invoke-ai/InvokeAI · error · RuntimeError

{source} is missing {key} after prefix strip and key convers

Error message

{source} is missing {key} after prefix strip and key conversion

What it means

_build_wan_transformer_config probes specific tensor keys (starting with patch_embedding.weight) to infer architecture parameters. If, after stripping diffusers prefixes and converting key names, a required tensor is absent from the state dict, the local `require` closure raises RuntimeError naming the missing key.

Source

Thrown at invokeai/backend/model_manager/load/model_loaders/wan.py:342

    release is described by its own weights rather than by a hard-coded table of
    known repos.

    ``source`` only flavours the error messages.
    """
    num_layers = 0
    for key in sd.keys():
        if isinstance(key, str) and key.startswith("blocks."):
            parts = key.split(".")
            if len(parts) >= 2:
                try:
                    num_layers = max(num_layers, int(parts[1]) + 1)
                except ValueError:
                    pass

    def require(key: str) -> tuple[int, ...]:
        tensor = sd.get(key)
        if tensor is None:
            raise RuntimeError(f"{source} is missing {key} after prefix strip and key conversion")
        return _tensor_shape(tensor)

    # Patch embedding gives us in_channels (16/36=A14B, 48=TI2V-5B) and inner dim.
    patch_shape = require("patch_embedding.weight")
    inner_dim = patch_shape[0]
    in_channels = patch_shape[1]

    # Wan uses head_dim=128 throughout the family; num_heads = inner_dim / 128.
    attention_head_dim = 128
    num_attention_heads = inner_dim // attention_head_dim

    ffn_dim = require("blocks.0.ffn.net.0.proj.weight")[0]

    text_w = sd.get("condition_embedder.text_embedder.linear_1.weight")
    text_dim = _tensor_shape(text_w)[1] if text_w is not None else 4096

    # out_channels is read from proj_out.weight directly rather than assumed
    # equal to in_channels: I2V-A14B has in_channels=36 (16 noise + 16

View on GitHub (pinned to 0b6a024f2f)

Solutions

Verify the file actually contains Wan transformer weights (not a VAE/text-encoder or unrelated model).
Re-download the checkpoint; a truncated file can lose early tensors.
Compare the file's key names against expected Wan keys; if naming is nonstandard, use a repackaged diffusers-compatible checkpoint.
Update InvokeAI so the latest prefix-strip/key-conversion rules apply.

Example fix

// before: wrong file registered as transformer
path = "wan_vae.safetensors"  # no patch_embedding.weight

// after
path = "wan2.1_t2v_1.3b_transformer.safetensors"

Defensive patterns

Strategy: validation

Validate before calling

from safetensors import safe_open

def validate_has_patch_embedding(path):
    with safe_open(path, framework='pt') as f:
        keys = set(f.keys())
    if not any('patch_embedding.weight' in k for k in keys):
        raise ValueError(f"{path} is not a Wan transformer checkpoint (no patch_embedding.weight)")

Try / catch

try:
    model = loader.load_model(config, SubModelType.Transformer)
except RuntimeError as e:
    if 'after prefix strip and key conversion' in str(e):
        verify_file_is_wan_transformer(config.path)  # correct the record or re-download
    else:
        raise

Prevention

Ensure model records point at transformer checkpoints, not VAE/T5 files.
Re-download files whose size differs from the published size.
Prefer diffusers-standard key naming repackaging for community checkpoints.
Update InvokeAI for the newest key-mapping rules.

When it happens

Trigger: Loading a single-file Wan checkpoint whose state dict lacks expected keys like patch_embedding.weight — e.g., a non-transformer file, a checkpoint with entirely different naming conventions, or a GGUF/compressed file misdetected as a standard checkpoint.

Common situations: Pointing a Wan checkpoint model record at a VAE or T5 encoder file by mistake; exotic community repackaging with unfamiliar key layout; corrupt or partially written safetensors file.

Related errors

AI-assisted analysis of invoke-ai/InvokeAI@0b6a024f2f (2026-08-29). Data as JSON: /api/errors/54ffe594351cc634. Report an issue: GitHub.