All pull requests

Pull requests

• qwen27 CUDA: pack weight tensors into 128 MiB blocks (~1 GiB back on a 24 GB card)

  #5·CanadaApollo6 opened Oct 13, 2026·

  ·

• qwen27 CUDA: a pre-sm_90 decode matmul, ~52% faster serving on an RTX 3090 (exact; stacked on #3)

  #4·CanadaApollo6 opened Oct 13, 2026·

  ·

• Qwen3.8-27B on CUDA: RTX 3090 (sm_86) support, the scratch sized by --parallel

  #3·CanadaApollo6 opened Oct 13, 2026·

  ·

• perf(qwen27 cuda): reduce TTFT with prompt caching and state reuse

  #2·jregan-beasley-smc opened Oct 11, 2026·

  ·

• feat(cuda): --keep-prefixes N for Qwen3.6-35B-A3B concurrent decoding (SEE-4099)

  #1·philip-pentatonic opened Oct 4, 2026·

  ·