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
