[Tokenizer] GLM4 tokenizer does not support `padding_side` in `.pad()` (transformers latest version)

[GitHub119ymj]

Describe the problem

I'm currently using GLM4 with the latest version of HuggingFace's transformers library in a P-Tuning experiment. While preparing input batches, I encountered the following error:

This happens when I try to use .pad(padding_side="right") — a common approach in HuggingFace to pad a batch of tokenized inputs.

My use case

I'm following the HuggingFace-style batching process for fine-tuning, where .pad() is typically used to ensure consistent input shapes. But the GLM4 tokenizer appears to lack support for padding_side, and perhaps even .pad() behavior in general.

What I’ve tried

• Looked into the tokenizer code — it seems that GLMTokenizer does not inherit the usual pad() method behavior from PreTrainedTokenizerFast.
• Tried manually padding the input sequences, but I’m concerned about whether that matches GLM4’s expected behavior, particularly for attention_mask and position_ids.

Questions

1. What's the recommended way to apply padding when using GLM4 tokenizer?
2. Is there a compatible data collator or tokenizer wrapper that supports HuggingFace-style padding?
3. Would manually implementing padding + masks be sufficient, or is there a better way to ensure compatibility?

Environment:

• GLM model: GLM4
• Transformers version: latest
• OS: Ubuntu

Thanks a lot for your help!