Changelog

User-visible changes only. Internal refactors and infrastructure work omitted.

v1.0 — Launch#

First generally-available release.

• REST submission for seven training types: SFT, LORA, QLORA, DPO, RLHF, CONTINUED_PRETRAIN, FULL_FINETUNE.
• Five framework targets: HF_TRAINER, DEEPSPEED, FSDP, MEGATRON, CUSTOM.
• Full lifecycle: submit, cancel, pause, resume, retry (with optional resource modification).
• Point-in-time job metrics plus Server-Sent Event streams for live logs and metric snapshots.
• Cluster operations: status, node listing, node drain (force-able) and enable, queue depth.
• Evaluation submission and retrieval, with multiple named benchmarks per run.
• Data operations: pre-flight validation of training sources, coordinator-side dataset cache inspection.
• Local job-state cache (mod_scaimind_jobs) auto-synced on list calls for fast dashboard reads.
• Downstream token forwarding for ScaiDrive and ScaiAtlas, scoped per request.
• gRPC error sanitisation — INTERNAL and UNKNOWN coordinator details are replaced with friendly messages and captured in structured logs.
• Admin UI: Training Dashboard, Job Creator, Training Monitor, Evaluation Center, Hardware Monitor.