All pull requests

Pull requests

• DeepSeek V4.1: make layer-major prefill chunk boundaries saveable

  #1345·emilianbold opened Nov 11, 2026Contributor·

  ·

• Keep the images of a conversation in the embedding cache

  #1344·fitchmultz opened Nov 10, 2026·

  ·
  1

• Save image conversations in the disk KV cache

  #1343·fitchmultz opened Nov 10, 2026·

  ·

• server: write a continued KV checkpoint at the first chunk past each step

  #1342·emilianbold opened Nov 10, 2026Contributor·

  ·

• server: align the continued KV checkpoint step with the prefill chunk schedule

  #1340·emilianbold opened Nov 10, 2026Contributor·

  ·

• Speed up DeepSeek V4 Flash prefill on ROCm gfx1151 (Strix Halo)

  #1339·felix920506 opened Nov 10, 2026·

  ·

• Copy model tensors into device memory on a discrete CUDA GPU when they fit

  #1338·n7olkachev opened Nov 9, 2026·

  ·

• Stage GLM decode attention scores in global memory past 12,032 rows

  #1337·n7olkachev opened Nov 9, 2026·

  ·
  1

• Clear transient buffers on shared-event drain success paths

  #1331·nazerim opened Nov 8, 2026·

  ·
  1

• 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·

  ·
  2

• 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