Sylvain Jeaugey
f9c3dc251e
Add local user buffer registration for NVLink SHARP. Add tuning plugin support. Increase net API to v7 to allow for device-side packet reordering; remove support for v4 plugins. Add support for RoCE ECE. Add support for C2C links. Better detect SHM allocation failures to avoid crash with Bus Error. Fix missing thread unlocks in bootstrap (Fixes #936). Disable network flush by default on H100. Move device code from src/collectives/device to src/device.
23 rader
828 B
C
23 rader
828 B
C
/*************************************************************************
|
|
* Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved.
|
|
* Copyright (c) 2023, Meta Platforms, Inc. and affiliates.
|
|
*
|
|
* See LICENSE.txt for license information
|
|
************************************************************************/
|
|
|
|
#ifndef NCCL_INT_TUNER_H_
|
|
#define NCCL_INT_TUNER_H_
|
|
|
|
#include "nccl_tuner.h"
|
|
|
|
// Tuning plugin to override NCCL's default algorithm/protocol tuning.
|
|
|
|
// Attempts to load NCCL tuner from environmental variable.
|
|
// Returns ncclSuccess if the correct tuner symbol has been found and
|
|
// successully loaded. Otherwise returns an error and also logs the error.
|
|
ncclResult_t ncclLoadTunerPlugin(ncclTuner_t** tuner);
|
|
|
|
// Cleans up NCCL tuner plugin.
|
|
ncclResult_t ncclCloseTunerPlugin(ncclTuner_t** tuner);
|
|
#endif
|