as titled. It's sometimes confusing to use PlacementStrategy as a name,
as we also have OpStrategy and TupleStrategy, the latter two contain
the former, so it is better to make the naming clearer.
Renaming PlacementStrategy -> OpSpec as it is an operator spec that
contains output_spec + input_specs.
Also found some utils can be merged to OpSchema so included together in
this PR
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155592
Approved by: https://github.com/awgu
Summary: This diff enhances the `get_process_group_ranks()` function to accept `group=None` as an optional argument. This allows the function to return all ranks associated with the default process group if no group is specified.
Test Plan:
contbuild & OSS CI
Rollback Plan:
Differential Revision: D75817800
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154902
Approved by: https://github.com/wz337
As we prepare to support re-sharding, the current approach of using BytesStorageMetadata to read safetenstors won't work anymore. Before, we didn't need to read the metadata of the safetensors file from its header because we were just loading the contents of the file directly into tensors with safetensor.load() that would handle the metadata and deserialization. But now, in preparation of handling re-sharding, we need to read the metadata directly from the header of the safetensors file and store it directly in TensorStorageMetadata objects so that we can perform re-sharding. Re-sharding won't currently work, as we need extra metadata to be stored on each save, so that will be added in a subsequent PR.
In addition this PR adds an integration test in addition to the unit tests.
It also removes the HfFileSystem import because that's only needed if users are using HfFileSystem, but we want to support any backend.
Differential Revision: [D74891998](https://our.internmc.facebook.com/intern/diff/D74891998/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154518
Approved by: https://github.com/saumishr
Summary:
As we move towards supporting saving partial tensors natively with HFStorageWriter, there are some simple changes that need to be made to make this happen.
- The current approach for distributed writes is that every rank has full tensors, but we split up the writing of these full tensors across all available ranks. We're removing this logic that was in the HFSavePlanner and instead assuming that every rank has a shard and saving every rank's local state
- as a result we can probably remove the HFSavePlanner, but keeping it as a placeholder for now
- the current naming of files doesn't support shards as its in the format "model-00001-of-00004.safetensors", but if every rank is writing the same file names they will overwrite eachother, so this adds a shard-00001 prefix, so that the rank files don't overwrite eachother
- don't save the metadata file models.safetensors.index.json if sharding is enabled. This file expects a 1 to 1 ratio between tensor and filename, but this doesn't make sense in the sharded saving approach, so we can just get rid of this file
- make the "fqn_to_file_index" map optional. This is to describe which files to save which tensors in, but if users don't want to provide this, we can just save all the tensors to one file. If they run into issues, they can choose how to split up their tensors to be more friendly with 5GB HF remote storage file size soft limit.
Test Plan: test_hf_storage.py
Reviewed By: saumishr
Differential Revision: D75099862
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155566
Approved by: https://github.com/saumishr
resolve https://github.com/pytorch/pytorch/issues/154655
`fully_shard(root, reshard_after_forward=True)` didn't really reshard parameters after forward, because we assumed root model will be used in backward immeidately. The assumption becomes invalid in 2 cases
* we have 3 roots for CLIP, T5, FLUX. we should reshard parameters are CLIP and T5 immeidately after their forward
for recommendation model, we may have mutiple root for dense part
Change default beahvior to always respect `reshard_after_forward=True`
Differential Revision: [D75663200](https://our.internmc.facebook.com/intern/diff/D75663200)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154704
Approved by: https://github.com/mori360
This is a fix when an unused kwarg is in the PP stage forward, we try to call `torch.autograd.grad()` and update its gradients when it shouldn't have gradients. Leading to this error:
```
[rank3]:[rank3]: File "/data/users/howardhuang/pytorch/torch/distributed/pipelining/stage.py", line 613, in
[rank3]:[rank3]: return lambda: stage_backward_input(
[rank3]:[rank3]: File "/data/users/howardhuang/pytorch/torch/distributed/pipelining/_backward.py", line 199, in stage_backward_input
[rank3]:[rank3]: dinputs = torch.autograd.grad(
[rank3]:[rank3]: File "/data/users/howardhuang/pytorch/torch/autograd/init.py", line 503, in grad
[rank3]:[rank3]: result = _engine_run_backward(
[rank3]:[rank3]: File "/data/users/howardhuang/pytorch/torch/autograd/graph.py", line 824, in _engine_run_backward
[rank3]:[rank3]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
[rank3]:[rank3]: RuntimeError: One of the differentiated Tensors does not require grad
```
related issues: https://github.com/pytorch/torchtitan/issues/1188
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153498
Approved by: https://github.com/kwen2501
I hit this in tests when calling `init_process_group(init_method="tcp://localhost:0", ...)`. You can't use port 0 due to the bug in the conditional and will get error `ValueError: Error initializing torch.distributed using tcp:// rendezvous: port number missing`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154156
Approved by: https://github.com/d4l3k, https://github.com/Skylion007
Summary: The error caused by the world size not being divisible by `group_size` is a common issue encountered by end-users when utilizing applications built on top of `new_subgroups()`. However, these applications may employ different variable names, such as `num_trainers_per_group`, which can make the current error messages less effective despite being correct. To address this, we have improved the error messages to display the actual numbers involved, thereby enhancing their clarity and usefulness.
Test Plan: contbuild & OSS CI
Differential Revision: D75226925
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154124
Approved by: https://github.com/wz337
Summary: The bug, introduced in https://github.com/pytorch/pytorch/pull/152765, was caused by passing the `group` parameter to the `get_rank()` function, which caused the function to return the rank of the entire group instead of the rank of the current process. The fix involves removing the `group` parameter from the `get_rank()` function call.
Test Plan: contbuild & OSS CI
Differential Revision: D74964213
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153798
Approved by: https://github.com/Skylion007
DSD currently will pop tensors if these tensors are on Meta device. This forbid the use cases that users would like to let DCP to directly initialize the tensors when loading.
This PR also removes test/distributed/checkpoint/e2e/test_pipeline.py which is based on the above feature that is not realistic and is not used anywhere.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153185
Approved by: https://github.com/mori360
Summary:
Currently things are hardcoded to only work with nccl backend. Extend it
to allow NCCL + custom plugin backend.
The split-specific methods/attributes have not been added to the base
Backend and Options as some of them are specific to backend implementations.
Instead, explicit checks have been added to the split_group method for the
expected methods and attributes.
I am open to making them part of base Backend based if folks prefer.
Test Plan:
CI
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152175
Approved by: https://github.com/shuqiangzhang, https://github.com/kwen2501
## Summary
- The unit test `pytest test/distributed/test_symmetric_memory.py -k test_fused_scaled_matmul_reduce_scatter_scatter` was not running for some reason when #149247 was merged, giving false green CI signals. When it was ran manually recently, the test failed, highlighting a bug causing incorrect numerics when `scatter_dim=1`.
- This PR fixes the bug, which was related to how we swap dims 0<=>scatter_dim at the beginning of the custom op (for more efficient cross-device data movement I believe), then swap it back prior to reduction.
## Test plan
- I confirmed the unit test `pytest test/distributed/test_symmetric_memory.py -k test_fused_scaled_matmul_reduce_scatter_scatter` is now passing.
- I confirmed e2e training w/ torchtitan looks good ([logs](https://www.internalfb.com/phabricator/paste/view/P1812054188))
- I analyzed the tlparse to verify the fused_all_gather_matmul and fused_scaled_matmul_reduce_scatter both appear at least once in the post grad graphs ([tlparse](https://manifold.edge.x2p.facebook.net/v0/read/tree/logs/.tmpVbUsdG/dedicated_log_torch_trace_65oh3qj_.log/index.html?bucketName=tlparse_reports&apiKey=tlparse_reports-key&withPayload=1&timeoutMsec=10000))
## Next steps
1. I think for async TP `fused_scaled_matmul_reduce_scatter` we may only need `scatter_dim_after_maybe_reshape` and not `orig_scatter_dim` after all. I can confirm this and refactor if it is the case.
2. This op is specifically designed for async TP, and many of the arguments don't make sense for a user trying to use this as a standalone op. IMO we should have separate standalone custom op without all the extra function args and internal logic that doesn't apply to non-async TP cases.
3. In a follow up PR I want to add shape annotations to each line (e.g. `# (B, T, H)` etc) to make this easier to debug in the future.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153595
Approved by: https://github.com/fegin
Change logging.error to logging.exception to log additional information when relevant. A few places have slipped in logging.errors in try except since I last did a clean up here and the rule is stabilized so I am enabling it codebase wide. I have NOQA'd much of our custom exception stack trace handling for RPC calls and distributed and tried to a fix a few errors based on whether we immediately reraised it or if we didn't print any exception handling where it could be useful.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153473
Approved by: https://github.com/albanD, https://github.com/cyyever
Summary:
### Diff Context
- PR introduces Pipes for multiprocess comms with checkpointer process.
- Pipes allow easier comms contract management due to close() API and catch-all feature when background process is dead (e.g. seg faults).
Test Plan: CI
Differential Revision: D74668559
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153488
Approved by: https://github.com/saumishr
as titled, this PR improves the device selection logic when user did not
set the device before calling the DeviceMesh constructor, as a device
manager, DeviceMesh should try to set the device for users in a good
way.
The behavior of set_device before:
* If user call init_process_group to init a world process group, we assume user already called set_device and we don't set the device for the user
* If user does not init a world process group by themselves, we init a world process group for the user and follow a heuristic to set the device.
This is ok but sometimes the set_device heuristic wouldn't work well (i.e. if user use TORCH_CUDA_VISBILE_DEVICES
So this PR improves the device selection logic to:
* If the default cuda context is initialized by the time we init DeviceMesh, then we assume user must called some cuda operation before therefore must have selected the device by themselves
* If not the above, then we check if envvars have "LOCAL_RANK" and "WORLD_SIZE" from the launcher (i.e. torchrun), if so, we use "LOCAL_RANK" to set the device for the current process, which is a very standard practice. (This solves the TORCH_CUDA_VISBILE_DEVICES issue)
* If not above, then we throw warning to users about situation, and fallback to the old heuristic.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150897
Approved by: https://github.com/tianyu-l
ghstack dependencies: #150898
In #117066, shutdown of the rendezvous was added if a worker shuts down. This is incorrect, because the rendezvous is actually shutdown in [this file](fa6f9eb2be/torch/distributed/launcher/api.py (L290)) but should not be shutdown if a signal is received. See also [this pull request](https://github.com/pytorch/pytorch/pull/67749).
#124819 then tried to remediate the situation by fixing the faulty shutdown for the restart case. But this is only triggered if the agent restarts the training, but not if the shutdown of the rendezvous happened before.
Removing both these changes restores the original behavior. The rendezvous should only be shutdown if a run completes or fails, not for a single worker leaving.
Fixes#150916Fixes#147064
Pull Request resolved: https://github.com/pytorch/pytorch/pull/152525
Approved by: https://github.com/kiukchung
