Lightning-AI/pytorch-lightning · error · ValueError

`.{fn}()` found no path for the best weights: {ckpt_path!r}.

Error message

`.{fn}()` found no path for the best weights: {ckpt_path!r}. Please specify a path for a checkpoint `.{fn}(ckpt_path=PATH)`

What it means

Final guard in _parse_ckpt_path: after all resolution logic (best/last/hpc/registry), the resulting ckpt_path is still empty/falsy. This means the requested resolution mode produced no usable path and Lightning cannot proceed to load weights.

Source

Thrown at src/lightning/pytorch/trainer/connectors/checkpoint_connector.py:213

            ckpt_path = max(candidates_ts, key=candidates_ts.get)  # type: ignore[arg-type]

        elif ckpt_path == "hpc":
            if not self._hpc_resume_path:
                raise ValueError(
                    f'`.{fn}(ckpt_path="hpc")` is set but no HPC checkpoint was found.'
                    f" Please pass an exact checkpoint path to `.{fn}(ckpt_path=...)`"
                )
            ckpt_path = self._hpc_resume_path

        elif _is_registry(ckpt_path) and module_available("litmodels"):
            ckpt_path = find_model_local_ckpt_path(
                ckpt_path,
                default_model_registry=self.trainer._model_registry,
                default_root_dir=self.trainer.default_root_dir,
            )

        if not ckpt_path:
            raise ValueError(
                f"`.{fn}()` found no path for the best weights: {ckpt_path!r}. Please"
                f" specify a path for a checkpoint `.{fn}(ckpt_path=PATH)`"
            )
        return ckpt_path

    def resume_end(self) -> None:
        """Signal the connector that all states have resumed and memory for the checkpoint object can be released."""
        assert self.trainer.state.fn is not None
        if self._ckpt_path:
            message = "Restored all states" if self.trainer.state.fn == TrainerFn.FITTING else "Loaded model weights"
            rank_zero_info(f"{message} from the checkpoint at {self._ckpt_path}")

        # free memory
        self._loaded_checkpoint = {}
        torch.cuda.empty_cache()

        # wait for all to catch up
        self.trainer.strategy.barrier("_CheckpointConnector.resume_end")

View on GitHub (pinned to 9fed5c27d2)

Solutions

Pass an explicit filesystem or registry path to ckpt_path
Ensure a checkpoint actually exists: fit first, keep enable_checkpointing=True, and configure save_last=True or a monitor
Use ckpt_path=None to use current in-memory weights if that is acceptable

Example fix

# before
trainer.test(model, ckpt_path="last")  # no last.ckpt exists
# after
trainer = Trainer(callbacks=[ModelCheckpoint(monitor="val_loss", save_last=True)])
trainer.fit(model)
trainer.test(model, ckpt_path="last")

Defensive patterns

Strategy: try-catch

Validate before calling

from pathlib import Path
if ckpt_path in ("best", "last"):
    cand = mc.best_model_path if ckpt_path == "best" else mc.last_model_path
    assert cand, f"no {ckpt_path} checkpoint saved yet"
    assert Path(cand).exists(), f"{cand} missing on disk"

Try / catch

try:
    trainer.test(model, ckpt_path="last")
except ValueError as e:
    if "found no path" in str(e):
        trainer.test(model)  # current weights
    else:
        raise

Prevention

Check that the checkpoint file exists before calling validate/test/predict
Save resolved paths (best/last) right after fit and reuse them
Configure save_last=True when you rely on 'last'

When it happens

Trigger: ckpt_path="last" when no last checkpoint was saved (checkpointing disabled or no run yet), a registry/model reference that resolved to nothing, or best-model resolution yielding "" from the checkpoint callback; then calling trainer.validate/test/predict.

Common situations: Calling .test(ckpt_path="last") before or without a prior fit; fresh output directory; ModelCheckpoint with save_top_k=0; registry lookup returning an empty path.

Related errors

AI-assisted analysis of Lightning-AI/pytorch-lightning@9fed5c27d2 (2026-08-28). Data as JSON: /api/errors/d15b18ccc04d44c4. Report an issue: GitHub.