neuroflow-cpp / tests /mla_step_test.cpp
cwenzi's picture
Upload folder using huggingface_hub
26d5b81 verified
Raw
History Blame Contribute Delete
4.98 kB
#include <iostream>
#include <exception>
#include "../include/neuroflow/tensor.hpp"
#include "../include/neuroflow/networks.hpp"
using namespace neuroflow;
int main() {
try {
std::cout << "MLA step by step..." << std::endl;
// MLA 参数
size_t d_model = 64;
size_t n_heads = 4;
size_t d_latent = 16;
size_t head_dim = d_model / n_heads; // 16
std::cout << "d_model=" << d_model << ", n_heads=" << n_heads
<< ", d_latent=" << d_latent << ", head_dim=" << head_dim << std::endl;
// 创建 Linear 层
Linear W_q(d_model, d_model, false);
Linear W_dkv(d_model, d_latent, false);
Linear W_uk(d_latent, d_model, false);
Linear W_uv(d_latent, d_model, false);
Linear W_o(d_model, d_model, false);
// 输入
Tensor input({1, d_model});
for (size_t i = 0; i < input.numel(); ++i) input.as_fp32()[i] = 0.1f * i;
size_t batch = 1;
size_t seq_len = 1;
std::cout << "1. Q projection..." << std::endl;
Tensor x_flat = input.reshape({batch * seq_len, d_model});
std::cout << " x_flat shape: [" << x_flat.shape_[0] << ", " << x_flat.shape_[1] << "]" << std::endl;
Tensor q = W_q.forward(x_flat);
std::cout << " q shape after linear: [" << q.shape_[0] << ", " << q.shape_[1] << "]" << std::endl;
// q 需要 reshape 到 {batch, seq_len, n_heads, head_dim}
std::cout << " reshaping q to {batch, seq_len, n_heads, head_dim} = {" << batch << ", " << seq_len << ", " << n_heads << ", " << head_dim << "}" << std::endl;
std::cout << " q numel: " << q.numel() << std::endl;
std::cout << " expected numel: " << (batch * seq_len * n_heads * head_dim) << std::endl;
q = q.reshape({batch, seq_len, n_heads, head_dim});
std::cout << " q reshaped: success" << std::endl;
std::cout << "2. KV compression..." << std::endl;
Tensor c_kv = W_dkv.forward(x_flat);
std::cout << " c_kv shape: [" << c_kv.shape_[0] << ", " << c_kv.shape_[1] << "]" << std::endl;
c_kv = c_kv.reshape({batch, seq_len, d_latent});
std::cout << " c_kv reshaped: success" << std::endl;
std::cout << "3. K/V decompression..." << std::endl;
size_t total_len = seq_len; // 第一次没有 cache
std::cout << " creating c_kv_flat {" << (batch * total_len) << ", " << d_latent << "}" << std::endl;
Tensor c_kv_flat({batch * total_len, d_latent});
float* ckf = c_kv_flat.as_fp32();
float* ck = c_kv.as_fp32();
std::cout << " copying " << (batch * total_len * d_latent) << " elements" << std::endl;
for (size_t i = 0; i < batch * total_len * d_latent; ++i) {
ckf[i] = ck[i];
}
std::cout << " copy done" << std::endl;
Tensor k = W_uk.forward(c_kv_flat);
Tensor v = W_uv.forward(c_kv_flat);
std::cout << " k shape: [" << k.shape_[0] << ", " << k.shape_[1] << "]" << std::endl;
std::cout << " v shape: [" << v.shape_[0] << ", " << v.shape_[1] << "]" << std::endl;
k = k.reshape({batch, total_len, n_heads, head_dim});
v = v.reshape({batch, total_len, n_heads, head_dim});
std::cout << " k reshaped, v reshaped" << std::endl;
std::cout << "4. Attention computation..." << std::endl;
Tensor output({batch, seq_len, d_model});
float* out = output.as_fp32();
float* qp = q.as_fp32();
float* kp = k.as_fp32();
float* vp = v.as_fp32();
float scale = 1.0f / std::sqrt(static_cast<float>(head_dim));
for (size_t b = 0; b < batch; ++b) {
for (size_t h = 0; h < n_heads; ++h) {
for (size_t s = 0; s < seq_len; ++s) {
std::vector<float> scores(total_len);
for (size_t t = 0; t < total_len; ++t) {
float dot = 0;
for (size_t d = 0; d < head_dim; ++d) {
size_t q_idx = b * seq_len * n_heads * head_dim + s * n_heads * head_dim + h * head_dim + d;
size_t k_idx = b * total_len * n_heads * head_dim + t * n_heads * head_dim + h * head_dim + d;
std::cout << " q_idx=" << q_idx << ", k_idx=" << k_idx << std::endl;
dot += qp[q_idx] * kp[k_idx];
}
scores[t] = dot * scale;
}
// softmax...
// output...
}
}
}
std::cout << "Success!" << std::endl;
return 0;
} catch (const std::exception& e) {
std::cout << "Error: " << e.what() << std::endl;
return 1;
}
}