def __init__()

in tensorrtllm/run_eval.py [0:0]

• 35 lines of code
• 3 McCabe index (conditional complexity)

    def __init__(self,
                 engine_dir,
                 assets_dir="assets",
                 batch_size=64):
        encoder_config = read_config('encoder', engine_dir)
        decoder_config = read_config('decoder', engine_dir)
        self.n_mels = encoder_config['n_mels']
        self.num_languages = encoder_config['num_languages']
        is_multilingual = (decoder_config['vocab_size'] >= 51865)
        if is_multilingual:
            tokenizer_name = "multilingual"
            assert (Path(assets_dir) / "multilingual.tiktoken").exists(
            ), "multilingual.tiktoken file is not existed in assets_dir"
        else:
            tokenizer_name = "gpt2"
            assert (Path(assets_dir) / "gpt2.tiktoken").exists(
            ), "gpt2.tiktoken file is not existed in assets_dir"
        self.text_prefix="<|startoftranscript|><|en|><|transcribe|><|notimestamps|>" if is_multilingual else "<|startoftranscript|><|notimestamps|>"
        self.tokenizer = get_tokenizer(name=tokenizer_name,
                                       num_languages=self.num_languages,
                                       tokenizer_dir=assets_dir)
        self.eot_id = self.tokenizer.encode(
            "<|endoftext|>",
            allowed_special=self.tokenizer.special_tokens_set)[0]
        json_config = GptJsonConfig.parse_file(Path(engine_dir) / 'decoder' / 'config.json')
        assert json_config.model_config.supports_inflight_batching
        runner_kwargs = dict(engine_dir=engine_dir,
                                is_enc_dec=True,
                                max_batch_size=batch_size,
                                max_input_len=3000,
                                max_output_len=96,
                                max_beam_width=1,
                                debug_mode=False,
                                kv_cache_free_gpu_memory_fraction=0.9)
        self.model_runner_cpp = ModelRunnerCpp.from_dir(**runner_kwargs)