✨ log all errored requests

[joerunde]

Drafting some changes to log all errors during validation or generation for a request like TGIS does. The goal is to get something like this:

2024-05-17T19:24:20.591834Z ERROR generate{input=["Hallo, wie heißt du? ..."] prefix_id=None correlation_id="<none>" input_bytes=[22] params=Some(Parameters { method: Sample, sampling: Some(SamplingParameters { temperature: 0.0, top_k: 3, top_p: 0.0, typical_p: 0.0, seed: None }), stopping: Some(StoppingCriteria { max_new_tokens: 20, min_new_tokens: 16, time_limit_millis: 0, stop_sequences: ["Peter ", "Timothy ", "joseph", "Corinthians"], include_stop_sequence: None }), response: Some(ResponseOptions { input_text: false, generated_tokens: false, input_tokens: false, token_logprobs: false, token_ranks: false, top_n_tokens: 0 }), decoding: Some(DecodingParameters { repetition_penalty: 2.0, length_penalty: Some(LengthPenalty { start_index: 0, decay_factor: 25.0 }) }), truncate_input_tokens: 0 })}: text_generation_router::grpc_server: src/grpc_server.rs:429: length_penalty must be >= 1.0 and <= 10.0

though probably without the file name/line number stuff (text_generation_router::grpc_server: src/grpc_server.rs:429 in the example above)

[njhill]

Thanks @joerunde! only minor comments.

It feels like it would be better for the response logging to happen in one place (whether success or error), not sure how realistic that is but in any case can save for possible later refactor.

[joerunde]

Thanks for the review @njhill, everything looks like it still works!

[njhill]

Thanks @joerunde!