All pull requests

Pull requests

• Pipeline parallelism: wait for the gathered output on the CPU, not in a GPU command buffer

  #2542·AlexCheema opened Nov 1, 2026Contributor·

  ·

• Faster MiniMax M3 decode: sparse attention in three kernels, one-kernel routing

  #2541·AlexCheema opened Nov 1, 2026Contributor·

  ·

• Faster GLM-5.3-Flash: fused decode kernels, cheaper sparse attention, vocabulary-parallel lm_head

  #2540·AlexCheema opened Oct 31, 2026Contributor·

  ·
  1

• GLM-5.x: faster prefill and decode (sparse attention, fused indexer and MoE kernels)

  #2539·AlexCheema opened Oct 31, 2026Contributor·

  ·

• DeepSeek V4.1 Flash: fused decode kernels, Engram tables on the GPU, lossless speculative decoding

  #2538·AlexCheema opened Oct 31, 2026Contributor·

  ·

• perf(deepseek_v4): 1.5-1.6x decode and up to 2x prefill for DeepSeek V4, same correctness

  #2537·AlexCheema opened Oct 31, 2026Contributor·8 tasks done·

  ·

• Faster Hy3: fused decoder kernels, bit for bit, and one all-reduce per layer over the TCP ring

  #2535·AlexCheema opened Oct 31, 2026Contributor·

  ·
  1

• runner: keep the wired weights resident while the GPU idles

  #2534·AlexCheema opened Oct 31, 2026Contributor·

  ·

• Flint: fused decoder-layer kernels and sliding-window prefill, bit for bit

  #2533·AlexCheema opened Oct 31, 2026Contributor·

  ·
  1

• fix(tools): a string argument stays a string, as the model wrote it

  #2532·AlexCheema opened Oct 31, 2026Contributor·

  ·

• perf(laguna): faster Laguna S 2.1 and XS 2.1, bit-identical results

  #2531·AlexCheema opened Oct 31, 2026Contributor·

  ·

• Inkling-Small: fused kernels, ~1.5x decode, same outputs

  #2530·AlexCheema opened Oct 31, 2026Contributor·

  ·

• Faster Qwen3.8-Flash-Next: fused decode kernels, packed projections, overlapped n-gram reads

  #2529·AlexCheema opened Oct 31, 2026Contributor·

  ·

• Faster Qwen3.8-27B decode: fused Gated DeltaNet, attention and norm kernels

  #2528·AlexCheema opened Oct 31, 2026Contributor·

  ·

• fix(mlx): penalties count the latest token, and only the tokens they're defined over

  #2527·AlexCheema opened Oct 31, 2026Contributor·

  ·

• Faster Nemotron-H: fused float32 Mamba-2 and MoE kernels, and pipeline ranks that build their decode graph early

  #2526·AlexCheema opened Oct 31, 2026Contributor·

  ·

• Pipeline parallelism: pipeline a prompt's prefill from two chunks on, counting its last token

  #2525·AlexCheema opened Oct 31, 2026Contributor·

  ·

• perf(mlx): faster exact sampling and less per-token work in the decode engine

  #2524·AlexCheema opened Oct 31, 2026Contributor·

  ·

• mlx: keep MLX's memory limit under the GPU working set

  #2523·AlexCheema opened Oct 31, 2026Contributor·

  ·
  1

• Keep MLX float32 matmuls in full float32 in runners

  #2522·AlexCheema opened Oct 31, 2026Contributor·

  ·

• Pipeline parallelism: build a rank's decode graph while its input is still on the way

  #2521·AlexCheema opened Oct 30, 2026Contributor·

  ·

• fix(mlx): wait for a collective on the CPU before the GPU uses it

  #2510·AlexCheema opened Oct 30, 2026Contributor·

  ·

• fix(worker): finish every Shutdown task so they don't pile up in the state

  #2506·AlexCheema opened Oct 30, 2026Contributor·

  ·

• Recover malformed Qwen tool calls instead of emitting empty arguments

  #2504·darinlarimore opened Oct 29, 2026·

  ·

• Keep the instance alive when a client cancels a request

  #2503·darinlarimore opened Oct 29, 2026·

  ·