Server docs/openapi.yaml returns bytes[] per token in Logprob/TokenLogprob, but ollama/_types.py TokenLogprob only has token/logprob (plus top_logprobs on Logprob), so raw token bytes are silently dropped on generate(..., logprobs=True) / chat(..., logprobs=True).
Expected: optional bytes preserved (backward compatible, defaults None), matching #601 logprobs support.
Server docs/openapi.yaml returns bytes[] per token in Logprob/TokenLogprob, but ollama/_types.py TokenLogprob only has token/logprob (plus top_logprobs on Logprob), so raw token bytes are silently dropped on generate(..., logprobs=True) / chat(..., logprobs=True).
Expected: optional bytes preserved (backward compatible, defaults None), matching #601 logprobs support.