Replace deprecated torch_dtype with dtype in llava_hf.py

[xyf5432]

Summary

Starting with transformers 4.56.0 (huggingface/transformers#39782), the torch_dtype keyword argument of PreTrainedModel.from_pretrained is deprecated in favor of dtype. Code that still uses torch_dtype emits the deprecation warning "torch_dtype is deprecated! Use dtype instead!" on every invocation, and the deprecated API is scheduled for removal in a future major release.

This repository still passes torch_dtype at the location below. Note that requirements/runtime.txt allows transformers >= 4.45.2, i.e. older versions that do not accept dtype yet. The proposed change therefore keeps a fallback for transformers < 4.56.0 so the code keeps working across the whole supported range: dtype is honored identically to torch_dtype on transformers >= 4.56.0.

Trigger point

• llmc/models/llava_hf.py line L30-33:

  self.vlm_model = LlavaForConditionalGeneration.from_pretrained(
      self.model_path,
      config=self.vlm_model_config,
      torch_dtype=self.torch_dtype,
      low_cpu_mem_usage=True,
  )

Proposed change

try:
    # transformers >= 4.56: `dtype` is the replacement for `torch_dtype`
    self.vlm_model = LlavaForConditionalGeneration.from_pretrained(
        self.model_path,
        config=self.vlm_model_config,
        dtype=self.torch_dtype,
        low_cpu_mem_usage=True,
    )
except TypeError:
    # transformers < 4.56: `dtype` is not accepted yet
    self.vlm_model = LlavaForConditionalGeneration.from_pretrained(
        self.model_path,
        config=self.vlm_model_config,
        torch_dtype=self.torch_dtype,
        low_cpu_mem_usage=True,
    )

Would a PR updating this call to dtype (with the fallback above) be welcome?