[BFCL] Adds support for parallel inference and batching

[TikZSZ]

Parallel Inference Support for berkeley-function-call-leaderboard

This PR adds support for running berkeley-function-call-leaderboard inference in parallel, reducing running time by 4x or more depending on --batch-size.

Changes

Modifies berkeley-function-call-leaderboard/model_handler/handler.py

• Modified write function to make it async using aiofiles
• Added sort_results function to sort the results based on idx after each individual test_cate is over
• sort_results function returns the indices after sorting, supporting resuming functionality

Modifies berkeley-function-call-leaderboard/openfunctions_evaluation.py

• Added --batch-size arg, defaults to 1 -> controls number of parallel requests
• Refactored processing and result-writing logic to fetch_and_process function
• Added make_async function to wrap sync functions as async (used for handler.inference)
• Added nested progress bar for tracking iterations
• Refactored core logic for processing under main function
• Implemented proper resume support, replacing num_existing_lines

Resume Support

Improved resuming functionality in async code:

• Addresses potential issues where some test cases could complete earlier than others, leading to inconsistent resume.
• Filters already saved test cases instead of using a simple line count
• For saved test cases, adds None as a placeholder, which becomes the conditional for skipping test cases
• This approach ensures consistent resuming even if execution is interrupted mid-test
• This is really important for models that are expensive to run, where re-running the whole test again is undesirable
• A test log screenshot is also uploaded bottom of this PR to confirm that it works as intended

Note: This PR automatically wraps inference calls as async to minimize code changes, but the calls are still synchronous and will block the event loop so we use loop.run_in_executor to run the calls in parallel on multiple threads instead it uses min(32, os.cpu_count() + 4) threads by default. If handlers are made async in future they would continue to work like normal async code

Testing

Tested on a custom OpenAI-compatible model running on vllm:

• Completed simple test in 40 seconds
• Hardware: RTX4090
• Model: LLAMA 8B BF16
• Batch size: 15-20

Benchmark Results

Debug Logs for new Resume System