协作头脑风暴
brainstorming
在实现前澄清需求、探索设计选择,并形成清晰的开发方向。
moe-and-acceleration【开发与工程】Use this repo skill for Swin-Transformer optional Swin-MoE, Tutel, Apex, fused window-process CUDA extension, distributed checkpoint, and backend probe workflows.
独立 GitHub 快照 · 自动分类 · 以原作者文档为准
Use this sub-skill when a task involves Swin-MoE, Tutel, optional Apex fused layernorm/optimizers, --fused_window_process, the CUDA window-process extension, or distributed MoE checkpoint behavior.
main_moe.py command shape and distributed expectations.MODEL.SWIN_MOE config fields: experts, top-k, capacity factor, router type, aux loss, and MoE block placement.swin_window_process CUDA extension build/probe guidance.fused_adam, fused_lamb.core-models.main.py commands: training-eval-cli.data-and-checkpoints.simmim-workflows.scripts/check_optional_backends.py to see which optional backends are actually available.references/moe-and-acceleration.md for command and config rules.scripts/swin_cli_command_builder.py --workflow moe-train or moe-eval.references/troubleshooting.md and avoid claiming runtime verification.CPU import checks are not evidence that Tutel MoE, Apex fused ops, or the custom CUDA extension works. Treat those as optional backend capabilities that require compatible CUDA/PyTorch/toolkit/runtime evidence.
references/moe-and-acceleration.md - MoE and optional acceleration workflow details.references/troubleshooting.md - missing Tutel/Apex/fused-kernel and distributed checkpoint issues.scripts/check_optional_backends.py - safe optional backend probe.