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}
163 lines
6.9 KiB
C++
163 lines
6.9 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 "ilu_ops_api.h"
|
|
#include "ixinfer.h"
|
|
#include "utils.h"
|
|
|
|
using namespace ixformer;
|
|
|
|
namespace xllm::kernel::ilu {
|
|
|
|
void reshape_paged_cache(torch::Tensor& key,
|
|
std::optional<torch::Tensor>& value,
|
|
torch::Tensor& key_cache,
|
|
std::optional<torch::Tensor>& value_cache,
|
|
torch::Tensor& slot_mapping) {
|
|
auto value_ = value.value_or(torch::Tensor());
|
|
auto value_cache_ = value_cache.value_or(torch::Tensor());
|
|
|
|
int64_t key_token_stride = key.stride(0);
|
|
int64_t value_token_stride = 0;
|
|
if (value_.defined()) {
|
|
value_token_stride = value_.stride(0);
|
|
}
|
|
slot_mapping = slot_mapping.to(at::kLong);
|
|
infer::xllm_reshape_and_cache(key,
|
|
value_,
|
|
key_cache,
|
|
value_cache_,
|
|
slot_mapping,
|
|
key_token_stride,
|
|
value_token_stride);
|
|
}
|
|
|
|
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) {
|
|
double softcap = 0.0;
|
|
bool sqrt_alibi = false;
|
|
auto q_cu_seq_lens_ = q_cu_seq_lens.value_or(torch::Tensor());
|
|
auto kv_cu_seq_lens_ = kv_cu_seq_lens.value_or(torch::Tensor());
|
|
auto q_quant_scale_ = q_quant_scale.value_or(torch::Tensor());
|
|
auto k_quant_scale_ = k_quant_scale.value_or(torch::Tensor());
|
|
auto v_quant_scale_ = v_quant_scale.value_or(torch::Tensor());
|
|
auto block_tables_ = block_tables;
|
|
auto key_ = key;
|
|
auto value_ = value.value();
|
|
infer::ixinfer_flash_attn_unpad_with_block_tables(query,
|
|
key_,
|
|
value_,
|
|
output,
|
|
block_tables_,
|
|
q_cu_seq_lens_,
|
|
kv_cu_seq_lens_,
|
|
max_query_len,
|
|
max_seq_len,
|
|
is_causal,
|
|
window_size_left,
|
|
window_size_right,
|
|
static_cast<double>(scale),
|
|
softcap,
|
|
sqrt_alibi,
|
|
alibi_slope,
|
|
c10::nullopt,
|
|
output_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) {
|
|
if (query.dim() == 4) {
|
|
query =
|
|
query
|
|
.view({query.size(0) * query.size(1), query.size(2), query.size(3)})
|
|
.contiguous();
|
|
}
|
|
if (output.dim() == 4) {
|
|
output = output
|
|
.view({output.size(0) * output.size(1),
|
|
output.size(2),
|
|
output.size(3)})
|
|
.contiguous();
|
|
;
|
|
}
|
|
auto v_cache_ = v_cache.value_or(torch::Tensor());
|
|
int64_t num_kv_heads = k_cache.size(1);
|
|
int64_t page_block_size = k_cache.size(2);
|
|
double softcap = 0.0;
|
|
bool enable_cuda_graph = false;
|
|
bool use_sqrt_alibi = false;
|
|
auto block_table_ = block_table;
|
|
auto k_cache_ = k_cache;
|
|
auto seq_lens_ = seq_lens;
|
|
infer::xllm_paged_attention(output,
|
|
query,
|
|
k_cache_,
|
|
v_cache_,
|
|
num_kv_heads,
|
|
scale,
|
|
block_table_,
|
|
seq_lens_,
|
|
page_block_size,
|
|
max_seq_len,
|
|
alibi_slope,
|
|
is_causal,
|
|
(int32_t)window_size_left,
|
|
(int32_t)window_size_right,
|
|
softcap,
|
|
enable_cuda_graph,
|
|
use_sqrt_alibi,
|
|
c10::nullopt);
|
|
}
|
|
|
|
} // namespace xllm::kernel::ilu
|