Categorize autograd engine, DDP reducer, and gradient helpers as BACKWARD#18
Closed
Chenguang-Zhu wants to merge 1 commit into
Closed
Categorize autograd engine, DDP reducer, and gradient helpers as BACKWARD#18Chenguang-Zhu wants to merge 1 commit into
Chenguang-Zhu wants to merge 1 commit into
Conversation
…WARD Summary: Three new BACKWARD rules in the categorization rule table, all using prefix matching against the demangled C++ symbol form (real frame names from PyTorch memory snapshots carry parameter signatures, so exact-name matching silently misses every real-world frame): 1. **C++ autograd engine / graph-task** — `torch::autograd::Engine::`, `python::PythonEngine::`, `Node::operator()`, `PyNode::`, `CppNode<`, `GraphTask::`, `AccumulateGrad::`, `deleteNode`, `utils::LambdaPostHook`. Account for a large share of allocations the old categorizer logged as UNKNOWN. 2. **DDP reducer bookkeeping** — `c10d::Reducer::mark_variable_ready` (also covers `_dense` overload), `mark_bucket_ready`, `autograd_hook`, `initialize_buckets`, `initialize_local_used_map`, `rebuild_buckets`, `set_static_graph`. Allreduce comm is categorized as NET separately. 3. **Grad-scale helpers** — `_instantiate_filtered_grads` (Distributed Shampoo's post-backward grad materialization). Differential Revision: D103992607
|
@Chenguang-Zhu has exported this pull request. If you are a Meta employee, you can view the originating Diff in D103992607. |
|
This pull request has been merged in 3155b30. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary:
Three new BACKWARD rules in the categorization rule table, all using prefix matching against the demangled C++ symbol form (real frame names from PyTorch memory snapshots carry parameter signatures, so exact-name matching silently misses every real-world frame):
C++ autograd engine / graph-task —
torch::autograd::Engine::,python::PythonEngine::,Node::operator(),PyNode::,CppNode<,GraphTask::,AccumulateGrad::,deleteNode,utils::LambdaPostHook. Account for a large share of allocations the old categorizer logged as UNKNOWN.DDP reducer bookkeeping —
c10d::Reducer::mark_variable_ready(also covers_denseoverload),mark_bucket_ready,autograd_hook,initialize_buckets,initialize_local_used_map,rebuild_buckets,set_static_graph. Allreduce comm is categorized as NET separately.Grad-scale helpers —
_instantiate_filtered_grads(Distributed Shampoo's post-backward grad materialization).Differential Revision: D103992607