All pull requests

Pull requests

• qwen3_5 cuda: a prompt chunk nothing reads stops its last layer at its keys

  #3060·javierbits opened Nov 12, 2026Contributor·8 tasks done·

  ·
  1

• train 1112172818: #2851 #2765 #2883 #2889 #2891 #2892 #2878 (CI only)

  #3059·ashhart opened Nov 12, 2026Owner·

  ·

• lfm2_moe metal: prefill runs a chunk a call, so a new prompt no longer stalls other streams' rounds (#3021)

  #3058·ashhart opened Nov 12, 2026Owner·4 tasks done·

  ·

• fix(qwen): refuse unqualified SM110 serving modes

  #3057·yguanrong opened Nov 12, 2026Contributor·6 of 8 tasks·

  ·

• train 1112171357: #3010 #3011 #2695 #2865 #2726 #2838 #2854 #2864 #2852 #2877 (CI only)

  #3056·ashhart opened Nov 12, 2026Owner·

  ·

• Ling-3.0-tiny KDA on Metal using existing k3_kda (#1247)

  #3055·ThinkOffApp opened Nov 12, 2026Contributor·

  ·

• nemotron metal: prefill runs a chunk a call, so a new prompt no longer stalls other streams' rounds (#3021)

  #3054·ashhart opened Nov 12, 2026Owner·

  ·
  1

• hip flashnext: a lone stream chains up to seven MTP drafts when asked

  #3053·gilby opened Nov 12, 2026Contributor·

  ·

• XPU Qwen3.8-27B: a lone DFlash2 stream draws its tree's width from the lane rule's lone pricing

  #3052·Mirrdhyn opened Nov 12, 2026Contributor·

  ·

• xpu: serve logprobs and top_logprobs on Qwen3.8-27B

  #3051·Mirrdhyn opened Nov 12, 2026Contributor·

  ·
  1

• app: a packaged build ignores VITE_DEV_SERVER_URL

  #3049·ashhart opened Nov 12, 2026Owner·4 tasks done·

  ·

• server: idle connections give their slots up under pressure, one address holds a quarter at most

  #3048·ashhart opened Nov 12, 2026Owner·4 tasks done·

  ·
  1

• xpu families: refuse a token id at or past the embedding table before the lookup

  #3047·ashhart opened Nov 12, 2026Owner·5 of 8 tasks·

  ·
  1

• hip families: refuse a token id at or past the embedding table before the lookup

  #3046·ashhart opened Nov 12, 2026Owner·5 of 8 tasks·

  ·
  1

• metal families: refuse a token id at or past the embedding table before the lookup

  #3045·ashhart opened Nov 12, 2026Owner·5 of 8 tasks·

  ·
  1

• cuda families: refuse a token id at or past the embedding table before the lookup

  #3044·ashhart opened Nov 12, 2026Owner·5 of 8 tasks·

  ·
  1

• cluster pull: a file pulled whole is checked against the manifest's hash

  #3043·ashhart opened Nov 12, 2026Owner·3 of 4 tasks·

  ·

• hip ranks: a follower checks a step's counts against its message before using them

  #3042·ashhart opened Nov 12, 2026Owner·5 of 8 tasks·

  ·
  1

• ranks: every rank connection proves a shared key before its first protocol byte

  #3041·ashhart opened Nov 12, 2026Owner·5 of 8 tasks·

  ·
  1

• app agent: a child agent's loop refuses a tool outside its list

  #3040·ashhart opened Nov 12, 2026Owner·4 tasks done·

  ·

• cluster: a node takes only the nodes its file names and believes a node only about itself

  #3039·ashhart opened Nov 12, 2026Owner·3 of 4 tasks·

  ·
  1

• server: a grammar over 512 KiB is refused before it is compiled

  #3038·ashhart opened Nov 12, 2026Owner·3 of 4 tasks·

  ·
  1

• grammar budget: exploring a grammar stops at a work limit and a time limit

  #3037·ashhart opened Nov 12, 2026Owner·6 of 8 tasks·

  ·
  1

• mobile: a pairing link from outside the app is an offer the user accepts first

  #3036·ashhart opened Nov 12, 2026Owner·4 tasks done·

  ·

• server media: a host with more addresses than the lookup's queue holds resolves

  #3035·ashhart opened Nov 12, 2026Owner·3 of 4 tasks·

  ·