Use `dtype` instead of deprecated `torch_dtype` for transformers >= 4.56

[xyf5432]

In train.py lines 331 and 344, ShowUIForConditionalGeneration.from_pretrained and Qwen2VLForConditionalGeneration.from_pretrained are called with torch_dtype=torch_dtype (e.g. torch_dtype=torch_dtype, low_cpu_mem_usage=True, _attn_implementation=..., quantization_config=bnb_config, device_map=...).

The torch_dtype keyword argument was deprecated in transformers 4.56 (PR #39782) and replaced by dtype. On transformers 4.56+ these calls emit a DeprecationWarning, and the argument will be removed in a future release, breaking training.

Suggested fix: choose the keyword based on the installed transformers version with packaging.version:

import transformers
from packaging.version import Version

def _dtype_kwargs(dtype):
    """`dtype` keyword of `from_pretrained` exists since transformers 4.56 (PR #39782);
    older versions use `torch_dtype`."""
    if Version(transformers.__version__) >= Version("4.56"):
        return {"dtype": dtype}
    return {"torch_dtype": dtype}

model = ShowUIForConditionalGeneration.from_pretrained(
    model_url,
    **_dtype_kwargs(torch_dtype),
    low_cpu_mem_usage=True,
    _attn_implementation=args.attn_imple,
    quantization_config=bnb_config,
    device_map=f"cuda:{args.local_rank}",
    lm_skip_layer=lm_skip_layer,
    lm_skip_ratio=args.lm_skip_ratio,
)

(Same pattern for the Qwen2VLForConditionalGeneration call on line 344.) This keeps compatibility with transformers < 4.56 and stops the deprecation warning on 4.56+.