All pull requests

Pull requests

• [Engine] Metal: decode and prompt kernels for the pair-trellis expert types TQP5/TQP6/TQP7

  #4127·mnott opened Oct 25, 2026·

  ·
  1

• [Engine] Metal: sampling on the decode path (temperature, top-k, top-p, min-p, penalties, Philox seed)

  #4126·mnott opened Oct 25, 2026·

  ·

• [Other] serve tests: VisionArgs passes again (fake process without pid)

  #4125·mnott opened Oct 25, 2026·

  ·

• batch-groups: MTP drafts in the pipelined slot groups (STRATA_PIPED_MTP=N, opt-in) — ported onto main

  needs-format
  #4121·wannagoxuf opened Oct 25, 2026·

  ·
  1

  1

• [Engine] Opt-in two-pass QSA prompt selection: 3xTF32 only near the top-k cut-off, the same cells (256K prompt read +5.2%)

  #4119·bitaccesscomau opened Oct 25, 2026·

  ·

• [Engine] Automatically use --resident-experts to prevent startup failure

  #4117·qwertie opened Oct 25, 2026·

  ·

• [Engine] Native RDNA4 (gfx1201) NVFP4 engine (--gfx1201-nvfp4)

  #4114·dimitri91209 opened Oct 25, 2026·

  ·
  2

• [Engine] GBNF constrained decoding: stable tool calls + JSON schema, now riding the MTP drafter

  #4113·sd133cn opened Oct 25, 2026·

  ·

• [Benchmark] Community results: 2x RTX 2080 Ti, and RTX 5060 Ti + RTX 2080 Ti (qwen3.8-flash-next iq3_s)

  needs-format
  #4112·SnoopTM opened Oct 25, 2026·

  ·
  1

• [GUI] the App Host's source-checkout mode no longer passes an empty STRATA_API_KEY

  #4111·demetree opened Oct 25, 2026Contributor·

  ·
  1

• [Benchmark] Community report: RTX 3090 + RTX 3060, Xeon W-2125 (v0.1.41 vs v0.2.3.1)

  #4107·flix2032 opened Oct 25, 2026·

  ·
  2

• [Engine] --ple-io ram: read the PLE table into page-locked RAM with the unbuffered reader instead of faulting it in through the page cache

  #4105·chenyb999-zhcn opened Oct 25, 2026·

  ·
  1

• [Benchmark] RTX 5090: two community reports from tools/strata_bench.py (Qwen3.8-Flash-Next IQ3_S on 0.1.50, Gyro-M TQ2_0 on 0.1.49.2)

  #4099·PierpaoloPernici opened Oct 25, 2026·

  ·

• [Benchmark] DeepSeek-V4.1-Flash-1017 IQ2_XS on one RTX PRO 5000 72GB (Strata Terminal 0.2.4.2, DSpark A/B, 256K near-window)

  #4096·realsquall1980-art opened Oct 25, 2026·

  ·
  3

• [Engine] add lossless REF and RHT+INT16 disk-state codecs

  #4086·G0-0000 opened Oct 25, 2026Contributor·

  ·

• [Benchmark] RTX 4090, 125.4 GiB RAM: Huihui abliterated UD-Q4_K_XL at 393,216 tokens (2026-10-23)

  #4084·TheHubUser4860938 opened Oct 25, 2026Contributor·

  ·

• [Other] tokenizer: llama.cpp's pre label set, a prerequisite for the DeepSeek-V4-Flash-0731 (deepseek4) port

  #4079·yoyz opened Oct 25, 2026·

  ·
  2

• [Engine] KV cache: the exact-cell window for q4lm8v (opt-in, hardened)

  #4068·monscoro opened Oct 25, 2026·

  ·

• [Benchmark] AMD Radeon RX 7900 XT - qwen3.8-flash-next iq3_s

  #4061·fsacrini opened Oct 25, 2026·

  ·
  2

• [Other] serve: logprobs on /v1/completions (stacks #3200 + #3332)

  #4056·gdevenyi opened Oct 25, 2026·

  ·
  1

• [Benchmark] RTX 3090 + i9-10900K, Flash-Next IQ3_S at 512K context (engines 0.1.44.1 / 0.1.48 / 0.1.50) + seven-configuration A/B

  #4054·smallmj opened Oct 25, 2026·

  ·
  2

• [Engine] alias_sampling: request defaults per model alias

  #4050·furkangkhsn opened Oct 25, 2026·

  ·

• [Engine] replicas: each replica its own server on its own port, with one shared expert arena

  #4049·furkangkhsn opened Oct 25, 2026·

  ·

• [Engine] add opt-in reference-order CPU mixed-IQ down dispatch

  #4046·CrashxZ opened Oct 25, 2026·

  ·
  1

• [Engine] --lookup-max-t: prompt-lookup windows up to 16 tokens while the MTP keeps --spec, bitwise, edits +19% decode

  #4040·merbanan opened Oct 25, 2026Contributor·

  ·