fix(calibration): interpolate the item index in the unsupported-sequence error

[Anai-Guo]

Problem

In prepare_calibration_dataset, the helper _tokenize_messages_value raises three
ValueErrors that all report which calibration item failed. Two of them are f-strings;
the middle one is a plain string literal, so its {idx} is emitted verbatim:

gptqmodel/utils/calibration.py

if templated is None:
    raise ValueError(f"tokenizer.apply_template returned None for calibration item {idx}.")   # L244  f-string
...
if isinstance(templated, (list, tuple)):
    if len(templated) > 0 and isinstance(templated[0], int):
        return _pack_ids(list(templated), None, idx)
    raise ValueError(
        "tokenizer.apply_template returned an unsupported sequence type for calibration item {idx}."  # L262  no f
    )
...
raise ValueError(
    f"tokenizer.apply_template returned unsupported type {type(templated)} for calibration item {idx}."  # L272  f-string
)

Rendering the two literals with idx = 3:

L244 (sibling) -> tokenizer.apply_template returned None for calibration item 3.
L262 (this fix) -> tokenizer.apply_template returned an unsupported sequence type for calibration item {idx}.

The index is the only actionable part of the message — it is what tells the user which
row of their calibration dataset to look at. This branch fires when a custom
apply_template returns a non-empty sequence whose first element is not an int
(e.g. a list of token strings), and the user is left without the item number.

Neither ruff nor flake8 flags this: F541 covers f-strings with no placeholders, not
the reverse.

Fix

Add the missing f prefix. One character, no behaviour change beyond the message text.

Note on AGENTS.md

AGENTS.md routes tokenizer/chat-template behaviour changes to Tokenicer. This patch
changes no tokenization, prompt rendering, or normalization logic — it only makes an
existing diagnostic interpolate the value it already names — so it stays in
GPT-QModel.

🤖 Generated with Claude Code