Block a user
Per-account top-up overrides, so one workload's appetite doesn't set everyone's limits
S1 — sampling defaults from generation_config, plus top_k, seed and tunable repetition
[epic] helexa.ai/games — nightly generative benchmark scored by people playing the results
[epic] Sampling correctness: honour the caller's parameters and the model's own generation_config
Converted to an epic — children filed
Sized against the code: 8 LogitsProcessor construction sites, 4 places
extracting sampling fields from a request, 6 signatures threading them as
loose…
S3 — stop lying about unsupported parameters: reject, and advertise what we support
S2 — stop sequences, and the penalties that need per-token logit work
S1 — sampling defaults from generation_config, plus top_k, seed and tunable repetition
Sampling correctness: standard inference parameters are silently ignored, and the model's own generation_config is never read
Prefix cache savings are invisible:
cached_tokens is never reported, so every client shows 0% cache hit
A streamed turn logs nothing when it ends, so it cannot be diagnosed afterwards
/no_think is honoured on chat/completions but ignored on /v1/responses
Measured: with no working control, reasoning expands to fill any budget
This issue's third point — populate incomplete_details — is no longer
the cosmetic half. Three runs today against…
Prefix-cache reuse looks low on long agentic contexts — 2068 tokens reused against a 42K prompt