All pull requests

Pull requests

• [Engine] content is "" (not null) when a reply has no text and no tool call

  #5249·festeraeb opened Nov 4, 2026·

  ·
  1

• [Engine] Support capacity-aware block-aligned GGUF tensor parallelism and fix K-quant head slicing

  #5246·Moeblack opened Nov 4, 2026·

  ·
  2

• [Engine] serve: accept Codex multi_agent_v2 agent_message and multi_agent_call(_output) items on /v1/responses

  #5244·bsorensen110 opened Nov 4, 2026Contributor·

  ·
  7

• [Engine] perf(cpu): AVX-512 (Zen 4/5 & Intel) and Arm64 NEON/DotProd (Apple M, Oryon, Grace/DGX GB10) expert & quantizer kernels

  #5243·bsorensen110 opened Nov 4, 2026Contributor·4 tasks·

  ·

• [Engine] hip: gfx1200 hipBLASLt tuning table for hipBLASLt 1.4.1 (100401)

  needs-format
  #5235·er34testarossa opened Nov 4, 2026·

  ·
  1

• [Benchmark] RTX 4070 Ti SUPER: ~5x faster prefill with RAM caching

  #5234·fm5454 opened Nov 4, 2026·

  ·

• [Engine] gfx906 builds again: map the DFlash2 stream-priority calls in the HIP compat layer

  #5233·zhuoran opened Nov 4, 2026·

  ·

• [Benchmark] RTX 4090: Qwen3.8-Flash-Next IQ2_XS, 128K, DFlash2 (standard-v1, 12/12)

  #5231·Have-Purple opened Nov 4, 2026·

  ·

• [Benchmark] RTX 4090: Qwen3.8-Flash-Next IQ3_S, 128K, DFlash2 (standard-v1, 12/12)

  #5230·Have-Purple opened Nov 4, 2026·

  ·

• [Benchmark] RTX 4090: GSQ-RCO Coder IQ1_M, 128K, MTP (standard-v1, 12/12)

  #5228·Have-Purple opened Nov 4, 2026·

  ·

• [Benchmark] RTX 4090: Qwen3.8-Flash-Next IQ3_XXS, 128K, DFlash2 (standard-v1, 12/12)

  #5227·Have-Purple opened Nov 4, 2026·

  ·
  2

• [Benchmark] RTX 4090 / Qwen3.8-Flash-Next-Gyro-M / TQ2_0 / strata-qwen38-thinking-v1-draft.1

  #5224·MoonbirdLu opened Nov 4, 2026·

  ·

• [Benchmark] RTX 3090 + RTX 3060 / Qwen3.8-Flash-Next / IQ2_XS / strata-qwen38-thinking-v1-draft.1

  #5223·erwanpaccard opened Nov 4, 2026·

  ·

• [Docs] Qualify the 32 GB RAM setup plan for Qwen3.6 UD-Q4_K_XL

  #5221·zapabob opened Nov 4, 2026Contributor·

  ·

• [Engine] serve: a reasoning loop written as one long hyphen run-on

  #5218·e041l4su3qu4u03 opened Nov 4, 2026·

  ·
  2

• [Other] Control: recommend_context counts the resident plan's expert arbitration

  #5217·Flo5k5 opened Nov 4, 2026·

  ·
  2

• [Engine] Refuse unsupported Qwen3.6 flags under --serve instead of ignoring them

  #5216·zapabob opened Nov 4, 2026Contributor·

  ·

• [Other] Serve: session_autosave, the conversation cache's L2 on disk

  #5211·Flo5k5 opened Nov 4, 2026·

  ·
  3

• [Benchmark] RTX 3070 Laptop 8GB / Qwen3.8-Flash-Next IQ3_XXS / standard-v1 draft (custom track, 8 GiB laptop)

  #5207·SK-Hardwired opened Nov 4, 2026·

  ·
  1

• [Benchmark] 2x Intel Arc Pro B70 / Qwen3.8-Flash-Next UD-IQ4_XS / strata-qwen38-thinking-v1-draft.1 (layer split, 262k, vfio VM)

  #5205·Maddy55555 opened Nov 4, 2026·

  ·

• [Benchmark] standard-v1: RTX 4090 Gyro-M TQ2_0 + RTX 5090 UD-Q4_K_XL (262k, k8s)

  #5204·em3rik opened Nov 4, 2026·

  ·

• [Engine] STRATA_SELECT_F16=1: the prompt path's block scores on FP16 tensor cores (opt-in, sm_80+, 1M read selection 95 -> 62 s)

  #5203·D-Ogi opened Nov 4, 2026·

  ·

• [Engine] One-read top-k: double the sampling stride past 524,288 cells instead of failing every query

  #5201·D-Ogi opened Nov 4, 2026·

  ·

• [Benchmark] RX 6700 XT: asymmetric PCIe, 4K–256K, KV reuse and 1C BSL

  #5195·Crowmen007 opened Nov 4, 2026·

  ·

• [GUI] Hotfix: Fixed an issue where chats could be trimmed and not stored

  #5192·MohSakhaei opened Nov 4, 2026·

  ·