@xiciyang

Quantize a checkpoint and prove it in DSH: PTQ with a leak-free calibration split, independent accuracy and performance gates through real serving endpoints, auto-degrade, signed-off audit or rollback.

Deploy a checkpoint as a reachable inference endpoint in DSH: probe, backend select, three-stage health check, registration, auto-degrade (memory/concurrency/precision/port), rollback, teardown — hash-chained audit log. 154 tests.

Lossless visual memory for DSH agents: store an image once, address it by sha256, read exact pixels, compare past frames, recall by text/tag/turn — no second vision model in the loop.

Find the root-cause wall, apply the top intervention, prove it with a two-sided gate — or roll back. Engine- and hardware-neutral: no implicit defaults, spec-sheet values are never load-bearing.

Diagnose an inference service from evidence: generate a sanitized collection script, parse it offline, rank root causes with falsifiers, re-collect only what is missing. Hash-chained audit; five read-only DSH tools.

@xiciyang/training-pretrain — DSH 训练插件：可独立安装，依赖 @xiciyang/* 共享库

@xiciyang/training-sft — DSH 训练插件：可独立安装，依赖 @xiciyang/* 共享库

@xiciyang/training-qat — DSH 训练插件：可独立安装，依赖 @xiciyang/* 共享库

@xiciyang/training-rl — DSH 训练插件：可独立安装，依赖 @xiciyang/* 共享库