欢迎光临
我们一直在努力

Qwen3.6-35B-A3B反复的参数调试可玩性很高

实测跑:100K的上下文,效果还可以

对显存暂用放非常友好,目前很稳定

效果图

这是结果

这是调优的参数:GTX 1060 6G 100K上下文,挺好玩的

@echo off
chcp 437 > nul
title Qwen3.6-35B-A3B-Uncensored

:: ========== Colors ==========
set "ESC=["
set "RED=%ESC%91m"
set "GREEN=%ESC%92m"
set "YELLOW=%ESC%93m"
set "BLUE=%ESC%94m"
set "PURPLE=%ESC%95m"
set "CYAN=%ESC%96m"
set "WHITE=%ESC%97m"
set "RESET=%ESC%0m"

:: ========== Paths ==========
set SERVER=C:\\Users\\AI\\Llama Server\\llama.cpp\\bin\\llama-server.exe
set MODEL=C:\\Users\\AI\\Llama Server\\llama.cpp\\models\\Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf
set MMPROJ=C:\\Users\\AI\\Llama Server\\llama.cpp\\models\\mmproj-Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-f16.gguf
set SLOTS=C:\\Users\\AI\\Llama Server\\llama.cpp\\data\\slots

:: ========== File Checks ==========
if not exist "%MODEL%"  (echo %RED%[X] Model file missing%RESET% & pause & exit /b 1)
if not exist "%MMPROJ%" (echo %RED%[X] mmproj file missing%RESET% & pause & exit /b 1)
if not exist "%SERVER%" (echo %RED%[X] Server binary missing%RESET% & pause & exit /b 1)
if not exist "%SLOTS%"  mkdir "%SLOTS%" 2>nul

cls
echo.
echo %PURPLE%  +—————————————————+%RESET%
echo %PURPLE%  ^|%RESET%  %GREEN%Qwen3.6-35B-A3B-Uncensored%RESET%                         %PURPLE%^|%RESET%
echo %PURPLE%  ^|%RESET%  %CYAN%HauhauCS Aggressive ^| MoE ^| Q4_K_M%RESET%               %PURPLE%^|%RESET%
echo %PURPLE%  +—————————————————+%RESET%
echo.
echo %CYAN%   CUDA 12.8 self-compiled  —  GPU: 999L  —  CPU MoE: 999  —  Context: 100K  —  np: 2%RESET%
echo %CYAN%   KV unified  —  reasoning: off  —  rope: 2.5  —  flash-attn%RESET%
echo %YELLOW%   Port: 8080  —  Host: 0.0.0.0 — Host: 127.0.0.1 %RESET%
echo.
echo %WHITE%   Model: Q4_K_M (~20 GB)  +  mmproj F16 (~860 MB)%RESET%
echo %WHITE%   Path:  C:\\Users\\AI\\Llama Server\\llama.cpp\\models\\%RESET%
echo.

"%SERVER%" ^
  -m "%MODEL%" ^
  –mmproj "%MMPROJ%" ^
  –host 0.0.0.0 –port 8080 ^
  –n-gpu-layers 999 ^
  –n-cpu-moe 999 ^
  -t 8 ^
  –ctx-size 200000 ^
  -np 2 ^
  –batch-size 512 ^
  –ubatch-size 256 ^
  –cache-type-k q4_0 ^
  –cache-type-v q4_0 ^
  –cache-ram 5120 ^
  –rope-scale 2.5 ^
  –reasoning on ^
  –no-mmap ^
  –slot-save-path "%SLOTS%" ^
  –no-warmup –prio 2 ^
  –temp 0.30 –top-k 60 –top-p 0.80 –min-p 0.15 –repeat-penalty 1.15 ^ –repeat-penalty 1.12 ^ –repeat-last-n 512 ^
  –alias "Qwen3.6-35B-Uncensored" ^
  –timeout 300 ^
  –ui –metrics ^
  –flash-attn auto
  –log-disable   

pause

赞(0)
未经允许不得转载:171主机测评 » Qwen3.6-35B-A3B反复的参数调试可玩性很高
分享到: 更多 (0)

评论 抢沙发

  • 昵称 (必填)
  • 邮箱 (必填)
  • 网址