All pull requests

Pull requests

• cli: tensorfold plan, the serve startup budget before weights load (#840)

  #1319·jvr0x opened Oct 25, 2026Contributor·

  ·

• cuda tests: guard prompt positions in window rows fixtures

  #1318·Maxrio9 opened Oct 25, 2026·5 of 9 tasks·

  ·
  1

• server: --name-priority ID=background gives a served id a background default

  #1317·jvr0x opened Oct 25, 2026Contributor·

  ·

• l2_ref: preserve signed-zero interval endpoints

  #1315·Maxrio9 opened Oct 25, 2026·5 of 9 tasks·

  ·
  1

• build: server-cpu builds the fake server and engine-less CLI through the shared graph

  #1314·jvr0x opened Oct 25, 2026Contributor·

  ·
  1

• server tests: keep listener announcements visible to selectors

  #1308·Maxrio9 opened Oct 25, 2026·5 tasks done·

  ·
  1

• Qwen3.8-27B Metal: kept states reuse spare buffers, saves are not waited for, and slot 0's hand-over leaves the release path

  #1304·jschmied opened Oct 25, 2026Contributor·

  ·

• Qwen3.8-27B Metal: prompts fill by prefill_step chunks, so live streams keep their rounds

  #1303·jschmied opened Oct 25, 2026Contributor·

  ·

• Metal: streams admitted at load under the load limit, for Qwen3.5-2B, Nemotron, LFM2 and Gemma 4

  #1302·jschmied opened Oct 25, 2026Contributor·

  ·

• spark2_5: Spark-X2.5/2.6-4B on the CUDA engine (stacked on #1296)

  #1297·BobClawblaw opened Oct 25, 2026Contributor·9 tasks done·

  ·

• Muse Spark Open on CUDA: EXL3, DFlash2, two GPUs

  #1295·ajensenwaud opened Oct 25, 2026Contributor·8 of 9 tasks·

  ·
  1

• Metal: KV pages from the core pool, piloted on Qwen3.5-2B behind TF_QWEN35_METAL_KV_PAGES=1

  #1294·jschmied opened Oct 25, 2026Contributor·

  ·

• prompt cache: a miss keeps the state where the next prompt forks

  #1291·paragontasx opened Oct 25, 2026Contributor·

  ·
  2

• lane host: three fairness fixes to #1188's prompt steps

  #1287·jschmied opened Oct 25, 2026Contributor·

  ·
  1

• gemma4 CUDA: a stepped drafted pass keeps its mark after the drafter absorbed the chunk

  #1286·jschmied opened Oct 25, 2026Contributor·

  ·

• dory: full-attention KV in pages of the core pool

  #1284·jschmied opened Oct 25, 2026Contributor·8 tasks done·

  ·
  2

• copy index: a chain follows the earlier match that runs furthest back

  #1283·outcastofmusic opened Oct 25, 2026Contributor·9 tasks done·

  ·
  1

  1

• New model family: MiniMax-M3.1-Flash VL (minimax_m3_vl) EXL3, CUDA, TP2-served

  #1281·styles01 opened Oct 25, 2026·7 of 8 tasks·

  ·
  1

  1

• CONTRIBUTING: a speed claim attaches a Grill A/B/A2 receipt

  #1279·plotarmordev opened Oct 25, 2026Contributor·

  ·
  1

• server: expose per-request Drama speech sampling controls

  #1277·krunkosaurus opened Oct 25, 2026·4 of 5 tasks·

  ·
  1

• server: accept dialogue speaker tags (depends on #1277)

  #1278·krunkosaurus opened Oct 25, 2026·3 of 4 tasks·

  ·
  1

• cosmos3: Cosmos3-Edge reasoner on CUDA (stacked on #1271, proposal for #1269)

  #1272·drowzeys opened Oct 25, 2026Contributor·

  ·

• cosmos3: Cosmos3-Edge reasoner (image + text) on Metal (new family, proposal for #1269)

  #1271·drowzeys opened Oct 25, 2026Contributor·

  ·
  1

• App: release beta.10 with TensorFold 1.1.4

  #1255·ashhart opened Oct 24, 2026Owner·6 tasks done·

  ·
  1

• Notes 4/5: fetch more notes while the reply decodes

  #1254·ThinkOffApp opened Oct 24, 2026Contributor·6 of 9 tasks·

  ·