6415249693
MoE call chain from ds_vllm (vllm-project/vllm latest): ex_engine/moe/ — 20 files, 8736 lines - modular_kernel.py (1630 lines) — base classes for modular MoE - experts/fused_batched_moe.py (972 lines) — NaiveBatchedExperts - prepare_finalize/batched.py (171 lines) — token grouping by expert - topk_weight_and_reduce.py (176 lines) — scatter-add finalize - fused_moe.py (1740 lines) — main fused_moe dispatch - config.py (1407 lines) — FusedMoEQuantConfig - activation.py, utils.py, layer.py, etc. xllm layer code (jd-opensource/xllm): ex_engine/xllm_layers/ — 39 files, 5859 lines - ilu/fused_moe.cpp (797 lines) — production ixformer 7-step MoE pipeline - ilu/attention.cpp (189 lines) — paged_attention + flash_attn bridge - npu_torch/qwen3_gated_delta_net_base.cpp (576 lines) — GDN reference - common/rms_norm.cpp, rotary_embedding.cpp, activation.cpp, dense_mlp.cpp xllm ILU kernels — synced 10 files to upstream (diffs from prior edits) These are reference implementations, NOT hand-written. Source repos: vllm-project/vllm, jd-opensource/xllm
74 lines
2.3 KiB
C++
74 lines
2.3 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 "util/env_var.h"
|
|
|
|
namespace xllm::kernel::ilu {
|
|
|
|
bool gemv_conditions(const torch::Tensor& input,
|
|
const torch::Tensor& weight,
|
|
const torch::Tensor& bias,
|
|
int64_t gemv_max_batch) {
|
|
// gemv input:[m,k] weight:[n,k]
|
|
// 1. m <= gemv_max_batch
|
|
// 2. k % 32 == 0 && n % 2 == 0
|
|
// 3. bias is None
|
|
|
|
torch::Tensor input_view = input.view({-1, input.size(-1)});
|
|
torch::Tensor weight_view = weight.view({-1, weight.size(-1)});
|
|
|
|
int64_t m = input_view.size(0);
|
|
int64_t k = input_view.size(1);
|
|
int64_t n = weight_view.size(0);
|
|
|
|
if (bias.defined() == false && m <= gemv_max_batch && k % 32 == 0 &&
|
|
n % 2 == 0) {
|
|
return true;
|
|
}
|
|
return false;
|
|
}
|
|
|
|
torch::Tensor matmul(torch::Tensor a,
|
|
torch::Tensor b,
|
|
std::optional<torch::Tensor> bias) {
|
|
int64_t act_type = -1;
|
|
bool persistent = false;
|
|
std::vector<int64_t> output_shape = a.sizes().vec();
|
|
if (!output_shape.empty()) {
|
|
output_shape[output_shape.size() - 1] = b.size(0);
|
|
}
|
|
torch::Tensor output = a.new_empty(output_shape);
|
|
|
|
bool use_gemv = true;
|
|
const int64_t gemv_max_batch = 1;
|
|
const bool disable_infer_gemm_ex =
|
|
xllm::util::get_bool_env("DISABLE_INFER_GEMM_EX", false);
|
|
|
|
use_gemv =
|
|
use_gemv &&
|
|
gemv_conditions(a, b, bias.value_or(at::Tensor()), gemv_max_batch) &&
|
|
!disable_infer_gemm_ex && (act_type == -1);
|
|
|
|
if (use_gemv) {
|
|
output = infer::ixformer_linear_ex(a, b, bias, output);
|
|
} else {
|
|
output = infer::ixformer_linear(a, b, act_type, bias, output, persistent);
|
|
}
|
|
return output;
|
|
}
|
|
|
|
} // namespace xllm::kernel::ilu
|