odysseus-dev/odysseus · error · RuntimeError

Could not load model from {model_path}. Check diffusers vers

Error message

Could not load model from {model_path}. Check diffusers version and model format.

What it means

RuntimeError from scripts/diffusion_server.py: every loader strategy for the requested model failed — the pipeline loop over diffusers classes (from_pretrained / from_single_file with config variants) raised, leaving loaded=False, so the server aborts with 'Could not load model from {model_path}. Check diffusers version and model format.' Preceding WARN lines ('<Cls>.from_single_file (config=...) failed: ...') carry the real per-strategy causes.

Source

Thrown at scripts/diffusion_server.py:638

                            kwargs["config"] = local_config
                            logger.info(f"Trying {cls_name}.from_single_file with config={config}")
                        _pipe = cls.from_single_file(single_file, **kwargs)
                        _fix_meta_tensors(_pipe, torch_dtype)
                        if use_offload and _can_cpu_offload(target_device):
                            _pipe.enable_model_cpu_offload()
                            logger.info(f"Loaded as {cls_name} (single file, config={config}) with CPU offload")
                        else:
                            _pipe = _pipe.to(target_device)
                            logger.info(f"Loaded as {cls_name} (single file, config={config}) on {target_device}")
                        loaded = True
                        break
                    except Exception as e:
                        logger.warning(f"{cls_name}.from_single_file (config={config}) failed: {e}")
                        _pipe = None
                        _cleanup()

    if not loaded:
        raise RuntimeError(f"Could not load model from {model_path}. Check diffusers version and model format.")

    # Memory optimizations
    if _args.attention_slicing:
        try:
            _pipe.enable_attention_slicing()
            logger.info("Attention slicing enabled")
        except Exception:
            pass
    if _args.vae_slicing:
        try:
            _pipe.enable_vae_slicing()
            logger.info("VAE slicing enabled")
        except Exception:
            pass

    logger.info(f"Model loaded: {_model_id}")

    # Load LoRA weights if specified

View on GitHub (pinned to f9235ebbf1)

Solutions

1. Read the WARNING lines just above the traceback — they name each class/config attempt and its exception; fix that root cause first
2. Upgrade/align the diffusers version to one supporting the model's pipeline class (check the model card's required version)
3. For single-file checkpoints, pass an explicit config (e.g. config='./configs/sdxl' style argument the script supports) or convert to a diffusers repo layout
4. Verify the download: re-run with a clean cache / check file sizes against the source repo
5. Confirm the path is a directory containing model_index.json when using repo-format models

Example fix

# before
pip install diffusers==0.24.0
python scripts/diffusion_server.py --model ./flux1-dev.safetensors
# RuntimeError: Could not load model ...

# after: use a diffusers release that supports the architecture
pip install -U diffusers transformers accelerate safetensors
python scripts/diffusion_server.py --model ./flux1-dev.safetensors

import importlib.metadata as md
from packaging.version import Version

required = {"FLuxPipeline": "0.30.0", "StableDiffusionXLPipeline": "0.24.0"}  # per model card
assert Version(md.version("diffusers")) >= Version(required[cls]), "upgrade diffusers"
assert Path(model_path).exists() and (Path(model_path)/'model_index.json').exists() or model_path.endswith(('.safetensors','.ckpt'))

def is_supported_checkpoint(path: str) -> bool:
    p = Path(path)
    return (p.is_dir() and (p / 'model_index.json').exists()) or p.suffix in {'.safetensors', '.ckpt'}

try:
    serve_model(model_path)
except RuntimeError as e:
    if 'Could not load model' in str(e):
        # read the preceding '<Cls>.from_single_file ... failed' warnings for the real cause
        align_diffusers_version(); verify_download_integrity(model_path)

When it happens

Trigger: Loading a single-file checkpoint (.safetensors/.ckpt) with a diffusers version lacking or breaking from_single_file support for that class; a model directory in an unexpected format (missing model_index.json); corrupted weights; wrong pipeline class guesses for the architecture; out-of-memory during .to(device) after load attempts.

Common situations: diffusers major-version upgrades changing from_single_file kwargs/behavior; SDXL/Flux checkpoints needing newer diffusers than pinned; downloading only part of a repo (interrupted snapshot); mixing safetensors-only expectations with pickle checkpoints.

Related errors

• Remote Windows Diffusers serving is not supported yet; use l
• {type(_pipe).__name__} does not support image edits. Use /v1

AI-assisted analysis of odysseus-dev/odysseus@f9235ebbf1 (2026-08-14). Data as JSON: /api/errors/935d82bbb3ed1e56. Report an issue: GitHub.