A few things to note:
1. Customers like vllm use a custom backend (e.g. VllmBackend), split the graph, and call standalone_compile for each split. If we let the bisector override the backend, we won't bisect thru the custom backend. `test_configs.bisect_keep_custom_backend_for_inductor` is used to keep the custom backend if we are bisecting for inductor.
2. pre_grad_graph bisecting and lowering bisecting so far does not compose well with each other since an issue may be just captured by the first one we try. `test_configs.bisect_pre_grad_graph` is used to enable the 'pre_grad_graph' bisecting.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/166344
Approved by: https://github.com/eellison
The gatherKthValue kernel had a race condition where multiple threads could write to the same output location without synchronization when duplicate k-th values exist, resulting in non-deterministic output.
Changes:
- aten/src/ATen/native/cuda/Sorting.cu: Use atomicMin with shared memory to deterministically find minimum index. Add early termination and remove redundant inRange checks. (We have to cast the index to `int32_t`, but this is already assumed to fit earlier in the kernel.)
- aten/src/ATen/native/cuda/Sorting.cpp: Remove non-deterministic alert since kthvalue is now deterministic on CUDA.
- torch/__init__.py: Remove kthvalue from non-deterministic operations list and remove kthvalue example from use_deterministic_algorithms() docstring.
- test/test_torch.py: Remove test_nondeterministic_alert_kthvalue since kthvalue no longer raises alerts on CUDA.
Benefits:
- Deterministic: always returns minimum index when duplicates exist
- Potential performance improvement on large arrays with repetitions
Test Results:
- All existing PyTorch tests pass (test_kthvalue)
- Custom determinism tests confirm consistent results
- Custom CUDA vs CPU correctness validated across 50+ scenarios
- Custom performance benchmarks show improvements with no visible regressions
Addresses #165227
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165762
Approved by: https://github.com/ngimel, https://github.com/eqy
This is follow-up of #165037. It generally recommended to use `is/is not` to compare types. Therefore this series of changes apply this suggestion in the code base, and it aims to finally enabling related linter checks.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/165142
Approved by: https://github.com/albanD
During 2.9 rc testing I am seeing an issue on Amazon Linux 2023 with CUDA 13.0 builds
This is related to:
https://github.com/pytorch/pytorch/issues/152756
Workflow: https://github.com/pytorch/test-infra/actions/runs/18324074610/job/52184079262
Error:
```
WARNING: There was an error checking the latest version of pip.
+ python3.11 .ci/pytorch/smoke_test/smoke_test.py --package torchonly
Traceback (most recent call last):
File "/usr/local/lib64/python3.11/site-packages/torch/__init__.py", line 333, in _load_global_deps
ctypes.CDLL(global_deps_lib_path, mode=ctypes.RTLD_GLOBAL)
File "/usr/lib64/python3.11/ctypes/__init__.py", line 376, in __init__
self._handle = _dlopen(self._name, mode)
^^^^^^^^^^^^^^^^^^^^^^^^^
OSError: libcudart.so.13: cannot open shared object file: No such file or directory
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/pytorch/pytorch/.ci/pytorch/smoke_test/smoke_test.py", line 12, in <module>
import torch
File "/usr/local/lib64/python3.11/site-packages/torch/__init__.py", line 425, in <module>
_load_global_deps()
File "/usr/local/lib64/python3.11/site-packages/torch/__init__.py", line 383, in _load_global_deps
_preload_cuda_deps(lib_folder, lib_name)
File "/usr/local/lib64/python3.11/site-packages/torch/__init__.py", line 317, in _preload_cuda_deps
raise ValueError(f"{lib_name} not found in the system path {sys.path}")
Traceback (most recent call last):
ValueError: libnvToolsExt.so.*[0-9] not found in the system path ['/pytorch/pytorch/.ci/pytorch/smoke_test', '/usr/lib64/python311.zip', '/usr/lib64/python3.11', '/usr/lib64/python3.11/lib-dynload', '/usr/local/lib64/python3.11/site-packages', '/usr/local/lib/python3.11/site-packages', '/usr/lib64/python3.11/site-packages', '/usr/lib/python3.11/site-packages']
File "/home/ec2-user/actions-runner/_work/test-infra/test-infra/test-infra/.github/scripts/run_with_env_secrets.py", line 102, in <module>
main()
File "/home/ec2-user/actions-runner/_work/test-infra/test-infra/test-infra/.github/scripts/run_with_env_secrets.py", line 98, in main
run_cmd_or_die(f"docker exec -t {container_name} /exec")
File "/home/ec2-user/actions-runner/_work/test-infra/test-infra/test-infra/.github/scripts/run_with_env_secrets.py", line 39, in run_cmd_or_die
raise RuntimeError(f"Command {cmd} failed with exit code {exit_code}")
RuntimeError: Command docker exec -t 7d9c5bd403cac9a9ee824d63a1d6f6057ecce89a7daa94a81617dbf8eff0ff2e /exec failed with exit code 1
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164870
Approved by: https://github.com/Camyll
Co-authored-by: Eli Uriegas <1700823+seemethere@users.noreply.github.com>
So this fixes at least two issues:
1) When we are invoking inductor backend, we apply pre-grad passes which try to find correct fake mode to use. In the nested case, we will run into clash when there is closure variable in the inductor region because non-strict would have fakified this variable before hand and inner torch.compile would have created a new fresh fake mode. This is not a problem in regular torch.compile because inner torch.compile gets ignored. I don't know if we are supposed to inherit fake mode from parent context in this case. But we can avoid this problem if we just default to eager backend which is fine in this case because the point of export is to capture aten operators. Going to inductor would mean we will lose inner torch.compile ops.
2) There is custom torch function modes in export that track number of torch fns executed and inner compile itself doesn't work because of guard failure as this mode state gets changed. I noticed torch.cond fixes this problem by carefully stashing the torch function mode and defer it in the backend. So the correct thing to do here is just re-use torch.cond implementation unconditionally.
So the things i did for fixing above were:
1) Always default to eager backend when compile is invoked inside export. I needed to make how torch.cond sets up the fresh tracing env into an util that can be shared.
2) The previous eager backend for torch.cond was wrong because the context managers didn't actually persist until the backend is invoked.
3) torch.cond used only disable TorchFunctionMetadata tf mode and stash it for later, but in fact, we should do both TorchFunctionMetadata and PreDispatchTorchFunctionMode.
With above fixes, we are able to export flex attention in export.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164171
Approved by: https://github.com/ydwu4
This PR is part of the work to deprecate torch::deploy in OSS. Effectively it does 3 things to get started.
1. Remove test_deploy_interaction as we no longer need to worry about this
2. Remove all torch._running_with_deploy checks and use the False path always (surfaced 1)
3. Remove `USE_DEPLOY` and switch to the default path always
Note: MyPy does fail on a bunch of things here as a bunch of older files are touched. It may be better to fix these things on a separate PR
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158288
Approved by: https://github.com/albanD
This PR is part of the work to deprecate torch::deploy in OSS. Effectively it does 3 things to get started.
1. Remove test_deploy_interaction as we no longer need to worry about this
2. Remove all torch._running_with_deploy checks and use the False path always (surfaced 1)
3. Remove `USE_DEPLOY` and switch to the default path always
Note: MyPy does fail on a bunch of things here as a bunch of older files are touched. It may be better to fix these things on a separate PR
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158288
Approved by: https://github.com/albanD
This PR is part of the work to deprecate torch::deploy in OSS. Effectively it does 3 things to get started.
1. Remove test_deploy_interaction as we no longer need to worry about this
2. Remove all torch._running_with_deploy checks and use the False path always (surfaced 1)
3. Remove `USE_DEPLOY` and switch to the default path always
Note: MyPy does fail on a bunch of things here as a bunch of older files are touched. It may be better to fix these things on a separate PR
Pull Request resolved: https://github.com/pytorch/pytorch/pull/158288
Approved by: https://github.com/albanD
The structure is
```
torch/
__init__.py
version.py
```
When we import torch, only `torch/__init__.py` is executed by default.
The submodules like `version.py` are not automatically imported or attached to the torch module.
So without anything in `__init__.py`, `torch.version` may not be found. So in this PR, we make the import explicit.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157584
Approved by: https://github.com/ezyang
#153622 introduced a hook for getting the relevant code objects after frame tracing. The idea is to have vLLM use this instead of monkey-patching `inline_call_()` to determine the source code files to hash. Unfortunately, the hook runs too late; the vLLM backend needs access to the set of source code filenames while it's running.
This PR replaces the newly-added hook with a utility function that a backend can call to get this information. I've made the change in vLLM and can verify that this allows the information to be queried at the right time.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155249
Approved by: https://github.com/zou3519
AMD is beginning to roll out ROCm distribution via Python wheels. This patch adds the `__init__.py` hook that is necessary to bootstrap ROCm correctly on Linux and Windows when built from these wheels.
See draft, developer documentation describing the mechanism here: https://github.com/ROCm/TheRock/blob/main/docs/packaging/python_packaging.md
This operates to similar effect as how Torch can depend on CUDA wheels, with some differences:
* ROCm libraries and checks are delegated to helpers in the `rocm_sdk` module, which knows how to find and configure access to the installed libraries. This limits the amount of plumbing and path machinations that must match up between the framework and ROCm.
* When building torch against ROCm, no ROCm system install is needed: instead the proper SDK development wheel is installed and the `CMAKE_PREFIX_PATH` is obtained via `rocm-sdk path --cmake`.
* It is expected that whoever produces such a build will also place a generated `_rocm_init.py` in the `torch` module with initialization logic to preload libraries, check versions, verify GPU compatibility, etc.
* See [build_prod_wheels.py](https://github.com/ROCm/TheRock/blob/main/external-builds/pytorch/build_prod_wheels.py) for an example build script that is being used to generate nightlies in this configuration.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/155285
Approved by: https://github.com/jeffdaily
Co-authored-by: Jeff Daily <jeff.daily@amd.com>
This PR:
* Expands `Hooks` with a new, optional `frame_traced_fn` field. It should be a callable receiving the list of traced code objects
* Maintains a list of `traced_code` objects in the `TracingContext` of an `OutputGraph`
* Whenever an `inline_call()` is encountered, the corresponding code object is added to this set
* `OutputGraph`'s associated `f_code` is added to the list just before the hook is called
I believe use of this hook should enable the source code hashing that vLLM does in a better way than monkey-patching `inline_call()`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153622
Approved by: https://github.com/jansel
### Changes
- Detect NVSHMEM install location via `sysconfig.get_path("purelib")`, which typically resolves to `<conda_env>/lib/python/site-packages`, and NVSHMEM include and lib live under `nvidia/nvshmem`
- Added link dir via `target_link_directories`
- Removed direct dependency on mlx5
- Added preload rule (following other other NVIDIA libs)
### Plan of Record
1. End user experience: link against NVSHMEM dynamically (NVSHMEM lib size is 100M, similar to NCCL, thus we'd like users to `pip install nvshmem` than torch carrying the bits)
2. Developer experience: at compile time, prefers wheel dependency than using Git submodule
General rule: submodule for small lib that torch can statically link with
If user pip install a lib, our CI build process should do the same, rather than building from Git submodule (just for its header, for example)
3. Keep `USE_NVSHMEM` to gate non-Linux platforms, like Windows, Mac
4. At configuration time, we should be able to detect whether nvshmem is available, if not, we don't build `NVSHMEMSymmetricMemory` at all.
For now, we have symbol dependency on two particular libs from NVSHMEM:
- libnvshmem_host.so: contains host side APIs;
- libnvshmem_device.a: contains device-side global variables AND device function impls.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/153010
Approved by: https://github.com/ngimel, https://github.com/fduwjj, https://github.com/Skylion007
Retry of https://github.com/pytorch/pytorch/pull/150957, which was reverted due to internal meta failures
Credit to @mgmtea who wrote the initial version of this PR: https://github.com/pytorch/pytorch/pull/146604
Context: CUPTI is the NVIDIA library that Kineto uses for collecting GPU-side info during profiling. The intended usage is to register a callback while you want profiling to occur, and then unregister the callback when you want profiling to stop. But a bug would cause crashes if CUPTI callbacks were de-registered when used with cudagraphs. The workaround was to disable "CUPTI_LAZY_REINIT" and "CUPTI_TEARDOWN" in Kineto - which prevents crashes, but can result in slower execution after profiling has occurred and completed.
This bug is believed to be fixed in CUDA >= 12.6, so this PR qualifies that DISABLE_CUPTI_LAZY_REINIT=1 and CUPTI_TEARDOWN=0 should only be applied if CUDA >= 12.6. Additionally, `profiler_allow_cudagraph_cupti_lazy_reinit_cuda12()` is added as an escape hatch so that we can add a killswitch in case we see more crashes related to this.
Differential Revision: [D72842114](https://our.internmc.facebook.com/intern/diff/D72842114/)
**NOTE FOR REVIEWERS**: This PR has internal Meta-specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D72842114/)!
Differential Revision: [D72842114](https://our.internmc.facebook.com/intern/diff/D72842114)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151124
Approved by: https://github.com/sraikund16
Credit to @mgmtea who wrote the initial version of this PR: https://github.com/pytorch/pytorch/pull/146604
Context: CUPTI is the NVIDIA library that Kineto uses for collecting GPU-side info during profiling. The intended usage is to register a callback while you want profiling to occur, and then unregister the callback when you want profiling to stop. But a bug would cause crashes if CUPTI callbacks were de-registered when used with cudagraphs. The workaround was to disable "CUPTI_LAZY_REINIT" and "CUPTI_TEARDOWN" in Kineto - which prevents crashes, but can result in slower execution after profiling has occurred and completed.
This bug is believed to be fixed in CUDA >= 12.6, so this PR qualifies that DISABLE_CUPTI_LAZY_REINIT=1 and CUPTI_TEARDOWN=0 should only be applied if CUDA >= 12.6. Additionally, `profiler_allow_cudagraph_cupti_lazy_reinit_cuda12()` is added as an escape hatch so that we can add a killswitch in case we see more crashes related to this.
Differential Revision: [D72745929](https://our.internmc.facebook.com/intern/diff/D72745929)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/150957
Approved by: https://github.com/aaronenyeshi, https://github.com/Skylion007
This PR remove the usage of guard_size_oblivious in vector_norm by inlining it in the runtime check,
this prevent any data dependent error from ever appearing here at the locations where guard_size_oblivious
used to exist. Before this PR it used to break potentially. This is NOT BC breaking or changing of semantics from eager.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/148809
Approved by: https://github.com/bobrenjc93
