mirror of
https://github.com/NVIDIA/TensorRT-LLM.git
synced 2026-01-14 06:27:45 +08:00
5e634dd1bd
This PR adds a customized allreduce to TensorRT-LLM. The new allreduce is used for communication on PCIe-based GPUs via low-precision quantization, which can accelerate the PCIe allreduce process. Signed-off-by: Hui Kang <hkang@nvidia.com> Co-authored-by: Hui Kang <hkang@nvidia.com> |
||
|---|---|---|
| .. | ||
| allReduceFusionKernels.cu | ||
| allReduceFusionKernels.h | ||
| allReduceWorkspace.cu | ||
| allReduceWorkspace.h | ||
| customLowPrecisionAllReduceKernels.cu | ||
| customLowPrecisionAllReduceKernels.h | ||
| moeAllReduceFusionKernels.cu | ||
| moeAllReduceFusionKernels.h | ||