chore(release): 0.58.0 — GPU QLoRA correctness goes from theater to enforced - #2264
Merged
Conversation
…nforced Bumps the workspace 0.57.0 → 0.58.0 (root + 33 crate manifests + Cargo.lock) and ships the v0.58.0 CHANGELOG for the 8-commit delta: the QLoRA eval-correctness half of the cascade (#2257 stale-adapter val_loss, #2258 rank-aware-lr default-converge, #2260 CPU QKV-bias + 111x GPU-eval) plus the cross-silicon CUDA nightly lane (#2262/#2263) that now enforces the six GPU falsifiers on RTX 4090 sm_89 + GB10 Blackwell sm_121 (both legs green 2026-07-03). Also closes the v0.57 "GPU falsifiers aren't CI-enforced" doctrine gap in ROADMAP.md (now CI-enforced nightly on two silicons), upgrades the P3 pillar line, and marks CUDA-CI-NIGHTLY-001 completed in roadmap.yaml. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
v0.58.0 — the QLoRA correctness release
v0.57 made GPU QLoRA training run correctly; v0.58 makes the validation it reports trustworthy, and stands the whole correctness suite up as a nightly cross-silicon CI gate. Ships the 8-commit
v0.57.0..maindelta.Fixed
val_lossnow reflects the trained adapters (was byte-identical every epoch → froze best-epoch at 0)evaluate()now runs the GPU forward (111× faster/sample)Changed
apr finetune -m qlorano longer diverges #2258 rank-aware default lr — defaultapr finetune -m qlorano longer diverges (epoch avg 1.58 vs 11.18)CI
CUDA-CI-NIGHTLY-001; both legs green end-to-end (run 28661079713)This PR
cargo metadataresolves clean)[0.58.0]CUDA-CI-NIGHTLY-001marked completed in roadmap.yamlfmt clean ·
pv lint contracts/PASS (0 errors) · PMAT pre-commit gates pass.Post-merge: tag
v0.58.0+ GH release. crates.io cascade handled separately (dev-dep-cycle preflight per the known blocker).🤖 Generated with Claude Code