12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182 |
- #include "cache.h"
- #include "cuda_utils.h"
- #include "ops.h"
- #include <torch/extension.h>
- PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
- // Aphrodite custom ops
- pybind11::module ops = m.def_submodule("ops", "Aphrodite custom operators");
- // Attention ops
- ops.def("paged_attention_v1", &paged_attention_v1,
- "Compute the attention between an input query and the cached "
- "keys/values using PagedAttention.");
- ops.def("paged_attention_v2", &paged_attention_v2, "PagedAttention V2.");
- // Activation ops
- ops.def("silu_and_mul", &silu_and_mul, "Activation function used in SwiGLU.");
- ops.def("gelu_and_mul", &gelu_and_mul,
- "Activation function used in GeGLU with `none` approximation.");
- ops.def("gelu_tanh_and_mul", &gelu_tanh_and_mul,
- "Activation function used in GeGLU with `tanh` approximation.");
- ops.def("gelu_new", &gelu_new, "GELU implementation used in GPT-2.");
- ops.def("gelu_fast", &gelu_fast, "Approximate GELU implementation.");
- // Layernorm
- ops.def("rms_norm", &rms_norm,
- "Apply Root Mean Square (RMS) Normalization to the input tensor.");
- ops.def("fused_add_rms_norm", &fused_add_rms_norm,
- "In-place fused Add and RMS Normalization");
- // Rotary embedding
- ops.def("rotary_embedding", &rotary_embedding,
- "Apply GPT-NeoX or GPT-J style rotary embedding to query and key");
- ops.def("batched_rotary_embedding", &batched_rotary_embedding,
- "Apply batched GPT-NeoX or GPT-J style rotary embedding to query and "
- "key");
- ops.def("moe_align_block_size", &moe_align_block_size,
- "Aligning the number of tokens to be processed by each expert such "
- "that it is divisible by the block size.");
- // Cache ops
- pybind11::module cache_ops =
- m.def_submodule("cache_ops", "Aphrodite cache ops");
- cache_ops.def("swap_blocks", &swap_blocks,
- "Swap in (out) the cache blocks from src to dst");
- cache_ops.def("copy_blocks", ©_blocks,
- "Copy the cache blocks from src to dst");
- cache_ops.def("reshape_and_cache", &reshape_and_cache,
- "Reshape the key and value tensors and cache them");
- cache_ops.def("reshape_and_cache_flash", &reshape_and_cache_flash,
- "Reshape the key and value tensors and cache them");
- cache_ops.def("convert_fp8", &convert_fp8,
- "Convert the key and value cache to fp8 data type");
- // Cuda utils
- pybind11::module cuda_utils =
- m.def_submodule("cuda_utils", "Aphrodite cuda utils");
- cuda_utils.def("get_device_attribute", &get_device_attribute,
- "Gets the specified device attribute.");
- cuda_utils.def("get_max_shared_memory_per_block_device_attribute",
- &get_max_shared_memory_per_block_device_attribute,
- "Gets the maximum shared memory per block device attribute.");
- #ifndef USE_ROCM
- // Custom all-reduce kernels
- pybind11::module custom_ar = m.def_submodule("custom_ar", "custom allreduce");
- custom_ar.def("init_custom_ar", &init_custom_ar, "init_custom_ar");
- custom_ar.def("should_custom_ar", &should_custom_ar, "should_custom_ar");
- custom_ar.def("all_reduce_reg", &all_reduce_reg, "all_reduce_reg");
- custom_ar.def("all_reduce_unreg", &all_reduce_unreg, "all_reduce_unreg");
- custom_ar.def("dispose", &dispose, "dispose");
- custom_ar.def("meta_size", &meta_size, "meta_size");
- custom_ar.def("register_buffer", ®ister_buffer, "register_buffer");
- custom_ar.def("get_graph_buffer_ipc_meta", &get_graph_buffer_ipc_meta,
- "get_graph_buffer_ipc_meta");
- custom_ar.def("register_graph_buffers", ®ister_graph_buffers,
- "register_graph_buffers");
- #endif
- }
|