Fix unintuitive --gen_kwargs behavior

[haileyschoelkopf]

This PR removes an override we had that would set do_sample=False if no value for it were set.

Users using --model hf with --gen_kwargs in the CLI will still have to pass --gen_kwargs do_sample=True alongside their other sampling parameters for tasks where greedy generation is expected, still, but this should lead to less weird unintuitive behavior where the gen kwargs seem not respected.

Addresses #1274 but won't close that issue yet.

[baberabb]

oh. Totally forgot about this CLI command in #1341 (and was mirroring HF). Maybe change the condition to something like:

do_sample = kwargs.pop("do_sample", None)
if do_sample is False or "temperature" not in kwargs:
    kwargs["temperature"] = 0.0

lm-evaluation-harness/lm_eval/models/vllm_causallms.py

Lines 439 to 441 in 081deb8

	do_sample = kwargs.pop("do_sample", False)
	if do_sample is not True:
	kwargs["temperature"] = 0.0

[baberabb]

made a PR #1345 with this as well as another nit.

[haileyschoelkopf]

I think actually, we want to still pass do_sample to HF, as it actually uses this parameter--just that we don't want both temperature=0.0 and do_sample=False both on, as this results in warning messages.

[haileyschoelkopf]

@lintangsutawika ran linters on this, should be good to go!