Memory snapshots from C++ #86190
Memory snapshots from C++ #86190
Conversation
Sometimes the driving process want to save memory snapshots but isn't Python. Add a simple API to turn it on without python stack traces. It still saves to the same format for the vizualization and summary scripts, using the C++ Pickler. [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/86190
Note: Links to docs will display an error until the docs builds have been completed. ✅ No Failures, 2 PendingAs of commit 871d880: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
oncall: jit
Sometimes the driving process want to save memory snapshots but isn't Python. Add a simple API to turn it on without python stack traces. It still saves to the same format for the vizualization and summary scripts, using the C++ Pickler. [ghstack-poisoned]
Sometimes the driving process want to save memory snapshots but isn't Python. Add a simple API to turn it on without python stack traces. It still saves to the same format for the vizualization and summary scripts, using the C++ Pickler. [ghstack-poisoned]
| std::unordered_map<std::string, uint32_t> memoized_globals_map_; | ||
| std::unordered_map<std::string, uint32_t> memoized_strings_map_; | ||
| std::unordered_map<std::string, uint32_t> memoized_devices_map_; | ||
| bool tag_aggregates_; |
There was a problem hiding this comment.
not sure what this field is doing
| try: | ||
| m.record(True) | ||
| t = torch.rand(311, 411, device='cuda') | ||
| mem = pickle.loads(m.do_snapshot()) |
torch/csrc/cuda/memory_snapshot.cpp
Outdated
| enabled ? blank_context : nullptr); | ||
| } | ||
|
|
||
| std::vector<char> _memory_snapshot_pickled() { |
There was a problem hiding this comment.
going into a vector is kinda weird but OK
There was a problem hiding this comment.
A std::string instead? I don't want it to be a file because these things might need to be written via API calls rather than directly to a file.
There was a problem hiding this comment.
going to a file stream would be the normal thing, but I think this is fine, we can reevaluate if this ends up being too big to store in memory (it shouldn't ever, tbh)
Sometimes the driving process want to save memory snapshots but isn't Python. Add a simple API to turn it on without python stack traces. It still saves to the same format for the vizualization and summary scripts, using the C++ Pickler. [ghstack-poisoned]
Sometimes the driving process want to save memory snapshots but isn't Python. Add a simple API to turn it on without python stack traces. It still saves to the same format for the vizualization and summary scripts, using the C++ Pickler. [ghstack-poisoned]
Sometimes the driving process want to save memory snapshots but isn't Python. Add a simple API to turn it on without python stack traces. It still saves to the same format for the vizualization and summary scripts, using the C++ Pickler. [ghstack-poisoned]
|
@pytorchbot merge |
|
@pytorchbot successfully started a merge job. Check the current status here. |
|
Hey @zdevito. |
Summary: Sometimes the driving process want to save memory snapshots but isn't Python. Add a simple API to turn it on without python stack traces. It still saves to the same format for the vizualization and summary scripts, using the C++ Pickler. Pull Request resolved: #86190 Approved by: https://github.com/ezyang Test Plan: contbuild & OSS CI, see https://hud.pytorch.org/commit/pytorch/pytorch/736adc08084814c13d57324f74adad091e304eb2 Reviewed By: seemethere Differential Revision: D40106329 fbshipit-source-id: 37912dc751d7e51cdcd0724623a9dde2617e4841
Stack from ghstack (oldest at bottom):
Sometimes the driving process want to save memory snapshots but isn't Python.
Add a simple API to turn it on without python stack traces. It still
saves to the same format for the vizualization and summary scripts, using
the C++ Pickler.