Handle profiler proxy initialization failures

[pawelchcki]

Summary

Do not enable profiler proxy events when proxy initialization fails. This prevents SaveProxyProfiler() from dereferencing a null proxy pool after /dev/shm allocation failure.

Validation

• Profiler v6 requesting ncclProfileColl | ncclProfileKernelCh, two T4 ranks.
• 64 MiB /dev/shm: previously segfaulted in ncclLocalOpAppend; now logs allocation warnings and exits 0.
• 1 GiB /dev/shm: exits 0.

Minimal reproducer

Save as repro.cc:

#ifdef PLUGIN
#include "nccl.h"
#include "nccl_common.h"
#include "plugin/nccl_profiler.h"
static ncclResult_t init(void** c, uint64_t, int* m, const char*, int, int, int,
                         ncclDebugLogger_t) {
  static int context; *c = &context;
  *m = ncclProfileColl | ncclProfileKernelCh;
  return ncclSuccess;
}
static ncclResult_t start(void* c, void** e, ncclProfilerEventDescr_t*) {
  *e = c; return ncclSuccess;
}
static ncclResult_t stop(void*) { return ncclSuccess; }
static ncclResult_t record(void*, ncclProfilerEventState_t,
                           ncclProfilerEventStateArgs_t*) { return ncclSuccess; }
extern "C" __attribute__((visibility("default"))) ncclProfiler_t ncclProfiler_v6 =
    {"shm-repro", init, start, stop, record, stop};
#else
#include <cuda_runtime.h>
#include <nccl.h>
#define OK(x) do { if ((x) != 0) return 1; } while (0)
int main() {
  int devs[] = {0, 1}; ncclComm_t comms[2]; float* ptrs[2];
  OK(ncclCommInitAll(comms, 2, devs));
  for (int i = 0; i < 2; ++i) {
    OK(cudaSetDevice(i)); OK(cudaMalloc(&ptrs[i], sizeof(float)));
  }
  OK(ncclGroupStart());
  for (int i = 0; i < 2; ++i) {
    OK(cudaSetDevice(i));
    OK(ncclAllReduce(ptrs[i], ptrs[i], 1, ncclFloat, ncclSum, comms[i], 0));
  }
  OK(ncclGroupEnd());
  for (int i = 0; i < 2; ++i) {
    OK(cudaSetDevice(i)); OK(cudaDeviceSynchronize());
  }
}
#endif

Build from the NCCL source root and run in a two-GPU container with 64 MiB /dev/shm:

g++ -DPLUGIN -DNCCL_OS_LINUX=1 -Isrc/include -Ibuild/include -I/usr/local/cuda/include -shared -fPIC repro.cc -o profiler.so
g++ -Ibuild/include -I/usr/local/cuda/include repro.cc -Lbuild/lib -L/usr/local/cuda/lib64 -lnccl -lcudart -o repro
NCCL_PROFILER_PLUGIN=$PWD/profiler.so ./repro

[xiaofanl-nvidia]

@pawelchcki please help add DCO signoff as per automatic checks. Thanks!