All pull requests

Pull requests

• Networked Pool support

  #482·fragtion opened Oct 2, 2026·

  ·
  1

• serve, web: a space-free vision temp dir, lazy vision, and a Chinese UI

  #480·ziphao2010 opened Oct 2, 2026·

  ·
  1

• prefill: MMQ prompt kernels for Q6_K experts (STRATA_MMQ_KQUANTS)

  #478·yannickloth opened Oct 2, 2026·

  ·

• native head: report the cudaHostAlloc error, not just "cannot pin N MiB"

  #476·yannickloth opened Oct 2, 2026·

  ·

• native experts: Q5_0 GPU kernel (#222)

  #473·Rhonstin opened Oct 2, 2026·

  ·
  1

  1

• tests: test_prebuilt_hip_zip passes on Linux (EXE as on Windows)

  #472·homeofe opened Oct 2, 2026·

  ·

• setup: a download stopped before its rename is finished without a request (no 5-minute 416 retry loop)

  #470·homeofe opened Oct 2, 2026·

  ·

• bench: community results, RTX 2080 Ti 11 GB + Threadripper 3960X, IQ3_S at 262K with KV streaming

  #469·homeofe opened Oct 2, 2026·

  ·

• bench: community report — RTX 5090, IQ3_S at 1M context (YaRN), agent-style workload

  #466·gravitomagnetic opened Oct 2, 2026·

  ·

• ple: read the n-gram table at full BF16 precision

  #464·constantindjonkam opened Oct 2, 2026·

  ·

• decode: wait for an adaptive expert swap before reading the residency table

  #463·constantindjonkam opened Oct 2, 2026·

  ·

• diagnostics: dump every verify window's logits, so greedy decode can be bisected

  #462·constantindjonkam opened Oct 2, 2026·

  ·

• Monitor: the conversation cache card (parked conversations from the engine's CACHE lines)

  #456·sebastianrcnt opened Oct 2, 2026·

  ·

• setup: the Unsloth UD-Q4_K_XL is NVIDIA only, stop on the AMD HIP backend before the download

  #455·zuraiz-anjum opened Oct 2, 2026·

  ·
  1

• serve: honor stop and stop_sequences

  #454·zuraiz-anjum opened Oct 2, 2026·

  ·

• prefill: the draft layer's batched K/V for a ring too (KV streaming) - prefill +4.6% with --kv-resident

  #453·architectds opened Oct 2, 2026Contributor·

  ·

• qsa_prompt_attn: Q4_0 KV on tensor cores (mode 4) - long prompts with --kv q4_0 ~19% faster

  #452·architectds opened Oct 2, 2026Contributor·

  ·
  1

• calibrate: a gpu list is not a layer split (#447)

  #450·EdderTalmor opened Oct 2, 2026·

  ·
  1

• fix: fall back when ROCm development files are missing

  #449·kuishou68 opened Oct 2, 2026·

  ·

• kernels: add opt-in small-window GR and one-warp MMVQ experiments

  #443·CC-David-CC opened Oct 2, 2026·

  ·

• hip: add community gfx1012 support with version-gated legacy compatibility

  #442·CC-David-CC opened Oct 2, 2026·

  ·

• Contrib/non mtp serving

  #441·CC-David-CC opened Oct 2, 2026·

  ·

• bench: community report, RTX 5090 + Ryzen 9 9950X3D, IQ3_S at the full 262K window; prefill auto:32768 A/B; conversation-cache 32 GiB demo

  #440·Ambolio opened Oct 2, 2026·

  ·

• prefill: a group's expert gathers in one launch (bit-identical, long prompts +4%)

  #439·architectds opened Oct 2, 2026Contributor·

  ·

• Add optional Strata GUI Manager

  #434·Ayshinko opened Oct 1, 2026·

  ·