feat: Add support for accelerate-wrapped models in simple_evaluate()

[DhruvaKashyap]

Summary

Adds HFLM_Accelerate class to support models that are already wrapped with accelerate, enabling the use of simple_evaluate() with distributed and memory-optimized model instances.

Problem

Currently, simple_evaluate() only accepts models as strings or base objects, making it incompatible with models wrapped using HuggingFace's accelerate library. This limits the library's utility for distributed evaluation scenarios. Addresses Issue #3312.

Solution

• Added new HFLM_Accelerate class in lm_eval/models/huggingface.py
• Accepts pre-initialized PreTrainedModel, PreTrainedTokenizerBase, and Accelerator instances
• Inherits from HFLM but skips model loading since model is already initialized
• Maintains full compatibility with existing evaluation workflows

Usage Example

from accelerate import Accelerator
from transformers import AutoModel, AutoTokenizer
from lm_eval.models.huggingface import HFLM_Accelerate

accelerator = Accelerator()
model = AutoModel.from_pretrained("model_name")
tokenizer = AutoTokenizer.from_pretrained("model_name")

# Prepare model with accelerate
model = accelerator.prepare(model)

# Use with lm-evaluation-harness
hf_model = HFLM_Accelerate(
    pretrained=model,
    tokenizer=tokenizer, 
    accelerator=accelerator
)

[CLAassistant]

All committers have signed the CLA.

[DhruvaKashyap]

Hoping that this would be a useful commit to the repository and looking forward to hearing feedback.