google-research/timesfm · error · FileNotFoundError

{self.WEIGHTS_FILENAME} not found in directory {path}

Error message

{self.WEIGHTS_FILENAME} not found in directory {path}

What it means

load_checkpoint() expects a checkpoint directory to contain a weights file named model.safetensors (cls.WEIGHTS_FILENAME). If path is a directory but that file is absent, a FileNotFoundError is raised. It exists so the loader fails fast instead of producing a confusing error deeper inside safetensors/torch loading.

Source

Thrown at src/timesfm/timesfm_2p5/timesfm_2p5_torch.py:292

  CONFIG_FILENAME = "config.json"

  def __init__(
    self,
    torch_compile: bool = True,
    config: Optional[dict] = None,
    **kwargs,
  ):
    self.model = TimesFM_2p5_200M_torch_module()
    self.torch_compile = torch_compile
    if config is not None:
      self._hub_mixin_config = config

  def load_checkpoint(self, path: str, **kwargs):
    """Loads a TimesFM model from a checkpoint directory or file."""
    if os.path.isdir(path):
      model_file_path = os.path.join(path, self.WEIGHTS_FILENAME)
      if not os.path.exists(model_file_path):
        raise FileNotFoundError(
          f"{self.WEIGHTS_FILENAME} not found in directory {path}"
        )
    else:
      model_file_path = path

    torch_compile = kwargs.pop("torch_compile", self.torch_compile)
    self.model.load_checkpoint(model_file_path, **kwargs)
    if torch_compile:
      logging.info("Compiling model...")
      self.model.forward = torch.compile(self.model.forward)

  @classmethod
  def _from_pretrained(
    cls,
    *,
    model_id: str = DEFAULT_REPO_ID,
    revision: Optional[str],
    cache_dir: Optional[Union[str, Path]],

View on GitHub (pinned to 331c6d33cb)

Solutions

1. Verify model.safetensors exists in the directory: ls <path>/model.safetensors; if missing, re-download the checkpoint (e.g. huggingface-cli download google/timesfm-2.5-200m-pytorch).
2. Ensure git-lfs is installed and pull the actual files: git lfs install && git lfs pull inside the cloned repo.
3. Alternatively pass the path of the .safetensors file itself instead of the directory — load_checkpoint accepts a file path when path is not a directory.
4. If pointing at a legacy checkpoint directory, locate the safetensors file and pass its full path, or copy/rename it to model.safetensors in the directory.

Example fix

// before: dir has old-style weights
model.load_checkpoint("/checkpoints/timesfm_v1")  # FileNotFoundError
// after: pass the weights file directly, or use a dir containing model.safetensors
model.load_checkpoint("/checkpoints/timesfm_v1/timesfm_weights.safetensors")

import os
if os.path.isdir(path) and not os.path.exists(os.path.join(path, "model.safetensors")):
    raise FileNotFoundError(f"model.safetensors missing in {path}; re-download checkpoint")

try:
    model.load_checkpoint(path)
except FileNotFoundError:
    model = TimesFM_2p5_200M_torch.from_pretrained("google/timesfm-2.5-200m-pytorch")

When it happens

Trigger: Calling model.load_checkpoint(path) where path is an existing directory that does not contain model.safetensors — e.g. the directory only holds flax checkpoint files, or an incomplete/interrupted download left the folder without the safetensors file.

Common situations: Downloading the Hugging Face repo without git-lfs (so model.safetensors is a pointer stub or missing), renaming checkpoint files manually, pointing at a directory holding an older TimesFM 1.x/2.0 checkpoint whose weights use a different filename, or copying only part of a checkpoint.

Related errors

• {cls.WEIGHTS_FILENAME} not found in directory {model_id}

AI-assisted analysis of google-research/timesfm@331c6d33cb (2026-08-29). Data as JSON: /api/errors/142864f97f9c280e. Report an issue: GitHub.