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 14, 05:30 UTC
[CI Failure] Fix test_flashinfer_cutlass_mxfp4_mxfp8_fused_moe (#24750)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Sep 13, 2025 at 03:29 UTC
59d7ffc17f4c948b4d25d014e4f90d0cd4c20990
2 files changed
+3
-2
csrc/attention/mla/sm100_cutlass_mla_kernel.cu
+1
@@ -43,6 +43,7 @@ void sm100_cutlass_mla_decode(
43
torch::Tensor const& seq_lens,
44
torch::Tensor const& page_table,
45
torch::Tensor const& workspace,
46
+ double sm_scale,
47
int64_t num_kv_splits) {
48
TORCH_CHECK(false, "CUDA version must be >= 12.4 for cutlass_mla_decode");
49
}
tests/kernels/moe/test_mxfp4_moe.py
+2
-2
@@ -771,11 +771,11 @@ def test_flashinfer_cutlass_mxfp4_mxfp8_fused_moe(
771
w13_ref = dequant_mxfp4_batches(
772
w13_q.view(torch.uint8),
773
w13_scale.view(torch.uint8).reshape(-1)).to(torch.float32).reshape(
774
- num_experts, 2 * intermediate_size, hidden_size)
774
+ num_experts, 2 * intermediate_size, hidden_size).to(device)
775
w2_ref = dequant_mxfp4_batches(
776
w2_q.view(torch.uint8),
777
w2_scale.view(torch.uint8).reshape(-1)).to(torch.float32).reshape(
778
- num_experts, hidden_size, intermediate_size)
778
+ num_experts, hidden_size, intermediate_size).to(device)
779
780
# Quantize activations for SM100 path and dequantize for reference
781
hidden_states_q, hidden_states_sf = mxfp8_quantize(hidden_states, True, 32)