Enable fp8 support (#1101)
* initial checkin
* resolve cr comments
* resolve the build issue
* fix the data correctless issue
* update fp8 header file and update the unit test for fp8 support
* remove fp16 from fp8 headers
* fix ut issue and catch up the latest code from develop
* udate according to cr comments
* update ut according to cr comments
* update num floats for each SumPostDiv from 4 to 6
* update fp8 header file name
* fix the typo
[ROCm/rccl commit: 6777e65c1d]
Цей коміт міститься в:
@@ -226,6 +226,10 @@ static ncclResult_t hostToDevRedOp(
|
||||
#if defined(RCCL_BFLOAT16)
|
||||
rccl_bfloat16 bf16;
|
||||
#endif
|
||||
#if defined(RCCL_FLOAT8)
|
||||
rccl_float8 fp8_e4m3;
|
||||
rccl_bfloat8 fp8_e5m2;
|
||||
#endif
|
||||
float f32;
|
||||
double f64;
|
||||
void *ptr;
|
||||
@@ -269,6 +273,16 @@ static ncclResult_t hostToDevRedOp(
|
||||
bf16 = (rccl_bfloat16)(float(1.0/comm->nRanks));
|
||||
break;
|
||||
#endif
|
||||
#if defined(RCCL_FLOAT8)
|
||||
case ncclFp8E4M3:
|
||||
opFull->op = ncclDevPreMulSum;
|
||||
fp8_e4m3 = (rccl_float8)(float(1.0/comm->nRanks));
|
||||
break;
|
||||
case ncclFp8E5M2:
|
||||
opFull->op = ncclDevPreMulSum;
|
||||
fp8_e5m2 = (rccl_bfloat8)(float(1.0/comm->nRanks));
|
||||
break;
|
||||
#endif
|
||||
case ncclFloat32:
|
||||
opFull->op = ncclDevPreMulSum;
|
||||
f32 = float(1.0/comm->nRanks);
|
||||
@@ -315,7 +329,9 @@ static ncclResult_t hostToDevRedOp(
|
||||
MSCCL_KERNEL_ENTRY_DEVREDOP_TYPE(devredop, half, fullOps), \
|
||||
MSCCL_KERNEL_ENTRY_DEVREDOP_TYPE(devredop, float, fullOps), \
|
||||
MSCCL_KERNEL_ENTRY_DEVREDOP_TYPE(devredop, double, fullOps), \
|
||||
MSCCL_KERNEL_ENTRY_DEVREDOP_TYPE(devredop, rccl_bfloat16, fullOps)
|
||||
MSCCL_KERNEL_ENTRY_DEVREDOP_TYPE(devredop, rccl_bfloat16, fullOps), \
|
||||
MSCCL_KERNEL_ENTRY_DEVREDOP_TYPE(devredop, rccl_float8, fullOps), \
|
||||
MSCCL_KERNEL_ENTRY_DEVREDOP_TYPE(devredop, rccl_bfloat8, fullOps)
|
||||
|
||||
#define MSCCL_KERNEL_ENTRY_DEVREDOP_NOFLOAT(devredop, fullOps) \
|
||||
MSCCL_KERNEL_ENTRY_DEVREDOP_TYPE(devredop, int8_t, fullOps), \
|
||||
@@ -327,6 +343,8 @@ static ncclResult_t hostToDevRedOp(
|
||||
MSCCL_KERNEL_ENTRY_DEVREDOP_NULL(), \
|
||||
MSCCL_KERNEL_ENTRY_DEVREDOP_NULL(), \
|
||||
MSCCL_KERNEL_ENTRY_DEVREDOP_NULL(), \
|
||||
MSCCL_KERNEL_ENTRY_DEVREDOP_NULL(), \
|
||||
MSCCL_KERNEL_ENTRY_DEVREDOP_NULL(), \
|
||||
MSCCL_KERNEL_ENTRY_DEVREDOP_NULL()
|
||||
|
||||
#define MSCCL_KERNEL_ENTRY() \
|
||||
|
||||
Посилання в новій задачі
Заблокувати користувача