[huggingface] NameError: name 'AutoConfig' is not defined in download_model()

Bug

Running any HuggingFace benchmark model that goes through download_model()
fails with:

File "torchbenchmark/util/framework/huggingface/basic_configs.py", line 304, in download_model
    config = eval(HUGGINGFACE_MODELS[model_name][2])
NameError: name 'AutoConfig' is not defined

Root Cause

basic_configs.py line 303:

exec(f"from transformers import {config_cls_name}")   # line 303
config = eval(HUGGINGFACE_MODELS[model_name][2])      # line 304 — NameError

In Python 3, exec() without an explicit globals dict imports into the
local scope of the enclosing function. The subsequent eval() resolves
names against module globals, not local scope — so AutoConfig (or any
other imported class) is invisible to it.

This is a well-known Python 3 scoping behaviour:
https://docs.python.org/3/library/functions.html#exec

Fix

exec(f"from transformers import {config_cls_name}", globals())
config = eval(HUGGINGFACE_MODELS[model_name][2])

Passing globals() makes the imported name available to eval().

Affected Models

All HuggingFace models whose config constructor string in
HUGGINGFACE_MODELS uses a class that must be imported (e.g.
AutoConfig.from_pretrained(...), LlamaConfig(...), etc.).

Reproduction

# Minimal reproduction of the bug
def broken():
    exec("from transformers import AutoConfig")
    return eval('AutoConfig.from_pretrained("albert-base-v2")')  # NameError

def fixed():
    exec("from transformers import AutoConfig", globals())
    return eval('AutoConfig.from_pretrained("albert-base-v2")')  # works

Environment

• Python 3.11+
• Any recent transformers version

Start in torchbenchmark/util/framework/huggingface/basic_configs.py around lines 303-304 and reproduce the failure through download_model(). Check the interaction between the dynamic import and eval(), then verify that affected HuggingFace models no longer raise NameError when their config constructor is evaluated.