frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-21 13:44:15 +08:00
Code Issues Packages Projects Releases Wiki Activity
94,690 Commits 5,457 Branches 1,239 Tags
3255e7872bc94d95c63db844f4279d50884741d7
Commit Graph

6 Commits

Author SHA1 Message Date
bobrenjc93
db00e1699a [pc] introduce ProgressiveCompilationState and clear callback (#157619) 
followup from https://github.com/pytorch/pytorch/pull/157305 where
@aorenste correctly suggested clearing callback. this refactor
introduces a new dataclass so we don't need to check nullability for
each field

Pull Request resolved: https://github.com/pytorch/pytorch/pull/157619
Approved by: https://github.com/aorenste
ghstack dependencies: #157305, #157614
 2025-07-05 07:55:11 +00:00
bobrenjc93
5ea832e5f6 [pc] migrate progression futures from list to deque (#157614) 
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157614
Approved by: https://github.com/aorenste
ghstack dependencies: #157305
 2025-07-05 07:55:03 +00:00
bobrenjc93
d58ed04d89 [async-compile] add progressive compile mode (#157305) 
Pull Request resolved: https://github.com/pytorch/pytorch/pull/157305
Approved by: https://github.com/aorenste
 2025-07-04 04:18:50 +00:00
bobrenjc93
9d677389cb [async compile] make it more obvious that we support backwards (#157204) 
current failing with

```
(/home/bobren/local/a/pytorch-env) [13:02] devgpu009:/home/bobren/local/a/pytorch python test/inductor/test_compile_subprocess.py -k GPUTests.test_async
/home/bobren/local/a/pytorch/torch/backends/cudnn/__init__.py:115: UserWarning: PyTorch was compiled without cuDNN/MIOpen support. To use cuDNN/MIOpen, rebuild PyTorch making sure the library is visible to the build system.
  warnings.warn(
/home/bobren/local/a/pytorch/torch/_inductor/ops_handler.py:741: UserWarning: undefined OpHandler.__getstate__, please add missing op schema
  warnings.warn(f"undefined OpHandler.{name}, please add missing op schema")
/home/bobren/local/a/pytorch/torch/_inductor/ops_handler.py:741: UserWarning: undefined OpHandler.__getstate__, please add missing op schema
  warnings.warn(f"undefined OpHandler.{name}, please add missing op schema")
W0628 13:02:30.666000 3610483 torch/_inductor/compile_fx_ext.py:491] [0/0] Unable to pickle input graph or example inputs
W0628 13:02:30.666000 3610483 torch/_inductor/compile_fx_ext.py:491] [0/0] Traceback (most recent call last):
W0628 13:02:30.666000 3610483 torch/_inductor/compile_fx_ext.py:491] [0/0]   File "/home/bobren/local/a/pytorch/torch/_inductor/compile_fx_ext.py", line 484, in serialize_compile
W0628 13:02:30.666000 3610483 torch/_inductor/compile_fx_ext.py:491] [0/0]     ).serialize()
W0628 13:02:30.666000 3610483 torch/_inductor/compile_fx_ext.py:491] [0/0]   File "/home/bobren/local/a/pytorch/torch/_inductor/compile_fx_ext.py", line 210, in serialize
W0628 13:02:30.666000 3610483 torch/_inductor/compile_fx_ext.py:491] [0/0]     return _WireProtocolPickledInput(GraphPickler.dumps(self))
W0628 13:02:30.666000 3610483 torch/_inductor/compile_fx_ext.py:491] [0/0]   File "/home/bobren/local/a/pytorch/torch/fx/_graph_pickler.py", line 124, in dumps
W0628 13:02:30.666000 3610483 torch/_inductor/compile_fx_ext.py:491] [0/0]     pickler.dump(obj)
W0628 13:02:30.666000 3610483 torch/_inductor/compile_fx_ext.py:491] [0/0] AttributeError: Can't pickle local object 'make_opaque_bitwise_fn.<locals>.BitwiseFn'
```

Pull Request resolved: https://github.com/pytorch/pytorch/pull/157204
Approved by: https://github.com/aorenste
 2025-06-29 05:38:54 +00:00
James Wu
fe954cdcbf Use correct boxed_forward_device_index when running CompiledFxGraph.post_compile (#148130) 
This PR threads through the correct boxed_forward_device_index from graph_kwargs to CompiledFXGraph.post_compile. This allows us to correctly update BoxedDeviceIndex from cache hits.

We don't actually need to save `boxed_forward_device_index` in CompiledFXGraph because its value is in the cache key, so it always matches to the ambient one anyway. On forward with cudagraphs enabled, derive `boxed_forward_device_index`'s value from `device_idxs`.

Testing:

```
python benchmarks/dynamo/cachebench.py --mode training --benchmark torchbench --model BERT_pytorch --device cuda --repeat 1 --dynamic --output="dynamic.json"
```

Now cache hits properly on FXGraphCache. AOTAutogradCache has a guard failure. Will look into that as a followup.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/148130
Approved by: https://github.com/eellison
 2025-03-23 02:57:58 +00:00
Aaron Orenstein
2fcfae72b4 async fx compile (#146135) 
Adds the ability to run the selected out-of-process fx compile scheme in async mode - where we kick off the compile and then run eagerly until the compile is finished.

Added a test which runs a tiny model in a loop making sure that we execute it both eagerly and then compiled.

Differential Revision: [D71135546](https://our.internmc.facebook.com/intern/diff/D71135546)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/146135
Approved by: https://github.com/jamesjwu, https://github.com/jansel
 2025-03-19 14:07:51 +00:00