Add scripts/reproduce.py + benchmark_latency.py; pin reqs to working transformers/torch range; refresh README with adapter overhead numbers
8e4e948 verified | torch>=2.1,<2.6 | |
| transformers>=4.40,<4.54 | |
| numpy>=1.24,<3.0 | |
| safetensors>=0.4 | |
| torch>=2.1,<2.6 | |
| transformers>=4.40,<4.54 | |
| numpy>=1.24,<3.0 | |
| safetensors>=0.4 | |