从头开始独立实现 Qwen-14B 步骤 (没有 PyTorch、没有 cuBLAS、自行编译器) — 什么匹配、成本如何、以及 LoRA 形状的发现

**Title:** Independent from-scratch Qwen-14B step reimplementation (no PyTorch, no cuBLAS, own compiler) — what matched, what it cost, and a LoRA-shape finding Not a bug report or a support request. We reimplemented the Qwen-14B forward/training step outside the PyTorch stack — own compiler, own PTX backend, own GEMM kernels, no cuBLAS call — as part of a self-hosted-language project. That project is now shelved, so I'm writing up the parts that seem useful to people working on Qwen rather than letting them sit in a dead repo. Close this if it's off-topic for your tracker.