This PR has multiple changes to `ProcessGroupNCCL` (which unfortunately are related):
1. When async_op=False, we directly launch the collective on "current" stream, instead of a trampoline stream and join back.
- Resolves#147729
- Resolves#146881
- Also saves two event syncs (which have overhead in case of HIP) and one pybind when we call `work.wait()` in distributed_c10d.py on behalf of user.
2. Entirely remove `record_stream` and use CPU-side stashing for managing tensor lifetime against recycling.
- Resolves#147168
3. Remove tensor life management when async_op=False; only use it when async_op=True.
4. To guard against user not calling `work.wait()`, we ask watchdog to unstash tensors after detecting completion of collectives, to prevent us from holding reference to tensors forever. This is a safety net, rather than a service guarantee, see discussion [here](https://github.com/pytorch/pytorch/issues/147168#issuecomment-2660142460).
5. Profile in async_op=False mode would look different -- collective kernels would show up in the same line and compute kernels.
Joint work with @cenzhaometa who wants to remove the event sync overhead.
Cc: @ngimel @awgu @Aidyn-A @skyw @wconstab @leonardo0lyj
Differential Revision: [D70937982](https://our.internmc.facebook.com/intern/diff/D70937982)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148590
Approved by: https://github.com/eqy, https://github.com/Aidyn-A, https://github.com/fduwjj
This PR has multiple changes to `ProcessGroupNCCL` (which unfortunately are related):
1. When async_op=False, we directly launch the collective on "current" stream, instead of a trampoline stream and join back.
- Resolves#147729
- Resolves#146881
- Also saves two event syncs (which have overhead in case of HIP) and one pybind when we call `work.wait()` in distributed_c10d.py on behalf of user.
2. Entirely remove `record_stream` and use CPU-side stashing for managing tensor lifetime against recycling.
- Resolves#147168
3. Remove tensor life management when async_op=False; only use it when async_op=True.
4. To guard against user not calling `work.wait()`, we ask watchdog to unstash tensors after detecting completion of collectives, to prevent us from holding reference to tensors forever. This is a safety net, rather than a service guarantee, see discussion [here](https://github.com/pytorch/pytorch/issues/147168#issuecomment-2660142460).
5. Profile in async_op=False mode would look different -- collective kernels would show up in the same line and compute kernels.
Joint work with @cenzhaometa who wants to remove the event sync overhead.
Cc: @ngimel @awgu @Aidyn-A @skyw @wconstab @leonardo0lyj
Differential Revision: [D70835197](https://our.internmc.facebook.com/intern/diff/D70835197)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148590
Approved by: https://github.com/eqy, https://github.com/Aidyn-A, https://github.com/fduwjj
This PR has multiple changes to `ProcessGroupNCCL` (which unfortunately are related):
1. When async_op=False, we directly launch the collective on "current" stream, instead of a trampoline stream and join back.
- Resolves#147729
- Resolves#146881
- Also saves two event syncs (which have overhead in case of HIP) and one pybind when we call `work.wait()` in distributed_c10d.py on behalf of user.
2. Entirely remove `record_stream` and use CPU-side stashing for managing tensor lifetime against recycling.
- Resolves#147168
3. Remove tensor life management when async_op=False; only use it when async_op=True.
4. To guard against user not calling `work.wait()`, we ask watchdog to unstash tensors after detecting completion of collectives, to prevent us from holding reference to tensors forever. This is a safety net, rather than a service guarantee, see discussion [here](https://github.com/pytorch/pytorch/issues/147168#issuecomment-2660142460).
5. Profile in async_op=False mode would look different -- collective kernels would show up in the same line and compute kernels.
Joint work with @cenzhaometa who wants to remove the event sync overhead.
Cc: @ngimel @awgu @Aidyn-A @skyw @wconstab @leonardo0lyj
Differential Revision: [D70835197](https://our.internmc.facebook.com/intern/diff/D70835197)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148590
Approved by: https://github.com/eqy, https://github.com/Aidyn-A, https://github.com/fduwjj
This adds `abort` and `shutdown` to `Backend` and `ProcessGroup` objects. This simplifies the logic in `distributed_c10d.py` by having a default noop implementation for all PGs.
This will be useful for torchft and upcoming versions of NCCL which will handle abort correctly. Currently `torchft` would have to call internal methods `_abort` on the PGNCCL object directly but with this change we can now just call `.abort()` and have it work for any PG implementation.
Test plan:
```
pytest distributed/test_backends.py distributed/test_c10d_common.py distributed/test_c10d_pypg.py
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148798
Approved by: https://github.com/kwen2501
1. My company is using privateuseone to connect new hardware device and requires the use of `batch_isend_irecv` function. However, `batch_isend_irecv` is currently only open to CUDA, so I add `supports_coalescing` property in `c10d::Backend` to determine whether backend supports coalescing.
2. If `pg._has_hooks` return True, We don't need to determine if the current device is CUDA. So privateuseone can also support `pg._wait_for_pending_works`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/135338
Approved by: https://github.com/kwen2501, https://github.com/albanD
So that NVLink SHARP comes with zero-copy on H100+ platforms, for DDP applications.
Less SM usage, less memory contention between NCCL kernel and compute kernels.
Added env `DDP_DISABLE_COMM_MEM` as a back-out option:
```
An environment variable to disable comm-optimized memory pool.
Default is 0, which means comm-optimized memory pool is enabled.
Users can set it to 1 in case of seeing regression or OOM (because this
comm MemPool may not share space with regular compute MemPool).
```
Differential Revision: [D69297766](https://our.internmc.facebook.com/intern/diff/D69297766)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/146589
Approved by: https://github.com/syed-ahmed, https://github.com/c-p-i-o, https://github.com/fduwjj
This PR implements a small UI improvement over #133603.
It prepares a NCCL memory allocator in torch cpp and then pybind's it out, so that user can directly use it.
UI:
```
pool = torch.cuda.MemPool(backend.mem_allocator)
with torch.cuda.use_mem_pool(pool):
tensor = torch.arange(1024 * 1024 * 2, device=device)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145675
Approved by: https://github.com/syed-ahmed, https://github.com/wconstab
This PR implements a small UI improvement over #133603.
It prepares a NCCL memory allocator in torch cpp and then pybind's it out, so that user can directly use it.
UI:
```
pool = torch.cuda.MemPool(backend.mem_allocator)
with torch.cuda.use_mem_pool(pool):
tensor = torch.arange(1024 * 1024 * 2, device=device)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145675
Approved by: https://github.com/syed-ahmed, https://github.com/wconstab
This PR implements a small UI improvement over #133603.
It prepares a NCCL memory allocator in torch cpp and then pybind's it out, so that user can directly use it.
UI:
```
pool = torch.cuda.MemPool(backend.mem_allocator)
with torch.cuda.use_mem_pool(pool):
tensor = torch.arange(1024 * 1024 * 2, device=device)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/145675
Approved by: https://github.com/syed-ahmed, https://github.com/wconstab
Summary:
This PR is basically a replacement of
https://github.com/pytorch/pytorch/pull/140087, which caused some perf
drop due to frequent TCPStore check in watchdog thread. The fix is to move the
tcpstore check in monitoring thread
If unhealthy, the user should be able to get the type of errors, e.g.,
timeout,nccl error or remote error.
This API is applied to PG level, compared to the
work.get_future_result() API which is applied to Work Level.
Error detection at PG level is much more convenient for users to handle
the PG failure as a whole, e.g, restarting the PG.
Error handling at the work level is still useful for users to attach
work specific context and debug the RC of the specific failing
work/collective
Note it is critical for all ranks in the PG to be notified about an
error as soon as it occurs, so we introduce an errorType of
REMOTE_ERROR, which is 'broadcasted' from a src rank (which detects a
local error) to all other ranks in the PG, the broadcast is done through
TCPStore currently
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/144498
Approved by: https://github.com/kwen2501
Reported in https://github.com/pytorch/pytorch/issues/143470, we have a dangling references in `CudaEventCache`. So we want to fix it.
1. We add a unit test to repro the issue mentioned in the issue.
2. Instead of converting variables to shared pointers as suggested in the issue, we then make the cache itself a shared pointer. So if the thread creates the cache dies before all events get recycled, the cache is still there until the last CudaEvent get deleted. (thanks for the suggestion from @kwen2501 )
Pull Request resolved: https://github.com/pytorch/pytorch/pull/144496
Approved by: https://github.com/kwen2501
Summary:
This is an attempt to improve the flight recorder efficacy.
We have a small subset of jobs that are timing out (i.e. failing to write out FR logs in 1 minute) and some that are throwing a `std::exception - broken promise`.
There are two changes in here.
1. We attempt to write out FR buffer with stack traces. If this fails, we attempt to capture FR buffer again - but this time without stack traces. The assumption here is that FR could be locking up when unwinding stack.
Note, to keep things simple, I'm re-using the same file name for both with/without stack_trace.
2. Add additional catch statements in the Manifold writer. There might be something going on in here - so we'll get a log statement if this is failing.
TODO:
- there's nothing differentiating in the output that says whether stack traces were omitted purposefully or not.
This info might be useful for the analyzer - so I'll add this in a follow on diff.
Test Plan: Unit tests.
Differential Revision: D66843194
Pull Request resolved: https://github.com/pytorch/pytorch/pull/142178
Approved by: https://github.com/kwen2501
We added `CudaEventCache` in https://github.com/pytorch/pytorch/pull/133727 and this is a feature which tries to reuse CudaEvent so that we don't call destroy of CudaEvent which causes hang in the past. We had a bunch of tests and testing on TorchTitan and internal workload already. So far no errors or crash are found at the moment so we decide to roll out to all OSS users. For internal workload, this PR would not affect it because of some internal gating.
Also we observed some multi-device use cases in OSS, so that we want to bring back multi-device support originally proposed in https://github.com/pytorch/pytorch/pull/122732/files.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/140975
Approved by: https://github.com/eqy, https://github.com/kwen2501
Summary:
If unhealthy, the user should be able to get the type of errors, e.g.,
timeout,nccl error or remote error.
This API is applied to PG level, compared to the work.get_future_result() API which is applied to Work Level.
Error detection at PG level is much more convenient for users to handle the PG failure as a whole, e.g, restarting the PG.
Error handling at the work level is still useful for users to attach work specific context and debug the RC of the specific failing work/collective
Note it is critical for all ranks in the PG to be notified about an error as soon as it occurs, so we introduce an errorType of REMOTE_ERROR, which is 'broadcasted' from a src rank (which detects a local error) to all other ranks in the PG, the broadcast is done through TCPStore currently
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/140087
Approved by: https://github.com/kwen2501
1. My company is using privateuseone to connect new hardware device and requires the use of `batch_isend_irecv` function. However, `batch_isend_irecv` is currently only open to CUDA, so I add `supports_coalescing` property in `c10d::Backend` to determine whether backend supports coalescing.
2. If `pg._has_hooks` return True, We don't need to determine if the current device is CUDA. So privateuseone can also support `pg._wait_for_pending_works`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/135338
Approved by: https://github.com/kwen2501
- Refactored traceback code into `work.printTraceback()`. cc @H-Huang @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o @shuqiangzhang
- Refactored desync debug code into `class DesyncDebugger`.
- Moved occurrences of `futureWorkResult_->markCompleted` into `checkAndSetException` and `checkTimeout`, respectively. cc @shuqiangzhang
- Modularized dump signal broadcast code into `ProcessGroupNCCL::broadcastDumpSignal`. cc @fduwjj @c-p-i-o
Pull Request resolved: https://github.com/pytorch/pytorch/pull/139834
Approved by: https://github.com/shuqiangzhang
### Why use non-blocking mode in eager init?
For overlapping comm init and model init, etc.
### Why can we set non-blocking as default?
If the setting is dangling -- i.e. not passed in by user nor set via env -- `ProcessGroupNCCL` can have some preferred logic. And torch-level API semantics does not change whether the NCCL comm is blocking or non-blocking (handled within `ProcessGroupNCCL`).
### Why not make non-blocking default for lazy mode as well?
PR https://github.com/pytorch/pytorch/pull/137544 tried it.
Two reasons why that's not preferred today:
1. It is hard -- too big a blast.
2. There is no gain by doing lazy init in non-blocking mode, because the right next CPU call is a collective, and we will block there waiting for comm to be ready, so same effect as blocked init, no "opening" compared to eager mode.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138527
Approved by: https://github.com/wconstab
ghstack dependencies: #138860
### Why use non-blocking mode in eager init?
For overlapping comm init and model init, etc.
### Why can we set non-blocking as default?
If the setting is dangling -- i.e. not passed in by user nor set via env -- `ProcessGroupNCCL` can have some preferred logic. And torch-level API semantics does not change whether the NCCL comm is blocking or non-blocking (handled within `ProcessGroupNCCL`).
### Why not make non-blocking default for lazy mode as well?
PR https://github.com/pytorch/pytorch/pull/137544 tried it.
Two reasons why that's not preferred today:
1. It is hard -- too big a blast.
2. There is no gain by doing lazy init in non-blocking mode, because the right next CPU call is a collective, and we will block there waiting for comm to be ready, so same effect as blocked init, no "opening" compared to eager mode.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138527
Approved by: https://github.com/wconstab
ghstack dependencies: #138860
Summary:
We always need to give the heartbeat monitor thread time to write out flight recorder dumps. Otherwise, the watchdog thread kills the heartbeat monitor thread too fast before it has time to write out the Flight Recorder logs.
This change:
1. Removes the "sleep after exception" JK. We don't need to sleep for 8 minutes.
2. Use a promise between watchdog thread and heartbeat monitor thread to delay, at most, one minute to give Flight Recorder time to write out it's log on timeout.
Test Plan:
Tested on my local job and flight recorder successfully executed for the job.
https://fburl.com/mlhub/38fj5yne
The watchdog thread gives heartbeat thread time to write out the logs.
In the logs we see:
```
[trainer4]:I1023 17:39:29.755507 12592 ProcessGroupNCCL.cpp:1950] [PG ID 0 PG GUID 0(precheck) Rank 12] slept for 1647ms giving time for flight recorder dumps to finish.
```
Differential Revision: D64857928
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138828
Approved by: https://github.com/d4l3k, https://github.com/fduwjj
### Why use non-blocking mode in eager init?
For overlapping comm init and model init, etc.
### Why can we set non-blocking as default?
If the setting is dangling -- i.e. not passed in by user nor set via env -- `ProcessGroupNCCL` can have some preferred logic. And torch-level API semantics does not change whether the NCCL comm is blocking or non-blocking (handled within `ProcessGroupNCCL`).
### Why not make non-blocking default for lazy mode as well?
PR https://github.com/pytorch/pytorch/pull/137544 tried it.
Two reasons why that's not preferred today:
1. It is hard -- too big a blast.
2. There is no gain by doing lazy init in non-blocking mode, because the right next CPU call is a collective, and we will block there waiting for comm to be ready, so same effect as blocked init, no "opening" compared to eager mode.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/138527
Approved by: https://github.com/wconstab
ghstack dependencies: #137855, #138488, #138374, #138384
