All pull requests

Pull requests

• [stage CI] 264 at 9fec7cb

  #27·LeoBorcherding opened Nov 6, 2026Owner·

  ·

• [stage CI] 251 at 184123b

  #26·LeoBorcherding opened Nov 6, 2026Owner·

  ·

• [stage CI] 240 at 9242345

  #25·LeoBorcherding opened Nov 6, 2026Owner·

  ·

• [fix for 251] common : fit: warn when --moe-cache-mib auto meets a user expert placement

  #24·LeoBorcherding opened Nov 6, 2026Owner·

  ·
  3

• [mirror #268] Host-weight prefetch in the scheduler, faster MoE cache prompt and decode, cache seeding under llama-server

  #15·LeoBorcherding opened Nov 2, 2026Owner·

  ·
  5

• [mirror #264] CUDA: faster MoE prompt processing (compact MMQ tiling, device expert sort, GET_ROWS)

  #14·LeoBorcherding opened Nov 2, 2026Owner·

  ·
  5

• [mirror #259] Show the MoE cache hit rate in perf output and llama-server /metrics

  #13·LeoBorcherding opened Nov 2, 2026Owner·

  ·
  13

• [mirror #251] common : add --moe-cache-mib auto to size the MoE expert cache from the fit

  #12·LeoBorcherding opened Nov 2, 2026Owner·

  ·
  13

• [mirror #240] qwen4exp: run MTP heads that share the target's token_embd and output

  #11·LeoBorcherding opened Nov 2, 2026Owner·

  ·
  7

• [mirror #238] cuda: grow the CUDA graph cache cap while it evicts graphs still in use

  #10·LeoBorcherding opened Nov 2, 2026Owner·

  ·
  5

• [mirror] unslothai/llama.cpp#345: MoE cache seeds a long first prompt chunk from the last prompt's routing

  #9·LeoBorcherding opened Nov 2, 2026Owner·

  ·
  3

• [mirror #321] common : --fit grows the ubatch of MoE models into spare VRAM, for up to 2.5x faster prompt processing

  #8·LeoBorcherding opened Oct 30, 2026Owner·

  ·
  6

• [needs design] Inkling: keep the memory-safe banded attention on for multi-slot, unified and non-CUDA caches

  #7·LeoBorcherding opened Oct 17, 2026Owner·

  ·
  3

• Fork RPC: refuse stock peers whose op numbers differ

  #6·LeoBorcherding opened Oct 17, 2026Owner·

  ·
  3

• Inkling chat handler: map "no thinking" to effort 0, and build continuations the way the template does

  #5·LeoBorcherding opened Oct 17, 2026Owner·

  ·
  3

• DiffusionGemma: count the prompt-KV store in the context auto-size, and fail the request instead of aborting

  #4·LeoBorcherding opened Oct 17, 2026Owner·

  ·
  7

• DiffusionGemma visual server: don't abort on multi-GPU hosts past about 2048 tokens

  #3·LeoBorcherding opened Oct 17, 2026Owner·

  ·
  6

• CUDA: honor the F32 precision Inkling asks for on the banded attention op

  #2·LeoBorcherding opened Oct 17, 2026Owner·

  ·
  3

• Prebuilt metadata: record source tarballs under the repo that actually hosts them

  #1·LeoBorcherding opened Oct 16, 2026Owner·

  ·
  3