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
114 lines
4.4 KiB
C++
114 lines
4.4 KiB
C++
#pragma once
|
|
|
|
#include "engine/framework/runtime/cache_slots.h"
|
|
#include "engine/framework/runtime/session_base.h"
|
|
#include "engine/models/vevo2/ar.h"
|
|
#include "engine/models/vevo2/assets.h"
|
|
#include "engine/models/vevo2/components.h"
|
|
#include "engine/models/vevo2/fm.h"
|
|
#include "engine/models/vevo2/prompt_builder.h"
|
|
#include "engine/models/vevo2/vocoder.h"
|
|
#include "engine/framework/modules/speech_encoders/whisper_frontend.h"
|
|
|
|
#include <cstddef>
|
|
#include <cstdint>
|
|
#include <memory>
|
|
#include <string>
|
|
#include <vector>
|
|
|
|
namespace engine::assets {
|
|
enum class TensorStorageType;
|
|
}
|
|
|
|
namespace engine::core {
|
|
class ExecutionContext;
|
|
}
|
|
|
|
namespace engine::models::vevo2 {
|
|
|
|
class Vevo2Session final
|
|
: public runtime::RuntimeSessionBase
|
|
, public runtime::IOfflineVoiceTaskSession {
|
|
public:
|
|
Vevo2Session(
|
|
runtime::TaskSpec task,
|
|
runtime::SessionOptions options,
|
|
std::shared_ptr<const Vevo2Assets> assets);
|
|
|
|
std::string family() const override;
|
|
runtime::VoiceTaskKind task_kind() const override;
|
|
runtime::RunMode run_mode() const override;
|
|
void prepare(const runtime::SessionPreparationRequest & request) override;
|
|
runtime::TaskResult run(const runtime::TaskRequest & request) override;
|
|
|
|
private:
|
|
struct AudioCacheKey {
|
|
uint64_t hash = 0;
|
|
int sample_rate = 0;
|
|
int channels = 0;
|
|
size_t samples = 0;
|
|
int64_t frames = 0;
|
|
};
|
|
|
|
struct AudioCacheKeyEqual {
|
|
bool operator()(const AudioCacheKey & lhs, const AudioCacheKey & rhs) const noexcept {
|
|
return lhs.hash == rhs.hash &&
|
|
lhs.sample_rate == rhs.sample_rate &&
|
|
lhs.channels == rhs.channels &&
|
|
lhs.samples == rhs.samples &&
|
|
lhs.frames == rhs.frames;
|
|
}
|
|
};
|
|
|
|
struct AudioFeatureCacheValue {
|
|
std::vector<float> features;
|
|
};
|
|
|
|
struct AudioTokenCacheValue {
|
|
Vevo2TokenSequence tokens;
|
|
};
|
|
|
|
Vevo2Request make_request(const runtime::TaskRequest & request) const;
|
|
std::vector<float> cached_whisper_features(
|
|
const runtime::AudioBuffer & audio,
|
|
int64_t target_frames,
|
|
size_t threads);
|
|
Vevo2TokenSequence cached_content_style_tokens(
|
|
const runtime::AudioBuffer & audio,
|
|
const std::vector<float> & whisper_features,
|
|
int64_t feature_frames);
|
|
|
|
runtime::TaskSpec task_;
|
|
std::shared_ptr<const Vevo2Assets> assets_;
|
|
size_t ar_weight_context_bytes_ = 2ull * 1024ull * 1024ull * 1024ull;
|
|
size_t ar_prefill_graph_context_bytes_ = 4ull * 1024ull * 1024ull;
|
|
size_t ar_decode_graph_context_bytes_ = 4ull * 1024ull * 1024ull;
|
|
engine::assets::TensorStorageType ar_weight_storage_type_;
|
|
engine::core::ExecutionContext reference_execution_context_;
|
|
size_t whisper_weight_context_bytes_ = 256ull * 1024ull * 1024ull;
|
|
size_t whisper_graph_context_bytes_ = 4ull * 1024ull * 1024ull;
|
|
engine::assets::TensorStorageType whisper_matmul_weight_storage_type_;
|
|
engine::assets::TensorStorageType whisper_conv_weight_storage_type_;
|
|
size_t tokenizer_weight_context_bytes_ = 768ull * 1024ull * 1024ull;
|
|
size_t tokenizer_graph_context_bytes_ = 4ull * 1024ull * 1024ull;
|
|
engine::assets::TensorStorageType tokenizer_matmul_weight_storage_type_;
|
|
engine::assets::TensorStorageType tokenizer_conv_weight_storage_type_;
|
|
size_t fm_weight_context_bytes_ = 2ull * 1024ull * 1024ull * 1024ull;
|
|
size_t fm_graph_context_bytes_ = 32ull * 1024ull * 1024ull;
|
|
engine::assets::TensorStorageType fm_matmul_weight_storage_type_;
|
|
engine::assets::TensorStorageType fm_conv_weight_storage_type_;
|
|
size_t vocoder_weight_context_bytes_ = 1024ull * 1024ull * 1024ull;
|
|
size_t vocoder_graph_context_bytes_ = 4ull * 1024ull * 1024ull;
|
|
engine::assets::TensorStorageType vocoder_matmul_weight_storage_type_;
|
|
engine::assets::TensorStorageType vocoder_conv_weight_storage_type_;
|
|
engine::modules::WhisperFrontendComponent whisper_frontend_;
|
|
Vevo2CocoProsodyTokenizerRuntime prosody_tokenizer_;
|
|
Vevo2CocoContentStyleTokenizerRuntime content_style_tokenizer_;
|
|
Vevo2Qwen2ARRuntime autoregressive_model_;
|
|
Vevo2DiffLlamaFlowMatchingRuntime flow_matching_model_;
|
|
Vevo2VocosRuntime vocoder_;
|
|
runtime::CacheSlots<AudioCacheKey, AudioFeatureCacheValue, AudioCacheKeyEqual> whisper_feature_cache_;
|
|
runtime::CacheSlots<AudioCacheKey, AudioTokenCacheValue, AudioCacheKeyEqual> content_style_token_cache_;
|
|
};
|
|
|
|
} // namespace engine::models::vevo2
|