Fixes#115331.
This PR increases the number of valid GPU devices to 512 (from 64) in order to future-proof PyTorch for providers that offer [single nodes with a large device count](https://www.tensorwave.com/). Until now, `DeviceIndex` was an `int8_t`, thus multiple changes were necessary:
- `DeviceIndex` changed to `int16_t`. Updated consumers that assume it to be an `int8_t`.
- Updated bounds checking for `torch.device()` in the Python frontend. Right now, we allow funny things like `torch.device('cpu', 200).index == -56`, which is undefined behavior. I inserted some checks to only allow values between 0 and `c10::Device::MAX_NUM_DEVICES - 1`.
- Updated the `ArgumentInfo` struct as it hardcodes the device index as 8 bit field [^1]. Might be a breaking change, not sure if users rely on this.
- Introduced `c10::Device::MAX_NUM_DEVICES` as a replacement for the old `C10_COMPILE_TIME_MAX_GPUS`
[^1]: This field was unsigned, so I guess this has also been undef behavior the whole time? Our default device index is -1, so this always wrapped around to 255 when written to the `ArgumentInfo` struct. When I switched the `DeviceIndex` to `int16_t`, it actually stayed 255 after unpacking from `ArgumentInfo` again, as the `DeviceIndex` was now wide enough that it didn't wrap back to -1.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/119639
Approved by: https://github.com/cyyever, https://github.com/albanD
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
On Linux and Mac `int64_t` is an alias to either `long` (Linux) or `long long` (Mac)
Because of that, attempt to construct `c10::Scalar` from the other type will fail with `conversion from ‘long long int’ to ‘c10::Scalar’ is ambiguous`.
I.e. attempt to compile:
```cpp
int main() {
c10::Scalar s = 1L;
}
```
on MacOS failed with:
```
foo.cpp:3:15: error: conversion from 'long' to 'c10::Scalar' is ambiguous
c10::Scalar s = 1L;
^ ~~
/Users/nshulga/git/pytorch/pytorch/torch/include/c10/core/Scalar.h:59:7: note: candidate constructor
DEFINE_IMPLICIT_CTOR)
^
/Users/nshulga/git/pytorch/pytorch/torch/include/c10/core/Scalar.h:59:7: note: candidate constructor
/Users/nshulga/git/pytorch/pytorch/torch/include/c10/core/Scalar.h:59:7: note: candidate constructor
/Users/nshulga/git/pytorch/pytorch/torch/include/c10/core/Scalar.h:59:7: note: candidate constructor
/Users/nshulga/git/pytorch/pytorch/torch/include/c10/core/Scalar.h:59:7: note: candidate constructor
/Users/nshulga/git/pytorch/pytorch/torch/include/c10/core/Scalar.h:59:7: note: candidate constructor
/Users/nshulga/git/pytorch/pytorch/torch/include/c10/core/Scalar.h:59:7: note: candidate constructor
/Users/nshulga/git/pytorch/pytorch/torch/include/c10/core/Scalar.h:62:3: note: candidate constructor
Scalar(uint16_t vv) : Scalar(vv, true) {}
^
/Users/nshulga/git/pytorch/pytorch/torch/include/c10/core/Scalar.h:63:3: note: candidate constructor
Scalar(uint32_t vv) : Scalar(vv, true) {}
^
/Users/nshulga/git/pytorch/pytorch/torch/include/c10/core/Scalar.h:64:3: note: candidate constructor
Scalar(uint64_t vv) {
^
```
Prevent this by providing missing constructors when needed. Alas one can not use SFINAE, as template constructors on Scalar mess up a lot of implicit conversions, so I use `static_asserts` to detect early on if premise for constructing this class holds.
Add ScalarTest::LongsAndLongLongs that is essentially a compile time test
Discovered while trying to enable AOTI on MacOS
Pull Request resolved: https://github.com/pytorch/pytorch/pull/118149
Approved by: https://github.com/ezyang, https://github.com/albanD
ghstack dependencies: #118077, #118076
Summary:
When using a custom deleter InefficientStdFunctionContext was using a
std::unique_ptr<> to store the pointer and call the deleter - but this failed to
call the deleter if the pointer was null. Since we have a separate holder class
anyway take out the std::unique_ptr<> and call the deleter directly.
Fixes#117273
Test Plan:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117418
Approved by: https://github.com/wjakob, https://github.com/yanboliang
Summary:
These dtypes are added since we see more demand for these sub byte dtypes, especially with
the popularity of LLMs (https://pytorch.org/blog/accelerating-generative-ai-2/#step-4-reducing-the-size-of-the-weights-even-more-with-int4-quantization-and-gptq-2021-toks)
Note these are just placeholders, the operator support for these dtypes will be implemented with tensor subclass.
e.g. torch.empty(..., dtype=torch.uint1) will return a tensor subclass of uint1, that supports different operations like bitwsise ops, add, mul etc. (will be added later)
Also Note that these are not quantized data types, we'll implement quantization logic with tensor subclass backed up by these dtypes as well.
e.g `Int4GroupedQuantization(torch.Tensor)` will be implemented with torch.uint4 Tensors (see https://github.com/pytorch-labs/ao/pull/13 as an example)
Test Plan:
CIs
python test/test_quantization.py -k test_uint1_7_dtype
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117208
Approved by: https://github.com/ezyang
Which is faster then ternary.
Following script
```python
import torch
from timeit import default_timer
global_setup = """
"""
setup = """
c10::SymInt a = c10::SymInt(123);
"""
code = """
-a;
"""
from torch.utils.benchmark import Timer
t = Timer(stmt=code, setup=setup, global_setup=global_setup, language="c++", timer=default_timer)
print(t.blocked_autorange())
```
reports 4.17 ns median type before and 3.61 ns after on x86_64 Linux and 2.02 ns before and 1.91 ns after on Apple M1
Pull Request resolved: https://github.com/pytorch/pytorch/pull/117015
Approved by: https://github.com/albanD
Here's the problem: if we support unsigned integer types, and in particular if we support uint64_t, we need a way to represent these integers in Scalar. However, Scalar currently stores all integral values inside int64_t, which is not wide enough to accommodate all possible uint64_t values. So we need to do something to Scalar to support it.
The obvious thing to do is add a uint64_t field to the union, and used it some situations. But when should we use it? The proposal is that we only use it if and only if the integer in question is not representable in int64_t. The historical precedent for this is our handling for uint8_t. Because this type is representable inside int64_t, we have historically stored it inside Scalar as an int64_t. In general, the concept behind Scalar is that it doesn't know the signedness/unsignedness/bitwidth of its input; in particular, we typically construct Scalar from Python int, which doesn't have any concept of how wide the integer is! So it doesn't make any sense to allow for a small integer like 255 to be representable under both the HAS_i tag and the HAS_u tag. So we forbid the latter case.
Although I have proposed this, the PR as currently written just chokes when you pass it a uint64_t that's too big. There's some more logic that would have to be written out for this. I'm putting this out to start to get some agreement that this is the way to do it.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/116595
Approved by: https://github.com/albanD
Instead of manually fixing the indices (extremely error prone when new
dtypes are added) we just setup a lookup table to map ScalarType to the
offsets table.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/116693
Approved by: https://github.com/albanD
ghstack dependencies: #116698
Before this change `-SymInt(std::numeric_limits<int64_t>::min()) == 0` would reliably crash with null pointer dereference, as `data_` of the SymInt returned by `operator-` would be `0x8000000000000000`, because of the carry/overflow flags set by `negq`.
Before the change x86_64 assembly generated for
4f02cc0670/c10/core/SymInt.cpp (L137)
looked as follows:
```
0x7ffff7f2f490 <+115>: movq %rax, %rdx
0x7ffff7f2f493 <+118>: negq %rdx
0x7ffff7f2f496 <+121>: movq %rdx, (%rbp)
0x7ffff7f2f49a <+125>: movabsq $0x4000000000000000, %rdx ; imm = 0x4000000000000000
0x7ffff7f2f4a4 <+135>: cmpq %rdx, %rax
0x7ffff7f2f4a7 <+138>: jle 0x7ffff7f2f520 ; <+259> at SymInt.cpp:141:1
```
`negq %rfx` correspond to unary minus and `cmpq %rdx, 0x4000000000000000` are inverted `check_range`
b6d0d0819a/c10/core/SymInt.h (L247-L249)
Flags raised by `negq` will affect the results of `cmpq`, and as result value would not be allocated on heap, but rather preserved as `nullptr`.
Not sure if it's worth benchmarking, but perhaps using `__builtin_sub_overflow` would be faster as it does not require an extra comparison, just guarantees that overflow flags is cleared after the op.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/116160
Approved by: https://github.com/Skylion007, https://github.com/colesbury
The 2MB thp pages provide better allocation latencies compared to the standard 4KB pages. This change has shown substantial improvement for batch mode usecases where the tensor sizes are larger than 100MB.
Only enabled if THP_MEM_ALLOC_ENABLE environment variable is set.
Relanding https://github.com/pytorch/pytorch/pull/93888 with functionality disabled for Android
Co-authored-by: Nikita Shulga <2453524+malfet@users.noreply.github.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107697
Approved by: https://github.com/malfet
If `cpuinfo_initalize` returns false, call to subsequent cpuinfo functions may result in `abort()`
Also, `defaultNumThreads()` method is assumption if one method fails then try another, and finally return 1.
Alas there are no good way to test it on x86 platform, but on ARM one can replicate it by running `sudo chmod 750 /sys` and then `python3 -c "import torch;torch._C.profiler.gather_traceback(True, True, True)"`
<!--
copilot:poem
-->
### <samp>🤖 Generated by Copilot at 4d942e8</samp>
> _`cpuinfo` fails_
> _avoid undefined behavior_
> _check before you count_
Partially addresses https://github.com/pytorch/pytorch/issues/113568
Pull Request resolved: https://github.com/pytorch/pytorch/pull/113771
Approved by: https://github.com/atalman
We found a scenario where ``tensor.device().is_privateuseone()`` is used to determine whether a tensor is privateuse1 but fails.
In the code of ``Autograd``, for example：
```
::std::tuple<at::Tensor,at::Tensor,at::Tensor> native_batch_norm(c10::DispatchKeySet ks, const at::Tensor & input, const c10::optional<at::Tensor> & weight, const c10::optional<at::Tensor> & bias, const c10::optional<at::Tensor> & running_mean, const c10::optional<at::Tensor> & running_var, bool training, double momentum, double eps) {
auto& input_ = unpack(input, "input", 0);
[[maybe_unused]] auto _any_requires_grad = compute_requires_grad( input, weight, bias );
[[maybe_unused]] auto _any_has_forward_grad_result0 = (isFwGradDefined(input) || isFwGradDefined(weight) || isFwGradDefined(bias));
check_no_requires_grad(running_mean, "running_mean", "native_batch_norm");
check_no_requires_grad(running_var, "running_var", "native_batch_norm");
std::shared_ptr<NativeBatchNormBackward0> grad_fn;
if (_any_requires_grad) {
grad_fn = std::shared_ptr<NativeBatchNormBackward0>(new NativeBatchNormBackward0(), deleteNode);
grad_fn->set_next_edges(collect_next_edges( input, weight, bias ));
grad_fn->eps = eps;
grad_fn->input_ = SavedVariable(input, false);
grad_fn->running_mean_ = SavedVariable(running_mean, false);
grad_fn->running_var_ = SavedVariable(running_var, false);
grad_fn->training = training;
grad_fn->weight_ = SavedVariable(weight, false);
}
...
}
```
When ``weight`` is ``None``, an empty tensor is automatically generated and will be transferred to the backward calculation:
c7e12c7427/torch/csrc/autograd/saved_variable.cpp (L121-L128)
At the beginning of the backward calculation in our scenario, we need to determine whether the input tensor is ``PrivateUse1`` . However, if we use ``tensor.device().is_privateuseone()``, we will get an error ``"tensor does not have a device"``:
c7e12c7427/c10/core/TensorImpl.h (L1223-L1235)
I think this part of the code can be optimized, what do you think?
Pull Request resolved: https://github.com/pytorch/pytorch/pull/113421
Approved by: https://github.com/albanD
Currently whenever the sizes or strides are modified for a `TensorImpl` we
eagerly recompute the numel and memory format flags. This is fine for static
shapes as it's all fast C++ code, but for symbolic shapes it runs slow python code.
This instead changes the `SymbolicShapeMeta` object to compute the derived
quantities lazily at the first request. This has the added benefit that we can
now pass assumptions in `empty_tensor_restride` which remove the need to compute
some contiguity flags at all.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/112785
Approved by: https://github.com/ezyang
ghstack dependencies: #112689, #112890
This should be the last of the "it used to work with static shapes but
it doesn't work with dynamic shapes" hard errors. Now we will just
specialize if you hit it from C++.
The strategy here is a bit clever. We shunt the size() call to Python
binding if an error would have occurred. Importantly, we already have
logic to make sure the newly allocated ints stay live for the duration
of the ArrayRef access.
storage_offset is intentionally omitted because there are some problems
with it. I will fix them next.
This should let us get rid of the aotautograd_static test configuration.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/111935
Approved by: https://github.com/zou3519
This PR supports sym_ite. This is useful for converting SymBool to SymInt in e.g. #109916. Internally, it uses sympy.Piecewise. We cannot use sympy.ITE because it expects the arguments and output all to be boolean type but we want return SymInt type when converting a SymBool to SymInt. So we use sympy.Piecewise to denote the symbolic relationship.
Note that this pr uses the range analysis for sympy.Piecewise implemented in https://github.com/pytorch/pytorch/blob/main/torch/utils/_sympy/value_ranges.py.
Test Plan:
See added test.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/111440
Approved by: https://github.com/ezyang
This PR relands #110022 but accounts for the changes in #110191. Also, the function for creating COW storages is called `lazy_clone_storage` in this PR, instead of `try_ensure`
NOTE: COW storages do not actually copy on write yet, they just have the COW deleter and deleter context applied to them
Part of #109833
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110192
Approved by: https://github.com/ezyang
`libshm.so` depends on the torch library exclusively for `at::RefcountedMapAllocator`,
so it makes sense to move it to c10 along with the other memory allocators.
This means `libshm.so` only depends on `c10` and we don't need to relink
`libshm.so` for every ATen change.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109881
Approved by: https://github.com/albanD
`libshm.so` depends on the torch library exclusively for `at::RefcountedMapAllocator`,
so it makes sense to move it to c10 along with the other memory allocators.
This means `libshm.so` only depends on `c10` and we don't need to relink
`libshm.so` for every ATen change.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109881
Approved by: https://github.com/albanD
The "import torch" crashes with following cpuinfo error on powerpc64.
==============================================================
>>> import torch
Error in cpuinfo: processor architecture is not supported in cpuinfo
Fatal error in cpuinfo: cpuinfo_get_processors_count called before cpuinfo is initialized
Aborted (core dumped)
==================================================================
The patch fixes this by excluding powerpc from using cpuinfo as it is not supported for ppc64.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110708
Approved by: https://github.com/ezyang
Fixes the string_view errors and reland the work. The previous changes in torch/csrc/utils/invalid_arguments.cpp were too aggressive and not tested thoroughly. They are discarded.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110518
Approved by: https://github.com/ezyang
There is an issue with float8 type promotion, because _promoteTypesLookup doesn't contain records for few types between bfloat16 and float8.
I have simply moved float8 types just after bfloat16, however I'm not sure if it doesn't break serialization.
Please, decide if it can stay like this, or should I insert missing records filled with "ud" into _promoteTypesLookup instead of moving types.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110279
Approved by: https://github.com/albanD
We want to be able to use SingletonSymNode to represent strides for Jagged layout tensor. The following is for 3D, but easily generalizable to higher dimensions.
Constraints:
- [B, x, D] (where x represents the "variably lengthed dim") can be strided in two ways [x, 1, sum(x)] and [dx, d, 1]. We need two different placeholder values depending on how the jagged tensor is strided.
- When doing operations we need the strides of output tensors to be expressable in terms of the strides and sizes of the inner tensors. Given [B, x, D] @ [D, D'], the output strides is [x * D', D', 1] rather than some opaque [x2, D', 1]. This constraint exists because if I'm tracing, I need a symint to represent the output stride. This symint needs to come from somewhere; I get it in several ways: (1) create a constant, (2) unbacked symint, (3) create a new input using a source, (4) output of an operation on an existing symint. It is clear that (4) is what we want here, which brings us to the design below.
Design:
Given the two constraints, the most straightforward way to implement this is actually to update SingletonSymNode to include some scalar factor, i.e. Morally, SingletonSymNode represents `factor * [s_0, s_1, …, s_n]` This enables us to symbolically compute strides from sizes.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110369
Approved by: https://github.com/ezyang
ghstack dependencies: #110044
Previously, something like j0 >= 3, would return False. In sympy however, it is not possible to make it so that both j0 >= 3 and j0 < 3 return False. In sympy, you only get to dispatch on Ge, and the remaining are derived, e.g. defining Ge(j0 >= 3) to be False would force Lt(j0, 3) to be True, which is not what we want.
In this PR, we make it so that both j0 >=3 and j0 < 3 error, so that in a future PR when we create the symbolic counterpart of this singleton, the behaviors can be the same.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110044
Approved by: https://github.com/ezyang
This PR does the following:
* Combine `cow/context.<h/cpp>` and `cow/deleter.<h/cpp>` into `cow/COWDeleter.<h/cpp>`
* Rename `Context` to `COWDeleterContext`
* Rename `delete_context` to `cow_deleter`
* Remove the separate `impl_cow_context` bazel library, combining it with the base c10 core library
* Rename `context_test.cpp` to `cow_test.cpp`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110191
Approved by: https://github.com/ezyang
This is reland of PRs #https://github.com/pytorch/pytorch/pull/108626 and #109564. We fixed the IOS build failure by changing
```
((CHECK) ? (EXPR) : ([] { assert(!#CHECK); }(), (EXPR)))
```
to
```
((CHECK) ? (EXPR) : ([] { assert(false); }(), (EXPR)))
```
in TR2_OPTIONAL_ASSERTED_EXPRESSION, since the former syntax was invalid on Apple Clang. Anyway, we could apply the simple fix hoping that c10::optional would be replaced by std::optional soon.
We also enabled -Wdeprecated on c10.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/110019
Approved by: https://github.com/clee2000
In cudagraph trees, we invalidate tensors at some point and drop their storage. Then, when they are accessed with .data_ptr(), a custom error message is thrown. Previously, this invalidation didn't also make untyped_storage()/storage() error which could result in a segfault.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109750
Approved by: https://github.com/zou3519
I added some tests for Conj, Neg and ZeroTensor for both python and C++ functionalization. This also fixes a nasty segfult when running a functorch `jacfwd` test with `torch.compile`, once AOTAutograd is using `FunctionalTensor`.
Changes:
(1) I use Jeffrey's `make_wrapper_subclass(extra_dispatch_keys)` kwarg to plumb extra dispatch keys ontoto the wrapper, mirroring what C++ functionalization does (C++ functionalization will mirror all dispatch keys from the inner tensor to the wrapper, except for python and functorch keys).
(2) FunctionalTensorMode will decompose CompositeImplicitAutograd ops, since (for example) ZeroTensor kernels can send ops like `.to()` directly to the Python key. We'll need a way to toggle this later for pre-dispatch functionalization
(3) Bound `_ForceDispatchKeyGuard` and BatchedTensorImpl's dispatch keyset to python
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109023
Approved by: https://github.com/zou3519
ghstack dependencies: #108654, #109662, #109632
This PR fixes the ownership/lifetime handling for tensor subclasses that override sizes/strides, when tensors get resized.
This is needed now, because `FunctionalTensor` is a subclass that has a custom size/stride (so it can plumb requests to its inner tensor), and is also a core piece of infra (it's used during tracing in AOTAutograd, which means that metadata mutation and resizing that happens to work with torch.compile today needs to work with FunctionalTensor).
After a bunch of discussion with @ezyang and @soulitzer, I updated `PyInterpreter::sym_sizes()` (and friends) so that:
(1) They allocate a py::capsule buffer and stash it on the tensor on the first call to size/stride
(2) On a size/stride call where we noticed that the number of **dimensions** on the tensor has changed (so our buffer it stale), we re-allocate the buffer
(3) On a size/strude cal where we notice that the number of dimensions is the same, but the values are different (this happens whenever a tensor experiences a metadata mutation, like `.transpose_()`), we inplace-modify the buffer and put the new ints/symints into it
I also ended up doing the SmallVector optimization, which was required to fix some tests in AOTAutograd. Ideally we should look into those tests, and nail down the parts of our codebase that rely on SmallVector not re-allocating on a resize... but I'm saving this for a followup.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108654
Approved by: https://github.com/ezyang
We want users to be able to define custom ops in C++ but put the
abstract impl in Python (since it is easier to write them in Python and
the abstract impl better models device semantics and data-dependent
operators).
`m.impl_abstract_pystub(opname, python_module, context)` declares the
abstract_impl of the operator to exist in the given python module.
When the abstract_impl needs to be accessed (either via FakeTensor or
Meta), and it does not exist, the PyTorch Dispatcher will yell
with a descriptive error message.
Some details:
- We construct a new global AbstractImplPyStub mapping in
Dispatcher.cpp. Read/write to this map is protected by the Dispatcher
lock.
- We add a new Meta Tensor fallback kernel. The fallback errors out if there is
no meta kernel, but also offers a nicer error message if we see that there is
a pystub.
- We create a `torch._utils_internal.throw_abstract_impl_not_imported_error`
helper function to throw errors. This way, we can throw different error
messages in OSS PyTorch vs internal PyTorch. To invoke this from C++, we
added a PyInterpreter::throw_abstract_impl_not_imported_error.
Differential Revision: [D49464753](https://our.internmc.facebook.com/intern/diff/D49464753/)
Differential Revision: [D49464753](https://our.internmc.facebook.com/intern/diff/D49464753)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109529
Approved by: https://github.com/ezyang, https://github.com/bdhirsh
In this PR:
- When Constant SymNode are detected in unary/binary ops demote them to plain int/bool before proceeding. Sometimes this means doing a unary op with a Constant SymNode would result in a plain bool.
- Introduce an is_symbolic method, only available from Python. We need this because isinstance(x, SymInt) is no longer sufficient to check whether a given int/SymInt is symbolic or not. See later PR in the stack to see how this is used.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109169
Approved by: https://github.com/ezyang
Unlike TORCH_CHECK, these always show C++ stacktrace on error. Put it
on errors where you frequently seem to need this information.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/109373
Approved by: https://github.com/bdhirsh
ghstack dependencies: #109372
This PR adds a new `FunctionalTensor` subclass, and `FunctionalTensorMode` torch dispatch mode. Together, this class/mode are a lightweight wrapper around our existing C++ functionalization logic.
This idea came from Ed - later in the stack, I want to be able to run functionalization **underneath** torch_dispatch, when performing tracing in AOTAutograd. I can't do this easily with vanilla C++ functionalization, because it has a dedicated dispatch key that always runs before TorchDispatch. However, by adding a torch_dispatch mode shim around functionalization, we can use functionalization as a torch_dispatch mode, which will make it easier to run underneath other modes later.
This PR provides the basic new classes, and some light testing.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/106404
Approved by: https://github.com/ezyang
In this PR:
- {in,}equality between singleton and plain ints returns false instead of erroring
- Morally define the semantic of j0 > c to be as if j0 represented an array [s_0, s_1, ... s_n] and s_k > c for all k
- Just like for equality, we don't actually want to do the comparison one by one, instead j0 is constrained to some range [min, max]. By default this range is [2, int64_t::max] so that it acts like a size and passes 0/1 specialization checks.
- In the future, we can define some API to allow users to constrain the range of their singletons
Pull Request resolved: https://github.com/pytorch/pytorch/pull/108315
Approved by: https://github.com/ezyang
PoC demonstrating vmap + NT based on the [design doc](https://docs.google.com/document/d/1dVVk6TOqz93PLTIneU2T3xaxCs9qZ0MaJyCvOAp_bC0). This PR:
* Allows `BatchedTensorImpl`s to contain NTs
* Introduces a `BatchedNestedTensor` dispatch key for NT-specific batching rules
* Provides a batching rule fallback that unbinds the NTs -> performs computation on constituent -> rebinds results into NT
Restrictions:
* Only supports one level of vmap
* Only supports vmapping over dim=0 for NTs
* For operations with mixed NT / dense inputs, support is also limited to dim=0 for the dense inputs
Pull Request resolved: https://github.com/pytorch/pytorch/pull/106786
Approved by: https://github.com/zou3519
**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
Adds `SingletonSymNodeImpl` (alternatively, `SkolemSymNodeImpl`). This is a int-like object that only allows the`eq` operation; any other operation produces an error.
The main complexity is that we require operations that dispatch to SymNode must take and return SymNodes, but when performing operations involving `SingletonSymNodeImpl`, operations involving SymNode can return non-SymNode bools. For more discussion see [here](https://docs.google.com/document/d/18iqMdnHlUnvoTz4BveBbyWFi_tCRmFoqMFdBHKmCm_k/edit)
- Introduce `ConstantSymNodeImpl` a generalization of `LargeNegativeIntSymNodeImpl` and replace usage of `LargeNegativeIntSymNodeImpl` in SymInt.
- Also use ConstantSymNodeImpl to enable SymBool to store its data on a SymNode. Remove the assumption that if SymBool holds a non-null SymNode, it must be symbolic.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107089
Approved by: https://github.com/ezyang
ghstack dependencies: #107839
This PR stops `SymNode` from mutating (i.e. simplifying) its expression. Instead, the
simplification (without mutation) is deferred to the `SymNode.maybe_as_int` method.
```python
- FakeTensor(size=(s0,), ...)
- FakeTensor(size=(s1, s2, s3), ...)
- Eq(s0, s1 + s2 + s3)
- FakeTensor(size=(s0,), ...)
- FakeTensor(size=(s1, s2, s3), ...)
```
In summary, this PR:
- Replaces `SymNode._expr` by `SymNode.expr`, removing the old property function
- This makes it so `SymNode` instances never update their expression
- Creates `SymNode.simplified_expr()` method for actually calling `ShapeEnv.replace` on
its expression. Note that this doesn't updates `SymNode.expr`
- Changes how `tensor.size()` gets converted to its Python `torch.Size` type
- Instead of calling `SymInt::maybe_as_int()` method, we create a new
`SymInt::is_symbolic()` method for checking whether it is actually a symbolic value
- This is needed so that when we call `tensor.size()` in the Python side, the returned
sequence is faithful to the actual data, instead of possibly simplifying it and
returning an integer
- 2 files needs this modification:
- _torch/csrc/Size.cpp_: for handling `torch.Tensor.size` Python calls
- _torch/csrc/utils/pybind.cpp_: for handling `symint.cast()` C++ calls
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107492
Approved by: https://github.com/ezyang
ghstack dependencies: #107523
Here's what it does from the comments:
```
Assume that a boolean is true for the purposes of subsequent symbolic
reasoning. This will keep track of corresponding runtime checks to verify
that the result is upheld: either as a regular guard, or as a special set
of asserts which are triggered when an unbacked SymInt is allocated.
DO NOT use this function for these cases:
- This is inappropriate for "branching" conditions (where both
true and false result in valid programs). We will always assume
the condition evaluates true, and so it will never be possible
to trace the false condition when you use it. For true branching
on unbacked SymInts, you must use torch.cond.
- This is inappropriate for situations where you know some other system
invariant guarantees that this property holds, since you don't
really need to insert a runtime check in that case. Use something
like constrain_range in that case.
This API has a hitch. To avoid having to reimplement error reporting
capabilities, this function CAN return False. The invariant is that
the surrounding code must raise an error when this function returns
False. This is quite low level, so we recommend using other functions
like check() which enforce this in a more intuitive way.
By the way, this name is a nod to the __builtin_expect likely macro,
which is used similarly (but unlike __builtin_expect, you MUST fail
in the unlikely branch.)
```
We don't do anything with this right now, except use it to discharge regular guards. Follow up PRs to (1) use it at important error checking sites, (2) actually ensure the runtime asserts make there way into the exported IR / inductor generated code.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/106720
Approved by: https://github.com/ysiraichi, https://github.com/voznesenskym
This PR prefers "logical processor number" (the cpu cores shown in htop) returned by cpuinfo for determining c10 thread number. If that fails, it uses hardware_concurrency exactly.
The motivation is that in a x86 host with 64 cores and Hyper-Threading disabled, the current behavior uses 32 threads, resulting half of cores being idle.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107010
Approved by: https://github.com/ezyang
Compiler behavior when non-zero offset is added to a null pointer is undefined and is a bad habit.
- When `lapackEig` is called with to estimate a workspace size, do not add matrix size to the W pointer.
- When `unpack_pivots_cpu_kernel` with zero `dim_size` exit early.
- When `topk_impl_loop` is called with `k` is zero, exit right away as output tensors are empty anyway.
- Ignore adding non-zero storage-offset in `TensorImpl::data_ptr_impl_impl`, which can be the case if tensor is created as `torch.empty(3)[4:]`.
- In `s_addmm_out_sparse_dense_worker` do not call `axpy` over an empty vector.
- In `_sparse_binary_op_intersection_kernel_impl` do skip computing `ptr_indices_dim` when `sparse_dim` is empty.
- Exit `grid_sample` forward/backward kernels earlier if either `input` or `grid` are empty tensors.
Found by asan in clang-12
Before the change UBSan report looks as follows:
```
ASAN_SYMBOLIZER_PATH=/usr/lib/llvm-12/bin/llvm-symbolizer UBSAN_OPTIONS=print_stacktrace=1 LD_PRELOAD=/usr/lib/llvm-12/lib/clang/12.0.1/lib/linux/libclang_rt.asan-x86_64.so python test_fx_experimental.py -v -k test_normalize_operator_exhaustive_linalg_eig_cpu_float32
Test results will be stored in test-reports/python-unittest/test_fx_experimental
Running tests...
----------------------------------------------------------------------
test_normalize_operator_exhaustive_linalg_eig_cpu_float32 (__main__.TestNormalizeOperatorsCPU) ... /opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/overrides.py:111: UserWarning: 'has_cuda' is deprecated, please use 'torch.backends.cuda.is_built()'
torch.has_cuda,
/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/overrides.py:112: UserWarning: 'has_cudnn' is deprecated, please use 'torch.backends.cudnn.is_available()'
torch.has_cudnn,
/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/overrides.py:118: UserWarning: 'has_mps' is deprecated, please use 'torch.backends.mps.is_built()'
torch.has_mps,
/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/overrides.py:119: UserWarning: 'has_mkldnn' is deprecated, please use 'torch.backends.mkldnn.is_available()'
torch.has_mkldnn,
/var/lib/jenkins/workspace/aten/src/ATen/native/BatchLinearAlgebra.cpp:937:17: runtime error: applying non-zero offset 20 to null pointer
#0 0x7f2025794888 in void at::native::lapackEig<float, float>(char, char, int, float*, int, float*, float*, int, float*, int, float*, int, float*, int*) (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so+0x9945888)
#1 0x7f20257da256 in void at::native::(anonymous namespace)::apply_linalg_eig<float>(at::Tensor&, at::Tensor&, at::Tensor&, at::Tensor&, bool) (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so+0x998b256)
#2 0x7f20257d902d in at::native::(anonymous namespace)::linalg_eig_kernel(at::Tensor&, at::Tensor&, at::Tensor&, at::Tensor const&, bool) (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so+0x998a02d)
#3 0x7f20257b5b3d in at::native::linalg_eig_out_info(at::Tensor const&, at::Tensor&, at::Tensor&, at::Tensor&, bool) (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so+0x9966b3d)
#4 0x7f20257b4770 in at::native::linalg_eig_out(at::Tensor const&, at::Tensor&, at::Tensor&) (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so+0x9965770)
#5 0x7f20280710e6 in c10::impl::wrap_kernel_functor_unboxed_<c10::impl::detail::WrapFunctionIntoFunctor_<c10::CompileTimeFunctionPointer<std::tuple<at::Tensor&, at::Tensor&> (at::Tensor const&, at::Tensor&, at::Tensor&), &(at::(anonymous namespace)::(anonymous namespace)::wrapper_CPU_out_linalg_eig_out(at::Tensor const&, at::Tensor&, at::Tensor&))>, std::tuple<at::Tensor&, at::Tensor&>, c10::guts::typelist::typelist<at::Tensor const&, at::Tensor&, at::Tensor&> >, std::tuple<at::Tensor&, at::Tensor&> (at::Tensor const&, at::Tensor&, at::Tensor&)>::call(c10::OperatorKernel*, c10::DispatchKeySet, at::Tensor const&, at::Tensor&, at::Tensor&) (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so+0xc2220e6)
#6 0x7f202727a045 in at::_ops::linalg_eig_out::call(at::Tensor const&, at::Tensor&, at::Tensor&) (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so+0xb42b045)
#7 0x7f20257b7e29 in at::native::linalg_eig(at::Tensor const&) (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so+0x9968e29)
#8 0x7f2028070bf0 in c10::impl::wrap_kernel_functor_unboxed_<c10::impl::detail::WrapFunctionIntoFunctor_<c10::CompileTimeFunctionPointer<std::tuple<at::Tensor, at::Tensor> (at::Tensor const&), &(at::(anonymous namespace)::(anonymous namespace)::wrapper_CPU__linalg_eig(at::Tensor const&))>, std::tuple<at::Tensor, at::Tensor>, c10::guts::typelist::typelist<at::Tensor const&> >, std::tuple<at::Tensor, at::Tensor> (at::Tensor const&)>::call(c10::OperatorKernel*, c10::DispatchKeySet, at::Tensor const&) (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so+0xc221bf0)
#9 0x7f2026b1f787 in std::tuple<at::Tensor, at::Tensor> c10::Dispatcher::redispatch<std::tuple<at::Tensor, at::Tensor>, at::Tensor const&>(c10::TypedOperatorHandle<std::tuple<at::Tensor, at::Tensor> (at::Tensor const&)> const&, c10::DispatchKeySet, at::Tensor const&) const (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so+0xacd0787)
#10 0x7f20273230a7 in at::_ops::linalg_eig::redispatch(c10::DispatchKeySet, at::Tensor const&) (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so+0xb4d40a7)
#11 0x7f202c3cc32d in torch::autograd::VariableType::(anonymous namespace)::linalg_eig(c10::DispatchKeySet, at::Tensor const&) (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so+0x1057d32d)
#12 0x7f202c3cba96 in c10::impl::wrap_kernel_functor_unboxed_<c10::impl::detail::WrapFunctionIntoFunctor_<c10::CompileTimeFunctionPointer<std::tuple<at::Tensor, at::Tensor> (c10::DispatchKeySet, at::Tensor const&), &(torch::autograd::VariableType::(anonymous namespace)::linalg_eig(c10::DispatchKeySet, at::Tensor const&))>, std::tuple<at::Tensor, at::Tensor>, c10::guts::typelist::typelist<c10::DispatchKeySet, at::Tensor const&> >, std::tuple<at::Tensor, at::Tensor> (c10::DispatchKeySet, at::Tensor const&)>::call(c10::OperatorKernel*, c10::DispatchKeySet, at::Tensor const&) (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so+0x1057ca96)
#13 0x7f20272798e0 in at::_ops::linalg_eig::call(at::Tensor const&) (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/lib/libtorch_cpu.so+0xb42a8e0)
#14 0x7f2043d97ae3 in torch::autograd::THPVariable_linalg_eig(_object*, _object*, _object*) (/opt/conda/envs/py_3.9/lib/python3.9/site-packages/torch/lib/libtorch_python.so+0x23feae3)
#15 0x5072d6 in cfunction_call /usr/local/src/conda/python-3.9.17/Objects/methodobject.c:543:19
...
SUMMARY: UndefinedBehaviorSanitizer: undefined-behavior /var/lib/jenkins/workspace/aten/src/ATen/native/BatchLinearAlgebra.cpp:937:17 in
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/106354
Approved by: https://github.com/huydhn, https://github.com/lezcano
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
Previously, x.size(0) could return a SymInt, even when the internal
sympy expression was actually already constant (e.g., due to an
introduced guard.) We now allow to query the Python object with
maybe_as_int which allows us to transmute these objects back to
int when possible.
It is still possible to end up with a constant SymInt even after this
change, e.g., if you get out a SymInt and while holding onto it
specialize it, but casual users are more likely to get ints when they
want to.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/104828
Approved by: https://github.com/Skylion007
Previously, x.size(0) could return a SymInt, even when the internal
sympy expression was actually already constant (e.g., due to an
introduced guard.) We now allow to query the Python object with
maybe_as_int which allows us to transmute these objects back to
int when possible.
It is still possible to end up with a constant SymInt even after this
change, e.g., if you get out a SymInt and while holding onto it
specialize it, but casual users are more likely to get ints when they
want to.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/104828
Approved by: https://github.com/Skylion007
Previously, x.size(0) could return a SymInt, even when the internal
sympy expression was actually already constant (e.g., due to an
introduced guard.) We now allow to query the Python object with
maybe_as_int which allows us to transmute these objects back to
int when possible.
It is still possible to end up with a constant SymInt even after this
change, e.g., if you get out a SymInt and while holding onto it
specialize it, but casual users are more likely to get ints when they
want to.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/104828
Approved by: https://github.com/Skylion007
Fixes#91338
Follow up from https://github.com/pytorch/pytorch/pull/91342
> 🚀 The feature, motivation and pitch
> We have an existing DeviceType class all over the place in our code base, and it conflicts with the one that is used in torch. > Thankfully the pytorch DeciceType enum class is under the c10 namespace.
```
In file included from /xxx/build/_deps/torch-src/../../aten/src/ATen/ops/view.h:5:
/xxx/_deps/torch-src/aten/src/ATen/Context.h:265:14: error: reference to 'DeviceType' is ambiguous
if (p == DeviceType::HIP) {
^
/xxx/include/Common_types.h:178:8: note: candidate found by name lookup is 'DeviceType'
struct DeviceType {
^
/xxx/build/_deps/torch-src/c10/../c10/core/DeviceType.h:32:12: note: candidate found by name lookup is 'c10::DeviceType'
enum class DeviceType : int8_t {
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/104364
Approved by: https://github.com/albanD
As part of this, a new `AutocastIPU` dispatch key has been added.
There's an existing PR, #85043, to make `Autocast` a proper per-backend functionality key, but it ran into issues with layering with other functionality keys and went stale.
This has been tested in the out-of-tree IPU PyTorch backend.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103890
Approved by: https://github.com/albanD
Fixes https://github.com/pytorch/pytorch/issues/103132
This is kind of annoying: Functionalization (and also vmap, I think?) manually figures out which ops have C++ CompositeImplicit decomps, and directly registers them to the Functionalize key. This is a problem for the PyDispatcher: We normally want the PyDispatcher to take precedence over the regular dispatcher. But in this case, we have a python decomp registered to `CompositeImplicitAutograd`, and a C++ decomp registered *directly* to the `Functionalize` key, so the C++ decomp gets precedence over the python decomp.
The way this showed up was that a model was running `matmul()` under inference mode, so we never hit the autograd dispatch key, and go straight to the functionalize dispatch key. Matmul has both a python decomp and a c++ decomp, but we were running the C++ decomp. That C++ decomp isn't meant to be used with dynamic shapes, so we were failing with the "tried to call `.sizes()` on a tensor with dynamic shapes" error.
For now, I had the PyDispatcher mimic the behavior of functionalization codegen: when you register a python decomp to the `CompositeImplicitAutograd` key, this PR just automatically registers that decomp to the `Functionalize` key at the same time.
I'm trying to remember now why we didn't just add `Functionalize` (and all of the other functorch transform keys) directly to the `CompositeImplicitAutograd` alias keyset, but I couldn't remember (@zou3519 any chance you remember?).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/103275
Approved by: https://github.com/ezyang, https://github.com/zou3519
We discussed in a composability meeting a few weeks ago that `pre_autograd` should probably be renamed to `pre_dispatch`.
One question in this PR was: should I re-use a dispatch key? Or should I create a new dispatch key (that yet again corresponds to "top of the dispatcher")?
~~For now, I ended up sticking our proxy mode on the mode stack corresponding to `PythonTLSSnapshot`, because it was simple and it works. It looks like one of the functorch dispatch keys has higher priority though, so it's possible that functorch will end up running first. Open to options, but we can consider adding a new dispatch key later if that becomes a problem~~
Update: I added a dedicated dispatch key, `PreDispatch`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/101818
Approved by: https://github.com/ezyang, https://github.com/Neilblaze, https://github.com/albanD, https://github.com/zou3519
Fixes #ISSUE_NUMBER
For the scenario where users inherit storageimpl to implement their own subclasses, the current storage creation method cannot correctly create storage objects.
Refer to the registration method of Allocator to expand the creation method of storageimpl, users can register their own custom storageimpl creation.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100237
Approved by: https://github.com/albanD
The PyTorch Dispatcher's "no kernel found for DispatchKey" error message
is a bit long and winded. This PR adds a way to add a custom error
callback and changes the CustomOp API to use the custom error callback
to deliver better error messages.
Test Plan:
- new tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/101015
Approved by: https://github.com/ezyang
implementation of DataPtr context for copy-on-write tensors
Summary:
Copy-on-write storage
=====================
This library adds support for copy-on-write storage, i.e. lazy copies,
to tensors. The design maintains the PyTorch invariant that tensors
alias if and only if they share a storage. Thus, tensors that are lazy
copies of one another will have distinct storages that share a data
allocation.
Thread-safety
-------------
The correctness of this design hinges on the pre-existing PyTorch user
requirement (and general default programming assumption) that users
are responsible for guaranteeing that writes do not take places
concurrently with reads and other writes.
Lazily copied tensors add a complication to this programming model
because users are not required to know if lazy copies exist and are
not required to serialize writes across lazy copies. For example: two
tensors with distinct storages that share a copy-on-write data context
may be given to different threads that may do whatever they wish to
them, and the runtime is required to guarantee its safety.
It turns out that this is not that difficult to protect because, due
to the copy-on-write requirement, we just need to materialize a tensor
upon writing. This could be done entirely without synchronization if
we materialized each copy, however, we have a common-sense
optimization to elide the copy for the last remaining reference. This
requires waiting for any pending copies.
### Thread-safety detailed design
There are two operations that affect the copy-on-write details of a
tensor:
1) lazy-clone (e.g. an explicit call or a hidden implementation detail
added through an operator like reshape)
2) materialization (i.e. any write to the tensor)
The key insight that we exploit is that lazy-clone is logically a read
operation and materialization is logically a write operation. This
means that, for a given set of tensors that share a storage, if
materialization is taking place, no other read operation, including
lazy-clone, can be concurrent with it.
However, this insight only applies within a set of tensors that share
a storage. We also have to be concerned with tensors with different
storages that share a copy-on-write context. In this world,
materialization can race with lazy-clone or even other
materializations. _However_, in order for this to be the case, there
must be _at least_ two references to the context. This means that the
context _can not_ vanish out from under you if you are performing a
lazy-clone, and hence, it only requires an atomic refcount bump.
The most complicated case is that all lazy-copies are concurrently
materializing. In this case, because a write is occurring, there are
no in-flight lazy-copies taking place. We must simply ensure that all
lazy-copies are able to materialize (read the data) concurrently. If
we didn't have the aforementioned optimization where the last copy
steals the data, we could get away with no locking whatsoever: each
makes a copy and decrements the refcount. However, because of the
optimization, we require the loser of the materializing race wait for
the pending copies to finish, and then steal the data without copying
it.
We implement this by taking a shared lock when copying the data and
taking an exclusive lock when stealing the data. The exclusive lock
acquisition ensures that all pending shared locks are finished before
we steal the data.
Test Plan: 100% code coverage.
---
Stack created with [Sapling](https://sapling-scm.com). Best reviewed with [ReviewStack](https://reviewstack.dev/pytorch/pytorch/pull/100818).
* #100821
* #100820
* #100819
* __->__ #100818
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100818
Approved by: https://github.com/ezyang
Cruise uses [clang static analyzer](https://clang-analyzer.llvm.org/) internally.
In the v2.0.0 release of PyTorch it found this problem
```
In file included from external/pytorch/aten/src/ATen/ATen.h:7:
In file included from external/pytorch/aten/src/ATen/Context.h:3:
In file included from external/pytorch/aten/src/ATen/CPUGeneratorImpl.h:3:
In file included from external/pytorch/aten/src/ATen/core/Generator.h:22:
In file included from external/pytorch/c10/core/GeneratorImpl.h:8:
In file included from external/pytorch/c10/core/TensorImpl.h:6:
external/pytorch/c10/core/InferenceMode.h:58:5: warning: Passed-by-value struct argument contains uninitialized data (e.g., field: 'view_replay_enabled_')
AutogradState::set_tls_state(AutogradState(
^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
1 warning generated.
```
In other words, the value of `view_replay_enabled_` could be initialized which may lead to subtle bugs later on.
This PR addresses the warning by explicitly initializing it to `false`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100822
Approved by: https://github.com/Skylion007
Previously the change to aten/src/ATen/native/LossNLL.cpp eventually resulted in a double / SymInt division, which ended up calling the int64_t / SymInt overload, truncating the double (bad!) By adding overloads for all the int/float types, we avoid this situation from happening in the future.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100008
Approved by: https://github.com/albanD
This PR introduces **-Wmissing-prototypes** of clang-tidy to prevent further coding errors such as the one fixed by PR #96714.
<!--
copilot:summary
-->
### <samp>🤖 Generated by Copilot at fd2cf2a</samp>
This pull request makes several internal functions static to improve performance and avoid name clashes. It also fixes some typos, formatting, and missing includes in various files. It adds a new .clang-tidy check to warn about missing prototypes for non-static functions.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/96805
Approved by: https://github.com/malfet, https://github.com/albanD
Why?
* To reduce the latency of hot path in https://github.com/pytorch/pytorch/pull/97377
Concern - I had to add `set_offset` in all instances of `GeneratorImpl`. I don't know if there is a better way.
~~~~
import torch
torch.cuda.manual_seed(123)
print(torch.cuda.get_rng_state())
torch.cuda.set_rng_state_offset(40)
print(torch.cuda.get_rng_state())
tensor([123, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0], dtype=torch.uint8)
tensor([123, 0, 0, 0, 0, 0, 0, 0, 40, 0, 0, 0, 0, 0,
0, 0], dtype=torch.uint8)
~~~~
Reland of https://github.com/pytorch/pytorch/pull/98965
(cherry picked from commit 8214fe07e8)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99565
Approved by: https://github.com/anijain2305
Fixes #ISSUE_NUMBER
1、torch.jit.load for custom device
```
# custom device named `foo`
ts_model = torch.jit.script(mode.to(device="foo"))
ts_model.save("./ts.pt") # it is a script model on device `foo`
# and then we want to load it and run it
torch.jit.load("./ts.pt")
```
2、 add some extra key for custom device with `privateuse1`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99535
Approved by: https://github.com/albanD
The strategy is that we will heap allocate a LargeNegativeIntSymNodeImpl whenever we have a large negative int, so that we can keep the old `is_symbolic` test (now called `is_heap_allocated`) on SymInt. Whenever we need to do something with these ints, though, we convert them back into a plain `int64_t` (and then, e.g., wrap it in whatever user specificed SymNodeImpl they need.) We cannot wrap directly in the user specified SymNodeImpl as we generally do not know what the "tracing context" is from C++. We expect large negative ints to be rare, so we don't apply optimizations like singleton-ifying INT_MIN. Here's the order to review:
* c10/core/SymInt.h and cpp
* `is_symbolic` renamed to `is_heap_allocated` as I needed to audit all use sites: the old `is_symbolic` test would return true for large negative int, but it would be wrong to then try to dispatch on the LargeNegativeIntSymNodeImpl which supports very few operations. In this file, I had to update expect_int,
* If you pass in a large negative integer, we instead heap allocate it in `promote_to_negative`. The function is written in a funny way to keep compact constructor code for SymInt (the heap allocation happens out of line)
* clone is now moved out-of-line
* New method maybe_as_int which will give you a constant int if it is possible, either because it's stored inline or in LargeNegativeIntSymNodeImpl. This is the preferred replacement for previous use of is_symbolic() and then as_int_unchecked().
* Rename toSymNodeImpl to toSymNode, which is more correct (since it returns a SymNode)
* Complete rewrite of `normalize_symints.cpp` to use new `maybe_as_int`. Cannot easily use the old code structure, so it's now done doing a macro and typing out each case manually (it's actually not that bad.)
* Reimplementations of all the unary operators by hand to use `maybe_as_int`, relatively simple.
* c10/core/LargeNegativeIntSymNodeImpl.h - Just stores a int64_t value, but it has to be big and negative. Most methods are not implemented, since we will rewrap the large negative int in the real SymNodeImpl subclass before doing operations with it
* The rest of the files are just rewriting code to use `maybe_as_int`. There is a nontrivial comment in c10/core/SymIntArrayRef.h
Very minor test adjustment in c10/test/core/SymInt_test.cpp . Plan to exercise this properly in next PR.
Companion XLA PR: https://github.com/pytorch/xla/pull/4882
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99157
Approved by: https://github.com/albanD
…eMeta
This modularizes ExtraMeta to bring down its creation cost when it is needed for other functions than syn shape handling.
Change-Id: Ife59b201b0c4fd75090fe8be5171a6dd73a10d10
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98399
Approved by: https://github.com/ezyang
BackendMeta offers a binary interface for the backend to attach arbitrary data to TensorImpl. TensorImpl has exactly one "slot" for backend metadata, however backend is free to compose any structure that is opaque to the framework beyond iheriting standard BackendMeta base.
Change-Id: I670fcdd16dd1c2b00f7eaa1cbc5b5dfea59a6221
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97429
Approved by: https://github.com/ezyang
BackendMeta offers a binary interface for the backend to attach arbitrary data to TensorImpl. TensorImpl has exactly one "slot" for backend metadata, however backend is free to compose any structure that is opaque to the framework beyond iheriting standard BackendMeta base.
Change-Id: I670fcdd16dd1c2b00f7eaa1cbc5b5dfea59a6221
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97429
Approved by: https://github.com/ezyang
`is_empty()` checks `numel() == 0`, but we don't need to access `numel_` at all (or the policy that `numel()` checks) in our happy path -- we just need the data pointer from `storage_`. Let's do the check we need to do using only the data we strictly need, rather than adding instructions loading other pieces of data.
Differential Revision: [D44586464](https://our.internmc.facebook.com/intern/diff/D44586464/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98090
Approved by: https://github.com/Skylion007
To implement the warning when transitioning reshape to copy-on-write
storage, we want to be able to detect a write to one view family
following by a read or a write to another one that shares the same
copy-on-write storage.
Because we have historically not been strict about the mutability of
our data pointers, any warning we have would likely be far too
aggressive.
Therefore, this is the first PR in a long series to ensure a strict
distinction between mutable and const data accessors in TensorBase,
TensorImpl, Storage, and StorageImpl.
The rough plan is to give the mutable accessor a new name that is
explicit about mutation, this will also force us to rewrite any code
that really needs a mutation.
Differential Revision: [D44409928](https://our.internmc.facebook.com/intern/diff/D44409928/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97647
Approved by: https://github.com/ezyang
Fixes #ISSUE_NUMBER
1、add amp support for custom backend
2、optimize the file `backend_registration.py`, and rename it with `custom_backend_registration.py`. And then we would register other funcs for custom backend.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/96188
Approved by: https://github.com/bdhirsh
Fixes #ISSUE_NUMBER
1、add amp support for custom backend
2、optimize the file `backend_registration.py`, and rename it with `custom_backend_registration.py`. And then we would register other funcs for custom backend.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/96188
Approved by: https://github.com/bdhirsh
Part of #91395
Also modifies how `StorageImpl`s are stored in JIT static runtime's `MemoryPlanner`, which used to `std::move` `StorageImpl`s into a vector. But `StorageImpl` can no longer be moved. Instead, `MemoryPlanner` now contains a malloced buffer to which we add new `StorageImpl`s using placement new
Pull Request resolved: https://github.com/pytorch/pytorch/pull/93342
Approved by: https://github.com/ezyang
This PR introduces some modifications:
1. We find out some const function parameters that can be passed by reference and add the reference.
2. We find more opportunists of passing by value and change them accordingly.
3. Some use-after-move errors are fixed.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95942
Approved by: https://github.com/Skylion007
Summary:
This is a retry of https://github.com/pytorch/pytorch/pull/94992 which was reverted due to CI issues.
This PR adds a set of unintrepreted data types on PyTorch which can be used to implement experimental functionality out of core (think fp8, int4, int16 quant, etc).
@bypass-github-export-checks
Test Plan:
```
python test/test_quantization.py -k TestBits
```
Reviewers:
Subscribers:
Tasks:
Tags:
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95860
Approved by: https://github.com/atalman
The 2MB thp pages provide better allocation latencies compared to the standard 4KB pages. This change has shown significant improvement for batch mode usecases where the tensor sizes are larger than 100MB.
Only enabled if `THP_MEM_ALLOC_ENABLE` environment variable is set.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/93888
Approved by: https://github.com/jgong5, https://github.com/malfet
Summary:
This PR adds a set of unintrepreted data types on PyTorch which can be used to implement experimental functionality out of core (think fp8, int4, int16 quant, etc).
Note: this is a copy-pasta of https://github.com/pytorch/pytorch/pull/89990 with a bug fix for clang9, easier to just to put up another PR since I'm not sure how comandeering works with Meta-only changes.
@bypass-github-export-checks
Test Plan:
```
python test/test_quantization.py -k TestBits
```
Reviewers:
Subscribers:
Tasks:
Tags:
Fixes #ISSUE_NUMBER
Pull Request resolved: https://github.com/pytorch/pytorch/pull/94992
Approved by: https://github.com/angelayi
This PR removes the unnecessary == 0 guard when constructing empty tensors, by ensuring that when we create a contiguous tensor we go directly to the C++ torch.empty implementation (instead of indirecting through empty_strided), where we can bypass doing zero tests when computing the size of the storage. This probably also speeds up trace time.
When I did this, I found out that `empty_tensor_restride_symint` was flagrantly wrong (we had never exercised it before because we redirected to `empty_strided` in PrimTorch decomp, which doesn't hit this codepath.) The bugs:
* Stride computation was wrong (only `last_idx` was ever written to)
* Using set_sizes_and_strides with `sym_sizes` input doesn't work, because there is some sort of ordering problem where `clone_symvec` isn't safe when you clone a vector into itself. Probably should fix this.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/94512
Approved by: https://github.com/ngimel
The basic idea behind this PR is that we want to continue using the guarding implementations of contiguity tests, if all of the elements are backend (aka, have hints). If they don't have hints, we'll have to do something slower (use the non-short circuiting, non guarding implementations of contiguity), but most of the time you aren't dealing with unbacked SymInts.
So this PR has three parts.
1. We expose `has_hint` on `SymNode`. This allows us to query whether or not a SymInt is backed or not from C++. Fairly self explanatory. Will require LTC/XLA updates; but for backends that don't support unbacked SymInts you can just always return true.
2. We update `compute_non_overlapping_and_dense` to test if the inputs are hinted. If they are all hinted, we use the conventional C++ implementation. Otherwise we call into Python. The Python case is not heavily tested right now because I haven't gotten all of the pieces for unbacked SymInts working yet. Coming soon.
3. We add stubs for all of the other contiguity tests. The intention is to apply the same treatment to them as well, but this is not wired up yet for safety reasons.
Signed-off-by: Edward Z. Yang <ezyang@meta.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/94431
Approved by: https://github.com/voznesenskym
The other `Autograd[Backend]` keys all have fallthrough kernels registered to them, but `AutogradMeta` was missing the fallthrough kernel.
This is a problem for custom ops that don't have autograd support, if you try to run them with meta tensors. If you have a custom op, and register a CPU and a Meta kernel, then:
(1) if you run the op with cpu tensors, it will dispatch straight to the CPU kernel (as expected)
(2) if you run the op with meta tensors, you will error - because we don't have a fallthrough registered to the AutogradMeta key, we will try to dispatch to the AutogradMeta key and error, since the op author hasn't provided an autograd implementation.
Here's a repro that I confirmed now works:
```
import torch
from torch._dispatch.python import enable_python_dispatcher
from torch._subclasses.fake_tensor import FakeTensorMode
lib = torch.library.Library("test", "DEF")
impl_cpu = torch.library.Library("test", "IMPL", "CPU")
impl_meta = torch.library.Library("test", "IMPL", "Meta")
def foo_impl(x):
return x + 1
lib.define("foo(Tensor a) -> Tensor")
impl_meta.impl("foo", foo_impl)
impl_cpu.impl("foo", foo_impl)
with enable_python_dispatcher():
a = torch.ones(2, device='meta')
print("@@@@@")
b = torch.ops.test.foo.default(a)
print(b)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/94603
Approved by: https://github.com/ezyang, https://github.com/albanD
tldr; this should fix some minor perf regressions that were caused by adding more as_strided() calls in aot autograd.
This PR adds a new context manager, `torch.autograd._set_view_replay_enabled()`.
Context: AOT Autograd has special handling for "outputs that alias graph intermediates". E.g. given this function:
```
def f(x):
y = torch.mul(x, 2)
out = y.view(-1)
return out
```
AOT Autograd will do the following:
```
def fn_to_compile(x):
y = torch.mul(x, 2)
out = y.view(-1)
# return the graph intermediate
return y, out
compiled_fn = compile(fn_to_compile)
def wrapper(x):
y, out = compiled_fn(x)
# regenerate the alias of the graph intermediate
return out._view_func(y)
```
What's annoying is that `out._view_func()` will result in a `.as_strided` call, because `out` is an ordinary runtime tensor. This (likely?) caused a perf regression, because when running the backward, out `as_strided_backward()` is slower than our `view_backward()`.
In this PR, I added some TLS for instructing autograd to do view replay instead of as_strided, even when given a normal tensor. I'm definitely interested in thoughts from autograd folks (cc @albanD @soulitzer). A few points that I want to bring up:
(1) One reason that this API seems generally useful to me is because of the case where you `torch.compile()` a function, and you pass in two inputs that alias each other, and mutate one of the inputs. Autograd is forced to add a bunch of as_strided() calls into the graph when this happens, but this would give users an escape hatch for better compiled perf in this situation
(2) To be fair, AOT Autograd probably won't need this TLS in the long term. There's a better (more complicated) solution, where AOT Autograd manually precomputes the view chain off of graph intermediates during tracing, and re-applies them at runtime. This is kind of complicated though and feels lower priority to implement immediately.
(3) Given all of that I made the API private, but lmk what you all think.
This is a followup of https://github.com/pytorch/pytorch/pull/92255.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/92588
Approved by: https://github.com/ezyang, https://github.com/albanD
