All pull requests

Pull requests

• [Engine] session-file saves: ping-pong blocks and a writer thread so the source copy overlaps the previous block write

  #4411·sakuumomo opened Oct 28, 2026Contributor·

  ·

• [Engine] fix KVA crash with STRATA_GDN_CHUNKED on HIP: the state-only GDN recurrence no longer takes the chunked WMMA scan

  #4408·bsvinay opened Oct 28, 2026Contributor·

  ·
  2

• [Engine] build v0.2.7 with nvcc 12.0: leave out the opt-in WideBatch prefill kernels (kernel parameters over 4096 bytes)

  #4406·nomishbhardwaj opened Oct 28, 2026·

  ·

• [Benchmark] RTX 5060 Ti 16 GB + i7-14700 / Qwen3.8-Flash-Next IQ3_S 256K / standard-v1 draft.1

  #4401·rifaniponk opened Oct 28, 2026·

  ·

• [Benchmark] RX 7900 XT 20 GB / qwen3.8-flash-next / IQ3_S / strata-qwen38-thinking-v1-draft.1 (0.2.7, engine 0.1.52)

  #4386·GustavoZiaugra opened Oct 28, 2026·

  ·

• [Engine] serve: a zero drafter ring (Metal --mtp) must not void the conversation cache or kill the engine

  #4384·catalinb13 opened Oct 28, 2026·

  ·

• [Engine] DeepSeek-V4.1: the hyper-connection split off the decode path, +1.4% decode, bit-exact

  #4382·Licenser opened Oct 28, 2026Contributor·

  ·

• [Engine] Prefill: cache the fused ring's layer check per CUDA device (layer split over unlike cards, refs #3577)

  #4380·moses-sun opened Oct 28, 2026·

  ·

• [Engine] Metal: moe_gu decodes gate then up rows for 3-4 token groups (MOE_SEQ); Q2_0/IQ2_XXS/IQ3_XXS only

  #4377·joseedson18jc opened Oct 27, 2026·

  ·

• [Engine] dsv41: the WMMA expert kernels compile on pre-Volta cards

  needs-format
  #4376·lyesrock opened Oct 27, 2026·

  ·
  1

• [Other] serve: opt-in "system_reminders": "tail" keeps a client's system-reminder blocks at the end of the prompt

  #4374·XinacS opened Oct 27, 2026·

  ·
  3

• [Benchmark] standard draft: RX 9070 XT, Windows 11 / Qwen3.8-Flash-Next / IQ3_XXS / strata-qwen38-thinking-v1-draft.1 (3 runs, engine 0.1.48)

  #4372·DE3000 opened Oct 27, 2026·

  ·

• [GUI] Report each native model's context window in /v1/models (n_ctx)

  #4370·msalsas opened Oct 27, 2026·

  ·
  1

• [Engine] DeepSeek-V4.1: stream the model load, 3x faster start, up to 7x faster restart

  #4367·Licenser opened Oct 27, 2026Contributor·

  ·

• [Engine] DeepSeek prompt: the busy experts take FP16 GEMMs, not the tiles (DS4_FAT_ROWS, opt-in)

  #4366·lyesrock opened Oct 27, 2026·

  ·

• [Engine] dsv41: charge the context to RAM only on an integrated GPU

  #4362·makefunstuff opened Oct 27, 2026·

  ·

• [GUI] Keep subagent rows in the host and repair old ghost durations (#4218)

  needs-format
  #4360·xvcezrqpbxtpzblrt opened Oct 27, 2026Contributor·

  ·
  1

  1

• [Engine] isa guard: the iq512 exact rows are wide on purpose (#4213)

  #4349·rwkeyes opened Oct 27, 2026Contributor·

  ·

• [Engine] GGUF reader: size the integer tensors, so a complete DeepSeek first shard is accepted and a cut one refused

  #4348·lu123bing opened Oct 27, 2026·

  ·

• [Engine] q36: give the packed Q8_0 decode copies back to the expert slots in the hybrid mode

  #4347·rynwang opened Oct 27, 2026·

  ·
  1

• [Engine] dsv41: a kept conversation's files are its owner's alone (0600)

  #4343·suifei opened Oct 27, 2026·

  ·
  2

• [Engine] serve: STRATA_ADMIT_SHORT_FIRST orders the single-stream queue too

  #4341·suifei opened Oct 27, 2026·

  ·
  2

• [Engine] serve: keep a long prompt read alive for clients that count events

  #4340·suifei opened Oct 27, 2026·

  ·
  2

• [Engine] size the conversation cache for one snapshot of the configured context

  #4338·xvcezrqpbxtpzblrt opened Oct 27, 2026Contributor·

  ·

• [Engine] q36: a 4096-token prompt chunk in the hybrid mode when the card has room

  #4335·rynwang opened Oct 27, 2026·

  ·
  1

  2