lllyasviel/Fooocus · error · ValueError

You have to specify either input_ids or inputs_embeds or enc

Error message

You have to specify either input_ids or inputs_embeds or encoder_embeds

What it means

NLVR encoder's forward requires one of input_ids, inputs_embeds, or encoder_embeds to infer input shape, batch size, and device; if all are None it raises this ValueError before any computation.

Source

Thrown at extras/BLIP/models/nlvr_encoder.py:767

        else:
            use_cache = False

        if input_ids is not None and inputs_embeds is not None:
            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
        elif input_ids is not None:
            input_shape = input_ids.size()
            batch_size, seq_length = input_shape
            device = input_ids.device
        elif inputs_embeds is not None:
            input_shape = inputs_embeds.size()[:-1]
            batch_size, seq_length = input_shape
            device = inputs_embeds.device
        elif encoder_embeds is not None:    
            input_shape = encoder_embeds.size()[:-1]
            batch_size, seq_length = input_shape 
            device = encoder_embeds.device
        else:
            raise ValueError("You have to specify either input_ids or inputs_embeds or encoder_embeds")

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

        if attention_mask is None:
            attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
            
        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
        # ourselves in which case we just need to make it broadcastable to all heads.
        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, 
                                                                                 device, is_decoder)

        # If a 2D or 3D attention mask is provided for the cross-attention
        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
        if encoder_hidden_states is not None:
            if type(encoder_hidden_states) == list:
                encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states[0].size()
            else:

View on GitHub (pinned to ae05379cc9)

Solutions

1. Pass input_ids (or inputs_embeds / encoder_embeds) explicitly
2. Assert the input tensor is not None right before the forward call
3. Log which input source your pipeline chose in debug builds to catch silent None

Example fix

// before
outputs = encoder(attention_mask=mask)

// after
assert ids is not None, 'input_ids missing from batch'
outputs = encoder(input_ids=ids, attention_mask=mask)

assert input_ids is not None or inputs_embeds is not None or encoder_embeds is not None

When it happens

Trigger: Calling the encoder with only auxiliary args (attention_mask, past_key_values, labels), or with an input tensor that a preprocessing step set to None.

Common situations: None tensors from a failing dataloader or tokenizer; refactors where encoder_embeds was renamed; wrapper code that conditionally builds inputs but all branches were skipped.

Related errors

• You cannot specify both input_ids and inputs_embeds at the s
• checkpoint url or path is invalid
• You cannot specify both input_ids and inputs_embeds at the s
• You have to specify either input_ids or inputs_embeds or enc
• The hidden size (%d) is not a multiple of the number of atte

AI-assisted analysis of lllyasviel/Fooocus@ae05379cc9 (2026-08-15). Data as JSON: /api/errors/80f2505fa96b7207. Report an issue: GitHub.