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 15, 03:30 UTC
[Kernel][Bugfix] Fixup some warnings in nvfp4_blockwise_moe when CUDA < 12.8 (#20324)
Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>
Tyler Michael Smith committed
Jul 1, 2025 at 21:05 UTC
3be8d312a216727dd3624afa92984dab6e170274
1 file changed
+2
csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu
+2
@@ -335,8 +335,10 @@ void run_fp4_blockwise_scaled_group_mm(
335
TORCH_CHECK(status == cutlass::Status::kSuccess, "Failed to run GEMM");
336
}
337
338
+#if defined ENABLE_NVFP4 && ENABLE_NVFP4
339
constexpr auto FLOAT4_E2M1X2 = at::ScalarType::Byte;
340
constexpr auto SF_DTYPE = at::ScalarType::Float8_e4m3fn;
341
+#endif
342
343
#define CHECK_TYPE(x, st, m) \
344
TORCH_CHECK(x.scalar_type() == st, ": Inconsistency of Tensor type:", m)