PyTorch Model / Weights

Identify first: AI model or Python configuration

Size and readable text are clues, not definitive identification. You can inspect a copy through File → Open in a text editor; readable content is not proof of safety.

Opening steps

Choose your environment or purpose to see its steps.

Menu labels and support vary by app, OS and version. Preserve the original before making changes.

Minimal loading example for a plain state_dict

This function is only for a trusted PTH containing a plain state_dict. Pass a model already constructed with the producer’s architecture and configuration. It neither creates an architecture nor establishes the safety of an unknown file.

import torch

def load_weights(model, path):
    state = torch.load(
        path, map_location="cpu", weights_only=True
    )
    model.load_state_dict(state, strict=True)
    model.eval()
    return model

The example is explanatory; this site does not upload or execute PTH files. eval() changes model behavior and is separate from disabling gradient tracking.

Source for these steps：PyTorch Project · Saving and Loading Models / PyTorch Project · torch.load

State dictionaries and checkpoints load differently

• A state_dict maps parameter names to tensors and related state; it does not necessarily supply the Python model class or preprocessing.
• A training checkpoint may also contain optimizer state, epoch or loss. Select the producer’s documented model-weight key instead of passing the whole checkpoint to load_state_dict.
• A serialized complete model may depend on original Python classes and libraries. A restricted-loading failure is not a reason to blindly allow unknown objects.

Source for these steps：PyTorch Project · Saving and Loading Models / PyTorch Project · Serialization semantics

Troubleshooting by symptom

• Missing/unexpected keys: check architecture, dictionary nesting and training wrappers. Verify the producer’s layout before removing prefixes such as module..
• Size mismatch: layer width, class count or other settings differ from saved tensor shapes. strict=False is not a universal fix or a guarantee of correct inference.
• CUDA device error: CPU mapping may help loading, but a GPU-dependent model or pipeline may still not run entirely on CPU.
• Restricted-loading failure: establish serialization and supported versions. Do not automatically disable weights_only or allowlist unknown classes just to bypass the error.
• ModuleNotFoundError: a saved complete model may require the original class. Verify dependencies with the trusted producer or request a weights-only release.
• Archive error or unexpectedly small file: check for an incomplete download or an HTML error page. Compare the publisher’s size or hash when available.
• Out of memory: required RAM/VRAM depends on architecture and inputs, not just file size. CPU mapping does not eliminate memory requirements.
• Import failure involving a path PTH: check the active interpreter and configured locations. Model loading or ONNX conversion will not repair path configuration.

PTH → ONNX / Safetensors: identify the export purpose

1. Establish trusted architecture, weights, preprocessing and input specifications; validate the original model first.
2. ONNX exports model computation. Review example inputs, supported operations, dynamic shapes and destination runtime requirements.
3. Safetensors primarily stores tensors. Select the required tensors and destination naming/layout; it does not replace the model’s complete Python implementation.
4. Save separately and compare outputs, shapes, accuracy and performance on the same inputs. Path-configuration PTH files are not model-conversion inputs.

Renaming .pth to .onnx or .safetensors is not conversion. The initial PTH loading step also carries risk; consider a supported export supplied by the producer.

Related-format differences

Handling and security

PyTorch object serialization can involve pickle-based code execution. weights_only=True restricts loading but does not guarantee protection against every risk or memory exhaustion. Verify origin, release terms, file identity and updated dependencies; do not load unknown models. Python configuration PTH can also contain import lines executed at startup, depending on the version.

Questions and answers

Primary and official sources

• Project documentationSaving and Loading Models

  Distinguishes state dictionaries, complete modules and checkpoints.

  PyTorch Project · Link checked：2026-10-10
• Project documentationtorch.load

  map_location, weights_only and warnings about untrusted input.

  PyTorch Project · Link checked：2026-10-10
• Project documentationSerialization semantics

  Restricted loading and the limits of its security guarantees.

  PyTorch Project · Link checked：2026-10-10
• Project documentationtorch.onnx

  ONNX export requirements, model definitions and example inputs.

  PyTorch Project · Link checked：2026-10-10
• Project documentationsite — Site-specific configuration hook

  Path-configuration PTH files, startup processing and version differences.

  Python Software Foundation · Link checked：2026-10-10
• Project documentationSafetensors

  Tensor storage and differences from PyTorch object serialization.

  Hugging Face · Link checked：2026-10-10