From 3be8d312a216727dd3624afa92984dab6e170274 Mon Sep 17 00:00:00 2001 From: Tyler Michael Smith Date: Tue, 1 Jul 2025 21:05:47 -0400 Subject: [PATCH] [Kernel][Bugfix] Fixup some warnings in nvfp4_blockwise_moe when CUDA < 12.8 (#20324) Signed-off-by: Tyler Michael Smith --- csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu | 2 ++ 1 file changed, 2 insertions(+) diff --git a/csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu b/csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu index 45ec3d29ce04..a21ee55b6586 100644 --- a/csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu +++ b/csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu @@ -335,8 +335,10 @@ void run_fp4_blockwise_scaled_group_mm( TORCH_CHECK(status == cutlass::Status::kSuccess, "Failed to run GEMM"); } +#if defined ENABLE_NVFP4 && ENABLE_NVFP4 constexpr auto FLOAT4_E2M1X2 = at::ScalarType::Byte; constexpr auto SF_DTYPE = at::ScalarType::Float8_e4m3fn; +#endif #define CHECK_TYPE(x, st, m) \ TORCH_CHECK(x.scalar_type() == st, ": Inconsistency of Tensor type:", m)