All pull requests

Pull requests

• [Engine] prefill: a padded activation is gfx11 only, and Gemm::native copies it unpadded instead of exiting

  #3484·John-Pi opened Oct 20, 2026·

  ·

• [Engine] stage buffers: fence a queued copy per stream, not per buffer (#1237)

  #3483·John-Pi opened Oct 20, 2026·

  ·

• [Engine] HIP: opt-in Q4_0-KV prompt attention on RDNA4 matrix cores (STRATA_HIP_WMMA_Q4)

  #3482·shenmintao opened Oct 20, 2026·

  ·

• [Benchmark] RX 9060 XT 16GB / Qwen3.8-Flash-Next / IQ2_XS / strata-qwen38-thinking-v1-draft.1 (custom track)

  #3481·shenmintao opened Oct 20, 2026·

  ·

• [Engine] HIP: verify-window routing for non-native routers (256 experts) on every HIP card

  #3476·pugolini opened Oct 20, 2026Contributor·

  ·

• [Benchmark] RX 9070 XT / 64 GB: IQ3 and Q4 performance and coding results

  #3475·visualizedigital opened Oct 20, 2026·

  ·

• [Other] setup: FP4-Local conversion no longer dies at step 6 on Python 3.15 (torch pin 2.9.1 -> 2.14.1)

  #3474·beplee opened Oct 20, 2026·

  ·

• [Engine] Zen 1/2: build ggml-cpu without BMI2 pdep (IQ1_S/IQ1_M CPU dots 16-19x faster, same results)

  needs-format
  #3472·Licenser opened Oct 20, 2026·

  ·
  1

• [Other] tools: the busbar gateway, one OpenAI address in front of several providers

  #3466·Buddychek opened Oct 20, 2026·

  ·

• [Engine] Metal: opt-in --dflash DFlash2 drafter in place of the MTP layer

  #3457·hotocoo opened Oct 20, 2026·

  ·

• [Engine] GP100 (sm_60): 4-bit slots for resident IQ2_S / IQ3_XXS / IQ3_S gate/up experts, decode +9.7% (opt-in, STRATA_EXPERT_NIBBLE=1)

  #3455·ruibeikaa opened Oct 20, 2026Contributor·

  ·

• [Engine] the sm_89 (Ada) MMVQ-IL table, measured on an RTX 4070

  #3453·rodi4eg opened Oct 20, 2026·

  ·

• [Benchmark] community benchmark: RTX 4070 12 GB (sm_89), Coder IQ1_M, 131,072-token context

  #3452·rodi4eg opened Oct 20, 2026·

  ·

• [Benchmark] RTX 3090 / Qwen3.8-Flash-Next UD-Q4_K_XL / strata-qwen38-thinking-v1-draft.1

  #3449·intosec opened Oct 20, 2026·

  ·

• [Engine] CUDA 12.0/12.1: widen the f32 bits by hand in the q36 VDEBUG print

  #3447·rodi4eg opened Oct 20, 2026·

  ·

• [Benchmark] tools/strata_bench.py: assign the port before the harness starts the server

  #3446·rodi4eg opened Oct 20, 2026·

  ·

• [Other] update the Coder model pin to the 2026-10-19 release

  #3445·rodi4eg opened Oct 20, 2026·

  ·

• [Benchmark] add rx 7900 xtx qwen3.8 flash-next iq3xxs community + standard-v1 results (v0.2.1, engine 0.1.45)

  #3436·GoldenLeafBird opened Oct 20, 2026·

  ·
  2

• [GUI] Use web API keys immediately when browser storage fails

  #3435·engharat opened Oct 20, 2026·

  ·

• [Benchmark] 2x Radeon AI PRO R9700 / Flash-Next IQ3_S / strata-qwen38…

  needs-format
  #3433·lvdesheng opened Oct 20, 2026·

  ·
  1

• [Engine] HIP: the expdec shared int4 arrays use the attribute-after-declarator form

  #3429·chenxyz129 opened Oct 20, 2026·

  ·

• [Other] Fix Windows child process permissions in adopt_user_defaults

  #3428·kossum opened Oct 20, 2026Contributor·

  ·

• [Engine] Metal: the control vector (experimental speed projection) on the Mac engine

  #3427·Flo5k5 opened Oct 20, 2026·

  ·

• [Benchmark] 2x RX 6900 XT (gfx1030), 0.2.0-beta stock, Flash-Next IQ3_S: strata_bench suite + real-text reads and decode

  #3426·Vop3p opened Oct 20, 2026Contributor·

  ·

• [Other] Set the policy baseline for the standalone GGML overlay test

  #3423·hulkbig opened Oct 20, 2026Contributor·

  ·
  1