ultravox/training/train.py (2 lines):
	- line 241: # TODO (Farzad): reconsider for multi-node
	- line 247: # tf32=dtype == torch.float32 and device.type == "cuda",  # TODO: check for Ampere GPU not just CUDA


ultravox/tools/infer_api.py (1 line):
	- line 143: ]  # TODO: change regarding multiple messages?


ultravox/inference/oaieval_registry/solvers/vllm_server_config.yaml (1 line):
	- line 11: # TODO: investigate and try to match HF/VLLM results without this


ultravox/training/ddp_utils.py (1 line):
	- line 42: # TODO: handle drop last gracefully


ultravox/training/configs/llama_70b.yaml (1 line):
	- line 4: # TODO: make sure to increase max_duration in mcloud.yaml to 18 hours instead of 6


ultravox/model/ultravox_processing.py (1 line):
	- line 242: # TODO: Add support for multiple text inputs.


ultravox/model/ultravox_data_proc.py (1 line):
	- line 75: # TODO: this might be slow due to calling audio_processor twice. We can compute modified input_text_len directly too.


ultravox/model/ultravox_model.py (1 line):
	- line 485: # TODO: refactor common parts to a shared module