Skip to content

chore: bump mlx-swift-lm to pick up TurboKV silent-data-loss fix - #166

Merged
solderzzc merged 1 commit into
mainfrom
chore/bump-mlx-swift-lm-turbokv-fix
Sep 5, 2026
Merged

chore: bump mlx-swift-lm to pick up TurboKV silent-data-loss fix#166
solderzzc merged 1 commit into
mainfrom
chore/bump-mlx-swift-lm-turbokv-fix

Conversation

@solderzzc

Copy link
Copy Markdown
Member

Summary

  • Bumps the mlx-swift-lm submodule to pick up SharpAI/mlx-swift-lm#63: compressedOffset wasn't reset across TurboKV eviction rounds, silently dropping whole 256-token windows from compressed KV history on any request past ~2560 tokens with default settings (turboMinActivationTokens=2048, hotWindowSize=256). No source changes needed on the SwiftLM side — this is purely the submodule pin.

Test plan

  • swift build --build-tests — clean build against the bumped submodule
  • swift test --skip-build --filter SwiftLMTests — 167 tests, 0 failures
  • swift test --skip-build --filter SwiftBuddyTests --disable-swift-testing — 127 tests (9 skipped), 0 failures

🤖 Generated with Claude Code

Pulls in SharpAI/mlx-swift-lm#63: compressedOffset wasn't reset across
TurboKV eviction rounds, silently dropping whole 256-token windows from
compressed KV history on any request past ~2560 tokens with default
settings. See that PR for the full root-cause writeup and the added
regression test.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
@solderzzc
solderzzc merged commit c70957c into main Sep 5, 2026
14 checks passed
@solderzzc
solderzzc deleted the chore/bump-mlx-swift-lm-turbokv-fix branch September 5, 2026 22:36
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