common_extension_rocm.cc

/* Copyright 2025 SGLang Team. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#include <ATen/core/dispatch/Dispatcher.h>
#include <torch/library.h>

#include "sgl_kernel_ops.h"

TORCH_LIBRARY_EXPAND(sgl_kernel, m) {

  /*
   * From FlashMLA
   */
  m.def("dcu_create_flashmla_kv_indices(Tensor req_to_token, Tensor req_pool_indices,Tensor page_kernel_lens, Tensor? kv_start_idx, Tensor kv_indices, int req_to_token_stride, int kv_indices_stride, int PAGED_SIZE) -> ()");
  m.impl("dcu_create_flashmla_kv_indices", torch::kCUDA, &dcu_create_flashmla_kv_indices);


  /*
   * From csrc/activation
   */
  m.def("silu_and_mul(Tensor! out, Tensor input) -> ()");
  m.impl("silu_and_mul", torch::kCUDA, &silu_and_mul);

  m.def("gelu_tanh_and_mul(Tensor! out, Tensor input) -> ()");
  m.impl("gelu_tanh_and_mul", torch::kCUDA, &gelu_tanh_and_mul);

  m.def("gelu_and_mul(Tensor! out, Tensor input) -> ()");
  m.impl("gelu_and_mul", torch::kCUDA, &gelu_and_mul);

  m.def("gelu_quick(Tensor! out, Tensor input) -> ()");
  m.impl("gelu_quick", torch::kCUDA, &gelu_quick);


  /*
   * From csrc/attention
   */
  m.def("merge_state_v2(Tensor v_a, Tensor s_a, Tensor v_b, Tensor s_b, Tensor! v_merged, Tensor! s_merged) -> ()");
  m.impl("merge_state_v2", torch::kCUDA, &merge_state_v2);
  /*
   * From csrc/allreduce
   */
  m.def(
      "init_custom_ar(Tensor meta, Tensor rank_data, "
      "str[] handles, int[] offsets, int rank, "
      "bool full_nvlink) -> int");
  m.impl("init_custom_ar", torch::kCUDA, &init_custom_ar);

  m.def("all_reduce_reg(int fa, Tensor inp, Tensor! out) -> ()");
  m.impl("all_reduce_reg", torch::kCUDA, &all_reduce_reg);

  m.def(
      "all_reduce_unreg(int fa, Tensor inp, Tensor reg_buffer, Tensor! out) -> "
      "()");
  m.impl("all_reduce_unreg", torch::kCUDA, &all_reduce_unreg);

  m.def("dispose", &dispose);

  m.def("meta_size", &meta_size);

  m.def(
      "register_buffer(int fa, Tensor t, str[] handles, "
      "int[] offsets) -> ()");
  m.impl("register_buffer", torch::kCUDA, &register_buffer);

  m.def("get_graph_buffer_ipc_meta", &get_graph_buffer_ipc_meta);
  m.def("register_graph_buffers", &register_graph_buffers);

  m.def("allocate_meta_buffer", &allocate_meta_buffer);
  m.impl("allocate_meta_buffer", torch::kCUDA, &allocate_meta_buffer);

  m.def("get_meta_buffer_ipc_handle", &get_meta_buffer_ipc_handle);
  m.impl("get_meta_buffer_ipc_handle", torch::kCPU, &get_meta_buffer_ipc_handle);

  // quick allreduce
#ifdef USE_ROCM
  m.def(
      "qr_all_reduce(int fa, Tensor inp, Tensor out, int quant_level, bool "
      "cast_bf2half) -> ()");
  m.impl("qr_all_reduce", torch::kCUDA, &qr_all_reduce);

  m.def("init_custom_qr", &init_custom_qr);
  m.def("qr_destroy", &qr_destroy);

  m.def("qr_get_handle", &qr_get_handle);

  m.def("qr_open_handles(int _fa, Tensor[](b!) handles) -> ()");
  m.impl("qr_open_handles", torch::kCPU, &qr_open_handles);

  // Max input size in bytes
  m.def("qr_max_size", &qr_max_size);
#endif

  /*
   * From csrc/moe
   */
  m.def(
      "moe_align_block_size(Tensor topk_ids, int num_experts, int block_size, Tensor! sorted_token_ids, Tensor! "
      "experts_ids, Tensor! num_tokens_post_pad, Tensor! cumsum_buffer, bool "
      "pad_sorted_token_ids) -> ()");
  m.impl("moe_align_block_size", torch::kCUDA, &moe_align_block_size);

  m.def("topk_softmax(Tensor! topk_weights, Tensor! topk_indices, Tensor gating_output, bool renormalize) -> ()");
  m.impl("topk_softmax", torch::kCUDA, &topk_softmax);

  /*
   * From csrc/speculative
   */
  m.def(
      "verify_tree_greedy(Tensor! predicts, Tensor! accept_index, Tensor! accept_token_num, "
      "Tensor candidates, Tensor retrive_index, Tensor retrive_next_token, Tensor retrive_next_sibling, "
      "Tensor target_predict, int cuda_stream) -> ()");
  m.impl("verify_tree_greedy", torch::kCUDA, &verify_tree_greedy);

  m.def(
      "build_tree_kernel_efficient(Tensor parent_list, Tensor selected_index, Tensor verified_seq_len, "
      "Tensor! tree_mask, Tensor! positions, Tensor! retrive_index, Tensor! retrive_next_token, "
      "Tensor! retrive_next_sibling, int topk, int depth, int draft_token_num, int tree_mask_mode) -> "
      "()");
  m.impl("build_tree_kernel_efficient", torch::kCUDA, &build_tree_kernel_efficient);

  /*
   * From XGrammar
   */
  m.def("apply_token_bitmask_inplace_cuda(Tensor logits, Tensor bitmask, Tensor? indices=None) -> ()");
  m.impl("apply_token_bitmask_inplace_cuda", &ApplyTokenBitmaskInplace);

  /*
   * From csrc/kvcacheio
   */
  m.def("dcu_create_extend_after_decode_spec_info(Tensor verified_id, Tensor seq_lens, Tensor accept_lens, Tensor positions, Tensor new_verified_id, int bs) -> ()");
  m.impl("dcu_create_extend_after_decode_spec_info", torch::kCUDA, &dcu_create_extend_after_decode_spec_info);
  m.def("dcu_create_chunked_prefix_cache_kv_indices(Tensor req_to_token, Tensor req_pool_indices, Tensor chunk_starts, Tensor chunk_seq_lens, Tensor chunk_cu_seq_lens, Tensor chunk_kv_indices, int col_num, int bs) -> ()");
  m.impl("dcu_create_chunked_prefix_cache_kv_indices", torch::kCUDA, &dcu_create_chunked_prefix_cache_kv_indices);
  m.def("dcu_assign_extend_cache_locs(Tensor req_pool_indices, Tensor req_to_token, Tensor start_offset, Tensor end_offset, Tensor out_cache_loc, int pool_len, int bs) -> ()");
  m.impl("dcu_assign_extend_cache_locs", torch::kCUDA, &dcu_assign_extend_cache_locs);
  m.def("dcu_get_last_loc(Tensor req_to_token, Tensor req_pool_indices, Tensor prefix_lens) -> Tensor");
  m.impl("dcu_get_last_loc", torch::kCUDA, &dcu_get_last_loc); 
  m.def("dcu_assign_req_to_token_pool(Tensor req_pool_indices_ptr,Tensor req_to_token_ptr,Tensor allocate_lens_ptr,Tensor new_allocate_lens,Tensor out_cache_loc_ptr,int shape,int bs) -> ()");
  m.impl("dcu_assign_req_to_token_pool",torch::kCUDA,&dcu_assign_req_to_token_pool);   


  m.def("dcu_alloc_extend_kernel(Tensor pre_lens_ptr, Tensor seq_lens_ptr, Tensor last_loc_ptr, Tensor free_page_ptr, Tensor out_indices, int bs, int page_size) -> ()");
  m.impl("dcu_alloc_extend_kernel", torch::kCUDA, &dcu_alloc_extend_kernel);
  m.def("dcu_alloc_decode_kernel(Tensor seq_lens_ptr, Tensor last_loc_ptr, Tensor free_page_ptr, Tensor out_indices, int bs, int page_size) -> ()");
  m.impl("dcu_alloc_decode_kernel", torch::kCUDA, &dcu_alloc_decode_kernel);
  m.def(
      "transfer_kv_per_layer(Tensor src_k, Tensor dst_k, Tensor src_v, Tensor dst_v, Tensor src_indices, Tensor "
      "dst_indices, int item_size, int block_quota, int num_warps_per_block) -> ()");
  m.impl("transfer_kv_per_layer", torch::kCUDA, &transfer_kv_per_layer);
  m.def(
      "transfer_kv_per_layer_pf_lf(Tensor src_k, Tensor dst_k, Tensor src_v, Tensor dst_v, Tensor src_indices, Tensor "
      "dst_indices, int layer_id, int item_size, int src_layout_dim, int block_quota, int num_warps_per_block) -> ()");
  m.impl("transfer_kv_per_layer_pf_lf", torch::kCUDA, &transfer_kv_per_layer_pf_lf);
  m.def(
      "transfer_kv_all_layer(Tensor src_k_layers, Tensor dst_k_layers, Tensor src_v_layers, Tensor dst_v_layers, "
      "Tensor src_indices, Tensor dst_indices, int item_size, int num_layers, int block_quota, int "
      "num_warps_per_block) -> ()");
  m.impl("transfer_kv_all_layer", torch::kCUDA, &transfer_kv_all_layer);
  m.def(
      "transfer_kv_all_layer_lf_pf(Tensor src_k_layers, Tensor dst_k, Tensor src_v_layers, Tensor dst_v, "
      "Tensor src_indices, Tensor dst_indices, int item_size, int dst_layout_dim, int num_layers, int block_quota, int "
      "num_warps_per_block) -> ()");
  m.impl("transfer_kv_all_layer_lf_pf", torch::kCUDA, &transfer_kv_all_layer_lf_pf);
  m.def(
      "transfer_kv_per_layer_mla(Tensor src, Tensor dst, Tensor src_indices, Tensor dst_indices, int item_size, int "
      "block_quota, int num_warps_per_block) -> ()");
  m.impl("transfer_kv_per_layer_mla", torch::kCUDA, &transfer_kv_per_layer_mla);
  m.def(
      "transfer_kv_per_layer_mla_pf_lf(Tensor src, Tensor dst, Tensor src_indices, Tensor dst_indices, int layer_id, "
      "int item_size, int src_layout_dim, int block_quota, int num_warps_per_block) -> ()");
  m.impl("transfer_kv_per_layer_mla_pf_lf", torch::kCUDA, &transfer_kv_per_layer_mla_pf_lf);
  m.def(
      "transfer_kv_all_layer_mla(Tensor src_layers, Tensor dst_layers, Tensor src_indices, Tensor dst_indices, int "
      "item_size, int num_layers, int block_quota, int num_warps_per_block) -> ()");
  m.impl("transfer_kv_all_layer_mla", torch::kCUDA, &transfer_kv_all_layer_mla);
  m.def(
      "transfer_kv_all_layer_mla_lf_pf(Tensor src_layers, Tensor dst, Tensor src_indices, Tensor dst_indices, "
      "int item_size, int dst_layout_dim, int num_layers, int block_quota, int num_warps_per_block) -> ()");
  m.impl("transfer_kv_all_layer_mla_lf_pf", torch::kCUDA, &transfer_kv_all_layer_mla_lf_pf);
  m.def(
      "transfer_kv_direct(Tensor[] src_layers, Tensor[] dst_layers, Tensor src_indices, Tensor dst_indices, int "
      "page_size) -> ()");
  m.impl("transfer_kv_direct", torch::kCUDA, &transfer_kv_direct);
  m.def(
      "transfer_kv_per_layer_direct_pf_lf(Tensor[] src_ptrs, Tensor[] dst_ptrs, Tensor src_indices, "
      "Tensor dst_indices, int layer_id, int page_size)->() ");
  m.impl("transfer_kv_per_layer_direct_pf_lf", torch::kCUDA, &transfer_kv_per_layer_direct_pf_lf);
  m.def(
      "transfer_kv_all_layer_direct_lf_pf(Tensor[] src_ptrs, Tensor[] dst_ptrs, Tensor src_indices, "
      "Tensor dst_indices, int page_size) ->() ");
  m.impl("transfer_kv_all_layer_direct_lf_pf", torch::kCUDA, &transfer_kv_all_layer_direct_lf_pf);
}

REGISTER_EXTENSION(common_ops)