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
78 lines
2.5 KiB
C++
78 lines
2.5 KiB
C++
#pragma once
|
|
|
|
#include "engine/framework/core/backend_weight_store.h"
|
|
#include "engine/framework/core/execution_context.h"
|
|
#include "engine/framework/modules/linear_module.h"
|
|
#include "engine/framework/modules/norm_modules.h"
|
|
|
|
#include <vector>
|
|
#include <memory>
|
|
#include <cstddef>
|
|
|
|
namespace engine::models::kugelaudio {
|
|
|
|
struct ConnectorWeights {
|
|
modules::LinearWeights fc1;
|
|
modules::NormWeights norm;
|
|
modules::LinearWeights fc2;
|
|
};
|
|
|
|
struct DiffusionLayerWeights {
|
|
modules::NormWeights norm;
|
|
modules::LinearWeights modulation;
|
|
modules::LinearWeights gate;
|
|
modules::LinearWeights up;
|
|
modules::LinearWeights down;
|
|
};
|
|
|
|
struct DiffusionWeights {
|
|
modules::LinearWeights input;
|
|
modules::LinearWeights condition;
|
|
modules::LinearWeights time1;
|
|
modules::LinearWeights time2;
|
|
std::vector<DiffusionLayerWeights> layers;
|
|
modules::LinearWeights final_modulation;
|
|
modules::LinearWeights output;
|
|
core::TensorValue ones;
|
|
};
|
|
|
|
struct DiffusionConfig {
|
|
int64_t hidden_size = 3584;
|
|
int64_t latent_size = 64;
|
|
int64_t intermediate_size = 10752;
|
|
int64_t layers = 4;
|
|
float norm_eps = 1e-5F;
|
|
};
|
|
|
|
ConnectorWeights load_connector_weights(
|
|
core::BackendWeightStore & store, const assets::TensorSource & source,
|
|
assets::TensorStorageType storage, int64_t latent_size, int64_t hidden_size);
|
|
|
|
DiffusionWeights load_diffusion_weights(
|
|
core::BackendWeightStore & store, const assets::TensorSource & source,
|
|
assets::TensorStorageType storage, const DiffusionConfig & config);
|
|
|
|
core::TensorValue build_connector(
|
|
core::ModuleBuildContext & ctx, const core::TensorValue & input,
|
|
const ConnectorWeights & weights, int64_t hidden_size);
|
|
|
|
// time_features is the schedule's precomputed 256-channel cosine/sine embedding.
|
|
core::TensorValue build_diffusion_head(
|
|
core::ModuleBuildContext & ctx, const core::TensorValue & latent,
|
|
const core::TensorValue & condition, const core::TensorValue & time_features,
|
|
const DiffusionWeights & weights, const DiffusionConfig & config);
|
|
|
|
class DiffusionHead {
|
|
public:
|
|
DiffusionHead(const assets::TensorSource & source, core::ExecutionContext & execution,
|
|
assets::TensorStorageType storage, DiffusionConfig config = {});
|
|
~DiffusionHead();
|
|
void begin_frame(const std::vector<float> & condition, const std::vector<int64_t> & timesteps);
|
|
std::vector<float> predict(const std::vector<float> & latent, size_t step);
|
|
|
|
private:
|
|
struct Impl;
|
|
std::unique_ptr<Impl> impl_;
|
|
};
|
|
|
|
} // namespace engine::models::kugelaudio
|