List view
Long-context SFT/QLoRA efficiency, template-safe response masking, and staged alignment-training expansion.
No due date•1/1 issues closedSingle-stream multimodal decode optimization, per-image feature reuse, and native quantized multimodal load paths.
No due date•0/1 issues closedRequest-scoped stream parsing, deterministic channel assembly, and continuity-safe reasoning handling for repeated turns.
No due date•0/1 issues closedPrefix reuse, partial-prefix resume, deferred cleanup, and scheduler-safe cache behavior for repeated local sessions.
No due date•0/1 issues closed