Pull requests

• vulkan: copy a misaligned top_k input to an aligned scratch instead of aborting

  #328·danielhanchen opened Oct 26, 2026Member·

  ·
  6

• server: preserve speculative drafts across sub-batch retries

  #326·Imagineer99 opened Oct 25, 2026Member·

  ·
  1

• model: add MiniMax-M3.1

  #320·danielhanchen opened Oct 23, 2026Member·

  ·
  6

• model : project mamba1, plamo2 and minimax-01 inputs in 2-D before the per-sequence reshape

  #319·danielhanchen opened Oct 23, 2026Member·

  ·
  1

• model : Beam applies its attention-gate and FFN scales; converter loads the tokenizer on transformers 4.x

  #317·danielhanchen opened Oct 22, 2026Member·

  ·
  1

• sampling: backend samplers match the CPU samplers, also under draft verification; ggml-alloc re-plans when a tensor becomes an output

  #316·danielhanchen opened Oct 21, 2026Member·

  ·
  7

• CUDA: FA MMA in-kernel K/V dequant for all quantized K/V types, head dims 64/96/192 and MLA, and Turing

  #315·danielhanchen opened Oct 21, 2026Member·

  ·
  2

• Model arch: Flint

  #314·danielhanchen opened Oct 21, 2026Member·

  ·
  2

• ggml-backend, llama: drop most host syncs per decode step (async sched inputs, lazy getter syncs)

  #313·danielhanchen opened Oct 21, 2026Member·

  ·
  2

• chat : end Beam analysis on <|constrain|> header variants

  #311·danielhanchen opened Oct 21, 2026Member·

  ·
  9

• rpc : one send per message and an optional mutual TLS 1.3 transport

  #309·danielhanchen opened Oct 21, 2026Member·

  ·
  29

• pr-set: pin qwen4exp DFlash layer-input fix (#307)

  #308·danielhanchen opened Oct 21, 2026Member·

  ·
  1

• qwen4exp : give drafters the mean of the hyper-connection streams as layer input

  #307·danielhanchen opened Oct 21, 2026Member·

  ·
  1

  11

• pr-set: pin #303 (CUDA FA MMA in-kernel q4_0/q8_0 K/V dequant for 3-16 tokens)

  #306·danielhanchen opened Oct 21, 2026Member·

  ·
  3

• pr-set: pin CUDA IQ1_M MMQ (#304)

  #305·danielhanchen opened Oct 21, 2026Member·

  ·
  1

• CUDA: add IQ1_M to MMQ

  #304·danielhanchen opened Oct 21, 2026Member·

  ·
  5

• CUDA: dequantize q4_0/q8_0 K/V tiles in the FA MMA kernel for small multi-sequence batches (pin branch on b11642)

  #303·danielhanchen opened Oct 20, 2026Member·

  ·
  3

• model : DeepSeek-V4.1-Flash-1017 full stack pin (all DeepSeek-V4.1 PRs except #292)

  #298·danielhanchen opened Oct 19, 2026Member·

  ·
  9

• mtmd : DeepSeek-V4.1-Flash vision (deepseek41v)

  #297·danielhanchen opened Oct 19, 2026Member·

  ·
  2

• model : deepseek41 DSpark speculative draft

  #296·danielhanchen opened Oct 19, 2026Member·

  ·
  3

• model : deepseek41 sparse attention, per-device KV mirrors, host gather

  #295·danielhanchen opened Oct 19, 2026Member·

  ·
  1

• chat : DeepSeek-V4.1 chat template, detection and DSML tool calls

  #294·danielhanchen opened Oct 19, 2026Member·

  ·
  11

• model : deepseek41 quantized K cache, shared prompt tokens, Engram host placement

  #293·danielhanchen opened Oct 19, 2026Member·

  ·
  4

• ggml-cuda : deterministic radix top-k, honour GGML_PREC_F32 for F32 x F32 MUL_MAT

  #292·danielhanchen opened Oct 19, 2026Member·

  ·
  4

• ggml-cpu : AMX only takes 2D weights and folds src1 batches into M

  #291·danielhanchen opened Oct 19, 2026Member·

  ·
  1