All pull requests

Pull requests

• feat(qwen4): add opt-in asynchronous exact expert LRU on gfx1201

  #1022·HUSRCF opened Oct 29, 2026Contributor·6 of 15 tasks·

  ·

• fix(receipts): drop the vs-ollama receipts orphaned by the v0.4.4 docs redirect

  #1017·amittell opened Oct 27, 2026Contributor·5 of 6 tasks·

  ·
  1

• fix(setup): refuse --tag mismatches and unsupported --gpu-arch values

  #1015·amittell opened Oct 27, 2026Contributor·5 of 6 tasks·

  ·
  1

• Multi-slot serve: VMM-backed shared paged KV, retained PM4 + batched MoE decode on gfx1100

  #1013·ghazni101 opened Oct 27, 2026Contributor·

  ·

• fix(llama): apply Llama-3 rope_scaling (rope_type llama3) in the LLaMA carrier

  #1007·fivetide opened Oct 27, 2026Collaborator·

  ·

• perf(gfx1100): reduce Qwen3.8 DFlash selector-to-verify handoff latency

  #1006·HUSRCF opened Oct 27, 2026Contributor·14 of 17 tasks·

  ·

• fix(client,install): refuse mixed CLI/daemon installs; name the unchanged install on failure

  #1002·amittell opened Oct 27, 2026Contributor·5 of 6 tasks·

  ·
  4

• feat(qwen4): add Flash Next vision sidecar export and image serving

  #970·HUSRCF opened Oct 25, 2026Contributor·13 of 15 tasks·

  ·
  4

• fix(dspark): llama DSpark verify captures the drafter context; gfx1151 XLDS for the drafter's Q8 gate_up and F16 head (HIPF-111)

  #959·fivetide opened Oct 24, 2026Collaborator·

  ·
  5

• perf(gfx1100): shard TP2 lm_head and keep bounded prefill reuse opt-in (measured +1.5% AR decode)

  #940·HUSRCF opened Oct 22, 2026Contributor·11 of 21 tasks·

  ·

• feat(lab): prototype opt-in tiered KV and Qwen position oracles

  #939·HUSRCF opened Oct 22, 2026Contributor·10 of 16 tasks·

  ·
  1

• feat(ds4): add transactional routed-expert paging

  #870·nwoolmer opened Oct 13, 2026Collaborator·

  ·