All pull requests

Pull requests

• CI: fit the prebuilt run in 60 runners: shard CUDA and ROCm windows, every job starts after resolve

  #302·danielhanchen opened Oct 20, 2026Member·

  ·
  2

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

  #298·danielhanchen opened Oct 19, 2026Member·

  ·
  7

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

  #297·danielhanchen opened Oct 19, 2026Member·

  ·
  1

• model : deepseek41 DSpark speculative draft

  #296·danielhanchen opened Oct 19, 2026Member·

  ·
  1

• 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·

  ·
  10

• 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

• ggml : use log1p in softplus so negative inputs do not underflow to 0

  #290·danielhanchen opened Oct 19, 2026Member·

  ·
  3

• model : add DeepSeek-V4.1-Flash-1017 (deepseek41), exact reference implementation

  #289·danielhanchen opened Oct 19, 2026Member·

  ·

• ggml-cuda: key the MUL_MAT_ID MMVQ batch cap on the compiled arch

  #287·danielhanchen opened Oct 18, 2026Member·

  ·
  1

  12

  

• pr-set: repin #264 and #266, drop ggml-org#30670 (merged upstream)

  #286·danielhanchen opened Oct 18, 2026Member·

  ·
  1

• [Draft] DeepSeek-V4.1-Flash-1017 support (preliminary, stacked on #61)

  #285·danielhanchen opened Oct 18, 2026Member·

  ·
  1

• server: add penalize_prompt to keep prompt tokens out of the penalty history

  #283·danielhanchen opened Oct 17, 2026Member·

  ·
  8

• model: add Beam (Reflection)

  #281·danielhanchen opened Oct 16, 2026Member·

  ·
  10

• unsloth: reconcile ggml op-count asserts after the pin merges

  #274·danielhanchen opened Oct 13, 2026Member·

  ·
  32

• vulkan: two-subgroup coopmat1 FA tile so HSK 512/576 fit 32 KiB shared memory

  #270·danielhanchen opened Oct 13, 2026Member·

  ·
  4

• server: preemption notices, asynchronous parks and exact concurrency (pin branch on b11577)

  #269·danielhanchen opened Oct 13, 2026Member·

  ·
  60

• Host-weight prefetch in the scheduler, faster MoE cache prompt and decode, cache seeding under llama-server

  #268·danielhanchen opened Oct 13, 2026Member·

  ·
  109

• llama-mmap: populate the non-lazy ranges when MAP_POPULATE is skipped for lazy tensors

  #266·danielhanchen opened Oct 13, 2026Member·

  ·
  6

• ggml-cpu: x86 SIMD vec_dot for Q2_0

  #265·danielhanchen opened Oct 13, 2026Member·

  ·
  6

• CUDA: faster MoE prompt processing (compact MMQ tiling, device expert sort, GET_ROWS)

  #264·danielhanchen opened Oct 13, 2026Member·

  ·
  10

• server: preemption notices, asynchronous parks and exact concurrency (pin branch on b11540)

  #260·danielhanchen opened Oct 11, 2026Member·

  ·
  10

• Show the MoE cache hit rate in perf output and llama-server /metrics

  #259·LeoBorcherding opened Oct 10, 2026·

  ·