**Update:** Made refactor of the original PR. See the original description below, but here I'll describe the updates:
(1) TLS changes in `TorchDispatchModeTLS.h/cpp`.
I added a `TorchDispatchModeKey` enum, that (for now) just contains PROXY and FAKE. The ModeTLS used to just contain a `std::vector<std::shared_ptr<c10::SafePyObject>>` corresponding to the mode stack. It now **also** contains a separate array of "infra modes", indexed by mode key (PROXY and FAKE, with a new addition, FUNCTIONAL, coming later in the stack).
`TorchDispatchModeTLS::push_onto_stack` and `TorchDispatchModeTLS::pop_stack` are now a bit more complicated. Pushing accepts an optional mode_key, which if set, tells us to add the given mode directly to our "infra_modes" array. Popping will first check the "user mode" stack, before trying to pop anything from the infra mode stack. It also optionally returns the mode key of the mode we popped if there was one - that way if we push that same mode back onto the TLS later, we know where it goes.
`TorchDispatchModeTLS::dispatch_mode_enabled()` now accepts an optional `skip_infra_modes` param, so you can separately query if there are "any modes at all", or if there are "any user modes".
`TorchDispatchModeTLS::get/set/unset_mode()` all take in a mode key, and get/set/unset the mode at that particular mode key (meaning they are only meant to be used for infra modes).
There were also some mild codegen changes to support the new enum
(2) `fake_tensor.py/proxy_tensor.py/_python_dispatch.py`
The way I tell the infra that certain subclasses/modes are "infra" is through the enum: I gave `FakeTensor` and `FakeTensorMode` a `self._mode_key = torch._C.TorchDispatchModeKey.FAKE`. `TorchDispatchMode.__enter/exit__()` (in `_python_dispatch.py` now check if the current mode has a mode key, and if so they plumb it into any `push_onto_stack()` calls (which eventually instructs `TorchDispatchModeTLS` where to put the mode). Same thing for `ProxyTorchDispatchMode`.
I also had to change both of these mode's enter/exit, to handle the fact that there can no longer be multiple proxy/fake modes on the mode stack at once. I updated them both to have a `self.enter_stack: List[Optional[TorchDispatchMode]]` - whenever we push a given mode in `__enter__`, we remove the current ambient fake/proxy mode from the mode stack, and save it in `enter_stack`, so that on exit we can reset the state properly.
(2) dispatching logic in `python_arg_parser.cpp`
This is where the core dispatching logic changes are. I added two helpers, `dispatch_on_subclass()` and `dispatch_on_mode()`. The overall dispatching order is now:
```
(a) dispatch_on_mode() # try user modes first (where the mode stack automatically considers infra modes last)
(b) dispatch_on_subclass() # try user subclasses next (skipping infra subclasses)
(c) dispatch_on_subclass() # try infra subclasses next (skipping user subclasses)
```
Note that we still want "user subclasses" to run before "infra modes". As Ed helped me realize, this will work today: If proxy/fake modes in step 1, they'll return NotImplemented if they see a user subclass, allowing us to redispatch to the user subclass.
How do (b) and (c) distinguish between user and infra subclasses? Infra subclasses (FakeTensor, and later FunctionalTensor) are required to have a `_mode_key` hidden on the subclass - so we filter via arguments that do/don't have the _mode_key.
(3) I also changed `DoubleTensor` to `TwoTensor` to minimize confusion (@albanD pointed out that DoubleTensor would be easily confused with `torch.FloatTensor` and friends).
----- original description below -----
The main purpose of this PR is to fix the "ordering problem" between torch_dispatch modes, where we want to ensure that our Fake and Proxy dispatch modes always run **after** any dispatch modes created by the user, regardless of where they are in the stack. See this doc for more details: https://docs.google.com/document/d/1COQ291nOZvtFnzGTQMJqoYZ3sttEYFw_7HbfSyL8gcA/edit
Full set of changes below. I ended up including a few semi-related changes in this PR that I documented - but if folks would rather I separate them out, happy to try to do that.
**(1) Add dedicated TLS slots for FakeTensorMode and ProxyTensorMode**
This is the main component of this PR. There are two new slots, `TorchDispatchModeTLS.fake_mode_` and `TorchDispatchModeTLS.proxy_mode_`, which correspond to a single "global" fake and proxy mode. There is now an invariant that `torchDispatchModeState.stack_` can never contain either of these modes.
I also added a `TorchDispatchModeTLS::maybe_highest_mode()` helper that consults the `stack_` as well as both the proxy and fake slots, and returns the highest priority mode - this is because there are a few places in the codebase where we legitimately want to get the highest priority mode, *including* fake or proxy, if one is set.
This also made the implementations of the existing `disable_proxy_modes_tracing()` and `get_innermost_proxy_mode()` marginally simpler.
**(2) Updated the dispatching logic in handle_torch_function_no_python_arg_parser()**
This is the function that actually figures out which torch_dispatch implementation to call, given the current mode stack and tensor subclass inputs. This function got marginally more complicated as part of the refactor: First we inspect the mode stack and any non-fake subclass inputs. Then we check for the proxy mode slot. Then we check for the Fake mode slot, before finally checking for any fake subclass inputs.
**(3) new python `_get_fake_tensor_mode()` and `_get_proxy_tensor_mode()` API's**
Before, if you wanted to see if proxy or fake modes were active in python, you would have to consult the mode stack. Since these two modes are no longer part of the actual mode stack, I added two new API's to directly check if either proxy or fake modes are active.
**(4) Allow traceable tensor subclasses to access storages from python**
This is convenient later in the stack, where AOTAutograd needs to detect aliasing of inputs and outputs, where those inputs and outputs might be tensor subclasses. Previously, `x.untyped_storage()` would raise an error if `x` was a subclass. In this PR, I tried to relax this constraint as little as possible: `THPVariable_storage()` will only try to return a storage to python if the tensor subclass that you are passing in is "traceable"
**(5) Fixed subclass fakeification**
@wanchaol recently added support to be able to fakeify tensor subclasses. That fakeification logic works in most cases, but there is one case it doesn't handle: autograd metadata. In particular, since autograd sees our tensor subclasses and not their desugared tensors, we need to make sure that our fakeified subclass has the same autograd metadata as the original subclass. I updated `meta_utils.py` to make sure that the autograd metadata is correct.
**(6) make tensor subclasses resizeable**
Previously we didn't allow tensor subclasses to be resizeable. I ran into an issue where fakeifying a tensor subclass occasionally requires swapping out its storage, which can involve resizing the tensor. Mechanically, this required updating `at::for_blob()` to expose a way to request that the tensor that you create has resizeable storage, and then using this new API in `_make_wrapper_tensor()`.
**(7) Added a basic DoubleTensor subclass for testing**
I use this subclass more later in this stack in my AOTAutograd tests - but it serves as a simple subclass example to test the dispatch ordering in this PR.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/104482
Approved by: https://github.com/ezyang
ghstack dependencies: #107415
Proposal of two float8 variants - e5m2 and e4m3 - based on https://arxiv.org/pdf/2209.05433.pdf
Hide all Float8 operator implementations behind `#if !defined(C10_MOBILE)` guard to keep Android build size almost unchanged
TODO:
- Refactor duplicated code
- Cleanup unbalanced pragma pop in dtype utils
- Add native implementation on the CUDA size
Co-authored-by: Nikita Shulga <nshulga@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/104242
Approved by: https://github.com/albanD
Proposal of two float8 variants - e5m2 and e4m3 - based on https://arxiv.org/pdf/2209.05433.pdf
Hide all Float8 operator implementations behind `#if !defined(C10_MOBILE)` guard to keep Android build size almost unchanged
TODO:
- Refactor duplicated code
- Cleanup unbalanced pragma pop in dtype utils
- Add native implementation on the CUDA size
Co-authored-by: Nikita Shulga <nshulga@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/104242
Approved by: https://github.com/albanD
I want to use torchgen to generate code, and my yaml file format is the same as `native_functions.yaml`.
I will use the PrivateUse1, but in my yaml file, I don't want to show PrivateUse1 to the user.
So I want to achieve the following result(e.g. my device is `YPU`):
```
>>>from torchgen.model import DispatchKey
>>>str(DispatchKey.PrivateUse1)
"YPU"
>>>DispatchKey.parse("YPU")
DispatchKey.PrivateUse1
```
I also thought that not everyone would need this feature, so I add a new func to handle this scenario.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99406
Approved by: https://github.com/ezyang
I applied some flake8 fixes and enabled checking for them in the linter. I also enabled some checks for my previous comprehensions PR.
This is a follow up to #94323 where I enable the flake8 checkers for the fixes I made and fix a few more of them.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/94601
Approved by: https://github.com/ezyang
The idea is to add a custom handler to Functionalize key in Python
dispatcher that runs the functionalized version along side a non
functionalized version, and checks that their outputs agree in the
end. (Technically, for metadata mutation we should also check the
inputs, but for now we're relying on those functions returning self.)
I turned this on for test_functionalize.py (new TestCrossRefFunctionalize)
and found a bunch of failures that look legit.
This probably doesn't interact that nicely if you're also tracing at
the same time, probably need more special logic for that (directly,
just disabling tracing for when we create the nested fake tensor mode,
but IDK if there's a more principled way to organize this.)
There are some misc fixups which I can split if people really want.
- xfail_inherited_tests moved to test common_utils
- Bindings for _dispatch_tls_set_dispatch_key_included,
_dispatch_tls_is_dispatch_key_included and _functionalization_reapply_views_tls
- Type stubs for _enable_functionalization, _disable_functionalization
- all_known_overloads utility to let you iterate over all OpOverloads
in all namespaces. Iterator support on all torch._ops objects to let
you iterate over their members.
- suspend_functionalization lets you temporarily disable functionalization mode
in a context
- check_metadata_matches for easily comparing outputs of functions and see
if they match (TODO: there are a few copies of this logic, consolidate!)
- _fmt for easily printing the metadata of a tensor without its data
- _uncache_dispatch for removing a particular dispatch key from the cache,
so that we force it to regenerate
- check_significant_strides new kwarg only_cuda to let you also do stride
test even when inputs are not CUDA
- Functionalize in torch._C.DispatchKey
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/89498
Approved by: https://github.com/malfet
This PR teaches PyDispatcher and PyOperator about functorch transforms.
It is important that PyDispatcher/PyOperator dispatch with functorch
transforms, because this is our plan for higher-order operators
(operators that accept functions as arguments). Examples of these
include:
- functorch transforms over the existing cond operator (control flow)
- autograd.Function support for functorch (which I am working towards),
- AOTDispatcher (should be a higher order operator)
Concretely, the problem with teaching PyDispatcher/PyOperator about
functorch is that the stack-based dispatching logic (DynamicLayerStack)
is hidden inside the fallbacks for two dispatch keys
(DynamicLayer{Front, Back}). PyDispatcher doesn't know about C++ boxed
fallbacks, our plan on record for that is that we need to reimplement
all of them in Python (but can call helper functions in C++ to make our
lives easier).
Instead of exposing all of what DynamicLayer{Front, Back} do to python,
this PR takes the approach of re-implementing part of the stack-based
dispatching in Python. The motivation is that this is more sane and
follows what the "ideal" implementation of functorch would have been:
- each transform should be a "mode"
- there should be no TLS dispatch key set hackery. functorch needs to do
this hackery today to re-use VariableType implementations.
This PR:
- exposes the DynamicLayerStack to Python
- The DynamicLayerStack is a stack of Interpreters.
These get exposed to Python as well.
- Interpreters can run operations (Interpreter.process) or lower them to
the next interpreter in the stack (Interpreter.lower)
- To use a PyOperator with functorch transforms, a developer needs to
register a rule for each transform (vmap, grad, jvp, ...).
- The PyOperator API is NOT user-facing. Things like autograd.Function
support for functorch will end up going through the autograd.Function
API.
Question for reviewers:
- Does this design make sense?
- I'm trying to split up the "functorch support for autograd.Function"
work into logical pieces. Would it be better if I didn't? (the full
thing is a bit long - 1000-2000 LOC).
Test Plan:
- new tests that construct PyOperator and compose them with functorch
transforms
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88785
Approved by: https://github.com/samdow, https://github.com/soulitzer
Summary:
Sometimes we want to extend an existing custom namespace library, instead of creating a new one,
but we don't have a namespace config right now, so we hardcode some custom libraries defined
in pytorch today, i.e. quantized and quantized_decomposed
Test Plan:
ci
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88229
Approved by: https://github.com/ezyang
Partially fixes: #66328
This PR:
- adds support for `ITensorList` to the dispatcher for:
- computing the dispatch key
- boxing and unboxing `ITensorList`
- modified the codegen for structured kernels:
- codegen APIs use `ITensorList` instead of `ArrayRef<Tensor>`
**Changes summary:**
- Signature changes due to the different APIs:
- dispatcher API (e.g. `BatchingRegistrations.cpp`)
- C++ API (e.g. `TensorShape.cpp`)
- Miscelaneous functions used by codegen'd functions (e.g. `FunctionalTensorWrapper.*`)
- Dispatcher changes for handling `ITensorList` correctly (e.g. `DispatchKeyExtractor.h`)
- Signature changes of `at::cat` due to the need of `const` inside `TensorBody.h`
- Forward declarations of `ITensorList` (e.g. `MethodOperators.h`)
- Codegen changes, special casing structured kernels (e.g. `gen.py`)
**Short description of structured kernels special casing:**
I introduced, mainly, 5 types of changes to the codegen for generating code depending on
whether the kernel is structured or not:
1. Added a `structured_type_override` flag to the `argument_type` function definition of
the affected APIs (mainly the dispatcher and C++ APIs).
- `api/cpp.py`, `api/dispatcher.py`, `api/native.py`
2. Added a `structured_type_override` member to the signature
classes (e.g. `CppSignature`), since `FunctionSchema` doesn't really know whether the
function is structured or not
- `api/types.py`
3. Added a `part_of_structured_group` to `NativeFunction` class, which is just a
convenient function to forward to `structured_type_override` wherever needed
- `model.py`
4. Appropriately changed the rest of the codegen, whenever it used either the signature
classes or the `arguments` function directly
5. Added a check for `const ITensorList&` type wherever there was a check for `TensorList`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/73350
Approved by: https://github.com/bdhirsh
Instead of calling into the Python dispatcher for EVERY dispatcher
call, we now have a two step process. First, we
getattr(op: OpOverload, dispatch_key) to "load" the handler for the
function. This can either be a conventional function (in which
case we will call it, in the same way the old Python dispatcher
worked), or it can be a DispatchKey, in which case we will directly
call that DispatchKey in C++, bypassing marshalling between Python
and C++ entirely. OpOverload.__getattr__ is carefully written so
that it will cache the
A further optimization would be to define __slots__ on OpOverload,
and ensuring that the DispatchKey strings are interned.
The resulting Python dispatcher is less flexible: after the first
lookup, the handler is cached and we won't recompute it. Furthermore,
by default, dispatches will not go into Python, and so you won't
get stack frames for the Python dispatcher by default. But we get
a huge performance improvement: on the following microbenchmark
we go from 2.5s to 1.9s.
```
import time
import torch
from functorch import make_fx
def f(x):
for i in range(1000):
x = x * x
return x
begin = time.time()
res = make_fx(f, tracing_mode="symbolic")(torch.randn(10, 20))
print(time.time()-begin)
```
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/85133
Approved by: https://github.com/wconstab
Something people found confusing was that whether or not a native::
signature would get SymInt or not in its type was based on the dispatch
key. This changes it so that SymInt or not in type is based on whether
or not you have _symint in the name of the kernel or not. This means
that even when we make operators support SymInt, you no longer have to
go and update all the preexisting definitions; instead, you now
selectively write _symint to opt individual kernels into SymInt support.
I then go and update a bunch of kernels that don't have proper SymInt
support to make use of this convention. There is some hacking around
for view generation code.
I also add support for external backends to specify 'symint' operators, for which we generate SymInt signatures instead of regular signatures.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Differential Revision: [D39310060](https://our.internmc.facebook.com/intern/diff/D39310060)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/84579
Approved by: https://github.com/wconstab
Also Back out "Revert D39075159: [acc_tensor] Use SymIntArrayRef for overloaded empty.memory_format's signature"
Original commit changeset: dab4a9dba4fa
Original commit changeset: dcaf16c037a9
Original Phabricator Diff: D38984222
Original Phabricator Diff: D39075159
Also update Metal registrations for C++ registration changes.
Also update NNPI registration to account for tightened schema checking
Differential Revision: [D39084762](https://our.internmc.facebook.com/intern/diff/D39084762/)
**NOTE FOR REVIEWERS**: This PR has internal Facebook specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D39084762/)!
Pull Request resolved: https://github.com/pytorch/pytorch/pull/84173
Approved by: https://github.com/Krovatkin
Previously, we introduced new SymInt overloads for every function we wanted. This led to a lot of boilerplate, and also a lot of confusion about how the overloads needed to be implemented.
This PR takes a simpler but more risky approach: just take the original function and changes its ints to SymInts.
This is BC-breaking in the following ways:
* The C++ API for registering implementations for aten operators will change from int64_t to SymInt whenever you make this change. Code generated registrations in PyTorch do not change as codegen handles the translation automatically, but manual registrations will need to follow the change. Typically, if you now accept a SymInt where you previously only took int64_t, you have to convert it back manually. This will definitely break XLA, see companion PR https://github.com/pytorch/xla/pull/3914 Note that not all dispatch keys get the automatic translation; all the composite keys and Meta keys are modified to take SymInt directly (because they should handle them directly), and so there are adjustments for this.
This is not BC-breaking in the following ways:
* The user facing C++ API remains compatible. Even if a function changes from int to SymInt, the default C++ binding still takes only ints. (e.g., at::empty(IntArrayRef, ...). To call with SymInts, you must call at::empty_symint instead. This involved adding two more signatures to CppSignatureGroup; in many cases I refactored code to iterate over all signatures in the group instead of hard-coding the two that previously existed.
* This is TorchScript compatible; internally we treat SymInts as ints so there is no change to what happens at runtime in TorchScript. In particular, it's OK to reference an empty schema by its old type (using int types), as long as you're not doing string equality (which you shouldn't be), these parse to the same underyling type.
Structure of the PR:
* The general strategy of this PR is that, even when you write `SymInt` inside `native_functions.yaml`, sometimes, we will treat it *as if* it were an `int`. This idea pervades the codegen changes, where we have a translation from SymInt to c10::SymInt or int64_t, and this is controlled by a symint kwarg which I added and then audited all call sites to decide which I wanted. Here are some of the major places where we pick one or the other:
* The C++ FunctionSchema representation represents `SymInt` as `int`. There are a few places we do need to know that we actually have a SymInt and we consult `real_type()` to get the real type in this case. In particular:
* When we do schema validation of C++ operator registration, we must compare against true schema (as the C++ API will provide `c10::SymInt`, and this will only be accepted if the schema is `SymInt`. This is handled with cloneWithRealTypes before we check for schema differences.
* In `toIValue` argument parsing, we parse against the true schema value. For backwards compatibility reasons, I do still accept ints in many places where Layout/SymInt/etc were expected. (Well, accepting int where SymInt is expected is not BC, it's just the right logic!)
* In particular, because SymInt never shows up as type() in FunctionSchema, this means that we no longer need a dedicated Tag::SymInt. This is good, because SymInts never show up in mobile anyway.
* Changes to functorch/aten are mostly about tracking changes to the C++ API registration convention. Additionally, since SymInt overloads no longer exist, registrations for SymInt implementations are deleted. In many cases, the old implementations did not properly support SymInts; I did not add any new functionality with this PR, but I did try to annotate with TODOs where this is work to do. Finally, because the signature of `native::` API changed from int to SymInt, I need to find alternative APIs for people who were directly calling these functions to call. Typically, I insert a new dispatch call when perf doesn't matter, or use `at::compositeexplicitautograd` namespace to handle other caes.
* The change to `make_boxed_from_unboxed_functor.h` is so that we accept a plain IntList IValue anywhere a SymIntList is expected; these are read-only arguments so covariant typing is OK.
* I change how unboxing logic works slightly. Previously, we interpret the C++ type for Layout/etc directly as IntType JIT type, which works well because the incoming IValue is tagged as an integer. Now, we interpret the C++ type for Layout as its true type, e.g., LayoutType (change to `jit_type.h`), but then we accept an int IValue for it anyway. This makes it symmetric with SymInt, where we interpret the C++ type as SymIntType, and then accept SymInt and int IValues for it.
* I renamed the `empty.names` overload to `empty_names` to make it less confusing (I kept mixing it up with the real empty overload)
* I deleted the `empty.SymInt` overload, which ended up killing a pile of functions. (This was originally a separate PR but the profiler expect test was giving me grief so I folded it in.)
* I deleted the LazyDynamicOpsTest tests. These were failing after these changes, and I couldn't figure out why they used to be passing: they make use of `narrow_copy` which didn't actually support SymInts; they were immediately converted to ints.
* I bashed LTC into working. The patches made here are not the end of the story. The big problem is that SymInt translates into Value, but what if you have a list of SymInt? This cannot be conveniently represented in the IR today, since variadic Values are not supported. To work around this, I translate SymInt[] into plain int[] (this is fine for tests because LTC dynamic shapes never actually worked); but this will need to be fixed for proper LTC SymInt support. The LTC codegen also looked somewhat questionable; I added comments based on my code reading.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/83628
Approved by: https://github.com/albanD, https://github.com/bdhirsh
Previously, we introduced new SymInt overloads for every function we wanted. This led to a lot of boilerplate, and also a lot of confusion about how the overloads needed to be implemented.
This PR takes a simpler but more risky approach: just take the original function and changes its ints to SymInts.
This is BC-breaking in the following ways:
* The C++ API for registering implementations for aten operators will change from int64_t to SymInt whenever you make this change. Code generated registrations in PyTorch do not change as codegen handles the translation automatically, but manual registrations will need to follow the change. Typically, if you now accept a SymInt where you previously only took int64_t, you have to convert it back manually. This will definitely break XLA, see companion PR https://github.com/pytorch/xla/pull/3914 Note that not all dispatch keys get the automatic translation; all the composite keys and Meta keys are modified to take SymInt directly (because they should handle them directly), and so there are adjustments for this.
This is not BC-breaking in the following ways:
* The user facing C++ API remains compatible. Even if a function changes from int to SymInt, the default C++ binding still takes only ints. (e.g., at::empty(IntArrayRef, ...). To call with SymInts, you must call at::empty_symint instead. This involved adding two more signatures to CppSignatureGroup; in many cases I refactored code to iterate over all signatures in the group instead of hard-coding the two that previously existed.
* This is TorchScript compatible; internally we treat SymInts as ints so there is no change to what happens at runtime in TorchScript. In particular, it's OK to reference an empty schema by its old type (using int types), as long as you're not doing string equality (which you shouldn't be), these parse to the same underyling type.
Structure of the PR:
* The general strategy of this PR is that, even when you write `SymInt` inside `native_functions.yaml`, sometimes, we will treat it *as if* it were an `int`. This idea pervades the codegen changes, where we have a translation from SymInt to c10::SymInt or int64_t, and this is controlled by a symint kwarg which I added and then audited all call sites to decide which I wanted. Here are some of the major places where we pick one or the other:
* The C++ FunctionSchema representation represents `SymInt` as `int`. There are a few places we do need to know that we actually have a SymInt and we consult `real_type()` to get the real type in this case. In particular:
* When we do schema validation of C++ operator registration, we must compare against true schema (as the C++ API will provide `c10::SymInt`, and this will only be accepted if the schema is `SymInt`. This is handled with cloneWithRealTypes before we check for schema differences.
* In `toIValue` argument parsing, we parse against the true schema value. For backwards compatibility reasons, I do still accept ints in many places where Layout/SymInt/etc were expected. (Well, accepting int where SymInt is expected is not BC, it's just the right logic!)
* In particular, because SymInt never shows up as type() in FunctionSchema, this means that we no longer need a dedicated Tag::SymInt. This is good, because SymInts never show up in mobile anyway.
* Changes to functorch/aten are mostly about tracking changes to the C++ API registration convention. Additionally, since SymInt overloads no longer exist, registrations for SymInt implementations are deleted. In many cases, the old implementations did not properly support SymInts; I did not add any new functionality with this PR, but I did try to annotate with TODOs where this is work to do. Finally, because the signature of `native::` API changed from int to SymInt, I need to find alternative APIs for people who were directly calling these functions to call. Typically, I insert a new dispatch call when perf doesn't matter, or use `at::compositeexplicitautograd` namespace to handle other caes.
* The change to `make_boxed_from_unboxed_functor.h` is so that we accept a plain IntList IValue anywhere a SymIntList is expected; these are read-only arguments so covariant typing is OK.
* I change how unboxing logic works slightly. Previously, we interpret the C++ type for Layout/etc directly as IntType JIT type, which works well because the incoming IValue is tagged as an integer. Now, we interpret the C++ type for Layout as its true type, e.g., LayoutType (change to `jit_type.h`), but then we accept an int IValue for it anyway. This makes it symmetric with SymInt, where we interpret the C++ type as SymIntType, and then accept SymInt and int IValues for it.
* I renamed the `empty.names` overload to `empty_names` to make it less confusing (I kept mixing it up with the real empty overload)
* I deleted the `empty.SymInt` overload, which ended up killing a pile of functions. (This was originally a separate PR but the profiler expect test was giving me grief so I folded it in.)
* I deleted the LazyDynamicOpsTest tests. These were failing after these changes, and I couldn't figure out why they used to be passing: they make use of `narrow_copy` which didn't actually support SymInts; they were immediately converted to ints.
* I bashed LTC into working. The patches made here are not the end of the story. The big problem is that SymInt translates into Value, but what if you have a list of SymInt? This cannot be conveniently represented in the IR today, since variadic Values are not supported. To work around this, I translate SymInt[] into plain int[] (this is fine for tests because LTC dynamic shapes never actually worked); but this will need to be fixed for proper LTC SymInt support. The LTC codegen also looked somewhat questionable; I added comments based on my code reading.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/83628
Approved by: https://github.com/albanD, https://github.com/bdhirsh
Add is_symint_like, by way of is_base_ty_like which generalizes
the pattern for is_tensor_like and is_generator_like. Now that
we can query if a signature contains a SymInt, we can enforce that
you must name the overload with SymInt if the signature contains
SymInt.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/83668
Approved by: https://github.com/bdhirsh, https://github.com/larryliu0820
- nondeterministic_seeded was not applied to enough functions. I added
some heuristics to codegen for identifying functions that are likely
to be random and added a bunch of these tags to functions. Not sure
I got all of them.
- Don't constant propagate through nondeterministic functions in FX
tracing.
It would be better to do some testing for the tag but this would be quite an effort.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/83650
Approved by: https://github.com/bdhirsh, https://github.com/eellison
Extending the current regex in `model.py` to support annotation alias set. See issue #83214.
Ideally we should have a full fledged lexer similar to `schema_type_parser.cpp`, since regex can be more and more difficult to read if we add more support to it.
Adding this to unblock this issue for now.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/83501
Approved by: https://github.com/SherlockNoMad
Summary:
Previously we don't generate out variant (both schema and kernel) for an operator with functional variant only. This adds support for that and adds test.
## Changes on `native_function_generation.py`
We are generating out variant for all functional variants if possible. This PR introduces a lot of newly generated out variants and `native_functions.yaml` needs to incorporate the changes by adding `autogen` keywords.
The logic for determining what operators we should generate an out variant for is the following:
1. No existing out variant for this `NativeFunction`
2. Contains an existing in place, mutable or functional variant
3. Contains at least 1 tensor like return(s)
For operators matching the first two conditions but failing the third, I listed them in `FUNCTIONAL_OPS_THAT_CANNOT_GET_AN_OUT_VARIANT`.
## Special handling
The following operators satisfy all 3 criteria above but we chose to not autogen them, with some reasons.
* `mkldnn_adaptive_avg_pool2d`, the generated out variant `mkldnn_adaptive_avg_pool2d.out` is colliding with the `mkldnn_adaptive_avg_pool2d_out` kernel in `adaptive_avg_pool2d.out` operator. I manually created `mkldnn_adaptive_avg_pool2d.out` and renamed `mkldnn_adaptive_avg_pool2d_out` to `mkldnn_adaptive_avg_pool2d_out_stub`.
* `min`, `max` and `mean`. There already exist `min.out`, `max.out` and `mean.out` but they are having different semantics with the functional ones. I manually created `min.unary_out`, `max.unary_out` and `mean.dtype_out` to disambiguate.
## Autograd Changes
We introduced a logic to not match derivatives info in `derivatives.yaml` to out variant, since we are generating `NOT_IMPLEMENTED` kernels for those out variants anyway. The issue we are seeing with the original logic is that it doesn't handle `TensorOption` arguments really well. For example we have these two operators:
* `_to_copy(Tensor self, *, ScalarType? dtype=None, Layout? layout=None, Device? device=None, bool? pin_memory=None, bool non_blocking=False, MemoryFormat? memory_format=None) -> Tensor`
* `_to_copy.out(Tensor self, *, bool non_blocking=False, MemoryFormat? memory_format=None, Tensor(a!) out) -> Tensor(a!)`
If we uses `_to_copy` derivative info, there will be compilation error since `dtype` is missing from `_to_copy.out` signature.
Test Plan: Rely on unit test
Differential Revision: D37832342
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81437
Approved by: https://github.com/iseeyuan, https://github.com/bdhirsh
`derivatives.yaml` can now take a `dispatch` entry which registers per-autograd dispatch key derivatives such as
```
name: foo(Tensor self, Tensor y) -> Tensor
dispatch:
Default:
x: grad
y: grad.expand(y.sizes())
AutogradNestedTensor:
x: grad
y: NestedTensor_foo_backward(grad, y)
output_differentiabilty: [True]
```
However the old schema where there is no `dispatch` entry is still supported.
Would greatly appreciate feedback on *how to improve the testing strategy* of this PR, currently have registered an aten test op in TestOps.cpp with dummy gradients in derivatives.yaml and have some tests in test_autograd.py:TestAutogradMultipleDispatch but I am not sure whether these are sufficiently rigorous.
Additionally, this PR also makes the assumption that sets like [VIEW_FUNCTIONS](ff5399e528/tools/autograd/gen_inplace_or_view_type.py (L60)) are per-native-function and not per-native-function-and-dispatch-key. I'm not sure whether this is necessarily the case, *would there ever be a situation where (e.g. a nested_tensor op is a view op but the aten function is not or vice versa?)*
* __->__ #82801
Pull Request resolved: https://github.com/pytorch/pytorch/pull/82801
Approved by: https://github.com/bhosmer, https://github.com/albanD
Differential Revision: [D38480514](https://our.internmc.facebook.com/intern/diff/D38480514/)
torchgen schema parser does not support parsing function schemas using custom class so far. Here is an example:
```
quantized::conv2d_relu.new(Tensor qx, __torch__.torch.classes.quantized.Conv2dPackedParamsBase packed_weight, float output_scale, int output_zero_point) -> (Tensor)
```
This PR parse custom class name and encapsulate that into an object of CustomClassType. The only thing we need right now is just store the string class name and return that in `__str__` method.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/82925
Approved by: https://github.com/ezyang, https://github.com/bdhirsh
The functional variant of one of the `arange` overloads has a schema mismatch with the out variant. The functional one has `Scalar step`, but the corresponding out variant has `Scalar step=1`. This isn't allowed, so it had to be special-cased in the python codegen and manually bound. This adds the default `step` value to the functional overload and removes the special-casing.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81380
Approved by: https://github.com/ngimel
`resize_()` is annoying because it needs special casing for functionalization. It's technically an inplace-view op, but it can't really have a pure view variant, since calling resize_() might bust the old storage. I gave it an `inplace_view` tag so that stuff like `FakeTensor` that relies on tags will pick it up properly, which required jumping through some codegen hoops.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/82667
Approved by: https://github.com/eellison
Summary:
Some quantized operators needs `QuantizedCPU` backend, due to an issue in namespace checking, currently if we have two backends as well as a custom namespaces in native function, codegen will hit assertion error. This PR fixes this issue
The root cause is that codegen right now asserts that a native function should only have one namespace. The current behavior is that If a native function is not found in a `BackendIndex`, we will use default namespace for that backend, for fallback kernels. However that default namespace may not be listed in the yaml file and it should not be counted when checking if we have two different namespaces for that backend. In our error case, we have 2 `BackendIndex`, one for `QuantizedCPU` and one for `CPU`. The native function doesn't have a kernel in `QuantizedCPU` but we still use a default namespace (`at::native`) for it. Since we have a custom namespace for dispatch key `CPU`, we ran into the assertion error.
This PR changes the assertion criteria. We only error out if a namespace has two or more kernels and they have two or more different namespaces.
Test Plan: rely on newly added unit test
Differential Revision: D38101345
Pull Request resolved: https://github.com/pytorch/pytorch/pull/82133
Approved by: https://github.com/iseeyuan
This also makes them not decompose when we switch Python key.
Note that CompositeExplicitAutogradNonFunctional maybe be overly
conservative for some implementations (which actually call into
other functional ops), but for now I just uniformly apply this
everywhere to avoid errors.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/82251
Approved by: https://github.com/bdhirsh, https://github.com/zou3519
Once CompositeImplicitAutograd gets registered to Python key, this will
ensure that tensor subclasses can interpose on these functions directly
rather than getting decomposed. We prefer not decomposing as these
functions are functional, but their implementations use inplace
operations (and are thus more difficult to deal with, unless you use
functionalization.)
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/82238
Approved by: https://github.com/zou3519, https://github.com/bdhirsh
This should fix the last issue that @anijain2305 hit when running ResNet with TorchDynamo <> functionalization.
Today if you try to call an `OpOverloadPacket` from python with some arguments, we will use the types of those arguments to perform overload resolution. With some functional variants of ops, this can be ambiguous.
Today this affects just one op: `_fused_moving_avg_obs_fq_helper`, although it would potentially affect e.g. `native_batch_norm` in the future.
Example:
```
# There are technically two overloads:
# torch.ops.aten._fused_moving_avg_obs_fq_helper.default (returns 2 argument, mutates 4 of its inputs inplace)
# torch.ops.aten._fused_moving_avg_obs_fq_helper.functional (returns 6 argument, mutates none of its inputs)
# We pick the wrong one - no way to know that we should pick the functional one, just from the call site.
outs = torch.ops.aten._fused_moving_avg_obs_fq_helper(a, a, a, a, a, a, a, 1.0, 0, 1, 0)
# raises an error - tries to call the overload with only 2 returns
return _fused_moving_avg_obs_fq_helper_functional[5]
```
Specifically, functionalization will bake `_fused_moving_avg_obs_fq_helper.functional` into the graph, but when AOTAutograd tries to compile with TorchScript, it needs to remove the overload name (TS doesn't know how to parse overload names directly, so we need to remove the overload name and let it infer the right overload at runtime later- so it picks the wrong one).
The situation is pretty similar to inplace; `ops.aten.add` and `ops.aten.add_` represent two different `OverloadPacket` objects; they can't be overloads of the same op, because their schemas would be ambiguous - the alias annotations are different, but that isn't enough to disambiguate).
In this PR, I try to fix the situation in a pretty similar way to how we handle `inplace` in the data model: `inplace` ops get their own base operator name, but they are represented as a flag inside of `BaseOperatorName` in the data model.
Two other important changes that I made as part of this PR:
(1) Originally, there were ~100 different `*_functional` operators: e.g. we had operators named `resize.functional` and `zero.functional`. The `_functional` bit isn't actually necessary in most cases: it's only necessary for operators that **also** have a `SchemaKind.mutable` variant, where `_fused_moving_avg_obs_fq_helper` is the only op that fits that description today. So I removed the unnecessary notion of "functional" from those other ops. I also added a bunch of assertions to force this restriction.
I think that makes more sense in the long run, because it eliminates an unnecessary difference in the model. E.g. we don't have `add_.Tensor` and `add.Tensor_functional`. We just have `add_.Tensor` and `add.Tensor`.
(2) I noticed that we actually still weren't pairing up a bunch of `_foreach` operators correctly, because their input arguments were different (`self` vs. `tensors`). Since they're private API's, I went ahead and changed the argument names directly so they get matched up. Before this PR, we were generating a separate `_foreach_add` and `_foreach_add.functional` variant in a bunch of cases, that really did the same thing (but happened to have a different name for the first argument).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/80556
Approved by: https://github.com/ezyang, https://github.com/albanD
Due to implicit conversion shenanigans, having both IntArrayRef
and SymIntArrayRef overloads makes {} ambiguous. While we could
fix this by making a single unified type that accepts all the overloads
we want, an easier fix was to just push the SymIntArrayRef overload
to its own name.
Signed-off-by: Edward Z. Yang <ezyangfb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/79281
Approved by: https://github.com/suo
Summary:
Adding a feature to allow user to specify namespaces for operator and kernels.
# Feature
There's a feature request to allow DSL to:
1. take in an operator namespace other than `aten`.
2. take in a kernel that is in a different namespace than `at::native`.
For both features, we only allow user to have a single layer of namespace for the sake of simplicity. If user specify `custom::function` as kernel, the codegen will depend on `custom::native::function` where `native` is hardcoded.
# Proposal
For feature 1, add a `namespace` attribute to data class `NativeFunction`. The namespace will be extract out by matching pattern "::" on the `func` variable. For `NativeFunctionsGroup` there's an assumption that all variants (function, inplace, out) will have the same namespace. By default (if not specified) the namespace will be "aten".
For feature 2, add a `namespace` attribute to `BackendMetadata` class, similarly match pattern "::" on the kernel field. Remove the `cpp_namespace` field from `register_dispatch_key` data class. By default (if not specified) the namespace for a kernel would be "at::native".
Test Plan:
Example yaml entries:
```
- func: custom::gelu.out(Tensor self, *, str approximate='none', Tensor(a!) out) -> Tensor(a!)
structured: True
structured_inherits: TensorIteratorBase
device_check: NoCheck # TensorIterator
python_module: nn
dispatch:
CPU: custom::gelu_out_cpu
CUDA: custom::gelu_out_cuda
MPS: custom::gelu_out_mps
- func: custom::gelu_(Tensor(a!) self, *, str approximate='none') -> Tensor(a!)
structured_delegate: gelu.out
device_check: NoCheck # TensorIterator
python_module: nn
dispatch:
NestedTensorCPU, NestedTensorCUDA: custom::NestedTensor_gelu_
- func: custom::gelu(Tensor self, *, str approximate='none') -> Tensor
structured_delegate: gelu.out
device_check: NoCheck # TensorIterator
python_module: nn
dispatch:
MkldnnCPU: custom::mkldnn_gelu
QuantizedCPU: custom::gelu_quantized_cpu
NestedTensorCPU, NestedTensorCUDA: custom::NestedTensor_gelu
```
see generated code:
`RegisterCPU.cpp`:
```
TORCH_LIBRARY_IMPL(aten, CPU, m) {
...
}
TORCH_LIBRARY_IMPL(custom, CPU, m) {
m.impl("gelu", TORCH_FN(wrapper_gelu));
m.impl("gelu.out", TORCH_FN(wrapper_gelu_out_out));
m.impl("gelu_", TORCH_FN(wrapper_gelu_));
};
```
```
struct structured_gelu_out_cpu_inplace final : public custom::native::structured_gelu_out_cpu {
structured_gelu_out_cpu_inplace(Tensor& self) : outputs_{std::ref(self)} {}
void set_output_strided(
int64_t output_idx, IntArrayRef sizes, IntArrayRef strides,
TensorOptions options, DimnameList names
) override {
const auto& out = outputs_[output_idx].get();
check_inplace(out, sizes, options);
auto maybe_proxy = maybe_create_proxy(out, sizes, strides, options);
if (C10_UNLIKELY(maybe_proxy.has_value())) {
proxy_outputs_[output_idx] = c10::ExclusivelyOwned<Tensor>(std::move(maybe_proxy).value());
}
if (!names.empty()) {
namedinference::propagate_names(outputs_[output_idx], names);
}
// super must happen after, so that downstream can use maybe_get_output
// to retrieve the output
custom::native::structured_gelu_out_cpu::set_output_raw_strided(output_idx, sizes, strides, options, names);
}
void set_output_raw_strided(
int64_t output_idx, IntArrayRef sizes, IntArrayRef strides,
TensorOptions options, DimnameList names
) override {
const auto& out = outputs_[output_idx].get();
check_inplace(out, sizes, options);
if (!names.empty()) {
namedinference::propagate_names(outputs_[output_idx], names);
}
// super must happen after, so that downstream can use maybe_get_output
// to retrieve the output
custom::native::structured_gelu_out_cpu::set_output_raw_strided(output_idx, sizes, strides, options, names);
}
const Tensor& maybe_get_output(int64_t output_idx) override {
return proxy_outputs_[output_idx].has_value() ? **proxy_outputs_[output_idx] : outputs_[output_idx].get();
}
std::array<std::reference_wrapper<Tensor>, 1> outputs_;
std::array<c10::optional<c10::ExclusivelyOwned<Tensor>>, 1> proxy_outputs_;
};
```
`RegisterSchema.cpp`
```
TORCH_LIBRARY(aten, m) {
...
}
TORCH_LIBRARY(custom, m) {
m.def("gelu.out(Tensor self, *, str approximate='none', Tensor(a!) out) -> Tensor(a!)");
m.def("gelu_(Tensor(a!) self, *, str approximate='none') -> Tensor(a!)");
m.def("gelu(Tensor self, *, str approximate='none') -> Tensor");
};
```
Differential Revision: D36558459
Pull Request resolved: https://github.com/pytorch/pytorch/pull/78015
Approved by: https://github.com/bdhirsh
Add codegen infrastructure to generate IR nodes for non-native ops.
The proposed change is to add a `non_native` key to the `{backend}_native_functions.yaml` file that contains schema definitions similar to what is found in `native_functions.yaml`. e.g.
```
non_native:
...
- func: expand(Tensor input, int[] size, bool is_scalar_expand) -> Tensor
...
```
these definitions are parsed into a `LazyIrSchema` that can be used for generating IR nodes using `GenLazyIR`.
Fixes#74628
CC: @wconstab @desertfire @henrytwo
Pull Request resolved: https://github.com/pytorch/pytorch/pull/76535
Approved by: https://github.com/wconstab
