invoke-ai/InvokeAI · critical · RuntimeError

Failed to load all parameters from SDNQ. The following remai

Error message

Failed to load all parameters from SDNQ. The following remain as meta tensors: {meta_names}.

What it means

Raised at the end of the SDNQ Z-Image Qwen3 encoder load path (_load_from_sdnq) when, after load_state_dict with assign=True, tied-weight handling, and meta-buffer re-initialization, model.named_parameters() still contains torch meta tensors. It means the SDNQ checkpoint did not supply materialized weights for every parameter the freshly-initialized Qwen3ForCausalLM declares, so the model would be unusable on device. The loader fails fast rather than producing a model with missing weights.

Source

Thrown at invokeai/backend/model_manager/load/model_loaders/z_image.py:1626

                if len(parts) == 2:
                    parent = model.get_submodule(parts[0])
                    buffer_name = parts[1]
                else:
                    parent = model
                    buffer_name = name

                if buffer_name == "inv_freq":
                    base = qwen_config.rope_theta
                    inv_freq = 1.0 / (base ** (torch.arange(0, head_dim, 2, dtype=torch.float32) / head_dim))
                    parent.register_buffer(buffer_name, inv_freq.to(dtype=compute_dtype), persistent=False)
                else:
                    logger.warning(f"Re-initializing unknown meta buffer: {name}")

        # Final check: ensure no meta tensors remain in parameters
        meta_params = [(name, p) for name, p in model.named_parameters() if p.is_meta]
        if meta_params:
            meta_names = [name for name, _ in meta_params]
            raise RuntimeError(
                f"Failed to load all parameters from SDNQ. The following remain as meta tensors: {meta_names}."
            )

        return model

View on GitHub (pinned to 0b6a024f2f)

Solutions

1. Re-download the SDNQ checkpoint from its official source; the file is likely truncated or from an incompatible export.
2. Check the meta_names list in the message and compare against the checkpoint's keys (safetensors header / gguf listing) to find which tensors are missing or renamed.
3. Verify the transformers/accelerate versions match what the model card requires, since Qwen3 parameter names can shift between versions.
4. If you re-saved the quantized file, redo the conversion so all parameters are written; lm_head.weight is the only acceptable omission (tied weights).
5. Regenerate the model with the correct Qwen3Config (layer count, vocab size) matching the checkpoint.

Example fix

// before
model = Qwen3ForCausalLM(wrong_config)  # e.g. vocab_size from a different tokenizer
missing, unexpected = model.load_state_dict(sd, strict=False, assign=True)
// after
model = Qwen3ForCausalLM(qwen_config_built_from_checkpoint)  # sizes derived from the SDNQ sd
missing, unexpected = model.load_state_dict(sd, strict=False, assign=True)
raise_on_incomplete_sdnq_load('SDNQ Qwen3 encoder', missing, unexpected, allowed_missing={'lm_head.weight'})

import torch

def assert_no_meta_params(model: torch.nn.Module) -> None:
    meta = [n for n, p in model.named_parameters() if p.is_meta]
    if meta:
        raise RuntimeError(f'Missing weights for: {meta}')

def has_meta_params(model: torch.nn.Module) -> bool:
    return any(p.is_meta for _, p in model.named_parameters())

try:
    model = loader._load_from_sdnq(...)
except RuntimeError as e:
    if 'remain as meta tensors' in str(e):
        logger.error(f'SDNQ checkpoint incomplete: {e}. Re-download the model file.')
        raise
    raise

When it happens

Trigger: Loading an SDNQ-quantized Z-Image text-encoder checkpoint whose state dict is missing required parameters (other than the allowed lm_head.weight), or contains keys from an incompatible/contaminated SDNQ export that don't match Qwen3ForCausalLM's expected names; also when load_state_dict(assign=True) leaves parameters unassigned because names/shapes mismatch.

Common situations: Using an SDNQ export of the wrong architecture or a partially-converted checkpoint; a library/transformers version where Qwen3 parameter names changed; hand-trimmed or re-saved quantized files that dropped tensors; mismatches between the config's num_hidden_layers/vocab_size and the checkpoint contents.

Related errors

• Only Main_SDNQ_ZImage_Config or Main_SDNQ_Diffusers_ZImage_C
• transformer is SDNQ-quantized; use Main_SDNQ_Diffusers_FLUX_
• transformer is SDNQ-quantized; use Main_SDNQ_Diffusers_Flux2
• transformer is SDNQ-quantized; use Main_SDNQ_Diffusers_ZImag
• folder is SDNQ-quantized; use Qwen3Encoder_SDNQ_Folder_Confi

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/ae78b86992f8df42. Report an issue: GitHub.