Support Qwen2ForCausalLM etc

[miyako]

Use Generator, not Decoder.
Use forward_batch(..., false), not generate_batch(...).
Extract the last token, not the first token.

[miyako]

#include <ctranslate2/generator.h>
#include <iostream>
#include <vector>
#include <cmath>

int main() {
    ctranslate2::Generator generator("path_to_exported_qwen2_ct2", ctranslate2::Device::CPU);

    std::vector<std::vector<std::string>> batch_tokens = {
        {" Hello", " world", "!", "<|im_end|>"}
    };

    // return_logits = false ensures we get the DENSE VECTORS, not vocabulary probabilities!
    ctranslate2::StorageView hidden_states = generator.forward_batch(batch_tokens, false);

    // hidden_states shape is [batch_size, seq_length, hidden_size]
    dim_t seq_length = hidden_states.dim(1);
    dim_t hidden_size = hidden_states.dim(2); // e.g., 3584 for Qwen2-7B

    const float* data = hidden_states.data<float>();

    std::vector<float> embeddings(hidden_size);
    float sum_sq = 0.0f;

    // Grab the last token's dense vector
    size_t last_token_idx = batch_tokens[0].size() - 1; 

    for (dim_t i = 0; i < hidden_size; ++i) {
        float val = data[last_token_idx * hidden_size + i];
        embeddings[i] = val;
        sum_sq += val * val;
    }

    // Normalize
    float norm = std::sqrt(std::max(sum_sq, 1e-12f));
    for (dim_t i = 0; i < hidden_size; ++i) {
        embeddings[i] /= norm;
    }

    std::cout << "Successfully generated Qwen2 embeddings." << std::endl;
    return 0;
}