| TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) { | |
| ops.def("fp4_quantize(Tensor x, Tensor(a!) resid, bool stochastic, int seed, int step) -> (Tensor, Tensor)"); | |
| ops.def("fp4_dequantize(Tensor packed, Tensor scale, int M, int K) -> Tensor"); | |
| ops.impl("fp4_quantize", torch::kCUDA, &fp4_quantize); | |
| ops.impl("fp4_dequantize", torch::kCUDA, &fp4_dequantize); | |
| } | |
| REGISTER_EXTENSION(TORCH_EXTENSION_NAME) | |