#pragma once
// pipeline-asr.h: wav -> text orchestration. Owns the backend pair, the loaded
// weights, the BPE tokenizer, and drives mel -> conv stem -> audio encoder ->
// audio prefix -> Qwen3 decoder loop -> detokenize. The public ABI in
// qwenasr.h wraps this; the CLI tools call it directly for staged debug dumps.

#include "qa-error.h"
#include "qwenasr.h"

#include <string>
#include <vector>

struct pipeline_asr;

struct pipeline_asr_params {
    std::string model_path;
    int         n_threads  = 0;
    bool        use_gpu    = true;
    bool        clamp_fp16 = false;
    bool        flash_attn = true;
    std::string dump_dir;  // when set, write the stage tensors there as f32 dumps
};

pipeline_asr * pipeline_asr_load(const pipeline_asr_params & params);
void           pipeline_asr_free(pipeline_asr * p);

// Full transcription. params carries the language hint, the optional context,
// max_new_tokens and the streaming callback. text receives the decoded UTF-8
// string.
qa_status pipeline_asr_run(pipeline_asr *               p,
                           const float *                pcm,
                           size_t                       n_samples,
                           int                          sample_rate,
                           const qa_transcribe_params * params,
                           std::string &                text);