All pull requests

Pull requests

• feat(qwen-flash): support mixed Q4_K_M weights and experimental matrix-core prefill

  #694·LMtx opened Oct 21, 2026Contributor·

  ·
  1

• fix(ds4): keep caller text out of DeepSeek control tokens

  #691·jyatesdotdev opened Oct 21, 2026Contributor·

  ·

• perf(qwen38_flash_next): take the wide prefill routes for every prefill pass

  #690·francescobozzo opened Oct 20, 2026Collaborator·

  ·

• perf(qwen27b): capture DFlash2 prompt features into bounded pinned host memory

  #688·headbouyJB opened Oct 20, 2026Contributor·

  ·
  1

  3

  

• Add support for gfx1150

  #682·lcpo opened Oct 20, 2026·

  ·

• fix(cache): free RAM before writing oversized checkpoints to disk

  #677·sdwilsh opened Oct 20, 2026Contributor·

  ·
  3

  

• perf(qwen-flash): token-history drafting over the Q8 stack

  #608·dardeaw opened Oct 15, 2026·

  ·
  1

  1

• perf(qwen-flash): persistent incremental Q8 staging for prefill

  #579·dardeaw opened Oct 13, 2026·

  ·
  1

• feat(qwen38-flash-next): offload dense prefill projections to the XDNA2 NPU

  #568·mikealanni opened Oct 12, 2026·

  ·
  1

  10

• feat(qwen-flash): add int8 WMMA attention for Q8 prefill

  #566·dardeaw opened Oct 12, 2026·

  ·
  8

• feat(qwen-flash): add Q8 KV cache precision mode

  #561·dardeaw opened Oct 12, 2026·

  ·
  7

• perf(ds4): keep the 16- and 32-column MMQ tiles spill-free

  #483·jtsylve opened Oct 8, 2026Contributor·

  ·
  7

• feat(gemma4): add Gemma 4 31B, 31B QAT and 26B-A4B support

  #426·franzmoca opened Oct 4, 2026Contributor·

  ·
  5

• build: add Debian packaging

  #372·avlis opened Oct 1, 2026·

  ·
  6

• feat(qwen-flash): opt-in YaRN context extension past 262,144 tokens

  #350·kh0pper opened Sep 30, 2026Contributor·

  ·
  1

  9

• feat(qwen): support Qwen3.6-35B-A3B with fast prefill, MTP and DFlash2 serving

  #299·slimsami opened Sep 27, 2026·

  ·
  14

• ROCm 10 evaluation for gfx1151, and a clang 23 prefill fix

  #228·fedeizzo opened Sep 5, 2026Collaborator·

  ·
  2

• feat(hrx): integrate AMD ROCm HRX runtime, Loom IR kernels, and macro-fusions

  #200·fedeizzo opened Aug 25, 2026Collaborator·2/2·

  ·
  10