[INIT] Fix fallback for unsupported user-specified runtime unroll factor (#1780)

* [INIT] Fix fallback for unsupported user-specified runtime unroll factor
* Add CollTrace guard
* Move `commSetUnrollFactor()` to rccl_wrap.cc
* Modify comments in the device-code generator script

[ROCm/rccl commit: 2c099fe29a]
このコミットが含まれているのは:
Nilesh M Negi
2025-07-10 10:56:18 -05:00
committed by GitHub
コミット ba31e4e846
6個のファイルの変更61行の追加45行の削除
+2
ファイルの表示
@@ -17,6 +17,8 @@
#define NCCL_SIMPLE_ALIGNMENT (WARP_SIZE * 8LL * 16LL)
#define NCCL_BYTES_ALIGNMENT 16
void* rcclGetKernelIndex(int unroll, bool useCollTrace, struct ncclTaskColl* task = NULL);
ncclResult_t ncclInitKernelsForDevice(int cudaArch, int maxSharedMem, size_t* maxStackSize);
ncclResult_t ncclEnqueueCheck(struct ncclInfo* info);
ncclResult_t ncclLaunchPrepare(struct ncclComm* comm);