All pull requests

Pull requests

• First-token logprobs on the Gemma 4 assistant pair (depends on #530)

  #543·jvmenen opened Sep 26, 2026·

  ·

• Encode seam-less scoped chat renders turn by turn (depends on #533)

  #542·jvmenen opened Sep 26, 2026·

  ·

• Check Gemma 4 markers without building the vocabulary

  #541·jvmenen opened Sep 26, 2026·

  ·

• Skip blank retries for greedy decodes

  #540·jvmenen opened Sep 26, 2026·

  ·

• Do not retry a well-formed bare tool call as orphan tool markup

  #539·jvmenen opened Sep 26, 2026·

  ·

• Update transformers requirement from !=5.13.0,<5.15,>=5.10.0 to >=5.10.0,!=5.13.0,<5.18

  dependenciespython
  #538·dependabot[bot] opened Sep 26, 2026Bot·

  ·

• Cap pending idle-persistence bytes

  #536·vahako opened Sep 26, 2026·

  ·

• Tests no longer read the user's ~/.mtplx/config.toml or depend on the Mac's memory size

  #535·jvmenen opened Sep 26, 2026·

  ·

• Session bank: report why the RAM lane refused a shared prompt start

  #534·jvmenen opened Sep 25, 2026·

  ·

• Memoize segmented chat encodes per segment

  #533·jvmenen opened Sep 25, 2026·

  ·

• Prompt scoring: top-K per chunk without a full-vocabulary log-softmax

  #532·jvmenen opened Sep 25, 2026·

  ·

• Drop the answer's leading separator from streamed content when it arrives after the close tag

  #531·pierre427 opened Sep 25, 2026·

  ·

• First-token logprobs on /v1/completions and /v1/chat/completions

  #530·jvmenen opened Sep 25, 2026·

  ·

• Lane handover: solo MTP when alone, batched on arrival, and back (stacked on the Flash-Next batching PR)

  #523·freddyhaddad opened Sep 23, 2026·

  ·

• Batched AR lane for the Flash-Next (QSA) cache family (#420 follow-up)

  #522·freddyhaddad opened Sep 23, 2026·

  ·
  1

• feat: mtplx supervise - multi-model JIT loading, idle unload, self-healing restart, admission control and a management API

  #520·bpmforge opened Sep 22, 2026Contributor·

  ·
  1

• Splash engine: serve the MTPLX API over Inco's Apple-silicon runtime

  #514·ksg98 opened Sep 20, 2026·

  ·
  2

• Bump anyio from 4.13.0 to 4.14.2

  dependenciespython:uv
  #512·dependabot[bot] opened Sep 19, 2026Bot·

  ·

• perf(qwen4): overlap streamed PLE decode without changing sampled AR (stacked on #475)

  #501·davidtai opened Sep 15, 2026Contributor·

  ·
  3

• [FEAT] Bare-Speed kernel adaptations with arm FR-Spec and the M=4 stage-3 combine on the Qwen3.8 Flash-Next Q4/g64 pack

  #488·davidtai opened Sep 9, 2026Contributor·

  ·

• [FEAT] Speculative-cascade acceptance (opt-in, off by default): token-specific rule holds HumanEval at +29% decode; peak rule does not

  #485·davidtai opened Sep 8, 2026Contributor·

  ·
  1

• [FEAT] Typical Acceptance Verifier from Medusa2 - perf(qwen4): typical acceptance behind --typical-threshold (default off)

  #478·davidtai opened Sep 7, 2026Contributor·

  ·

• [FEAT] 4 unimplemented optimizations from PR391 plus 2 new ones - six serving optimizations for Qwen3.8 Flash-Next on MTPLX 2.11.2 (four PR-391 remainder optimizations + two exact decode optimizations)

  #475·davidtai opened Sep 6, 2026Contributor·

  ·
  3

• Desktop/CLI: default to 6875 (MTPL) and keep 8000 as fallback — avoid 8000 collisions

  #419·Fe2-O3 opened Sep 1, 2026·

  ·
  1

• perf(qsa): multi-threadgroup scoring, quantized pooled keys, and deep MTP throttling for 546k+ decode speed

  #413·maceip opened Aug 31, 2026Contributor·

  ·
  34