Skip to content

vMLX 1.6.53

Choose a tag to compare

View all tags
@jjang-ai jjang-ai released this 04 Sep 14:44
· 69 commits to main since this release
v1.6.53
7ad06d7

vMLX 1.6.53

Changes

  • Publish vMLX 1.6.52 updater manifest (ddfa5f66)
  • feat(mtp): stock-MLX projection padding for native-MTP verify (default off) (d90e8758)
  • feat(mtp): parameterize the native-MTP depth ceiling for measurement (021f2451)
  • docs(mtp): record measured -38% full-graph loss of verify-pad dispatcher (23fcc9bf)
  • feat(mtp): complete depth-ceiling parameterization (verifier + profile caps) (7aae84f6)
  • fix(mtp): remove the residual depth-3 caps the first parameterization missed (98a85df7)
  • perf(mtp): default the qwen4_exp q4 proposal head ON (e2ae6bd8)
  • feat(mtp): q4 proposal head for the Qwen3.5/3.8 VLM MTP lane (default off) (edf713d7)
  • feat(mtp): one-time per-bundle proposal-head stamp, honored on later launches (bd279d74)
  • fix(mtp): confirm bundle type before stamping the proposal-head verdict (4f9a479b)
  • fix(mtp): stamp confirmation gates bundle TYPE only — layout truth is the loaded head (2693d6f5)
  • fix(mtp): require the JANG calibration marker before any proposal-head verdict (be8f9014)
  • fix(mtp): absolute imports in qwen4_exp prepare_mtp_draft_head — relative import crashed server startup under the mlx_vlm namespace (8bb04b01)
  • fix(mtp): a malformed eligible stamp is treated as absent, not a launch failure (09a03959)
  • feat(mtp): consume converter-calibrated proposal-head tensors from the stamp (187cd1dd)
  • feat(mtp): honor the converter draft_artifact schema with a single gated activation path (3fba93b7)
  • Revert the calibrated proposal-artifact consumption (keep it simple) (260623fa)
  • feat(panel): one-time dated redownload warning for Flash-Next MTP fix (84155f30)
  • feat(qwen4_exp): final Flash-Next MTP bundle contract (index-driven weights + calibrated sidecar) (d45aad0d)
  • feat(panel): MTP redownload warning localized + conformance-aware (ddf4a161)
  • fix(panel): Qwen3.8 MTP sessions default to FIXED depth 3, not adaptive (ed2e56d7)
  • fix(panel): language dropdown was clipped/covered — render in a body portal (b84f3a5d)
  • Prepare vMLX 1.6.53 (a1b9e2f5)
  • feat(mtp): policy-independent on-the-fly AR safety valve (windowed, context-scaled) (4974c9d4)
  • fix(mtp): log the pre-trip depth in the AR-safety demotion line (240abf88)
  • fix(mtp): AR-safety governor measures the real verify wall and ignores single stalls (fb44f2db)
  • fix(mtp): AR-safety context scale uses the cycle wall bounded by context growth (b8b4cb88)
  • docs: note the MTP AR safety valve in the 1.6.53 changelog (0495eef9)
  • fix(mtp): AR safety valve runs in the text lane too, shared with the multimodal lane (ef711efc)
  • perf(mtp): AR safety valve judges from cycle 16, no restored-prefix multiplier (6ff72661)
  • fix(mtp): AR safety baseline follows the cycle wall so batching cannot false-trip (7c8236ca)
  • fix(mtp): AR safety judges wall cost against the seed scaled by own-span growth (05aefe2f)
  • feat(mtp): reversible AR fallback — measured AR tier with MTP re-entry probes (5369c9be)
  • fix(mtp): AR safety baseline is a floor again; keep the last measured AR (80a05662)
  • docs: describe AR-safety re-entry in the 1.6.53 changelog (7d8fb2b4)
  • fix(mtp): a settled MTP re-entry that trips again counts a fallback and backs off (82a6e93a)
  • fix(mtp): key the prompt-priming sidecar at the N-1 boundary the prefix cache restores (4341c9b7)
  • test(mtp): priming tests follow the N-1 restore boundary contract (e7d9b3a7)
  • docs: note the priming boundary fix in the 1.6.53 changelog (7ad06d79)

Runtime provenance

  • vMLX source: 7ad06d79a744126a3d08869167dfda516c02e3d5
  • JANG runtime: f3c79081c3d99e8a26fa28444934f7567acb00c6 (jang 2.5.47)
  • Tahoe/macOS 26 is the default build; Sequoia is the compatibility build.
  • Both DMGs are Developer ID signed, independently notarized, stapled, and Gatekeeper checked by the production release scripts.

Downloads

  • Tahoe: vMLX-1.6.53-tahoe-arm64.dmg
    • SHA-256: af975564b0992b968080f22c195709b0b42f243e3a205e7885dea8c9c573707d
    • Size: 556191724 bytes
  • Sequoia: vMLX-1.6.53-sequoia-arm64.dmg
    • SHA-256: 078c9e8c4c2fd9dd82b92b1fbc8b2cc6fb669adfb04a53348f98e06587129517
    • Size: 534125852 bytes

Python distributions

  • Wheel: vmlx-1.6.53-py3-none-any.whl — SHA-256 4c954ff67b58d06611eed2fe917a66dd0de04649f977df286609edea4b135b8f
  • Source: vmlx-1.6.53.tar.gz — SHA-256 48cdbe226c80a5df0b86d8dcb6d6391e3cbcda960d661dd229ac0eef567b46a7
Assets 9
Loading