Convert torch.bfloat16, torch.float16, etc. to vLLM valid dtypes

[rishabh135]

Convert torch.bfloat16, torch.float16, etc. to valid CLI string

Checks dtype params when building the subprocess_command for vllm.

Passing torch.bfloat16 values seems to stop vllm from deploying properly, and result in

vllm serve: error: argument --dtype: invalid choice: 'torch.bfloat16' (choose from 'auto', 'bfloat16', 'float', 'float16', 'float32', 'half')

The update makes it such that vLLM will deploy even when the correct dtype isn't specified isn't specified in the model name

[danielhanchen]

Oh thank you! Could you remove some extra spaces and newlines - thanks :)

[rishabh135]

removed newline characters as requested. Kindly merge this submission!

[danielhanchen]

Thank you!