Summary:
Revive https://github.com/pytorch/pytorch/pull/138406. Only limit the scope to files in c10.
Summary from the original PR,
```
Looking in the code I see
// NB: __cplusplus doesn't work for MSVC, so for now MSVC always uses
// the "__declspec(deprecated)" implementation and not the C++14
// "[[deprecated]]" attribute. We tried enabling "[[deprecated]]" for C++14 on
// MSVC, but ran into issues with some older MSVC versions.
But looking at the MSVC C++ support table I see that the [[deprecated]] attribute is supported as of MSVC 2015 and that the vast majority of C++17 features became supported in MSVC 2015 or later.
Since PyTorch is C++17 now, I infer that PyTorch must not support versions of MSVC earlier than MSVC 2015, so the versions of MSVC supported by PyTorch must support [[deprecated]].
Therefore, since we are finished deprecating old MSVCs we can deprecate C10_DEPRECATED.
```
Test Plan: CI
Differential Revision: D72762767
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151058
Approved by: https://github.com/r-barnes
Instead of `setup-miniconda`
- Remove `CONDA_RUN` macro...
- Hack the search path in `macos-test.sh` to put both python and python3 aliases first in the path (not sure what other action are messing with path environment variable)
- Skip `TestMultiprocessing.test_fs_sharing` as even though it completes, it hangs on the shutdown both in CI and in all local setups I have
- Skip `TestCppExtensionOpenRgistration.test_base_device_registration` as it hangs on the shutdown as well
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155698
Approved by: https://github.com/atalman
ghstack dependencies: #155476, #155493, #155601, #155515, #155697
Summary:
Moves DelegateExecutor base class to PyTorch core. It provides the extension point of backend delegation for NativeRT.
Torch Native Runtime RFC: pytorch/rfcs#72
Test Plan:
This is only a virtual base class. So relying on internal CI is sufficient.
Rollback Plan:
Differential Revision: D76351984
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155581
Approved by: https://github.com/zhxchen17
vLLM profiler sets with_stack=True that shows the dict_getitem on the profiler, both inflating the numbers and confusing compile users. This PR keeps BINARY_SUBSCR for regular dicts, while using `dict.__getitem__` only for dict subclasses.
Using binary_subscr is little bit faster, but not enough to make any major latency improvements.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155727
Approved by: https://github.com/zou3519, https://github.com/StrongerXi, https://github.com/jansel
Summary:
urrently the node.meta["stack_trace"] is not preserved when we torch package/load GraphModule, which means the original stack trace is lost. When we re-trace the packaged graph module, we just get a stack trace like fx-generated._0......
Adding the node.meta["stack_trace"] to torch packaged graph module
Test Plan:
```
buck2 run @//mode/dev-nosan fbcode//caffe2/test:package -- -r TestPackageFX
```
Rollback Plan:
Differential Revision: D76379692
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155638
Approved by: https://github.com/angelayi
**Summary**
GEMM templates for INT4 weights are used for lowering `aten._weight_int4pack_mm_for_cpu` with Inductor when max-autotune is on. Currently, AMX-based microkernels are used only when M >= 16 if input tensor has shape [M, K]. However, we find that AMX kernel brings performance benefit when 4 < M < 16. For example, on a 6th gen of Intel(R) Xeon(R) platform, E2E latency can be improved by up to > 20% when running Llama-3.1-8B on 32 cores for M = 8. So, this PR changes the threshold so that AMX is used when M > 4.
**Test plan**
```
pytest test/inductor/test_cpu_select_algorithm.py -k test_int4_woq_mm
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155444
Approved by: https://github.com/sanchitintel, https://github.com/leslie-fang-intel
Handles GC for non-strict draft export; GPU memory usage shouldn't be much more than eager mode + input tensors now.
While trying to do draft export CPU offloading, I found out GC is feasible, because in non-strict, there's 2 places holding references to a `.real_tensor` attribute:
1) the FakeTensors in fake tensor prop, but these are held by the actual variables in the model's forward call, and so the real tensor gets gc-ed along with the fake one when the variable goes out of scope.
2) A clone of the fake tensor in 1) stored in `proxy.node.meta["val"]`, which was added in https://github.com/pytorch/pytorch/pull/150948. But we didn't actually need to store them on intermediate values; the placeholders are enough for retracing/lowering.
Avoiding storing the intermediate values in 2), the values in 1) should be naturally GC-ed, and the real-tensor memory usage for non-strict should be pretty similar to eager computation?
Strict still OOMs; dynamo still holds these in variable tracking, and not sure how to GC those.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154630
Approved by: https://github.com/angelayi, https://github.com/yushangdi
as titled. It's sometimes confusing to use PlacementStrategy as a name,
as we also have OpStrategy and TupleStrategy, the latter two contain
the former, so it is better to make the naming clearer.
Renaming PlacementStrategy -> OpSpec as it is an operator spec that
contains output_spec + input_specs.
Also found some utils can be merged to OpSchema so included together in
this PR
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155592
Approved by: https://github.com/awgu
Summary: This diff enhances the `get_process_group_ranks()` function to accept `group=None` as an optional argument. This allows the function to return all ranks associated with the default process group if no group is specified.
Test Plan:
contbuild & OSS CI
Rollback Plan:
Differential Revision: D75817800
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154902
Approved by: https://github.com/wz337
As we prepare to support re-sharding, the current approach of using BytesStorageMetadata to read safetenstors won't work anymore. Before, we didn't need to read the metadata of the safetensors file from its header because we were just loading the contents of the file directly into tensors with safetensor.load() that would handle the metadata and deserialization. But now, in preparation of handling re-sharding, we need to read the metadata directly from the header of the safetensors file and store it directly in TensorStorageMetadata objects so that we can perform re-sharding. Re-sharding won't currently work, as we need extra metadata to be stored on each save, so that will be added in a subsequent PR.
In addition this PR adds an integration test in addition to the unit tests.
It also removes the HfFileSystem import because that's only needed if users are using HfFileSystem, but we want to support any backend.
Differential Revision: [D74891998](https://our.internmc.facebook.com/intern/diff/D74891998/)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/154518
Approved by: https://github.com/saumishr
