[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
Αυτή η υποβολή περιλαμβάνεται σε:
Nilesh M Negi
2025-07-10 10:56:18 -05:00
υποβλήθηκε από GitHub
γονέας 68d6f99e0f
υποβολή 2c099fe29a
6 αρχεία άλλαξαν με 61 προσθήκες και 45 διαγραφές
+4 -4
Προβολή Αρχείου
@@ -29,7 +29,7 @@
using namespace rccl;
/* [RCCL] Determine which GPU kernel to execute */
void* rcclGetKernelIndex(int unroll, bool useCollTrace, struct ncclTaskColl* task = NULL)
void* rcclGetKernelIndex(int unroll, bool useCollTrace, struct ncclTaskColl* task)
{
// At this time, unroll factor is controlled only by passed in unroll argument
// After more investigation, this may be further tuned by the actual task being processed
@@ -48,9 +48,9 @@ void* rcclGetKernelIndex(int unroll, bool useCollTrace, struct ncclTaskColl* tas
return rcclKernelTable[firstKernel + kernelIdx].funcPtr;
}
}
// Fall back to default unroll
WARN("Requested RCCL_UNROLL_FACTOR: %d does not exist in `rcclKernelTable`. Falling back to default unroll: %d", unroll, rcclKernelTable[firstKernel].unroll);
return rcclKernelTable[firstKernel].funcPtr;
// If does not match, return null
return nullptr;
}
static int rcclProtoGrainSize(int proto, ncclComm *comm){