All pull requests

Pull requests

• poly: store Flint's and LFM2.5's KV caches as Q8_0 blocks

  #152·smashdog opened Nov 6, 2026Contributor·

  ·

• Add ROCm dockerfile

  #151·jintakhan opened Nov 6, 2026·

  ·

• feat: opt-in STRATA_POLY_STAGED_TILES: the staged experts of a prompt batch in size classes

  #149·drakarah opened Nov 6, 2026Contributor·

  ·
  2

• Feat: HIP - FP32 prompt attention with custom kernel: +30% prompt processing at 90K to 112K context (opt-in, q8_0 cache)

  #147·drakarah opened Nov 5, 2026Contributor·

  ·
  4

• feat: opt-in split-key decode attention for HIP (D256, GQA 8, q8_0), +53% decode at 90K to 112K context

  #145·drakarah opened Nov 5, 2026Contributor·

  ·

• feat: VRAM Pool

  #143·drakarah opened Nov 5, 2026Contributor·

  ·
  1

• hip: use the SDWA signed dot4 on gfx1010 too

  #140·drakarah opened Nov 4, 2026Contributor·

  ·
  1

• metal: expand Apple Silicon model support and enable optimized defaults

  #135·berkkarahan opened Nov 4, 2026Contributor·

  ·
  1

• setup: add Groot 19B-A3B (--family groot) on strata-poly

  #126·SRZhu26 opened Nov 3, 2026·

  ·
  4

• docs: README in Turkish

  #123·enderkus opened Nov 2, 2026·

  ·
  3

• Disable the spill of parked conversations to a file

  #119·vehystrix opened Nov 1, 2026Contributor·

  ·
  2

• Volta (sm_70) arm for the GDN tensor-core path

  #100·Csavoldi opened Oct 26, 2026·

  ·

• feat(web): allow editing user messages with tail removal

  #75·Ziomd opened Oct 22, 2026Contributor·

  ·
  3

• setup: add Windows max-compatibility build option

  #58·berkkarahan opened Oct 20, 2026Contributor·

  ·
  1

• setup: compile the AMD engine on Windows with --build

  #57·berkkarahan opened Oct 20, 2026Contributor·

  ·
  10