Skip to content

ScaleRLT: record three-family H20 CUDA update correctness - #33

Closed
0z5a wants to merge 1 commit into
ops/scalerlt-iquest-contractfrom
ops/scalerlt-cuda-correctness
Closed

0z5a wants to merge 1 commit into
ops/scalerlt-iquest-contractfrom
ops/scalerlt-cuda-correctness

Conversation

@0z5a

@0z5a 0z5a commented Oct 6, 2026

Copy link
Copy Markdown
Owner

The three-family prefix schedule now has an actual H20 CUDA check using MCore DDP and AdamW. This records the frozen twelve-case run for Ouro and Nanbeige ordinary/joint replay and Huginn independent/mixed latent groups, with token-mean and response-mean losses.

All twelve cases pass two paired updates, including full gradients, parameters, Adam moments and sample accounting, under the original tolerances. The Markdown table reports errors; the raw archive and independent offbox receipt are included with SHA-256 hashes. Verification took 22.66s; performance, official checkpoint learning and reward convergence remain unmeasured.

Validation: 12 passed, 36 warnings. All 84 payload hashes match, all recorded processes exited naturally, and a fresh check confirms their absence. Stack base: #32; this diff adds four evidence files.

Signed-off-by: 0z5a <dezhen.lu@student.uni-tuebingen.de>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant