invoke-ai/InvokeAI · error · TypeError

Expected torch.Tensor for prompt embeddings, got {type(promp

Error message

Expected torch.Tensor for prompt embeddings, got {type(prompt_embeds).__name__}. Text encoder returned unexpected type.

What it means

_encode_prompt in the Z-Image text encoder invocation expects the text encoder to return prompt embeddings as a torch.Tensor (after masking/batch-indexing). If the underlying encoder or pipeline path returns another type (list, tuple, nested tensors), this TypeError is thrown so downstream diffusion code never receives malformed embeddings. It is a defensive type check modeled on diffusers' ZImagePipeline expectations.

Source

Thrown at invokeai/app/invocations/z_image_text_encoder.py:194

                raise RuntimeError(
                    "Text encoder did not return hidden_states. "
                    "Ensure output_hidden_states=True is supported by this model."
                )
            if len(outputs.hidden_states) < 2:
                raise RuntimeError(
                    f"Expected at least 2 hidden states from text encoder, got {len(outputs.hidden_states)}. "
                    "This may indicate an incompatible model or configuration."
                )
            prompt_embeds = outputs.hidden_states[-2]

            # Z-Image expects a 2D tensor [seq_len, hidden_dim] with only valid tokens
            # Based on diffusers ZImagePipeline implementation:
            # embeddings_list.append(prompt_embeds[i][prompt_masks[i]])
            # Since batch_size=1, we take the first item and filter by mask
            prompt_embeds = prompt_embeds[0][prompt_mask[0]]

        if not isinstance(prompt_embeds, torch.Tensor):
            raise TypeError(
                f"Expected torch.Tensor for prompt embeddings, got {type(prompt_embeds).__name__}. "
                "Text encoder returned unexpected type."
            )
        return prompt_embeds

    def _lora_iterator(self, context: InvocationContext) -> Iterator[PatchSpec]:
        """Iterate over LoRA models to apply to the Qwen3 text encoder."""
        for lora in self.qwen3_encoder.loras:
            lora_info = context.models.load(lora.lora)
            if not isinstance(lora_info.model, ModelPatchRaw):
                raise TypeError(
                    f"Expected ModelPatchRaw for LoRA '{lora.lora.key}', got {type(lora_info.model).__name__}. "
                    "The LoRA model may be corrupted or incompatible."
                )
            yield (lora_info.model, lora.weight, lora_info.model_in_ram())

View on GitHub (pinned to 0b6a024f2f)

Solutions

1. Inspect the actual return type of the text encoder (print(type(prompt_embeds))) at the call site and convert with torch.stack()/torch.cat() before calling _encode_prompt
2. Verify the diffusers ZImagePipeline version matches what InvokeAI expects; upgrade/downgrade diffusers so encode_prompt returns a Tensor
3. Ensure the mask-slicing path (prompt_embeds[0][prompt_mask[0]]) is applied to a Tensor, not to a list of per-item outputs
4. If using a custom encoder, wrap it so it returns a single torch.Tensor for the embeddings

Example fix

// before
prompt_embeds = text_encoder(input_ids)  # returns list of tensors
// after
prompt_embeds = torch.stack(text_encoder(input_ids)) if isinstance(text_encoder_out, list) else text_encoder_out

import torch
assert isinstance(prompt_embeds, torch.Tensor), f"got {type(prompt_embeds).__name__}"

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

try:
    embeds = invocation.invoke(context)
except TypeError as e:
    if 'prompt embeddings' in str(e):
        prompt_embeds = torch.as_tensor(raw_output)
    else:
        raise

When it happens

Trigger: Calling invoke() on the Z-Image text encoder when the text encoder output is not a torch.Tensor, e.g. an encoder wrapper returning a list of tensors, a tuple like (embeddings, attn_mask) that was not unpacked, or a mocked/stubbed encoder in tests returning a list.

Common situations: Using an incompatible or custom text-encoder implementation with the Z-Image pipeline; a diffusers version change altering the return type of encode_prompt; passing raw tokenizer output instead of encoded embeddings; batch/mask slicing producing a list when batch_size != 1 assumptions break.

Related errors

• Unsupported control_lllite type: {type(control_lllite)}
• Expected PreTrainedModel for Gemma encoder, got {type(gemma_
• Expected ModelPatchRaw for LoRA '{lora.lora.key}', got {type
• Model config discriminator value must be computed from a dic
• missing config.json at {config_path}

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/6d779067cc7b1014. Report an issue: GitHub.