跳转至

INDEX

按任务选一篇。找不到就先读 map/repos.md,再决定新建哪一类。

任务 打开
不知道该改哪个仓 map/repos.md
编译→量化→部署→推理怎么串 map/pipeline.md
Thor / Orin / Ada / 目标硬件 map/platforms.md
TVM / Relax / TRT 怎么编 domains/compile/README.md
TVM GQA broadcast_to + RoPE 写回拖慢 decode domains/optimize/tvm-gqa-rope-materialize.md
FP8 / NVFP4 / AWQ / 标定 domains/quantize/README.md
融合、few-core、kernel、图优化 domains/optimize/README.md
engine / 车端包 / 产物布局 domains/deploy/README.md
runtime、对拍、延迟 domains/infer/README.md
pi05 多后端 models/pi05/README.md
FlashRT 对 PI05 的 ViT / LLM / expert / 调度对照 models/pi05/flashrt-opt-map.md
TensorRT-Edge-LLM 对 Thor 该搬核策略、不换栈 models/pi05/trt-edge-llm-opt-map.md
Thor GEMM 按 SM 数分 tile,写死 128/256 打输小 M lessons/thor-sm-count-tile-dispatch.md
PI05 视觉 INT8 / QKV NVFP4 / 1-view NVFP4 不要当默认 lessons/pi05-flashrt-quant-dont.md
cosmos3 / cosmos3-edge models/cosmos3/README.md
精度对点从哪进 playbooks/precision-align.md(入口仓 Chameleon)
跑 eval / compare / bench playbooks/eval-compare.md
π0.5 当前 TVM vs TRT 缺口在哪 lessons/pi05-tvm-trt-gap-nsys.md
π0.5 TVM FP8 权重必须静态、还不是 tensor core lessons/pi05-tvm-fp8-static-not-tc.md
Thor 延迟测前(空闲 + MAXN + 锁时钟) lessons/thor-bench-maxn-idle.md
CUTLASS attention_bias 被错派到 flash(pad mask 无效) lessons/cutlass-attention-bias-not-flash.md
把 FlashRT GeGLU / GeGLU+Q 编进 TVM playbooks/flashrt-geglu-tvm.md
nn.SourceModule 接多输出 extern(shape 函数返回 list 会报错) lessons/tvm-sourcemodule-multi-output.md
拿 nsys 对比两个栈前先折成同 iter 口径(否则会打错地方) lessons/nsys-same-iter-attribution.md
接一个新模型进栈 playbooks/onboard-model.md
跨仓接口以谁为准 contracts/index.md
查已拍板结论 decisions/README.md
查踩坑 lessons/README.md