All pull requests

Pull requests

• [Engine] add opt-in huge-page backing to DeepSeek's registered RAM store

  #4286·yhfgyyf opened Oct 27, 2026·

  ·

• [Engine] Opt-in routed prompt stream: skip streamed experts no token routes to (follow-up read time -18%, same output)

  #4283·bitaccesscomau opened Oct 27, 2026·

  ·

• [Benchmark] RX 7700XT + 32GB DDR5 6000

  #4277·losdock opened Oct 27, 2026·

  ·
  2

• [Engine] prevent SM60 VMAD input accumulator aliasing

  #4274·Joe11221 opened Oct 27, 2026·

  ·

• [Engine] Metal: opt-in routed-expert tiles (STRATA_METAL_MOE_GU/_DN); Q2_0 8,1,2 + 4,2,4 is 17-24% faster on M1 Max

  #4272·joseedson18jc opened Oct 26, 2026·

  ·

• [Engine] qwen35: FP16 K/V cache by default, int8 with --kv-q8 (262144 fits a 16 GiB card)

  needs-format
  #4270·benthic-io opened Oct 26, 2026·

  ·
  1

• [Benchmark] STRATA_PROMPT_ATTN_IMMA and STRATA_HC_UPMIX on 2x RTX 4060 Ti (sm_89), Strata 0.2.6

  #4268·fioruccione opened Oct 26, 2026Contributor·

  ·

• [Other] update test start: skip DeepSeek models when choosing the model to test with

  #4267·ai-bits opened Oct 26, 2026·

  ·
  1

• [Engine] Vision: keep image ownership through request preparation

  #4264·Tsury opened Oct 26, 2026Contributor·

  ·

• Arm64 gb10

  needs-format
  #4259·nings opened Oct 26, 2026·

  ·
  1

• [Engine] --drop-file-cache: hand the model files' page cache back once the startup reads are done

  #4258·G0-0000 opened Oct 26, 2026Contributor·

  ·

• [Other] serve: refuse a port another program holds on 127.0.0.1, and say which program

  #4255·ByteJuggler opened Oct 26, 2026·

  ·

• [Other] setup: an interrupted engine update is tried again instead of skipped for good (#2159)

  #4254·ByteJuggler opened Oct 26, 2026·

  ·
  1

• [GUI] record child absence before publishing managed host state

  #4250·zapabob opened Oct 26, 2026·

  ·

• [GUI] accept local GGUF models with 1K and 2K context windows

  #4249·zapabob opened Oct 26, 2026·

  ·

• [GUI] allow an explicit local inference silence timeout

  #4248·zapabob opened Oct 26, 2026·

  ·

• [GUI] Add opt-in answerer for the agent's ask_user_question events

  #4245·devplex69 opened Oct 26, 2026·

  ·

• [Engine] Metal: the mmp expert path for Q4_K and Q5_K packs

  #4240·catalinb13 opened Oct 26, 2026·

  ·

• [Engine] Metal: hot kernel reload - POST /reload_kernels rebuilds the window without a restart

  #4239·catalinb13 opened Oct 26, 2026·

  ·

• [Engine] Metal: faster routed-expert gate/up tiles on M1/M2 (8,1,2 for the K quants)

  #4238·catalinb13 opened Oct 26, 2026·

  ·

• [GUI] API monitor: show the tokens held in the engine's KV cache

  #4232·qwertie opened Oct 26, 2026·

  ·

• [Other] gitignore the engine-feed folder

  #4231·qwertie opened Oct 26, 2026·

  ·

• [TUI] strata connect: use bigger output caps on bigger context windows

  #4230·qwertie opened Oct 26, 2026·

  ·

• [Engine] dsv41: read Engram tables stored as ggml Q5_K

  #4228·melchisedekko opened Oct 26, 2026·

  ·

• [TUI] canonicalize models in setup args

  #4227·Cyneric opened Oct 26, 2026·

  ·
  1