Summary: In TS converter, tensor constants are traced as BUFFER and later we will convert them back to CONSTANT_TENSOR. So we need to prevent naming conflicts during lift constant pass.
Test Plan: CI
Differential Revision: D70826426
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148803
Approved by: https://github.com/angelayi
#### Description
Transform quantized operation properly. Add de/quantization before and after the quantized operation.
#### Test Plan
`pytest test/export/test_converter.py -s -k test_ts2ep_convert_quantized_model`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/133026
Approved by: https://github.com/angelayi
#### Description
Transform quantized operation properly. Add de/quantization before and after the quantized operation.
#### Test Plan
`pytest test/export/test_converter.py -s -k test_ts2ep_convert_quantized_model`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/131915
Approved by: https://github.com/angelayi
#### Issue
ScriptObject was treated as normal attribute by the converter previously. This PR lifts it to be a constant and convert it directly to a GetAttr fx node. ScriptObject would also trigger `CallMethod` and this PR adds that support as well.
#### Test Plan
Add test case for ScriptObject.
`pytest test/export/test_converter.py -s -k test_convert_script_object`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/130952
Approved by: https://github.com/angelayi
Summary:
#### Description
Add support for aten::append with a python function that returns a new list with the appended element. We then update the `fx_node` in the `name_to_node` mapping.
aten::append contributed by Jiashen Cao <jiashenc@meta.com>
Fix conversion for csr_ranker_test
```
model_name: csr_ranker_test_4.ptl
has_ts_model: True
has_sample_inputs: True
ops_maybe_missing_meta: set()
script_objects: set()
ts_can_run: True
ts_run_exception: None
can_convert: True
convert_exception: None
ep_result_correct: True
ep_run_exception: None
can_package: True
package_exception: None
sigmoid_can_run: False
sigmoid_run_exception: RuntimeError('not for symbolics')
sigmoid_result_correct: None
```
Test Plan:
test_aten_add_t
test_aten_append_t
test_aten_to_dtype_with_mutating_storage
buck2 run mode/opt sigmoid/inference/ts_migration:main -- --mode test_one --model_name csr_ranker_test
Differential Revision: D60635893
Pull Request resolved: https://github.com/pytorch/pytorch/pull/132529
Approved by: https://github.com/jiashenC
Add similar semantics for creating a buffer object similar to creating a parameter. This is done by introducing a new Buffer class that can be used for type disambiguation. The underlying functionality of registering a buffer remains the same as the register_buffer method has not been changed. The persistent parameter in the Buffer type is to indicate whether a buffer object should be persistent or not. Other non-test changes have to do with getting the new Buffer type recognized by inductor and dynamo. Remaining changes are test changes to make sure that the Buffer type can be used as a drop in replacement for register_buffer as it just leads to register_buffer being called. The addition of this new functionality still allows for normal tensors to be used as buffers so these changes are intended to be backwards compatible.
Fixes#35735
Co-authored-by: Mikayla Gawarecki <mikaylagawarecki@gmail.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125971
Approved by: https://github.com/albanD, https://github.com/anijain2305, https://github.com/mlazos
Summary: Currently, running explain with TORCH_LOGS enabled will cause duplicate loggings because explain uses the exact same code path for covnersion. This PR just disables logging when it is running explain. And move all logging to convert() to prevent from logging from __init__ when we are just using explain.
Test Plan: Manual testing with attached outputs.
Reviewed By: SherlockNoMad, angelayi
Differential Revision: D60199007
Pull Request resolved: https://github.com/pytorch/pytorch/pull/132082
Approved by: https://github.com/ydwu4
This is an updated PR to equip cond with the autograd feature and replaces the old [PR](https://github.com/pytorch/pytorch/pull/126007)
@ydwu4 I tried to incorporate your requests already.
Currently there are two problems that I struggle with solving:
1. There seems to be an import issue when trying to import cond in `torch/__init__.py`, see [here](8a704035c9/torch/__init__.py (L1914-L1916)). Therefore, I had to comment those lines, which resolved the import issues, but I believe cond is not proberly exposed as torch.cond.
2. I am not entirely sure how to deal with the opinfo test in `hop_db.py`
Co-authored-by: Yidi Wu <yidi@meta.com>
Co-authored-by: Xuehai Pan <XuehaiPan@outlook.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/126911
Approved by: https://github.com/ydwu4
#### Issue
Model parameters sometime do not appear in the `named_parameters()` function. For example, when trying to jit.trace an already jit.scripted model. This PR fixes that by relying on `state_dict` to get both parameters`requires_grad=True` and buffers.
#### Test Plan
* `pytest test/export/test_converter.py -s -k test_convert_retrace_nested_scripted_modules`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/129787
Approved by: https://github.com/angelayi
#### Issue
Fix two issues related to inputs lifting when there are sub-blocks.
* Some inputs may appear in the nested sub-blocks, which need a recursive search to identify which arguments need to be lifted / passed in the top-level block.
* Some inputs to the sub-block are intermediate results, meaning their names are only number. This will cause issue during code generation (i.e., invalid argument name). We rename those to valid names.
#### Test Plan
* `pytest test/export/test_converter.py -s -k test_convert_nn_module_with_nested_if_and_param`
* `test/export/test_converter.py -s -k test_hidden_input_name`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128386
Approved by: https://github.com/angelayi
- Lifting Tensor Constant attributes to buffers: TorchScript does not automatically lift tensor constant attributes to buffers. So previous converter cannot access tensor constant attributes. This PR fixed the issue.
- Add SetAttr support for tensor attributes by copy_.
- Add SetAttr support for non-tensor attributes. In particular, we maintain the current value of non-tensor attributes in `name_to_non_tensor_attribute_node`, similar to an interpreter pass on non-tensor attributes. So we can support the following use case:
```python
def forward(self, x):
c1 = self.count
self.count += 1
c2 = self.count
return x + c1 + c2
```
- Fixed a bug in GetAttr to support the following use case:
```python
def forward(self, inp):
x = self.buffer
self.buffer += 1
y = self.buffer
return x + y + inp
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/129440
Approved by: https://github.com/angelayi
#### Issue
In jit.trace, torch.numel() is automatically cast to a `LongTensor`. But during conversion, we lost the casting part. `prim::NumToTensor` was previously converted to `torch.ops.aten.scalar_tensor`, which uses the same `dtype` as the input tensor instead of `LongTensor`. in this PR, we add a casting to convert it to the correct `dtype`.
#### Test Plan
We activate previously failing test case.
* `pytest test/export/test_converter.py -s -k test_implicit_constant_to_tensor_handling`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128761
Approved by: https://github.com/angelayi
#### Issue
Tensor constant was previously lifted directly as an input in the fx graph, which results errors for multiple test cases with tensor constant. This PR introduces a fix to convert tensor constant to a `GetAttr` in the fx graph.
This PR also introduces other fixes to maintain a valid `state_dict` for exported program when there are tensor constants. In short, after tensor constants are converted as `GetAttr`, they are treated as buffers during retracing. The fix will convert those back from buffer to constant.
#### Test Plan
Add new test cases that generate tensor constants
* `pytest test/export/test_converter.py -s -k test_implicit_constant_to_tensor_handling`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128442
Approved by: https://github.com/angelayi
Summary: prim::dtype has the signature `(Tensor a) -> int`, where it gets the dtype of the tensor and returns the integer corresponding to this dtype based on the enum in ScalarType.h. Previously we were converting prim::dtype by returning the actual dtype of the tensor (ex. torch.float32). This causes some incorrect control flow to behavior, specifically where it checks if `prim::dtype(tensor) in [3, 5, 7]`, where [3, 5, 7] correspond to torch.int32, torch.float16, torch.float64. This control flow would always returns False because we would be comparing torch.float32 against the integers [3, 5, 7], which is a type mismatch.
Test Plan: 7/22 internal models now are convertable and runnable in eager and sigmoid! P1410243909
Reviewed By: jiashenC
Differential Revision: D58469232
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128517
Approved by: https://github.com/jiashenC
#### Description
Handle custom ops during TorchScript to ExportedProgram covnersion
```python
torch.library.define(
"mylib::foo",
"(Tensor x) -> Tensor",
lib=lib,
)
# PyTorch custorm op implementation
@torch.library.impl(
"mylib::foo",
"CompositeExplicitAutograd",
lib=lib,
)
def foo_impl(x):
return x + x
# Meta function of the custom op.
@torch.library.impl_abstract(
"mylib::foo",
lib=lib,
)
def foo_meta(x):
return x + x
class M(torch.nn.Module):
def forward(self, x):
return torch.ops.mylib.foo(x)
```
#### Test Plan
* Add a test case where custom op is called and converted. `pytest test/export/test_converter.py -s -k test_ts2ep_converter_custom_op`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/127580
Approved by: https://github.com/angelayi
#### Description
Add support for converting `prim::__contains__` from TorchScript IR to ExportedProgram, e.g.,
```python
class MIn(torch.nn.Module):
def forward(self, x: torch.Tensor):
return x.dtype in [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
```
#### Test Plan
* Add test cases to cover both contains IR resulted from primitive types or Tensor. `pytest test/export/test_converter.py -s -k test_ts2ep_converter_contains`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/127544
Approved by: https://github.com/angelayi
Summary:
Initial commit for TorchScript 2 ExportedProgram Converter.
TODO:
- Improve TorchScript IR coverage
- parameter and buffers should be owned by output ExportedProgram
- Experiment on conditional op conversion
Test Plan: buck2 run mode/dev-nosan fbcode//caffe2/test:test_export -- -r TestConverter
Differential Revision: D57694784
Pull Request resolved: https://github.com/pytorch/pytorch/pull/126920
Approved by: https://github.com/angelayi, https://github.com/tugsbayasgalan
