decoding

游雁

2024-06-19 45d7aa9004763684fb748ee17942ecba81042201

decoding

 docs/images/wechat.png

 examples/industrial_data_pretraining/ctc/demo.py

@@ -6,8 +6,11 @@
import sys
from funasr import AutoModel

model_dir=sys.argv[1]
input_file=sys.argv[2]

model_dir = "/Users/zhifu/Downloads/modelscope_models/ctc_model"
input_file = (
    "https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav"
)

model = AutoModel(
    model=model_dir,

 examples/industrial_data_pretraining/ctc/infer_from_local.sh

 examples/industrial_data_pretraining/llm_asr/conf/whisper_qwen_linear2.yaml

@@ -69,7 +69,7 @@
  batch_size_scale_ratio_max: 2
  num_workers: 4
  audio_adaptor_downsample_rate: ${audio_adaptor_conf.downsample_rate}
  audio_encoder_downsample_rate: 2
  audio_encoder_downsample_rate: 4
  data_split_num: 512
  batch_size_sample_max: 15
  retry: 20

 examples/industrial_data_pretraining/llm_asr/demo_speech2text.py

@@ -28,6 +28,9 @@
    init_param=f"{os.path.join(ckpt_dir, ckpt_id)}",
    output_dir=output_dir,
    device=device,
    fp16=False,
    bf16=False,
    llm_dtype="bf16",
)



 examples/industrial_data_pretraining/llm_asr/demo_speech2text.sh

@@ -12,6 +12,7 @@
out_dir="${ckpt_dir}/inference-${ckpt_id}"
mkdir -p ${out_dir}
for data_set in "librispeech_test_clean_speech2text.jsonl" "librispeech_test_other_speech2text.jsonl"; do
{
    jsonl=${jsonl_dir}/${data_set}
    output_dir=${out_dir}/${data_set}
    mkdir -p ${output_dir}
@@ -22,10 +23,12 @@

    python /mnt/workspace/zhifu.gzf/codebase/FunASR/funasr/metrics/wer.py ++ref_file=${ref_file} ++hyp_file=${pred_file} ++cer_file=${pred_file}.cer ++cn_postprocess=false

}&
done
wait


for data_set in "aishell1_test_speech2text.jsonl" "aishell2_ios_test_speech2text.jsonl" "librispeech_test_other_speech2text.jsonl"; do
for data_set in "aishell1_test_speech2text.jsonl" "aishell2_ios_test_speech2text.jsonl"; do
{
    jsonl=${jsonl_dir}/${data_set}
    output_dir=${out_dir}/${data_set}
    mkdir -p ${output_dir}
@@ -36,9 +39,12 @@

    python /mnt/workspace/zhifu.gzf/codebase/FunASR/funasr/metrics/wer.py ++ref_file=${ref_file} ++hyp_file=${pred_file} ++cer_file=${pred_file}.cer ++cn_postprocess=true

}&
done
wait

for data_set in "s2tt_en2zh.v20240605.test.jsonl"; do
for data_set in "common_voice_zh-CN_speech2text.jsonl" "common_voice_en_speech2text.jsonl"; do
{
    jsonl=${jsonl_dir}/${data_set}
    output_dir=${out_dir}/${data_set}
    mkdir -p ${output_dir}
@@ -47,19 +53,13 @@

    python ./demo_speech2text.py ${ckpt_dir} ${ckpt_id} ${jsonl} ${output_dir} ${device}

    python /mnt/workspace/zhifu.gzf/codebase/FunASR/funasr/metrics/wer.py ++ref_file=${ref_file} ++hyp_file=${pred_file} ++cer_file=${pred_file}.cer ++cn_postprocess=true
    cn_postprocess=false
    if [ $data_set = "common_voice_zh-CN_speech2text.jsonl" ];then
      cn_postprocess=true
    fi

    python /mnt/workspace/zhifu.gzf/codebase/FunASR/funasr/metrics/wer.py ++ref_file=${ref_file} ++hyp_file=${pred_file} ++cer_file=${pred_file}.cer ++cn_postprocess=${cn_postprocess}

}&
done

for data_set in "s2tt_zh2en.v20240605.test.jsonl"; do
    jsonl=${jsonl_dir}/${data_set}
    output_dir=${out_dir}/${data_set}
    mkdir -p ${output_dir}
    pred_file=${output_dir}/1best_recog/text_tn
    ref_file=${output_dir}/1best_recog/label

    python ./demo_speech2text.py ${ckpt_dir} ${ckpt_id} ${jsonl} ${output_dir} ${device}

    python /mnt/workspace/zhifu.gzf/codebase/FunASR/funasr/metrics/wer.py ++ref_file=${ref_file} ++hyp_file=${pred_file} ++cer_file=${pred_file}.cer ++cn_postprocess=false

done
wait

 examples/industrial_data_pretraining/sense_voice/demo_ctc.py

New file
@@ -0,0 +1,25 @@
#!/usr/bin/env python3
# -*- encoding: utf-8 -*-
# Copyright FunASR (https://github.com/alibaba-damo-academy/FunASR). All Rights Reserved.
#  MIT License  (https://opensource.org/licenses/MIT)

import sys
from funasr import AutoModel

model_dir = "/Users/zhifu/Downloads/modelscope_models/SenseVoiceCTC"
input_file = (
    "https://isv-data.oss-cn-hangzhou.aliyuncs.com/ics/MaaS/ASR/test_audio/asr_example_zh.wav"
)

model = AutoModel(
    model=model_dir,
)

res = model.generate(
    input=input_file,
    cache={},
    language="zh",
    text_norm="wotextnorm",
)

print(res)

 funasr/auto/auto_model.py

@@ -213,7 +213,6 @@
        deep_update(model_conf, kwargs.get("model_conf", {}))
        deep_update(model_conf, kwargs)
        model = model_class(**model_conf, vocab_size=vocab_size)
        model.to(device)

        # init_param
        init_param = kwargs.get("init_param", None)
@@ -236,6 +235,7 @@
            model.to(torch.float16)
        elif kwargs.get("bf16", False):
            model.to(torch.bfloat16)
        model.to(device)
        return model, kwargs

    def __call__(self, *args, **cfg):
@@ -324,7 +324,7 @@
            input, input_len=input_len, model=self.vad_model, kwargs=self.vad_kwargs, **cfg
        )
        end_vad = time.time()
            

        #  FIX(gcf): concat the vad clips for sense vocie model for better aed
        if kwargs.get("merge_vad", False):
            for i in range(len(res)):
@@ -466,7 +466,7 @@
                            result[k] = restored_data[j][k]
                        else:
                            result[k] += restored_data[j][k]
                            

            if not len(result["text"].strip()):
                continue
            return_raw_text = kwargs.get("return_raw_text", False)
@@ -481,7 +481,7 @@
                if return_raw_text:
                    result["raw_text"] = raw_text
                result["text"] = punc_res[0]["text"]
                

            # speaker embedding cluster after resorted
            if self.spk_model is not None and kwargs.get("return_spk_res", True):
                if raw_text is None:
@@ -602,6 +602,6 @@
        )

        with torch.no_grad():
            export_dir = export_utils.export(model=model, data_in=data_list,  **kwargs)
            export_dir = export_utils.export(model=model, data_in=data_list, **kwargs)

        return export_dir

 funasr/bin/train_ds.py

@@ -84,6 +84,8 @@
        dist.init_process_group(backend=kwargs.get("backend", "nccl"), init_method="env://")
        torch.cuda.set_device(local_rank)

    # rank = dist.get_rank()

    logging.info("Build model, frontend, tokenizer")
    device = kwargs.get("device", "cuda")
    kwargs["device"] = "cpu"
@@ -124,6 +126,7 @@
        use_ddp=use_ddp,
        use_fsdp=use_fsdp,
        device=kwargs["device"],
        excludes=kwargs.get("excludes", None),
        output_dir=kwargs.get("output_dir", "./exp"),
        **kwargs.get("train_conf"),
    )
@@ -143,7 +146,7 @@
    dataloader = dataloader_class(**kwargs)
    # dataloader_tr, dataloader_val = dataloader_class(**kwargs)

    scaler = GradScaler(enabled=trainer.use_fp16) if trainer.use_fp16 else None
    scaler = GradScaler(enabled=True) if trainer.use_fp16 or trainer.use_bf16 else None
    scaler = ShardedGradScaler(enabled=trainer.use_fp16) if trainer.use_fsdp else scaler

    trainer.resume_checkpoint(
@@ -182,7 +185,7 @@

            time_escaped = (time.perf_counter() - time_slice_i) / 3600.0
            logging.info(
                f"rank: {local_rank}, "
                f"\n\nrank: {local_rank}, "
                f"time_escaped_epoch: {time_escaped:.3f} hours, "
                f"estimated to finish {dataloader.data_split_num} data_slices, remaining: {dataloader.data_split_num-data_split_i} slices, {(dataloader.data_split_num-data_split_i)*time_escaped:.3f} hours, "
                f"epoch: {trainer.max_epoch - epoch} epochs, {((trainer.max_epoch - epoch - 1)*dataloader.data_split_num + dataloader.data_split_num-data_split_i)*time_escaped:.3f} hours\n"
@@ -199,7 +202,7 @@
        time2 = time.perf_counter()
        time_escaped = (time2 - time1) / 3600.0
        logging.info(
            f"rank: {local_rank}, "
            f"\n\nrank: {local_rank}, "
            f"time_escaped_epoch: {time_escaped:.3f} hours, "
            f"estimated to finish {trainer.max_epoch} "
            f"epoch: {(trainer.max_epoch - epoch) * time_escaped:.3f} hours\n"

 funasr/datasets/openai_datasets/datasets.py

@@ -64,6 +64,8 @@
        self.max_token_length = kwargs.get("max_token_length", 1024)
        self.batch_size_scale_ratio_max = kwargs.get("batch_size_scale_ratio_max", 1.5)
        self.batch_size_token_max = kwargs.get("batch_size_token_max", 2500)
        self.audio_adaptor_downsample_rate = kwargs.get("audio_adaptor_downsample_rate", 2)
        self.audio_encoder_downsample_rate = kwargs.get("audio_encoder_downsample_rate", 4)

    def get_source_len(self, index):
        item = self.index_ds[index]
@@ -136,10 +138,13 @@
                                speech = speech.permute(0, 2, 1)
                            # if speech_lengths > self.batch_size:
                            #     continue
                            if self.audio_encoder_downsample_rate == 4:
                                olens = 1 + (speech_lengths[0].item() - 3 + 2 * 1) // 2
                                olens = 1 + (olens - 3 + 2 * 1) // 2
                            elif self.audio_encoder_downsample_rate == 1:
                                olens = speech_lengths[0].item()

                            olens = 1 + (speech_lengths[0].item() - 3 + 2 * 1) // 2
                            olens = 1 + (olens - 3 + 2 * 1) // 2
                            sub_token_len = (olens - 1) // 2 + 1
                            sub_token_len = (olens - 1) // self.audio_adaptor_downsample_rate + 1
                            sub_token = [0] * sub_token_len
                            fbank_beg_i = [len(source_ids)]
                            source_ids += sub_token
@@ -222,3 +227,247 @@
            break

        return outputs


@tables.register("dataset_classes", "OpenAIDatasetMultiTurn")
class OpenAIDatasetMultiTurn(torch.utils.data.Dataset):
    """
    SenseVoiceDataset
    """

    def __init__(
        self,
        path,
        index_ds: str = None,
        frontend=None,
        tokenizer=None,
        int_pad_value: int = -1,
        float_pad_value: float = 0.0,
        **kwargs,
    ):
        super().__init__()
        index_ds_class = tables.index_ds_classes.get(index_ds)
        self.index_ds = index_ds_class(path, **kwargs)
        preprocessor_speech = kwargs.get("preprocessor_speech", None)
        if preprocessor_speech:
            preprocessor_speech_class = tables.preprocessor_classes.get(preprocessor_speech)
            preprocessor_speech = preprocessor_speech_class(
                **kwargs.get("preprocessor_speech_conf")
            )
        self.preprocessor_speech = preprocessor_speech
        preprocessor_text = kwargs.get("preprocessor_text", None)
        if preprocessor_text:
            preprocessor_text_class = tables.preprocessor_classes.get(preprocessor_text)
            preprocessor_text = preprocessor_text_class(**kwargs.get("preprocessor_text_conf"))
        self.preprocessor_text = preprocessor_text

        self.frontend = frontend
        self.fs = 16000 if frontend is None else frontend.fs
        self.data_type = "sound"
        self.tokenizer = tokenizer

        self.int_pad_value = int_pad_value
        self.float_pad_value = float_pad_value
        self.sos = kwargs.get("sos", "<|startoftranscript|>")
        self.eos = kwargs.get("eos", "<|endoftext|>")
        self.batch_size = kwargs.get("batch_size")
        self.batch_type = kwargs.get("batch_type")
        self.prompt_ids_len = 0
        self.retry = kwargs.get("retry", 100)

        self.permute = False
        from funasr.frontends.whisper_frontend import WhisperFrontend

        if isinstance(self.frontend, WhisperFrontend):
            self.permute = True

        self.pattern = re.compile(r"(<\|startofspeech\|>.*?<\|endofspeech\|>)")
        # self.kwargs = kwargs
        self.max_token_length = kwargs.get("max_token_length", 1024)
        self.batch_size_scale_ratio_max = kwargs.get("batch_size_scale_ratio_max", 1.5)
        self.batch_size_token_max = kwargs.get("batch_size_token_max", 2500)
        self.multiturn_num_max = kwargs.get("multiturn_num_max", 5)

    def get_source_len(self, index):
        item = self.index_ds[index]
        return self.index_ds.get_source_len(item)

    def get_target_len(self, index):
        item = self.index_ds[index]
        return self.index_ds.get_target_len(item)

    def __len__(self):
        return len(self.index_ds)

    def __getitem__(self, index):
        # import pdb
        #
        # pdb.set_trace()

        output = None

        for idx in range(self.retry):
            badcase_flag = False
            if idx == 0:
                index_cur = index
            else:
                index_cur = torch.randint(0, len(self.index_ds), ()).item()

            item = self.index_ds[index_cur]

            system = item["system"]
            user = item["user"]
            assistant = item["assistant"]

            input_ids, labels, fbank, fbank_lens, fbank_mask, fbank_beg, fake_token_len = (
                [],
                [],
                [],
                [],
                [],
                [],
                [],
            )

            for i, (system_prompt, user_prompt, target_out) in enumerate(
                zip(system, user, assistant)
            ):
                if i >= self.multiturn_num_max:
                    break
                if i == 0:
                    source_input = f"<|im_start|>system\n{system_prompt}<|im_end|>\n<|im_start|>user\n{user_prompt}<|im_end|>\n<|im_start|>assistant\n"
                else:
                    source_input = (
                        f"<|im_start|>user\n{user_prompt}<|im_end|>\n<|im_start|>assistant\n"
                    )

                splits = self.pattern.split(source_input)
                source_ids = []
                fbank_i = []
                fbank_mask_i = []
                fake_token_len_i = 0
                fbank_beg_i = -1
                fbank_lens_i = []
                for k, sub_str in enumerate(splits):
                    if not sub_str.startswith("<|startofspeech|>"):
                        sub_token = self.tokenizer.encode(sub_str)
                        source_ids += sub_token
                        fbank_mask_i += [0] * len(sub_token)
                    else:
                        sub_str = sub_str.replace("<|startofspeech|>", "").replace(
                            "<|endofspeech|>", ""
                        )
                        if sub_str.startswith("!"):
                            try:
                                data_src = load_audio_text_image_video(sub_str[1:], fs=self.fs)
                            except Exception as e:
                                logging.error(
                                    f"Loading wav failed! {str(e)}, {traceback.format_exc()}"
                                )
                                badcase_flag = True
                                continue
                            speech, speech_lengths = extract_fbank(
                                data_src,
                                data_type=self.data_type,
                                frontend=self.frontend,
                                is_final=True,
                            )  # speech: [b, T, d]
                            if self.permute:
                                speech = speech.permute(0, 2, 1)
                            # if speech_lengths > self.batch_size:
                            #     continue

                            olens = 1 + (speech_lengths[0].item() - 3 + 2 * 1) // 2
                            olens = 1 + (olens - 3 + 2 * 1) // 2
                            fake_token_len_i = (olens - 1) // 2 + 1
                            fake_token = [0] * fake_token_len_i
                            fbank_beg_i = len(source_ids)
                            source_ids += fake_token
                            fbank_mask_i += [1] * len(fake_token)

                if badcase_flag:
                    continue

                fbank_beg += [fbank_beg_i + len(input_ids)]
                fake_token_len += [fake_token_len_i]
                source_mask = [-100] * len(source_ids)
                target_out = f"{target_out}<|im_end|>"
                target_ids = self.tokenizer.encode(target_out)
                input_ids += source_ids + target_ids
                labels += source_mask + target_ids
                fbank.append(speech[0, :, :])
                fbank_mask += fbank_mask_i
                fbank_lens.append(speech_lengths)

            if len(input_ids) > self.max_token_length:
                logging.info(
                    f"input_ids > max_token_length: {len(input_ids)}>{self.max_token_length}, {item}"
                )
                badcase_flag = True
            if badcase_flag:
                continue
            input_ids = torch.tensor(input_ids, dtype=torch.int64)  # [: self.max_token_length]
            attention_mask = torch.tensor([1] * len(input_ids), dtype=torch.int32)
            labels = torch.tensor(labels, dtype=torch.int64)  # [: self.max_token_length]

            # fbank = speech[0, :, :]
            # fbank_lens = torch.tensor(fbank_lens, dtype=torch.int32)
            fbank_mask = torch.tensor(fbank_mask, dtype=torch.float32)
            fbank_beg = torch.tensor(fbank_beg, dtype=torch.int32)
            fake_token_len = torch.tensor(fake_token_len, dtype=torch.int32)

            output = {
                "speech": fbank,
                "speech_lengths": fbank_lens,
                "fbank_mask": fbank_mask,
                "fbank_beg": fbank_beg,
                "fake_token_len": fake_token_len,
                "input_ids": input_ids,
                "attention_mask": attention_mask,
                "labels_ids": labels,
            }
            break

        return output

    def collator(self, samples: list = None):

        for idx in range(self.retry):
            badcase_flag = False

            outputs = {}
            for sample in samples:
                if sample is None:
                    continue
                for key in sample.keys():
                    if key not in outputs:
                        outputs[key] = []
                    if isinstance(sample[key], (list, tuple)):
                        outputs[key].extend(sample[key])
                    else:
                        outputs[key].append(sample[key])

            for key, data_list in outputs.items():
                if isinstance(data_list[0], torch.Tensor):
                    if data_list[0].dtype == torch.int64 or data_list[0].dtype == torch.int32:

                        pad_value = self.int_pad_value
                    else:
                        pad_value = self.float_pad_value

                    outputs[key] = torch.nn.utils.rnn.pad_sequence(
                        data_list, batch_first=True, padding_value=pad_value
                    )

            if self.batch_type != "example":
                b, t = outputs["input_ids"].shape
                if b > 1 and b * t > self.batch_size_token_max:
                    logging.info(
                        f"Warning, {idx}th, b*t: {b}*{t}={b * t} > batch_size_sample_max: {self.batch_size_token_max}, drop last data"
                    )
                    samples = samples[:-1]
                    continue

            break

        return outputs

 funasr/datasets/openai_datasets/index_ds.py

@@ -15,11 +15,6 @@

    def __init__(self, path: str, **kwargs):
        super().__init__()
        self.max_source_length = kwargs.get("max_source_length", 2048)
        self.min_source_length = kwargs.get("min_source_length", 0)
        self.max_target_length = kwargs.get("max_target_length", 2048)
        self.min_target_length = kwargs.get("min_target_length", 0)
        self.max_token_length = kwargs.get("max_token_length", 2200)

        is_training = kwargs.get("is_training", True)
        if not (path.endswith(".jsonl") or path.endswith(".json")):

 funasr/download/download_from_hub.py

@@ -56,13 +56,13 @@
                config = OmegaConf.load(cfg["config"])
                kwargs = OmegaConf.merge(config, cfg)
                kwargs["model"] = config["model"]
    elif os.path.exists(os.path.join(model_or_path, "config.yaml")) and os.path.exists(
        os.path.join(model_or_path, "model.pt")
    ):
    elif os.path.exists(os.path.join(model_or_path, "config.yaml")):
        config = OmegaConf.load(os.path.join(model_or_path, "config.yaml"))
        kwargs = OmegaConf.merge(config, kwargs)
        init_param = os.path.join(model_or_path, "model.pb")
        kwargs["init_param"] = init_param
        init_param = os.path.join(model_or_path, "model.pt")
        if "init_param" not in kwargs or not os.path.exists(kwargs["init_param"]):
            kwargs["init_param"] = init_param
            assert os.path.exists(kwargs["init_param"]), "init_param does not exist"
        if os.path.exists(os.path.join(model_or_path, "tokens.txt")):
            kwargs["tokenizer_conf"]["token_list"] = os.path.join(model_or_path, "tokens.txt")
        if os.path.exists(os.path.join(model_or_path, "tokens.json")):
@@ -122,7 +122,7 @@
    ):
        config = OmegaConf.load(os.path.join(model_or_path, "config.yaml"))
        kwargs = OmegaConf.merge(config, kwargs)
        init_param = os.path.join(model_or_path, "model.pb")
        init_param = os.path.join(model_or_path, "model.pt")
        kwargs["init_param"] = init_param
        if os.path.exists(os.path.join(model_or_path, "tokens.txt")):
            kwargs["tokenizer_conf"]["token_list"] = os.path.join(model_or_path, "tokens.txt")

 funasr/models/llm_asr/model.py

@@ -21,6 +21,8 @@
from funasr.train_utils.device_funcs import to_device
import traceback

dtype_map = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}


@tables.register("model_classes", "LLMASR")
class LLMASR(nn.Module):
@@ -394,7 +396,9 @@
            # frontend = model.kwargs.get("frontend")
            audio_encoder_output_size = model.model.encoder_output_size

            audio_encoder = model.model.model.encoder
            audio_encoder = (
                model.model.model.encoder if hasattr(model.model, "model") else model.model.encoder
            )

            # self.frontend = frontend

@@ -405,38 +409,60 @@
            audio_encoder = encoder_class(input_size=input_size, **audio_encoder_conf)
            audio_encoder_output_size = audio_encoder.output_size()
        freeze = audio_encoder_conf.get("freeze", True)
        freeze_layer_num = int(audio_encoder_conf.get("freeze_layer_num", -1))
        # if freeze_layer_num > 0:
        #     freeze_layer_num = range(freeze_layer_num)

        if freeze:
            for name, param in audio_encoder.named_parameters():
                param.requires_grad = False
                if freeze_layer_num > 0:
                    idx = re.search(r"\.\d+\.", name)
                    if idx is not None:
                        beg, end = idx.regs[0]
                        layer_id = int(name[beg + 1 : end - 1])
                        if layer_id < freeze_layer_num:
                            param.requires_grad = False
                    elif "ln_post." not in name:
                        param.requires_grad = False
                else:
                    param.requires_grad = False

            audio_encoder.eval()

        self.audio_encoder = audio_encoder

        # 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")
        from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig

            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
        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_dtype = llm_conf.get("llm_dtype", "fp32")
        self.llm = model.to(dtype_map[self.llm_dtype])
        llm_dim = model.get_input_embeddings().weight.shape[-1]

        # adaptor
        adaptor_class = tables.adaptor_classes.get(audio_adaptor)
        audio_adaptor_conf["encoder_dim"] = audio_encoder_output_size
        audio_adaptor_conf["llm_dim"] = llm_dim
        audio_adaptor = adaptor_class(**audio_adaptor_conf)
        init_param_path = audio_adaptor_conf.get("init_param_path", None)
        if init_param_path is not None:
            src_state = torch.load(init_param_path, map_location="cpu")
            flag = audio_adaptor.load_state_dict(src_state, strict=False)
            logging.info(f"Loading audio_adaptor ckpt: {init_param_path}, status: {flag}")

        self.audio_adaptor = audio_adaptor

@@ -470,11 +496,12 @@

        batch_size, frames, _ = speech.shape

        # audio encoder
        encoder_out, encoder_out_lens = self.audio_encoder(speech.permute(0, 2, 1), speech_lengths)
        with torch.cuda.amp.autocast(enabled=False):
            # audio encoder
            encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)

        # audio_adaptor
        encoder_out, encoder_out_lens = self.audio_adaptor(encoder_out, encoder_out_lens)
            # audio_adaptor
            encoder_out, encoder_out_lens = self.audio_adaptor(encoder_out, encoder_out_lens)

        input_ids[input_ids < 0] = 0
        inputs_embeds = self.llm.model.get_input_embeddings()(input_ids)
@@ -504,12 +531,17 @@
                    batch_idx, :min_len, :
                ]

        labels_ids[labels_ids == -1] = -100

        model_outputs = self.llm(
            inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels_ids
        )
        loss = model_outputs.loss
        with torch.cuda.amp.autocast(
            enabled=True if self.llm_dtype != "fp32" else False, dtype=dtype_map[self.llm_dtype]
        ):
            labels_ids[labels_ids == -1] = -100
            attention_mask[attention_mask < 0] = 0
            model_outputs = self.llm(
                inputs_embeds=inputs_embeds.to(dtype_map[self.llm_dtype]),
                attention_mask=attention_mask,
                labels=labels_ids,
            )
            loss = model_outputs.loss

        stats = {}
        with torch.no_grad():
@@ -531,6 +563,519 @@
            batch_size = int((labels_ids > 0 + 1).sum())
        loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
        return loss, stats, weight

    def encode(self, speech, speech_lengths):
        # audio encoder
        encoder_out, encoder_out_lens = self.audio_encoder(speech.permute(0, 2, 1), speech_lengths)

        return encoder_out, encoder_out_lens

    def data_template(self, data):
        system, user, assistant = [], [], []
        for i, item in enumerate(data):
            role = item["role"]
            content = item["content"]
            if role == "system":
                system.append(content)
            elif role == "user":
                user.append(content)
            elif role == "assistant":
                assistant.append(content)

        system = system * len(user)

        contents = {
            "system": system,
            "user": user,
            "assistant": assistant,
        }

        return contents

    def data_load_speech(self, contents: dict, tokenizer, frontend, meta_data={}, **kwargs):

        system = contents["system"]
        user = contents["user"]
        assistant = contents["assistant"]
        pattern = re.compile(r"(<\|startofspeech\|>.*?<\|endofspeech\|>)")
        input_ids, labels, source_ids, target_ids, fbank, fbank_lens, fbank_mask, fbank_beg = (
            [],
            [],
            [],
            [],
            [],
            [],
            [],
            [],
        )

        for i, (system_prompt, user_prompt, target_out) in enumerate(zip(system, user, assistant)):

            source_input = f"<|im_start|>system\n{system_prompt}<|im_end|>\n<|im_start|>user\n{user_prompt}<|im_end|>\n<|im_start|>assistant\n"

            splits = pattern.split(source_input)
            source_ids_i = []
            fbank_mask_i = []
            fbank_beg_i = []
            fbank_lens_i = []
            # target_ids_i = []
            for k, sub_str in enumerate(splits):
                if not sub_str.startswith("<|startofspeech|>"):
                    sub_token = tokenizer.encode(sub_str)
                    source_ids_i += sub_token
                    fbank_mask_i += [0] * len(sub_token)
                else:
                    sub_str = sub_str.replace("<|startofspeech|>", "").replace(
                        "<|endofspeech|>", ""
                    )
                    if sub_str.startswith("!"):
                        try:
                            time1 = time.perf_counter()
                            data_src = load_audio_text_image_video(sub_str[1:], fs=frontend.fs)
                            time2 = time.perf_counter()
                            meta_data["load_data"] = f"{time2 - time1:0.3f}"
                        except Exception as e:
                            logging.error(f"Loading wav failed! {str(e)}, {traceback.format_exc()}")

                        speech, speech_lengths = extract_fbank(
                            data_src,
                            data_type=kwargs.get("data_type", "sound"),
                            frontend=frontend,
                            is_final=True,
                        )  # speech: [b, T, d]

                        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
                        )

                        if hasattr(frontend, "permute") and not frontend.permute:
                            # if kwargs.get("permute", True):
                            speech = speech.permute(0, 2, 1)

                        if (
                            kwargs.get("dataset_conf", {}).get("audio_encoder_downsample_rate", 1)
                            == 4
                        ):
                            olens = 1 + (speech_lengths[0].item() - 3 + 2 * 1) // 2
                            olens = 1 + (olens - 3 + 2 * 1) // 2
                        elif (
                            kwargs.get("dataset_conf", {}).get("audio_encoder_downsample_rate", 1)
                            == 1
                        ):
                            olens = speech_lengths[0].item()

                        sub_token_len = (olens - 1) // kwargs.get("dataset_conf", {}).get(
                            "audio_adaptor_downsample_rate", 1
                        ) + 1
                        sub_token = [0] * sub_token_len
                        fbank_beg_i = [len(source_ids_i)]
                        source_ids_i += sub_token
                        fbank_mask_i += [1] * len(sub_token)

            source_mask = [-100] * len(source_ids_i)
            target_out = f"{target_out}<|im_end|>"
            target_ids = tokenizer.encode(target_out)
            input_ids += source_ids_i + target_ids
            labels += source_mask + target_ids
            fbank_mask += fbank_mask_i
            fbank_beg.append(fbank_beg_i)

        input_ids = torch.tensor(input_ids, dtype=torch.int64)  # [: self.max_token_length]
        attention_mask = torch.tensor([1] * len(input_ids), dtype=torch.int32)
        labels = torch.tensor(labels, dtype=torch.int64)  # [: self.max_token_length]
        source_ids = torch.tensor(source_ids_i, dtype=torch.int64)
        target_ids = torch.tensor(target_ids, dtype=torch.int64)

        fbank = speech[0, :, :]
        fbank_lens = speech_lengths
        fbank_mask = torch.tensor(fbank_mask, dtype=torch.float32)
        fbank_beg = torch.tensor(fbank_beg, dtype=torch.int32)

        output = {
            "speech": fbank[None, :, :],
            "speech_lengths": fbank_lens[:, None],
            "fbank_mask": fbank_mask[None, :],
            "fbank_beg": fbank_beg[None,],
            "input_ids": input_ids[None, :],
            "attention_mask": attention_mask[None, :],
            "labels_ids": labels[None, :],
            "source_ids": source_ids[None, :],
            "target_ids": target_ids[None, :],
        }

        return output

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

        meta_data = {}
        prompt = kwargs.get("prompt", None)

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

        contents = self.data_template(data_in[0])
        output = self.data_load_speech(contents, tokenizer, frontend, meta_data=meta_data, **kwargs)
        batch = to_device(output, kwargs["device"])

        # audio encoder
        speech = batch["speech"]
        speech_lengths = batch["speech_lengths"][:, 0]
        # fp16
        if kwargs.get("fp16", False):
            speech = speech.to(torch.float16)
        elif kwargs.get("bf16", False):
            speech = speech.to(torch.bfloat16)
        # audio encoder
        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)

        # audio_adaptor
        encoder_out, encoder_out_lens = self.audio_adaptor(encoder_out, encoder_out_lens)

        input_ids = batch["input_ids"]
        source_ids = batch["source_ids"]
        if not kwargs.get("tearchforing", False):
            input_ids = source_ids
        input_ids[input_ids < 0] = 0
        inputs_embeds = self.llm.model.get_input_embeddings()(input_ids)

        batch_size, token_num, dims = inputs_embeds.shape
        fbank_beg = batch["fbank_beg"]
        for batch_idx in range(batch_size):

            min_len = encoder_out_lens[batch_idx].item()
            fbank_beg_idx = fbank_beg[batch_idx]
            inputs_embeds[batch_idx, fbank_beg_idx : fbank_beg_idx + min_len, :] = encoder_out[
                batch_idx, :min_len, :
            ]

        llm_dtype = kwargs.get("llm_dtype", "fp32")
        if llm_dtype == "fp32":
            llm_dtype = "fp16" if kwargs.get("fp16", False) else llm_dtype
            llm_dtype = "bf16" if kwargs.get("bf16", False) else llm_dtype

        with torch.cuda.amp.autocast(
            enabled=True if llm_dtype != "fp32" else False, dtype=dtype_map[llm_dtype]
        ):
            label = contents["assistant"][0]
            self.llm = self.llm.to(dtype_map[llm_dtype])
            inputs_embeds = inputs_embeds.to(dtype_map[llm_dtype])

            if not kwargs.get("tearchforing", False):

                generated_ids = self.llm.generate(
                    inputs_embeds=inputs_embeds, max_new_tokens=kwargs.get("max_length", 512)
                )
                # generated_ids = [
                #     output_ids[len(input_id) :]
                #     for input_id, output_ids in zip(input_ids, generated_ids)
                # ]
                response = tokenizer.batch_decode(
                    generated_ids, skip_special_tokens=kwargs.get("skip_special_tokens", True)
                )[0]

                loss = None
            else:

                labels_ids = batch["labels_ids"]
                labels_ids[labels_ids == -1] = -100
                attention_mask = batch.get("attention_mask", None)
                # attention_mask = attention_mask.to(dtype_map[llm_dtype])
                model_outputs = self.llm(
                    inputs_embeds=inputs_embeds, attention_mask=attention_mask, labels=labels_ids
                )

                preds = torch.argmax(model_outputs.logits, -1)[:, source_ids.shape[1] :]
                response = tokenizer.batch_decode(
                    preds,
                    add_special_tokens=False,
                    skip_special_tokens=kwargs.get("skip_special_tokens", True),
                )[0]
                loss = model_outputs.loss.item()

        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 = []
        response_clean = re.sub("[^\w\s\u3000\u4e00-\u9fff]+", "", response)
        result_i = {"key": key[0], "text": response, "text_tn": response_clean, "label": label}
        if loss is not None:
            result_i["loss"] = loss
        results.append(result_i)

        if ibest_writer is not None:
            ibest_writer["text"][key[0]] = response
            ibest_writer["label"][key[0]] = label
            ibest_writer["text_tn"][key[0]] = response_clean

        return results, meta_data


@tables.register("model_classes", "LLMASR3")
class LLMASR3(LLMASR2):
    """ """

    def __init__(
        self,
        *args,
        **kwargs,
    ):

        super().__init__(*args, **kwargs)

    def encode(self, speech, speech_lengths):
        # audio encoder
        encoder_out, encoder_out_lens = self.audio_encoder(speech, speech_lengths)
        return encoder_out, encoder_out_lens


@tables.register("model_classes", "LLMASR4")
class LLMASR4(nn.Module):
    """ """

    def __init__(
        self,
        specaug: str = None,
        specaug_conf: dict = None,
        normalize: str = None,
        normalize_conf: dict = None,
        audio_encoder: str = None,
        audio_encoder_conf: dict = None,
        audio_adaptor: str = None,
        audio_adaptor_conf: dict = None,
        decoder: str = None,
        decoder_conf: dict = None,
        ctc: str = None,
        ctc_conf: dict = None,
        ctc_weight: float = 0.5,
        llm: str = None,
        llm_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,
        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__()

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

            model = AutoModel(model=audio_encoder, model_revision="master")
            # frontend = model.kwargs.get("frontend")
            audio_encoder_output_size = model.model.encoder_output_size

            audio_encoder = (
                model.model.model.encoder if hasattr(model.model, "model") else model.model.encoder
            )

            # self.frontend = frontend

        elif hub == "hf":
            pass
        else:
            encoder_class = tables.encoder_classes.get(audio_encoder)
            audio_encoder = encoder_class(input_size=input_size, **audio_encoder_conf)
            audio_encoder_output_size = audio_encoder.output_size()
        freeze = audio_encoder_conf.get("freeze", True)
        freeze_layer_num = int(audio_encoder_conf.get("freeze_layer_num", -1))
        # if freeze_layer_num > 0:
        #     freeze_layer_num = range(freeze_layer_num)

        if freeze:
            for name, param in audio_encoder.named_parameters():
                if freeze_layer_num > 0:
                    idx = re.search(r"\.\d+\.", name)
                    if idx is not None:
                        beg, end = idx.regs[0]
                        layer_id = int(name[beg + 1 : end - 1])
                        if layer_id < freeze_layer_num:
                            param.requires_grad = False
                    elif "ln_post." not in name:
                        param.requires_grad = False
                else:
                    param.requires_grad = False

            audio_encoder.eval()

        self.audio_encoder = audio_encoder

        # llm
        self.llm = None

        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_dtype = llm_conf.get("llm_dtype", "fp32")
        self.llm = model.to(dtype_map[self.llm_dtype])
        llm_dim = model.get_input_embeddings().weight.shape[-1]

        # adaptor
        adaptor_class = tables.adaptor_classes.get(audio_adaptor)
        audio_adaptor_conf["encoder_dim"] = audio_encoder_output_size
        audio_adaptor_conf["llm_dim"] = llm_dim
        audio_adaptor = adaptor_class(**audio_adaptor_conf)
        init_param_path = audio_adaptor_conf.get("init_param_path", None)
        if init_param_path is not None:
            src_state = torch.load(init_param_path, map_location="cpu")
            flag = audio_adaptor.load_state_dict(src_state, strict=False)
            logging.info(f"Loading audio_adaptor ckpt: {init_param_path}, status: {flag}")

        self.audio_adaptor = audio_adaptor

        self.error_calculator = None

        self.length_normalized_loss = length_normalized_loss
        self.beam_search = None

    def forward(
        self,
        speech: torch.Tensor,
        speech_lengths: torch.Tensor,
        input_ids: torch.Tensor,
        attention_mask: torch.Tensor,
        labels_ids: torch.Tensor,
        fbank_beg: torch.Tensor,
        fbank_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,)
        """
        import pdb

        pdb.set_trace()
        if len(speech_lengths.size()) > 1:
            speech_lengths = speech_lengths[:, 0]

        batch_size_speech, frames, _ = speech.shape
        batch_size, token_num = input_ids.shape

        with torch.cuda.amp.autocast(enabled=False):
            # audio encoder
            encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)

            # audio_adaptor
            encoder_out, encoder_out_lens = self.audio_adaptor(encoder_out, encoder_out_lens)

        input_ids[input_ids < 0] = 0
        inputs_embeds = self.llm.model.get_input_embeddings()(input_ids)

        batch_size, token_num, dims = inputs_embeds.shape
        fake_token_len = kwargs.get("fake_token_len")
        fake_token_len[fake_token_len < 0] = 0
        fbank_beg[fbank_beg < 0] = 0
        speech_idx = 0
        for batch_idx in range(batch_size):

            for turn_id in range(fbank_beg.shape[1]):
                fbank_beg_idx = fbank_beg[batch_idx, turn_id].item()
                if fbank_beg_idx > 0:
                    speech_token_len = fake_token_len[batch_idx, turn_id]
                    speech_token = encoder_out[speech_idx, :speech_token_len, :]

                    try:
                        inputs_embeds[
                            batch_idx, fbank_beg_idx : fbank_beg_idx + speech_token_len, :
                        ] = speech_token
                    except Exception as e:
                        logging.error(f"{str(e)}, {traceback.format_exc()}")
                        logging.info(
                            f"batch_idx: {batch_idx}, inputs_embeds: {inputs_embeds.shape}, fbank_beg_idx: {fbank_beg_idx}, speech_token_len: {speech_token_len}, encoder_out: {encoder_out.shape}, encoder_out_lens: {encoder_out_lens[speech_idx].item()}"
                        )
                        speech_token_len = encoder_out_lens[speech_idx].item()
                        speech_token = encoder_out[speech_idx, turn_id, :speech_token_len, :]
                        inputs_embeds[
                            batch_idx, fbank_beg_idx : fbank_beg_idx + speech_token_len, :
                        ] = speech_token

                    speech_idx += 1

        with torch.cuda.amp.autocast(
            enabled=True if self.llm_dtype != "fp32" else False, dtype=dtype_map[self.llm_dtype]
        ):
            labels_ids[labels_ids == -1] = -100
            attention_mask[attention_mask < 0] = 0
            model_outputs = self.llm(
                inputs_embeds=inputs_embeds.to(dtype_map[self.llm_dtype]),
                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())
        stats["batch_size"] = batch_size
        stats["batch_size_speech"] = batch_size_speech
        stats["batch_size_x_frames"] = frames * batch_size_speech
        stats["batch_size_real_frames"] = speech_lengths.sum().item()
        stats["padding_frames"] = stats["batch_size_x_frames"] - stats["batch_size_real_frames"]
        stats["batch_size_x_tokens"] = token_num * batch_size
        stats["batch_size_real_tokens"] = attention_mask.sum().item()
        stats["padding_tokens"] = stats["batch_size_x_tokens"] - stats["batch_size_real_tokens"]

        # force_gatherable: to-device and to-tensor if scalar for DataParallel
        if self.length_normalized_loss:
            batch_size = int((labels_ids > 0 + 1).sum())
        loss, stats, weight = force_gatherable((loss, stats, batch_size), loss.device)
        return loss, stats, weight

    def encode(self, speech, speech_lengths):
        # audio encoder
        encoder_out, encoder_out_lens = self.audio_encoder(speech.permute(0, 2, 1), speech_lengths)

        return encoder_out, encoder_out_lens

    def data_template(self, data):
        system, user, assistant = [], [], []
@@ -685,11 +1230,10 @@
        # fp16
        if kwargs.get("fp16", False):
            speech = speech.to(torch.float16)
            encoder_out_lens = encoder_out_lens.to(torch.float16)
        elif kwargs.get("bf16", False):
            speech = speech.to(torch.bfloat16)
            encoder_out_lens = encoder_out_lens.to(torch.bfloat16)
        encoder_out, encoder_out_lens = self.audio_encoder(speech.permute(0, 2, 1), speech_lengths)
        # audio encoder
        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)

        # audio_adaptor
        encoder_out, encoder_out_lens = self.audio_adaptor(encoder_out, encoder_out_lens)
@@ -712,11 +1256,16 @@
            ]

        llm_dtype = kwargs.get("llm_dtype", "fp32")
        dtype_map = {"bf16": torch.bfloat16, "fp16": torch.float16, "fp32": torch.float32}
        with torch.cuda.amp.autocast(dtype=dtype_map[llm_dtype]):
        if llm_dtype == "fp32":
            llm_dtype = "fp16" if kwargs.get("fp16", False) else llm_dtype
            llm_dtype = "bf16" if kwargs.get("bf16", False) else llm_dtype

        with torch.cuda.amp.autocast(
            enabled=True if llm_dtype != "fp32" else False, dtype=dtype_map[llm_dtype]
        ):
            label = contents["assistant"][0]
            # self.llm = self.llm.to(dtype_map[llm_dtype])
            # inputs_embeds = inputs_embeds.to(dtype_map[llm_dtype])
            self.llm = self.llm.to(dtype_map[llm_dtype])
            inputs_embeds = inputs_embeds.to(dtype_map[llm_dtype])

            if not kwargs.get("tearchforing", False):


 funasr/models/paraformer/cif_predictor.py

@@ -494,6 +494,8 @@
        token_num_floor = torch.floor(token_num)



        return hidden, alphas, token_num_floor





@torch.jit.script

def cif_v1_export(hidden, alphas, threshold: float):

    device = hidden.device

@@ -504,7 +506,7 @@
    frames = torch.zeros(batch_size, len_time, hidden_size, dtype=dtype, device=device)

    fires = torch.zeros(batch_size, len_time, dtype=dtype, device=device)



    prefix_sum = torch.cumsum(alphas, dim=1, dtype=torch.float64).to(torch.float32) # cumsum precision degradation cause wrong result in extreme 
    prefix_sum = torch.cumsum(alphas, dim=1)

    prefix_sum_floor = torch.floor(prefix_sum)

    dislocation_prefix_sum = torch.roll(prefix_sum, 1, dims=1)

    dislocation_prefix_sum_floor = torch.floor(dislocation_prefix_sum)

@@ -516,9 +518,7 @@
    fires[fire_idxs] = 1

    fires = fires + prefix_sum - prefix_sum_floor



    prefix_sum_hidden = torch.cumsum(

        alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1

    )

    prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1)



    frames = prefix_sum_hidden[fire_idxs]

    shift_frames = torch.roll(frames, 1, dims=0)

@@ -530,25 +530,21 @@
    shift_frames[shift_batch_idxs] = 0



    remains = fires - torch.floor(fires)

    remain_frames = (

        remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]

    )

    remain_frames = remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]



    shift_remain_frames = torch.roll(remain_frames, 1, dims=0)

    shift_remain_frames[shift_batch_idxs] = 0



    frames = frames - shift_frames + shift_remain_frames - remain_frames



    max_label_len = alphas.sum(dim=-1)

    max_label_len = torch.floor(max_label_len).max().to(dtype=torch.int64)

    max_label_len = batch_len.max()



    frame_fires = torch.zeros(

        batch_size, max_label_len, hidden_size, dtype=dtype, device=device

    )

    frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)

    indices = torch.arange(max_label_len, device=device).expand(batch_size, -1)

    frame_fires_idxs = indices < batch_len.unsqueeze(1)

    frame_fires[frame_fires_idxs] = frames

    return frame_fires, fires





@torch.jit.script

def cif_export(hidden, alphas, threshold: float):

@@ -671,7 +667,7 @@


    fires = torch.zeros(batch_size, len_time, dtype=dtype, device=device)



    prefix_sum = torch.cumsum(alphas, dim=1, dtype=torch.float64).to(torch.float32) # cumsum precision degradation cause wrong result in extreme 
    prefix_sum = torch.cumsum(alphas, dim=1)

    prefix_sum_floor = torch.floor(prefix_sum)

    dislocation_prefix_sum = torch.roll(prefix_sum, 1, dims=1)

    dislocation_prefix_sum_floor = torch.floor(dislocation_prefix_sum)

@@ -693,11 +689,8 @@
    device = hidden.device

    dtype = hidden.dtype

    batch_size, len_time, hidden_size = hidden.size()

    frames = torch.zeros(batch_size, len_time, hidden_size,

                         dtype=dtype, device=device)

    prefix_sum_hidden = torch.cumsum(

        alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1

    )

    frames = torch.zeros(batch_size, len_time, hidden_size, dtype=dtype, device=device)

    prefix_sum_hidden = torch.cumsum(alphas.unsqueeze(-1).tile((1, 1, hidden_size)) * hidden, dim=1)



    frames = prefix_sum_hidden[fire_idxs]

    shift_frames = torch.roll(frames, 1, dims=0)

@@ -709,21 +702,16 @@
    shift_frames[shift_batch_idxs] = 0



    remains = fires - torch.floor(fires)

    remain_frames = (

        remains[fire_idxs].unsqueeze(-1).tile((1,

                                               hidden_size)) * hidden[fire_idxs]

    )

    remain_frames = remains[fire_idxs].unsqueeze(-1).tile((1, hidden_size)) * hidden[fire_idxs]



    shift_remain_frames = torch.roll(remain_frames, 1, dims=0)

    shift_remain_frames[shift_batch_idxs] = 0



    frames = frames - shift_frames + shift_remain_frames - remain_frames



    max_label_len = torch.round(alphas.sum(-1)).int().max() # torch.round to calculate the max length

    max_label_len = batch_len.max()



    frame_fires = torch.zeros(

        batch_size, max_label_len, hidden_size, dtype=dtype, device=device

    )

    frame_fires = torch.zeros(batch_size, max_label_len, hidden_size, dtype=dtype, device=device)

    indices = torch.arange(max_label_len, device=device).expand(batch_size, -1)

    frame_fires_idxs = indices < batch_len.unsqueeze(1)

    frame_fires[frame_fires_idxs] = frames


 funasr/models/sense_voice/model.py

@@ -16,6 +16,7 @@
from . import whisper_lib as whisper
from funasr.utils.load_utils import load_audio_text_image_video, extract_fbank
from funasr.utils.datadir_writer import DatadirWriter
from funasr.models.ctc.ctc import CTC

from funasr.register import tables

@@ -1035,6 +1036,7 @@
        self.length_normalized_loss = length_normalized_loss
        self.beam_search = None
        self.activation_checkpoint = kwargs.get("activation_checkpoint", False)
        self.encoder_output_size = encoder_output_size

    def forward(
        self,
@@ -1256,7 +1258,7 @@
        if isinstance(task, str):
            task = [task]
        task = "".join([f"<|{x}|>" for x in task])
        

        sos = kwargs.get("model_conf").get("sos")
        if isinstance(sos, str):
            initial_prompt = kwargs.get("initial_prompt", f"<|startoftranscript|>{task}")
@@ -1270,7 +1272,9 @@
            language = DecodingOptions.get("language", None)
            language = None if language == "auto" else language
            initial_prompt = kwargs.get("initial_prompt", f"{task}")
            initial_prompt_lid = f"{initial_prompt}<|{language}|>" if language is not None else initial_prompt
            initial_prompt_lid = (
                f"{initial_prompt}<|{language}|>" if language is not None else initial_prompt
            )
            initial_prompt_lid_int = tokenizer.encode(initial_prompt_lid, allowed_special="all")
            sos_int = [sos] + initial_prompt_lid_int
        eos = kwargs.get("model_conf").get("eos")
@@ -1303,9 +1307,7 @@
        )
        self.beam_search.event_score_ga = DecodingOptions.get("gain_tokens_score", [1, 1, 1, 1])

        encoder_out, encoder_out_lens = self.encode(
            speech[None, :, :], speech_lengths
        )
        encoder_out, encoder_out_lens = self.encode(speech[None, :, :], speech_lengths)

        if text_token_int is not None:
            i = 0
@@ -1384,3 +1386,279 @@
                    ibest_writer["text"][key[i]] = text

        return results, meta_data


from funasr.models.paraformer.search import Hypothesis
from funasr.utils import postprocess_utils


@tables.register("model_classes", "SenseVoiceSANMCTC")
class SenseVoiceSANMCTC(nn.Module):
    """CTC-attention hybrid Encoder-Decoder model"""

    def __init__(
        self,
        specaug: str = None,
        specaug_conf: dict = None,
        normalize: str = None,
        normalize_conf: dict = None,
        encoder: str = None,
        encoder_conf: dict = None,
        ctc_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,
        length_normalized_loss: bool = False,
        **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)
        encoder_class = tables.encoder_classes.get(encoder)
        encoder = encoder_class(input_size=input_size, **encoder_conf)
        encoder_output_size = encoder.output_size()

        if ctc_conf is None:
            ctc_conf = {}
        ctc = CTC(odim=vocab_size, encoder_output_size=encoder_output_size, **ctc_conf)

        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.error_calculator = None

        self.ctc = ctc

        self.length_normalized_loss = length_normalized_loss
        self.encoder_output_size = encoder_output_size

        self.lid_dict = {"zh": 3, "en": 4, "yue": 7, "ja": 11, "ko": 12, "nospeech": 13}
        self.textnorm_dict = {"withtextnorm": 14, "wotextnorm": 15}
        self.embed = torch.nn.Embedding(8 + len(self.lid_dict) + len(self.textnorm_dict), 560)

    def forward(
        self,
        speech: torch.Tensor,
        speech_lengths: torch.Tensor,
        text: torch.Tensor,
        text_lengths: torch.Tensor,
        **kwargs,
    ):
        """Encoder + Decoder + Calc loss
        Args:
                speech: (Batch, Length, ...)
                speech_lengths: (Batch, )
                text: (Batch, Length)
                text_lengths: (Batch,)
        """
        # import pdb;
        # pdb.set_trace()
        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]

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

        loss_ctc, cer_ctc = None, None
        stats = dict()

        loss_ctc, cer_ctc = self._calc_ctc_loss(encoder_out, encoder_out_lens, text, text_lengths)

        loss = loss_ctc

        # Collect total loss stats
        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,
    ):
        """Frontend + Encoder. Note that this method is used by asr_inference.py
        Args:
                speech: (Batch, Length, ...)
                speech_lengths: (Batch, )
                ind: int
        """

        # Data augmentation
        if self.specaug is not None and self.training:
            speech, speech_lengths = self.specaug(speech, speech_lengths)

        # Normalization for feature: e.g. Global-CMVN, Utterance-CMVN
        if self.normalize is not None:
            speech, speech_lengths = self.normalize(speech, speech_lengths)

        # Forward encoder
        # feats: (Batch, Length, Dim)
        # -> encoder_out: (Batch, Length2, Dim2)
        encoder_out, encoder_out_lens = self.encoder(speech, speech_lengths)

        return encoder_out, encoder_out_lens

    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)

        # 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,
    ):

        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=tokenizer,
            )
            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"])

        language = kwargs.get("language", None)
        if language is not None:
            language_query = self.embed(
                torch.LongTensor(
                    [[self.lid_dict[language] if language in self.lid_dict else 0]]
                ).to(speech.device)
            ).repeat(speech.size(0), 1, 1)
        else:
            language_query = self.embed(torch.LongTensor([[0]]).to(speech.device)).repeat(
                speech.size(0), 1, 1
            )
        textnorm = kwargs.get("text_norm", "wotextnorm")
        textnorm_query = self.embed(
            torch.LongTensor([[self.textnorm_dict[textnorm]]]).to(speech.device)
        ).repeat(speech.size(0), 1, 1)
        speech = torch.cat((textnorm_query, speech), dim=1)
        speech_lengths += 1

        event_emo_query = self.embed(torch.LongTensor([[1, 2]]).to(speech.device)).repeat(
            speech.size(0), 1, 1
        )
        input_query = torch.cat((language_query, event_emo_query), dim=1)
        speech = torch.cat((input_query, speech), dim=1)
        speech_lengths += 3

        # Encoder
        encoder_out, encoder_out_lens = self.encode(speech, speech_lengths)
        if isinstance(encoder_out, tuple):
            encoder_out = encoder_out[0]

        # c. Passed the encoder result and the beam search
        ctc_logits = self.ctc.log_softmax(encoder_out)

        results = []
        b, n, d = encoder_out.size()
        if isinstance(key[0], (list, tuple)):
            key = key[0]
        if len(key) < b:
            key = key * b
        for i in range(b):
            x = ctc_logits[i, : encoder_out_lens[i], :]
            yseq = x.argmax(dim=-1)
            yseq = torch.unique_consecutive(yseq, dim=-1)
            yseq = torch.tensor([self.sos] + yseq.tolist() + [self.eos], device=yseq.device)
            nbest_hyps = [Hypothesis(yseq=yseq)]

            for nbest_idx, hyp in enumerate(nbest_hyps):
                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"{nbest_idx + 1}best_recog"]

                # remove sos/eos and get results
                last_pos = -1
                if isinstance(hyp.yseq, list):
                    token_int = hyp.yseq[1:last_pos]
                else:
                    token_int = hyp.yseq[1:last_pos].tolist()

                # remove blank symbol id, which is assumed to be 0
                token_int = list(
                    filter(
                        lambda x: x != self.eos and x != self.sos and x != self.blank_id, token_int
                    )
                )

                # Change integer-ids to tokens
                text = tokenizer.decode(token_int)

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

                if ibest_writer is not None:
                    ibest_writer["token"][key[i]] = " ".join(token)
                    ibest_writer["text"][key[i]] = text_postprocessed

        return results, meta_data

 funasr/train_utils/load_pretrained_model.py

@@ -10,36 +10,6 @@
import pdb


def filter_state_dict(
    dst_state: Dict[str, Union[float, torch.Tensor]],
    src_state: Dict[str, Union[float, torch.Tensor]],
):
    """Filter name, size mismatch instances between dicts.

    Args:
            dst_state: reference state dict for filtering
            src_state: target state dict for filtering

    """
    match_state = {}
    for key, value in src_state.items():
        if key in dst_state and (dst_state[key].size() == src_state[key].size()):
            match_state[key] = value
        else:
            if key not in dst_state:
                logging.warning(
                    f"Filter out {key} from pretrained dict"
                    + " because of name not found in target dict"
                )
            else:
                logging.warning(
                    f"Filter out {key} from pretrained dict"
                    + " because of size mismatch"
                    + f"({dst_state[key].size()}-{src_state[key].size()})"
                )
    return match_state


def load_pretrained_model(
    path: str,
    model: torch.nn.Module,
@@ -62,7 +32,7 @@
    obj = model
    dst_state = obj.state_dict()

    print(f"ckpt: {path}")
    logging.info(f"ckpt: {path}")

    if oss_bucket is None:
        src_state = torch.load(path, map_location=map_location)
@@ -77,8 +47,24 @@
    if isinstance(scope_map, str):
        scope_map = scope_map.split(",")
    scope_map += ["module.", "None"]
    logging.info(f"scope_map: {scope_map}")

    if excludes is not None:
        if isinstance(excludes, str):
            excludes = excludes.split(",")

    logging.info(f"excludes: {excludes}")

    for k in dst_state.keys():
        excludes_flag = False
        if excludes is not None:
            for k_ex in excludes:
                if k.startswith(k_ex):
                    logging.info(f"key: {k} matching: {k_ex}, excluded")
                    excludes_flag = True
                    break
        if excludes_flag:
            continue

        k_src = k

@@ -92,25 +78,25 @@
                if dst_prefix == "" and (src_prefix + k) in src_state.keys():
                    k_src = src_prefix + k
                    if not k_src.startswith("module."):
                        print(f"init param, map: {k} from {k_src} in ckpt")
                        logging.info(f"init param, map: {k} from {k_src} in ckpt")
                elif (
                    k.startswith(dst_prefix)
                    and k.replace(dst_prefix, src_prefix, 1) in src_state.keys()
                ):
                    k_src = k.replace(dst_prefix, src_prefix, 1)
                    if not k_src.startswith("module."):
                        print(f"init param, map: {k} from {k_src} in ckpt")
                        logging.info(f"init param, map: {k} from {k_src} in ckpt")

        if k_src in src_state.keys():
            if ignore_init_mismatch and dst_state[k].shape != src_state[k_src].shape:
                print(
                logging.info(
                    f"ignore_init_mismatch:{ignore_init_mismatch}, dst: {k, dst_state[k].shape}, src: {k_src, src_state[k_src].shape}"
                )
            else:
                dst_state[k] = src_state[k_src]

        else:
            print(f"Warning, miss key in ckpt: {k}, mapped: {k_src}")
            print(f"Warning, miss key in ckpt: {k}, {path}")

    flag = obj.load_state_dict(dst_state, strict=True)
    # print(flag)
    logging.info(f"Loading ckpt: {path}, status: {flag}")

 funasr/train_utils/trainer_ds.py

@@ -29,9 +29,10 @@
        with torch.cuda.amp.autocast(enabled=True, dtype=dtype, cache_enabled=False):
            yield
    else:
        if dtype == torch.float16:
            with autocast(enabled=True):
                yield
        if dtype == torch.float16 or dtype == torch.bfloat16:
            yield
            # with autocast(enabled=True, dtype=dtype):
            #     yield
        else:
            yield

@@ -60,6 +61,7 @@
        use_ddp: bool = False,
        use_fsdp: bool = False,
        use_fp16: bool = False,
        use_bf16: bool = False,
        use_deepspeed: bool = False,
        output_dir: str = "./",
        **kwargs,
@@ -78,7 +80,7 @@
                      output_dir (str): The directory where model checkpoints will be saved. Default is './'.
                      resume (str, optional): The file path to a checkpoint to resume training from.
        """
        self.rank = kwargs.get("rank", 0)
        self.rank = rank
        self.local_rank = local_rank
        self.world_size = world_size
        self.use_ddp = use_ddp
@@ -98,8 +100,11 @@
        self.batch_total = 0
        self.dtype = torch.float32
        self.use_fp16 = use_fp16
        self.use_bf16 = use_bf16
        if self.use_fp16:
            self.dtype = torch.float16
        if self.use_bf16:
            self.dtype = torch.bfloat16
        self.save_checkpoint_interval = kwargs.get("save_checkpoint_interval", 5000)
        self.validate_interval = kwargs.get("validate_interval", 5000)
        self.keep_nbest_models = kwargs.get("keep_nbest_models", 500)
@@ -147,6 +152,16 @@

        self.use_deepspeed = use_deepspeed
        self.deepspeed_config = kwargs.get("deepspeed_config", "")
        excludes = kwargs.get("excludes", None)
        if excludes is not None:
            if isinstance(excludes, str):
                excludes = excludes.split(",")
        self.excludes = excludes
        effective_save_name_excludes = kwargs.get("effective_save_name_excludes", None)
        if effective_save_name_excludes is not None:
            if isinstance(effective_save_name_excludes, str):
                effective_save_name_excludes = effective_save_name_excludes.split(",")
        self.effective_save_name_excludes = effective_save_name_excludes

    def save_checkpoint(
        self,
@@ -277,11 +292,12 @@
        elif self.use_fsdp:
            pass
        elif self.rank == 0:
            logging.info(f"Save checkpoint: {epoch}, rank: {self.local_rank}\n")
            logging.info(
                f"Save checkpoint: {epoch}, rank: {self.rank}, local_rank: {self.local_rank}\n"
            )
            # self.step_or_epoch += 1
            state = {
                "epoch": epoch,
                "state_dict": model.state_dict(),
                "optimizer": optim.state_dict(),
                "scheduler": scheduler.state_dict(),
                "saved_ckpts": self.saved_ckpts,
@@ -299,7 +315,24 @@
            }
            step = step_in_epoch
            if hasattr(model, "module"):
                state["state_dict"] = model.module.state_dict()
                state_dict = model.module.state_dict()
            else:
                state_dict = model.state_dict()

            if self.effective_save_name_excludes is not None:
                logging.info(f"effective_save_name_excludes: {self.effective_save_name_excludes}")
                dst_state_dict = {}
                for k in state_dict.keys():
                    for k_ex in self.effective_save_name_excludes:
                        k_tmp = k.replace("module.", "")
                        if k.startswith(k_ex):
                            logging.info(f"key: {k} matching: {k_ex}, not save it")
                            break
                    else:
                        dst_state_dict[k] = state_dict[k]
                state["state_dict"] = dst_state_dict
            else:
                state["state_dict"] = state_dict

            if scaler:
                state["scaler_state"] = scaler.state_dict()
@@ -440,6 +473,16 @@
                    src_state = checkpoint["state_dict"]
                    dst_state = model.state_dict()
                    for k in dst_state.keys():
                        excludes_flag = False
                        if self.excludes is not None:
                            for k_ex in self.excludes:
                                k_tmp = k.replace("module.", "")
                                if k_tmp.startswith(k_ex):
                                    logging.info(f"key: {k} matching: {k_ex}, excluded")
                                    excludes_flag = True
                                    break
                        if excludes_flag:
                            continue
                        if not k.startswith("module.") and "module." + k in src_state.keys():
                            k_ddp = "module." + k
                        elif k.startswith("module.") and "module." + k not in src_state.keys():
@@ -640,7 +683,7 @@
            scaled_loss = model.backward(loss)
        else:
            loss = loss / self.accum_grad
            if self.use_fp16:
            if self.use_fp16 or self.use_bf16:
                scaler.scale(loss).backward()
            else:
                loss.backward()
@@ -668,7 +711,7 @@
                # Execute an optimization step (update model parameters)
                if self.use_ddp or self.use_fsdp:
                    dist.barrier()
                if self.use_fp16:
                if self.use_fp16 or self.use_bf16:
                    scaler.step(optim)
                    scaler.update()
                else: