8d75652949
Sources cloned and tree'd (no --depth):
- jd-opensource/xllm: ILU kernels, CUDA kernels, MoE kernels
- NVIDIA/cccl: CUB tuning/dispatch headers (block-level primitives)
- fla-org/flash-linear-attention: Triton GDN kernels
- NVIDIA/cutlass: grouped GEMM reference (read, not copied)
- Dao-AILab/flash-attention: attention kernel reference (SM80+, read only)
New CUDA kernels (from xllm, SM-agnostic, portable to BI-V100):
ex_engine/xllm_kernels/cuda/activation.cu (188 lines) — silu_and_mul, gelu
ex_engine/xllm_kernels/cuda/norm.cu (600 lines) — rms_norm, fused_add_rms_norm
ex_engine/xllm_kernels/cuda/rope.cu (258 lines) — rotary_embedding
ex_engine/xllm_kernels/cuda/block_copy.cu (209 lines) — copy_blocks, swap_blocks
ex_engine/xllm_kernels/cuda/reshape_paged_cache.cu (101 lines) — KV cache ops
ex_engine/xllm_kernels/cuda/headers/ (5 headers for compilation)
ILU bridge kernel sources (from xllm, verified SAME as upstream):
ex_engine/xllm_kernels/ilu/ (10 files, 925 lines total)
— activation.cpp, attention.cpp, fused_moe.cpp, group_gemm.cpp,
matmul.cpp, norm.cpp, rope.cpp, ilu_ops_api.h, ixformer.h, utils.h
FLA Triton GDN kernels (for GatedDeltaNet without SM90+ FlashQLA):
ex_engine/fla_kernels/gated_delta_rule/ (7 files, 2370 lines)
— chunk_fwd.py (428), chunk.py (487), wy_fast.py (409),
fused_recurrent.py (392), naive.py (161), gate.py (380)
CCCL sync (12 tuning + 14 dispatch headers updated from NVIDIA/cccl):
cccl_upstream/cub/cub/device/dispatch/tuning/ — 12 changed files synced
cccl_upstream/cub/cub/device/dispatch/ — 14 changed dispatch files synced
Compilation targets for real machine (ivcore10):
1. CUDA kernels: --cuda-gpu-arch=ivcore10 via corex clang/16
2. ILU bridges: torch.utils.cpp_extension linking ixformer .so
3. FLA kernels: Triton JIT (if Triton works on BI-V100)
154 lines
6.1 KiB
C++
154 lines
6.1 KiB
C++
/* Copyright 2025-2026 The xLLM Authors.
|
|
|
|
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
|