For a masked `tl.load` operation, the Triton language specifies that values masked out (i.e. where the mask evaluates to false) are undefined in the output of the load. Triton provides an optional `other` parameter which, when included, provides an explicit value to use for masked out values from the load. If the output from a masked load without the `other` parameter is used in a conditional, unexpected behavior can occur.
Despite the language specification, all Triton backends currently in use by PyTorch Inductor (NVIDIA, AMD, and Intel) 0-initialize masked loads if `other` is not present (we recently changed the Intel backend behavior to match NVIDIA and AMD because that's what our users expect, even if we are not following the Triton spec to the tee). This PR attempts to "future-proof" Inductor for new backends (or perhaps changes in the current backends? - we did not see any performance change from 0-initializing in the Intel XPU backend but one could imagine compiler optimizations to remove paths that depend on undefined) to add an explicit `other` in instances where later conditionals depend on the `tl.load` output. I also removed an exception to `other` behavior for boolean loads, which was put in place for a Triton bug that should be fixed. I added `other` to the getting started documentation as a clue that masked load behavior requires explicit initialization if, even though I don't expect `undef` values to cause the example code to fail if the underlying output is not 0-initialized. Finally, I added other to the `make_load` function in `select_algorithm.py`, though I wasn't able to determine if that function was actually being called.
Fixes#126535
Pull Request resolved: https://github.com/pytorch/pytorch/pull/127311
Approved by: https://github.com/jansel
We create a new landing page for PyTorch custom ops (suggested by
jansel). All of our error messages will link here, and I'll work with
the docs team to see if we can boost SEO for this page.
NB: the landing page links some non-searchable webpages. Two of those
(the Python custom ops tutorial and C++ custom ops tutorial) will turn
into actual webpages when PyTorch 2.4 comes around. I'll make the third one
(the Custom Operators Manual) once it stabilizes (we continously add new
things to it and the length means that we might want to create a custom
website for it to make the presentation more ingestable).
Test Plan:
- view docs preview.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/127400
Approved by: https://github.com/jansel
ghstack dependencies: #127291, #127292
The `usort` config in `pyproject.toml` has no effect due to a typo. Fixing the typo make `usort` do more and generate the changes in the PR. Except `pyproject.toml`, all changes are generated by `lintrunner -a --take UFMT --all-files`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/127126
Approved by: https://github.com/kit1980
The `usort` config in `pyproject.toml` has no effect due to a typo. Fixing the typo make `usort` do more and generate the changes in the PR. Except `pyproject.toml`, all changes are generated by `lintrunner -a --take UFMT --all-files`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/127126
Approved by: https://github.com/kit1980
ghstack dependencies: #127122, #127123, #127124, #127125
The `usort` config in `pyproject.toml` has no effect due to a typo. Fixing the typo make `usort` do more and generate the changes in the PR. Except `pyproject.toml`, all changes are generated by `lintrunner -a --take UFMT --all-files`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/127124
Approved by: https://github.com/Skylion007
ghstack dependencies: #127122, #127123
# Motivation
## for `torch.amp.GradScaler`,
- `torch.cpu.amp.GradScaler(args...)` is completely equivalent to `torch. amp.GradScaler("cpu", args...)`.
- `torch.cuda.amp.GradScaler(args...)` is completely equivalent to `torch.amp.GradScaler("cuda", args...)`.
So, we intend to depreate them and **strongly recommend** developer to use `torch.amp.GradScaler`.
## for `custom_fwd` and `custom_bwd`,
this is a good solution to make the custom function run with or without effect even in an autocast-enabled region and can be shared by other backends, like CPU and XPU.
So we generalize it to be device-agnostic and put them int `torch/amp/autocast_mode.py` and re-expose to `torch.amp.custom_fwd` and `torch.amp.custom_bwd`. Meanwhile, we deprecate `torch.cuda.amp.custom_fwd` and `torch.cuda.amp.custom_bwd`.
# Additional Context
Add UT to cover the deprecated warning.
No need for more UTs to cover the functionality of `torch.amp.custom_f/bwd`, the existing UTs that previously covered the functionality of `torch.cuda.amp.custom_f/bwd` can cover them.
To facilitate the review, we separate these code changes to two PRs. The first PR cover `torch.amp.GradScaler`. The follow-up covers `custom_fwd` and `custom_bwd`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/126527
Approved by: https://github.com/jgong5, https://github.com/gujinghui, https://github.com/janeyx99, https://github.com/EikanWang
Summary:
1. Define explicit `use_agent_store` on rdzv handlers. Handlers that set is true can share the store.
2. Instead of agent coordinating master_add/master_port values, the logic is now encapsulated by a *rdzv_handler* where `RendezvousInfo` will have `RendezvousStoreInfo` object that handlers must return.
- Depending on the implementation they can either:
- point to existing store (and expected to `use_agent_store` as true - point 1). Client code will rely on `TORCHELASTIC_USE_AGENT_STORE` env variable to know if the store is shared.
- build args that `torch.distributed.init_process_group` can bootstrap by creating new store.
Additional points:
- When TCPStore is shared, it should be wrapped in PrefixStore to qualify/scope namespace for other usecases.
- `next_rendezvous` signature changed to return instance of `RendezvousInfo` instead of a (store, rank, world_size) tuple for extensibility purposes.
Why:
- Reduce moving parts
- easier to swap implementation
- improve tractability
- addressing perf/debug-ability will benefit all usecases
-
Test Plan: CI
Differential Revision: D57055235
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125743
Approved by: https://github.com/d4l3k
Rule is enforced by #126103.
The rule:
- If `torch.a.b` defines a public class `C` (i.e. to be exposed in torch API namespace), then `torch.a.b` must be a public path, i.e. no `_`.
- `torch.a.b` should ideally have an `__all__` that defines what should be imported from this file when it is imported.
- All other definitions in `torch.a.b` that you don't want to expose should have a `_` prefix.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/126812
Approved by: https://github.com/wconstab
The padded dense -> jagged conversion op has the signature:
```
_fbgemm_dense_to_jagged_forward(Tensor dense, Tensor[] offsets, SymInt? total_L=None) -> Tensor
```
when `total_L` is not specified, the meta registration has a data-dependent output shape (based on `offsets[0][-1]`). Returning an unbacked SymInt here should work in theory, but traceable wrapper subclass support is missing in later code to handle deferred runtime asserts. This PR fixes this.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/126198
Approved by: https://github.com/ezyang
#### Conditions for allowlisting tensor subclasses
We allow tensor subclasses types that
(1) Do not override `__setstate__`, `__getattr__`, `__setattr__`, `__get__`, `__set__` or `__getattribute__` of `torch.Tensor` (`torch.Tensor` does not have a definition of `__getattr__`, `__get__` or `__set__` so we check that these are `None`)
(2) Use the generic `tp_alloc`
(3) Are in a module that *has been imported by the user*
to be pushed onto the stack as strings by `GLOBAL` instructions, while storing the type in a dict
The strings will be converted to the classes as appropriate when executing `REBUILD` with `_rebuild_from_type_v2`
*Note that we use `inspect.getattr_static(sys.modules[module], name)` to get the class/function as this method claims to have no code execution.
The rationale for the 3 conditions above is as follows:
The rebuild func provided by `Tensor.__reduce_ex__` is `torch._tensor._rebuild_from_type_v2`, which is defined as such (note the call to `getattr`, `Tensor.__setstate__` and the call to `as_subclass` as well as the call to `_set_obj_state` which calls `setattr`)
4e66aaa010/torch/_tensor.py (L57-L71)
`as_subclass` is implemented with a call to `THPVariable_NewWithVar`
that will eventually call `tp_alloc` here
4e66aaa010/torch/csrc/autograd/python_variable.cpp (L2053)
The `func` arg to `_rebuild_from_type_v2` for wrapper subclasses is `Tensor.rebuild_wrapper_subclass`, which will similarly call into `THPVariable_NewWithVar` and hit the above `tp_alloc`
**Note that we do not call `tp_init` or `tp_new` (i.e. `cls.__init__` or `cls.__new__`) when unpickling**
### How do we check something is a tensor subclass/constraints around imports
In order to check whether `bla` is a tensor subclass in the bytecode `GLOBAL module.name`, we need to do an `issubclass` check, which entails converting the global string to the appropriate type. We *do not* arbitrarily import modules but will perform this check as long as the given subclass (given by `module.name`) has already been imported by the user (i.e. `module in sys.modules` and `issubclass(getattr(sys[modules], name), torch.Tensor)`
This PR also allowlisted `torch._utils._rebuild_wrapper_subclass` and `torch.device` (used by `_rebuild_wrapper_subclass`)
### API for allow listing
This PR also added `torch.serialization.{add/get/clear}_safe_globals` that enables user to allowlist globals they have deemed safe and manipulate this list (for example they could allowlist a tensor subclass with a custom `__setstate__` if they have checked that this is safe).
Next steps:
- Add testing and allowlist required classes for all in-core tensor subclasses (e.g. `DTensor`, `FakeTensor` etc.)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124331
Approved by: https://github.com/albanD
This does a few things that were originally a few PRs but I am on a new machine and don't have ghstack.
If it is too problematic to review, I can re-split, just let me know.
This does:
- Cleanup context manager use in test_flop_counter
- Remove need for mod argument in FlopCounterMode, warning about it
- Re-implement a Module tracker from scratch using global forward Module use and multi_grad_hook (we cannot use global backward Module hook because they don't look for nested Tensor and they're custom Function based instead of multi_grad_hook).
- Update FlopCouterMode to use the new ModuleTracker. All the existing test suite passes as-is (only changes there are new tests and refactoring mentioned above)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125352
Approved by: https://github.com/mikaylagawarecki
`torch.utils.benchmark.Compare` is not directly exposed in torch.utils.benchmark documentation.
I think this is a valuable resource to add since it can help people embracing the torch benchmark way of doing things, and help people building documentation towards it.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125009
Approved by: https://github.com/mikaylagawarecki
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #124944
* #124939
* __->__ #122965
Differential Revision: [D55493240](https://our.internmc.facebook.com/intern/diff/D55493240/)
*This PR is now ready for merge and is not an RFC*
Major choices are:
-- the introduction of the AsyncStager protocol
-- removed `executor` from param.
-- leave async as a separate method (for now)
This proposal seeks to add extension points to dcp.async_save, allowing users to:
- Specify a specific staging method when calling async_save
- Allow a vehicle for also making the staging method async, to allow for cases where we may want to overlap with the training loop (e.g., overlap d2h with and only synchronize at the optim.step)
- Potentially specify the execution method for doing async_save in parallel. For example some users may prefer a subprocess over a thread to avoid GIL issues.
A totally reasonable alternative to this entire proposal is to expect users who want this level of customization
to write their own custom async save methods. Here's an example which addresses the issues mentioned
in PR comments.
```
def custom_async_save(...):
# this step accomplishes staging and includes the usual 'planning' calls (issue 1)
buffered_writer = CpuBufferedWriter() # this is stateful, contains a copy of state_dict
dcp.save(state_dict, storage_writer=buffered_writer)
final_storage_writer = FileSystemWriter()
mp.spawn( # issue2 is gone, do whatever you want here
dcp.save, # or some custom sub-process method which calls dcp.save under the hood
buffered_writer.state_dict, # lot's of way's to do this, not really the most important part
checkpoint_id=checkpoint_id,
storage_writer=storage_writer,
planner=planner,
process_group=process_group, # this actually wouldn't work, but again not the pt.
)
# leaving out the rest of the details for managing your extra special subprocess.
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122965
Approved by: https://github.com/daulet-askarov
Fixes#124528
Going over the options for our MapAllocator and what they do, I don't think any other of them need to be piped up to `torch.load`
4f29103749/aten/src/ATen/MapAllocator.h (L8-L16)
~However, I wonder if this `MmapVisibility(Enum)` is a good way to represent "or-ing" together of `mmap` flags if we want to extend it in the future. I looked over the flags for [`mmap(2)`](https://man7.org/linux/man-pages/man2/mmap.2.html), and could not immediately see how most of them would be useful for `torch.load` (would maybe `MAP_LOCKED` (like `mlock`) or `MAP_HUGE` ever be worthwhile?)~
Using the flags provided by the python `mmap` library so that we can extend the allowed flags and pipe them down to the cpp `mmap` call if there is a need for other flags in the future
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124889
Approved by: https://github.com/albanD
This PR introduces a new way of building `dynamic_shapes` for export. The idea is to build up a mapping from input tensors to the dynamic shapes that should be assigned to their corresponding fake tensors.
This mapping is automatically converted to the current form of `dynamic_shapes`, which must exactly match the structure of inputs. We do this by using pytree utils.
With the current `dynamic_shapes`, we had to be careful about user-defined classes that are registered with pytree, since such classes are not necessarily polymorphic containers; they may be fine containing tensors, but not dynamic shapes. Thus we had decided to allow input instances of such classes to be associated with dynamic shapes in flattened form. This decision needs to be mirrored in this PR as well. To make it easier to keep these code paths in sync, we refactor the current recursive procedure for associating inputs with dynamic shapes to use the same pytree utils. This needs minor fixes to a few tests where `dynamic_shapes` were not exactly matching the structure of inputs.
Differential Revision: D56551992
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124898
Approved by: https://github.com/zhxchen17
Summary:
This makes barrier and rank operations linear instead of quadratic with the number of workers. This drastically improves performance for rendezvous when running with over 1000 hosts.
This uses 2 approaches for different areas:
* local rank assignment: each worker does 1 set and 1 get, local ranks are assigned on the rank 0 host in a O(n) operation which reduces total store operations to be linear with number of workers.
* exit_barrier: use a counter and a final flag so each worker has to do max 1 set, 1 get and 1 add.
At 4000 hosts we see torchelastic be able to run in as little as 10 seconds down from 373 seconds.
Test Plan:
This is testing using many small tests running on a remote cluster.
{D56549942}
```
torchx run --scheduler mast -- --image=torchelastic_benchmark --j=4000x1
```
Differential Revision: D56605193
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124982
Approved by: https://github.com/kiukchung, https://github.com/kurman
MTIA device has its own Module in PyTorch now.
torch.mtia has following APIs similar to other backends. The lazy_init is also supported.
```
__all__ = [
"init",
"is_available",
"synchronize",
"device_count",
"current_device",
"current_stream",
"default_stream",
"set_stream",
"stream",
"device",
]
```
------------
For device management. We expand AccleratorHooksInterface to support generic device management and it can be used in both C++ and PyThon.
```
def _accelerator_hooks_device_count() -> _int: ...
def _accelerator_hooks_set_current_device(device_index: _int) -> None: ...
def _accelerator_hooks_get_current_device() -> _int : ...
def _accelerator_hooks_exchange_device(device_index: _int) -> _int : ...
def _accelerator_hooks_maybe_exchange_device(device_index: _int) -> _int : ...
```
---------
Adding get_device_module API to retrieve device modules for different device types.
```
def get_device_module(device: Optional[Union[torch.device, str]] = None)
```
---------
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123612
Approved by: https://github.com/albanD
ghstack dependencies: #123611
Previously, unbacked SymInts would gradually get larger and larger as we kept rebinding them. Now, we do the replacement to preserve the old symbol.
Actually doing this is a bit tricky. Here’s the order things happen when retracing data dependent:
1. Run fake tensor prop: allocate new unbacked SymInt
2. Run proxy tensor mode, calculate bindings and associate them with FX node
3. Run PropagateUnbackedSymInts, rename unbacked bindings to their old ones so they are consistent
So the problem is when we calculate bindings in step (2), we don't know
what the original names are yet, we only find out later at (3). But by
the time (3) runs, we've already stuffed some new bindings in
meta["unbacked_bindings"] and we don't know how to update them! To fix
this, I introduce resolve_unbacked_bindings which post facto applies any
of the renamings we discovered in (3).
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124782
Approved by: https://github.com/lezcano
ghstack dependencies: #124310, #124314, #124316, #124394, #124739
This PR has a lot of "draw the rest of the fucking owl" energy. Here's how to break it down.
1. **torch/_inductor/graph.py** - We start by tightening unbacked symbol invariants. Specifically, as we lower FX nodes, we check whether or not every unbacked_binding recorded on the FX node meta, actually ends up getting bound (according to get_unbacked_symbol_defs) in all the buffers generated by the lowering. Hopefully this invariant is self evident. This leads to a lot of failures.
2. **torch/_inductor/ir.py** - Problem 1: There is softness in how Inductor computes defs of unbacked symbols in IR node. Previously, we tried to infer it by looking at the output sizes/strides/etc and see if new unbacked symbols popped up that we hadn't seen in the inputs. I don't know exactly what was buggy about the old code, but sometimes we would fail to notice an unbacked symbol had been bound, or rebind an unbacked symbol multiple times. Fortunately, thanks to the earlier PRs in our stack, we now have a nice list of unbacked symbol bindings from FX, so we now just store it directly on ExternKernel and use it directly to report defs. This has to be done twice: once for FallbackKernel (e.g., nonzero) and once for DynamicScalar (e.g., item) (see also **torch/_inductor/lowering.py**, **torch/_inductor/codegen/wrapper.py** and **torch/_inductor/codegen/cpp_wrapper_cpu.py** for the lowering and codegen changes for item)
* **process_kernel** - Sidequest! It turns out that Inductor lowering can reallocate unbacked symbols. This happens specifically when we repropagate fake tensors through the operator in `process_kernel`. This repropagation process is necessary because Inductor may have changed the strides of input tensors, and it must now recompute the strides so that it can continue to appropriately plan the rest of the lowering process. This is fine: we just make sure we do the rebind unbacked + compute_unbacked_bindings dance we've been doing previously in the PR stack. But instead of putting unbacked_bindings on a new FX node, they go straight into our unbacked_bindings on the Inductor IR node.
* **codegen_unbacked_symbol_defs** - Sidequest! FallbackKernel lowering is done in two steps. First, you emit the FallbackKernel buffer. Then, you emit MultiOutput buffers which actually give access to the individual outputs of FallbackKernel, which may have been multi-output. There is a design decision here: does the FallbackKernel bind the unbacked symbols, or the MultiOutput buffer? Historically, we put the binding on MultiOutput buffer, because it's more convenient: the FallbackKernel buffer is fake, in fact, it doesn't even get a name in C++ codegen. But it's kind of inconsistent with the keypath model that we've been tracking unbacked bindings with: if you have a multi-output node, you'd expect a keypath like `[0].size()[0]` representing the first output's first dimension size. That suggests that it's the FallbackKernel that should define the things. So that was my first implementation. Unfortunately, the C++ codegen is too cursed and I could not understand how to make it work in that case. So now we just unsoundly assume you cannot have multi-output data dependent output, and do the codegen in MultiOutput. There are some comments explaining exactly what we are improperly assuming.
3. **_rename_unbacked_to** in **torch/fx/experimental/symbolic_shapes.py** - Previously, when we renamed unbacked symbols, we clobbered any facts we previously knew about them. So for example, if we had a replacement `u0 -> s0` but then we renamed u0 to u1, we would now setup the replacement `u0 -> u1`, clobbering the old replacement. This apparently didn't matter in earlier PRs in the stack, but with Inductor now on the ball, there were some tests that indicated this was a problem. The solution is easy: if u0 had a preexisting replacement, reapply it to u1. However...
* **torch/_functorch/_aot_autograd/collect_metadata_analysis.py** - When we run forward analysis, this triggers fake tensor repropagation and fresh allocations. Previously, we just cleared out the pending symbols when finished the analysis. But with the change above, this would also migrate replacements to the new symbols... which are now dead. So now we explicitly suppress generation of these symbols with `ignore_fresh_unbacked_symbols` so that no rebinding happens at all.
* **torch/_dynamo/eval_frame.py** - same deal; I just searched for all sites we called clear() on pending
4. The last step is fixing the long tail of extra problems that show up, now that unbacked_bindings are load bearing into Inductor
* **torch/_dynamo/eval_frame.py** - Some of the exports are making copies of nodes without repropagating fake tensors, so in this case, it is important to also copy the `unbacked_bindings` (apparently this didn't matter before without the Inductor changes)
* **torch/_export/pass_base.py** - I discover that this is doing fake tensor repropagation via a test suite failure. Do the same playbook as AOTAutograd: PropagateUnbackedSymInts too! Actually, they also have implemented their own tracer as well, so do the same playbook as proxy_tensor: record unbacked_bindings on the newly traced nodes. UGH code duplication.
* **torch/_subclasses/fake_tensor.py**, **torch/_subclasses/fake_impls.py** (with call site updates at **torch/_functorch/_aot_autograd/traced_function_transforms.py** and **torch/fx/passes/fake_tensor_prop.py**) - What's this new epoch thing? I noticed that sometimes I would be retracing, call nonzero() on a fake tensor, and not allocate a new unbacked symbol. This is actually bad, because if I don't get a new unbacked symbol, I don't know there's a binding site, and `unbacked_bindings` is now missing a binding. The reason for this is memoization: if I reuse the exact same fake tensor on my retrace, it will already have an unbacked symint memoized on it and we will short circuit allocation. Well, that's no good. So I associate the memos with a fake tensor epoch, and every time you start a new fake tensor propagation from scratch, you bump the epoch so that I clear all the memos.
* **torch/_inductor/scheduler.py** - I notice in unit tests that V.current_node is not always set when we call process_kernel. So I save it into the IR node and restore it when we are running `get_estimated_runtime`.
* **torch/fx/experimental/symbolic_shapes.py** - A few things
* **rebind_unbacked** (re **_tensor_version**). Ordinarily, when you have an unbacked SymInt, you persistently hvae it all the way to the end of the program. `_tensor_version` violates this: this generates an unbacked SymInt (for reasons I don't quite understand?) and then gets rid of it later. This triggered an assert violation. I think this op is kind of misusing unbacked SymInt, but I didn't know how to refactor it, so it gets a special case.
* **rebind_unbacked** (re **Simplify SymBool binding**). Ugh, SymBool, what a pain in the butt. I have an assert that you can only rebind unbacked symbol to another unbacked symbol. This assert fails when a boolean is involved, because the result of running keypath on the result is not `u1`, it's `sympy.Piecewise(... sympy.Eq(u1, 1) ...)`. This is actually just `u1`, but Sympy doesn't know it because it doesn't know that `u1` value range is `[0, 1]`. So we manually implement the simplification needed to get the assert to pass.
* **compute_unbacked_bindings** (re **This is pretty fragile**). There is a really funny disaster involving memoization and Inductor process kernel. Ordinarily when I retrace, if there was a memo hit in the old trace, there will be a memo hit in the new trace. However, Inductor process kernel breaks this, because it recreates fake tensor inputs to the operator call from scratch (since they might have different strides), and obviously these tensor inputs don't have the memo from the old one. I tried a little bit to try to manually transplant the memo to the new fake tensor but it seemed hopeless, so I just let the fresh symbol ride, allocating a new unbacked symbol. However, in one of our tests, we rely on knowing that the first nonzero call is equal to the second (memoized) nonzero call. The equality test looked pretty easy to discharge, so I just went ahead and added a deferred runtime assert to this effect and it worked.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124394
Approved by: https://github.com/jansel
ghstack dependencies: #124310, #124314, #124316
This is a partial revert of https://github.com/pytorch/pytorch/pull/124059
Like in #124297, profiling has revealed that testing equality on *every* output is kind of expensive. So we only test equality when we know there is an unbacked binding. This is the same playbook as the previous PR, just on FakeTensorProp instead of PropagateUnbackedSymInts. Note that we also need to populate `unbacked_bindings` in proxy_tensor.py, since we're generating an entirely new graph in that case.
We now have enough propagation that we're able to trigger a bug related to divisibility replacement. In https://github.com/pytorch/pytorch/pull/113165 we allowed to replace `u0` with `u1 * c` for some constant c, when we have determined that u0 is divisible by c. However, where does the binding for u1 come from? What we will have in practice is that there is some node that is supposed to have bound u1, but which actually is getting a `u1 * c` in its output. So, to get u1, we must divide out c. Fortunately, under the divisibility condition, this is always possible (but remember, we must test divisibility at runtime!)
Because we have tightened up asserts, it is now an error to allocate unbacked SymInts and then fail to track them under unbacked_bindings. In torch/_dynamo/eval_frame.py and torch/_functorch/_aot_autograd/collect_metadata_analysis.py there are examples of benign cases where we repropagated fake tensors but then immediately threw away the results. In these cases, it's not appropriate to rebind, since we're still using the old FX graph that has all of the old symbols. So we just manually clear it. It is possible that other cases will need to be updated, so this PR is "risky" from the perspective of hitting fbcode.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124310
Approved by: https://github.com/lezcano
Summary:
Adding function to log additional debug information before killing the expired watchdog timers.
Additional information like stack trace can be added in the debug function using worker process IDs from expired timers.
Test Plan: buck test mode/opt caffe2/test/distributed/elastic/timer:file_based_timer_test
Differential Revision: D56044153
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123883
Approved by: https://github.com/kurman
MTIA device has its own Module in PyTorch now.
torch.mtia has following APIs similar to other backends. The lazy_init is also supported.
```
__all__ = [
"init",
"is_available",
"synchronize",
"device_count",
"current_device",
"current_stream",
"default_stream",
"set_stream",
"stream",
"device",
]
```
------------
For device management. We expand AccleratorHooksInterface to support generic device management and it can be used in both C++ and PyThon.
```
def _accelerator_hooks_device_count() -> _int: ...
def _accelerator_hooks_set_current_device(device_index: _int) -> None: ...
def _accelerator_hooks_get_current_device() -> _int : ...
def _accelerator_hooks_exchange_device(device_index: _int) -> _int : ...
def _accelerator_hooks_maybe_exchange_device(device_index: _int) -> _int : ...
```
---------
Adding get_device_module API to retrieve device modules for different device types.
```
def get_device_module(device: Optional[Union[torch.device, str]] = None)
```
---------
Differential Revision: [D56443356](https://our.internmc.facebook.com/intern/diff/D56443356)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123612
Approved by: https://github.com/albanD
ghstack dependencies: #123611
This subsumes https://github.com/pytorch/pytorch/pull/124069
In the original PR, my idea was that when we run PropagateUnbackedSymInts, we check that the sizes before and after are exactly the same. This ended up turning up lots of bugs that I didn't feel like fixing. Separately, Ivan let me know that this pass was quite expensive in terms of compile time, since we spent a lot of time thinking about the equalities.
To kill two birds with one stone, we now only check for equality precisely when an unbacked SymInt was bound (thanks to the previous PR in this stack, we now have this information). Specifically, we look to see if `meta["unbacked_bindings"]` is set on the old node, and if it is, we assert the old value is equal to the new value from the repropagation. Note that the pytree key is used to actually extract the new value from the example value, as it may be nested inside an, e.g., tensor size.
We do something a bit naughty at the end: we use `defer_runtime_assert` to actually teach ShapeEnv about the equality. This is implementationally equivalent to what we used to do, but we're going to change this later soon.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124297
Approved by: https://github.com/lezcano
ghstack dependencies: #124290
The important comment:
```
# Whenever we allocate a fresh unbacked Symbol, we add it to this
# pending list. Unbacked symbol allocation can occur at unpredictable
# points during meta tensor propagation, but at some point, the we
# have to know what the binding site for an unbacked symbol is, and
# this is computed when we actually place the node in the graph. The
# important thing is that we always actually handle every unaccounted
# for unbacked symbol, so this list helps us keep track of them and
# then make sure they are all accounted for.
#
# We could potentially give rise to errors earlier by lexically
# scoping when we do propagation, and only allowing unbacked symbols
# to be allocated at this point in time. However this is inconvenient
# to do in Dynamo, because fake tensor propagation is far from when we
# analyze binding sites (set_example_value), so we do it in a more
# mutatey way.
#
# NB: fresh unbacked symbols NEVER get substitutions applied to them,
# they are binding sites!
```
The compute_unbacked_bindings is the other half of the equation: the thing that actually consumes the pending_fresh_unbacked_symbols and does something with them. Important comment:
```
After having run fake tensor propagation and producing example_value
result, traverse example_value looking for freshly bound unbacked
symbols and record their paths for later. It is an error if
we have allocated an unbacked SymInt but it cannot be found in
example_value. (NB: this means if you have a multi-output
function, you must call this on the tuple of tensor output, you
cannot wait!)
```
For example, if I return a tensor with size `[u0, u1]`, and u1 is a fresh unbacked SymInt, then I'll have `{u1: KeyPath(".size(1)")}`, telling me I can get u1 by running `size(1)` on the result of this node. u0 is not fresh (it probably flowed in as an argument), so I don't generate a binding for it.
I eventually intend to propagate this information all the way to Inductor lowering, where extra metadata about unbacked symbol binding will be canonically used for codegen, instead of trying to infer it from defs/uses.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124290
Approved by: https://github.com/lezcano
This PR:
- exposes torch.testing._internal.optests.opcheck as
torch.library.opcheck
- Adds support for CustomOpDef (aka functions decorated with
torch.library.custom_op) to opcheck.
Test Plan:
- Updated tests
- We validated opcheck's design internally.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/124496
Approved by: https://github.com/williamwen42
Following the example of PyTorch supporting a preferred Linalg library (cusolver or magma), this PR introduces a preferred blas library selector of either cublas or cublaslt for CUDA and hipblas or hipblaslt for ROCm via normal hipification of sources.
The default blas implementation remains cublas or hipblas. cublaslt or hipblaslt can be enabled using environment variable TORCH_BLAS_PREFER_CUBLASLT=1 (or TORCH_BLAS_PREFER_HIPBLASLT=1 as an alias) or by calling `torch.backends.cuda.preferred_blas_library(backend="cublaslt")` or as an alias `backend="hipblaslt"`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122106
Approved by: https://github.com/lezcano
MTIA device has its own Module in PyTorch now.
torch.mtia has following APIs similar to other backends. The lazy_init is also supported.
```
__all__ = [
"init",
"is_available",
"synchronize",
"device_count",
"current_device",
"current_stream",
"default_stream",
"set_stream",
"stream",
"device",
]
```
------------
For device management. We expand AccleratorHooksInterface to support generic device management and it can be used in both C++ and PyThon.
```
def _accelerator_hooks_device_count() -> _int: ...
def _accelerator_hooks_set_current_device(device_index: _int) -> None: ...
def _accelerator_hooks_get_current_device() -> _int : ...
def _accelerator_hooks_exchange_device(device_index: _int) -> _int : ...
def _accelerator_hooks_maybe_exchange_device(device_index: _int) -> _int : ...
```
---------
Adding get_device_module API to retrieve device modules for different device types.
```
def get_device_module(device: Optional[Union[torch.device, str]] = None)
```
---------
@exported-using-ghexport
Differential Revision: [D52923602](https://our.internmc.facebook.com/intern/diff/D52923602/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123612
Approved by: https://github.com/albanD
ghstack dependencies: #123611
This PR:
- adds a new torch.library.register_fake and deprecates
torch.library.impl_abstract. The motivation is that we have a lot of
confusion around the naming so we are going to align the naming with
the actual subsystem (FakeTensor).
- renames `m.impl_abstract_pystub("fbgemm_gpu.sparse_ops")` to
`m.has_python_registration("fbgemm_gpu.sparse_ops")`. No deprecation
here yet; I need to test how this works with static initialization.
- Renames a bunch of internals to match (e.g. abstractimplpystub ->
pystub)
I'm scared to rename the Python-side internal APIs (e.g.
torch._library.abstract_impl) because of torch.package concerns. I'll do
that in its own isolated PR next just in case it causes problems.
DEPRECATION NOTE: torch.library.impl_abstract was renamed to to
torch.library.register_fake. Please use register_fake. We'll delete
impl_abstract in a future version of PyTorch.
Test Plan:
- existing tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123937
Approved by: https://github.com/albanD
Summary:
Building hook for external mechanism to monitor the health of torch elastic launcher. Health check server takes dependency on FileTimerServer to check if launcher is healthy or not. It will be always healthy if FileTimerServer is disabled.
Implementation of start_healthcheck_server is unsupported, however tcp/http server can be started on specific port which can monitor the aliveness of worker_watchdog and accordingly take the action.
Test Plan: buck test mode/opt caffe2/test/distributed/elastic/agent/server/test:local_agent_test
Differential Revision: D55837899
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123504
Approved by: https://github.com/kurman
Summary:
Building hook for external mechanism to monitor the health of torch elastic launcher. Health check server takes dependency on FileTimerServer to check if launcher is healthy or not. It will be always healthy if FileTimerServer is disabled.
Implementation of start_healthcheck_server is unsupported, however tcp/http server can be started on specific port which can monitor the aliveness of worker_watchdog and accordingly take the action.
Test Plan: buck test mode/opt caffe2/test/distributed/elastic/agent/server/test:local_agent_test
Differential Revision: D55108182
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122750
Approved by: https://github.com/kurman
Fixes https://github.com/pytorch/pytorch/issues/123068
Fixes https://github.com/pytorch/pytorch/issues/111256
While investigating the flaky doc build failure .w.r.t duplicated `torch.ao.quantization.quantize` docstring warning, i.e. https://github.com/pytorch/pytorch/actions/runs/8532187126/job/23376591356#step:10:1260, I discover an old but still open bug in Sphinx https://github.com/sphinx-doc/sphinx/issues/4459. These warnings have always been there, but they are hidden because we are using `-j auto` to build docs with multiple threads. It's just by chance that they start to surface now.
The issue can be reproduced by removing `-j auto` from https://github.com/pytorch/pytorch/blob/main/docs/Makefile#L5 and run `make html` locally. Then, these warnings shows up consistently. As `make html` treats warnings as errors, they will fail the build.
```
...
/data/users/huydo/conda/py3.8/lib/python3.8/site-packages/torch/ao/quantization/quantize.py:docstring of torch.ao.quantization.quantize.quantize:1: WARNING: duplicate object description of torch.ao.quantization.quantize, other instance in quantization, use :noindex: for one of them
/data/users/huydo/conda/py3.8/lib/python3.8/site-packages/torch/nn/parallel/data_parallel.py:docstring of torch.nn.parallel.data_parallel.data_parallel:1: WARNING: duplicate object description of torch.nn.parallel.data_parallel, other instance in nn, use :noindex: for one of them
/data/users/huydo/conda/py3.8/lib/python3.8/site-packages/torch/nn/utils/spectral_norm.py:docstring of torch.nn.utils.spectral_norm.spectral_norm:1: WARNING: duplicate object description of torch.nn.utils.spectral_norm, other instance in nn, use :noindex: for one of them
/data/users/huydo/conda/py3.8/lib/python3.8/site-packages/torch/nn/utils/weight_norm.py:docstring of torch.nn.utils.weight_norm.weight_norm:1: WARNING: duplicate object description of torch.nn.utils.weight_norm, other instance in nn, use :noindex: for one of them
/data/users/huydo/github/pytorch/docs/source/nn.rst:579: WARNING: duplicate object description of torch.nn.parallel.data_parallel, other instance in generated/torch.nn.functional.torch.nn.parallel.data_parallel, use :noindex: for one of them
/data/users/huydo/github/pytorch/docs/source/nn.rst:594: WARNING: duplicate object description of torch.nn.utils.spectral_norm, other instance in generated/torch.nn.utils.spectral_norm, use :noindex: for one of them
/data/users/huydo/github/pytorch/docs/source/nn.rst:595: WARNING: duplicate object description of torch.nn.utils.weight_norm, other instance in generated/torch.nn.utils.weight_norm, use :noindex: for one of them
/data/users/huydo/github/pytorch/docs/source/quantization.rst:1348: WARNING: duplicate object description of torch.ao.quantization.quantize, other instance in generated/torch.ao.quantization.quantize, use :noindex: for one of them
...
```
The fix is just to clean up those duplicated placeholder py:module docs, which were there because these modules didn't have any docs originally.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/123244
Approved by: https://github.com/andrewor14, https://github.com/malfet
This is the entrypoint for defining an opaque/blackbox (e.g. PyTorch will
never peek into it) custom op. In this PR, you can specify backend impls
and the abstract impl for this op.
NB: most of this PR is docstrings, please don't be intimidated by the
line count.
There are a number of interesting features:
- we infer the schema from type hints. In a followup I add the ability
to manually specify a schema.
- name inference. The user needs to manually specify an op name for now.
In a followup we add the ability to automatically infer a name (this
is a little tricky).
- custom_op registrations can override each other. This makes them
more pleasant to work with in environments like colab.
- we require that the outputs of the custom_op do not alias any inputs
or each other. We enforce this via a runtime check, but can relax this
into an opcheck test if it really matters in the future.
Test Plan:
- new tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122344
Approved by: https://github.com/ezyang, https://github.com/albanD
This supersedes the previous `Guards Overview" as a more comprehensive
approach to most of the main topics within Dynamo.
In the future, we could add specific sections for each of the topics
discussed here.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122305
Approved by: https://github.com/msaroufim
`dynamo.explain()` was updated to return a structure but the docs weren't updated to match.
- Update the docs to use the new API
- Remove some dead code left when `explain` was updated.
- Drive-by: Fix some `nopython` uses that I noticed
- Drive-by: I noticed an ignored error coming from CleanupHook on shutdown - make it check the global before setting it.
Fixes#122573
Pull Request resolved: https://github.com/pytorch/pytorch/pull/122745
Approved by: https://github.com/jansel
Previously, when we applied a replacement, a SymInt that was
previously an unbacked SymInt would then transmute into whatever
we replaced it into (e.g., a constant).
This has a major downside: we often look at SymInts associated with
FX nodes (e.g., the meta of x.item() return) to find out where the
unbacked SymInt was allocated. If we replace it, we no longer can
find out where, e.g., u1 was allocated! But we need to know this
so we can generate deferred runtime asserts like u1 == s0.
To solve this problem, I have a special mode for replace, resolve_unbacked=False, which lets you disable substitutions on unbacked SymInts. When reporting node.expr, we preferentially avoid applying unbacked SymInt substitutions. To understand if we might accidentally reapply the substitution later, before we have reached the deferred runtime assert, we must study the calls to simplify() in ShapeEnv. My audit turns up these sites:
* `produce_guards`: this is fine, deferred runtime asserts never show up here, we must NOT have unbacked SymInts show up here. Similarly `get_nontrivial_guards`.
* `_maybe_evaluate_static`: this is fine, we are using this to determine if it is necessary to produce a guard/runtime assert. We don't want to reissue a runtime assert if we've already asserted on it, and replacements can help us understand if this has occurred.
* `_simplify_floor_div`: this is a legitimate bug, it needs to be `resolve_unbacked=False`
* `_refine_ranges`: this is fine, a refined range doesn't affect what runtime asserts we issue
* `_update_divisible`: this updates the `self.divisible` set, which specifies when we can simplify away divisibility constraints. Since this affects replacements only, it won't cause us to oversimplify a user provided expression.
There are some situations where we DO want to always apply the substitution, specifically when we have the duplicate symbol problem (we retrace an item call and get u0 and u1 which refer to the same thing.) I don't want two symbols in this case, so a special `rename_unbacked_to` is provided which sets up the unconditional renaming.
Along the way, I make a refinement to `_update_var_to_range`: if you update a var range for a size-like unbacked SymInt, you are now no longer allowed to set its lower bound below 2. This is because if you could, then our size oblivious tests for it would be inconsistent. Actually, I think there is still some inconsistency, because if you assert `u0 == 0` we will still end up with this in deferred runtime asserts, and we will then use this to simplify these statements to be True everywhere else. Maybe we should forbid this kind of refinement; not done in this PR.
Fixes https://github.com/pytorch/pytorch/issues/119689
Fixes https://github.com/pytorch/pytorch/issues/118385
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/120816
Approved by: https://github.com/lezcano
See #113541
The PR allows for registering and controlling multiple RNG states using indices, ensuring cudagraph-safe operations, and includes both C++ and Python API changes to support this functionality.
cc @eellison @anijain2305 @jansel @ezyang @ptrblck @csarofeen @mcarilli
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114068
Approved by: https://github.com/ezyang
This PR corrects the example in the AOTInductor example which currently fails with:
```
/home/ubuntu/test/inference.cpp:21:62: error: cannot bind non-const lvalue reference of type ‘std::vector<at::Tensor>&’ to an rvalue of type ‘std::vector<at::Tensor>’
21 | std::cout << runner.run({torch::randn({2, 10}, at::kCPU)})[0] << std::endl;
|
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121672
Approved by: https://github.com/desertfire
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
As discussed on [slack](https://pytorch.slack.com/archives/C3PDTEV8E/p1703699711772289) adding Andrew Gu's advanced FSDP design notes with a few additions from myself based on our discussion.
I hope I did the RST right, I haven't done RST in a while.
- The first section is Andrew's words verbatim + formatting
- The second section is Andrew's words verbatim + formatting + a few of my additions that were confirmed by Andrew, and which hopefully should help understand the process better.
tagging @albanD as requested.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117323
Approved by: https://github.com/awgu
As discussed on [slack](https://pytorch.slack.com/archives/C3PDTEV8E/p1703699711772289) adding Andrew Gu's advanced FSDP design notes with a few additions from myself based on our discussion.
I hope I did the RST right, I haven't done RST in a while.
- The first section is Andrew's words verbatim + formatting
- The second section is Andrew's words verbatim + formatting + a few of my additions that were confirmed by Andrew, and which hopefully should help understand the process better.
tagging @albanD as requested.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117323
Approved by: https://github.com/albanD, https://github.com/awgu
This PR adds the bare minimum functionality to get torchbind working in an e2e testable way on PT2.
It implements:
* ProxyTensor support
* Simple torch.export support (proxytensor-only path, e.g. non-strict).
* add some tests exercising the path.
Because all this is not fully baked, I hide the functionality behind a feature flag (`enable_torchbind_tracing()`) so it does not affect regular users for now.
Still on the agenda:
* Dynamo support
* Actual FakeMode support
* Mutability support
Hoping to get this first bit in as a standalone, as it will unblock some more extensive experimentation/testing going on internally.
Differential Revision: [D51825372](https://our.internmc.facebook.com/intern/diff/D51825372/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117697
Approved by: https://github.com/SherlockNoMad
This adds a function `statically_known_true` for `SymBool` that works
like inductor's `is_expr_static_and_true`. That is, it tries to simplify the
expression to a constant or returns `False` if it cannot be simplified.
This is useful in cases that can be optimized if the condition is met,
otherwise it doesn't effect correctness so we can avoid adding guards.
I also use this new function in inductor for `FakeTensorUpdater` and
`remove_noop_pass` which both generated unexpected guards previously.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117359
Approved by: https://github.com/lezcano
A less general version of this wrapper was used in the keynote on
`torch.compile(numpy)`. We expose a generic version of the wrapper
that works seamlessly with `torch.compile`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114610
Approved by: https://github.com/albanD
When exporting a model with a convolution kernel on cpu, if mkldnn is disabled and nnpack is enabled, export will go down the nnpack optimized convolution kernel for certain shapes ((code pointer)[cd449e260c/aten/src/ATen/native/Convolution.cpp (L542-L552)]). This means that we will automatically create a guard on that certain shape. If users want to export without any restrictions, one option is to disable nnpack. However, no config function exists for this, so this PR is adding a config function, similar to the `set_mkldnn_enabled` function.
Original context is in https://fb.workplace.com/groups/1075192433118967/posts/1349589822345892/?comment_id=1349597102345164&reply_comment_id=1349677642337110.
To test the flag, the following script runs successfully:
```
import os
import torch
from torchvision.models import ResNet18_Weights, resnet18
torch.set_float32_matmul_precision("high")
model = resnet18(weights=ResNet18_Weights.DEFAULT)
model.eval()
with torch.no_grad():
# device = "cuda" if torch.cuda.is_available() else "cpu"
torch.backends.mkldnn.set_flags(False)
torch.backends.nnpack.set_flags(False) # <--- Added config
device = "cpu"
model = model.to(device=device)
example_inputs = (torch.randn(2, 3, 224, 224, device=device),)
batch_dim = torch.export.Dim("batch", min=2, max=32)
so_path = torch._export.aot_compile(
model,
example_inputs,
# Specify the first dimension of the input x as dynamic
dynamic_shapes={"x": {0: batch_dim}},
# Specify the generated shared library path
options={
"aot_inductor.output_path": os.path.join(os.getcwd(), "resnet18_pt2.so"),
"max_autotune": True,
},
)
```
I'm not sure who to add as reviewer, so please feel free to add whoever is relevant!
Pull Request resolved: https://github.com/pytorch/pytorch/pull/116152
Approved by: https://github.com/malfet
This PR is proposing a new approach to solve the nn/optim only linked by python object identity problem.
The idea is to have a function that can swap the content of two Tensors t1 and t2 while preserving all the old references.
This would allow us to swap the `model.weight` with a new Tensor (can be any subclass of Tensor and any TensorImpl (xla, sparse, nested tensorimpl would work)). The use within nn will be done in a follow up.
This is done by swapping the whole content of the PyObject and then putting back the fields associated with external references (refcount, gc tracking and weakrefs).
Note that we have to properly handle all the cases where there is memory used before the public pointer PyObject* and where the PyObject is bigger due to dict/weakref being inlined (older CPython version) or due to slots.
The main limitation of this approach is that the number of slots need to match for the objects being swapped and thus limit usage of slots in subclasses.
Draft right now to see what @colesbury thinks about doing this?
Pull Request resolved: https://github.com/pytorch/pytorch/pull/111747
Approved by: https://github.com/colesbury
Summary:
Rename _device_mesh.py to device_mesh.py, update all callsites, add documentation.
We created stubs for public class and methods in torch.distributed.device_mesh so that torch.distributed.device_mesh can be imported with or without distributed is available().
Original diff reverted: D51629761
Original PR reverted: https://github.com/pytorch/pytorch/pull/115099
Prior to landing, CI signals are all passed. Shipit added the "ci/trunk" label to the PR and DID NOT wait for it and went ahead committing. More context can be found in the reverted PR above.
Test Plan: CI.
Differential Revision: D51861018
Pull Request resolved: https://github.com/pytorch/pytorch/pull/115193
Approved by: https://github.com/fegin
Adds a useful high level wrapper for calling `dist.save/load` with the correct storage readers and writers.
Instead of doing:
```
DCP.save(
state_dict={...},
storage_writer=StorageWriter(...)
)
DCP.load(
state_dict={...},
storage_reader=StorageReader(...)
)
```
We can now do:
```
checkpointer = Checkpointer(...)
checkpointer.save(state_dict={...})
checkpointer.load(state_dict={...})
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114603
Approved by: https://github.com/fegin, https://github.com/wz337
Summary:
Documenting the `Work` object
For a collective (broadcast, all_reduce, etc.) when async_op=True we return a `Work` object to which users can call `.wait()`, `.is_success()`, among other things but this class is not documented
Test Plan: Preview the docs build in OSS
Differential Revision: D51854974
Pull Request resolved: https://github.com/pytorch/pytorch/pull/115172
Approved by: https://github.com/wconstab
# Summary
This PR introduces a new Tensor subclass that is designed to be used with torch.nn.functional.scaled_dot_product_attention. Currently we have a boolean `is_causal` flag that allows users to do do causal masking without the need to actually create the "realized" attention bias and pass into sdpa. We originally added this flag since there is native support in both fused kernels we support. This provides a big performance gain ( the kernels only need to iterate over ~0.5x the sequence, and for very large sequence lengths this can provide vary large memory improvements.
The flag was introduced when the early on in the kernel development and at the time it was implicitly meant to "upper_left" causal attention. This distinction only matters when the attention_bias is not square. For a more detailed break down see: https://github.com/pytorch/pytorch/issues/108108. The kernels default behavior has since changed, largely due to the rise of autogressive text generation. And unfortunately this would lead to a BC break. In the long term it may actually be beneficial to change the default meaning of `is_causal` to represent lower_right causal masking.
The larger theme though is laid here: https://github.com/pytorch/pytorch/issues/110681. The thesis being that there is alot of innovation in SDPA revolving around the attention_bias being used. This is the first in hopefully a few more attention_biases that we would like to add. The next interesting one would be `sliding_window` which is used by the popular mistral model family.
Results from benchmarking, I improved the meff_attention perf hence the slightly decreased max perf.
```Shell
+---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+
| Type | Speedup | batch_size | num_heads | q_seq_len | k_seq_len | embed_dim | dtype | head_dim |
+---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+
| Average | 1.2388050062214226 | | | | | | | |
| Max | 1.831672915579016 | 128 | 32 | 1024 | 2048 | 2048 | torch.bfloat16 | 64 |
| Min | 0.9430534166730135 | 1 | 16 | 256 | 416 | 2048 | torch.bfloat16 | 128 |
+---------+--------------------+------------+-----------+-----------+-----------+-----------+----------------+----------+
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114823
Approved by: https://github.com/cpuhrsch
Continuation of #112185, following the design in this [doc](https://docs.google.com/document/d/1ipSxcTzEMMOAPvxP-YJlD5JBZZmIGgh8Q34ixtOUCRo).
Summary:
* Introduce `SubclassSymbolicPolicy` containing separate dynamic dim / constraint policies for the outer and inner tensors
* Expand the automatic dynamic algorithm to recurse into inner tensors and produce one of these for a subclass instance
* Maintain legacy behavior for subclasses by recursively calling `mark_dynamic()` on inner tensors *of the same dim as outer* when `mark_dynamic(outer, ...)` is called
* Addresses this: 6a86cf00ad/torch/_dynamo/variables/builder.py (L1750)
* Add `outer_size` and `outer_stride` arguments to `__tensor_unflatten__()` so that you can find out what symbols were allocated for the outer size / stride (you are expected to return a tensor that compares equal to the outer symbols)
* Signatures now:
```python
# attrs is a list of inner tensor attributes on x; inner_tensor = getattr(x, attr)
# ctx is anything useful for rebuilding the class we want to guard on
attrs, ctx = x.__tensor_flatten__()
...
# inner_tensors is a dict of {attr -> tensor}
# ctx is taken unmodified from flattening and (eventually) guarded on
# outer_size is the expected size of the output; possibly symbolic
# outer_stride is the expected strides of the output; possibly symbolic
y = MySubclass.__tensor_unflatten__(inner_tensors, ctx, outer_size, outer_stride)
# at the __tensor_unflatten__() call-site in PT2, we assert y.shape == outer_size and y.stride() == outer_stride
# the assert simplifies symbols when there are relationships between outer and inner symbols
```
* Size info needed for `NestedTensor` at least, stride info needed for `DTensor` at least
* Punting on `outer_storage_offset` because storage_offset handling is horribly broken in PT2 right now
* ~~Add new `__tensor_mark_dynamic__()` to allow overriding the behavior of mark_dynamic on a per-subclass basis~~ (booted to future work)
* ~~Add guards for tensor subclasses by calling `__tensor_flatten__()` in the guard to test equality on `ctx`~~
* Now handled in #114469
* Next PR: add TENSOR_MATCH guards on inner tensors
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114311
Approved by: https://github.com/ezyang, https://github.com/drisspg, https://github.com/voznesenskym, https://github.com/bdhirsh
Summary:
Rename _device_mesh.py to device_mesh.py, update all callsites, adds documentation.
Original diff reverted: D51629761
Original PR reverted: https://github.com/pytorch/pytorch/pull/114991
It was failing because failing a public module binding tests in MacOS, and this is due to the change in import order for torch/distributed/fsdp/_common_utils.py. Since this original import would still work, we remove the changes in this file.
Test Plan: CI.
Differential Revision: D51825114
Pull Request resolved: https://github.com/pytorch/pytorch/pull/115099
Approved by: https://github.com/wanchaol, https://github.com/fegin
This PR rewrites Tensor Parallel implementation. Tensor Parallel APIs
supposed to be a very thin-wrapper to DTensor APIs, but the current
implementation got too messy and buggy. It's really hard to debug what
went wrong when using it. It's crucially important for advanced users or
developers to understand the API and its implementation easily without
going through all different types of functions and utils, so that
they could trust what happen under the hood.
In particular this PR:
* Make ParallelStyle to be a real contract API for parallelize_module to
take, each concrete ParallelStyle only needs to implement `apply` to
apply the sharding to nn.Module, remove all non-necessary fields. This
also enable easier ParallelStyle authoring going forward.
* Keep the ColwiseParallel and RowwiseParallel public interface, but
refactor them in a way that makes the parameter sharding, inputs and
outputs handling lives within the style itself, so that it's easy to
understand how Linear/Embedding layers are sharded and how the inputs/outputs
transformations are performed.
* remove all those private _prepare_input/_prepare_output_fn fields for
both ColwiseParallel/RowwiseParallel. Since we throw deprecation
messages in nightly for a while and TP is on prototype release, the
fields are also private, it should be safe to remove them
* Refactor the recently landed PrepareModuleInput/Output style, change
output_layouts to desired_input/output_layouts, group
the function inside the style itself, no default arguments for these
two styles and user need to specify them to think about the sharding
layouts. Fixed bugs about not handling
`use_local_output` flag.
* Make default arguments be None instead of Placement object, this is
standard python practice to not have custom object instance as default
argument
* Remove all dead APIs (i.e. PairwiseParallel and SequenceParallel
style, all prepare input/output functions) as we throw deprecation
msgs for a while, and in the progress of removing all of them from the tests.
* throw deprecation warning for `tp_mesh_dim` as we recomemnd use device
mesh slice/indexing instead of manually specify mesh dim
* Rewrite all documentations for every ParallelStyle and make the
documentation more clear about what each style is doing
TODOs:
* Rewrite TP tests to adjust for the changes we have in this PR
* add more tests to guard the bug fixes
Differential Revision: [D51761183](https://our.internmc.facebook.com/intern/diff/D51761183)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114732
Approved by: https://github.com/wz337, https://github.com/fduwjj
Summary:
This is a util for numeric suite in pt2 export so that we can build
a more streamlined UX for numerical debugging in quant + executorch stack
Test Plan:
python test/test_quantization.py TestGenerateNumericDebugHandle
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114315
Approved by: https://github.com/zhxchen17
Summary:
The primary problem we are setting out to solve here is fake tensor freshness. Before this PR, fake tensors after dynamo represented fake tensors *at the end* of trace, so subsequent retraces like aot_autograd would start off with fake tensors in the wrong (end result) state, rather than their expected fresh state. The solution here is to start a fresh fake mode, and re-fakify the tensors. The nuance comes from ensuring that symbols are uniformly created for the symbolic sizes and strides of the tensor.
This PR is the result of *a lot* of back and forth with ezyang and eellison. Initially, the first pass at this was not super different from what we have in the PR - the broad strokes were the same:
1) We cache source->symbol in shape_env
2) We pass policy objects around, stored at dynamo fakificaiton time, and reused for later fakification
3) We create a new fake mode for backends
(from https://github.com/pytorch/pytorch/pull/113605/files)
This is ugly, and has some layering violations. We detoured our decision making through a few other alternatives. Immutable/mutable fake tensor mode was the most interesting alternative, https://github.com/pytorch/pytorch/pull/113653, and was struck down on concerns of complexity in fake mode combined with it not covering all edge cases. We also detoured on what to do about tensor memoization returning back potentially different tensors than requested, and if that was an anti pattern (it is) we want to hack in with the symbol cache (we don't).
We went back to the drawing board here, but with a few concessions:
1) the cache for source->symbol must live outside of shape_env, for both lifecycle, and layering reasons
2) A good amount of work needs to be done to pipe policy around fake_mode and meta_utils correctly, to cover all the cases (ezyang did this)
cc penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 aakhundov kadeng
imported-using-ghimport
Test Plan: Imported from OSS
Reviewed By: huydhn, Chillee
Differential Revision: D51566250
Pulled By: voznesenskym
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114526
Approved by: https://github.com/Chillee, https://github.com/huydhn
If you copy and paste the env var in the docs:
```console
TORCHDYNAMO_REPRO_AFTER=“aot”
```
it leads to this error:
```python
@functools.wraps(unconfigured_compiler_fn)
def debug_wrapper(gm, example_inputs, **kwargs):
compiler_fn = functools.partial(unconfigured_compiler_fn, **kwargs)
> assert config.repro_after in ("dynamo", "aot", None)
E torch._dynamo.exc.BackendCompilerFailed: backend='inductor' raised:
E AssertionError:
```
because `config.repro_after` is being `'“aot”'` but not `'aot'`.
---
It would've saved a few minutes of my time 😄
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114530
Approved by: https://github.com/Chillee
