Fsmn_vad支持多线程并发调用 (#2519)

王梦迪

2025-05-20 fe588bc508c0076bb007d6ed36c18ac8ecb341ac

 funasr/models/llm_asr_nar/model.py

@@ -12,10 +12,12 @@
from funasr.models.ctc.ctc import CTC
from funasr.models.transformer.utils.add_sos_eos import add_sos_eos
from funasr.metrics.compute_acc import th_accuracy, compute_accuracy

# from funasr.models.e2e_asr_common import ErrorCalculator
from funasr.train_utils.device_funcs import force_gatherable
from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank
from funasr.utils import postprocess_utils
from funasr.models.paraformer.cif_predictor import mae_loss
from funasr.utils.datadir_writer import DatadirWriter
from funasr.register import tables

@@ -23,7 +25,7 @@
@tables.register("model_classes", "LLMASRNAR")
class LLMASRNAR(nn.Module):
    """ """
    

    def __init__(
        self,
        specaug: str = None,
@@ -59,28 +61,32 @@
        # postencoder: Optional[AbsPostEncoder] = None,
        **kwargs,
    ):
        

        super().__init__()
        

        if specaug is not None:
            specaug_class = tables.specaug_classes.get(specaug)
            specaug = specaug_class(**specaug_conf)
        if normalize is not None:
            normalize_class = tables.normalize_classes.get(normalize)
            normalize = normalize_class(**normalize_conf)
        

        # audio encoder
        hub = encoder_conf.get("hub", None)
        if hub == "funasr":
            from funasr import AutoModel
            init_param_path = encoder_conf.get("init_param_path", "iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch")
            model = AutoModel(model=init_param_path, model_revision="v2.0.4")

            init_param_path = encoder_conf.get(
                "init_param_path",
                "iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch",
            )
            model = AutoModel(model=init_param_path, model_revision="master")
            # frontend = model.kwargs.get("frontend")
            model.model.decoder = None
            

            self.audio_encoder = model.model
            # self.frontend = frontend
            

        elif hub == "hf":
            pass
        else:
@@ -107,14 +113,13 @@
                    param.requires_grad = False
                model.eval()
            self.llm = model
        

        # adaptor
        adaptor_class = tables.adaptor_classes.get(adaptor)
        adaptor = adaptor_class(**adaptor_conf)
        

        self.adaptor = adaptor
        
        

        self.blank_id = blank_id
        self.sos = sos if sos is not None else vocab_size - 1
        self.eos = eos if eos is not None else vocab_size - 1
@@ -123,7 +128,6 @@
        self.specaug = specaug
        self.normalize = normalize
        self.encoder = encoder


        self.criterion_att = LabelSmoothingLoss(
            size=vocab_size,
@@ -141,7 +145,7 @@

        self.length_normalized_loss = length_normalized_loss
        self.beam_search = None
    

    def forward(
        self,
        speech: torch.Tensor,
@@ -149,7 +153,362 @@
        text: torch.Tensor,
        text_lengths: torch.Tensor,
        input_ids: torch.Tensor,
        attention_mask:torch.Tensor,
        attention_mask: torch.Tensor,
        labels_ids: torch.Tensor,
        label_mask: torch.Tensor,
        audio_mask: torch.Tensor,
        **kwargs,
    ) -> Tuple[torch.Tensor, Dict[str, torch.Tensor], torch.Tensor]:
        """Encoder + Decoder + Calc loss
        Args:
                speech: (Batch, Length, ...)
                speech_lengths: (Batch, )
                text: (Batch, Length)
                text_lengths: (Batch,)
        """
        if len(text_lengths.size()) > 1:
            text_lengths = text_lengths[:, 0]
        if len(speech_lengths.size()) > 1:
            speech_lengths = speech_lengths[:, 0]

        batch_size = speech.shape[0]

        # audio encoder
        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths, audio_mask=audio_mask)

        # adaptor
        encoder_out = self.adaptor(encoder_out)

        if input_ids is not None:
            input_ids[input_ids == -1] = 0
            input_ids[input_ids == -100] = 0
            if hasattr(self.llm.model, "embed_tokens"):
                inputs_embeds = self.llm.model.embed_tokens(input_ids)
            elif hasattr(self.llm.model.model, "embed_tokens"):
                inputs_embeds = self.llm.model.model.embed_tokens(input_ids)
            else:
                inputs_embeds = self.llm.model.model.model.embed_tokens(input_ids)

            if audio_mask is not None:
                batch_size, token_num, dims = inputs_embeds.shape
                _, l, _ = encoder_out.shape
                encoder_outs_pad = F.pad(encoder_out, (0, 0, token_num - l - 1, 1, 0, 0), value=0.0)
                inputs_embeds = encoder_outs_pad * audio_mask[:, :, None] + inputs_embeds * (
                    1.0 - audio_mask[:, :, None]
                )
                inputs_embeds = F.pad(inputs_embeds[:, 1:, :], (0, 0, 0, 1, 0, 0), value=0.0)

        model_outputs = self.llm(
            inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels_ids
        )
        loss = model_outputs.loss

        stats = {}
        with torch.no_grad():
            preds = torch.argmax(model_outputs.logits, -1)
            acc_att = compute_accuracy(preds[:, :-1], labels_ids[:, 1:], ignore_label=-100)
            stats["acc"] = acc_att

        stats["loss"] = torch.clone(loss.detach())

        # force_gatherable: to-device and to-tensor if scalar for DataParallel
        if self.length_normalized_loss:
            batch_size = int((text_lengths + 1).sum())
        loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
        return loss, stats, weight

    def encode(
        self,
        speech: torch.Tensor,
        speech_lengths: torch.Tensor,
        **kwargs,
    ):

        audio_mask = kwargs.get("audio_mask", None)
        audio_token_lengths = audio_mask.sum(-1) if audio_mask is not None else None
        text_token_int = kwargs.get("text_token_int", None)
        if audio_token_lengths is None:
            audio_token_lengths = torch.tensor([len(text_token_int)], dtype=torch.int64)

        batch = {"speech": speech, "speech_lengths": speech_lengths}
        enc, enc_lens = self.audio_encoder.encode(**batch)
        with autocast(False):
            enc_mask = sequence_mask(enc_lens, enc.size(1), device=enc.device)[:, None, :]
            pre_acoustic_embeds, pre_token_length, _, _ = self.audio_encoder.predictor(
                enc,
                mask=enc_mask,
                target_label_length=audio_token_lengths,
            )

        return pre_acoustic_embeds, pre_token_length

    def inference(
        self,
        data_in,
        data_lengths=None,
        key: list = None,
        tokenizer=None,
        frontend=None,
        **kwargs,
    ):

        prompt = kwargs.get("prompt", "Transcribe speech to text.")

        if kwargs.get("batch_size", 1) > 1:
            raise NotImplementedError("batch decoding is not implemented")

        meta_data = {}
        if (
            isinstance(data_in, torch.Tensor) and kwargs.get("data_type", "sound") == "fbank"
        ):  # fbank
            speech, speech_lengths = data_in, data_lengths
            if len(speech.shape) < 3:
                speech = speech[None, :, :]
            if speech_lengths is None:
                speech_lengths = speech.shape[1]
        else:
            # extract fbank feats
            time1 = time.perf_counter()
            audio_sample_list = load_audio_text_image_video(
                data_in,
                fs=frontend.fs,
                audio_fs=kwargs.get("fs", 16000),
                data_type=kwargs.get("data_type", "sound"),
                tokenizer=None,
            )
            if len(kwargs.get("data_type", [])) > 1:
                audio_sample_list, text_token_int_list = audio_sample_list
                text_token_int = text_token_int_list[0].replace(" ", "")
                text_token_int = tokenizer.encode(text_token_int)
            else:
                text_token_int = None
            time2 = time.perf_counter()
            meta_data["load_data"] = f"{time2 - time1:0.3f}"
            speech, speech_lengths = extract_fbank(
                audio_sample_list, data_type=kwargs.get("data_type", "sound"), frontend=frontend
            )
            time3 = time.perf_counter()
            meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
            meta_data["batch_data_time"] = (
                speech_lengths.sum().item() * frontend.frame_shift * frontend.lfr_n / 1000
            )

        speech = speech.to(device=kwargs["device"])
        speech_lengths = speech_lengths.to(device=kwargs["device"])

        # Encoder
        encoder_out, encoder_out_lens = self.encode(
            speech, speech_lengths, text_token_int=text_token_int
        )

        # adaptor
        encoder_out = self.adaptor(encoder_out)

        prompt_pre = "USER: \nINSTRUCTION: {}\nINPUT: ".format(prompt)
        prompt_ids = tokenizer.encode(prompt_pre)
        prompt_length = len(prompt_ids)
        prompt_ids = torch.tensor(prompt_ids, dtype=torch.int64).to(kwargs["device"])

        if hasattr(self.llm.model, "embed_tokens"):
            inputs_embeds = self.llm.model.embed_tokens(prompt_ids)
        elif hasattr(self.llm.model.model, "embed_tokens"):
            inputs_embeds = self.llm.model.model.embed_tokens(prompt_ids)
        else:
            inputs_embeds = self.llm.model.model.model.embed_tokens(prompt_ids)

        inputs_embeds = torch.cat(
            (inputs_embeds[None, :, :], encoder_out), dim=1
        )  # [prompt, audio]
        attention_mask = torch.ones(inputs_embeds.size()[:-1], dtype=torch.long).to(
            kwargs["device"]
        )

        # model_outputs = self.llm.generate(
        #     inputs_embeds=inputs_embeds,
        #     max_length=kwargs.get("max_length", 200),
        #     max_new_tokens=kwargs.get("max_new_tokens", 200),
        #     num_beams=kwargs.get("num_beams", 4),
        #     do_sample=kwargs.get("do_sample", False),
        #     min_length=kwargs.get("min_length", 1),
        #     top_p=kwargs.get("top_p", 1.0),
        #     repetition_penalty=kwargs.get("repetition_penalty", 1.0),
        #     length_penalty=kwargs.get("length_penalty", 1.0),
        #     temperature=kwargs.get("temperature", 1.0),
        #     attention_mask=attention_mask,
        #     bos_token_id=tokenizer.bos_token_id,
        #     eos_token_id=tokenizer.eos_token_id,
        #     pad_token_id=tokenizer.pad_token_id
        # )

        model_outputs = self.llm(
            inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=None
        )
        preds = torch.argmax(model_outputs.logits, -1)
        text = tokenizer.batch_decode(preds, add_special_tokens=False, skip_special_tokens=True)

        text = text[0].split(": ")[-1]
        text = text.strip()

        # preds = torch.argmax(model_outputs.logits, -1)

        ibest_writer = None
        if kwargs.get("output_dir") is not None:
            if not hasattr(self, "writer"):
                self.writer = DatadirWriter(kwargs.get("output_dir"))
            ibest_writer = self.writer[f"{0 + 1}best_recog"]

        results = []
        result_i = {"key": key[0], "text": text}
        results.append(result_i)

        if ibest_writer is not None:
            ibest_writer["text"][key[0]] = text

        return results, meta_data


@tables.register("model_classes", "LLMASRNARPrompt")
class LLMASRNARPrompt(nn.Module):
    """ """

    def __init__(
        self,
        specaug: str = None,
        specaug_conf: dict = None,
        normalize: str = None,
        normalize_conf: dict = None,
        encoder: str = None,
        encoder_conf: dict = None,
        decoder: str = None,
        decoder_conf: dict = None,
        ctc: str = None,
        ctc_conf: dict = None,
        ctc_weight: float = 0.0,
        llm: str = None,
        llm_conf: dict = None,
        adaptor: str = None,
        adaptor_conf: dict = None,
        input_size: int = 80,
        vocab_size: int = -1,
        ignore_id: int = -1,
        blank_id: int = 0,
        sos: int = 1,
        eos: int = 2,
        lsm_weight: float = 0.0,
        length_normalized_loss: bool = False,
        predictor_weight: int = 1.0,
        report_cer: bool = True,
        report_wer: bool = True,
        sym_space: str = "<space>",
        sym_blank: str = "<blank>",
        # extract_feats_in_collect_stats: bool = True,
        share_embedding: bool = False,
        # preencoder: Optional[AbsPreEncoder] = None,
        # postencoder: Optional[AbsPostEncoder] = None,
        **kwargs,
    ):

        super().__init__()

        if specaug is not None:
            specaug_class = tables.specaug_classes.get(specaug)
            specaug = specaug_class(**specaug_conf)
        if normalize is not None:
            normalize_class = tables.normalize_classes.get(normalize)
            normalize = normalize_class(**normalize_conf)

        # audio encoder
        hub = encoder_conf.get("hub", None)
        if hub == "funasr":
            from funasr import AutoModel

            init_param_path = encoder_conf.get(
                "init_param_path",
                "iic/speech_paraformer-large_asr_nat-zh-cn-16k-common-vocab8404-pytorch",
            )
            model = AutoModel(model=init_param_path, model_revision="master")
            # frontend = model.kwargs.get("frontend")
            model.model.decoder = None

            self.audio_encoder = model.model
            # self.frontend = frontend
            self.predictor_weight = predictor_weight

        elif hub == "hf":
            pass
        else:
            encoder_class = tables.encoder_classes.get(encoder)
            encoder = encoder_class(input_size=input_size, **encoder_conf)
            encoder_output_size = encoder.output_size()

        # llm
        hub = llm_conf.get("hub", "hf")
        self.llm = None
        if hub == "hf":
            from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig

            init_param_path = llm_conf.get("init_param_path", "vicuna-7b-v1.5")
            model = AutoModelForCausalLM.from_pretrained(
                init_param_path,
                load_in_8bit=None,
                device_map=None,
                use_cache=None,
            )
            freeze = llm_conf.get("freeze", True)
            if freeze:
                for name, param in model.named_parameters():
                    param.requires_grad = False
                model.eval()
            self.llm = model

        # adaptor
        adaptor_class = tables.adaptor_classes.get(adaptor)
        adaptor = adaptor_class(**adaptor_conf)

        self.adaptor = adaptor

        self.blank_id = blank_id
        self.sos = sos if sos is not None else vocab_size - 1
        self.eos = eos if eos is not None else vocab_size - 1
        self.vocab_size = vocab_size
        self.ignore_id = ignore_id
        self.specaug = specaug
        self.normalize = normalize
        self.encoder = encoder

        self.criterion_att = LabelSmoothingLoss(
            size=vocab_size,
            padding_idx=ignore_id,
            smoothing=lsm_weight,
            normalize_length=length_normalized_loss,
        )
        self.criterion_pre = mae_loss(normalize_length=length_normalized_loss)
        #
        # if report_cer or report_wer:
        #     self.error_calculator = ErrorCalculator(
        #         token_list, sym_space, sym_blank, report_cer, report_wer
        #     )
        #
        self.error_calculator = None

        self.length_normalized_loss = length_normalized_loss
        self.beam_search = None
        if ctc_weight > 0.0:
            if ctc_conf is None:
                ctc_conf = {}

            ctc = CTC(odim=vocab_size, encoder_output_size=adaptor_conf["encoder_dim"], **ctc_conf)
        self.ctc_weight = ctc_weight
        self.ctc = ctc

    def forward(
        self,
        speech: torch.Tensor,
        speech_lengths: torch.Tensor,
        text: torch.Tensor,
        text_lengths: torch.Tensor,
        input_ids: torch.Tensor,
        attention_mask: torch.Tensor,
        labels_ids: torch.Tensor,
        label_mask: torch.Tensor,
        audio_mask: torch.Tensor,
@@ -168,12 +527,23 @@
            text_lengths = text_lengths[:, 0]
        if len(speech_lengths.size()) > 1:
            speech_lengths = speech_lengths[:, 0]
        

        batch_size = speech.shape[0]
        

        stats = {}
        # audio encoder
        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths, audio_mask=audio_mask)
        
        outs = self.encode(speech, speech_lengths, audio_mask=audio_mask)
        enc, enc_lens = outs[0], outs[1]
        encoder_out, encoder_out_lens, loss_pre = outs[2], outs[3], outs[4]

        # decoder: CTC branch

        if self.ctc_weight != 0.0:
            loss_ctc, cer_ctc = self._calc_ctc_loss(enc, enc_lens, text, text_lengths)

            # Collect CTC branch stats
            stats["loss_ctc"] = torch.clone(loss_ctc.detach()) if loss_ctc is not None else None

        # adaptor
        encoder_out = self.adaptor(encoder_out)

@@ -188,67 +558,118 @@
                inputs_embeds = self.llm.model.model.model.embed_tokens(input_ids)

            if audio_mask is not None:
                # inputs_embeds： [bos, prompt, input, pad, target]
                prompt_bos_length = kwargs.get("prompt_bos_length", None)
                assert prompt_bos_length is not None
                prompt_bos_length = prompt_bos_length[0].item()
                batch_size, token_num, dims = inputs_embeds.shape
                _, l, _ = encoder_out.shape
                encoder_outs_pad = F.pad(encoder_out, (0, 0, token_num-l-1, 1, 0, 0), value=0.0)
                inputs_embeds = encoder_outs_pad * audio_mask[:, :, None] + inputs_embeds * (1.0-audio_mask[:, :, None])
                inputs_embeds = F.pad(inputs_embeds[:, 1:, :], (0, 0, 0, 1, 0, 0), value=0.0)
                encoder_outs_pad = F.pad(
                    encoder_out,
                    (0, 0, prompt_bos_length, token_num - prompt_bos_length - l, 0, 0),
                    value=0.0,
                )
                inputs_embeds = encoder_outs_pad * audio_mask[:, :, None] + inputs_embeds * (
                    1.0 - audio_mask[:, :, None]
                )
                inputs_embeds = F.pad(
                    inputs_embeds[:, 1:, :], (0, 0, 0, 1, 0, 0), value=0.0
                )  # [prompt, input, pad, target, 0.0]

        model_outputs = self.llm(inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels_ids)
        loss = model_outputs.loss
        # labels_ids: [bos, prompt, input, target, eos] -> [-1, -1, input, target, eos]
        # loss:
        # inputs_embeds[:-1] -> [prompt, input, pad, target]
        # labels_ids[1:] ->  [prompt, input, target, eos] -> [-1, input, target, eos];
        model_outputs = self.llm(
            inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels_ids
        )
        loss_llm = model_outputs.loss
        stats["loss_llm"] = torch.clone(loss_llm.detach())
        if self.ctc_weight > 0.0:
            loss_llm = self.ctc_weight * loss_ctc + loss_llm
        loss = loss_llm + loss_pre * self.predictor_weight


        stats = {}
        with torch.no_grad():
            preds = torch.argmax(model_outputs.logits, -1)
            acc_att = compute_accuracy(preds[:, :-1], labels_ids[:, 1:], ignore_label=-100)
            stats["acc"] = acc_att

        stats["loss_pre"] = torch.clone(loss_pre.detach())
        stats["loss"] = torch.clone(loss.detach())
        stats["batch_size"] = batch_size

        # force_gatherable: to-device and to-tensor if scalar for DataParallel
        if self.length_normalized_loss:
            batch_size = int((text_lengths + 1).sum())
        loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
        return loss, stats, weight
    

    def encode(
        self, speech: torch.Tensor, speech_lengths: torch.Tensor, **kwargs,
        self,
        speech: torch.Tensor,
        speech_lengths: torch.Tensor,
        **kwargs,
    ):
    

        audio_mask = kwargs.get("audio_mask", None)
        audio_token_lengths = audio_mask.sum(-1) if audio_mask is not None else None
        text_token_int = kwargs.get("text_token_int", None)
        if audio_token_lengths is None and text_token_int is not None:
            audio_token_lengths = torch.tensor([len(text_token_int)], dtype=torch.int64)

        batch = {"speech": speech, "speech_lengths": speech_lengths}
        enc, enc_lens = self.audio_encoder.encode(**batch)
        with autocast(False):
            enc_mask = sequence_mask(enc_lens, enc.size(1), device=enc.device)[:, None, :]
            pre_acoustic_embeds, pre_token_length, _, _ = self.audio_encoder.predictor(enc,
                                                                               mask=enc_mask,
                                                                               target_label_length=audio_token_lengths,
                                                                               )
            pre_acoustic_embeds, pre_token_length, _, _ = self.audio_encoder.predictor(
                enc,
                mask=enc_mask,
                target_label_length=audio_token_lengths,
            )
            loss_pre = 0.0
            if audio_token_lengths is not None:
                loss_pre = self.criterion_pre(
                    audio_token_lengths.type_as(pre_token_length), pre_token_length
                )

        return pre_acoustic_embeds, pre_token_length
        return enc, enc_lens, pre_acoustic_embeds, pre_token_length, loss_pre

    def _calc_ctc_loss(
        self,
        encoder_out: torch.Tensor,
        encoder_out_lens: torch.Tensor,
        ys_pad: torch.Tensor,
        ys_pad_lens: torch.Tensor,
    ):
        # Calc CTC loss
        loss_ctc = self.ctc(encoder_out, encoder_out_lens, ys_pad, ys_pad_lens)

    def inference(self,
                  data_in,
                  data_lengths=None,
                  key: list = None,
                  tokenizer=None,
                  frontend=None,
                  **kwargs,
                  ):
        
        # Calc CER using CTC
        cer_ctc = None
        if not self.training and self.error_calculator is not None:
            ys_hat = self.ctc.argmax(encoder_out).data
            cer_ctc = self.error_calculator(ys_hat.cpu(), ys_pad.cpu(), is_ctc=True)
        return loss_ctc, cer_ctc

    def inference(
        self,
        data_in,
        data_lengths=None,
        key: list = None,
        tokenizer=None,
        frontend=None,
        **kwargs,
    ):

        prompt = kwargs.get("prompt", "Transcribe speech to text.")
        

        if kwargs.get("batch_size", 1) > 1:
            raise NotImplementedError("batch decoding is not implemented")


        
        meta_data = {}
        if isinstance(data_in, torch.Tensor) and kwargs.get("data_type", "sound") == "fbank":  # fbank
        if (
            isinstance(data_in, torch.Tensor) and kwargs.get("data_type", "sound") == "fbank"
        ):  # fbank
            speech, speech_lengths = data_in, data_lengths
            if len(speech.shape) < 3:
                speech = speech[None, :, :]
@@ -257,43 +678,67 @@
        else:
            # extract fbank feats
            time1 = time.perf_counter()
            audio_sample_list = load_audio_text_image_video(data_in, fs=frontend.fs, audio_fs=kwargs.get("fs", 16000),
                                                            data_type=kwargs.get("data_type", "sound"),
                                                            tokenizer=tokenizer)
            audio_sample_list = load_audio_text_image_video(
                data_in,
                fs=frontend.fs,
                audio_fs=kwargs.get("fs", 16000),
                data_type=kwargs.get("data_type", "sound"),
                tokenizer=None,
            )
            if len(kwargs.get("data_type", [])) > 1:
                audio_sample_list, text_token_int_list = audio_sample_list
                text_token_int = text_token_int_list[0]
                text_token_int = tokenizer.encode(text_token_int)
                if text_token_int[0] == tokenizer.bos_token_id:
                    text_token_int = text_token_int[1:]
            else:
                text_token_int = None
            time2 = time.perf_counter()
            meta_data["load_data"] = f"{time2 - time1:0.3f}"
            speech, speech_lengths = extract_fbank(audio_sample_list, data_type=kwargs.get("data_type", "sound"),
                                                   frontend=frontend)
            speech, speech_lengths = extract_fbank(
                audio_sample_list, data_type=kwargs.get("data_type", "sound"), frontend=frontend
            )
            time3 = time.perf_counter()
            meta_data["extract_feat"] = f"{time3 - time2:0.3f}"
            meta_data["batch_data_time"] = speech_lengths.sum().item() * frontend.frame_shift * frontend.lfr_n / 1000
        
            meta_data["batch_data_time"] = (
                speech_lengths.sum().item() * frontend.frame_shift * frontend.lfr_n / 1000
            )

        speech = speech.to(device=kwargs["device"])
        speech_lengths = speech_lengths.to(device=kwargs["device"])
        

        # Encoder
        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
        res = self.encode(speech, speech_lengths, text_token_int=text_token_int)
        encoder_out = res[0]

        # adaptor
        encoder_out = self.adaptor(encoder_out)
        
    

        prompt_pre = "USER: \nINSTRUCTION: {}\nINPUT: ".format(prompt)
        prompt_ids = tokenizer.encode(prompt_pre)
        if prompt_ids[0] == tokenizer.bos_token_id:
            prompt_ids = prompt_ids[1:]
        # prompt_ids = prompt_ids + [tokenizer.pad_token_id]
        prompt_length = len(prompt_ids)
        prompt_ids = torch.tensor(prompt_ids, dtype=torch.int64).to(kwargs["device"])

        pad = torch.tensor([tokenizer.pad_token_id], dtype=torch.int64).to(kwargs["device"])

        if hasattr(self.llm.model, "embed_tokens"):
            inputs_embeds = self.llm.model.embed_tokens(prompt_ids)
            pad = self.llm.model.embed_tokens(pad)
        elif hasattr(self.llm.model.model, "embed_tokens"):
            inputs_embeds = self.llm.model.model.embed_tokens(prompt_ids)
        else:
            inputs_embeds = self.llm.model.model.model.embed_tokens(prompt_ids)

        inputs_embeds = torch.cat((inputs_embeds[None, :, :], encoder_out), dim=1)  # [prompt, audio]
        attention_mask = torch.ones(inputs_embeds.size()[:-1], dtype=torch.long).to(kwargs["device"])
        
        # inputs_embeds = torch.cat((inputs_embeds[None, :, :], encoder_out, pad[None, :, :]), dim=1)  # [prompt, audio, pad]
        inputs_embeds = torch.cat(
            (inputs_embeds[None, :, :], encoder_out), dim=1
        )  # [prompt, audio]
        attention_mask = torch.ones(inputs_embeds.size()[:-1], dtype=torch.long).to(
            kwargs["device"]
        )

        # model_outputs = self.llm.generate(
        #     inputs_embeds=inputs_embeds,
        #     max_length=kwargs.get("max_length", 200),
@@ -311,16 +756,20 @@
        #     pad_token_id=tokenizer.pad_token_id
        # )


        model_outputs = self.llm(inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=None)
        model_outputs = self.llm(
            inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=None
        )
        preds = torch.argmax(model_outputs.logits, -1)
        text = tokenizer.batch_decode(preds, add_special_tokens=False, skip_special_tokens=True)

        text = text[0].split(': ')[-1]
        text = text[0].split(":")[-1]
        text = text.strip()
        
        if text.startswith("Please\n "):
            text = text.replace("Please\n ", "")
            text = text.strip()

        # preds = torch.argmax(model_outputs.logits, -1)
        

        ibest_writer = None
        if kwargs.get("output_dir") is not None:
            if not hasattr(self, "writer"):
@@ -333,9 +782,5 @@

        if ibest_writer is not None:
            ibest_writer["text"][key[0]] = text
        
        
        
        
        return results, meta_data

        return results, meta_data