Strata

2026-09-24-final

Strata: Qwen3.8-Flash-Next on a 12-24 GB NVIDIA GPU + 64 GB RAM, one-click install

2026-09-24 18:42:20 +02:00

2026-09-24-vision

Strata: Qwen3.8-Flash-Next on a 12-24 GB NVIDIA GPU + 64 GB RAM, one-click install

2026-09-24 18:42:20 +02:00

2026-09-27-cache-parity

Merge PR #21 : Q4_0 KV with Hadamard rotation (--kv q4_0), integrated with KV streaming; accurate expert-cache note (#23 )

2026-09-27 11:36:28 +02:00

2026-09-27-esp

Experimental speed projection: an optional control vector (EXPERIMENTAL, off by default)

2026-09-27 16:47:45 +02:00

2026-09-27-kv-q4

Merge PR #21 : Q4_0 KV with Hadamard rotation (--kv q4_0), integrated with KV streaming; accurate expert-cache note (#23 )

2026-09-27 11:36:28 +02:00

2026-09-28-coder

The GSQ-RCO Coder (256 of 512 experts): setup, profile, speeds (PR #54 )

2026-09-28 14:20:58 +02:00

2026-09-28-prefill-speed

engine 0.1.13: about 2x faster long prompts; stall reports for issue #31

2026-09-28 02:03:00 +02:00

2026-09-28-rope-scaling

docs: rope scaling is experimental and opt-in; the measured perplexity at 293K

2026-09-30 16:44:48 +02:00

2026-09-28-speed-0114

docs: speed tables measured with engine 0.1.14 (prompts 1.2-2.4x faster than the old tables)

2026-09-28 13:07:35 +02:00

2026-09-29-layer-split

Layer split: measured results (bench/results/2026-09-29-layer-split) and the recommendations in docs/MULTI_GPU.md

2026-09-29 01:17:30 +02:00

2026-09-29-layer-split-limits

Layer split limits: docs/MULTI_GPU.md and bench/results/2026-09-29-layer-split-limits

2026-09-29 08:44:35 +02:00

2026-09-29-rtx3090-epyc-milan

Bench: RTX 3090 / dual RTX 3090 on EPYC 7453

2026-10-01 02:52:39 +02:00

2026-09-29-speed-0122

Speed tables: prompts with engine 0.1.22 (bench/results/2026-09-29-speed-0122)

2026-09-29 13:38:54 +02:00

2026-09-29-speed-0126

docs: speed tables for engine 0.1.26 (prompts 8-28% faster at 32K-128K than 0.1.22)

2026-09-29 22:09:44 +02:00

2026-09-30-community-rtx-5090

bench: community report, RTX 5070 Ti + Ryzen 9 5900X, 64 GB DDR4-3600 (IQ2_XS / IQ3_XXS / IQ3_S, six-arm A/B) (#993 )

2026-10-06 13:19:04 +02:00

2026-09-30-tc-gemv-sm75

bench: BF16 GEMV via Turing tensor cores on sm_75 - measured negative

2026-10-01 08:15:27 +02:00

2026-10-01-0.1.31-release

bench(gfx1100): 0.1.31 vs 0.1.30 prefill/decode, curated entry + raw trials (#404 )

2026-10-06 13:19:04 +02:00

2026-10-01-community-2x-rtx-pro-4500

bench: community entry, 2x RTX PRO 4500, Swift IQ3_XXS, engine 0.1.30 and 0.1.36 (#418 )

2026-10-06 13:19:04 +02:00

2026-10-01-community-2x-titan-rtx

bench: community report, 2x TITAN RTX (sm_75), IQ3_S at 262K (#389 )

2026-10-06 13:19:03 +02:00

2026-10-01-community-rtx-5090-5950x

bench: community report, RTX 5090 + Ryzen 9 5950X (AVX2), UD-Q4_K_XL / IQ3_S / Swift IQ3_XXS on engines 0.1.31-0.1.39 (#433 )

2026-10-06 13:19:03 +02:00

2026-10-01-community-rtx-pro-4500

bench: community report, RTX PRO 4500 x1/x2 + RTX PRO 4000, Threadripper PRO 3975WX, engine 0.1.31 (#417 )

2026-10-06 13:19:03 +02:00

2026-10-01-ctx-ladder

Community benchmark: RTX 4070 Ti SUPER 16 GiB (Shin-BlackMamba, upstream 6f32ec07) (#995 )

2026-10-06 13:19:05 +02:00

2026-10-01-iq4xs-64gb

bench: IQ4_XS on a 64 GB PC - pinned arena vs experts.bin mmap vs the RAM budget (#416 )

2026-10-06 13:19:04 +02:00

2026-10-02-community-2x5060ti

bench: community results, 2x RTX 5060 Ti 16 GB (PCIe gen3), Strata 0.1.34 (#483 )

2026-10-06 13:19:04 +02:00

2026-10-02-community-2x-rtx-pro-4500-engine-0.1.36

bench: community entry, 2x RTX PRO 4500, Swift IQ3_XXS, engine 0.1.30 and 0.1.36 (#418 )

2026-10-06 13:19:04 +02:00

2026-10-02-community-rtx2080ti-11gb

bench: community results, RTX 2080 Ti 11 GB + Threadripper 3960X, IQ3_S at 262K with KV streaming (#469 )

2026-10-06 13:19:03 +02:00

2026-10-02-community-rtx4000ada-layer-split

bench: add community RTX 4000 Ada 2-GPU vs 3-GPU layer-split results (#521 )

2026-10-06 13:19:04 +02:00

2026-10-02-community-rtx5090-1m-agent

bench: community report — RTX 5090, IQ3_S at 1M context (YaRN), agent-style workload (#466 )

2026-10-06 13:19:05 +02:00

2026-10-02-community-rtx-5060-ti-x2

Community benchmark: 2x RTX 5060 Ti (16 GB), Xeon E5-2690 v4, IQ3_XXS, engine 0.1.35 (#508 )

2026-10-06 13:19:03 +02:00

2026-10-02-community-rtx-5090-iq3s

bench: community report, RTX 5090 + Ryzen 9 9950X3D, IQ3_S at the full 262K window; prefill auto:32768 A/B; conversation-cache 32 GiB demo (#440 )

2026-10-06 13:19:04 +02:00

2026-10-02-community-rx9070xt-windows

bench: community results, RX 9070 XT 16 GB on Windows 11, ready-made AMD engine 0.1.35, IQ3_S (#499 )

2026-10-06 13:19:03 +02:00

2026-10-02-gdn-chunk

bench (v100): gdn_chunk_bench - the GDN prompt recurrence variants and the chunked path, real shapes

2026-10-14 07:05:18 +02:00

2026-10-03-ashley-2080ti

Bench: RTX 5090 / 9950X3D (IQ3_S, three-pass A/B) + RTX 2080 Ti / 9900KF (IQ3_XXS, Turing) (#698 )

2026-10-06 13:19:03 +02:00

2026-10-03-community-2x-mi50

bench: community report, 2x AMD Instinct MI50 (gfx906), Coder IQ1_M, 4K to 128K prompt tokens

2026-10-03 23:33:55 +02:00

2026-10-03-community-r9700-windows

bench: Windows 11 + Radeon AI PRO R9700 (gfx1201), IQ2_XS, 4K to 248K prompt tokens (#618 )

2026-10-06 13:19:04 +02:00

2026-10-03-community-rtx4090-iq3s-140k-code

bench: community report — RTX 4090, Ryzen 9 7950X, Flash-Next IQ3_S (#624 )

2026-10-06 13:19:04 +02:00

2026-10-03-community-rtx4090-iq3s-140k-ru

bench: community report — RTX 4090, Ryzen 9 7950X, Flash-Next IQ3_S (#624 )

2026-10-06 13:19:04 +02:00

2026-10-03-community-rtx4090-iq3xxs-200k-code

bench: community report — RTX 4090, Ryzen 9 7950X, Flash-Next IQ3_S (#624 )

2026-10-06 13:19:04 +02:00

2026-10-03-community-rtx4090-iq3xxs-200k-ru

bench: community report — RTX 4090, Ryzen 9 7950X, Flash-Next IQ3_S (#624 )

2026-10-06 13:19:04 +02:00

2026-10-03-community-rx7900xtx-rocm102-nightly

Community benchmark: RX 7900 XTX 24 GB — ROCm 7.1.1 vs ROCm 10.2 nightly, serve-path decode (Qwen3.8-Flash-Next IQ3_S) (#745 )

2026-10-06 13:19:04 +02:00

2026-10-03-community-v100

docs: community benchmark row - Tesla V100-PCIE-32GB (sm_70 source build, before/after calibration) (#707 )

2026-10-06 13:19:03 +02:00

2026-10-03-community-z840-rtx3090ti

bench: community report, RTX 3090 Ti + 2x Xeon E5-2699 v3 (HP Z840), IQ3_S at 262K (#674 )

2026-10-06 13:19:04 +02:00

2026-10-03-dev-5090

Bench: RTX 5090 / 9950X3D (IQ3_S, three-pass A/B) + RTX 2080 Ti / 9900KF (IQ3_XXS, Turing) (#698 )

2026-10-06 13:19:03 +02:00

2026-10-03-rx6900xt-coder

bench: report Coder IQ1_M on Threadripper 3990X and RX 6900 XT (#628 )

2026-10-06 13:19:03 +02:00

2026-10-03-v100-kv-int8-read

kv_q8: the gather decodes 8 int8 per thread (bitwise, +11% on the 32k microbench)

2026-10-14 07:05:04 +02:00

2026-10-03-v100-prompt-attn

docs: point the BF16 note at #540 (the BF16-via-FP16 PR #593 was closed in its favor)

2026-10-03 23:28:14 +02:00

2026-10-03-windows-groups

bench: report Coder IQ1_M on Threadripper 3990X and RX 6900 XT (#628 )

2026-10-06 13:19:03 +02:00

2026-10-04-community-2x-arc-pro-b60

docs: the host has 64 GB of RAM (62 GiB usable), not 62 GB

2026-10-05 12:18:48 +02:00

2026-10-04-community-2x-rtx3060

Community benchmark: 2× RTX 3060 12 GB, Qwen3.8-Flash-Next IQ3_S (three prompt sizes, 3 runs each) (#721 )

2026-10-06 13:19:04 +02:00

2026-10-04-community-4090-32gb-ram-512k

bench: community report, RTX 4090 + 32 GB RAM at 512K context (IQ2_XS) (#834 )

2026-10-06 13:19:04 +02:00

2026-10-04-community-4090-agents-vs-dense-27b

bench: community report, Flash-Next IQ2_XS vs a dense 27B as coding agents (RTX 4090 + 32 GB RAM) (#882 )

2026-10-06 13:19:04 +02:00

2026-10-04-community-mi50

bench: community MI50 (gfx906) — Strata 0.1.38 speed and recall (#758 )

2026-10-06 13:19:03 +02:00

2026-10-04-community-rtx4090-iq3xxs-200k-code

bench: community report — RTX 4090, Ryzen 9 7950X, Flash-Next IQ3_S (#624 )

2026-10-06 13:19:04 +02:00

2026-10-04-community-rtx4090-iq3xxs-200k-ru

bench: community report — RTX 4090, Ryzen 9 7950X, Flash-Next IQ3_S (#624 )

2026-10-06 13:19:04 +02:00

2026-10-04-community-rtx-4080s-iq3s

Community benchmark: Flash-Next IQ3_S on an RTX 4080 SUPER (Windows), with the expert-pool worker sweep (#780 )

2026-10-06 13:19:04 +02:00

2026-10-04-community-rtx-4090-laptop

bench: community results for an RTX 4090 Laptop GPU (IQ3_S, 128K and 256K) (#757 )

2026-10-06 13:19:04 +02:00

2026-10-04-community-rtx-5070ti-windows

bench: community report, RTX 5070 Ti + Ryzen 7 9800X3D on Windows 11, IQ3_S, 5K to 164K prompt tokens (#832 )

2026-10-06 13:19:05 +02:00

2026-10-04-community-rtx-5090-9700x

Add an RTX 5090 + Ryzen 7 9700X IQ3_XXS benchmark on engine 0.1.39 (#777 )

2026-10-06 13:19:03 +02:00

2026-10-04-community-rx7900xtx-hipblaslt-100500

Community benchmark: RX 7900 XTX 24 GB — ROCm 7.1.1 vs ROCm 10.2 nightly, serve-path decode (Qwen3.8-Flash-Next IQ3_S) (#745 )

2026-10-06 13:19:04 +02:00

2026-10-04-community-rx-6800-windows

bench: community results for an RX 6800 on Windows (0.1.39 against 0.1.38) (#815 )

2026-10-06 13:19:04 +02:00

2026-10-04-community-rx-7900-gre

Community benchmark: RX 7900 GRE (gfx1100, 16 GB), three Flash-Next quants (#740 )

2026-10-06 13:19:04 +02:00

2026-10-04-community-v100-16gb-ram

bench: community report, Tesla V100 32 GB with 16 GB of RAM (Coder IQ1_M) (#823 )

2026-10-06 13:19:04 +02:00

2026-10-04-community-v100-rtx4070-split

bench: community report, Tesla V100 32 GB + RTX 4070 layer split with 16 GB of RAM (IQ2_XS) (#850 )

2026-10-06 13:19:04 +02:00

2026-10-04-rdna2-fp16-prompt

hip gfx103x FP16 prompt: STRATA_F16_RANGE counts what enters and leaves FP16; the range measured (outputs peak at 410 of 65504) and the FP32 route compared

2026-10-14 07:08:33 +02:00

2026-10-04-rdna2-helper-pcie-share

bench, docs: the helper GPU's experts and the PCIe share on 2x RX 6900 XT; SECOND_GPU.md states the HIP measurement

2026-10-05 13:19:38 +02:00

2026-10-04-rdna2-pre75-attention

HIP gfx103x: PR #540 's attention kernel with 8 cells per step and DPP lane exchanges (bit-exact)

2026-10-06 14:58:47 +02:00

2026-10-04-windows-mapped-release

Windows mapped-page release: the switch is STRATA_FILE_RELEASE (#818 )

2026-10-05 15:58:38 +02:00

2026-10-05-community-2x-p40

Community benchmark: 2x Tesla P40 22 GB (Pascal, experimental CUDA 12), single card and layer split (#1028 )

2026-10-06 13:19:05 +02:00

2026-10-05-community-2x-rtx-4060ti

bench: community report, 2x RTX 4060 Ti 16 GB + Threadripper PRO 3975WX, UD-IQ4_XS at 131K, layer split (#1062 )

2026-10-06 13:19:05 +02:00

2026-10-05-community-2x-rx-6900xt

bench: index the three AMD / Intel community folders in bench/results/COMMUNITY.md, list the 6900 XT trim

2026-10-06 14:59:04 +02:00

2026-10-05-community-arc-b65

bench: community report, Intel Arc Pro B65 32 GB (SYCL), IQ2_XS (PR #955 )

2026-10-06 14:59:04 +02:00

2026-10-05-community-e5-2673v3-rtx-4060ti

bench: community report, RTX 4060 Ti + Xeon E5-2673 v3 (DDR3, PCIe 3.0 x8), Q2_0 and IQ3_S, 0.1.39 path vs #706 + #764 + a busier tier (#913 )

2026-10-06 13:19:04 +02:00

2026-10-05-community-gfx1151

bench: community report, Radeon 8060S (gfx1151) on a Ryzen AI Max+ 395 (PR #917 )

2026-10-06 14:59:04 +02:00

2026-10-05-community-rtx5080-9900x3d

bench: community report, RTX 5080, Ryzen 9 9900X3D, Coder IQ1_M and others (PR #1243 )

2026-10-07 03:04:53 +02:00

2026-10-05-community-rtx-4070-ti-super

Community benchmark: RTX 4070 Ti SUPER 16 GiB (Shin-BlackMamba, upstream 6f32ec07) (#995 )

2026-10-06 13:19:05 +02:00

2026-10-05-community-rtx-5060-ti-x2

Community benchmark: 2x RTX 5060 Ti (16 GB), Xeon E5-2690 v4, IQ3_XXS, engine 0.1.35 (#508 )

2026-10-06 13:19:03 +02:00

2026-10-05-community-rtx-5090-9700x-iq3s

Add an RTX 5090 + Ryzen 7 9700X IQ3_S benchmark on engine 0.1.39 (#811 )

2026-10-06 13:19:03 +02:00

2026-10-05-community-rx-6700-xt

bench: community report, RX 6700 XT 12 GB (gfx1031), Ryzen 7 5700G, I… (#1027 )

2026-10-06 13:19:05 +02:00

2026-10-05-community-v100-sxm2

Community benchmark: Tesla V100-SXM2-32GB on Windows (CUDA 12.6 / MSVC) (#902 )

2026-10-06 13:19:04 +02:00

2026-10-05-rdna2-rocblas-table

hip: a rocBLAS solution table for the FP16 prompt GEMMs on gfx103x (the GDN projection 7x below 1,152 tokens; short prompts -10 to -14%)

2026-10-14 07:12:35 +02:00

2026-10-06-community-2x-p100

bench: community report, 2x Tesla P100 16 GB, Flash-Next IQ3_S, 128K (PR #1157 )

2026-10-07 03:04:45 +02:00

2026-10-06-community-2x-rtx-pro-4500-engine-0.1.40.1

bench: community report, 2x RTX PRO 4500 Blackwell, Swift IQ3_XXS (PR #1226 )

2026-10-07 03:04:49 +02:00

2026-10-06-community-2x-rx-6900xt-0.1.40.1

bench: community report, 2x RX 6900 XT (gfx1030), Ryzen 5 5600X, IQ3_S, 131K (PR #1270 )

2026-10-07 03:04:56 +02:00

2026-10-06-community-2x-rx-7900-gre

bench: community report, 2x RX 7900 GRE (gfx1100), layer split, #848 / #854 (PR #1140 )

2026-10-07 03:04:44 +02:00

2026-10-06-community-4x-tesla-p100

bench: community report, 4x Tesla P100 16 GB (CUDA 12 engine), IQ3_XXS (PR #1192 )

2026-10-07 03:04:47 +02:00

2026-10-06-community-r9-2x-rtx-2080-ti

bench: community report, 2x RTX 2080 Ti (R9), see README (PR #1225 )

2026-10-07 03:04:48 +02:00

2026-10-06-community-rtx3090-egpu-64gb

bench: community report, RTX 3090 eGPU, 64 GB RAM, IQ2_XS vs IQ3_XXS (PR #1199 )

2026-10-07 03:04:48 +02:00

2026-10-06-community-rtx4090-opt-ins

bench: community report, RTX 4090, IQ3_S 140K / IQ3_XXS 200K opt-ins and tool-call hotfix (PR #1158 )

2026-10-07 03:04:46 +02:00

2026-10-06-community-rtx4090-toolcall-hotfix

bench: community report, RTX 4090, IQ3_S 140K / IQ3_XXS 200K opt-ins and tool-call hotfix (PR #1158 )

2026-10-07 03:04:46 +02:00

2026-10-06-community-rtx5070ti-5900x

bench: community report, RTX 5070 Ti + Ryzen 9 5900X, 64 GB DDR4-3600 (IQ2_XS / IQ3_XXS / IQ3_S, six-arm A/B) (#993 )

2026-10-06 13:19:04 +02:00

2026-10-06-community-rtx-5060ti

Bench: RTX 5060 Ti 16 GB on Ryzen 9 7945HX, IQ3_S (#1016 )

2026-10-06 13:19:05 +02:00

2026-10-06-community-rtx-5090-laptop

bench: community report, RTX 5090 Laptop 24 GB, Windows 11, see README (PR #1263 )

2026-10-07 03:04:54 +02:00

2026-10-06-community-rtxpro-v0140

bench: community report, RTX PRO 6000, Q4 / Q8 (MTP and ngram runs) (PR #1137 )

2026-10-07 03:04:42 +02:00

2026-10-06-community-rx-6800m

bench: RX 6800M (gfx1031) community report - self-built HIP engine on Windows (#1078 )

2026-10-06 13:19:05 +02:00

2026-10-06-community-rx-7900-xtx

bench: community report, RX 7900 XTX, Windows 11, IQ3_S (PR #1173 )

2026-10-07 03:04:47 +02:00

2026-10-06-llamacpp-0.6.0

bench: Strata vs llama.cpp v0.6.0 head-to-head script and results (3060, P100, 5070 one round)

2026-10-07 01:33:18 +02:00

2026-10-06-q8-prefill8192-rtxpro

bench: community report, RTX PRO 6000, Q8, --prefill 8192 (PR #1171 )

2026-10-07 03:04:46 +02:00

2026-10-07-community-2x-rtx-4000

Add benchmark results for 2x RTX 4000 configuration

2026-10-08 02:34:01 +02:00

2026-10-07-community-arc-b65-v01402

0.1.41 merge cleanups: bench key placeholder, B65 README author text, calibration time text, fixups.py regenerates the run-time spin bound

2026-10-08 09:58:59 +02:00

2026-10-07-community-gfx1150

hip: hipBLASLt tuning table for gfx1150 (Radeon 890M, ROCm 7.14.1 / hipBLASLt 100401)

2026-10-07 20:32:04 +02:00

2026-10-07-community-rtx5090-pro4000-helper

community bench: RTX 5090 + RTX PRO 4000 Blackwell, UD-Q4_K_XL helper cache, engine 0.1.40.3

2026-10-08 02:33:57 +02:00

2026-10-07-community-rtx-5080-docker-wsl2

Add community benchmark: RTX 5080 in Docker Desktop on Windows (WSL2)

2026-10-08 02:34:04 +02:00

2026-10-07-community-rtx-5090-laptop-ud-iq4xs

bench: community report, RTX 5090 Laptop 24 GB, UD-IQ4_XS (first NVIDIA measurement) (PR #1245 )

2026-10-07 03:04:53 +02:00

2026-10-07-community-v100-p100-helper

bench: community report, Tesla V100 32 GB + P100 16 GB expert-helper (mixed Volta/Pascal), Flash-Next IQ3_XXS (#1079 )

2026-10-08 02:34:01 +02:00

2026-10-07-community-xtx-6800xt-1m-helper

bench: mixed AMD helper layout at1M, long-output and quality results

2026-10-08 02:34:02 +02:00

2026-10-07-rdna2-qsa-select-simt

hip: the FP32 tiled QSA block scorer on HIP cards (opt-in STRATA_SELECT_SIMT=1, as on CUDA), tiled 32/128/16/8 on gfx103x

2026-10-14 07:12:53 +02:00

2026-10-07-v100-decode-kernels

bench/results: V100 decode kernels (GPU work per verify window 22.8 -> 21.2 ms); docs/NVIDIA_V100.md: the decode side; tools/ab_engine.py

2026-10-07 20:42:39 +02:00

2026-10-07-v100-prompt-experts

bench/results: V100 prompt experts on FP16 tensor cores (+9.6% / +10.3% prompt speed); docs/NVIDIA_V100.md

2026-10-14 07:08:09 +02:00

2026-10-08-community-2x-rtx3080-20gb-ud-q4-k-xl-512k-pool

bench: drop the tool-attribution line from the 2x RTX 3080 report (the numbers are the submitter's)

2026-10-09 15:36:37 +02:00

2026-10-08-community-2x-titan-rtx-0.1.40.3

0.1.41 merge cleanups: bench key placeholder, B65 README author text, calibration time text, fixups.py regenerates the run-time spin bound

2026-10-08 09:58:59 +02:00

2026-10-08-community-r9700-linux

bench: single R9700 performance report (PR #1638 )

2026-10-09 15:36:36 +02:00

2026-10-08-community-rtx4070-iq2-xs

bench: COMMUNITY.md rows and the docs list entries for #1739 #1751 #1784 #1793 #1799 #1802 #1811

2026-10-10 10:47:13 +02:00

2026-10-08-community-rtx4090-iq3xxs-200k-code

bench: community report - RTX 4090, Flash-Next IQ3_XXS at 204800, engine 0.1.40.3

2026-10-08 02:34:06 +02:00

2026-10-08-community-rtx4090-iq3xxs-200k-ru

bench: community report - RTX 4090, Flash-Next IQ3_XXS at 204800, engine 0.1.40.3

2026-10-08 02:34:06 +02:00

2026-10-08-community-rtx-5090-laptop-engine-0.1.40.3

bench: RTX 5090 Laptop GPU follow-up for 0.1.40.2 and 0.1.40.3

2026-10-08 02:34:05 +02:00

2026-10-08-community-v100-sm70-table

bench: community report, a Tesla V100 (sm_70) confirms the opt-in Volta decode kernels (PR #1647 )

2026-10-09 15:36:36 +02:00

2026-10-08-gfx906-qsa-query-swizzle

docs(gfx906): record 200K and API qualification for the combined candidate

2026-10-09 16:48:03 +02:00

2026-10-09-community-2x-rx-6900xt-0.1.41

bench: community report, 2x RX 6900 XT (gfx1030) + Ryzen 5 5600X, IQ3_S at 131K on 0.1.41 (PR #1657 )

2026-10-09 15:36:36 +02:00

2026-10-09-community-4x-rx-7900-xt

bench: community report, 4x RX 7900 XT (gfx1100), Flash-Next IQ3_S at 262K, five opt-ins off vs on, engine 0.1.41 (PR #1655 )

2026-10-09 15:36:36 +02:00

2026-10-09-community-gfx1151-ud-iq4xs-0.1.41

bench: community report, Radeon 8060S (gfx1151) on a Ryzen AI Max+ 395, UD-IQ4_XS, engine 0.1.41 (PR #1717 )

2026-10-09 15:36:36 +02:00

2026-10-09-community-p4-prefill-roofline

bench: P4 prefill 78.8 -> 114.8 tok/s with existing auto sizing; expert reuse roofline (PR #1695 )

2026-10-09 15:36:36 +02:00

2026-10-09-community-radeon-pro-vii-16gb

bench: community report, Radeon Pro VII 16 GB (gfx906), ROCm 6.2 (PR #1793 ; build and server logs left out)

2026-10-10 10:47:13 +02:00

2026-10-09-community-rtx5090-9800x3d-swift-1m

bench: RTX 5090 Swift IQ3_XXS 1M tuning and native NVMe restore (PR #1722 )

2026-10-09 16:44:46 +02:00

2026-10-09-community-rtx-3060-0141

bench: community report, RTX 3060 12 GB, Flash-Next Q2_0/IQ2_XS/IQ3_XXS/IQ3_S, engine 0.1.41 (PR #1725 )

2026-10-09 16:44:46 +02:00

2026-10-09-community-rtx-5060-ti-e5-2680-v4

bench: community report, RTX 5060 Ti / Xeon E5-2680 v4 (PR #1784 )

2026-10-10 10:47:13 +02:00

2026-10-09-community-rtx-5060ti-wsl2-iq3_s-0141

bench: COMMUNITY.md rows and the docs list entries for #1739 #1751 #1784 #1793 #1799 #1802 #1811

2026-10-10 10:47:13 +02:00

2026-10-09-community-rtx-5080-4060ti-0.1.41

bench: community report, RTX 5080 16 GB + RTX 4060 Ti 8 GB (helper cache), IQ3_S at 4K / 32K / 128K on 0.1.41 (PR #1681 )

2026-10-09 15:36:36 +02:00

2026-10-09-community-rtx-5080-w7-2475x-ud-iq4xs-786k

bench #1726 : keep the reporter's background-workload list complete (a desktop chat app was open)

2026-10-09 16:51:42 +02:00

2026-10-09-community-strix-halo-laptop-70w

bench: community report, Strix Halo laptop (Radeon 8060S, 70 W), v0.1.41, UD-IQ4_XS (PR #1716 )

2026-10-09 15:36:36 +02:00

2026-10-09-rope-yarn-long-context

bench: existing RoPE vs YaRN at 512K and 1M, with setup guide (PR #1692 )

2026-10-09 15:36:36 +02:00

2026-10-10-community-2x-p40-iq3_s

bench: community report, 2x Tesla P40, IQ3_S, 131K (PR #1799 )

2026-10-10 10:47:13 +02:00

2026-10-10-community-rx-9060-xt

bench: COMMUNITY.md rows and the docs list entries for #1739 #1751 #1784 #1793 #1799 #1802 #1811

2026-10-10 10:47:13 +02:00

2026-10-10-community-rx-9070

bench: community report, RX 9070 16 GB, IQ3_XXS, engine 0.1.41 (PR #1811 )

2026-10-10 10:47:12 +02:00

2026-10-11-community-rtx3090-5600x-arena-async

serve: --adapt-async 1 in the arena mode (one GPU, copy-in only, opt-in)

2026-10-13 14:50:46 +02:00

2026-10-11-community-rtx3090-5600x-arena-pin-by-need

arena: pin by need (STRATA_PIN_BY_NEED, opt-in) + PCIe share floor (STRATA_PCIE_MIN1, opt-in)

2026-10-13 14:50:46 +02:00

2026-10-11-gfx1100-qsa-scores

hip: add opt-in exact gfx1100 QSA prefill scorer

2026-10-14 09:59:02 +02:00

2026-10-11-rtx-5090-callback-affinity

Start the GPU host-callback thread before the host is pinned, by default (STRATA_ADAPT_CALLBACK_WARMUP=0 turns it off)

2026-10-11 21:05:23 +02:00

2026-10-12-community-gfx1103

hip: a gfx1103 hipBLASLt tuning table (Radeon 780M)

2026-10-14 09:59:02 +02:00

2026-10-12-hc-persist-gfx1030

hc persist results: ten interleaved pairs on 0.1.43.1 with no other switch (-6.6 % / -4.4 % ms per verify window at 4K / 32K, 10 of 10 pairs)

2026-10-14 09:59:02 +02:00

2026-10-12-p40-sm61-kernels

docs + bench: the Tesla P40 (sm_61) profile, the four changes and their measurements

2026-10-13 04:12:24 +02:00

2026-10-20-0.2.1-bench

bench: results of the first real runs on the 0.2.1 release (RTX 3060, 2x RTX A4000), protocol text from the measured counts

2026-10-20 12:36:35 +02:00

2026-10-20-0.2.1-idle-gpu

2507-0: measure the idle GPUs with the existing switches (z440 2x A4000, P2003 RTX 3060)

2026-10-20 12:35:59 +02:00

COMMUNITY.md

hip: a gfx1103 hipBLASLt tuning table (Radeon 780M)

2026-10-14 09:59:02 +02:00

exchange-rotation-ab.json

bench: document the standard Q2 token-identical rotation pair

2026-10-05 13:19:38 +02:00

exchange-rotation-token-ids.json

bench: verify rotation-only build with an exact native 1K pair

2026-10-05 13:19:38 +02:00