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}
154 lines
6.2 KiB
C++
154 lines
6.2 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.
|
|
==============================================================================*/
|
|
|
|
#pragma once
|
|
|
|
#include <ATen/DynamicLibrary.h>
|
|
#include <ATen/core/dispatch/Dispatcher.h>
|
|
#include <cuda_runtime.h>
|
|
#include <glog/logging.h>
|
|
#include <torch/all.h>
|
|
|
|
#include <optional>
|
|
|
|
#include "ATen/Tensor.h"
|
|
#include "ATen/cuda/CUDAEvent.h"
|
|
#include "c10/core/Device.h"
|
|
#include "c10/core/DeviceGuard.h"
|
|
#include "c10/core/GradMode.h"
|
|
#include "c10/core/InferenceMode.h"
|
|
#include "c10/core/MemoryFormat.h"
|
|
#include "c10/core/ScalarType.h"
|
|
#include "c10/core/TensorOptions.h"
|
|
#include "c10/cuda/CUDAFunctions.h"
|
|
#include "c10/cuda/CUDAGuard.h"
|
|
#include "c10/cuda/CUDAStream.h"
|
|
#include "ixformer.h"
|
|
#include "kernels/kernels.h"
|
|
|
|
// #include "utils.h"
|
|
using namespace ixformer;
|
|
|
|
namespace xllm::kernel::ilu {
|
|
|
|
void apply_rope_pos_ids_cos_sin_cache(torch::Tensor& query,
|
|
torch::Tensor& key,
|
|
torch::Tensor& cos_sin_cache,
|
|
torch::Tensor& positions,
|
|
bool interleave);
|
|
|
|
// act_mode only support silu, gelu, gelu_tanh
|
|
void act_and_mul(torch::Tensor out,
|
|
torch::Tensor input,
|
|
const std::string& act_mode);
|
|
|
|
void reshape_paged_cache(
|
|
torch::Tensor& key, // (num_tokens, num_heads, head_size)
|
|
std::optional<torch::Tensor>& value, // (num_tokens, num_heads, head_size)
|
|
torch::Tensor& key_cache, // (num_blocks, num_heads, block_size, head_size)
|
|
std::optional<torch::Tensor>&
|
|
value_cache, // (num_blocks, num_heads, block_size, head_size)
|
|
torch::Tensor& slot_mapping); //(num_tokens)
|
|
|
|
void batch_prefill(torch::Tensor& query,
|
|
const torch::Tensor& key,
|
|
const std::optional<torch::Tensor>& value,
|
|
torch::Tensor& output,
|
|
std::optional<torch::Tensor>& output_lse,
|
|
const std::optional<torch::Tensor>& q_cu_seq_lens,
|
|
const std::optional<torch::Tensor>& kv_cu_seq_lens,
|
|
const std::optional<torch::Tensor>& alibi_slope,
|
|
const std::optional<torch::Tensor>& attn_bias,
|
|
const std::optional<torch::Tensor>& q_quant_scale,
|
|
const std::optional<torch::Tensor>& k_quant_scale,
|
|
const std::optional<torch::Tensor>& v_quant_scale,
|
|
const torch::Tensor& block_tables,
|
|
int64_t max_query_len,
|
|
int64_t max_seq_len,
|
|
float scale,
|
|
bool is_causal,
|
|
int64_t window_size_left,
|
|
int64_t window_size_right,
|
|
const std::string& compute_dtype,
|
|
bool return_lse);
|
|
|
|
void batch_decode(torch::Tensor& query,
|
|
const torch::Tensor& k_cache,
|
|
torch::Tensor& output,
|
|
const torch::Tensor& block_table,
|
|
const torch::Tensor& seq_lens,
|
|
const std::optional<torch::Tensor>& v_cache,
|
|
std::optional<torch::Tensor>& output_lse,
|
|
const std::optional<torch::Tensor>& q_quant_scale,
|
|
const std::optional<torch::Tensor>& k_cache_quant_scale,
|
|
const std::optional<torch::Tensor>& v_cache_quant_scale,
|
|
const std::optional<torch::Tensor>& out_quant_scale,
|
|
const std::optional<torch::Tensor>& alibi_slope,
|
|
const std::optional<torch::Tensor>& mask,
|
|
const std::string& compute_dtype,
|
|
int64_t max_seq_len,
|
|
int64_t window_size_left,
|
|
int64_t window_size_right,
|
|
float scale,
|
|
bool return_lse,
|
|
bool is_causal,
|
|
int64_t kv_cache_quant_bit_size);
|
|
|
|
void residual_layer_norm(torch::Tensor& input,
|
|
torch::Tensor& output,
|
|
std::optional<torch::Tensor>& residual,
|
|
torch::Tensor& weight,
|
|
std::optional<torch::Tensor>& bias,
|
|
std::optional<torch::Tensor>& residual_out,
|
|
double eps);
|
|
|
|
void rms_norm(torch::Tensor& output,
|
|
torch::Tensor& input,
|
|
torch::Tensor& weight,
|
|
double eps);
|
|
|
|
torch::Tensor matmul(torch::Tensor a,
|
|
torch::Tensor b,
|
|
std::optional<torch::Tensor> bias);
|
|
|
|
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);
|
|
|
|
std::vector<torch::Tensor> moe_gen_idx(torch::Tensor& expert_id,
|
|
int64_t expert_num);
|
|
|
|
torch::Tensor moe_expand_input(const torch::Tensor& input,
|
|
const torch::Tensor& gather_index,
|
|
const torch::Tensor& combine_idx,
|
|
int64_t topk);
|
|
|
|
torch::Tensor group_gemm(torch::Tensor& input,
|
|
torch::Tensor& weight,
|
|
torch::Tensor& tokens_per_experts,
|
|
const std::optional<torch::Tensor>& dst_to_src,
|
|
torch::Tensor& output);
|
|
|
|
torch::Tensor moe_combine_result(torch::Tensor& input, torch::Tensor& weight);
|
|
} // namespace xllm::kernel::ilu
|