Skip to content

vMLX v1.3.0

Choose a tag to compare

View all tags
@jjang-ai jjang-ai released this 20 Mar 18:55
· 4127 commits to main since this release
v1.3.0
60c80da

vMLX v1.3.0

Critical Bug Fixes

  • Metal crash on disk cache store (P0) — pre-materialize arrays on main thread before background write, preserves bfloat16
  • Paged cache layer mismatch (P1) — block reuse now checks cumulative SSM state for last-block position
  • Hybrid cache reconstruction (P1) — text scheduler applies _fix_hybrid_cache (was VLM-only)
  • Fresh-cache fallback detection — detects empty cache (offset=0), treats as miss instead of silent corruption
  • Reasoning OFF not working — _template_always_thinks() checks unclosed <think> only (Nemotron, MiniMax, all models)
  • Image generation interval leak — clearInterval in finally block
  • JANG VLM config detection — uses _find_config_path() instead of hardcoded path
  • JIT wake for /v1/rerank — added to inference endpoint list

New

  • Nemotron-H JANG support — gate dequant (8→2 bit), fc1/fc2 rename, MTP filter (42GB GPU, 46 tok/s)
  • Hybrid SSM full cache — prefix, paged, and disk caching work with Qwen3.5-A3B, Nemotron-H
  • Session status banners — loading, sleeping, stopped states in chat
  • Smooth token streaming — renderer-side typewriter animation (rAF)

Removed

  • All think-completion seed injection
  • Dead STREAM_THROTTLE_MS, _is_vlm_config(), vestigial weights param

Audit

  • 7-section 76-check audit: 0 FAIL across reasoning ON/OFF (11 models), Anthropic API, OpenAI API, streaming pipeline, cache system, JANG/Nemotron engine, sleep/wake/JIT
  • 2021 Python + 1545 panel = 3566 tests, 0 regressions

Install

pip install --upgrade vmlx

Desktop: MLX Studio v1.3.0

Assets 2
Loading