unslothai/unsloth · error · HTTPException

Selected cached model is no longer available.

Error message

Selected cached model is no longer available.

What it means

Raised as a 404 by the model-config helper when latest_snapshot_from_cache_path() cannot find a usable HF-hub snapshot for the requested model under the given local_path. It means the model cache entry the user selected previously no longer resolves to a loadable snapshot (missing config.json/adapter_config.json or the snapshot directory vanished).

Source

Thrown at studio/backend/routes/models.py:2125

def _model_config_inspection_target(
    model_name: str, prefer_local_cache: bool, local_path: Optional[str]
) -> str:
    if not prefer_local_cache or is_local_path(model_name):
        return model_name
    from hub.utils.hf_cache_state import (
        latest_snapshot_from_cache_path,
        with_load_subdirs,
    )

    snapshot = latest_snapshot_from_cache_path(
        local_path,
        "model",
        canonical_model_repo_id(model_name),
        with_load_subdirs(model_name, ("config.json", "adapter_config.json")),
    )
    if snapshot is None:
        raise HTTPException(
            status_code = 404,
            detail = "Selected cached model is no longer available.",
        )
    return snapshot


@router.get("/config/{model_name:path}")
async def get_model_config(
    model_name: str,
    hf_token: Optional[str] = Query(None),
    prefer_local_cache: bool = False,
    local_path: Optional[str] = None,
    header_hf_token: Optional[str] = Depends(get_hf_token),
    current_subject: str = Depends(get_current_subject),
):
    """Get configuration for a specific model (wraps load_model_defaults)."""
    hf_token = _normalize_hf_token(header_hf_token) or _normalize_hf_token(hf_token)
    from core.inference.llama_cpp import _hf_offline_if_unreachable_for

View on GitHub (pinned to 203007d190)

Solutions

1. Re-download the model snapshot (huggingface-cli download <repo_id>) so the cache contains config.json/adapter_config.json again.
2. Inspect local_path/snapshots to confirm a complete snapshot directory exists for canonical_model_repo_id(model_name).
3. If the cache was intentionally pruned, clear the stale selection in the UI and re-pick the model so local_path points at a valid entry.
4. Fall back to fetching config from the Hub (drop prefer_local_cache / local_path).

from pathlib import Path

def snapshot_complete(local_path: str, repo_id: str) -> bool:
    snaps = Path(local_path).expanduser() / 'snapshots'
    if not snaps.is_dir():
        return False
    return any(
        (s / 'config.json').is_file() or (s / 'adapter_config.json').is_file()
        for s in snaps.iterdir() if s.is_dir()
    )

if not snapshot_complete(local_path, model_name):
    raise FileNotFoundError('cached snapshot missing configs — re-download before requesting config')

try:
    cfg = client.get(f'/api/models/config/{model}', params={'local_path': lp, 'prefer_local_cache': True})
except HTTPError as e:
    if e.response.status_code == 404 and 'no longer available' in e.response.json()['detail']:
        cfg = client.get(f'/api/models/config/{model}')  # fall back to Hub fetch
    else: raise

When it happens

Trigger: GET /api/models/config/{model_name}?local_path=...&prefer_local_cache=true where the HF cache snapshot directory was deleted, pruned (huggingface-cli delete-cache), partially downloaded, or the expected config.json / adapter_config.json files are absent from every candidate snapshot under with_load_subdirs.

Common situations: Cache cleanup tools removing snapshots; interrupted downloads leaving partial snapshots; multi-revision caches where the 'latest' snapshot lacks the needed config; repo id casing mismatch in canonical_model_repo_id.

Related errors

• STT model '{model_id}' is not downloaded. Download it in Set
• dataset_local_cache_miss
• Path does not exist: {os.path.basename(requested_path)}
• Invalid model snapshot repository ID.
• The Hugging Face cache folder is invalid.

AI-assisted analysis of unslothai/unsloth@203007d190 (2026-08-15). Data as JSON: /api/errors/59268fcbba413968. Report an issue: GitHub.