PaddlePaddle/PaddleOCR · error · ValueError

You have to specify either decoder_input_ids or decoder_inpu

Error message

You have to specify either decoder_input_ids or decoder_inputs_embeds

What it means

The decoder forward in the UniMERNet head requires exactly one input source: decoder_input_ids or decoder_inputs_embeds. If both are None there is nothing to decode, so the model raises this ValueError rather than producing a confusing downstream shape error.

Source

Thrown at ppocr/modeling/heads/rec_unimernet_head.py:1049

        )
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        return_dict = (
            return_dict if return_dict is not None else self.config.use_return_dict
        )

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError(
                "You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time"
            )
        elif input_ids is not None:
            input = input_ids
            input_shape = input.shape
            input_ids = input_ids.reshape([-1, input_shape[-1]])
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.shape[:-1]
            input = inputs_embeds[:, :, -1]
        else:
            raise ValueError(
                "You have to specify either decoder_input_ids or decoder_inputs_embeds"
            )

        past_key_values_length = (
            past_key_values[0][0].shape[2] if past_key_values is not None else 0
        )

        if inputs_embeds is None:
            inputs_embeds = self.embed_tokens(input_ids) * self.embed_scale

        if self._use_flash_attention_2:
            attention_mask = (
                attention_mask
                if (attention_mask is not None and 0 in attention_mask)
                else None
            )
        else:
            attention_mask = _prepare_4d_causal_attention_mask(

View on GitHub (pinned to 2661c7c0ef)

Solutions

Pass decoder_input_ids (usually a [batch,1] start token) to the forward call
If using generate, ensure generation_config.decoder_start_token_id / bos_token_id is set so ids are prepared
Or pass inputs_embeds if you embed inputs yourself

Example fix

# before
out = model_decoder(encoder_hidden_states=enc)  # no decoder input
# after
start = paddle.full([bsz, 1], decoder_start_token_id, dtype='int64')
out = model_decoder(input_ids=start, encoder_hidden_states=enc)

Defensive patterns

Strategy: validation

Validate before calling

if input_ids is None and inputs_embeds is None:
    input_ids = paddle.full([bsz, 1], decoder_start_token_id, dtype='int64')

Type guard

def has_decoder_input(input_ids, inputs_embeds) -> bool:
    return input_ids is not None or inputs_embeds is not None

Prevention

Centralize decoder input creation (start token) in one helper
Log model_kwargs keys before generate to confirm decoder ids exist

When it happens

Trigger: Calling the decoder/model forward with input_ids=None and inputs_embeds=None; typical when a generator fails to inject decoder_start_token_id or when a wrapper drops the argument.

Common situations: Custom generate() implementations that build model_kwargs but forget decoder_input_ids; encoder-decoder setups where the start token id is undefined so no ids are created.

Related errors

AI-assisted analysis of PaddlePaddle/PaddleOCR@2661c7c0ef (2026-08-14). Data as JSON: /api/errors/70f9aa0bf859ce04. Report an issue: GitHub.