`train_tokens_per_second` is wrong after continuing from checkpoint

[AmitMY]

System Info

• transformers version: 4.55.4
• Platform: macOS-15.6.1-arm64-arm-64bit
• Python version: 3.12.2
• Huggingface_hub version: 0.34.3
• Safetensors version: 0.5.3
• Accelerate version: 1.9.0
• Accelerate config: not found
• DeepSpeed version: not installed
• PyTorch version (accelerator?): 2.7.1 (NA)
• Tensorflow version (GPU?): not installed (NA)
• Flax version (CPU?/GPU?/TPU?): not installed (NA)
• Jax version: not installed
• JaxLib version: not installed
• Using distributed or parallel set-up in script?:

Who can help?

@zach-huggingface @SunMarc

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

When continuing run_clm.py from a checkpoint, it starts with insane tokens per second numbers, then slowly, eventually, goes down.

c

Expected behavior

It should be calculated with the local training run information, not the global one. The numbers should roughly be the same since it is the same exact setup.

[ArkVex]

@ArthurZucker @SunMarc is this an issue ?

[ArthurZucker]

Yep probably!

[ArkVex]

Okk can you assign this to me plz @ArthurZucker ?

[ArthurZucker]

Hey! We never assign issue @ArkVex just feel free to open a PR and link it here!