v1.2.23

1.2.23 (2026-10-12)

Full Changelog: v1.2.22...v1.2.23

Features

• training-sdk: record each completed sampling attempt's client-measured latency and generated-token count in ServerMetrics, so callers can compare client and serving time from DeploymentSampler.drain_metrics()
• training-sdk: DeploymentSampler.drain_metrics() now returns one record per completed or failed attempt: every record carries its 1-based retry_attempt, and failed attempts (including non-retryable HTTP errors, truncated streams and responses the parser rejects) set the new ServerMetrics.error

Bug Fixes

• dependencies: update runtime and development locks to multidict 6.9.1 to fix a native reference leak
• training-sdk: bound the per-attempt ServerMetrics buffer kept for drain_metrics(); samplers that are never drained no longer grow memory for the life of the run
• training-sdk: measure client_latency and client_ttft from the completion request itself with a monotonic clock, excluding the shared R3 format probe and a pre-generation fallback
• training-sdk: draining metrics no longer clears the recent history that sampler timeout diagnostics report; that history keeps only attempts with timings, for at most ten minutes