fix: Handle tuple and set types in HfArgumentParser

[dajiaohuang]

Description

HfArgumentParser was not handling tuple and set dataclass fields correctly. When parsing --layers 4,5 for a tuple[int, int] field, argparse would call tuple("4,5") which iterates character-by-character, producing ('4', ',', '5') instead of (4, 5).

The fix adds handling for tuple and set types similar to list:

• tuple[int, int] (fixed-length): nargs=2, type=int
• tuple[int, ...] (variable-length): nargs="+", type=int
• set[str]: nargs="+", type=str

Example

from dataclasses import dataclass, field
from transformers import HfArgumentParser

@dataclass
class ModelArgs:
    layers: tuple[int, int] = (1, 2)
    tags: set[str] = field(default_factory=set)

parser = HfArgumentParser(ModelArgs)

# Before fix: layers=('4', ',', '5'), tags={'1', '2', ',', 'a', 'g', 't'}
# After fix: layers=(4, 5), tags={'tag1', 'tag2'}
args = parser.parse_args_into_dataclasses(["--layers", "4", "5", "--tags", "tag1", "tag2"])[0]

Related Issue

Fixes #48419

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 33615004147
Result: success | Grafana metrics are not available yet.