sampler: greedy paths report log of raw logit instead of log of probability

[dx2102]

Symptom

Greedy request with logprobs=5 returns positive numbers:

'were'  logprob=+1.322
'was'   logprob=+1.314

Positive values cannot be natural log probabilities since ln(p) <= 0 for p <= 1. After #2327 lands these become ln(raw_logit) instead of log10(raw_logit); neither is ln(softmax(logits)[token]).

Root cause

Sampler::sample has two None (greedy) arms:

• non-speculative greedy passes the post-penalty logits into sample_argmax
• speculative greedy passes the post-penalty logits into sample_speculative_top_kp_min_p

Neither arm applies softmax_last_dim, while both Some(temperature) arms do. Both callees then bind the input to let probs: Vec<f32> = logits.to_vec1()?; and treat that variable as a probability distribution: .ln() at the end (both functions), plus top-p and min-p filters that assume values in [0, 1] (only sample_speculative_top_kp_min_p).

The sampled token is usually still correct because argmax_f32 is invariant under softmax. The wrong logprob value is only surfaced to the caller when return_logprobs=true; otherwise it is filled and never read.

Repro

import requests
r = requests.post("http://127.0.0.1:8766/v1/completions", json={
    "model": "Qwen/Qwen3-0.6B",
    "prompt": "Once upon a time, there",
    "max_tokens": 1,
    "temperature": 0.0, "top_p": 1.0, "top_k": 1,
    "logprobs": 5,
})
for t in r.json()["choices"][0]["logprobs"]["content"][0]["top_logprobs"]:
    print(f"  {t['bytes']!r:12}  logprob={t['logprob']:+.4f}")

HF transformers log_softmax reference on the same model and prompt:

'were'   ln(prob)=-0.809
'was'    ln(prob)=-1.059
'are'    ln(prob)=-1.934

Other issues found in sampler.rs while looking at this

• sample_speculative_top_kp_min_p has "sample" in its name but its last step is argmax_f32(&probs). Its top-k, top-p, min-p filters therefore never change the sampled token, only the probability distribution returned to the user.
• Both sample_speculative_top_kp_min_p and sample_top_kp_min_p apply all three sampling filters (top-k, top-p, min-p) to their local probs variable, then pass that filtered variable into get_top_logprobs. The API's top_logprobs list (the user-requested count of alternatives) is therefore drawn from post-filter values rather than from the model's real distribution. On temperature paths that filtered variable is a real probability distribution; on the greedy speculative path (see primary root cause) it is post-penalty logits treated as probs and filtered by comparing logit magnitudes against thresholds in [0, 1], so top_logprobs there is doubly wrong: neither derived from probabilities nor filtered on meaningful values.
• sample_speculative_top_kp_min_p's min-p block computes threshold = max_p * min_p. After penalty application probs can be negative; when max_p is negative, threshold > max_p, so the check zeros every entry including the max, and the subsequent argmax_f32 falls back to index 0. In this corner case the sampled token itself is wrong, not just the reported logprob.
• sample_speculative_top_kp_min_p lacks the "top_p / min_p out of range means early return to multinomial" guards that sample_top_kp_min_p has.
• test_gumbel_speculative has no Gumbel sampling in the tested code path.