(cherry picked from commit 94da229a7788d74685d1591a4e75a8341de64f41)
[ROCm/rccl commit: a7fcd58a97]
Этот коммит содержится в:
@@ -120,7 +120,13 @@ ncclResult_t NpKit::Dump(const std::string& dump_dir) {
|
||||
dump_file_path = dump_dir;
|
||||
dump_file_path += "/gpu_clock_rate_rank_";
|
||||
dump_file_path += std::to_string(rank_);
|
||||
constexpr int vega_gpu_rtc_freq_in_khz = 25000;
|
||||
hipDeviceProp_t devProp;
|
||||
int vega_gpu_rtc_freq_in_khz;
|
||||
CUDACHECK(hipGetDeviceProperties(&devProp, 0));
|
||||
if (devProp.gcnArch/10 == 94)
|
||||
vega_gpu_rtc_freq_in_khz = 100000;
|
||||
else
|
||||
vega_gpu_rtc_freq_in_khz = 25000;
|
||||
std::string clock_rate_str = std::to_string(vega_gpu_rtc_freq_in_khz);
|
||||
auto gpu_clock_rate_file = std::fstream(dump_file_path, std::ios::out);
|
||||
gpu_clock_rate_file.write(clock_rate_str.c_str(), clock_rate_str.length());
|
||||
|
||||
Ссылка в новой задаче
Block a user