Repository navigation

Expose bundle-aware speculative defaults before model load - #3029

Merged

jjang-ai merged 35 commits into

mainfrom

review/qwen-spec-defaults-oct7

Oct 8, 2026

Commits on Oct 6, 2026
- Pin Flash native MTP verification corrections
  Eric
  committed
  fedb853
- Pin vmlx-swift perf/claude-swift-port-oct6 for speed proof
  Eric
  committed
  d08eb33
Commits on Oct 7, 2026
- Speculation on by default for Flash-Next and bundled 27B DFlash2
  Eric
  committed
  de69800
- Pin vmlx-swift 0e26bcb5: Qwen3.8-27B-JANGH2 loads
  Eric
  committed
  f154244
- Integrate current schema support before speculative runtime qualification
  Eric
  committed
  2fdd211
- Honor AR across loaded drafters and expose bundled speculation before load
  Eric
  committed
  97cb72b
- Describe bundle-aware speculative defaults and Off behavior
  Eric
  committed
  ba87ee9
- Show compatible DFlash capability consistently before model loading
  Eric
  committed
  d7287a2
- Clarify advisory capacity diagnostics without changing explicit limits
  Eric
  committed
  5394dc1
- Migrate only provenance-tagged defaults to bundle-aware speculation
  Eric
  committed
  7119be8
- Expect bundle defaults only for fresh or provenance-owned settings
  Eric
  committed
  0c12a53
- Use complete metadata fixture for bundled drafter detection
  Eric
  committed
  f3be35b
- Describe bundled DFlash selection accurately in picker help
  Eric
  committed
  23a4958
- Restore bundle policy when resetting speculative decoding
  Eric
  committed
  136aac4
- Reject non-string speculative option values
  Eric
  committed
  4590c61
- Test picker reset persists family speculation default
  Eric
  committed
  7c9e4a2
- Align local model guides with bundle-aware speculation defaults
  Eric
  committed
  d7967fb
- Import engine metadata API in preload fixture tests
  Eric
  committed
  94d9bce
- Release unshared DFlash drafters after model teardown
  Eric
  committed
  0370584
- Pin reviewed Qwen defaults and complete drafter validation
  Eric
  committed
  e50fa3f
- Pin reviewed Qwen runtime portability fixes
  Eric
  committed
  ffb03f2
- Pin vmlx-swift a7123840: sampled MTP, DFlash 2 trees, media speculation
  Eric
  committed
  3627d4f
- Keep the admitted scratch floor in MLX's freed-buffer pool
  Eric
  committed
  492c66f
- Pin vmlx-swift e40f9fa4: media conversation restore, direct 8-bit lane, DFlash 2 AR width
  Eric
  committed
  dde94c0
- Pin vmlx-swift a5a0689c (docs-only over the live-proven e40f9fa4)
  Eric
  committed
  53231ce
- Pin vmlx-swift 90eecf3c (JANGH dense split-K verify tile)
  Eric
  committed
  5380ba8
- Pin vmlx-swift e54e7388 (DFlash 2 drafter cache after prefix restore)
  Eric
  committed
  4c9e5e4
- Reload when leaving or entering the speculative family default
  Eric
  committed
  f41de0a
- fix: avoid mislabeling DFlash completion telemetry as plain decode
  Eric
  committed
  9167afe
- build: pin audited native MTP cache boundary candidate
  Eric
  committed
  0a4428f
- fix: inspect bundle architecture before native MTP exemptions
  Eric
  committed
  a96dae9
- Delegation usage reports the runtime's prompt tokens
  Eric
  committed
  d546b3a
- Pin vmlx-swift dc139e2c (native MTP head completeness)
  Eric
  committed
  f6700d3
- Pin vmlx-swift 7c5a8a08 (n-gram on SSD, residency on GPU weights)
  Eric
  committed
  1383948
- Pin Qwen speculative runtime to merged engine audit
  Eric
  committed
  15e917f

Commits

Commits on Oct 6, 2026

Commits on Oct 7, 2026