| 1 | # Common dependencies |
| 2 | -r common.txt |
| 3 | |
| 4 | numba == 0.65.0 # Required for N-gram speculative decoding |
| 5 | |
| 6 | # Dependencies for NVIDIA GPUs |
| 7 | torch==2.13.0 |
| 8 | torchaudio==2.11.0 |
| 9 | # These must be updated alongside torch |
| 10 | torchvision==0.28.0 # Required for phi3v processor. See https://github.com/pytorch/vision?tab=readme-ov-file#installation for corresponding version |
| 11 | torchcodec >= 0.14 |
| 12 | PyNvVideoCodec==2.0.4 |
| 13 | # FlashInfer should be updated together with the Dockerfile |
| 14 | # flashinfer-cubin is not on PyPI since 0.6.14; setup.py excludes it from |
| 15 | # install_requires so the published wheel does not carry an unresolvable pin |
| 16 | --extra-index-url https://flashinfer.ai/whl/ |
| 17 | flashinfer-python==0.7.0.post1 |
| 18 | flashinfer-cubin==0.7.0.post1 |
| 19 | apache-tvm-ffi==0.1.11 |
| 20 | tilelang==0.1.12 |
| 21 | nvidia-cudnn-frontend>=1.19.1 |
| 22 | # Required for LLM_NVTX_SCOPES_FOR_PROFILING=1 |
| 23 | nvtx==0.2.15 |
| 24 | # Required for faster safetensors model loading |
| 25 | fastsafetensors >= 0.3.3 |
| 26 | instanttensor >= 0.1.9 |
| 27 | |
| 28 | # QuACK and Cutlass DSL for FA4 (cute-DSL implementation) |
| 29 | nvidia-cutlass-dsl[cu13]==4.7.1 |
| 30 | quack-kernels==0.6.5 # Required by MSA |
| 31 | |
| 32 | # Tokenspeed_MLA for faster mla with spec decode |
| 33 | tokenspeed-mla==0.1.8; platform_system == "Linux" |
| 34 | |
| 35 | # Humming kernels for quantization gemm |
| 36 | humming-kernels[cu13]==0.1.16 |