All pull requests

Pull requests

• feat(qwen3.8-mtp): speculative sampling for every request

  #1045·LeoBorcherding opened Nov 13, 2026·

  ·

• chore: refresh all engine libs (HRX and XRT)

  #1044·zaneni6 opened Nov 13, 2026Collaborator·

  ·
  1

• ci: run the oga-hipep unit tests on a hosted Windows runner

  #1043·Allegro-ma-non-troppo opened Nov 13, 2026Collaborator·

  ·
  1

• oga-hipep: keep up to four conversations warm, one generator each (WP-22)

  #1042·Allegro-ma-non-troppo opened Nov 13, 2026Collaborator·

  ·
  1

• fix(gemma4): current Gemma 4 chat template for E2B/E4B; 400 for non-object tool arguments (#975)

  #1039·zaneni6 opened Nov 12, 2026Collaborator·

  ·
  1

  1

• feat(qwen3.8-mtp): speculative sampling for non-greedy requests

  #1038·zaneni6 opened Nov 12, 2026Collaborator·

  ·
  1

• chore(deps): bump the github-actions group across 1 directory with 2 updates

  dependenciesgithub_actions
  #1028·dependabot[bot] opened Nov 12, 2026Bot·

  ·
  1

• fix(server): keep the reasoning between tool calls

  #981·NVolcz opened Nov 7, 2026Contributor·

  ·

• fix(server): stop merging consecutive user and system messages

  #980·NVolcz opened Nov 7, 2026Contributor·

  ·

• Flm aie next

  #937·zaneni6 opened Nov 4, 2026Collaborator·1 task·

  ·

• feat(hrx): add hrx::Runtime::shutdown()

  #795·quine00 opened Oct 15, 2026Contributor·

  ·
  3

  

• fix(qwen3.5): capture logprobs on the prefill draw when the think block is prefilled

  #780·manueldomke opened Oct 12, 2026·1 task done·

  ·
  2

• perf(qwen3.5): prefill the empty think block for single-turn models

  #779·manueldomke opened Oct 12, 2026·1 task done·

  ·
  2

• fix(qwen3.5): stop at max_tokens instead of one token past it

  #778·manueldomke opened Oct 12, 2026·1 task done·

  ·
  1

• feat(server): report logprobs when the request asks for them

  #777·manueldomke opened Oct 12, 2026·1 task done·

  ·
  1

• fix: don't count a discarded sample as generated text

  #776·manueldomke opened Oct 12, 2026·1 task done·

  ·
  1

• fix: drop a tool call that generation never finished

  #742·Javinator9889 opened Sep 18, 2026Contributor·

  ·
  1

  1

• fix: split the think block in non-streaming responses

  #740·Javinator9889 opened Sep 18, 2026Contributor·

  ·
  1

  1

• fix: send SSE headers before prefill and beat during it

  #735·Javinator9889 opened Sep 17, 2026Contributor·

  ·
  1

• fix: let an unchanged conversation hit the prompt cache

  #734·Javinator9889 opened Sep 17, 2026Contributor·

  ·
  2

• feat: granite 4.2 - kernels

  #701·vegah opened Sep 2, 2026·

  ·
  1

• feat: granite 4.2

  #700·vegah opened Sep 2, 2026·

  ·

• whisper: don't force a timestamp after 16 tokens — truncated every transcription (#698)

  #699·Anjielon opened Sep 1, 2026·

  ·
  1

  1

• docs(embeddinggemma): quantization is bf16, not Q4_1

  #693·Aitschend3251 opened Aug 30, 2026·

  ·
  5

• docs: remove stale OOC note from Gemma 4 benchmark

  #667·ei-grad opened Aug 12, 2026·1 task done·

  ·