Summary:
## No Functional Change
- Refactor Subprocess Handler into a separate folder for easier subclassing
- SubprocessHandler
- added `local_rank_id` in `SubprocessHandler` to make it available as a field in the class
- pass in `local_rank_id` from subprocess start
Test Plan: No functional changes.
Differential Revision: D54038627
#suppress-api-compatibility-check
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120373
Approved by: https://github.com/kurman
As titled, this PR introduces a dedicated `ParallelStyle` to shard the
nn.LayerNorm/nn.Dropout/RMSNorm layers. We were mainly using a manual
distribute_module calls before when sharding the RMSNorm layer, but I
think we should have a dedicate TP API to easily shard those layers,
instead of user manually using DTensors.
I call this SequenceParallel, which might bring some confusion that we
technically "deprecated" a SequenceParallel style months ago. But this
time the SeuqenceParallel style is significantly different with the
previous ones (which used to shard two consecutive Linear layers). I
believe making it the right name is the first priority, instead of
worrying about the issue of reusing the old name
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121295
Approved by: https://github.com/awgu, https://github.com/tianyu-l
ghstack dependencies: #121294
Summary: WrapperModule seems a good idea but may introduce some surprising behavior to users, for example, it never registers enclosed modules as submodules and therefore it's unclear that's the state dict for the exported program should look like, because some people may argue to include every state in state dict but others want to keep them as constants.
Test Plan: CI
Reviewed By: tugsbayasgalan
Differential Revision: D54326331
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121042
Approved by: https://github.com/angelayi
Loss parallel is the last piece of sequence parallelism to enable. It enables efficient distributed cross entropy computation when the input is sharded on the class dimension (in a classification problem with many classes). The implementation is via a context manager `loss_parallel`, after enabling which users can directly use `torch.nn.functional.cross_entropy` or `torch.nn.CrossEntropyLoss` without modifying other parts of their code.
Here are the underlying rationales why we are going through these op replacements:
1. `nn.functional.cross_entropy` is the common method that OSS user is using for things like transformer training, to avoid changing user code, we want user to still use this function for loss calculation if they are already using it.
2. `nn.functional.cross_entropy` boils down into `aten.log_softmax` and `aten.nll_loss_foward/backward`, and DTensor now supports those ops already (#117723#119255#118917#119256). They are doing computation with input *replicated* on the class dimension.
3. However when the input of this loss calculation is **sharded on the class dimension**, to run sharded computation efficiently, we need to run both `aten.log_softmax` and `aten.nll_loss_foward` with multiple all-reduce collectives **in the middle of** those aten ops. This is not possible if we are just overriding these two ops, so we need to have some way to **decompose** these two ops into smaller ops to have collectives run in the middle of these two ops.
4. We explored the existing decompositions (#118950). It seems working, except that `log_softmax_backward` and `nll_loss_backward` combined together in aten are implemented in a inefficient way, which would trigger an additional expensive collective. Recently some user also reported similar issues https://github.com/pytorch/pytorch/issues/119261.
5. Therefore, currently we are doing our own decomposition inside a context manager for sequence parallelism specifically. Once we have a better decomposition in core, we can possibly take that instead of reinventing the wheels here.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119877
Approved by: https://github.com/wanchaol
Summary:
Pulling out logging parameters into a logging specs that can be overridden (follow-up changes on possible mechanism)
Why?
Right now the logging approach is quite rigid:
- Requires for log directory to exist and not be empty
- Will create tempdir otherwise,
- Creates subdir for a run
- creates subdir for each attempt
- creates files named as stdout.log, stderr.log, error.json
In some instances some of the users would like to customize the behavior including file names based on context. And we do have right now a mechanism to template multiplexed teed output prefix.
With current changes, users can create custom log spec that can use env variables to change the behavior.
Notes:
Made `LaunchConf.logs_specs` as an optional field that will be bound to `DefaultLogsSpecs` instance. There are large number of clients (code) that use the API directly without using torchrun API. For those cases, we have to explicitly pass LogSpecs implementation if we would like to override the implementation. For the regular torchrun users, we can use pluggable approach proposed in the follow up change.
Test Plan: CI + unit tests
Differential Revision: D54176265
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120691
Approved by: https://github.com/ezyang
Summary: In non-strict mode of torch.export() we didn't set those `is_compiling()` to `True` which is needed by some models.
Test Plan: Unit tests and manual testing.
Differential Revision: D53624452
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119602
Approved by: https://github.com/suo
This PR adds a couple of missing words in the Checkpointing documentation, it doesn't have a specific issue number related to it.
Changes are:
- "backward." -> "backward propagation."
- "to be advanced than" -> "to be more advanced than"
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120196
Approved by: https://github.com/soulitzer
Summary: This commit adds the `model_is_exported` util function
for users to be able to easily tell what APIs to call to move
their models between train and eval modes. This has the
additional advantage of hiding the implementation of how we
detect a model is exported, in case the metadata format changes
in the future.
Test Plan:
python test/test_quantization.py TestQuantizePT2E.test_model_is_exported
Differential Revision: [D53812972](https://our.internmc.facebook.com/intern/diff/D53812972)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119726
Approved by: https://github.com/tugsbayasgalan, https://github.com/albanD
# Motivation
According to [[RFC] Intel GPU Runtime Upstreaming](https://github.com/pytorch/pytorch/issues/114842) and [[RFC] Intel GPU Runtime Upstreaming for Allocator](https://github.com/pytorch/pytorch/issues/116322), we will upstream the key functionality of device `Allocator` dedicated for XPU to PyTorch. And following our design prepare to generalize `Allocator` in parallel.
# Design
In the current design, XPU uses an `XPUAllocator` class, inherited from `c10::Allocator`. `XPUAllocator` is a manager to handle `DeviceCachingAllocator`, which is a per-device implementation of the caching mechanism to manage the already cached or newly allocated memory. The caching mechanism is similar to other backends, like CUDA. We can visualize the design as below.
<p align="center">
<img width="162" alt="image" src="https://github.com/pytorch/pytorch/assets/106960996/6b17b8cf-e7d1-48b4-b684-f830c409d218">
</p>
# Additional Context
We're going to implement our design gradually. This PR covers the device `Allocator` dedicated to XPU. The second PR covers the host `Allocator`.
Besides these PRs, we plan to generalize the device `Allocator` device-agnostic through another PR.
In this PR, our device `Allocator` has the same memory management mechanism as CUDA, but lacks features such as expendable segments and statistics. We will add these features back in the subsequent PR which intend to generalize `Allocator`.
The differences with CUDA:
only key functionality, and lack of AsyncAllocator, gpu_trace, history_record, graph functionality, memory snapshot, memory statistics, expandable segment...
Pull Request resolved: https://github.com/pytorch/pytorch/pull/118091
Approved by: https://github.com/EikanWang, https://github.com/gujinghui, https://github.com/jgong5, https://github.com/albanD
ghstack dependencies: #117611, #117619, #117734
# Motivation
As mentioned in [[RFC] Intel GPU Runtime Upstreaming](https://github.com/pytorch/pytorch/issues/114842), the next runtime component we would like to upstream is `Event` which handles the status of an operation that is being executed. Typically, in some circumstances, we can fine-grain control of the operation execution via `Event`.
# Design
`XPUEvent` is a movable but not a copyable wrapper around sycl event. It should be created lazily on an XPU device when recording an `XPUStream`. Meanwhile, `XPUEvent` can wait for another `XPUEvent` or all the submitted kernels on an `XPUStream` to complete. Align to the other backend, the C++ files related to `Event` will be placed in `aten/src/ATen/xpu` folder. For frontend code, `XPUEvent` runtime API will be bound to Python `torch.xpu.Event`. The corresponding C++ code will be placed in `torch/csrc/xpu/Event.cpp` and Python code will be placed in `torch/xpu/streams.py` respectively.
# Additional Context
It is worth mentioning that the `elapsed_time` method is temporarily not supported by `XPUEvent`. We will be adding support for it soon. Meanwhile `XPUEvent` doesn't support IPC from different processes. For the other parts, we have almost a 1:1 mapping with CUDA.
lack of the below APIs:
- `torch.cuda.Event.ipc_handle`
- `CUDAEvent`'s constructor with `IpcEventHandle`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117734
Approved by: https://github.com/EikanWang, https://github.com/gujinghui, https://github.com/jgong5, https://github.com/malfet
ghstack dependencies: #117611, #117619
# Motivation
According to [[1/2] Intel GPU Runtime Upstreaming for Stream](https://github.com/pytorch/pytorch/pull/117611), as mentioned in [[RFC] Intel GPU Runtime Upstreaming](https://github.com/pytorch/pytorch/issues/114842), the second PR covers the changes under `python frontend`.
# Design
Currently, it primarily offers stream-related APIs, including
- `torch.xpu.StreamContext`
- `torch.xpu.current_stream`
- `torch.xpu.set_stream`
- `torch.xpu.synchronize`
- `torch._C._xpu_getCurrentRawStream`
# Additional Context
We will implement functions like `torch.xpu.Stream.wait_event`, `torch.xpu.Stream.wait_stream`, and `torch.xpu.Stream.record_event` in the next PR related with `Event`.
The differences with CUDA:
no default and external stream in XPU and lack of below APIs:
- `torch.cuda.ExternalStream`
- `torch.cuda.default_stream`
- `toch.cuda.is_current_stream_capturing`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117619
Approved by: https://github.com/EikanWang, https://github.com/jgong5, https://github.com/gujinghui, https://github.com/albanD
ghstack dependencies: #117611
Added a `torch.Tensor` method that defines how to transform `other`, a value in the state dictionary, to be loaded into `self`, a param/buffer in an `nn.Module` before swapping via `torch.utils.swap_tensors`
* `param.module_load(sd[key])`
This method can be overridden using `__torch_function__`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117913
Approved by: https://github.com/albanD
# Motivation
According to [[1/4] Intel GPU Runtime Upstreaming for Device](https://github.com/pytorch/pytorch/pull/116019), as mentioned in [[RFC] Intel GPU Runtime Upstreaming](https://github.com/pytorch/pytorch/issues/114842), this last PR covers the changes under lazy initialization.
# Design
This PR primarily offers the support of multi-processing via lazy initialization. We lazily initialize our runtime avoiding initializing XPU until the first time it is accessed. In our design, we extend `cuda_lazy_init` to `device_lazy_init` which is a device-agnostic API that can support any backend. And change `maybe_initialize_cuda` to `maybe_initialize_device` to support lazy initialization for both CUDA and XPU while maintaining scalability.
# Additional Context
We adopt a similar design to CUDA. So we share some code with CUDA.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/116869
Approved by: https://github.com/EikanWang, https://github.com/jgong5, https://github.com/gujinghui, https://github.com/malfet
ghstack dependencies: #119248
Added `torch.__future__.{get/set}_swap_module_params_on_conversion` that defaults to `False` for now, but we probably want to modify to override this and default to `True` in `nn.Module._apply` if input is a tensor subclass.
From offline discussion, for now we are **not** allowing `swap_tensor` after the first module forward has been run*** if the autograd graph is still alive. The reason being that `torch.utils.swap_tensors(t1, t2)` requires the `use_count` of both `TensorImpl`s associated with `t1` and `t2` to be 1. The first forward pass will install `AccumulateGrad` nodes on each param, which [bump the refcount of the associated TensorImpl](6cf1fc66e3/torch/csrc/autograd/variable.cpp (L307)). **Future work might be to swap the refs that the `AccumulateGrad` nodes hold if it is necessary.**
***From this, it might seem like we don't need to handle gradients. However, I still handle the grads for the edge case that the grads are set via `p.grad = grad` OR the autograd graph is no longer alive because the output has been garbage collected.
If any `swap_tensors` fails on any of the parameters in the `nn.Module` we raise an error.
**`RNNBase` overrides `nn.Module._apply()` and installs weakrefs on some parameters. As a result, all modules that inherit from `RNNBase` (`RNN`, `GRU` and `LSTM`) cannot use the`swap_tensors` path as of now**
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117167
Approved by: https://github.com/albanD
ghstack dependencies: #118028
Fixes https://github.com/pytorch/pytorch/issues/117361
The implementation here slightly diverges from what was proposed in the issue, so I will recap what this PR is doing here. Today, when doing computations involving size-like unbacked SymInts, we assume for all operations that the compile time range of the integer is `[2, inf]`, even though at runtime we also accept zero and one.
This PR removes the carte blanche assumption, and instead does the analysis in a much more limited and controlled fashion: only for guards which we have designated as "size oblivious" are we willing to do the analysis under the assumption that the range of all size-like unbacked SymInts is `[2, inf]`; otherwise, we will faithfully only do analysis with `[0, inf]` (or whatever the user provided) bounds.
The infra pieces of this PR are:
* Remove runtime_var_to_range from torch/fx/experimental/symbolic_shapes.py; modify `_constrain_range_for_size` to refine the range without clamping min to 2, and instead add the symbol to a `size_like` set in the ShapeEnv
* When evaluating an expression, if the expression is requested to be evaluated in a `size_oblivious` way, we attempt to statically compute the value of the expression with the assumption that all symbols in `size_like` are updated to assume that they are `>= 2`.
* Add Python and C++ APIs for guarding on a SymBool in a size-oblivious way. In C++, I also need to add some helpers for performing symbolic comparisons, since the stock comparisons immediately specialize in the "normal" way.
The rest of the changes of the PR are marking various spots in PyTorch framework code as size oblivious, based on what our current test suite exercises.
As you review the places where we have marked things as size oblivious, it may become clear why I ended up not opting for the "designate a branch as the default branch when it's not statically obvious which way to go": for some of the conditions, this answer is rather non-obvious. I think potentially there is another refinement on top of this PR, which is something like "I don't care if you can't figure it out with ValueRange analysis, go down this path anyway if there are unbacked sizes involved." But even if we add this API, I think we are obligated to attempt the ValueRange analysis first, since it can lead to better outcomes sometimes (e.g., we are able to figure out that something is contiguous no matter what the unbacked size is.)
When is it permissible to mark something as size oblivious? Heuristically, it is OK anywhere in framework code if it gets you past a guard on unbacked SymInt problem. It is somewhat difficult to provide a true semantic answer, however. In particular, these annotations don't have any observational equivalence guarantee; for example, if I have `torch.empty(u0, 1).squeeze()`, we will always produce a `[u0]` size tensor, even though if `u0 == 1` PyTorch will actually produce a `[]` size tensor. The argument that I gave to Lezcano is that we are in fact defining an alternate semantics for a "special" size = 0, 1, for which we have these alternate eager mode semantics. In particular, suppose that we have a constant `special1` which semantically denotes 1, but triggers alternate handling rules. We would define `torch.empty(special1, 1).squeeze()` to always produce a `[special1]` size tensor, making its semantics coincide with unbacked SymInt semantics. In this model, the decision to designate guards as size oblivious is simply a user API question: you put them where ever you need some handling for special1! As we conservatively error out whenever it is not obvious what `special1` semantics should be, it is always valid to expand these semantics to cover more cases (although you can always choose the wrong semantics!)
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/118579
Approved by: https://github.com/eellison, https://github.com/lezcano
# Motivation
According to [[1/4] Intel GPU Runtime Upstreaming for Device](https://github.com/pytorch/pytorch/pull/116019), As mentioned in [[RFC] Intel GPU Runtime Upstreaming](https://github.com/pytorch/pytorch/issues/114842), this third PR covers the changes under `libtorch_python`.
# Design
This PR primarily offers device-related APIs in python frontend, including
- `torch.xpu.is_available`
- `torch.xpu.device_count`
- `torch.xpu.current_device`
- `torch.xpu.set_device`
- `torch.xpu.device`
- `torch.xpu.device_of`
- `torch.xpu.get_device_name`
- `torch.xpu.get_device_capability`
- `torch.xpu.get_device_properties`
- ====================
- `torch.xpu._DeviceGuard`
- `torch.xpu._is_compiled`
- `torch.xpu._get_device`
# Additional Context
We will implement the support of lazy initialization in the next PR.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/116850
Approved by: https://github.com/EikanWang, https://github.com/jgong5, https://github.com/gujinghui, https://github.com/malfet
Enables the deduplication of saved entries by load balancing duplicates across ranks.
Tested with existing and modified tests. Additionally tested with the following code snippet, which saves a 20GB DDP model in **~3 seconds on 8 ranks**. Before this PR, the same operation has been measured at ~19 seconds.
```
def run(local_rank, world_size, param_size, num_params, work_dir):
os.environ["RANK"] = str(local_rank)
os.environ["MASTER_ADDR"] = "localhost"
os.environ["MASTER_PORT"] = "12355"
device = torch.device(f"cuda:{local_rank}")
torch.cuda.set_device(device)
dist.init_process_group(backend="nccl", rank=local_rank, world_size=world_size)
model = Model(param_size=param_size, num_params=num_params)
model = DistributedDataParallel(model, gradient_as_bucket_view=True)
_patch_model_state_dict(model)
sz = sum(t.nelement() * t.element_size() for t in model.parameters())
rank_0_print(f"Model size: {sz / 1_000_000_000.0} GB")
rank_0_print("Saving the model with DCP...")
checkpointer = _FileSystemCheckpointer(
f"{args.work_dir}/dcp",
sync_files=False,
single_file_per_rank=False,
thread_count=1
)
begin_ts = time.monotonic()
checkpointer.save(state_dict={"model": model})
end_ts = time.monotonic()
rank_0_print(f"Took {end_ts - begin_ts} seconds with DCP")
```
Differential Revision: [D52435926](https://our.internmc.facebook.com/intern/diff/D52435926/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/116469
Approved by: https://github.com/fegin, https://github.com/wz337
# Summary
Simplification of Backend Selection
This PR deprecates the `torch.backends/cuda/sdp_kernel` context manager and replaces it with a new context manager `torch.nn.attention.sdpa_kernel`. This context manager also changes the api for this context manager.
For `sdp_kernel` one would specify the backend choice by taking the negation of what kernel they would like to run. The purpose of this backend manager was to only to be a debugging tool, "turn off the math backend" and see if you can run one of the fused implementations.
Problems:
- This pattern makes sense if majority of users don't care to know anything about the backends that can be run. However, if users are seeking to use this context manager then they are explicitly trying to run a specific backend.
- This is not scalable. We are working on adding the cudnn backend and this API makes it so so that more implementations will need to be turned off if user wants to explicitly run a given backend.
- Discoverability of the current context manager. It is somewhat un-intutive that this backend manager is in backends/cuda/init when this now also controls the CPU fused kernel behavior. I think centralizing to attention namespace will be helpful.
Other concerns:
- Typically backends (kernels) for operators are entirely hidden from users and implementation details of the framework. We have exposed this to users already, albeit not by default and with beta warnings. Does making backends choices even more explicit lead to problems when we potentially want to remove existing backends, (perhaps inputs shapes will get covered by newer backends).
A nice side effect is now that we aren't using the `BACKEND_MAP` in test_transformers many, many dynamo failures are passing for CPU tests.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114689
Approved by: https://github.com/cpuhrsch
Summary: Adding an experimental API to FX graph module to place "hooks" every time when we are changing or replacing nodes in a graph, so that we can properly update the new name in graph signature and potentially other places.
Test Plan:
buck test mode/opt -c fbcode.enable_gpu_sections=true caffe2/test/distributed/_tensor/experimental:tp_transform
buck test mode/opt caffe2/test:test_export -- -r test_replace_hook
Differential Revision: D52896531
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117825
Approved by: https://github.com/avikchaudhuri
