Temperature 0 still gives different answers — is that expected?

Same prompt, temperature: 0, five runs, three slightly different answers. I assumed zero meant deterministic.

Zero means greedy, not deterministic. Batched inference on GPUs can change floating-point order between runs, which occasionally flips a near-tie between two tokens. Use the seed parameter for best-effort reproducibility and compare system_fingerprint across runs.

Makes sense. With seed I get the same answer 5/5 now.

Same here — following this thread.