All pull requests

Pull requests

• Overlap streamed Qwen decode layers with their expert loads on Metal

  #1328·mschneider opened Nov 6, 2026·

  ·

• prompt prefix: close imported Qwen assistant turns once

  #1326·koxudaxi opened Nov 6, 2026·

  ·

• server: evaluate Qwen's <|im_end|> after a thinking answer

  #1325·koxudaxi opened Nov 6, 2026·

  ·

• metal: select the shader language version explicitly

  #1324·koxudaxi opened Nov 6, 2026·

  ·

• server: --kv-cache-cold-anchor last-user cuts the cold checkpoint before the prompt's final user turn

  #1323·maciekd opened Nov 5, 2026·

  ·

• server: let a request with an image use the disk KV cache

  #1322·maciekd opened Nov 5, 2026·

  ·

• DeepSeek V4.1: seed the SSD expert cache from a routing hotlist

  #1321·emilianbold opened Nov 5, 2026Contributor·

  ·

• MiniMax M3.1 on M3 Ultra: expert scales, tool calls, and DFlash2

  #1320·jjakemaness opened Nov 5, 2026Contributor·

  ·

• server: add --loop-guard to steer replies out of short-cycle repetition loops

  #1319·nomishbhardwaj opened Nov 4, 2026·

  ·

• cuda: run the DeepSeek V4.1 vision encoder

  #1315·francoisp opened Nov 3, 2026·

  ·

• CUDA: V4.1 routed experts on extra GPUs, parallel SSD expert reads, tested on 3090s with the 2bit quant

  #1313·francoisp opened Nov 2, 2026·

  ·

• ci: add a model-free pull request gate

  #1312·giuliowaitforitdavide opened Nov 1, 2026Contributor·

  ·
  1

• server: mask DeepSeek V4 (pre-4.1) chat syntax against control-token injection (stacked on #1288)

  #1310·Flor1an-B opened Nov 1, 2026·

  ·
  1

• Raise the server request body cap to 256 MiB

  #1309·ivanfioravanti opened Oct 31, 2026Contributor·

  ·
  4

• server: add -a/--alias to rename the OpenAI model id

  #1307·dineshr93 opened Oct 31, 2026·

  ·

• docs(jpeg): describe the malformed scan fixture in the fixture README

  #1306·giuliowaitforitdavide opened Oct 31, 2026Contributor·

  ·
  1

• server: accept OpenAI token-ID prompt arrays in /v1/completions

  #1304·potoior opened Oct 30, 2026·

  ·

• docs: add Jetson AGX Thor platform guide

  #1300·potoior opened Oct 29, 2026·2 tasks done·

  ·

• Accept published V4.1 DSpark metadata and Markov tensor names

  #1299·amittell opened Oct 28, 2026·

  ·
  1

• Add opt-in V4.1 decode timing breakdown (DS4_V41_STEP_TIMING, DS4_V41_LAYER_TIMING)

  #1298·amittell opened Oct 28, 2026·

  ·
  7

• server: render blank reasoning as absent (fixes full re-prefill after reasoning-free turns)

  #1296·timlawrenz opened Oct 28, 2026Contributor·

  ·
  2

• Agent: save exact token ids with KV sessions

  #1295·HDYS-TTBYS opened Oct 28, 2026Contributor·

  ·

• Add experimental native V4.1 weights for Metal SSD inference

  #1293·jomplox opened Oct 27, 2026·

  ·

• Add IQ1_M as a routed-expert tensor type

  #1291·yijunyu opened Oct 27, 2026·

  ·

• Metal: reuse f16 weight rows for 2-8 decode rows

  #1290·HDYS-TTBYS opened Oct 26, 2026Contributor·

  ·