All pull requests

Pull requests

• feat(server): sampled requests speculate in batched DFlash2 chain rounds

  #910·Graffioh opened Oct 29, 2026Collaborator·

  ·

• fix(qwen35): concurrent engine penalises generated tokens only, like single-lane

  #909·Graffioh opened Oct 29, 2026Collaborator·

  ·

• fix(qwen35): sampled verify seeds the next round with the walk's own draw

  #908·Graffioh opened Oct 29, 2026Collaborator·

  ·

• fix(ggml-cuda,deepseek4): size integrated GPUs from what their allocations can take (stacked on #893)

  #907·Graffioh opened Oct 29, 2026Collaborator·

  ·

• feat(server): the model cards carry Lucebox's launch profiles and serve greedy by default

  #906·Graffioh opened Oct 29, 2026Collaborator·

  ·
  12

• feat(server): run tool-use checks on Windows and recover dropped tool calls

  #902·ElderOrb opened Oct 29, 2026Contributor·

  ·
  3

• fix(server): the built-in Qwen renderer writes an empty thinking block where the template does (stacked on #898)

  #900·Graffioh opened Oct 28, 2026Collaborator·

  ·

• fix(server): the built-in Qwen renderer writes one tool per line, as the template does (stacked on #898)

  #899·Graffioh opened Oct 28, 2026Collaborator·

  ·

• fix(server): the built-in Qwen renderer follows the model's template for past thinking (stacked on #896)

  #898·Graffioh opened Oct 28, 2026Collaborator·

  ·

• fix(server): continuous batching remembers a tool-call turn as it was generated (stacked on #895)

  #897·Graffioh opened Oct 28, 2026Collaborator·

  ·

• fix(server): the built-in Qwen renderer replays thinking after the last user message

  #896·Graffioh opened Oct 28, 2026Collaborator·

  ·

• fix(server): replay a thinking tool-call turn exactly as it was generated

  #895·Graffioh opened Oct 28, 2026Collaborator·

  ·
  3

• fix(ggml-cuda): fail managed buffers that RAM cannot back on integrated GPUs

  #893·Graffioh opened Oct 28, 2026Collaborator·

  ·
  13

• fix(qwen35): trim legacy device pools during long prefills by default

  #890·Graffioh opened Oct 27, 2026Collaborator·

  ·
  5

• fix(server): drop tool calls under tool_choice "none" without tool checks

  #885·Graffioh opened Oct 23, 2026Collaborator·

  ·
  3

• fix(server): post-close loop watchdog needs a 256-token loop, not one repeat

  #881·Graffioh opened Oct 22, 2026Collaborator·

  ·

• fix(qwen35): sampled-verify seeds with the miss draw after fast rollback

  #874·Graffioh opened Oct 22, 2026Collaborator·

  ·

• perf(draft): the batched draft step keeps its lanes packed

  #865·Graffioh opened Oct 20, 2026Collaborator·

  ·
  4

• feat(server): continuous batching fixes for agent subagents and greedy thinking

  #864·Graffioh opened Oct 20, 2026Collaborator·

  ·

• fix(gemma4): preserve sliding-window KV history during verification

  #824·pculaf opened Oct 7, 2026Contributor·

  ·

• perf(qwen4exp): enable guarded GPU argmax on gfx1151

  #823·dusterbloom opened Oct 7, 2026Collaborator·

  ·

• feat(vulkan): add native LFM2.5 MoE alongside Qwen

  #807·Amidwestnoob opened Oct 3, 2026Contributor·

  ·
  18

• Add JEV compatible interface

  #790·howard0su opened Sep 30, 2026Contributor·

  ·
  9

• feat(profiling): emit lucegraph concurrency capture in-tree

  #783·Graffioh opened Sep 29, 2026Collaborator·

  ·

• ci(e2e): run Qwen and DS4 end to end on lucebox3

  #756·Graffioh opened Sep 23, 2026Collaborator·4 tasks·

  ·
  24