fix: restore global_step from .states when resuming training

[ltianyi992]

Summary

Closes #18108

load_model() in ppocr/utils/save_load.py correctly restores start_epoch from the .states file, but silently drops global_step. Because program.train() initialises global_step from pre_best_model_dict["global_step"] (lines 232–234), the counter resets to 0 on every resume — corrupting log timestamps, eval scheduling, VisualDL step indices, and checkpoint metadata.

Root cause

Both code paths inside load_model() read the .states file and populate best_model_dict, but neither included:

if "global_step" in states_dict:
    best_model_dict["global_step"] = states_dict["global_step"]

Fix

Add the two-line restore to both paths:

1. The KIE path (metric.states — line ~92)
2. The main path (.states — line ~162)

No other logic changed. program.train() already has the consumer code; this patch just feeds it the correct value.

Test plan

• Resume a training run from a checkpoint saved at step N; verify logs show global_step starting at N+1 instead of 1
• Verify no regression for fresh-start training (no .states file present)
• Verify KIE model resume also restores global_step correctly

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

james seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[Copilot]

Pull request overview

This PR fixes resume-training state propagation by restoring global_step from the checkpoint .states/metric.states files inside ppocr/utils/save_load.py, preventing tools/program.py from resetting the step counter to 0 on resume.

Changes:

• Restore global_step from states_dict into best_model_dict in the KIE resume path (metric.states).
• Restore global_step from states_dict into best_model_dict in the standard resume path (.states).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.