All pull requests

Pull requests

• Back off speculative drafting on consecutive rejected proposals

  #1168·emilianbold opened Oct 2, 2026Contributor·

  ·

• Add n-gram prompt-lookup speculative drafting

  #1167·emilianbold opened Oct 1, 2026Contributor·

  ·
  3

• Server: advertise loaded model modalities

  #1165·kyuz0 opened Oct 1, 2026Contributor·

  ·
  1

• Unterminated tool call: defer to the parse-failure recovery instead of erroring the turn

  #1163·nazerim opened Oct 1, 2026·

  ·

• server: report max_output_tokens as the Responses incomplete reason

  #1162·Laurianti opened Oct 1, 2026·

  ·
  1

• server: report Anthropic stop_sequence when a client stop matches

  #1161·Laurianti opened Oct 1, 2026·

  ·
  1

• Render GLM 5.3 low reasoning effort as Low

  #1160·chopchop-jiahao opened Oct 1, 2026·

  ·

• Fix GLM imatrix collector reading stale MoE buffers

  #1158·chopchop-jiahao opened Oct 1, 2026·

  ·

• Find the cold KV anchor in ChatML prompts

  #1157·paperniuk opened Sep 30, 2026·

  ·

• Fix the iq1_s and iq4_nl block sizes in the GGUF type table

  #1156·paperniuk opened Sep 30, 2026·

  ·
  1

• Fix misleading prefill progress and false live cache hits after a backend state reset

  #1155·emilianbold opened Sep 30, 2026Contributor·

  ·
  1

  1

• Metal: overlap Qwen3.8 PLE reads with prefill

  #1154·mvdbos opened Sep 29, 2026·

  ·

• server: --image-prefix-snapshot reuses the text before a request's image (#1007)

  #1153·denovich opened Sep 29, 2026·

  ·

• server: recover GLM argument types using existing property schemas

  #1152·gvocale opened Sep 29, 2026·

  ·
  1

• Metal: defer hyper-connection residual writes into the next stream norm

  #1150·Cuteshy opened Sep 29, 2026·

  ·
  1

• Metal: run Qwen3.8 QSA prefill attention on the tensor units

  #1149·Cuteshy opened Sep 29, 2026·

  ·

• Drop incorrect claims about model download resumability with the hf client

  #1148·nolith opened Sep 29, 2026·

  ·

• Recover GLM Metal prefill speed with validated selection prefixes

  #1147·emilianbold opened Sep 29, 2026Contributor·

  ·
  1

  1

• Refuse duplicate tensor names and name the file of a missing tensor

  #1145·Avicennasis opened Sep 28, 2026·

  ·

• Document the expert hotlist and locality profile formats

  #1144·Avicennasis opened Sep 28, 2026·

  ·

• Clamp the Linux host memory probe to the cgroup limit

  #1143·Avicennasis opened Sep 28, 2026·

  ·

• speed-bench: M4 Max 64 GB — Qwen3.8 Flash Next Q2 (full 2K–64K sweep) and Flash Q2 with SSD streaming

  #1142·pswai opened Sep 28, 2026·

  ·
  5

• Let Qwen and GLM make several tool calls per turn

  #1140·kernelpool opened Sep 28, 2026·

  ·
  1

• Reject malformed chunked HTTP request framing

  #1139·moritzburgard opened Sep 28, 2026·

  ·
  1

• cuda: make stream top-512 compaction decision uniform across warps

  #1138·flaccagora opened Sep 27, 2026·

  ·
  1