build.toml

[general]
name = "quantization"

[torch]
src = [
  "core/registration.h",
  "core/scalar_type.hpp",
  "torch-ext/torch_binding.cpp",
  "torch-ext/torch_binding.h"
]
include = [ "." ]

[kernel.cutlass_w8a8]
cuda-capabilities = [ "7.5", "8.0", "8.6", "8.7", "8.9", "9.0", "10.0", "10.1", "12.0" ]
src = [
  "core/math.hpp",
  "cutlass_w8a8/common.hpp",
  "cutlass_w8a8/scaled_mm_c2x.cu",
  "cutlass_w8a8/scaled_mm_c2x.cuh",
  "cutlass_w8a8/scaled_mm_c2x_sm75_dispatch.cuh",
  "cutlass_w8a8/scaled_mm_c2x_sm80_dispatch.cuh",
  "cutlass_w8a8/scaled_mm_c2x_sm89_fp8_dispatch.cuh",
  "cutlass_w8a8/scaled_mm_c2x_sm89_int8_dispatch.cuh",
  "cutlass_w8a8/scaled_mm_entry.cu",
  "cutlass_extensions/epilogue/scaled_mm_epilogues_c2x.hpp",
  "cutlass_extensions/epilogue/broadcast_load_epilogue_c2x.hpp",
]
include = [ "." ]
depends = [ "cutlass_3_6", "torch" ]

[kernel.cutlass_w8a8_hopper]
cuda-capabilities = [ "9.0", "9.0a" ]
src = [
  "core/math.hpp",
  "cutlass_w8a8/common.hpp",
  "cutlass_w8a8/scaled_mm_c3x.cu",
  "cutlass_w8a8/scaled_mm_c3x.cuh",
  "cutlass_w8a8/scaled_mm_c3x_sm90_fp8_dispatch.cuh",
  "cutlass_w8a8/scaled_mm_c3x_sm90_int8_dispatch.cuh",
  "cutlass_extensions/common.cpp",
  "cutlass_extensions/common.hpp",
  "cutlass_extensions/epilogue/scaled_mm_epilogues_c3x.hpp",
  "cutlass_extensions/epilogue/broadcast_load_epilogue_c3x.hpp",
]
include = [ "." ]
depends = [ "cutlass_3_6", "torch" ]

[kernel.fp8_common]
language = "cuda-hipify"
cuda-capabilities = [ "7.0", "7.2", "7.5", "8.0", "8.6", "8.7", "8.9", "9.0", "10.0", "10.1", "12.0" ]
rocm-archs = [ "gfx906", "gfx908", "gfx90a", "gfx940", "gfx941", "gfx942", "gfx1030", "gfx1100", "gfx1101" ]
src = [
  "fp8/amd/hip_float8.h",
  "fp8/amd/hip_float8_impl.h",
  "fp8/common.cu",
  "fp8/common.cuh",
  "dispatch_utils.h",
  "vectorization.cuh"
]
include = [ "." ]
depends = [ "torch" ]

[kernel.fp8_marlin]
cuda-capabilities = [ "8.0", "8.6", "8.7", "8.9", "9.0", "10.0", "10.1", "12.0" ]
src = [
  "fp8/fp8_marlin.cu",
  "gptq_marlin/marlin.cuh",
  "gptq_marlin/marlin_dtypes.cuh",
]
depends = [ "torch" ]

[kernel.int8_common]
language = "cuda-hipify"
cuda-capabilities = [ "7.5", "8.0", "8.6", "8.7", "8.9", "9.0", "10.0", "10.1", "12.0" ]
rocm-archs = [ "gfx906", "gfx908", "gfx90a", "gfx940", "gfx941", "gfx942", "gfx1030", "gfx1100", "gfx1101" ]
src = [
  "compressed_tensors/int8_quant_kernels.cu",
  "dispatch_utils.h"
]
include = [ "." ]
depends = [ "torch" ]

[kernel.gptq_marlin]
cuda-capabilities = [ "8.0", "8.6", "8.7", "8.9", "9.0", "10.0", "10.1", "12.0" ]
src = [
  "core/scalar_type.hpp",
  "gptq_marlin/awq_marlin_repack.cu",
  "gptq_marlin/gptq_marlin.cu",
  "gptq_marlin/gptq_marlin_repack.cu",
  "gptq_marlin/marlin.cuh",
  "gptq_marlin/marlin_dtypes.cuh"
]
include = [ "." ]
depends = [ "torch" ]

[kernel.marlin]
cuda-capabilities = [ "8.0", "8.6", "8.7", "8.9", "9.0", "10.0", "10.1", "12.0" ]
src = [
  "core/scalar_type.hpp",
  "marlin/dense/common/base.h",
  "marlin/dense/common/mem.h",
  "marlin/dense/marlin_cuda_kernel.cu",
  "marlin/qqq/marlin_qqq_gemm_kernel.cu",
  "marlin/sparse/common/base.h",
  "marlin/sparse/common/mem.h",
  "marlin/sparse/common/mma.h",
  "marlin/sparse/marlin_24_cuda_kernel.cu"
]
include = [ "." ]
depends = [ "torch" ]