In recent versions NCCL introduced support for "user buffer registration", i.e., allowing user-owned memory (such as regular PyTorch tensors) to be "registered" (pinned, page-locked, etc.) with all the various hardware (NVLink, InfiniBand, ...) in order to support zero-copy transfers and thus accelerate communication and reduce resource footprint of NCCL's kernels (which reduces contention).
This was already exposed in PyTorch through a custom allocator provided by the NCCL process group. DDP already uses this, via a memory pool to allow caching and reusing.
FSDP2 is also particularly suited to leverage user buffer registration because the buffers it passes to NCCL are allocated by FSDP2 itself, since it anyways needs to (de)interleave the parameters to/from these private buffers.
This PR adds an extra flag to FSDP2 that tells it to use the ProcessGroup allocator for these private buffers, thus allowing it to leverage NCCL zero-copy (when supported).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150564
Approved by: https://github.com/kwen2501, https://github.com/weifengpy, https://github.com/syed-ahmed
We arrive at a point when so many files are related to symmetric memory and files are scattered around in the cpp side. Let's first put all related code (symmetric memory related) into a separate folder. We can do further refactoring later if needed.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155573
Approved by: https://github.com/fegin, https://github.com/d4l3k
Summary: Gloo PG doesn't release GIL, which results in python code hanging until the destructor completes. The destructor waits for all work on the PG to complete which can take a long time.
Test Plan: Ran
```
$ pytest --log-cli-level=INFO -vs torchft/local_sgd_integ_test.py
```
with a large timeout on the async work. Call to `gil_scoped_release` doesn't show up in the gdb stack trace.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154976
Approved by: https://github.com/d4l3k, https://github.com/dcci, https://github.com/fduwjj
C++ Reducer is silently incorrect under CA, its implementation is no-oping the collective. I'm guessing that it was no-op'd because in DDP + python reducer, the C++ reducer is still being initialized.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152735
Approved by: https://github.com/fegin
ghstack dependencies: #153300, #152689
This adds a non-blocking mode to queue_pop. This allows for workers to poll if work is ready without blocking the main loop. This is useful for the case where you want to have a GPU have maximum utilization when something only periodically is sent on the queue.
We also expose a `torch.distributed.QueueEmptyError` so users can catch the error and handle it accordingly.
Test plan:
```
pytest test/distributed/test_store.py -k queue -v -s -x
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151485
Approved by: https://github.com/fduwjj, https://github.com/tianfengfrank
Summary: add one option to allow skipping all reduce unused parameters, this could help improve training throughput significantly when the number of unused parameters is large in the model.
Test Plan: unit tests, CI
Differential Revision: D72282069
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151503
Approved by: https://github.com/mrshenli
This adds queue operations as described in https://github.com/pytorch/pytorch/issues/150943.
This works by adding two new operations `queue_push` and `queue_pop`. The semantics are designed to be blocking with a timeout. Pushing will always succeed as the queue is infinite size. Popping will first call `wait` until the key is ready and then pop the value from the queue.
This implements queues for only: HashStore, TCPStore w/ libuv. FileStore and the legacy backends are not supported.
`wait` and `check` work for queue operations though queue_push will only wake up the first waiter rather than all of them.
This also has a few cleanups to error types/documentation in related code.
Example trace:
```
[I409 16:51:43.963833529 TCPStoreLibUvBackend.cpp:829] [c10d - trace] validate magic:1015412686 address:[localhost]:55816
[I409 16:51:43.963845838 TCPStoreLibUvBackend.cpp:842] [c10d - trace] ping nonce:2840795 address:[localhost]:55816
[I409 16:51:43.963902914 TCPStoreLibUvBackend.cpp:911] [c10d - trace] add key:init/ val:1 address:[localhost]:55816
[I409 16:51:43.963939389 TCPStoreLibUvBackend.cpp:977] [c10d - trace] wait key_count:1 keys[0]:init/ address:[localhost]:55816
[I409 16:51:43.963974842 TCPStoreLibUvBackend.cpp:893] [c10d - trace] get key:init/ address:[localhost]:55816
[I409 16:51:43.964071909 TCPStoreLibUvBackend.cpp:1121] [c10d - trace] queue_push key:/test_prefix/test_queue_support address:[localhost]:55816
[I409 16:51:43.964080221 TCPStoreLibUvBackend.cpp:940] [c10d - trace] check key_count:1 keys[0]:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964108584 TCPStoreLibUvBackend.cpp:1121] [c10d - trace] queue_push key:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964123207 TCPStoreLibUvBackend.cpp:1121] [c10d - trace] queue_push key:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964128194 TCPStoreLibUvBackend.cpp:940] [c10d - trace] check key_count:1 keys[0]:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964156347 TCPStoreLibUvBackend.cpp:977] [c10d - trace] wait key_count:1 keys[0]:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964187493 TCPStoreLibUvBackend.cpp:977] [c10d - trace] wait key_count:1 keys[0]:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964217709 TCPStoreLibUvBackend.cpp:1133] [c10d - trace] queue_pop key:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964324300 TCPStoreLibUvBackend.cpp:977] [c10d - trace] wait key_count:1 keys[0]:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964354495 TCPStoreLibUvBackend.cpp:1133] [c10d - trace] queue_pop key:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964416299 TCPStoreLibUvBackend.cpp:940] [c10d - trace] check key_count:1 keys[0]:/test_prefix/foo address:[localhost]:55816
[I409 16:51:43.964458733 TCPStoreLibUvBackend.cpp:977] [c10d - trace] wait key_count:1 keys[0]:/test_prefix/non_existant address:[localhost]:55816
[W409 16:51:43.974516585 socket.cpp:460] [c10d] waitForInput: poll for socket SocketImpl(fd=75, addr=[localhost]:55816, remote=[localhost]:46641) returned 0, likely a timeout
[W409 16:51:43.974559169 socket.cpp:485] [c10d] waitForInput: socket SocketImpl(fd=75, addr=[localhost]:55816, remote=[localhost]:46641) timed out after 10ms
[I409 16:51:43.974600451 TCPStoreLibUvBackend.cpp:1101] [c10d - trace] cancel_wait address:[localhost]:55816
```
Test plan:
```
$ pytest test/distributed/test_store.py -k queue -v -s
test/distributed/test_store.py::FileStoreTest::test_queues SKIPPED [0.4351s] (Store does not support queues)
test/distributed/test_store.py::HashStoreTest::test_queues PASSED [0.0009s]
test/distributed/test_store.py::PrefixFileStoreTest::test_queues SKIPPED [0.0006s] (Store does not support queues)
test/distributed/test_store.py::TCPStoreTest::test_queues SKIPPED [0.0012s] (Store does not support queues)
test/distributed/test_store.py::LibUvTCPStoreTest::test_queues PASSED [0.0014s]
test/distributed/test_store.py::PrefixTCPStoreTest::test_queues PASSED [0.0014s]
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150969
Approved by: https://github.com/XilunWu, https://github.com/fduwjj
This adds a new `clone()` method to Store which will return a new Store instance that can be used from a different thread.
This is intended to better support multiple threads with stores such as when ProcessGroupNCCL needs a store to do error propagation.
Related issue: https://github.com/pytorch/pytorch/issues/150943
Test plan:
```
pytest test/distributed/test_store.py -k PythonStore
pytest test/distributed/test_store.py -k clone
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150966
Approved by: https://github.com/fduwjj
This adds lazy initialization support to ProcessGroupGloo via `TORCH_GLOO_LAZY_INIT` or via `create_device(..., lazy_init=True)`
This is still a draft PR as there's one race condition when doing coalesced operations that needs to be fixed upstream in Gloo first. Depends on https://github.com/facebookincubator/gloo/pull/427 landing first
This also updates the gloo submodule to include the required changes.
Test plan:
added lazy init test variants
```
pytest -v test/distributed/test_c10d_gloo.py -k Lazy
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150801
Approved by: https://github.com/fduwjj
Relanding #148590 due to merge conflict.
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.
Squashed contents:
* [ptd][nccl] use current-stream as nccl-stream under async=False mode (#147820)
PTD current workflow:
- PTD creates its own dedicated `ncclStream` for comm operation
- it will first add a dependency on current-stream (typically the compute stream) to ensure tensors are ready before invoking collective
such stream synchronization become expensive in Inference world (cpu overhead: 70us vs GPU kernel time: 160us).
This diff:
- async=False [default], will use current-stream as nccl-stream and avoid the stream-sync overhead
- async=True, will retain existing logic: create new nccl-stream, let it wait on current-stream to ensure tensors are ready
- pass down async from c10d down to NCCL-PG
this helps shave off 50% CPU overhead **(70us -> 35us)**, which reduce total CPU/GPU from **230us to 195us by 15%**
* [PGNCCL] Make avoid-record-stream default
* [c10d] Add asyncOp argument to Ops
* Change python side wait
* Pass asyncOp at ProcessGroup level
* Watchdog unstashing tensors as a safety net
* Stash tensors for reduce_scatter_v and all_gather_v
Pull Request approved: https://github.com/pytorch/pytorch/pull/149753
* [c10d] Move unstashing from watchdog to main thread
Pull Request approved: https://github.com/pytorch/pytorch/pull/150079
* [PGNCCL][BE] Merge mutex into TensorShelf for encapsulation
Pull Request approved: https://github.com/pytorch/pytorch/pull/150130
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150398
Approved by: https://github.com/atalman
This supports `masterListenFd` which is required for full compatibility with the non-libuv TCPStore. The code was just missing a `uv_listen` call and now it works just fine.
This is required to migrate the last remaining uses of TCPStore off of the non-libuv backend.
Test plan:
```
pytest -v test/distributed/test_store.py -k test_take_over_listen_socket
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150215
Approved by: 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: [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 has two fixes to improve IPC tensor release performance when using torchft's BabyProcessGroupNCCL.
1. release the IpcMutex when deleting the `ExpandableSegements` object to avoid synchronizing under the lock
2. release the GIL in WorkNCCL destructor since the shared tensor will be destructed there
Test plan:
Run with torchft + torchtitan
```
REPLICA_GROUP_ID=0 NGPU=2 CUDA_VISIBLE_DEVICES=0,1 CONFIG_FILE=./torchtitan/models/llama/train_configs/llama3_8b.toml ./run_train.sh --training.data_par
allel_shard_degree=2 --fault_tolerance.enable --fault_tolerance.group_size=2 --fault_tolerance.replica_id=0 --metrics.log_freq=1 --training.seq_len 4096
...
[rank0]:[titan] 2025-03-07 17:51:31,387 - root - INFO - step: 61 loss: 7.4825 memory: 79.73GiB(83.89%) tps: 317 tflops: 16.34 mfu: 1.65%
```
Check py-spy to verify no bottleneck on IPC lock when creating new shared tensors
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148805
Approved by: https://github.com/Skylion007, https://github.com/fegin, https://github.com/zdevito
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
This PR adds support for non-functional collectives under `FakeTensorMode` and `fake_pg`. It helps eliminate the patching of collectives for memory and runtime estimation.
It also modifies the `ModTracker` to enable the post-backward hook call for modules whose inputs don't require gradients but parameters do.
For the memory tracking, we now enable tracking DTensor dispatcher for custom dispatch functions like `entropy_loss`.
Dispatcher is only enabled for the memory tracking part and disabled as soon as it is done.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/147566
Approved by: https://github.com/weifengpy
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
@fegin found an issue where torchft is not compatible with functional collectives.
Found in https://github.com/pytorch/torchtitan/pull/806
The root cause is because PyProcessGroup/PyWork are not compatible with functional collectives due to a nasty ownership bug.
PyWork relies on a pybind trampoline to propagate requests to Python unfortunately the way Pybind works is that the Python object owns the C++ object rather than some form of shared ownership. Thus what happens is that the PyWork Python object will collected when returned to C++ from the PyProcessGroup but the C++ PyWork object still exists. When the PyWork object is used, this causes a deadlock as the corresponding Python object no longer exists
To solve this, we introduce a new `PyWorkHolder` class which holds a reference to the `py::object` as well as the trampoline class. This resolves any dependency issues since we can now hold ownership in C++ to both the Python and C++ objects.
To make this cleaner we introduce a `WORK_OVERRIDE` macro which is a patched version of `PYBIND11_OVERRIDE` that returns a `PyWorkHolder` rather than just `PyWork` and use for all collectives in PyProcessGroup.
Test plan:
```
cd pytorch
pytest test/distributed/test_c10d_functional_native.py
```
```
cd torchft
pytest torchft/process_group_test.py -k functional -v -x -s
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/146376
Approved by: https://github.com/yifuwang
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
`py::call_guard<py::gil_scoped_release>` is not safe when using multiple threads. This instead moves it into the init function which is safe.
For more details see #143593https://github.com/pybind/pybind11/issues/5473
Test plan:
```
python setup.py develop
```
CI
```py
import time
from concurrent.futures import ThreadPoolExecutor
from torch import distributed as dist
def run():
store = dist.TCPStore(
host_name="localhost",
port=0,
is_master=True,
wait_for_workers=False,
)
# this sleep is required to trigger the crash
time.sleep(0.1)
del store
futures = []
with ThreadPoolExecutor(
max_workers=100,
) as executor:
for i in range(100000):
print(i)
futures.append(executor.submit(run))
if len(futures) > 100:
futures.pop(0).result()
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/143598
Approved by: https://github.com/c-p-i-o
