Activate model on GPU (WARM → ACTIVE).
Per-token log-probability evaluation result.
std::vector< float > logprobs
Log-prob for each token transition (N-1 values)
int n_logprobs
Number of logprob values (n_tokens - 1)
int n_tokens
Number of input tokens.
float total_logprob
Sum of all logprob values.
float perplexity
exp(-mean(logprobs)) — lower = less surprising
std::vector< int32_t > tokens
Input tokens echoed back for verification.