Summary: If there is only one safetensors file, we don't need users to have a metadata file and we can just construct it from the keys of that file. This is a use-case for some HuggingFace models, so adding support for it
Test Plan:
ensure existing tests pass
tested e2e in a notebook
Differential Revision: D72472490
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150701
Approved by: https://github.com/joecummings
Summary: Introduce barrier util in the DistWrapper for rank local checkpointing. This barrier will be used at the end of the rank local checkpointing to ensure all ranks synchronize.
Test Plan: UTs
Differential Revision: D72541431
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150748
Approved by: https://github.com/MeetVadakkanchery
Summary: Adding new ops, support for empty shards, and fixed initializations for downstream checkpointing.
Test Plan: buck2 run 'fbcode//mode/dev-nosan' fbcode//torchrec/distributed/tests:test_shards_wrapper
Differential Revision: D72271275
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150469
Approved by: https://github.com/XilunWu
Summary: add a param to specify to the storage writer how to save tensors. Write now the only options are safetensors and torch.save.
Test Plan:
(lintrunner) [ankitageorge@devgpu003.cco3 /data/users/ankitageorge/fbsource/fbcode/caffe2 (1d57cb27b)]$ buck2 test 'fbcode//mode/opt' fbcode//caffe2/test/distributed/checkpoint:test_hf_storage
File changed: fbcode//caffe2/torch/distributed/checkpoint/filesystem.py
Buck UI: https://www.internalfb.com/buck2/e80cc963-e34a-4876-b6f4-7ce2794e48dd
Test UI: https://www.internalfb.com/intern/testinfra/testrun/3659174965882569
Network: Up: 32KiB Down: 1.9KiB (reSessionID-ef9fa764-a40a-451b-ab58-08eabe7a9422)
Executing actions. Remaining 0/4 3.4s exec time total
Command: test. Finished 2 local
Time elapsed: 19.6s
Tests finished: Pass 4. Fail 0. Fatal 0. Skip 0. Build failure 0
Reviewed By: saumishr
Differential Revision: D70271943
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150025
Approved by: https://github.com/saumishr
Needed this class for because `parallelize_module` takes a dict, which doesn't allow `PrepareModuleInput` and `PrepareModuleOutput` to be applied at the same time.
The `PrepareModuleInputOutput` in this PR initializes two variables `prepare_module_input` and `prepare_module_output` and uses them to process module / inputs / outputs.
I had another implementation which put all code in `PrepareModuleInputOutput` and let `PrepareModuleInput` and `PrepareModuleOutput` inherit the monolithic `PrepareModuleInputOutput`. But it is
1. less cleaner
2. conceptually abusing inheritance because `PrepareModuleInput` shouldn't be able to access class methods of `PrepareModuleOutput` and vice versa
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150372
Approved by: https://github.com/wanchaol
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
Summary:
The _load_state_dict_from_keys method specifies that `Loads any key specified in this set. If no keys are specified, the entire checkpoint is loaded.`
But this isn't happening right now, because an empty keys arg is passed in as a set() to `_load_state_dict` and keys is expected to be None for it to actually be included in the state_dict https://fburl.com/code/l8yzojyx. So with the set() argument, the state_dict is always going to be empty
Test Plan: ensure existing tests pass
Differential Revision: D71930712
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150058
Approved by: https://github.com/saumishr
Part of https://github.com/pytorch/torchtitan/issues/866
## Context
- Async TP needs to support the "reshape -> scaled_mm -> reshape" pattern because scaled mm only supports 2D input tensors and 2D scales.
- (a,b,c) => (a*b,c)
- (a\*b,c) @ (c,d) = (a\*b,d)
- (a\*b,d) => (a,b,d)
- Currently the implementation does not support scaled mm with rowwise scales **for all cases** of the reshape -> scaled_mm -> reshape pattern. The minimal example of this pattern is confirmed to work via this [unit test](00a2c68f67/test/distributed/tensor/parallel/test_micro_pipeline_tp.py (L406)), but more involved e2e examples in torchtitan fail silently (more context in final bullet point).
- Previously, the "A tensor" **node** referenced in the async TP graph manipulation code is the 3D+ node before the reshape, but the "A_scale" node is the 2d node from after the reshape, so they are incompatible.
- I previously implemented a simpler solution to this problem in https://github.com/pytorch/pytorch/pull/148001, with a [unit test](https://github.com/pytorch/pytorch/pull/148001/files#diff-115f1d0852382c9b58f22640d80999d879b33618e5f6c633fc9e4d0ca9781cecR406) confirming the fused node is indeed in the graph for the minimal example of the reshape->mm->reshape pattern. I also confirmed via manual e2e testing w/ torchtitan that the crash I was fixing no longer occurred. However, it turns out due to this [bug in torchtitan](https://github.com/pytorch/torchtitan/issues/866) it was causing async TP to fail silently and fall back to vanilla TP, hiding the fact that this original solution fixed the crash but the fusion would not occur for rowwise scales. Thus, more robust solution is needed to support all cases.
## Solution TL;DR
- Use the 2D 'A' tensor and corresponding 2D scales as input to the fused_matmul_reduce_scatter implementation, instead of the 3D+ tensor/scales.
- Track the "pre mm reshape" and "post mm reshape" separately, to be referenced in the `fused_scaled_matmul_reduce_scatter` implementation, to update the scatter dim through the pre-mm reshape, and apply the post-mm reshape before applying the reduce scatter and returning the output tensor.
- Separate the `fused_matmul_reduce_scatter` and the `fused_scaled_matmul_reduce_scatter` code paths, to simplify them both.
- By fixing the bug in torchtitan (PR https://github.com/pytorch/torchtitan/pull/965) and implementing support for rowwise scales in pytorch in this PR, together these changes will solve the problem of how to support rowwise scales with all types of AC.
## Additional details for reviewers
To use the 2D A tensor while also supporting the "reshape -> mm -> reshape" pattern, the following other changes were needed:
- Track the pre-mm reshape, as it will affect the scatter dim used in the fused_matmul_reduce_scatter impementation.
- Track the post-mm reshape, as it will affect the output shape used in the fused_matmul_reduce_scatter impementation
- Based on the pre-mm reshape and the original scatter dim, calculate the new scatter dim for the 2D tensor. This is needed because during the pipelined producer mm implementation, the scatter dim is moved to dim 0 (so it can be sharded along the first dim and then get chunks to do mm ops on by indexing into the first dim), then moved back to it's original place before the reduce-scatter.
- Use the tracked post-mm reshape to reshape the stacked partial 2D outputs of the mm ops into 3D outputs needed for 1) the reduce-scatter w/ the original scatter dim, and 2) the expected output shape to prevent shape errors with subsequent ops.
## Test plan
- All existing unit tests passing.
- Expand unit tests for rowwise scales to test more scatter dims
- Added unit tests enforcing that async TP fails fast / throws an error if it fails to perform any fusions. Previously it just "failed silently" (fell back to vanilla TP without the user knowing) which has led to confusion, so this will improve the UX.
- Compared loss curves of bf16 vs float8 w/ rowwise scales to confirm integrity of numerics
- Confirmed via manual testing with torchtitan and inspecting the compile graph that the fusion is working as intended for:
- bfloat16
- float8 with tensorwise scales
- float8 with rowwise scales
## Loss curves
Loss curves are virtually identical for bf16 + vanilla TP versus float8 with rowwise scales + async TP:
<img width="1017" alt="loss_async_tp" src="https://github.com/user-attachments/assets/4995db78-7012-490f-a370-f4fecc289a22" />
## Performance
#### Per op SAC
Performance benchmarks for torchtitan Llama3 8b training runs on 4 H100s with per op SAC, using FSDP degree=2, TP degree=2:
- bf16 (vanilla TP): TPS 5161.5, peak memory 50.53 GB
- bf16 (async TP): TPS 5229.5, peak memory 50.68 GB
- float8 tensorwise (vanilla TP): TPS: 5959.5, peak memory: 50.47 GB
- float8 tensorwise (async TP): TPS 5964.5, peak memory 50.47 GB
- float8 rowwise (vanilla TP): TPS: 4962.0, peak memory: 50.55 GB
- float8 rowwise (async TP): TPS 4966.5, peak memory 50.65 GB
#### Full AC
Llama3 70b training runs on 128 H100s with full AC, using FSDP=16, TP=8
- bf16 (vanilla TP): 598 TPS, peak memory 71.51 GB
- bf16 (async TP): TPS 673, peak memory 71.08 (+12.54% TPS vs vanilla TP)
- float8 tensorwise (vanilla TP): 820 TPS, peak memory 55.26 GB
- float8 tensorwise (async TP): 950 TPS, peak memory 55.91 GB (+15.85% TPS vs vanilla TP)
- float8 rowwise (vanilla TP): TPS: 540 TPS, peak memory 71.46 GB
- float8 rowwise (async TP): 560 TPS, peak memory 70.65 GB (+3.7% TPS vs vanilla TP but still unexpectedly lower than bf16)
As you can see, float8 rowwise is working but performance needs to be improved further.
## Other changes
- Added logging so the user will know why fusion failed if it does.
- Remove logic which inserted a reshape node targeting "A scale" to get it to be in 3D like the "A tensor" since it's no longer needed.
## Long term plan
- Add a `scaled_matmul` op in pytorch, which will natively support a 3D+ "A tensor" and allow us to simplify the async TP implementation by avoiding the reshape -> scaled_mm -> reshape pattern and the special handling for it.
## Visualizing fused nodes in graphs for torchtitan training runs
Below are examples of the visualized graph generated by torch compile for torchtitan llama3 8b training runs with per op SAC. These graphs provide additional evidence (beyond the new unit tests added) that the implementation is working correctly.
### bf16
<img width="900" alt="bf16-fusion" src="https://github.com/user-attachments/assets/a3bed917-28eb-4a56-8d6e-2d2bf498385c" />
### float8 with tensorwise scales
<img width="900" alt="tensorwise-node" src="https://github.com/user-attachments/assets/b212ec4a-1899-44de-a4de-18c74e1de68a" />
### float8 with rowwise scales
<img width="900" alt="rowwise" src="https://github.com/user-attachments/assets/ed3354a3-894b-4ec9-86d0-f80364bf3d83" />
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149247
Approved by: https://github.com/kwen2501
Summary: This diff adds the ability for HF reader/writer to read/write in a distributed way. We do this by sending all the tensors meant for the same file to the same rank.
Test Plan:
ensure existing tests pass
I also ran a full end to end test on my devserver to read/write from my HF repo
Differential Revision: D70096439
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148189
Approved by: https://github.com/joecummings, https://github.com/saumishr
Summary:
Cache save plan metadata to reduce the collective overhead.
Global plan dedupe and metadata creation are the main overheads on Rank 0. This change saves all this cost for the subsequent saves if the plans do not change. A quick experiment with the 256 rank job, Global step overhead drops by ~99%, from 90s+ to mere 1.5s. 1.5s was mostly spent on creating the checkpoint module directories and near empty collective.
Differential Revision: D71631441
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149785
Approved by: https://github.com/MeetVadakkanchery
This adds a `reduce_scatter` implementation for ProcessGroupGloo. This is a pretty naive implementation as it does 1 allreduce per rank but may be useful for testing in FSDP etc. There was an existing implementation of reduce_scatter_tensor/reduce_scatter_tensor_coalesed that has a very similar implementation but requires a fixed tensor size per rank.
If users find these functions to be too slow we can address them as issues arise.
Gloo now supports all major distributed operations. Quite a few of these were added by @rohan-varma and @yifuwang but they didn't update the support chart. We also have `CUDAWork` variants of most operations so those were also added to the chart.
Test plan:
```
pytest -v test/distributed/test_c10d_gloo.py -k reduce_scatter
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149869
Approved by: https://github.com/fduwjj
Before, we would take the first argument with the largest number of shards, regardless if it had fewer dims than another arg with the same number of shards but more dimensions. This would lead to potentially fewer sharding options
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149721
Approved by: https://github.com/tianyu-l
Today, if you run DTensor (or any tensor subclass) under __torch_dispatch__, you will start seeing `CompositeImplicitAutograd` ops show up in the torch_dispatch.
"handling" these ops is trivial: you can just tell them to decompose into their constituent ops. Normally this decomposing happens in autograd, above DTensor, but inference_mode turns autograd off, forcing the subclass to handle the op directly.
It looks like previously we manually added a few CompositeImplicitAutograd entries to DTensor (e.g. linear), but this PR tries to support these ops a bit more generically.
The main difference is that DTensor now needs to check if a given op is `CompositeImplicitAutograd` before attempting to run sharding prop. I ran a quick microbenchmark for the below code with `timeit`, which gave me overhead on the order of ~1us, which is hopefully not too bad for eager mode:
```
def fast_function():
return torch._C._dispatch_has_kernel_for_dispatch_key(op_call.name(), torch._C.DispatchKey.CompositeImplicitAutograd)
import timeit
time_taken = timeit.timeit(fast_function, number=1000)
# printed 0.12..., aka 1.2us
print(f'func={str(op_call)}, time={str(time_taken)}')
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149514
Approved by: https://github.com/kwen2501, https://github.com/albanD, https://github.com/wanchaol
Summary:
1\ The current write item structure does not contain the amount of data that needs to be written.
2\ the planner.item already has a size primitive 'tensor_storage_size'. https://fburl.com/code/7a0gsmw7 But only for tensors.
3\ Right now, the only way the writer layer get hold of this property (fro non tensor data)
first do a lookup in to the actual tensor/bytes
then calculate the nbytes.
This change introduce a way to capture non-tensor data size within a write-plan item.
Test Plan: Existing UT.
Differential Revision: D71599725
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149699
Approved by: https://github.com/MeetVadakkanchery
Summary:
1\ The current write item structure does not contain the amount of data that needs to be written.
2\ the planner.item already has a size primitive 'tensor_storage_size'. https://fburl.com/code/7a0gsmw7 But only for tensors.
3\ Right now, the only way the writer layer get hold of this property (fro non tensor data)
- first do a lookup in to the actual tensor/bytes
- then calculate the nbytes.
This change introduce a way to capture non-tensor data size within a write-plan item.
Reviewed By: daulet-askarov
Differential Revision: D70497442
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149434
Approved by: https://github.com/MeetVadakkanchery
PR does following
* Turns `inference_mode` to False and `no_grad` for `convert_frame`, if the inference_mode is on globally.
* Turns off inference_mode for fake tensor prop. This ensures that converting from real inference tensor to a fake tensor removes the inference-ness.
* Graph breaks on is_inference and is_inference_mode_enabled.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149321
Approved by: https://github.com/jansel, https://github.com/zou3519
Summary:
Avoid in-place update and deepcopy during dudpe. Deepcopy becomes prohibitively expensive with models having a huge number of FQNs. This was manifestd in the Ads 2K experiment as well. Here are the results from the TextRay model in Mitra:
#### Control job with deepcopy regression:
First save ~24.8s
Global step latency is ~7-8s
Test job with the new fix to avoid deepcopy:
First save is ~21s
global step latency ~2s
Test Plan:
```
buck test 'fbcode//mode/dev-nosan' fbcode//caffe2/test/distributed/checkpoint:test_planner
```
https://www.internalfb.com/intern/testinfra/testrun/3940649945104822
Differential Revision: D71245218
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149320
Approved by: https://github.com/MeetVadakkanchery
This is a trivial rule that for most cases isn't needed, but if we want to consider that the input data is actually `Shard(0)` (instead of `Replicated()` as it is currently assumed), then we need this rule.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/149253
Approved by: https://github.com/XilunWu
Fixes #ISSUE_NUMBER
When attempting to reconfigure the environment without properly handling the PyTorch-related settings, you may encounter the following message.
```
│ /root/.cache/pypoetry/virtualenvs/app-rag-sample-9TtSrW0h-py3.10/lib/python3.10/site-packages/torch/distributed/distribut │
│ ed_c10d.py:1215 in get_backend │
│ │
│ 1212 │ if _rank_not_in_group(pg): │
│ 1213 │ │ raise ValueError("Invalid process group specified") │
│ 1214 │ pg_store = _world.pg_map[pg] if pg in _world.pg_map else None │
│ ❱ 1215 │ return Backend(not_none(pg_store)[0]) │
│ 1216 │
│ 1217 │
│ 1218 def _get_process_group_uid(pg: ProcessGroup) -> int: │
│ │
│ /root/.cache/pypoetry/virtualenvs/app-rag-sample-9TtSrW0h-py3.10/lib/python3.10/site-packages/torch/utils/_typing_utils.p │
│ y:13 in not_none │
│ │
│ 10 │
│ 11 def not_none(obj: Optional[T]) -> T: │
│ 12 │ if obj is None: │
│ ❱ 13 │ │ raise TypeError("Invariant encountered: value was None when it should not be") │
│ 14 │ return obj │
│ 15 │
╰───────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯
TypeError: Invariant encountered: value was None when it should not be
Exception ignored in: <function Vllm.__del__ at 0x7f35f96b6dd0>
```
Since this message can cause confusion for multiple developers, the purpose of this PR is to suggest additional details to help clarify the situation.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/141796
Approved by: https://github.com/kwen2501
Fixes#138842
`device` is always the device of the `local_state_dict`, which may or may not be CPU, which is not supported by NCCL backend.
Instead, create broadcasted tensors on one of `pg._device_types` and then move the tensors back if `local_state_dict`'s `device` was not supported by the `ProcessGroup`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148865
Approved by: https://github.com/mori360
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
# Fix typo errors across PyTorch codebase
This PR fixes various spelling errors throughout the PyTorch codebase to improve documentation quality and code readability.
## Changes Made
### Documentation Fixes
- Changed "seperate" to "separate" in multiple files:
- `setup.py`: Build system documentation
- `torch/_library/triton.py`: AOT compilation comments
- `torch/csrc/dynamo/compiled_autograd.h`: Node compilation documentation
- `torch/export/_unlift.py`: Pass population comments
- `torch/export/exported_program.py`: Decomposition table notes
### Code Comments and Error Messages
- Changed "occured" to "occurred" in:
- `test/mobile/test_lite_script_module.py`: Exception handling comments
- `torch/export/_draft_export.py`: Error message text
- `aten/src/ATen/native/cuda/linalg/BatchLinearAlgebra.cpp`: MAGMA bug comment
- `torch/csrc/utils/python_numbers.h`: Overflow handling comment
- `torch/csrc/jit/OVERVIEW.md`: Graph compilation documentation
- `torch/_dynamo/symbolic_convert.py`: Error explanation
### API Documentation
- Changed "fullfill" to "fulfill" in `torch/distributed/checkpoint/state_dict_loader.py`
- Changed "accross" to "across" in:
- `torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp`
- `torch/distributed/distributed_c10d.py`
## Motivation
These changes improve code readability and maintain consistent spelling throughout the codebase. No functional changes were made; this is purely a documentation and comment improvement PR.
## Test Plan
No testing required as these changes only affect comments and documentation.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148262
Approved by: https://github.com/janeyx99
Co-authored-by: Jane (Yuan) Xu <31798555+janeyx99@users.noreply.github.com>
