AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

2.21.5-1

Browse Source 
Add support for IB SHARP 1PPN operation with user buffers.
Improve support for MNNVL, add NVLS support and multi-clique support.
 * Detect the NVLS clique through NVML
 * Exchange XML between peers in the same NVLS clique and fuse XMLs
   before creating the topology graph.
 * Rework bootstrap allgather algorithms to allow for large allgather
   operations intra-node (XML exchange).
Net/IB: add support for dynamic GID detection.
 * Automatically select RoCEv2/IPv4 interface by default. Allow to
   select IPv6 or even the network/mask.
Reduce NVLS memory usage.
 * Add stepSize as property of a connection to allow for different
   sizes on different peers; set it to 128K for NVLink SHARP.
Improve tuner loading
 * Look for more paths, be more consistent with the network device
   plugin.
 * Also search for tuner support inside the net plugin.
Improve tuner API
 * Add context to support multi-device per process.
Add magic number around comm object to detect comm corruption.
 * Add some basic check around communicators so that we can report a
   problem when a communicator gets corrupted or a wrong comm pointer
   is passed to NCCL.
Fix net/IB error path. Github PR #1164
Fix collnet rail mapping with split comm.
Fix packet reordering issue causing bootstrap mismatch
 * Use a different tag in ncclTransportP2pSetup for the connectInfo
   exchange and the following barrier.
Fix hang when crossNic is inconsistent between ranks.
Fix minCompCap/maxCompCap computation. Github issue #1184
This commit is contained in:
Sylvain Jeaugey
2024-03-26 06:08:55 -07:00
parent 6dd51f15bf
commit ab2b89c4c3
52 changed files with 2124 additions and 944 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/device/all_reduce.h
+72 -29
View File
@@ -297,13 +297,21 @@ struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_COLLNET_DIRECT, NCC
}
} else {
// Directly send to network
Primitives<T, RedOp, FanAsymmetric<0, 1>, /*Direct=*/0, Proto, 0>
if (args->regUsed == NCCL_COLLNET_REG_BUFFER) {
if (tid == tidStartReduce) {
int steps = (int)divUp(size * sizeof(T), NCCL_MAX_COLLNET_SIZE);
Primitives<T, RedOp, FanAsymmetric<0, 1>, /*Direct=*/0, Proto, 0>::sendPeerNotify(direct->out, 1, steps);
}
__syncwarp();
} else {
Primitives<T, RedOp, FanAsymmetric<0, 1>, /*Direct=*/0, Proto, 0>
prims(tid-tidStartReduce, nThreadsReduce, nullptr, &direct->out, args->sendbuff, args->recvbuff,
args->redOpArg, 3*Proto::MaxGroupWidth, 1, 1);
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
ssize_t offset = gridOffset + (bid*direct->nHeads+direct->headRank)*chunkSize;
int nelem = min(chunkSize, size-offset);
prims.send(offset, nelem);
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
ssize_t offset = gridOffset + (bid*direct->nHeads+direct->headRank)*chunkSize;
int nelem = min(chunkSize, size-offset);
prims.send(offset, nelem);
}
}
}
} else if (tid < tidStartBcast && hasUp) {
@@ -328,14 +336,22 @@ struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_COLLNET_DIRECT, NCC
prims.recvCopyDirectSend(offset, nelem, /*postOp=*/true);
}
} else {
// Recv from network (no post thread needed)
Primitives<T, RedOp, FanAsymmetric<1, 0>, /*Direct=*/0, Proto, 0>
prims(tid-tidStartBcast, nThreadsBcast, &direct->out, nullptr, args->sendbuff, args->recvbuff,
args->redOpArg, 1*Proto::MaxGroupWidth, 0, 0);
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
ssize_t offset = gridOffset + (bid*direct->nHeads+direct->headRank)*chunkSize;
int nelem = min(chunkSize, size-offset);
prims.recv(offset, nelem, /*postOp=*/true);
if (args->regUsed == NCCL_COLLNET_REG_BUFFER) {
if (tid == tidStartBcast) {
int steps = (int)divUp(size * sizeof(T), NCCL_MAX_COLLNET_SIZE);
Primitives<T, RedOp, FanAsymmetric<1, 0>, /*Direct=*/0, Proto, 0>::recvPeerNotify(direct->out, 0, steps);
}
__syncwarp();
} else {
// Recv from network (no post thread needed)
Primitives<T, RedOp, FanAsymmetric<1, 0>, /*Direct=*/0, Proto, 0>
prims(tid - tidStartBcast, nThreadsBcast, &direct->out, nullptr, args->sendbuff, args->recvbuff,
args->redOpArg, 1 * Proto::MaxGroupWidth, 0, 0);
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
ssize_t offset = gridOffset + (bid * direct->nHeads + direct->headRank) * chunkSize;
int nelem = min(chunkSize, size - offset);
prims.recv(offset, nelem, /*postOp=*/true);
}
}
}
}
@@ -616,21 +632,31 @@ struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_COLLNET_CHAIN, NCCL
groupNthreads = nthreads-nthreadsSplit;
}
Primitives<T, RedOp, FanSymmetric<1>, /*Direct=*/1, Proto, 0>
prims(groupTid, groupNthreads, &recv, &send, args->sendbuff, args->recvbuff,
args->redOpArg, group*Proto::MaxGroupWidth, connIndex, connIndex);
if (tid < nthreadsSplit) {
if (recv == -1) {
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
ssize_t offset = gridOffset + bid*int(chunkSize);
int nelem = min(chunkSize, size-offset);
prims.send(offset, nelem);
if (args->regUsed == NCCL_COLLNET_REG_BUFFER) {
if (groupTid == 0) {
int steps = (int)divUp(size * sizeof(T), NCCL_MAX_COLLNET_SIZE);
Primitives<T, RedOp, FanSymmetric<1>, /*Direct=*/1, Proto, 0>::sendPeerNotify(send, connIndex, steps);
}
__syncwarp();
} else {
Primitives<T, RedOp, FanSymmetric<1>, /*Direct=*/1, Proto, 0>
prims(groupTid, groupNthreads, &recv, &send, args->sendbuff, args->recvbuff,
args->redOpArg, group * Proto::MaxGroupWidth, connIndex, connIndex);
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
ssize_t offset = gridOffset + bid * int(chunkSize);
int nelem = min(chunkSize, size - offset);
prims.send(offset, nelem);
}
}
} else {
Primitives<T, RedOp, FanSymmetric<1>, /*Direct=*/1, Proto, 0>
prims(groupTid, groupNthreads, &recv, &send, args->sendbuff, args->recvbuff,
args->redOpArg, group * Proto::MaxGroupWidth, connIndex, connIndex);
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
ssize_t offset = gridOffset + bid*int(chunkSize);
int nelem = min(chunkSize, size-offset);
ssize_t offset = gridOffset + bid * int(chunkSize);
int nelem = min(chunkSize, size - offset);
prims.recvReduceSend(offset, nelem);
}
}
@@ -639,19 +665,36 @@ struct RunWorkElement<ncclFuncAllReduce, T, RedOp, NCCL_ALGO_COLLNET_CHAIN, NCCL
if (recv == nranks) {
// I'm the first in the broadcast chain, I need to perform the division (postOp)
if (send == -1) {
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
ssize_t offset = gridOffset + bid*int(chunkSize);
int nelem = min(chunkSize, size-offset);
prims.recv(offset, nelem, /*postOp*/true);
if (args->regUsed == NCCL_COLLNET_REG_BUFFER) {
if (groupTid == 0) {
int steps = (int)divUp(size * sizeof(T), NCCL_MAX_COLLNET_SIZE);
Primitives<T, RedOp, FanSymmetric<1>, /*Direct=*/1, Proto, 0>::recvPeerNotify(recv, connIndex, steps);
}
__syncwarp();
} else {
Primitives<T, RedOp, FanSymmetric<1>, /*Direct=*/1, Proto, 0>
prims(groupTid, groupNthreads, &recv, &send, args->sendbuff, args->recvbuff,
args->redOpArg, group * Proto::MaxGroupWidth, connIndex, connIndex);
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
ssize_t offset = gridOffset + bid * int(chunkSize);
int nelem = min(chunkSize, size - offset);
prims.recv(offset, nelem, /*postOp*/true);
}
}
} else {
Primitives<T, RedOp, FanSymmetric<1>, /*Direct=*/1, Proto, 0>
prims(groupTid, groupNthreads, &recv, &send, args->sendbuff, args->recvbuff,
args->redOpArg, group * Proto::MaxGroupWidth, connIndex, connIndex);
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
ssize_t offset = gridOffset + bid*int(chunkSize);
int nelem = min(chunkSize, size-offset);
ssize_t offset = gridOffset + bid * int(chunkSize);
int nelem = min(chunkSize, size - offset);
prims.recvCopyDirectSend(offset, nelem, /*postOp*/true);
}
}
} else {
Primitives<T, RedOp, FanSymmetric<1>, /*Direct=*/1, Proto, 0>
prims(groupTid, groupNthreads, &recv, &send, args->sendbuff, args->recvbuff,
args->redOpArg, group * Proto::MaxGroupWidth, connIndex, connIndex);
if (send == -1) {
for (ssize_t gridOffset = 0; gridOffset < size; gridOffset += loopSize) {
ssize_t offset = gridOffset + bid*int(chunkSize);