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
89 lines
3.1 KiB
C++
89 lines
3.1 KiB
C++
#pragma once
|
|
|
|
#include "engine/framework/assets/tensor_source.h"
|
|
#include "engine/framework/core/backend_weight_store.h"
|
|
#include "engine/framework/modules/attention/types.h"
|
|
#include "engine/framework/modules/conv_modules.h"
|
|
#include "engine/framework/modules/linear_module.h"
|
|
#include "engine/framework/modules/norm_modules.h"
|
|
#include "engine/framework/modules/streaming_conv_modules.h"
|
|
#include "engine/models/hviske_asr/assets.h"
|
|
|
|
#include <memory>
|
|
#include <vector>
|
|
|
|
namespace engine::models::hviske_asr {
|
|
|
|
struct HviskeBatchNormEvalWeights {
|
|
engine::core::TensorValue scale;
|
|
engine::core::TensorValue bias;
|
|
};
|
|
|
|
struct HviskeSubsamplingWeights {
|
|
engine::modules::Conv2dWeights conv0;
|
|
engine::core::TensorValue depthwise1_weight;
|
|
engine::core::TensorValue depthwise1_bias;
|
|
engine::modules::Conv2dWeights pointwise1;
|
|
engine::core::TensorValue depthwise2_weight;
|
|
engine::core::TensorValue depthwise2_bias;
|
|
engine::modules::Conv2dWeights pointwise2;
|
|
engine::modules::LinearWeights linear;
|
|
};
|
|
|
|
struct HviskeConformerLayerWeights {
|
|
engine::modules::NormWeights norm_feed_forward1;
|
|
engine::modules::NormWeights norm_self_att;
|
|
engine::modules::NormWeights norm_conv;
|
|
engine::modules::NormWeights norm_feed_forward2;
|
|
engine::modules::NormWeights norm_out;
|
|
engine::modules::LinearWeights ff1_linear1;
|
|
engine::modules::LinearWeights ff1_linear2;
|
|
engine::modules::LinearWeights ff2_linear1;
|
|
engine::modules::LinearWeights ff2_linear2;
|
|
engine::modules::RelativeAttentionWeights self_attn;
|
|
engine::modules::LinearWeights conv_pointwise1;
|
|
engine::modules::DepthwiseConv1dWeights conv_depthwise;
|
|
HviskeBatchNormEvalWeights conv_norm;
|
|
engine::modules::LinearWeights conv_pointwise2;
|
|
};
|
|
|
|
struct HviskeConformerWeights {
|
|
HviskeSubsamplingWeights subsampling;
|
|
std::vector<HviskeConformerLayerWeights> layers;
|
|
engine::modules::LinearWeights encoder_projector;
|
|
};
|
|
|
|
struct HviskeDecoderLayerWeights {
|
|
engine::modules::NormWeights self_norm;
|
|
engine::modules::AttentionWeights self_attn;
|
|
engine::modules::NormWeights cross_norm;
|
|
engine::modules::AttentionWeights cross_attn;
|
|
engine::modules::NormWeights ff_norm;
|
|
engine::modules::LinearWeights ff_in;
|
|
engine::modules::LinearWeights ff_out;
|
|
};
|
|
|
|
struct HviskeDecoderWeights {
|
|
engine::core::TensorValue token_embedding;
|
|
engine::core::TensorValue position_embedding;
|
|
engine::modules::NormWeights embedding_norm;
|
|
std::vector<HviskeDecoderLayerWeights> layers;
|
|
engine::modules::NormWeights final_norm;
|
|
engine::modules::LinearWeights classifier;
|
|
};
|
|
|
|
struct HviskeWeights {
|
|
std::shared_ptr<engine::core::BackendWeightStore> store;
|
|
HviskeConformerWeights encoder;
|
|
HviskeDecoderWeights decoder;
|
|
};
|
|
|
|
std::shared_ptr<const HviskeWeights> load_hviske_weights(
|
|
const HviskeASRAssets & assets,
|
|
ggml_backend_t backend,
|
|
engine::core::BackendType backend_type,
|
|
engine::assets::TensorStorageType matmul_storage_type,
|
|
engine::assets::TensorStorageType conv_storage_type,
|
|
size_t weight_context_bytes);
|
|
|
|
} // namespace engine::models::hviske_asr
|