Repository navigation
Conversation
Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Nanbeige inherited the differentiable prefix provider, but logical-step planning rejected its traces. Pass the actual provider family into the MCore prefix schedule and validate family, revision, depth and latent identity before reuse. Nanbeige now supports opt-in FP32 prefix training and the whole-depth joint replay while retaining its own loop normalization and attention math.
Stacked on #28 for ScaleRLT. Huginn latent-aware fallback and official online reward qualification remain separate work.
Validation: 188 CPU checks pass; 8 real MCore CUDA cases are skipped and 12 known strict Ouro BF16 xfails remain unchanged. In 24 Nanbeige dense/MCore-adapter/prefix update comparisons, maximum gradient relative L2 is 2.0740e-7 and parameter-delta error is 3.2783e-7. Another 24 B/joint Adam updates have zero gradient error and matching trainable Adam moments. Wrong/mixed model families are rejected. Raw collection failures from iCloud-evicted sources are retained; the final run uses all 490 SHA-verified files from the existing frozen MCore packet in the unchanged environment.
Result and speed comparison tables explicitly leave new CUDA speed/memory and online reward convergence unmeasured.