AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

2.20.3-1

Browse Source 
Add support for alternating rings, allow for cross-nic rings without
cross-rail communication.
Add support for user buffer registration for network send/recv.
Optimize aggregated operations to better utilize all channels.
Add flattening for BCM PCI gen5 switches.
Add support for inter-node NVLink communication
Add support for port fusion in NET/IB.
Add support for ReduceScatter and AllGather using Collnet.
Update net API to v8.
Fix hang during A2A connection.
This commit is contained in:
Sylvain Jeaugey
2024-02-05 05:06:02 -08:00
parent b6d7438d31
commit b6475625fb
74 changed files with 4632 additions and 2165 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/misc/nvmlwrap.cc
+14 -1
View File
@@ -39,6 +39,8 @@ namespace {
NCCL_NVML_FN(nvmlDeviceGetCudaComputeCapability, nvmlReturn_t, (nvmlDevice_t device, int* major, int* minor))
NCCL_NVML_FN(nvmlDeviceGetP2PStatus, nvmlReturn_t, (nvmlDevice_t device1, nvmlDevice_t device2, nvmlGpuP2PCapsIndex_t p2pIndex, nvmlGpuP2PStatus_t* p2pStatus))
NCCL_NVML_FN(nvmlDeviceGetFieldValues, nvmlReturn_t, (nvmlDevice_t device, int valuesCount, nvmlFieldValue_t *values))
// MNNVL support
NCCL_NVML_FN(nvmlDeviceGetGpuFabricInfoV, nvmlReturn_t, (nvmlDevice_t device, nvmlGpuFabricInfoV_t *gpuFabricInfo))
std::mutex lock; // NVML has had some thread safety bugs
bool initialized = false;
@@ -82,7 +84,9 @@ ncclResult_t ncclNvmlEnsureInitialized() {
{(void**)&pfn_nvmlDeviceGetNvLinkCapability, "nvmlDeviceGetNvLinkCapability"},
{(void**)&pfn_nvmlDeviceGetCudaComputeCapability, "nvmlDeviceGetCudaComputeCapability"},
{(void**)&pfn_nvmlDeviceGetP2PStatus, "nvmlDeviceGetP2PStatus"},
{(void**)&pfn_nvmlDeviceGetFieldValues, "nvmlDeviceGetFieldValues"}
{(void**)&pfn_nvmlDeviceGetFieldValues, "nvmlDeviceGetFieldValues"},
// MNNVL support
{(void**)&pfn_nvmlDeviceGetGpuFabricInfoV, "nvmlDeviceGetGpuFabricInfoV"},
};
for(Symbol sym: symbols) {
*sym.ppfn = dlsym(libhandle, sym.name);
@@ -269,3 +273,12 @@ ncclResult_t ncclNvmlDeviceGetFieldValues(nvmlDevice_t device, int valuesCount,
NVMLTRY(nvmlDeviceGetFieldValues, device, valuesCount, values);
return ncclSuccess;
}
// MNNVL support
ncclResult_t ncclNvmlDeviceGetGpuFabricInfoV(nvmlDevice_t device, nvmlGpuFabricInfoV_t *gpuFabricInfo) {
NCCLCHECK(ncclNvmlEnsureInitialized());
std::lock_guard<std::mutex> locked(lock);
gpuFabricInfo->version = nvmlGpuFabricInfo_v2;
NVMLTRY(nvmlDeviceGetGpuFabricInfoV, device, gpuFabricInfo);
return ncclSuccess;
}