贡献 / #2387
#2387 bench: community report 2026-10-13 CPU-only Swift 1.5 IQ2_XS (Ryzen 9 3900X)
open · @monscoro · 0 评论 · 去 GitHub 看
BenchmarksSetup & installAMD / HIPModels & quantsDocumentationWindows
说明
Adds `bench/results/2026-10-13-community-cpu-swift-1-5-iq2-xs/` per docs/COMMUNITY_BENCHMARKS.md. Results-only submission: no engine or setup changes. Hardware and software - CPU only (no GPU): AMD Ryzen 9 3900X (12 cores, 24 threads), 63.9 GiB RAM, Windows 11 (10.0.26200) - Strata project 0.1.43.1, commit a0116b27, engine 0.1.43.1, install's own configuration, hip backend Model and configuration - swift-1.5-iq2_xs (Swift-Qwen3.8-Flash-Next-GSQ-RCO-abliterated IQ2_XS), 262,144-token context - --expert-cache auto --prefill auto --spec 4 --spec-min-p 0.5 --mtp rt --kv int8 --kv-resident 32768 Method - tools/strata_bench.py suite, 3 runs per test, medians with ranges; model loading not timed; one warm-up chat and a 1K prompt discarded - short chat, fresh 4K and 16K prompts, and an agent session (system prompt, four tools, five turns that each add about 1500 words) Results (medians with ranges) - short chat: decode 37.9 tok/s (34.5-41.9), TTFT 0.84 s (0.74-1.00) - 4K fresh prompt: prompt 455.6 tok/s (453.2-459.8), TTFT 8.86 s (8.78-8.91) - 16K fresh prompt: prompt 542.8 tok/s (542.0-580.9), TTFT 29.63 s (27.69-29.68) - agent session follow-up turns: decode 43.0 tok/s (33.1-56.1), conversation cache hits 68.6 % Limitations - Speed only; no correctness check was run - Background workloads, power limits, storage and PCIe link not recorded
本站相关内容
相关页面的快捷入口。