All pull requests

Pull requests

• Batch K: Qwen-Image-2.1 from unsloth's GGUFs (base and Turbo) and the official Turbo

  #2736·JustVugg opened Nov 10, 2026Owner·

  ·

• tests: run C test binaries by absolute path

  #2735·tylnexttime opened Nov 10, 2026·3 of 5 tasks·

  ·

• deepseek-v41: join a message's content blocks on a blank line, as encoding.py does

  #2732·KairaCode opened Nov 9, 2026·2 of 3 tasks·

  ·

• tests: build the CUDA chain tests with nvcc's MSVC host (setenv shim)

  #2731·KairaCode opened Nov 9, 2026·3 of 4 tasks·

  ·

• coli: forward --ram as RAM_GB to qwen38 too (follow-up to #2356)

  #2730·KairaCode opened Nov 9, 2026·

  ·

• deepseek-v4: upload the exact CUDA head through a 1024-row staging buffer

  #2729·KairaCode opened Nov 9, 2026·3 of 4 tasks·

  ·

• deepseek-v4: cap the batched MoE union's loader lanes at the layer's lease capacity

  #2728·KairaCode opened Nov 9, 2026·2 of 4 tasks·

  ·

• Local inference toolkit: eval, NPU embeddings, benchmarks, training

  #2720·Liso7150 opened Nov 9, 2026·7 tasks done·

  ·

• release: a linux-aarch64 archive for Raspberry Pi and other ARM boards

  #2714·ZacharyZcR opened Nov 8, 2026Contributor·

  ·

• dsv41: the DeepInfer-style packed KV cache, on by default -- COLI_VK_KV_FP32=1 is the exact fp32 opt-out

  #2712·jameswitney opened Nov 8, 2026Contributor·

  ·
  5

• Recover unresponsive text engines after request deadlines

  #2653·ZacharyZcR opened Nov 6, 2026Contributor·

  ·
  4

• dsv41: trunk-first VRAM placement — place trunk first before Expert Cache

  #2637·jameswitney opened Nov 6, 2026Contributor·

  ·
  3

• deepseek-v4: AVX2 router, bit-exact with the scalar route (V4_ROUTER_AVX2)

  #2636·KairaCode opened Nov 6, 2026·3 of 5 tasks·

  ·
  1

• deepseek_v41: carry the published index list across the prompt tail

  #2618·lBroth opened Nov 5, 2026Contributor·

  ·
  1

• fix(windows): point at libgomp when -lgomp is missing under MSYS2 gcc 16

  #2617·Marouane-elfadli opened Nov 5, 2026·2 of 6 tasks·

  ·
  1

  2

• DeepSeek V4: native HIP tier and a single GPU placement plan (RX 7600M XT, 8 GB)

  #2603·kflirik opened Nov 5, 2026·

  ·

• Add native IQ2_XXS Metal GEMV acceleration

  #2602·kreem22 opened Nov 5, 2026·

  ·

• coli gateway: Ollama's API in front of coli serve, for Open WebUI and the other Ollama clients

  #2597·kreuzzelg opened Nov 4, 2026Contributor·

  ·

• Add file tools to the dashboard chat (read, write, edit, list, run)

  #2582·linuxrebel opened Nov 4, 2026·2 of 6 tasks·

  ·
  4

• api: preserve a single newline before a Qwen tool call

  #2557·bavanvrmk opened Nov 2, 2026·

  ·
  1

• Detect Apple GPUs and enable explicit Metal setup

  #2538·bharadwajr567 opened Nov 1, 2026·

  ·

• Mps backend

  #2533·pjsheini opened Nov 1, 2026·6 tasks·

  ·
  1

• Refactor nix code

  #2531·tfc opened Oct 31, 2026·6 tasks·

  ·

• feat: routed-expert parallelism across multiple CUDA devices (DeepSeek V4)

  #2487·xcjs opened Oct 29, 2026·

  ·
  1

• fix(eval): isolate MMLU prompt mode and greedy temperature

  #2483·rudycelekli opened Oct 29, 2026Contributor·

  ·