|
| 1 | +// Licensed under the BSD 3-Clause License (the "License"); |
| 2 | +// you may not use this file except in compliance with the License. |
| 3 | +// You may obtain a copy of the License at |
| 4 | +// |
| 5 | +// Unless required by applicable law or agreed to in writing, software |
| 6 | +// distributed under the License is distributed on an "AS IS" BASIS, |
| 7 | +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 8 | +// See the License for the specific language governing permissions and |
| 9 | +// limitations under the License. |
| 10 | +#include "defines.h" |
| 11 | +#include "build_tree_tiling.h" |
| 12 | +#include "tiling/platform/platform_ascendc.h" |
| 13 | +#include "aclrtlaunch_build_tree_efficient.h" |
| 14 | +#include "torch_helper.h" |
| 15 | + |
| 16 | +namespace sglang { |
| 17 | +namespace npu_kernel { |
| 18 | +constexpr uint32_t PADDING_BYTE = 32U; |
| 19 | + |
| 20 | +at::Tensor get_tiling(int32_t &block_dim, int32_t &workspace_size, int32_t batch_size, int32_t mask_size, |
| 21 | + int64_t topk, int64_t depth, int64_t draft_token_num, int64_t tree_mask_mode) |
| 22 | +{ |
| 23 | + auto ascendc_platform = platform_ascendc::PlatformAscendCManager::GetInstance(); |
| 24 | + int32_t max_aiv_core = static_cast<int32_t>(ascendc_platform->GetCoreNumAiv()); |
| 25 | + block_dim = std::min(max_aiv_core, batch_size); |
| 26 | + workspace_size = static_cast<int32_t>(ascendc_platform->GetLibApiWorkSpaceSize()); |
| 27 | + |
| 28 | + // align to 32 bytes |
| 29 | + int32_t tiling_size = (sizeof(BuildTreeTilingData) + PADDING_BYTE - 1) / PADDING_BYTE * PADDING_BYTE; |
| 30 | + auto tiling_buffer = at::empty({tiling_size}, at::TensorOptions().dtype(at::kByte).device(at::kCPU)); |
| 31 | + |
| 32 | + BuildTreeTilingData *tiling_data = reinterpret_cast<BuildTreeTilingData *>(tiling_buffer.data_ptr()); |
| 33 | + tiling_data->batch_size = batch_size; |
| 34 | + tiling_data->mask_size = mask_size; |
| 35 | + tiling_data->topk = topk; |
| 36 | + tiling_data->depth = depth; |
| 37 | + tiling_data->draft_token_num = draft_token_num; |
| 38 | + tiling_data->tree_mask_mode = tree_mask_mode; |
| 39 | + |
| 40 | + auto num_big_core = batch_size % max_aiv_core; |
| 41 | + tiling_data->big_core_num = num_big_core == 0 ? block_dim : num_big_core; |
| 42 | + tiling_data->big_core_tile_num = (batch_size + num_big_core - 1) / num_big_core; |
| 43 | + tiling_data->small_core_tile_num = batch_size / num_big_core; |
| 44 | + |
| 45 | + auto tiling_tensor = TorchNpuHepler::CopyTensorHostToDevice(tiling_buffer); |
| 46 | + return tiling_tensor; |
| 47 | +} |
| 48 | + |
| 49 | +HOST_API void build_tree_efficient(const at::Tensor &parent_list, |
| 50 | + const at::Tensor &selected_index, |
| 51 | + const at::Tensor &verified_seq_len, |
| 52 | + const at::Tensor &tree_mask, |
| 53 | + const at::Tensor &positions, |
| 54 | + const at::Tensor &retrive_index, |
| 55 | + const at::Tensor &retrive_next_token, |
| 56 | + const at::Tensor &retrive_next_sibling, |
| 57 | + int64_t topk, |
| 58 | + int64_t depth, |
| 59 | + int64_t draft_token_num, |
| 60 | + int64_t tree_mask_mode) |
| 61 | +{ |
| 62 | + if (QLEN_ONLY_BITPACKING == tree_mask_mode) { |
| 63 | + throw std::runtime_error("Not implemented"); |
| 64 | + } |
| 65 | + |
| 66 | + if (parent_list.options().dtype() != at::kLong |
| 67 | + || selected_index.options().dtype() != at::kLong |
| 68 | + || verified_seq_len.options().dtype() != at::kLong |
| 69 | + || tree_mask.options().dtype() != at::kBool |
| 70 | + || positions.options().dtype() != at::kLong |
| 71 | + || retrive_index.options().dtype() != at::kLong |
| 72 | + || retrive_next_token.options().dtype() != at::kLong |
| 73 | + || retrive_next_sibling.options().dtype() != at::kLong) { |
| 74 | + throw std::invalid_argument("Invaild input datetype. " \ |
| 75 | + "Support combo: int64, int64, int64, bool, int64, int64, int64, int64"); |
| 76 | + } |
| 77 | + int32_t block_dim; |
| 78 | + int32_t workspace_size; |
| 79 | + int32_t batch_size = parent_list.sizes()[0]; |
| 80 | + int32_t mask_size = tree_mask.size(0); |
| 81 | + |
| 82 | + at::Tensor tiling_tensor = get_tiling(block_dim, workspace_size, batch_size, mask_size, topk, depth, draft_token_num, |
| 83 | + tree_mask_mode); |
| 84 | + |
| 85 | + auto workspace_tensor = |
| 86 | + at::empty({workspace_size}, at::TensorOptions().dtype(at::kByte).device(parent_list.options().device())); |
| 87 | + /* lauch the kernal function via torch */ |
| 88 | + EXEC_KERNEL_CMD(build_tree_efficient, block_dim, parent_list, selected_index, verified_seq_len, tree_mask, |
| 89 | + positions, retrive_index, retrive_next_token, retrive_next_sibling, workspace_tensor, tiling_tensor); |
| 90 | +} |
| 91 | + |
| 92 | +} |
| 93 | +} |
0 commit comments