decoding

游雁

2024-06-13 6ca0b838d48106030984eacf204e8f1f2f05985b

 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):
@@ -407,38 +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
                    else:
                        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 = model
        llm_dim = model.get_input_embeddings().weight.shape[-1]
        self.llm_dtype = llm_conf.get("llm_dtype", "fp32")

        # 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

@@ -506,12 +530,15 @@
                    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, attention_mask=attention_mask, labels=labels_ids
            )
            loss = model_outputs.loss

        stats = {}
        with torch.no_grad():
@@ -687,10 +714,8 @@
        # 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_adaptor
@@ -714,12 +739,17 @@
            ]

        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])
            attention_mask = attention_mask.to(dtype_map[llm_dtype])

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

                generated_ids = self.llm.generate(
@@ -739,6 +769,7 @@
                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
                )