…Handle. (pytorch#159989)

Summary:
Today users outside of pytorch core cannot `#include <torch/nativert/ModelRunner.h>`.

It turns out that we should place a header inside `torch/csrc/api/include/`. Placing every single nativert header here would pollute the namespace a lot and that's not what we want in general. Therefore here we just create a Handle type which hold a pointer to decouple the actual type from header definition.

Test Plan:
CI

Rollback Plan:

Differential Revision: D79751098

Pull Request resolved: pytorch#159989
Approved by: https://github.com/dolpm

Fixes #ISSUE_NUMBER

Pull Request resolved: pytorch#158340
Approved by: https://github.com/seemethere

Co-authored-by: xinan.lin <[email protected]>

The test filter was wrong, it should not start with "test/".

Test Plan:
- wait for CI
- Tested locally with `python test/run_test.py --einops --verbose`
Pull Request resolved: pytorch#159776
Approved by: https://github.com/atalman, https://github.com/StrongerXi

numba currently doesn't build from source due to
numba/numba#10073
Pull Request resolved: pytorch#158636
Approved by: https://github.com/malfet

**Summary**
Some thoughts on view-op and `_StridedShard` interaction:
1. `_StridedShard` has no impact on sharding (i.e. how tensor is partitioned)
compared to `Shard`. It only changes how shards permute across the devices.
2. `view()` op on DTensor strictly forbids shard redistribution which means if
`view()` may cause shard permutation across devices, it should be rejected.
This is enforced in today's sharding prop for `view()`.
3. Since DTensor `view()` won't introduce any redistribution, it's certain that
`placements` won't change except the inner `dim` attribute of `Shard`
or `_StridedShard`.

Therefore, to support `_StridedShard` in `view()` op, the only change required
is to keep `_StridedShard` as `_StridedShard` in the output spec.

**Test**
`pytest test/distributed/tensor/test_view_ops.py`

Pull Request resolved: pytorch#159656
Approved by: https://github.com/wconstab

Disables the job on PRs completely, so that we don't litter people's CI signals and use machines unnecessarily.

If you want to run these xla tests, add the ciflow/unstable label to your PR
Pull Request resolved: pytorch#159272
Approved by: https://github.com/atalman, https://github.com/malfet

…)"

This reverts commit 4604f04.

Reverted pytorch#155200 on behalf of https://github.com/jithunnair-amd due to Broke ROCm periodic runs on MI300 e.g. https://github.com/pytorch/pytorch/actions/runs/16764977800/job/47470050573 ([comment](pytorch#138222 (comment)))

This reverts commit 15f1173.

Reverted pytorch#152932 on behalf of https://github.com/jithunnair-amd due to Broke ROCm periodic runs on MI300 e.g. https://github.com/pytorch/pytorch/actions/runs/16764977800/job/47470050573 ([comment](pytorch#138222 (comment)))

)"

This reverts commit f7a66da.

Reverted pytorch#138222 on behalf of https://github.com/jithunnair-amd due to Broke ROCm periodic runs on MI300 e.g. https://github.com/pytorch/pytorch/actions/runs/16764977800/job/47470050573 ([comment](pytorch#138222 (comment)))

Fixed `test_dynamo_timed `:
<img width="1030" height="389" alt="image" src="https://github.com/user-attachments/assets/02d84dd8-6a65-4f91-8d4c-48ba0a81fac1" />

Pull Request resolved: pytorch#159981
Approved by: https://github.com/angelayi

Unification inductor debug build, follow @desertfire 's suggestion: pytorch#159938 (review)

Pull Request resolved: pytorch#159998
Approved by: https://github.com/angelayi

Summary: Updated README with code structure and explanation of core features within profiler

Test Plan:
N/A

Rollback Plan:

Differential Revision: D79604189

Pull Request resolved: pytorch#159816
Approved by: https://github.com/sanrise, https://github.com/aaronenyeshi

…ytorch#155121) (pytorch#158758)

Fixes pytorch#155121

Pull Request resolved: pytorch#158758
Approved by: https://github.com/EikanWang, https://github.com/eellison

…on (pytorch#155360) (pytorch#158983)

Pull Request resolved: pytorch#158983
Approved by: https://github.com/eellison
ghstack dependencies: pytorch#158758

Update HF components to not inherit from fsspec components and instead use filesystem writer/reader. The reason is because there doesn't seem to be much of a need for fsspec, since users are using mounted storage. Using local storage will allow for performance improvements because we can take advantage of the safe_open API provided by HF safetensors (30s vs 4s for load of 8b model), which is signifcant performance wins over reading bytes and converting to tensors which is what we are doing now. Also, we can use the official methods provided by HF instead of relying on reading the metadata by bytes and loading it

Differential Revision: [D78993550](https://our.internmc.facebook.com/intern/diff/D78993550/)

Pull Request resolved: pytorch#159405
Approved by: https://github.com/saumishr

…9406)

Reading the bytes and converting to tensors is much slower than using safe_open. For a 8B model across 8 ranks, took ~30s to load before this change and ~4s after.

Differential Revision: [D78994259](https://our.internmc.facebook.com/intern/diff/D78994259/)

Pull Request resolved: pytorch#159406
Approved by: https://github.com/saumishr
ghstack dependencies: pytorch#159405

Get rid of the logic to read the metadata from the header of the safetensors file manually and use the functions as part of safe_open() to get the metadata. This is much cleaner and allows us to not rely on our own custom methods to get metadata, but use safetensors provided APIs

Differential Revision: [D79460272](https://our.internmc.facebook.com/intern/diff/D79460272/)

Pull Request resolved: pytorch#159681
Approved by: https://github.com/saumishr
ghstack dependencies: pytorch#159405, pytorch#159406

…160070)

This is a follow up on pytorch#159800 as other tests are still failing.

Pull Request resolved: pytorch#160070
Approved by: https://github.com/aorenste

Since switching from wheel 0.34.2 to wheel 0.45.1
python symlinks are no longer correctly created.

Migrate to packaging package for symlink creation
Pull Request resolved: pytorch#158634
Approved by: https://github.com/malfet

Discussed with @jianan-gu and @Valentine233 , disable flex decoding on Windows.

Pull Request resolved: pytorch#160072
Approved by: https://github.com/angelayi

)

If the user provides a generator kwarg to a random op (e.g.
nn.init.uniform_(..., generator=my_generator)), we can still advance
that generator's state in a SPMD-global way so that each local-tensor
gets appropriate values and the generator advances to the same state as
if it had operated on the full tensor.

Pull Request resolved: pytorch#159933
Approved by: https://github.com/fduwjj, https://github.com/XilunWu, https://github.com/wanchaol

Previously we only applied this move_to_device_pass to the toplevel graph. However if we have HOO, this pass will not be applied on the HOO submodules. This PR modifies the pass to run on all submodules.
Pull Request resolved: pytorch#159992
Approved by: https://github.com/yiming0416

Summary: In qembeddingbag_byte_prepack_meta, weight.sizes() would return a concrete int. we should use .sym_size() to return a SymInt instead.

Test Plan:
CI

Rollback Plan:

Reviewed By: kqfu, henryoier

Differential Revision: D79744512

Pull Request resolved: pytorch#159985
Approved by: https://github.com/jerryzh168, https://github.com/henryoier

…59691)

partially generated with
```
for TESTCASE in $(ls | cut -f1 -d'.' | grep -v CPython | uniq); do if grep "$TESTCASE" -m 1 .. -r; then echo; else   sl rm "$TESTCASE"* ; fi; done
```

Pull Request resolved: pytorch#159691
Approved by: https://github.com/xmfan

…mismatches in tracing and take a preferred device. (pytorch#159931)

Summary:
Device mismatches in tracing can most often be ignored. These are only logical mismatches not physical.

Take any intermediate computation, and that computation will not actually materialize in a compiled binary execution. So a device mismatch in the middle of the program is not real. The runtime will never materialize those tensors on CPU device during the execution, as they are temporary allocations.

If a user knows his tensors at graph input are all on the correct device, then he can ignore all tracing errors.

Users who know what they are doing should have an escape hatch to ignore any device mismatch in tracing.

Users can set
```
  torch._functorch.config.fake_tensor_prefer_device_type = 'mtia'
```
to forcefully override any mismatch and prefer the non cpu device. This unblocks vLLM graph mode for MTIA.

Test Plan:
Added two unit tests.

Rollback Plan:

Differential Revision: D79698438

Pull Request resolved: pytorch#159931
Approved by: https://github.com/jansel

…h#160128)

Pull Request resolved: pytorch#160128
Approved by: https://github.com/mori360

…ytorch#159691)"

This reverts commit 36f46d0.

Reverted pytorch#159691 on behalf of https://github.com/izaitsevfb due to breaking dynamo tests ([comment](pytorch#159691 (comment)))

…(but not insides of list) (pytorch#145089)

Signed-off-by: Edward Z. Yang <[email protected]>

Pull Request resolved: pytorch#145089
Approved by: https://github.com/albanD, https://github.com/zou3519

Summary:

Exceptions during autotune kernel precompilation are now systematically captured and reported via the chromium_event_logger, enabling better debugging and analysis of autotune failures.

Currently, exceptions are dumped to the console in the following format::
```
[0/0] RuntimeError: No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 262144 Hardware limit:232448 Reducing block sizes or `num_stages` may help.
[0/0] Runtime error during autotuning:
[0/0] No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 262144 Hardware limit:232448 Reducing block sizes or `num_stages` may help..
[0/0] Ignoring this choice.
```

The exception tracebacks:
```
# inner exception
traceback:
  File "/torch/_inductor/runtime/triton_heuristics.py", line 603, in _make_launchers
    launchers.append(result.make_launcher())
                     ^^^^^^^^^^^^^^^^^^^^^^
  File "/torch/_inductor/runtime/triton_heuristics.py", line 1503, in make_launcher
    self.kernel.load_kernel(device)
  File "/torch/_inductor/runtime/static_cuda_launcher.py", line 113, in load_kernel
    (self.function, self.n_regs, self.n_spills) = _StaticCudaLauncher._load_kernel(

# wrapped exception
traceback:
  File "/usr/local/fbcode/platform010/lib/python3.12/concurrent/futures/thread.py", line 59, in run
    result = self.fn(*self.args, **self.kwargs)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "<trimmed>#link-tree/torch/_inductor/select_algorithm.py", line 2596, in precompile_with_captured_stdout
    choice.precompile()
  File "<trimmed>#link-tree/torch/_inductor/select_algorithm.py", line 1881, in precompile
    self.bmreq.precompile()
  File "<trimmed>#link-tree/torch/_inductor/autotune_process.py", line 660, in precompile
    getattr(mod, self.kernel_name).precompile()
  File "<trimmed>#link-tree/torch/_inductor/runtime/triton_heuristics.py", line 440, in precompile
    self._make_launchers()
  File "<trimmed>#link-tree/torch/_inductor/runtime/triton_heuristics.py", line 608, in _make_launchers
    raise RuntimeError(f"No valid triton configs. {type(exc).__name__}: {exc}")
```

With this change, the exception details will also be logged in the metadata of the `{name}_template_precompiling` event.

The format:
```
{
  "exceptions": [
    {
      "choice_type": "triton",
      "choice": "ACC_TYPE='tl.float32', ALLOW_TF32=False, BLOCK_K=128, BLOCK_M=64, BLOCK_N=64, EVEN_K=True, GROUP_M=8, USE_FAST_ACCUM=False, num_stages=5, num_warps=4, num_consumer_groups=0, num_buffers_warp_spec=0",
      "exception_message": "No valid triton configs. OutOfMemoryError: out of resource: triton_mm Required: 262144 Hardware limit:232448 Reducing block sizes or `num_stages` may help.",
      "exception": "OutOfMemoryError",
      "required_memory": "262144",
      "hardware_limit": "232448"
    }
  ]
}
```

Test Plan:
buck2 run //scripts/wychi:test_autotune_mm 2>&1 > /tmp/mylog.txt

Rollback Plan:

Differential Revision: D79420953

Pull Request resolved: pytorch#159688
Approved by: https://github.com/stashuk-olek

…gitsLoss` (pytorch#150282)"

This reverts commit f990490.

Reverted pytorch#150282 on behalf of https://github.com/facebook-github-bot due to Diff reverted internally ([comment](pytorch#150282 (comment)))

…ubmodules (pytorch#157979)

Fixes issue with HF Gen AI models where we mark a param as static and a get_attr node gets put in the region.

The effect of this is lifting get_attr nodes to be inputs.

Pull Request resolved: pytorch#157979
Approved by: https://github.com/williamwen42

Before we would topologically sort each region individually, this works well except if some nodes have no arguments, then their order may change. To rectify this, we sort the first region as the reference region and use that sort order to sort the remaining regions.

Pull Request resolved: pytorch#158814
Approved by: https://github.com/williamwen42

get_free_symbol_uses is used to know what unbacked symbols are used by a given node.
not having correct get_free_symbol_uses defined properly leads to :

- eliminating of some nodes due to not detection of any users. (See the added unit test)
- Incorrect topological sort.

Fix get_free_symbol_uses , NopKernel , ConcarKernel, InputsKerenl, external kernel.
for ComputedBuffer with NonOwningLayout its interesting case.
when layout is NonOwningLayout we need to access the actual view op base layout and use
detect symbols in it. Because when we codegen the ComputedBuffer we uses those symbols.

Pull Request resolved: pytorch#160314
Approved by: https://github.com/eellison

XCCL backend no pytorch#62300 issue, add xccl path here.

Pull Request resolved: pytorch#159240
Approved by: https://github.com/guangyey, https://github.com/Skylion007, https://github.com/EikanWang

For [pytorch#114850](pytorch#114850), we will port distributed tests to Intel GPU.

We could enable Intel GPU with following methods and try the best to keep the original code styles:

1. use "torch.accelerator.current_accelerator()" to determine the accelerator backend
2. enabled XPU for some test path
3. skip some test cases which Intel GPU does not support

Pull Request resolved: pytorch#159543
Approved by: https://github.com/guangyey, https://github.com/d4l3k

Co-authored-by: Yu, Guangye <[email protected]>

- Add pytorch_overview.md
- Add pytorch_main_components.md
- Reorganize top nav to have Get Started, User Guide, Reference API, Community, Tutorials
- Move notes under user guide

Pull Request resolved: pytorch#159379
Approved by: https://github.com/albanD

Co-authored-by: sekyondaMeta <[email protected]>
Co-authored-by: Nikita Shulga <[email protected]>

…-> default val of 240 (pytorch#160500)

10 hours is very long
Pull Request resolved: pytorch#160500
Approved by: https://github.com/huydhn

Fixed `test_nccl_user_buffer_registration ` due to pytorch#160145, somehow CI didn't capture it.

Pull Request resolved: pytorch#160497
Approved by: https://github.com/ngimel

…tion (pytorch#160357)"

This reverts commit cbffde7.

Reverted pytorch#160357 on behalf of https://github.com/clee2000 due to broke a bunch of internal builds due to not being able to find the file  No such file or directory: torch/_inductor/kernel/flex/templates/flex_decode.py.jinja D80145761, might need a buck targets change? ([comment](pytorch#160357 (comment)))

Fixes builds for old compilers

Pull Request resolved: pytorch#160499
Approved by: https://github.com/Skylion007

… contain attributes (pytorch#160436)

Summary: Fixes internal test failures of D80037015

Test Plan:
CI

Rollback Plan:

Differential Revision: D80094187

Pull Request resolved: pytorch#160436
Approved by: https://github.com/clee2000

Hi, @malfet
Based on the previous discussion:

[RISCV CI support · Issue pytorch#141550 · pytorch/pytorch](pytorch#141550)

I have cross-compiled PyTorch for the RISC-V architecture on x86_64 Ubuntu 24.04 and created a new PR for it. Could you please help review it?

Pull Request resolved: pytorch#143979
Approved by: https://github.com/malfet

Co-authored-by: Nikita Shulga <[email protected]>

…ch#147758)

Summary:
DCP metadata collectives become prohibitively expensive as the job scale grows. This PR introduces rank-local checkpointing which basically saves and loads the checkpoint without any collective. The trade off for now is the dedupe and re-sharding. Support for these would be introduced soon.

Differential Revision: D70112642

Pull Request resolved: pytorch#147758
Approved by: https://github.com/meetv18

## Summary
- register conv3d with MPS autocast to ensure bias dtypes match under AMP
- add regression test chaining two Conv3d layers on MPS autocast

Written by Codex, see https://chatgpt.com/codex/tasks/task_e_689b64192df883278648935963d2776d

Pull Request resolved: pytorch#160423
Approved by: https://github.com/dcci

pytorch#159130)

Fixes pytorch#159129

Pull Request resolved: pytorch#159130
Approved by: https://github.com/soulitzer

…h#160435)

remove aten::contiguous for NHWC convolutions on ROCm

Tests:
- nn/test_convolution.py::TestConvolutionNNDeviceTypeCUDA::test_conv_cudnn_nhwc_cuda_float32
- nn/test_convolution.py::TestConvolutionNNDeviceTypeCUDA::test_conv_cudnn_nhwc_cuda_float16

Before:
<img width="1255" height="228" alt="image"
src="https://github.com/user-attachments/assets/b125ccab-00c2-4d3a-a341-4583e51d8d57" />

After:
<img width="874" height="153" alt="image"
src="https://github.com/user-attachments/assets/ec200754-3622-488e-8762-bff1c2d22818" />

Pull Request resolved: pytorch#160435
Approved by: https://github.com/jeffdaily

`_transform_cuda_paths` intentionally includes the CUDA stubs folder.

However this path must not be added to the rpath as otherwise any CUDA command will fail at runtime with
> CUDA_ERROR_STUB_LIBRARY: "CUDA driver is a stub library"

This results in e.g. non-descriptive errors like
```
cutlass_library/source/tools/util/include/cutlass/util/device_memory.h:67  cutlass::device_memory::allocate: cudaMalloc failed: bytes=4096
terminate called after throwing an instance of 'cutlass::cuda_exception'
  what():  std::exception
```

Pull Request resolved: pytorch#160179
Approved by: https://github.com/jansel

Fixes pytorch#159616

Pull Request resolved: pytorch#159617
Approved by: https://github.com/lezcano, https://github.com/jansel

MPS backend does not support double, so errors should be different
Pull Request resolved: pytorch#160378
Approved by: https://github.com/dcci

* Use input vectorization for reduction_on_fastest_striding_dimension when dim0 >= 128

**Reproducer:**
```
import time
import torch

shapes = [
    (5079670, 128)
]

dims = [
    (1)
]

for i, shape in enumerate(shapes):
    x = torch.randn(shape, device='cuda', dtype=torch.float)
    for _ in range(10):
        w = torch.sum(x, dims[i])
    torch.cuda.synchronize()
    print(w.size())

    start_time = time.time()
    for _ in range(50):
        _ = torch.sum(x, dims[i])
    torch.cuda.synchronize()
    end_time = time.time()
    mean_time = (end_time - start_time)/50
    print(f"Avg time for shape {shape}: {mean_time * 1e6:.2f} us")
```

**Before (MI300X):**
Avg time for shape (5079670, 128): 1629.99 us

**After (MI300X)**
Avg time for shape (5079670, 128): 1008.59 us

Pull Request resolved: pytorch#160466
Approved by: https://github.com/petrex, https://github.com/jeffdaily

Fixes maintenance of triton packaging script when library versions change from one ROCm version to next.

Pull Request resolved: pytorch#158408
Approved by: https://github.com/jeffdaily

Co-authored-by: Ethan Wee <[email protected]>

…h#159824)

Follow up to pytorch#159580

Pull Request resolved: pytorch#159824
Approved by: https://github.com/williamwen42

…#160356)

Differential Revision: [D80035771](https://our.internmc.facebook.com/intern/diff/D80035771/)

The motivation and the original change is to reduce the number parameters we pass into the kernel, which was motivated by aesthetic reasons only.

But seeing the need to use different batch stride, we should just pass in the batch stride. That would be a good long term fix.

Pull Request resolved: pytorch#160356
Approved by: https://github.com/mlazos

Adds `OperatorEntry::getComputedKernelForDispatchKey` which returns the KernelFunction corresponding to `OperatorEntry.dispatchTable_[dispatch_ix]` for a given dispatch key
- Specifically it returns a `SafeKernelFunction` that holds a `KernelToken`. This `KernelToken` is registered to the `KernelFunction` in `OperatorEntry.kernels_` and will be invalidated when the `KernelFunction` is destructed (i.e. when the `AnnotatedKernel` that holds this `KernelFunction` is removed from `kernels_`, which happens when the corresponding impl is deregistered).
- `SafeKernelFunction` can be called via `callBoxed`, the validity of the token will be checked before this happens
- `SafeKernelFunction` is pybinded and `getComputedKernelForDispatchKey` is exposed to the frontend ia `torch.library.get_kernel`

Related to pytorch#155330

Pull Request resolved: pytorch#158393
Approved by: https://github.com/albanD

…_testing_IFU_08132025

# Conflicts:
#	.ci/docker/requirements-ci.txt
#	aten/src/ATen/Context.cpp
#	test/distributed/_tools/test_fsdp2_mem_tracker.py
#	test/dynamo/test_activation_checkpointing.py
#	test/dynamo/test_structured_trace.py
#	test/inductor/test_combo_kernels.py
#	torch/_higher_order_ops/triton_kernel_wrap.py
#	torch/_inductor/choices.py
#	torch/_inductor/codegen/triton.py