Fabio
c36510c21e
Some checks failed
Linux Build / Linux cpu (push) Has been cancelled
Linux Build / Linux cpu-clang (push) Has been cancelled
Linux Build / Linux vulkan (push) Has been cancelled
mac-build / macOS CPU (push) Has been cancelled
Nix Build / Nix (cpu) (push) Has been cancelled
Nix Build / Nix (vulkan) (push) Has been cancelled
windows-build / Windows CPU (push) Has been cancelled
110 lines
3.2 KiB
C++
110 lines
3.2 KiB
C++
#pragma once
|
|
|
|
#include "engine/framework/assets/resource_bundle.h"
|
|
#include "engine/framework/assets/tensor_source.h"
|
|
|
|
#include <cstdint>
|
|
#include <filesystem>
|
|
#include <memory>
|
|
#include <optional>
|
|
#include <string>
|
|
#include <vector>
|
|
|
|
namespace engine::models::vibevoice {
|
|
|
|
struct VibeVoiceTokenizerConfig {
|
|
int64_t channels = 1;
|
|
bool causal = true;
|
|
int64_t vae_dim = 0;
|
|
float fix_std = 0.0F;
|
|
std::string std_dist_type;
|
|
std::string mixer_layer;
|
|
std::string conv_norm;
|
|
std::string pad_mode;
|
|
bool disable_last_norm = true;
|
|
std::string layernorm;
|
|
float layernorm_eps = 1.0e-5F;
|
|
bool layernorm_elementwise_affine = true;
|
|
bool conv_bias = true;
|
|
float layer_scale_init_value = 1.0e-6F;
|
|
float weight_init_value = 1.0e-2F;
|
|
int64_t encoder_n_filters = 0;
|
|
std::vector<int64_t> encoder_ratios;
|
|
std::string encoder_depths;
|
|
int64_t decoder_n_filters = 0;
|
|
std::vector<int64_t> decoder_ratios;
|
|
std::string decoder_depths;
|
|
};
|
|
|
|
struct VibeVoiceQwen2Config {
|
|
int64_t hidden_size = 0;
|
|
int64_t intermediate_size = 0;
|
|
int64_t max_position_embeddings = 0;
|
|
int64_t max_window_layers = 0;
|
|
int64_t num_attention_heads = 0;
|
|
int64_t num_hidden_layers = 0;
|
|
int64_t num_key_value_heads = 0;
|
|
int64_t vocab_size = 0;
|
|
int64_t head_dim = 0;
|
|
float rms_norm_eps = 1.0e-6F;
|
|
float rope_theta = 1000000.0F;
|
|
bool tie_word_embeddings = true;
|
|
bool use_cache = true;
|
|
bool use_sliding_window = false;
|
|
};
|
|
|
|
struct VibeVoiceDiffusionHeadConfig {
|
|
int64_t ddpm_batch_mul = 4;
|
|
std::string ddpm_beta_schedule;
|
|
int64_t ddpm_num_inference_steps = 20;
|
|
int64_t ddpm_num_steps = 1000;
|
|
std::string diffusion_type;
|
|
float head_ffn_ratio = 3.0F;
|
|
int64_t head_layers = 4;
|
|
int64_t hidden_size = 0;
|
|
int64_t latent_size = 0;
|
|
std::string prediction_type;
|
|
float rms_norm_eps = 1.0e-5F;
|
|
int64_t speech_vae_dim = 0;
|
|
};
|
|
|
|
struct VibeVoiceConfig {
|
|
int64_t acoustic_vae_dim = 0;
|
|
int64_t semantic_vae_dim = 0;
|
|
std::string model_type;
|
|
std::string torch_dtype;
|
|
VibeVoiceTokenizerConfig acoustic_tokenizer;
|
|
VibeVoiceTokenizerConfig semantic_tokenizer;
|
|
VibeVoiceQwen2Config decoder;
|
|
VibeVoiceDiffusionHeadConfig diffusion_head;
|
|
};
|
|
|
|
struct VibeVoiceAudioProcessorConfig {
|
|
int sample_rate = 24000;
|
|
bool normalize_audio = true;
|
|
float target_db_fs = -25.0F;
|
|
float eps = 1.0e-6F;
|
|
};
|
|
|
|
struct VibeVoiceProcessorConfig {
|
|
int64_t speech_tok_compress_ratio = 3200;
|
|
bool db_normalize = true;
|
|
std::string language_model_pretrained_name = "Qwen/Qwen2.5-1.5B";
|
|
VibeVoiceAudioProcessorConfig audio_processor;
|
|
};
|
|
|
|
struct VibeVoiceAssets {
|
|
assets::ResourceBundle resources;
|
|
VibeVoiceConfig config;
|
|
VibeVoiceProcessorConfig processor;
|
|
std::shared_ptr<const assets::TensorSource> model_weights;
|
|
float speech_scaling_factor = 1.0F;
|
|
float speech_bias_factor = 0.0F;
|
|
// Set once a fine-tune adapter has been overlaid onto model_weights, so the same adapter is
|
|
// not applied twice (e.g. via both --load-option and --session-option).
|
|
bool fine_tune_applied = false;
|
|
};
|
|
|
|
std::shared_ptr<const VibeVoiceAssets> load_vibevoice_assets(const std::filesystem::path & model_path);
|
|
|
|
} // namespace engine::models::vibevoice
|