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
29 lines
1,002 B
C++
29 lines
1,002 B
C++
#pragma once
|
|
|
|
// MOSS-TTSD: the dialogue member of the moss_tts_delay family (8B, 16 codebooks).
|
|
//
|
|
// The geometry, token ids, backbone, heads, delay decoder and prompt assembly
|
|
// are all the family's, shared with MOSS-TTS-v1.5 and MOSS-VoiceGenerator. What
|
|
// belongs to this checkpoint is the rendering of its <user_inst> block and a
|
|
// session that clones by continuation rather than by reference alone.
|
|
|
|
#include "engine/framework/assets/resource_bundle.h"
|
|
#include "engine/framework/decoders/moss_tts_delay/config.h"
|
|
|
|
#include <filesystem>
|
|
#include <memory>
|
|
|
|
namespace engine::models::moss_ttsd {
|
|
|
|
using Config = decoders::MossTtsDelayConfig;
|
|
|
|
struct Assets {
|
|
assets::ResourceBundle resources;
|
|
Config config;
|
|
std::shared_ptr<const assets::TensorSource> model_weights;
|
|
std::shared_ptr<const assets::TensorSource> audio_tokenizer_weights;
|
|
};
|
|
|
|
std::shared_ptr<const Assets> load_moss_ttsd_assets(const std::filesystem::path & model_path);
|
|
|
|
} // namespace engine::models::moss_ttsd
|