[Fix] conversation_to_ids: truncation guard crashes with "'list' object has no attribute 'shape'" (#581)

[yushuosun]

Motivation

Finetuning crashes on any sample longer than max_length with AttributeError: 'list' object has no attribute 'shape', reported in #581 (and #578). The bug is still present on current main (8a2db68).

Root cause

In finetune/dataset.py, conversation_to_ids() hstacks the per-segment arrays into the tensor ids:

ids = torch.from_numpy(np.hstack(input_ids, dtype=np.int32))

but the over-length guard still inspects input_ids, which is the raw Python list returned by conversation_to_ids_minicpm/llama3/qwen2 and has no .shape:

if input_ids.shape[-1] > max_length:                                   # line 147
    ids = ids[:max_length]
    context = context[:max_length]
    logger.warning(f"The input length ({input_ids.shape[-1]}) ...")     # line 150

So instead of truncating and warning, the run aborts the moment a tokenized sample exceeds max_length (default 2048).

Modifications

finetune/dataset.py: use the ids tensor in both the guard (line 147) and the warning message (line 150). Two-token change; the only behavioral effect is that the intended truncation path now works.

Duplicate-check

• Track A — open PRs referencing the issue:

  gh api "search/issues?q=repo:OpenBMB/MiniCPM-V+is:pr+581+in:body"  ->  no open PR
  

  Other open PRs touching finetune/ do not touch this block: fix(docs): correct LLM_TYPE for MiniCPM-V-4 from 'llama' to 'llama3' #1088 (docs), Update trainer.py #1010 (trainer.py), feat: Added judgment logic to support training with plain text data. #281 (__len__/data_collator).
• Prior PR fix finetune minicpm error #579 proposed a partial fix (line 147 only) but was closed unmerged by its author; line 150 still carries the identical bug.
• Track B — issue thread has no open/in-progress claim.

[Copilot]

Pull request overview

Fixes a finetuning crash in conversation_to_ids() when tokenized samples exceed max_length by ensuring the truncation guard operates on the stacked tensor rather than the original Python list of segments.

Changes:

• Switches the over-length check from input_ids.shape (invalid for lists) to ids.shape.
• Updates the warning message to reference ids instead of input_ids.

---

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

[tc-mb]

Hi @yushuosun, thanks for the PR.

Just a heads-up — the finetune code in the main MiniCPM-V repo is no longer actively maintained. The official finetune scripts have been moved to the new MiniCPM-V Cookbook repo:

• Repo: https://github.com/OpenSQZ/MiniCPM-V-Cookbook
• Corresponding file: finetune/official/dataset.py
• CookBook docs: https://opensqz.github.io/MiniCPM-V-CookBook/site/en/shared/finetune/official.html

The conversation_to_ids function in the Cookbook needs the same verification and fix — you can confirm over there.

Two options:

1. You submit the PR to the Cookbook repo: patch finetune/official/dataset.py in OpenSQZ/MiniCPM-V-Cookbook. The PR and commits remain entirely your contribution.
2. We submit it on our end: if we file it, we can only add a note in the commit message crediting you, which may not fully reflect your contribution.

Which would you prefer?