For boolean checks, it's more idiomatic in Python to use the truthiness of the value directly rather than comparing withFalse. Theload_in_fp8 parameter can beTrue,False, or a string like'block'. BothTrue and non-empty strings are truthy, whileFalse is falsy. Usingif load_in_fp8: is more concise and readable, and achieves the same result.

Yeah agree with gemini here :)

sure I can change it, I just had it this way because I saw that's how Daniel wrote it in a few existing places

Similar to my other comments, this check can be made more Pythonic. Instead ofif load_in_fp8 != False:, you can useif load_in_fp8:. This leverages Python's truthiness evaluation and is generally considered better style for readability and conciseness.

This checkif load_in_fp8 != False: can be simplified toif load_in_fp8:. This is the more idiomatic and preferred way to check for truthiness in Python, improving code readability.

Enforce FP8/4bit mutual exclusion for vLLM >=0.12

Whenload_in_fp8 is true,get_model_name now returns the original name as soon as vLLM ≥ 0.12.0 (loader_utils.py lines 110-118), so thenew_model_name is None branch here is never taken and_get_fp8_mode_and_check_settings no longer runs. With the defaultload_in_4bit=True, the code now proceeds to fast inference with bothload_in_fp8 anduse_bitsandbytes=load_in_4bit set, even though_get_fp8_mode_and_check_settings used to reject FP8 together with 4/8/16-bit loads. This yields conflicting quantization paths (fp8 on-the-fly plus bitsandbytes 4bit) and is likely to fail at runtime for users who simply enableload_in_fp8 without also disabling 4bit.

Useful? React with 👍 / 👎.