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
86 lines
2.6 KiB
C++
86 lines
2.6 KiB
C++
#pragma once
|
|
|
|
#include "engine/framework/assets/tensor_source.h"
|
|
#include "engine/framework/core/backend.h"
|
|
#include "engine/community_models/outetts/assets.h"
|
|
|
|
#include <cstddef>
|
|
#include <cstdint>
|
|
#include <memory>
|
|
#include <string_view>
|
|
#include <vector>
|
|
|
|
namespace engine::models::outetts {
|
|
|
|
struct OuteTTSGenerateOptions {
|
|
int64_t max_new_tokens = 4096;
|
|
float temperature = 0.4F;
|
|
float repetition_penalty = 1.1F;
|
|
int64_t repetition_window = 64;
|
|
int64_t top_k = 40;
|
|
float top_p = 0.9F;
|
|
float min_p = 0.05F;
|
|
uint32_t seed = 0;
|
|
int64_t minimum_new_tokens = 0;
|
|
int32_t allowed_token_min = -1;
|
|
int32_t allowed_token_max = -1;
|
|
int32_t allowed_special_token = -1;
|
|
bool repetition_aware_sampling = false;
|
|
int64_t repetition_aware_window = 10;
|
|
int64_t repetition_aware_threshold = 1;
|
|
// Some upstream samplers sort and slice the nucleus into a compact
|
|
// probability tensor before torch.multinomial. Keep this opt-in because
|
|
// Hugging Face-style generation samples a full-vocabulary tensor with
|
|
// filtered logits instead.
|
|
bool compact_sorted_multinomial = false;
|
|
};
|
|
|
|
enum class OuteTTSStopReason {
|
|
Eos,
|
|
AudioEnd,
|
|
MaxTokens,
|
|
ContextLimit,
|
|
};
|
|
|
|
std::string_view outetts_stop_reason_name(OuteTTSStopReason reason) noexcept;
|
|
|
|
struct OuteTTSGenerateResult {
|
|
std::vector<int32_t> tokens;
|
|
OuteTTSStopReason stop_reason = OuteTTSStopReason::MaxTokens;
|
|
};
|
|
|
|
class OuteTTSLlamaRuntime final {
|
|
public:
|
|
OuteTTSLlamaRuntime(
|
|
std::shared_ptr<const OuteTTSAssets> assets,
|
|
core::BackendType backend_type,
|
|
int device,
|
|
int threads,
|
|
#if defined(INTPTR_MAX) && (INTPTR_MAX == INT32_MAX)
|
|
size_t weight_context_bytes = 1024ull * 1024ull * 1024ull,
|
|
#else
|
|
size_t weight_context_bytes = 4ull * 1024ull * 1024ull * 1024ull,
|
|
#endif
|
|
size_t constant_context_bytes = 256ull * 1024ull * 1024ull,
|
|
assets::TensorStorageType weight_storage_type = assets::TensorStorageType::Native);
|
|
~OuteTTSLlamaRuntime();
|
|
|
|
OuteTTSLlamaRuntime(const OuteTTSLlamaRuntime &) = delete;
|
|
OuteTTSLlamaRuntime & operator=(const OuteTTSLlamaRuntime &) = delete;
|
|
|
|
OuteTTSGenerateResult generate(
|
|
const std::vector<int32_t> & prompt,
|
|
const OuteTTSGenerateOptions & options,
|
|
int32_t eos_id,
|
|
int32_t audio_end_id) const;
|
|
|
|
// Releases the reusable cached-step graph while keeping model weights
|
|
// resident. Returns the released KV-cache capacity in tokens.
|
|
int64_t release_cached_step_graph();
|
|
|
|
private:
|
|
struct Impl;
|
|
std::unique_ptr<Impl> impl_;
|
|
};
|
|
|
|
} // namespace engine::models::outetts
|