RTX5060Ti16G第一次跑Strata+Qwen3.8-Flash-Next-GSQ-RCO-IQ3 体验好丝滑
-
首字吞吐(Prefill):1242 tok/s 的极速读入!Prompt Tokens: 39266 tokens(约 4 万 Token)
耗时: 31s。一口气给它喂了一整部 181KB(约 3 万汉字)的完整小说《无声》,它仅仅花了 31 秒钟,就把整整 4 万 Token 的全部细节彻底消化完毕!

狂飙到了 57.46 t/s(在长达 1.4 万 token 的超长思考链输出下,稳定输出 4 分钟不掉速)。

下面是我的详细硬件配置和模型配置参数!


下面是strata-iq3_s.json参数
{ "exe": "E:\\strata-alesha-proQ3.8-Flash-abliterated-G-R\\Strata\\engine\\strata.exe", "args": [ "--pack", "E:\\strata-alesha-proQ3.8-Flash-abliterated-G-R\\Strata-data\\packs\\iq3_s", "--native", "..\\Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "--ple-gguf", "..\\Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00002-of-00002.gguf", "--expert-profile", "E:\\strata-alesha-proQ3.8-Flash-abliterated-G-R\\Strata\\data\\expert-profile.bin", "--expert-cache", "auto", "--prefill", "auto", "--spec", "4", "--spec-min-p", "0.5", "--mtp", "E:\\strata-alesha-proQ3.8-Flash-abliterated-G-R\\Strata-data\\mtp\\rt", "--max-context", "65536", "--kv", "k8v4", "--resident-experts", "--kv-resident", "32768", "--vision", "--vram-reserve-mib", "700", "--control-vector-scaled", "E:\\strata-alesha-proQ3.8-Flash-abliterated-G-R\\Huihui-Qwen3.8-Flash-Next-refusal-direction-r.gguf:1.0", "--control-vector-layer-range", "4", "44", "--cvec-mode", "project", "--cvec-dir", "per-layer" ], "cwd": "E:\\strata-alesha-proQ3.8-Flash-abliterated-G-R\\Strata", "tokenizer": "E:\\strata-alesha-proQ3.8-Flash-abliterated-G-R\\Strata-data\\packs\\iq3_s\\tokenizer", "model_name": "qwen3.8-flash-next-iq3_s", "log": "E:\\strata-alesha-proQ3.8-Flash-abliterated-G-R\\Strata\\strata-iq3_s.log", "lib_dirs": [ "E:\\strata-alesha-proQ3.8-Flash-abliterated-G-R\\Strata\\.venv\\Lib\\site-packages\\nvidia\\cu13\\bin\\x86_64" ], "port": 8080, "gpu": 0, "gpus_asked": true, "vision": { "exe": "E:\\strata-alesha-proQ3.8-Flash-abliterated-G-R\\Strata\\engine\\strata-vision.exe", "mmproj": "..\\mmproj-Qwen3.8-Flash-Next-BF16.gguf", "model": "..\\Qwen3.8-Flash-Next-GSQ-RCO-IQ3_S-00001-of-00002.gguf", "gpu": true, "max_tokens": 1024 } }我还没对接Agent,下一步准备保持这个速度的同时,先把上下文开大128K。今天就先到这里吧!献丑了,欢迎各位大佬批评指导!
-
-