[abhishekcpp007]

Description

sample_logits checks freq_penalty, temperature and repetition_penalty before it branches,
so they raise at any temperature. top_k and top_p were checked inside the sampling branch
instead, which greedy sampling never enters. A call with temperature=0.0 therefore accepted any
filter value without complaint.

Measured on dev before this change, with logits = [[1.0, 3.0, 2.0, 0.5]]:

call	before
temperature=1.0, top_k=0	AssertionError: top_k has to be greater than 0
temperature=1.0, top_k=-5	AssertionError: top_k has to be greater than 0
temperature=1.0, top_p=0.0	AssertionError: top_p has to be in (0, 1]
temperature=1.0, top_p=1.5	AssertionError: top_p has to be in (0, 1]
temperature=0.0, top_k=0	accepted, returns [1]
temperature=0.0, top_k=-5	accepted, returns [1]
temperature=0.0, top_p=0.0	accepted, returns [1]
temperature=0.0, top_p=1.5	accepted, returns [1]

freq_penalty=-1 raises at both temperatures, which is the behaviour the filters should match.

Someone sweeping a sampling config over temperatures gets a silent pass on the greedy rows and a
failure everywhere else, from the same arguments.

What changed

The two checks move above the temperature == 0.0 branch, next to the other parameter checks.
They still raise AssertionError with the same messages, so callers and existing tests see the
same exception type; the checks simply also fire for greedy sampling now. Clamping top_k to the
vocabulary size stays where it was, since it needs the logits' shape.

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

test_top_k_rejects_non_positive and test_top_p_rejects_out_of_range now run over
temperature of 1.0 and 0.0, matching how test_non_positive_penalty_raises and
test_negative_freq_penalty_raises already cover both. They also assert the message rather than
just the type. Both fail on dev for the 0.0 rows.

[jlarson4]

Looks good @abhishekcpp007, merging