Train manual impl - #8
Draft
rgodha24 wants to merge 59 commits into
Draft
Conversation
Compress terminal rewards with a log scale, bound model outputs with tanh, and bump replay/training defaults so overnight runs reuse data more while keeping losses sane.
This makes it easier to see why training stalls in shallow openings and avoids overweighting old tree data when the same subtree is reused across moves.
This raises queue capacity for eager eval submission, lowers the default worker fanout, and surfaces outstanding eval pressure so queue wraparound risks are visible before they corrupt results.
This reverts commit df13b0c.
This reverts commit 7342047.
This reverts commit 6b27c11.
we preserve game state between training steps. we were, before this commit, deadlocking when we were preserving gpu job queue futures in flight over train boundaries. this fixes that issue by basically just completing all batches in the wait_for function on SelfPlay. essentially, after all the threads stop sending GPU eval requests, we clear out the queue and potentially run it with trash values to ensure that we never get into a weird state with a half/almost full queue.
run the value net directly inside alpha_paint with embedded exported weights, unpacked board features, and the training-style log terminal shaping so UBFM can use the model on cpu. Made-with: Cursor
This reverts commit 6b3d24a.
This reverts commit 3d163dd.
Align role-relative model outputs with white-centric search targets, fix terminal child labeling, and honor training model/LR config so self-play stops learning from contradictory values.
Track a single wall-clock budget across the whole turn and let elapsed time, not an old heuristic iteration estimate, stop Descent so model-backed search stays within the intended move budget.
Switch the default trainer to the 90k model and the faster 16x8/2048 setup, and shrink the production GPU queue batch to 128 so CUDA graph lane memory stays under control during longer runs.
Blend hill control, territory margin, and depth into terminal targets and normalize territory by playable cells so training and search get a less collapse-prone value signal on wall-heavy maps.
This reverts commit 7d5fd63.
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.
just ignore tbh