invoke-ai/InvokeAI · error · TypeError

Expected torch.Tensor for attention_mask, got

Error message

Expected torch.Tensor for attention_mask, got {type(attention_mask).__name__}. Tokenizer returned unexpected type.

What it means

Immediately after the input_ids check, _encode_prompt validates that text_inputs.attention_mask is a torch.Tensor. A non-tensor attention_mask raises a TypeError stating the tokenizer returned an unexpected type. The attention mask is required for the padded Qwen3 forward pass, so its type must be guaranteed.

Solutions

Use the stock transformers Qwen3 tokenizer so attention_mask is tensorized with return_tensors="pt".
Upgrade transformers if return_tensors is being ignored.
Validate the tokenizer output before passing to the invocation (isinstance check on attention_mask).
Convert manually if required: attention_mask = torch.tensor(raw_mask).

Example fix

// before
text_inputs = tok(prompt)  # returns lists
// after
text_inputs = tok(prompt, padding="longest", return_tensors="pt")  # tensors

Defensive patterns

Strategy: type-guard

Validate before calling

inputs = tok(prompt, padding="longest", return_tensors="pt")
if not isinstance(inputs.attention_mask, torch.Tensor):
    fail_fast(inputs.attention_mask)

Type guard

def is_tensor(x) -> bool:
    import torch
    return isinstance(x, torch.Tensor)

Try / catch

try:
    encode(context)
except TypeError as e:
    if "Expected torch.Tensor for attention_mask" in str(e):
        swap_to_stock_tokenizer()
    else:
        raise

Prevention

Use stock transformers tokenizers so attention_mask is tensorized automatically.
Never construct BatchEncoding manually without tensor conversion.
Pin a transformers version known to work with your tokenizer files.

When it happens

Trigger: Z-Image text encoding where tokenizer(..., padding="longest", return_tensors="pt").attention_mask is not a torch.Tensor, e.g. returned as a list/ndarray by a custom tokenizer.

Common situations: Custom tokenizer subclass bypassing tensor conversion; transformers version mismatch; manually constructed BatchEncoding lacking tensorized fields.

Related errors

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/ecad3745ef5eba35. Report an issue: GitHub.

Appendix: source

Thrown at invokeai/app/invocations/z_image_text_encoder.py:149

            # Tokenize the formatted prompt
            text_inputs = tokenizer(
                prompt_formatted,
                padding="max_length",
                max_length=max_seq_len,
                truncation=True,
                return_attention_mask=True,
                return_tensors="pt",
            )

            text_input_ids = text_inputs.input_ids
            attention_mask = text_inputs.attention_mask
            if not isinstance(text_input_ids, torch.Tensor):
                raise TypeError(
                    f"Expected torch.Tensor for input_ids, got {type(text_input_ids).__name__}. "
                    "Tokenizer returned unexpected type."
                )
            if not isinstance(attention_mask, torch.Tensor):
                raise TypeError(
                    f"Expected torch.Tensor for attention_mask, got {type(attention_mask).__name__}. "
                    "Tokenizer returned unexpected type."
                )

            # Check for truncation
            untruncated_ids = tokenizer(prompt_formatted, padding="longest", return_tensors="pt").input_ids
            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
                text_input_ids, untruncated_ids
            ):
                removed_text = tokenizer.batch_decode(untruncated_ids[:, max_seq_len - 1 : -1])
                context.logger.warning(
                    f"The following part of your input was truncated because `max_sequence_length` is set to "
                    f"{max_seq_len} tokens: {removed_text}"
                )

            # Get hidden states from the text encoder
            # Use the second-to-last hidden state like diffusers does
            prompt_mask = attention_mask.to(device).bool()

View on GitHub (pinned to 0b6a024f2f)