# engine: sglang==0.5.12.post1
# sglang 0.5.12.post1 on CUDA 12 (torch 2.11 cu129).
#
# The PyPI sglang-kernel and sgl-deep-gemm wheels link libcudart.so.13 /
# libcublas.so.13, so they need a CUDA 13 driver; upstream also publishes cu129
# builds of the same versions, which link the .so.12 sonames and still carry
# sm_80 (A100) SASS. Those are only on sglang's own index, so they are pinned by
# URL here rather than through an extra index that would also offer nightlies.
# nvidia-cutlass-dsl and flashinfer both express the CUDA line as an extra, so
# dropping [cu13] / asking for [cu12] is all they need. cuda-python has no such
# split and has to be held below 13.
-r sglang_0.5.12.post1_common.txt

cuda-python>=12.9,<13
flash-attn-4>=4.0.0b9
flashinfer-cubin==0.6.11.post1
flashinfer-python[cu12]==0.6.11.post1
nvidia-cutlass-dsl==4.5.1
quack-kernels>=0.4.1
tilelang==0.1.8
sglang-kernel @ https://github.com/sgl-project/whl/releases/download/v0.4.2.post2/sglang_kernel-0.4.2.post2%2Bcu129-cp310-abi3-manylinux2014_x86_64.whl ; platform_machine == "x86_64"
sglang-kernel @ https://github.com/sgl-project/whl/releases/download/v0.4.2.post2/sglang_kernel-0.4.2.post2%2Bcu129-cp310-abi3-manylinux2014_aarch64.whl ; platform_machine == "aarch64"
sgl-deep-gemm @ https://github.com/sgl-project/whl/releases/download/v0.1.0/sgl_deep_gemm-0.1.0%2Bcu129-py3-none-manylinux2014_x86_64.whl ; platform_machine == "x86_64"
sgl-deep-gemm @ https://github.com/sgl-project/whl/releases/download/v0.1.0/sgl_deep_gemm-0.1.0%2Bcu129-py3-none-manylinux2014_aarch64.whl ; platform_machine == "aarch64"
