pytorch

b1a83c4d - [memory history] cleanup recording API (#97406)

[memory history] cleanup recording API (#97406) This makes the options for recording memory history easier to understand and makes the default to record the most information. <!-- copilot:summary --> ### <samp>🤖 Generated by Copilot at 4706acf</samp> This pull request enhances the memory profiling and debugging capabilities of PyTorch on CUDA devices. It introduces a new API for memory history recording in `torch/cuda/memory.py` and `test/test_cuda.py`, and adds new functions for memory snapshot management and visualization in `torch/cuda/memory.py`. Also adds a quick _dump_snapshot function to make it easier to look at the common visualizations. <!-- copilot:walkthrough --> ### <samp>🤖 Generated by Copilot at 4706acf</samp> * Modify the `_record_memory_history` function to use a new API that accepts a string argument for the `enabled` parameter and more parameters to control the stack trace collection and memory event history ([link](https://github.com/pytorch/pytorch/pull/97406/files?diff=unified&w=0#diff-80bd98caafb20d758f45a4d23711810f7e0b9ce7a6505094f9dbb0e00a657377L620-R696)) * Add a new function `_dump_snapshot` that allows users to dump a memory snapshot to a directory with HTML plots of the memory segments and events ([link](https://github.com/pytorch/pytorch/pull/97406/files?diff=unified&w=0#diff-80bd98caafb20d758f45a4d23711810f7e0b9ce7a6505094f9dbb0e00a657377R703-R713)) * Update the test cases in `test/test_cuda.py` to use the new API for memory history recording and check the expected output of the memory plots ([link](https://github.com/pytorch/pytorch/pull/97406/files?diff=unified&w=0#diff-893b1eea27352f336f4cd832919e48d721e4e90186e63400b8596db6b82e7450L4946-R4946), [link](https://github.com/pytorch/pytorch/pull/97406/files?diff=unified&w=0#diff-893b1eea27352f336f4cd832919e48d721e4e90186e63400b8596db6b82e7450L4984-R4984), [link](https://github.com/pytorch/pytorch/pull/97406/files?diff=unified&w=0#diff-893b1eea27352f336f4cd832919e48d721e4e90186e63400b8596db6b82e7450L5000-R5000), [link](https://github.com/pytorch/pytorch/pull/97406/files?diff=unified&w=0#diff-893b1eea27352f336f4cd832919e48d721e4e90186e63400b8596db6b82e7450L5015-R5015), [link](https://github.com/pytorch/pytorch/pull/97406/files?diff=unified&w=0#diff-893b1eea27352f336f4cd832919e48d721e4e90186e63400b8596db6b82e7450L5035-R5038), [link](https://github.com/pytorch/pytorch/pull/97406/files?diff=unified&w=0#diff-893b1eea27352f336f4cd832919e48d721e4e90186e63400b8596db6b82e7450R5045-R5046), [link](https://github.com/pytorch/pytorch/pull/97406/files?diff=unified&w=0#diff-893b1eea27352f336f4cd832919e48d721e4e90186e63400b8596db6b82e7450L5060-R5059), [link](https://github.com/pytorch/pytorch/pull/97406/files?diff=unified&w=0#diff-893b1eea27352f336f4cd832919e48d721e4e90186e63400b8596db6b82e7450L5068-R5065), [link](https://github.com/pytorch/pytorch/pull/97406/files?diff=unified&w=0#diff-893b1eea27352f336f4cd832919e48d721e4e90186e63400b8596db6b82e7450L5088-R5085)) * Add missing imports and types to the `torch/cuda/memory.py` module ([link](https://github.com/pytorch/pytorch/pull/97406/files?diff=unified&w=0#diff-80bd98caafb20d758f45a4d23711810f7e0b9ce7a6505094f9dbb0e00a657377L5-R15)) Pull Request resolved: https://github.com/pytorch/pytorch/pull/97406 Approved by: https://github.com/ezyang