unslothai/unsloth · error · ValueError

Local base_repo is not a diffusers pipeline directory (no {i

Error message

Local base_repo is not a diffusers pipeline directory (no {indexes}): {base}

What it means

Raised by the local-path validation for diffusion base repos: the path exists on disk and is a directory (or a non-directory path) but contains neither model_index.json nor, when allow_modular, modular_model_index.json — so it is not a diffusers pipeline directory and cannot serve as base_repo.

Source

Thrown at studio/backend/core/inference/diffusion.py:593

    pipeline (MiniMax-H3 ships no ``model_index.json`` at all), and the local-model scanners
    already count either index. Off by default -- a conventional ``DiffusionPipeline`` load still
    needs the conventional index, and accepting a modular directory there would only move the
    failure back into the loader."""
    base = (base_repo or "").strip()
    if not base:
        return
    try:
        root = Path(base).expanduser()
        exists = root.exists()
    except OSError:
        return  # invalid path characters -> a remote id, not a local path
    if not exists:
        return
    indexes = ["model_index.json"]
    if allow_modular:
        indexes.append("modular_model_index.json")
    if not root.is_dir() or not any((root / name).is_file() for name in indexes):
        raise ValueError(
            f"Local base_repo is not a diffusers pipeline directory "
            f"(no {' or '.join(indexes)}): {base}"
        )


def _repo_access_message(repo: str, *, gated: bool) -> str:
    """The repo id AND its licence page: the worker's 401/403 names neither, and the base comes from a
    card tag, so the user never saw which repo it is."""
    url = f"https://huggingface.co/{repo}"
    if gated:
        return (
            f"'{repo}' is gated on Hugging Face and this model cannot be downloaded without it. "
            f"Accept its licence at {url}, then add a Hugging Face token that has access in "
            "Studio settings and try again."
        )
    return (
        f"'{repo}' could not be read from Hugging Face (private, renamed or removed) and this "
        f"model cannot be downloaded without it. Check {url}, then add a Hugging Face token that "

View on GitHub (pinned to 203007d190)

Solutions

Point base_repo at the directory that directly contains model_index.json (the true pipeline root)
If the weights are a single .safetensors/.gguf checkpoint, pass it as single-file input instead of a pipeline base_repo
Re-download or complete the pipeline so model_index.json and component subfolders are present
Check for a nested wrapper directory (downloads often add one level)

Example fix

# before
{"base_repo": "/models/stable-diffusion/checkpoints"}

# after
{"base_repo": "/models/stable-diffusion/pipeline"}  # dir containing model_index.json

Defensive patterns

Strategy: validation

Validate before calling

from pathlib import Path
def is_diffusers_pipeline_dir(base: str, allow_modular: bool = False) -> bool:
    root = Path(base).expanduser()
    if not root.is_dir():
        return False
    names = ["model_index.json"] + (["modular_model_index.json"] if allow_modular else [])
    return any((root / n).is_file() for n in names)

Try / catch

try:
    validate_local_base(base, allow_modular=True)
except ValueError as e:
    if "not a diffusers pipeline directory" in str(e):
        hint = "point base_repo at the folder containing model_index.json"

Prevention

Check for model_index.json at the exact path before passing base_repo
Use the single-file path for raw .safetensors checkpoints
Watch for extra wrapper directories created by downloads

When it happens

Trigger: Pointing base_repo at a checkpoint folder holding only .safetensors + a config.yaml (SingleFile layout), a LoRA directory, a GGUF folder, or a hub cache blob directory; a typo'd subfolder inside a real pipeline dir; an empty/failed download directory.

Common situations: Mixing up A1111/ComfyUI-style checkpoint layouts with diffusers layouts; moving/renaming pipeline folders so model_index.json ends up outside; interrupted downloads leaving a directory with only partial files.

Related errors

AI-assisted analysis of unslothai/unsloth@203007d190 (2026-08-15). Data as JSON: /api/errors/e8c90a815f8ab280. Report an issue: GitHub.