Files
qwen-test/vsp
16337 4ac406572e fix: 修复模型加载方式,改用 FP16+CPU offload
RTX 3050 8GB 无法完整加载 Qwen3.5-9B,即使量化也不行:
- bitsandbytes 4-bit 不支持 CPU offload
- bitsandbytes 8-bit 与 accelerate 存在版本兼容问题
- FP16 + CPU offload 可以加载但推理质量极差(输出乱码)
- 推理速度仅 0.4 tokens/s

结论:RTX 3050 8GB 不适合运行 Qwen3.5-9B

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-03-16 13:05:20 +08:00
..