All pull requests

Pull requests

• Session bank: anchor a GDN boundary at the end of the prompt head (opt-in, depends on #550)

  #559·jvmenen opened Sep 28, 2026·

  ·

• One-kernel MoE combine for Qwen3.5/3.6 MoE prefill on the invariant lane (opt-in, depends on #549)

  #558·jvmenen opened Sep 28, 2026·

  ·

• Finish the chat response before the session postcommit (opt-in)

  #557·jvmenen opened Sep 28, 2026·

  ·

• SSD cold tier: incremental encode reuses unchanged KV blocks (opt-in)

  #556·jvmenen opened Sep 28, 2026·

  ·

• Prefill draft-head history appends evaluate the cache only (opt-in)

  #555·jvmenen opened Sep 28, 2026·

  ·

• Session bank: postcommit restores do not refresh entry recency

  #554·jvmenen opened Sep 28, 2026·

  ·

• MTPLX_AGENT_REWRITES=off also disables the steering stream retries

  #553·jvmenen opened Sep 27, 2026·

  ·

• Stream recovery: the reasoning repair continues the retry prompt, and stats count every pass

  #552·jvmenen opened Sep 27, 2026·

  ·

• Prompt scoring on the Gemma 4 assistant pair

  #551·jvmenen opened Sep 27, 2026·

  ·

• In-forward GDN boundaries for Qwen3.5/3.6 on the invariant lane (depends on #549)

  #550·jvmenen opened Sep 27, 2026·

  ·

• Batch-invariant prefill for MoE models (opt-in)

  #549·jvmenen opened Sep 27, 2026·

  ·

• Thinking guard: opt-in chat scope (MTPLX_THINKING_GUARD_SCOPE=all)

  #545·bpmforge opened Sep 26, 2026Contributor·

  ·

• 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·

  ·