| 1 | # Common dependencies |
| 2 | -r common.txt |
| 3 | |
| 4 | # The version of gRPC libraries should be consistent with each other |
| 5 | grpcio==1.78.0 |
| 6 | grpcio-reflection==1.78.0 |
| 7 | |
| 8 | numba == 0.65.0 # Required for N-gram speculative decoding |
| 9 | |
| 10 | # Dependencies for AMD GPUs |
| 11 | datasets |
| 12 | peft |
| 13 | pytest-asyncio |
| 14 | tensorizer==2.10.1 |
| 15 | packaging>=24.2 |
| 16 | setuptools>=78.1.1,<80.0.0 |
| 17 | setuptools-scm>=8 |
| 18 | setuptools-rust>=1.9.0 |
| 19 | runai-model-streamer[s3,gcs,azure]==0.15.7 |
| 20 | conch-triton-kernels==1.2.1 |
| 21 | timm>=1.0.17 |
| 22 | # amd-quark: required for Quark quantization on ROCm |
| 23 | # To be consistent with test_quark.py |
| 24 | amd-quark==0.12.post1 |
| 25 | tilelang==0.1.10 |
| 26 | # Required apache-tvm-ffi matching tilelang version |
| 27 | apache-tvm-ffi==0.1.11 |
| 28 | # Required for faster safetensors model loading |
| 29 | fastsafetensors >= 0.3.3 |
| 30 | # Mooncake transfer engine for KV cache offloading/transfer (ROCm build, published on PyPI) |
| 31 | mooncake-transfer-engine-rocm >= 0.3.13 |