fix "can not find model issue when running libtorch runtime" (#2504)

Yu Cao

2025-10-01 c4ac64fd5d24bb3fc8ccc441d36a07c83c8b9015

 funasr/models/seaco_paraformer/export_meta.py

@@ -9,173 +9,194 @@


class ContextualEmbedderExport(torch.nn.Module):
    def __init__(self,
                 model,
                 max_seq_len=512,
                 feats_dim=560,
                 **kwargs,):
    def __init__(
        self,
        model,
        max_seq_len=512,
        feats_dim=560,
        **kwargs,
    ):
        super().__init__()
        self.embedding = model.decoder.embed # model.bias_embed
        self.embedding = model.decoder.embed  # model.bias_embed
        model.bias_encoder.batch_first = False
        self.bias_encoder = model.bias_encoder
    

    def forward(self, hotword):
        hotword = self.embedding(hotword).transpose(0, 1) # batch second
        hotword = self.embedding(hotword).transpose(0, 1)  # batch second
        hw_embed, (_, _) = self.bias_encoder(hotword)
        return hw_embed
    

    def export_dummy_inputs(self):
        hotword = torch.tensor([
                                [10, 11, 12, 13, 14, 10, 11, 12, 13, 14], 
                                [100, 101, 0, 0, 0, 0, 0, 0, 0, 0],
                                [1, 0, 0, 0, 0, 0, 0, 0, 0, 0],
                                [10, 11, 12, 13, 14, 10, 11, 12, 13, 14], 
                                [100, 101, 0, 0, 0, 0, 0, 0, 0, 0],
                                [1, 0, 0, 0, 0, 0, 0, 0, 0, 0],
                               ], 
                                dtype=torch.int32)
        hotword = torch.tensor(
            [
                [10, 11, 12, 13, 14, 10, 11, 12, 13, 14],
                [100, 101, 0, 0, 0, 0, 0, 0, 0, 0],
                [1, 0, 0, 0, 0, 0, 0, 0, 0, 0],
                [10, 11, 12, 13, 14, 10, 11, 12, 13, 14],
                [100, 101, 0, 0, 0, 0, 0, 0, 0, 0],
                [1, 0, 0, 0, 0, 0, 0, 0, 0, 0],
            ],
            dtype=torch.int32,
        )
        # hotword_length = torch.tensor([10, 2, 1], dtype=torch.int32)
        return (hotword)
        return hotword

    def export_input_names(self):
        return ['hotword']
        return ["hotword"]

    def export_output_names(self):
        return ['hw_embed']
        return ["hw_embed"]

    def export_dynamic_axes(self):
        return {
            'hotword': {
                0: 'num_hotwords',
            "hotword": {
                0: "num_hotwords",
            },
            'hw_embed': {
                0: 'num_hotwords',
            "hw_embed": {
                1: "num_hotwords",
            },
        }
        

    def export_name(self):
        return 'model_eb.onnx'
    
        return "model_eb.onnx"


def export_rebuild_model(model, **kwargs):
        model.device = kwargs.get("device")
        is_onnx = kwargs.get("type", "onnx") == "onnx"
        encoder_class = tables.encoder_classes.get(kwargs["encoder"]+"Export")
        model.encoder = encoder_class(model.encoder, onnx=is_onnx)
        
        predictor_class = tables.predictor_classes.get(kwargs["predictor"]+"Export")
        model.predictor = predictor_class(model.predictor, onnx=is_onnx)
    model.device = kwargs.get("device")
    is_onnx = kwargs.get("type", "onnx") == "onnx"
    encoder_class = tables.encoder_classes.get(kwargs["encoder"] + "Export")
    model.encoder = encoder_class(model.encoder, onnx=is_onnx)

        # before decoder convert into export class
        embedder_class = ContextualEmbedderExport
        embedder_model = embedder_class(model, onnx=is_onnx)
    predictor_class = tables.predictor_classes.get(kwargs["predictor"] + "Export")
    model.predictor = predictor_class(model.predictor, onnx=is_onnx)

        decoder_class = tables.decoder_classes.get(kwargs["decoder"]+"Export")
        model.decoder = decoder_class(model.decoder, onnx=is_onnx)
        
        seaco_decoder_class = tables.decoder_classes.get(kwargs["seaco_decoder"]+"Export")
        model.seaco_decoder = seaco_decoder_class(model.seaco_decoder, onnx=is_onnx)
        
        from funasr.utils.torch_function import sequence_mask
        model.make_pad_mask = sequence_mask(kwargs["max_seq_len"], flip=False)
    # before decoder convert into export class
    embedder_class = ContextualEmbedderExport
    embedder_model = embedder_class(model, onnx=is_onnx)

    decoder_class = tables.decoder_classes.get(kwargs["decoder"] + "Export")
    model.decoder = decoder_class(model.decoder, onnx=is_onnx)

    seaco_decoder_class = tables.decoder_classes.get(kwargs["seaco_decoder"] + "Export")
    model.seaco_decoder = seaco_decoder_class(model.seaco_decoder, onnx=is_onnx)

    from funasr.utils.torch_function import sequence_mask

    model.make_pad_mask = sequence_mask(kwargs["max_seq_len"], flip=False)

    from funasr.utils.torch_function import sequence_mask

    model.make_pad_mask = sequence_mask(kwargs["max_seq_len"], flip=False)
    model.feats_dim = 560
    model.NOBIAS = 8377

    import copy
    import types

    backbone_model = copy.copy(model)

    # backbone
    backbone_model.forward = types.MethodType(export_backbone_forward, backbone_model)
    backbone_model.export_dummy_inputs = types.MethodType(
        export_backbone_dummy_inputs, backbone_model
    )
    backbone_model.export_input_names = types.MethodType(
        export_backbone_input_names, backbone_model
    )
    backbone_model.export_output_names = types.MethodType(
        export_backbone_output_names, backbone_model
    )
    backbone_model.export_dynamic_axes = types.MethodType(
        export_backbone_dynamic_axes, backbone_model
    )
    
        from funasr.utils.torch_function import sequence_mask
        model.make_pad_mask = sequence_mask(kwargs["max_seq_len"], flip=False)
        model.feats_dim = 560
        model.NOBIAS = 8377
    embedder_model.export_name = "model_eb"
    backbone_model.export_name = "model"

        import copy
        import types
        backbone_model = copy.copy(model)

        # backbone
        backbone_model.forward = types.MethodType(export_backbone_forward, backbone_model)
        backbone_model.export_dummy_inputs = types.MethodType(export_backbone_dummy_inputs, backbone_model)
        backbone_model.export_input_names = types.MethodType(export_backbone_input_names, backbone_model)
        backbone_model.export_output_names = types.MethodType(export_backbone_output_names, backbone_model)
        backbone_model.export_dynamic_axes = types.MethodType(export_backbone_dynamic_axes, backbone_model)
        backbone_model.export_name = types.MethodType(export_backbone_name, backbone_model)
        
        return backbone_model, embedder_model
    return backbone_model, embedder_model


def export_backbone_forward(
        self,
        speech: torch.Tensor,
        speech_lengths: torch.Tensor,
        bias_embed: torch.Tensor,
        # lmbd: float,
   ):
   # a. To device
   batch = {"speech": speech, "speech_lengths": speech_lengths}
    self,
    speech: torch.Tensor,
    speech_lengths: torch.Tensor,
    bias_embed: torch.Tensor,
    # lmbd: float,
):
    # a. To device
    batch = {"speech": speech, "speech_lengths": speech_lengths}

   enc, enc_len = self.encoder(**batch)
   mask = self.make_pad_mask(enc_len)[:, None, :]
   pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index = self.predictor(enc, mask)
   pre_token_length = pre_token_length.floor().type(torch.int32)
    enc, enc_len = self.encoder(**batch)
    mask = self.make_pad_mask(enc_len)[:, None, :]
    pre_acoustic_embeds, pre_token_length, alphas, pre_peak_index = self.predictor(enc, mask)
    pre_token_length = pre_token_length.floor().type(torch.int32)

   decoder_out, decoder_hidden, _ = self.decoder(enc, enc_len, pre_acoustic_embeds, pre_token_length, return_hidden=True, return_both=True)
   decoder_out = torch.log_softmax(decoder_out, dim=-1)
   # seaco forward
   B, N, D = bias_embed.shape
   _contextual_length = torch.ones(B) * N
    decoder_out, decoder_hidden, _ = self.decoder(
        enc, enc_len, pre_acoustic_embeds, pre_token_length, return_hidden=True, return_both=True
    )
    decoder_out = torch.log_softmax(decoder_out, dim=-1)
    # seaco forward
    B, N, D = bias_embed.shape
    _contextual_length = torch.ones(B) * N

   # ASF
   hotword_scores = self.seaco_decoder.forward_asf6(bias_embed, _contextual_length, decoder_hidden, pre_token_length)
   hotword_scores = hotword_scores[0].sum(0).sum(0)
   # _ = self.decoder2(bias_embed, _contextual_length, decoder_hidden, pre_token_length)
   # hotword_scores = self.decoder2.model.decoders[-1].attn_mat[0][0].sum(0).sum(0)
   dec_filter = torch.sort(hotword_scores, descending=True)[1][:51]
   contextual_info = bias_embed[:,dec_filter]
   num_hot_word = contextual_info.shape[1]
   _contextual_length = torch.Tensor([num_hot_word]).int().repeat(B).to(enc.device)
    # ASF
    hotword_scores = self.seaco_decoder.forward_asf6(
        bias_embed, _contextual_length, decoder_hidden, pre_token_length
    )
    hotword_scores = hotword_scores[0].sum(0).sum(0)
    # _ = self.decoder2(bias_embed, _contextual_length, decoder_hidden, pre_token_length)
    # hotword_scores = self.decoder2.model.decoders[-1].attn_mat[0][0].sum(0).sum(0)
    dec_filter = torch.sort(hotword_scores, descending=True)[1][:51]
    contextual_info = bias_embed[:, dec_filter]
    num_hot_word = contextual_info.shape[1]
    _contextual_length = torch.Tensor([num_hot_word]).int().repeat(B).to(enc.device)

   # again
   cif_attended, _ = self.seaco_decoder(contextual_info, _contextual_length, pre_acoustic_embeds, pre_token_length)
   dec_attended, _ = self.seaco_decoder(contextual_info, _contextual_length, decoder_hidden, pre_token_length)
   merged = cif_attended + dec_attended
   dha_output = self.hotword_output_layer(merged)
   dha_pred = torch.log_softmax(dha_output, dim=-1)
   # merging logits
   dha_ids = dha_pred.max(-1)[-1]
   dha_mask = (dha_ids == self.NOBIAS).int().unsqueeze(-1)
   decoder_out = decoder_out * dha_mask + dha_pred * (1-dha_mask)
   return decoder_out, pre_token_length, alphas
    # again
    cif_attended, _ = self.seaco_decoder(
        contextual_info, _contextual_length, pre_acoustic_embeds, pre_token_length
    )
    dec_attended, _ = self.seaco_decoder(
        contextual_info, _contextual_length, decoder_hidden, pre_token_length
    )
    merged = cif_attended + dec_attended
    dha_output = self.hotword_output_layer(merged)
    dha_pred = torch.log_softmax(dha_output, dim=-1)
    # merging logits
    dha_ids = dha_pred.max(-1)[-1]
    dha_mask = (dha_ids == self.NOBIAS).int().unsqueeze(-1)
    decoder_out = decoder_out * dha_mask + dha_pred * (1 - dha_mask)

    # get predicted timestamps
    us_alphas, us_cif_peak = self.predictor.get_upsample_timestmap(enc, mask, pre_token_length)
    
    return decoder_out, pre_token_length, us_alphas, us_cif_peak


def export_backbone_dummy_inputs(self):
   speech = torch.randn(2, 30, self.feats_dim)
   speech_lengths = torch.tensor([15, 30], dtype=torch.int32)
   bias_embed = torch.randn(2, 1, 512)
   return (speech, speech_lengths, bias_embed)
    speech = torch.randn(2, 30, self.feats_dim)
    speech_lengths = torch.tensor([15, 30], dtype=torch.int32)
    bias_embed = torch.randn(2, 1, 512)
    return (speech, speech_lengths, bias_embed)


def export_backbone_input_names(self):
   return ['speech', 'speech_lengths', 'bias_embed']
    return ["speech", "speech_lengths", "bias_embed"]


def export_backbone_output_names(self):
   return ['logits', 'token_num', 'alphas']
    return ["logits", "token_num", "us_alphas", "us_cif_peak"]


def export_backbone_dynamic_axes(self):
   return {
      'speech': {
         0: 'batch_size',
         1: 'feats_length'
      },
      'speech_lengths': {
         0: 'batch_size',
      },
      'bias_embed': {
         0: 'batch_size',
         1: 'num_hotwords'
      },
      'logits': {
         0: 'batch_size',
         1: 'logits_length'
      },
      'pre_acoustic_embeds': {
         1: 'feats_length1'
      }
   }
	
def export_backbone_name(self):
   return 'model.onnx'
    
    return {
        "speech": {0: "batch_size", 1: "feats_length"},
        "speech_lengths": {
            0: "batch_size",
        },
        "bias_embed": {0: "batch_size", 1: "num_hotwords"},
        "logits": {0: "batch_size", 1: "logits_length"},
        "pre_acoustic_embeds": {1: "feats_length1"},
        "us_alphas": {0: "batch_size", 1: "alphas_length"},
        "us_cif_peak": {0: "batch_size", 1: "alphas_length"},
    }