All pull requests

Pull requests

• qwen38: a GGUF on the CUDA chain, the experts the card leaves in a host pool

  #2254·JustVugg opened Oct 20, 2026Owner·

  ·

• vulkan: DeepSeek V4's chain prompt rows may take the tiled GEMM

  #2253·jameswitney opened Oct 20, 2026·

  ·

• fix(benchmarks): force primary greedy temperature in UX runs

  #2181·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(benchmarks): propagate UX engine failures

  #2180·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(benchmarks): handle early EOS in serving interference

  #2179·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(benchmarks): exclude prefill boundary token from decode rate

  #2178·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(benchmarks): reap continuous batch engines after failures

  #2177·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(doctor): fail decision-model checks when memory is exhausted

  #2176·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(doctor): respect CPU selection in Vulkan placement

  #2175·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(experiments): detect hidden null and boolean configuration changes

  #2174·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(gguf): validate expert cardinality before narrowing

  #2173·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(image): validate checkpoint tensor layouts before planning

  #2172·rudycelekli opened Oct 17, 2026Contributor·2 of 3 tasks·

  ·

• fix(image): reject unpaired surrogate prompts before generation

  #2171·rudycelekli opened Oct 17, 2026Contributor·2 of 3 tasks·

  ·

• fix(image): bound PNG inflation to declared scanline bytes

  #2170·rudycelekli opened Oct 17, 2026Contributor·2 of 3 tasks·

  ·

• fix(image): count VAE weights as resident float32 allocations

  #2169·rudycelekli opened Oct 17, 2026Contributor·2 of 3 tasks·

  ·

• fix(setup): resolve pagination links within the hub origin

  #2168·rudycelekli opened Oct 17, 2026Contributor·2 of 3 tasks·

  ·

• fix(audio): reject non-finite WAV samples before DMel encoding

  #2167·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(audio): decode extensible WAV by its subformat GUID

  #2166·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(api): reject misplaced Anthropic tool blocks

  #2165·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(api): honor history JSON error policy at decoder limits

  #2164·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(mcp): keep stdio session alive after JSON nesting errors

  #2163·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(api): preserve trailing Kimi tool argument data in fallback

  #2162·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(api): preserve malformed MiMo tool calls instead of fabricating arguments

  #2161·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(usage): keep prior history when closing a save fails

  #2160·rudycelekli opened Oct 17, 2026Contributor·

  ·

• fix(qwen36): fall back to finite argmax on invalid logits

  #2159·rudycelekli opened Oct 17, 2026Contributor·

  ·