f4e2264a83
Copied from upstream_ref (NOT rewritten — exact upstream code):
ixformer C++ API (the authoritative header):
include/ixformer.h — ixformer::infer namespace: topk_softmax,
moe_compute_token_index_api, moe_w16a16_group_gemm, moe_expand_input,
moe_output_reduce_sum, silu_and_mul, rms_norm, xllm_paged_attention, etc.
include/ilu_ops_api.h — xllm::kernel::ilu namespace: moe_active_topk,
moe_gen_idx, moe_expand_input, group_gemm, moe_combine_result,
batch_prefill, batch_decode, rms_norm, matmul, act_and_mul, etc.
ILU kernel wrappers (call ixformer::infer directly):
csrc/ilu_kernel_fused_moe.cpp — topk routing + gen_idx + expand + combine
csrc/ilu_kernel_group_gemm.cpp — batched expert GEMM
csrc/ilu_kernel_{activation,norm,rope,matmul,attention}.cpp
ILU layer implementations (full pipeline):
csrc/ilu_layer_fused_moe.{cpp,h} — 797 lines, the complete MoE pipeline
that competitor 168 ran as corex_moe.py
csrc/ilu_layer_attention.{cpp,h} — prefill/decode attention dispatch
CUDA MoE kernels (from xllm + ds_vllm):
csrc/moe/moe_topk_softmax_kernels.cuh — CUB BlockReduce + warp topk
csrc/moe/moe_topk_sigmoid_kernels.cuh — sigmoid scoring variant
csrc/moe/moe_topk.cuh + moe_fused_topk.cu — entry points
csrc/moe/moeTopKFuncs.cuh — TRT-LLM derived vllm-compatible topk
csrc/moe/moe_ops.h + moe_align_sum_kernels.cu — alignment kernels
Common layer headers:
csrc/common_fused_moe{,_base}.h + common_moe_fused_topk.{cpp,h}
100 lines
3.9 KiB
C++
100 lines
3.9 KiB
C++
/* Copyright 2025 The xLLM Authors. 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
|
|
|
|
https://github.com/jd-opensource/xllm/blob/main/LICENSE
|
|
|
|
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 <glog/logging.h>
|
|
|
|
#include "ilu_ops_api.h"
|
|
|
|
namespace xllm::kernel::ilu {
|
|
|
|
std::tuple<torch::Tensor, torch::Tensor> moe_active_topk(
|
|
const torch::Tensor& input,
|
|
int64_t topk,
|
|
int64_t num_expert_group,
|
|
int64_t topk_group,
|
|
bool normalize,
|
|
const std::optional<torch::Tensor>& mask,
|
|
const std::string& normed_by,
|
|
const std::string& scoring_func,
|
|
double route_scale,
|
|
const std::optional<torch::Tensor>& e_score_correction_bias) {
|
|
torch::Tensor input_ = input.to(torch::kFloat32);
|
|
auto reduce_weight =
|
|
torch::empty({input.size(0), topk},
|
|
torch::dtype(torch::kFloat).device(input.device()));
|
|
auto topk_indices =
|
|
torch::empty({input.size(0), topk},
|
|
torch::dtype(torch::kInt32).device(input.device()));
|
|
auto token_expert_indices =
|
|
torch::empty({input.size(0), topk},
|
|
torch::dtype(torch::kInt32).device(input.device()));
|
|
|
|
infer::topk_softmax(
|
|
reduce_weight, topk_indices, token_expert_indices, input_, false);
|
|
|
|
auto tt = reduce_weight.sum(-1);
|
|
if (normalize) {
|
|
reduce_weight = reduce_weight / reduce_weight.sum(-1).unsqueeze(-1);
|
|
}
|
|
return std::make_tuple(reduce_weight, topk_indices);
|
|
}
|
|
|
|
std::vector<torch::Tensor> moe_gen_idx(torch::Tensor& expert_id,
|
|
int64_t expert_num) {
|
|
auto src_dst = expert_id.new_empty({expert_id.numel()});
|
|
auto dst_src = torch::empty_like(src_dst);
|
|
auto expert_sizes_gpu = expert_id.new_empty({expert_num});
|
|
auto expert_sizes_gpu_cumsum = expert_id.new_zeros({expert_id.numel() + 1});
|
|
infer::moe_compute_token_index_api(expert_id,
|
|
src_dst,
|
|
dst_src,
|
|
expert_sizes_gpu,
|
|
/*expert_mask=*/std::nullopt,
|
|
/*expert_sizes_cpu*/ std::nullopt,
|
|
/*expert_sizes_gpu*/ std::nullopt,
|
|
0,
|
|
expert_num,
|
|
expert_num);
|
|
|
|
expert_sizes_gpu_cumsum = expert_sizes_gpu.cumsum(-1);
|
|
return {src_dst, dst_src, expert_sizes_gpu, expert_sizes_gpu_cumsum};
|
|
}
|
|
|
|
torch::Tensor moe_expand_input(const torch::Tensor& input,
|
|
const torch::Tensor& gather_index,
|
|
const torch::Tensor& combine_idx,
|
|
int64_t topk) {
|
|
int64_t dst_tokens = input.size(0) * topk;
|
|
auto output = input.new_empty({dst_tokens, input.size(1)});
|
|
infer::moe_expand_input(
|
|
output, input, combine_idx, gather_index, dst_tokens, topk);
|
|
|
|
return output;
|
|
}
|
|
|
|
torch::Tensor moe_combine_result(torch::Tensor& input, torch::Tensor& weight) {
|
|
input = input.view({-1, weight.size(1), input.size(1)});
|
|
auto output = input.new_empty({input.size(0), input.size(2)});
|
|
infer::moe_output_reduce_sum(output,
|
|
input,
|
|
weight,
|
|
/*mask=*/std::nullopt,
|
|
/*extra_residual*/ std::nullopt,
|
|
/*scaling_factor=*/1.0);
|
|
return output;
|
|
}
|
|
|
|
} // namespace xllm::kernel::ilu
|