All pull requests

Pull requests

• [Engine] HIP: alignas(16) int4 staging arrays do not compile (the two sites #778 left)

  #5518·rwkeyes opened Nov 7, 2026Contributor·

  ·
  1

• [Engine] qwen3next family: geometry-derived trunk + GDN grouped-head support (Groot lane byte-identical)

  #5517·archadrian opened Nov 7, 2026·

  ·
  2

• [Engine] serve: in-place retune (TUNE) and conversation-cache routing by traffic shape

  #5516·rwkeyes opened Nov 7, 2026Contributor·

  ·

• [Benchmark] RTX 5060 Ti / Qwen3.8-Flash-Next / UD-IQ4_XS-v31 / strata-qwen38-thinking-v1-draft.1

  #5514·jkenner111 opened Nov 7, 2026·

  ·

• [Engine] serve: skip the thinking block on prompts that ask for a short direct answer (opt-in)

  #5512·rwkeyes opened Nov 7, 2026Contributor·

  ·

• [Engine] Credit clean active_file cache in the conversation memory guard

  #5504·giphoo opened Nov 7, 2026·

  ·

• [Engine] Allow session snapshots without a draft layer

  #5503·giphoo opened Nov 7, 2026·

  ·

• [Engine] Prototype incremental and prefix-readable session persistence

  #5500·giphoo opened Nov 7, 2026·

  ·

• [Engine] prefill: STRATA_IQ4_W4A8 on sm_75 (Turing): MMQ bits, prompt reads -2 % on 2x CMP 50HX

  #5498·xrip opened Nov 7, 2026Contributor·

  ·

• [Benchmark] 3x RTX 3090 24 GB, all-cards split engine (v0.2.11), Qwen3.8-Flash-Next GSQ-RCO IQ3_S

  #5497·chaerle opened Nov 7, 2026·

  ·

• [Engine] serve: a picture's encode whose client left no longer holds every request

  #5495·Keoian opened Nov 7, 2026·

  ·

• [Other] Windows: recognise the Radeon Pro V620 (PCI id 0x73A1) as gfx1030

  #5494·LouisClt opened Nov 7, 2026·

  ·

• [Other] serve/frontend: deduplicate rewritten non-string XML params, coerce Python bools, and skip empty/null placeholders

  #5493·OleksiyPenkov opened Nov 7, 2026·

  ·
  2

• [Benchmark] AMD Radeon RX 7600 XT 16 GB (gfx1102), CachyOS Linux, Ryzen 5 3600

  #5486·DamirLukina88 opened Nov 7, 2026·

  ·
  9

• [Engine] serve: Cloudflare Access zero-trust JWT authentication (STRATA_CF_ACCESS_TEAM)

  #5485·DamirLukina88 opened Nov 7, 2026·

  ·
  5

• [Engine] clarify SYCL CPU diagnostics for the VNNI tier

  #5481·nekorobi-0 opened Nov 7, 2026·

  ·

• [Engine] allow reusing standalone PLE n-gram shards across different quant splits

  #5479·code-martin opened Nov 7, 2026Contributor·

  ·

• [Engine] Add GET /v1/models/{model_id} single model retrieve endpoint

  #5477·DamirLukina88 opened Nov 7, 2026·

  ·
  1

  4

• [Engine] gfx103x decode: opt-in one-warp MMVQ for K <= 4096 and a per-format expert layout table (STRATA_MMVQ_NW1, STRATA_EXP_TABLE)

  #5476·Vop3p opened Nov 7, 2026Contributor·

  ·

• [Engine] opt-in int8 copies for the persistent hc read (STRATA_HC_PERSIST_Q8)

  #5475·Vop3p opened Nov 7, 2026Contributor·

  ·

• [Engine] HIP prompt path: opt-in FP16 reads of the hc up projection and attn_qkv (STRATA_F16_NOWIDEN) and the fused SwiGLU + Q8 pass on HIP (STRATA_SWIGLU_Q8)

  #5474·Vop3p opened Nov 7, 2026Contributor·

  ·

• [Engine] link stubs for fused::want_dm_f16 / dm_f16_used without the fused prompt kernels

  #5473·YamineRL opened Nov 7, 2026Contributor·

  ·

• [Engine] DeepSeek on a DGX Spark: STRATA_DS4_DISK_SKIP_RANK / PROMPT_SKIP_RANK, a store miss left out by its pick rank (opt-in, changes the answers)

  #5472·R-Okauchi opened Nov 7, 2026·

  ·

• [Engine] DeepSeek: faster Q4_K / Q5_K decode GEMVs and light experts with the same bits, the default only on integrated GPUs

  #5471·R-Okauchi opened Nov 7, 2026·

  ·

• [Engine] DS4_PROF_WIN: the decode profile of the drafter's verify windows, and the attention by part

  #5470·R-Okauchi opened Nov 7, 2026·

  ·