transformers - 💡(How to fix) Fix Loading checkpoint trained on v4.57 cannot be reload after upgraded to v5.2 & v5.3 [2 comments, 2 participants]

Fix Plan

1. Update Qwen3 to the latest version

Update the Qwen3 library to the latest version (v4.58 or later) to ensure you have the latest bug fixes and improvements.

2. Check for incompatible dependencies

Verify that all dependencies are compatible with the latest Qwen3 version. You can do this by running pip freeze and checking for any outdated packages.

3. Use from_pretrained with a remote URL

Try loading the checkpoint using a remote URL instead of a local path. This can help identify if the issue is specific to local file loading.

from_pretrained("path/to/model", local_files_only=False)

4. Disable caching

Disable caching to ensure that the model is reloaded from scratch. You can do this by setting cache_dir to None.

from_pretrained("path/to/model", local_files_only=False, cache_dir=None)

5. Check for file system issues

Verify that the file system is not causing the issue. Try loading the checkpoint on a different file system or disk.

Verification

• Try loading the checkpoint using the latest Qwen3 version and a remote URL.
• Verify that the model loads successfully without hanging.
• If the issue persists, try disabling caching and checking for file system issues.

Example Code

import torch
from qwen3 import AutoModelForSequenceClassification

model = AutoModelForSequenceClassification.from_pretrained(
    "path/to/model",
    local_files_only=False,
    cache_dir=None
)

Extra Tips

• Always use the latest version of Qwen3 to ensure you have the latest bug fixes and improvements.
• Verify that all dependencies are compatible with the latest Qwen3 version.
• Use remote URLs instead of local paths to load checkpoints.
• Disable caching to ensure that the model is reloaded from scratch.