All pull requests

Pull requests

• deepseek-v4: two-GPU decode and VRAM-first expert lookups

  #2405·bobsummerwill opened Oct 26, 2026·6 tasks done·

  ·

• vulkan: COLI_VK_SUBGROUP=n pins the pipelines at subgroup size n

  #2404·bobsummerwill opened Oct 26, 2026·4 tasks done·

  ·
  1

• tests: test_vk_chain initializes the KV16 and host-expert scale fields

  #2403·bobsummerwill opened Oct 26, 2026·5 tasks done·

  ·

• vulkan: the uploader's host images align inside a calloc'd block, not posix_memalign

  #2402·bobsummerwill opened Oct 26, 2026·5 tasks done·

  ·
  1

• deepseek-v4: the decode mirror cache keeps its VRAM reserve between free-VRAM probes

  #2401·bobsummerwill opened Oct 26, 2026·6 tasks done·

  ·
  1

• test_setup_flow: the server cleanup waits for the stand-in to exit

  #2398·CoTi1938 opened Oct 26, 2026·3 of 6 tasks·

  ·

• resource_plan: a cgroup's active file cache is room too

  #2397·CoTi1938 opened Oct 26, 2026·3 of 6 tasks·

  ·

• compat_meminfo_gb: a container's cgroup limit bounds the RAM it reports

  #2396·CoTi1938 opened Oct 25, 2026·3 of 6 tasks·

  ·

• qwen38 CUDA chain: up to 128 promotions a forward, with a brake

  #2395·CoTi1938 opened Oct 25, 2026·6 tasks done·

  ·

• fix(api): keep an Optional string tool argument a string

  #2394·rudycelekli opened Oct 25, 2026Contributor·

  ·

• fix(api): keep a nullable Muse string argument a string

  #2393·rudycelekli opened Oct 25, 2026Contributor·

  ·

• qwen36 CUDA chain: the MTP head's draft picked on the device

  #2392·CoTi1938 opened Oct 25, 2026·6 tasks done·

  ·

• fix(serve): keep the running serve's pidfile when a second one fails

  #2391·rudycelekli opened Oct 25, 2026Contributor·3 tasks done·

  ·

• fix(stop): find a coli web launcher by its command line

  #2390·rudycelekli opened Oct 25, 2026Contributor·3 tasks done·

  ·

• fix(mcp): read the backend build log for logs which=build

  #2389·rudycelekli opened Oct 25, 2026Contributor·3 tasks done·

  ·

• fix(cluster): register a worker again after the coordinator restarts

  #2388·rudycelekli opened Oct 25, 2026Contributor·3 tasks done·

  ·

• fix(cli): open the dashboard on loopback when coli web binds 0.0.0.0

  #2387·rudycelekli opened Oct 25, 2026Contributor·3 tasks done·

  ·

• qwen36 CUDA chain: one kernel for small grid copies, one ids copy a frame

  #2385·CoTi1938 opened Oct 25, 2026·6 tasks done·

  ·

• Update manual download instruction in README

  #2354·sparshsingh12312 opened Oct 24, 2026·

  ·

• Fix numbering and wording in setup instructions

  #2353·sparshsingh12312 opened Oct 24, 2026·

  ·
  1

• qwen36: Ollama's qwen35moe GGUFs (head_count_kv per layer, ssm_dt, the embedded MTP head and vision tower)

  #2333·kreuzzelg opened Oct 23, 2026Contributor·

  ·
  3

• fix: make landing page responsive on mobile

  #2320·sparshsingh12312 opened Oct 23, 2026·6 tasks·

  ·
  7

• qwen36: keep a GGUF's dense matrices in memory through a prompt

  #2308·CoTi1938 opened Oct 22, 2026·8 of 9 tasks·

  ·

• qwen36: faster Metal kernels for GGUF blocks

  #2307·CoTi1938 opened Oct 22, 2026·9 of 10 tasks·

  ·

• qwen36: the Metal graph reads a GGUF

  #2306·CoTi1938 opened Oct 22, 2026·8 of 10 tasks·

  ·