Method run_step

src/models/vevo2/ar.cpp:808–842

        }
    }

    ggml_backend_t backend = nullptr;
    core::BackendType backend_type = core::BackendType::Cpu;
    int threads = 1;
    std::shared_ptr<const Vevo2ARWeights> weights;
    int64_t cache_steps = 0;
    std::unique_ptr<ggml_context, GgmlContextDeleter> ctx;
    ggml_tensor * token_id = nullptr;
    ggml_tensor * positions = nullptr;
    ggml_tensor * attention_mask = nullptr;
    ggml_tensor * logits = nullptr;
    std::vector<ggml_tensor *> key_sources;
    std::vector<ggml_tensor *> value_sources;
    std::vector<std::vector<ggml_tensor *>> key_destinations;
    std::vector<std::vector<ggml_tensor *>> value_destinations;
    std::vector<ggml_fp16_t> attention_mask_values;
    runtime::TransformerKVCache step_cache;
    ggml_cgraph * graph = nullptr;
    ggml_backend_buffer_t buffer = nullptr;
};

Vevo2AutoregressiveRuntime::Vevo2AutoregressiveRuntime(
    std::shared_ptr<const Vevo2Assets> assets,
    core::ExecutionContext & execution_context,
    size_t weight_context_bytes,
    size_t prefill_graph_context_bytes,
    size_t decode_graph_context_bytes,
    assets::TensorStorageType weight_storage_type)
    : assets_(require_assets(std::move(assets))),
      execution_context_(execution_context),
      prefill_graph_context_bytes_(prefill_graph_context_bytes),
      decode_graph_context_bytes_(decode_graph_context_bytes),
      tokenizer_(assets_),
      weight_source_(assets_->ar_weights),
      weights_(load_ar_weights(
          *assets_,
          execution_context_.backend(),
          execution_context_.backend_type(),