AI/rocm-systems
2
0
Garpu 0
Kode Masalah Tarik Permintaan Actions Packages Projects Rilis Wiki Kegiatan

2.24.3-1

Telusuri Sumber 
Network user buffer support for collectives
 * Leverage user buffer registration to achieve zero-copy
   inter-node communications for Ring, NVLS and Collnet

Add RAS subsystem
 * Create a RAS thread keeping track of all NCCL communicators.
 * Add a ncclras tool contacting the RAS thread and getting a
   report.

Add fp8 support
 * Add support for e5m2 and e4m3 8-bit floating point operations.
 * Use Tree/PAT algorithms when possible for better numerical
   stability.

Add NIC fusion
 * Add a NET API to ask the network plugin to fuse a set of
   interfaces together.
 * Fuse multiple NICs under the same PCI switch as a single,
   larger NIC.

Socket connection failure retry
 * Retry in case of socket connection failure (unreachable host)
 * Avoid "Software caused connection abort" errors on retries

QP connection failure retry
 * Retry in case of IB QP connection failure during ibv_modify_qp.

NET API improvements
 * Allow plugins to force a flush in case data and completion
   ordering is not guaranteed.
 * Indicate when completion is not needed (e.g. for the LL128
   protocol), allowing plugins to skip generating a completion.
 * Allow for full offload of allgather operations when using one
   GPU per node.

NCCL_ALGO/NCCL_PROTO strict enforcement
 * Extend NCCL_ALGO/NCCL_PROTO syntax to be able to specify
   ALGO/PROTO filters for each collective operation.
 * Strictly enforce the ALGO/PROTO filters, no longer fall back
   on the ring algorithm when the filtering leaves no option and
   error out instead.

Enable CUMEM host allocations
 * Use cumem functions for host memory allocation by default.

Improved profiler plugin API
 * Avoid dependencies with NCCL includes.
 * Add information on whether the buffer is registered or not

Adjust PAT tuning
 * Improve transition between PAT and ring at scale.

Fix hangs when running with different CPU architectures
 * Detect when we use a mix of GPU architectures
 * Ensure Algo/Proto decisions are made based on that unified
   state.

Fix FD leak in UDS
 * Fix a leak when mapping buffers intra-node with cumem IPCs.

Fix crash when mixing buffer registration and graph buffer registration.
 * Separate local and graph registration to avoid crashes when we free
   buffers.

Fix user buffer registration with dmabuf
 * Make ncclSend/ncclRecv communication with buffer registration functional
   on network plugins relying on dmabuf for buffer registration.

Fix crash in IB code caused by uninitialized fields.

Fix non-blocking ncclSend/ncclRecv
 * Fix case where ncclSend/ncclRecv would return ncclSuccess in non-blocking
   mode even though the operation was not enqueued onto the stream.
 * Issue #1495

Various compiler tweaks and fixes
 * PR #758

Fix typo in ncclTopoPrintGraph
 * Issue #1468
This commit is contained in:
Sylvain Jeaugey
2024-12-18 08:26:06 -08:00
orang tua 2ea4ee94bf
melakukan 6aae379278
97 mengubah file dengan 12588 tambahan dan 3127 penghapusan
Download Patch File Download Diff File Expand all files Collapse all files
src/device/reduce_kernel.h
+139 -32
Melihat File
@@ -20,6 +20,12 @@ struct IsFloatingPoint<half>: std::true_type {};
template<>
struct IsFloatingPoint<__nv_bfloat16>: std::true_type {};
#endif
#if defined(__CUDA_FP8_TYPES_EXIST__)
template<>
struct IsFloatingPoint<__nv_fp8_e4m3>: std::true_type {};
template<>
struct IsFloatingPoint<__nv_fp8_e5m2>: std::true_type {};
#endif
template<>
struct IsFloatingPoint<float>: std::true_type {};
template<>
@@ -298,6 +304,24 @@ SPECIALIZE_REDUCE(FuncMinMax, double, 1, double, fn.isMinNotMax ? fmin(x, y) : f
#endif
#endif
#if defined(__CUDA_FP8_TYPES_EXIST__)
#if __CUDA_ARCH__ >= 900
SPECIALIZE_REDUCE(FuncSum, __nv_fp8_e4m3, 1, __nv_fp8_e4m3, __nv_fp8_e4m3(__hadd(__half(x),__half(y))))
SPECIALIZE_REDUCE(FuncSum, __nv_fp8_e4m3, 2, __nv_fp8x2_e4m3, __nv_fp8x2_e4m3(__hadd2(__half2(x),__half2(y))))
SPECIALIZE_REDUCE(FuncProd, __nv_fp8_e4m3, 1, __nv_fp8_e4m3, __nv_fp8_e4m3(__hmul(__half(x),__half(y))))
SPECIALIZE_REDUCE(FuncProd, __nv_fp8_e4m3, 2, __nv_fp8x2_e4m3, __nv_fp8x2_e4m3(__hmul2(__half2(x),__half2(y))))
SPECIALIZE_REDUCE(FuncMinMax, __nv_fp8_e4m3, 1, __nv_fp8_e4m3, __nv_fp8_e4m3(fn.isMinNotMax ? __hmin(__half(x),__half(y)) : __hmax(__half(x),__half(y))))
SPECIALIZE_REDUCE(FuncMinMax, __nv_fp8_e4m3, 2, __nv_fp8x2_e4m3, __nv_fp8x2_e4m3(fn.isMinNotMax ? __hmin2(__half2(x),__half2(y)) : __hmax2(__half2(x),__half2(y))))
SPECIALIZE_REDUCE(FuncSum, __nv_fp8_e5m2, 1, __nv_fp8_e5m2, __nv_fp8_e5m2(__hadd(__half(x),__half(y))))
SPECIALIZE_REDUCE(FuncSum, __nv_fp8_e5m2, 2, __nv_fp8x2_e5m2, __nv_fp8x2_e5m2(__hadd2(__half2(x),__half2(y))))
SPECIALIZE_REDUCE(FuncProd, __nv_fp8_e5m2, 1, __nv_fp8_e5m2, __nv_fp8_e5m2(__hmul(__half(x),__half(y))))
SPECIALIZE_REDUCE(FuncProd, __nv_fp8_e5m2, 2, __nv_fp8x2_e5m2, __nv_fp8x2_e5m2(__hmul2(__half2(x),__half2(y))))
SPECIALIZE_REDUCE(FuncMinMax, __nv_fp8_e5m2, 1, __nv_fp8_e5m2, __nv_fp8_e5m2(fn.isMinNotMax ? __hmin(__half(x), __half(y)) : __hmax(__half(x), __half(y))))
SPECIALIZE_REDUCE(FuncMinMax, __nv_fp8_e5m2, 2, __nv_fp8x2_e5m2, __nv_fp8x2_e5m2(fn.isMinNotMax ? __hmin2(__half2(x), __half2(y)) : __hmax2(__half2(x), __half2(y))))
#endif
#endif
#undef SPECIALIZE_REDUCE
////////////////////////////////////////////////////////////////////////////////
@@ -416,9 +440,9 @@ template<>
struct FuncPreMulSum<half> {
using EltType = half;
#if __CUDA_ARCH__ >= 530 && __CUDA_ARCH__ != 610
half2 scalar;
__half2 scalar;
__device__ FuncPreMulSum(uint64_t opArg=0) {
union { uint64_t u64; half val; };
union { uint64_t u64; __half val; };
u64 = opArg;
scalar.x = val;
scalar.y = val;
@@ -426,9 +450,9 @@ struct FuncPreMulSum<half> {
#else
float scalar;
__device__ FuncPreMulSum(uint64_t opArg=0) {
union { uint64_t u64; half val; };
union { uint64_t u64; __half val; };
u64 = opArg;
scalar = __half2float(val);
scalar = (float)val;
}
#endif
};
@@ -459,11 +483,39 @@ struct FuncPreMulSum<half> {
};
#endif
template<typename T>
struct Apply_Reduce<FuncPreMulSum<T>, /*EltPerPack=*/1> {
__device__ static BytePack<sizeof(T)> reduce(FuncPreMulSum<T> fn, BytePack<sizeof(T)> a, BytePack<sizeof(T)> b) {
#if defined(__CUDA_FP8_TYPES_EXIST__)
#if __CUDA_ARCH__ >= 900
template<>
struct FuncPreMulSum<__nv_fp8_e4m3> {
using EltType = __nv_fp8_e4m3;
__half2 scalar2;
__device__ FuncPreMulSum(uint64_t opArg) {
union { uint64_t u64; __nv_fp8_storage_t val; };
u64 = opArg;
scalar2.x = __half(__nv_cvt_fp8_to_halfraw(val, __NV_E4M3));
scalar2.y = scalar2.x;
}
};
template<>
struct FuncPreMulSum<__nv_fp8_e5m2> {
using EltType = __nv_fp8_e5m2;
__half2 scalar2;
__device__ FuncPreMulSum(uint64_t opArg) {
union { uint64_t u64; __nv_fp8_storage_t val; };
u64 = opArg;
scalar2.x = __half(__nv_cvt_fp8_to_halfraw(val, __NV_E5M2));
scalar2.y = scalar2.x;
}
};
#endif
#endif
template<typename T, int EltPerPack>
struct Apply_Reduce<FuncPreMulSum<T>, EltPerPack> {
__device__ static BytePack<EltPerPack*sizeof(T)> reduce(FuncPreMulSum<T> fn, BytePack<EltPerPack*sizeof(T)> a, BytePack<EltPerPack*sizeof(T)> b) {
// FuncPreMulSum reduce dispatches to FuncSum.
return Apply_Reduce<FuncSum<T>, 1>::reduce(FuncSum<T>(), a, b);
return Apply_Reduce<FuncSum<T>, EltPerPack>::reduce(FuncSum<T>(), a, b);
}
};
@@ -530,6 +582,51 @@ struct Apply_PreOp<FuncPreMulSum<half>, /*EltPerPack=*/1> {
#endif
#endif
////////////////////////////////////////////////////////////////////////////////
// Apply_PreOp of FuncPreMulSum for fp8.
#if defined(__CUDA_FP8_TYPES_EXIST__)
#if __CUDA_ARCH__ >= 900
template<>
struct Apply_PreOp<FuncPreMulSum<__nv_fp8_e4m3>, /*EltPerPack=*/1> {
static constexpr bool IsIdentity = false;
__device__ static BytePack<sizeof(__nv_fp8_e4m3)> preOp(
FuncPreMulSum<__nv_fp8_e4m3> fn, BytePack<sizeof(__nv_fp8_e4m3)> a
) {
return toPack<__nv_fp8_e4m3>(__nv_fp8_e4m3(__hmul(__half(fromPack<__nv_fp8_e4m3>(a)), fn.scalar2.x)));
}
};
template<>
struct Apply_PreOp<FuncPreMulSum<__nv_fp8_e4m3>, /*EltPerPack=*/2> {
static constexpr bool IsIdentity = false;
__device__ static BytePack<sizeof(__nv_fp8x2_e4m3)> preOp(
FuncPreMulSum<__nv_fp8_e4m3> fn, BytePack<sizeof(__nv_fp8x2_e4m3)> a
) {
return toPack<__nv_fp8x2_e4m3>(__nv_fp8x2_e4m3(__hmul2(__half2(fromPack<__nv_fp8x2_e4m3>(a)), fn.scalar2)));
}
};
template<>
struct Apply_PreOp<FuncPreMulSum<__nv_fp8_e5m2>, /*EltPerPack=*/1> {
static constexpr bool IsIdentity = false;
__device__ static BytePack<sizeof(__nv_fp8_e5m2)> preOp(
FuncPreMulSum<__nv_fp8_e5m2> fn, BytePack<sizeof(__nv_fp8_e5m2)> a
) {
return toPack<__nv_fp8_e5m2>(__nv_fp8_e5m2(__hmul(__half(fromPack<__nv_fp8_e5m2>(a)), fn.scalar2.x)));
}
};
template<>
struct Apply_PreOp<FuncPreMulSum<__nv_fp8_e5m2>, /*EltPerPack=*/2> {
static constexpr bool IsIdentity = false;
__device__ static BytePack<sizeof(__nv_fp8x2_e5m2)> preOp(
FuncPreMulSum<__nv_fp8_e5m2> fn, BytePack<sizeof(__nv_fp8x2_e5m2)> a
) {
return toPack<__nv_fp8x2_e5m2>(__nv_fp8x2_e5m2(__hmul2(__half2(fromPack<__nv_fp8x2_e5m2>(a)), fn.scalar2)));
}
};
#endif
#endif
////////////////////////////////////////////////////////////////////////////////
// FuncSumPostDiv
@@ -541,34 +638,44 @@ struct RedOpArg<FuncSumPostDiv<T>> {
}
};
template<typename T, bool IsFloating=IsFloatingPoint<T>::value>
struct FuncSumPostDiv_IntOnly;
template<typename T>
struct FuncSumPostDiv: FuncSumPostDiv_IntOnly<T> {
__device__ FuncSumPostDiv(uint64_t opArg=0):
FuncSumPostDiv_IntOnly<T>(opArg) {
struct FuncSumPostDiv {
static_assert(T(0) < T(-1), "FuncSumPostDiv is only for implementing ncclAvg on uint types.");
using EltType = T;
using UintType = typename std::conditional<sizeof(T)==8, uint64_t, uint32_t>::type;
uint32_t divisor:31, isSigned:1;
UintType recip;
__device__ FuncSumPostDiv(uint64_t opArg=0) {
isSigned = opArg & 1;
divisor = opArg >> 1;
recip = UintType(-1)/divisor;
}
__device__ T divide(T x) {
// x is negative iff we are in signed mode and the top bit is set
bool xneg = isSigned && (x & ~(T(-1)>>1));
// Compute abs(x):
// T(-x) vs -T(x) is critical. We have to negate then truncate the bits. Consider
// if we are doing signed 8-bit types, thus T=uint8_t. The value -1 is encoded
// as 0xff. -T(0xff) when promoted to 32-bit (which is implicit by compiler)
// gives 0xffffff01, but T(-0xff) is 0x1, and that is the abs value we want.
UintType xabs = xneg ? T(-x) : x;
// Compute quotient by multiplying by reciprical.
UintType q = sizeof(T)==8 ? __umul64hi(xabs, recip) : __umulhi(xabs, recip);
// Quotient may be off by one so do a fixup.
if (xabs - q*divisor >= divisor) q += 1;
// If original x was negative then we have to negate it back since we were
// working with its abs val.
return xneg ? -T(q) : T(q);
}
};
template<typename T>
struct FuncSumPostDiv_IntOnly<T, /*IsFloating=*/false>: FuncSum<T> {
using EltType = T;
int divisor;
__device__ FuncSumPostDiv_IntOnly(uint64_t opArg=0): divisor(opArg) {}
};
template<typename T>
struct FuncSumPostDiv_IntOnly<T, /*IsFloating=*/true> {
static_assert(sizeof(T)!=sizeof(T), "FuncSumPostDiv is only for implementing ncclAvg on integral types.");
};
template<typename T>
struct Apply_Reduce<FuncSumPostDiv<T>, /*EltPerPack=*/1>:
Apply_Reduce<FuncSum<T>, 1> {
__device__ static BytePack<sizeof(T)> reduce(FuncSumPostDiv<T> fn, BytePack<sizeof(T)> a, BytePack<sizeof(T)> b) {
template<typename T, int EltPerPack>
struct Apply_Reduce<FuncSumPostDiv<T>, EltPerPack>:
Apply_Reduce<FuncSum<T>, EltPerPack> {
__device__ static BytePack<EltPerPack*sizeof(T)> reduce(FuncSumPostDiv<T> fn, BytePack<EltPerPack*sizeof(T)> a, BytePack<EltPerPack*sizeof(T)> b) {
// FuncSumPostDiv reduce dispatches to FuncSum.
return Apply_Reduce<FuncSum<T>, 1>::reduce(FuncSum<T>(), a, b);
return Apply_Reduce<FuncSum<T>, EltPerPack>::reduce(FuncSum<T>(), a, b);
}
};
@@ -576,7 +683,7 @@ template<typename T>
struct Apply_PostOp<FuncSumPostDiv<T>, /*EltPerPack=*/1> {
static constexpr bool IsIdentity = false;
__device__ static BytePack<sizeof(T)> postOp(FuncSumPostDiv<T> fn, BytePack<sizeof(T)> a) {
return toPack<T>(fromPack<T>(a) / fn.divisor);
return toPack<T>(fn.divide(fromPack<T>(a)));
}
};