.nn.functional.softmax( router_logits, dim=1, dtype=torch.float32