All pull requests

Pull requests

• fix(ui): keep model job identities unique across simultaneous submissions

  #2560·rudycelekli opened Nov 2, 2026Contributor·

  ·

• qwen38 cuda chain: the MTP head on the card, on by default (decode 512 tok +34 %)

  #2559·JustVugg opened Nov 2, 2026Owner·

  ·

• fix(responses): cancel upstream work when clients disconnect before output

  #2558·rudycelekli opened Nov 2, 2026Contributor·

  ·

• api: preserve a single newline before a Qwen tool call

  #2557·bavanvrmk opened Nov 2, 2026·

  ·

• fix(responses): forward reasoning and structured output settings

  #2555·rudycelekli opened Nov 2, 2026Contributor·

  ·

• fix(api): select the full GLM-5.3 prompt contract from checkpoint metadata

  #2552·rudycelekli opened Nov 2, 2026Contributor·

  ·

• fix(api): keep Muse tool arguments valid JSON for nonfinite values

  #2551·rudycelekli opened Nov 2, 2026Contributor·

  ·

• fix(ui): preserve existing Pi providers on name collisions

  #2550·rudycelekli opened Nov 2, 2026Contributor·

  ·

• fix(tools): reject invalid routing counts before cross-validation

  #2549·rudycelekli opened Nov 2, 2026Contributor·

  ·

• fix(responses): preserve failed and incomplete upstream outcomes

  #2548·rudycelekli opened Nov 2, 2026Contributor·

  ·

• fix(responses): preserve forced function and custom tool selection

  #2547·rudycelekli opened Nov 2, 2026Contributor·

  ·

• Vulkan performance levers, default-on and probe-gated, the measured prefill shape self-tuned at defaults

  #2539·jameswitney opened Nov 1, 2026Contributor·5 tasks done·

  ·

• Detect Apple GPUs and enable explicit Metal setup

  #2538·bharadwajr567 opened Nov 1, 2026·

  ·

• qwen36: an attention-only model resumes where a prompt leaves the held state, as olmoe does

  #2537·dcoldeira opened Nov 1, 2026·

  ·
  2

• api: qwen3_coder keeps the model's own newline before a tool call, so prefix reuse holds

  #2535·dcoldeira opened Nov 1, 2026·

  ·
  3

• Mps backend

  #2533·pjsheini opened Nov 1, 2026·6 tasks·

  ·
  1

• Refactor nix code

  #2531·tfc opened Oct 31, 2026·6 tasks·

  ·

• glm53: add opt-in exact bounded nucleus selection

  #2518·appleguy opened Oct 30, 2026Contributor·

  ·

• feat: routed-expert parallelism across multiple CUDA devices (DeepSeek V4)

  #2487·xcjs opened Oct 29, 2026·

  ·

• fix(eval): isolate MMLU prompt mode and greedy temperature

  #2483·rudycelekli opened Oct 29, 2026Contributor·

  ·

• fix(eval): reject impossible MMLU subset sample counts

  #2482·rudycelekli opened Oct 29, 2026Contributor·

  ·

• fix(mcp): reject non-JSON constants before dispatch

  #2481·rudycelekli opened Oct 29, 2026Contributor·

  ·

• fix(chat): assemble SSE data fields before decoding JSON

  #2480·rudycelekli opened Oct 29, 2026Contributor·

  ·

• fix(chat): report SSE failures and premature response closure

  #2479·rudycelekli opened Oct 29, 2026Contributor·

  ·

• fix(chat): retain answer text alongside reasoning deltas

  #2478·rudycelekli opened Oct 29, 2026Contributor·

  ·