Repository navigation
Record partial actual fullweight FP32 optimizer evidence - #63
Merged
Merged
Conversation
2 of 5 tasks
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Partial actual-weight evidence
Record one actual pretrained rank0 FP32 optimizer update using the explicit CPU-state backend. Reloaded checkpoint has 294 CPU states / 3,952,722,584 bytes. Separate stock CUDA AdamW comparison found exact weights, moments and step counters across all 294 parameters. Saved checkpoint and all 294 expected shard hashes were independently rechecked for this portable receipt.
Limits retained
Original oracle completion reports 14,821,785,600 available host bytes, below the 16 GiB floor; comparison boundaries were unchecked. Exact optimizer arithmetic is proven for this step, while complete resource compliance, resume and BF16 remain unproven. Failed guarded attempts and the independent derivative failure remain preserved. Worker numerical_oracle_status=not_run and the later oracle receipt stay distinct. No production source, tolerances or existing acceptance flags changed. No issue closure claim.
Validation
Full pytest: 875 passed, 62 skipped in 45.07 seconds; required compileall, feature-matrix verification and snapshot/index source audits passed. SSH commit signature verified locally. Model/checkpoint/gradient binaries and private local paths excluded from Git.
Refs #3. Native training/resume/precision/branch/resource acceptance remains unfinished.