Repository navigation
Add the Sophia optimizer - #550
Open
shaneraphel wants to merge 4 commits into
Open
shaneraphel wants to merge 4 commits into
shaneraphel wants to merge 4 commits into
Conversation
get_trace averaged the Hessian diagonal over the spatial dims only for 4D kernels. Conv1d (3D) and Conv3d (5D) weights left tmp_output unbound. Average over dims 2..ndim-1 instead; 4D is unchanged.
Each mode factor of an order-k tensor enters to the power -1/(2k): -1/4 per side for matrices. The code used -1/k, twice the published exponent. The 7 pre-existing test_optimizer Shampoo failures are a state-dict precision issue on this torch version and fail identically without this change.
warmup_rate (default 1e-6) replaces the hard-coded warm-up slope and min_step_size (default 1e-2) the floor. Old checkpoints without the new group keys fall back to the old values. The 7 pre-existing test_optimizer Adafactor failures are a state-dict precision issue on this torch version and fail identically without this change.
Second-order optimizer with a diagonal Hessian estimate (SophiaG). Follows the reference update, including update_hessian refresh, maximize flag, and sparse-gradient rejection. CUDA-graph capturable execution is not ported. Wired into the test harness lists and README.
This was referenced Sep 25, 2026
This branch has not been deployed
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Adds
Sophia(SophiaG), the scalable stochastic second-order optimizer from https://arxiv.org/abs/2305.14342. The port follows the reference implementation at https://github.com/Liuhong99/Sophia:update_hessian()(EMA withbeta2overgrad * grad)sign(m) * min(|m| / (rho * bs * h), 1)with decoupled-style weight decayp *= 1 - lr * wdmaximizeflag, sparse-gradient rejection,__setstate__defaults for old checkpointsCUDA-graph capturable execution is intentionally not ported. Without any
update_hessian()call the estimate stays zero and the update degrades to a signed momentum step, same as the reference.Fixes #501.
Test plan
pytest tests/test_sophia.py: 5 passed — update matches a hand-computed reference step, Hessian EMA accumulates, no-Hessian step equals a sign step, sparse rejected, maximize flips the signpytest tests/test_optimizer_with_nn.py -k Sophia: 1 passed (logistic regression loss decreases)pytest tests/test_param_validation.py -k Sophia: 2 passedtests/test_optimizer.py -k Sophia: 7 failures, all in the state-dict round-trip assertion. Existing optimizers fail the same assertions on torch 2.11 (verified Lion, Apollo, Adahessian on the clean tree):load_state_dictshares state tensors with the input dict, so the "snapshot" mutates during the parallel run. Pre-existing harness issue, out of scope here.