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
65 lines
2.1 KiB
C++
65 lines
2.1 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 <torch/torch.h>
|
|
|
|
#include "core/framework/model_context.h"
|
|
#include "framework/state_dict/state_dict.h"
|
|
#include "framework/state_dict/utils.h"
|
|
|
|
namespace xllm {
|
|
namespace layer {
|
|
|
|
class RMSNormImpl : public torch::nn::Module {
|
|
public:
|
|
RMSNormImpl(int64_t dim, double eps, const torch::TensorOptions& options);
|
|
RMSNormImpl(const ModelContext& context);
|
|
|
|
// Standard forward: returns (normalized_output, updated_residual)
|
|
std::tuple<torch::Tensor, std::optional<torch::Tensor>> forward(
|
|
torch::Tensor& input,
|
|
std::optional<torch::Tensor> residual = std::nullopt,
|
|
std::optional<torch::Tensor> inplace_output = std::nullopt);
|
|
|
|
// Fused forward with FP8 quantization output (for static quantization)
|
|
// Returns: (fp8_quantized_output, updated_residual)
|
|
// This combines RMSNorm + FP8 quantization to reduce memory bandwidth
|
|
std::tuple<torch::Tensor, std::optional<torch::Tensor>> forward_fp8(
|
|
torch::Tensor& input,
|
|
const torch::Tensor& fp8_scale,
|
|
std::optional<torch::Tensor> residual = std::nullopt);
|
|
|
|
void set_layernorm_mode();
|
|
|
|
void load_state_dict(const StateDict& state_dict);
|
|
|
|
torch::Tensor weight() const { return weight_; }
|
|
torch::Tensor bias() const { return bias_; }
|
|
double eps() const { return eps_; }
|
|
|
|
private:
|
|
DEFINE_WEIGHT(weight);
|
|
DEFINE_WEIGHT(bias);
|
|
int64_t norm_dim_;
|
|
double eps_;
|
|
std::string mode_;
|
|
};
|
|
TORCH_MODULE(RMSNorm);
|
|
|
|
} // namespace layer
|
|
} // namespace xllm
|