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
85 lines
2.8 KiB
C++
85 lines
2.8 KiB
C++
#pragma once
|
|
|
|
#include "engine/framework/assets/tensor_source.h"
|
|
#include "engine/framework/core/backend.h"
|
|
#include "engine/framework/core/module.h"
|
|
#include "engine/framework/modules/linear_module.h"
|
|
#include "engine/framework/modules/transformers/gemma_decoder.h"
|
|
#include "engine/models/dramabox/assets.h"
|
|
#include "engine/models/dramabox/gemma_tokenizer.h"
|
|
|
|
#include <ggml-backend.h>
|
|
|
|
#include <cstddef>
|
|
#include <memory>
|
|
#include <vector>
|
|
|
|
namespace engine::core {
|
|
class BackendWeightStore;
|
|
class ExecutionContext;
|
|
}
|
|
|
|
namespace engine::models::dramabox {
|
|
|
|
struct DramaBoxGemma3PromptWeights {
|
|
std::shared_ptr<core::BackendWeightStore> store;
|
|
modules::GemmaDecoderComponent encoder;
|
|
std::vector<modules::LinearWeights> aggregate_layers;
|
|
core::TensorValue aggregate_bias;
|
|
};
|
|
|
|
struct DramaBoxPromptEncoding {
|
|
int64_t batch = 0;
|
|
int64_t tokens = 0;
|
|
int64_t hidden_size = 0;
|
|
std::vector<float> audio_features;
|
|
std::vector<int32_t> attention_mask;
|
|
};
|
|
|
|
DramaBoxGemma3PromptWeights load_dramabox_gemma3_prompt_weights(
|
|
const DramaBoxAssets & assets,
|
|
ggml_backend_t backend,
|
|
core::BackendType backend_type,
|
|
size_t weight_context_bytes,
|
|
assets::TensorStorageType gemma_weight_storage_type,
|
|
assets::TensorStorageType projection_weight_storage_type);
|
|
|
|
core::TensorValue build_dramabox_gemma3_prompt_encoder(
|
|
core::ModuleBuildContext & ctx,
|
|
const core::TensorValue & input_ids,
|
|
const core::TensorValue & positions,
|
|
const core::TensorValue & additive_attention_mask,
|
|
const core::TensorValue & hidden_attention_mask,
|
|
const DramaBoxGemma3PromptWeights & weights,
|
|
const DramaBoxConfig & config);
|
|
|
|
class DramaBoxGemma3PromptRuntime {
|
|
public:
|
|
DramaBoxGemma3PromptRuntime(
|
|
core::ExecutionContext & execution,
|
|
std::shared_ptr<const DramaBoxAssets> assets,
|
|
assets::TensorStorageType gemma_weight_storage_type,
|
|
assets::TensorStorageType projection_weight_storage_type,
|
|
int64_t max_batch);
|
|
~DramaBoxGemma3PromptRuntime();
|
|
|
|
DramaBoxGemma3PromptRuntime(const DramaBoxGemma3PromptRuntime &) = delete;
|
|
DramaBoxGemma3PromptRuntime & operator=(const DramaBoxGemma3PromptRuntime &) = delete;
|
|
|
|
void prepare(int64_t batch) const;
|
|
DramaBoxPromptEncoding encode(const DramaBoxGemmaTokenBatch & tokens) const;
|
|
void release_runtime_state() const;
|
|
|
|
private:
|
|
class Graph;
|
|
|
|
core::ExecutionContext * execution_ = nullptr;
|
|
std::shared_ptr<const DramaBoxAssets> assets_;
|
|
assets::TensorStorageType gemma_weight_storage_type_ = assets::TensorStorageType::Native;
|
|
assets::TensorStorageType projection_weight_storage_type_ = assets::TensorStorageType::Native;
|
|
int64_t max_batch_ = 1;
|
|
mutable std::unique_ptr<DramaBoxGemma3PromptWeights> weights_;
|
|
mutable std::unique_ptr<Graph> graph_;
|
|
};
|
|
|
|
} // namespace engine::models::dramabox
|