decoding

游雁

2024-06-13 d72df6cd2f1914a749db0778233f3a66f447894a

 funasr/models/llm_asr/model.py

@@ -449,9 +449,9 @@
            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")
        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)
@@ -496,11 +496,14 @@

        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.audio_encoder(
                speech.permute(0, 2, 1), 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)
@@ -536,7 +539,9 @@
            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
                inputs_embeds=inputs_embeds.to(dtype_map[self.llm_dtype]),
                attention_mask=attention_mask,
                labels=labels_ids,
            )
            loss = model_outputs.loss