Files
CTI-Inference-Opt/代码
Serendipity f3fe2df610 revert: 移除所有 torch.compile(四战全败),回到稳定版 58.49
torch.compile 全模式验证:
- reduce-overhead: 199s (+126%)
- default 全模型: 118s (+34%)
- default Expert: 108.6s (+23%)
- dynamic=True: 102.6s (+17%)
MoE 动态路由 + 可变序列长度,与任何 JIT 编译不兼容
2026-06-13 14:45:32 +08:00
..