| # Cluster environment backups | |
| Per-cluster runtimes for Qwen3.5 / Nemotron post-training. Public on purpose: | |
| these are recipes and build artifacts, no credentials. | |
| Each cluster directory holds `manifest.json` (python, torch, CUDA, driver, GPU, | |
| `torch.cuda.get_arch_list()`), `pip-freeze-*.txt`, the build scripts, and any | |
| expensively-built CUDA extensions under `overlays/`. | |
| ## Arch compatibility, measured | |
| b200 is compute capability 10.0, b300 is 10.3. Neither cluster's torch lists | |
| `sm_103`, yet a torch built for `sm_100` runs correctly on a B300: matmul, | |
| transformer-engine Linear and causal_conv1d CUDA ops were all verified on a | |
| B300 SXM6. Blackwell `sm_103` is binary compatible with `sm_100` cubins, so | |
| these artifacts are portable across b200 and b300. | |
| ## Overlay rule | |
| Install overlays with `pip install --target ... --no-deps` and prepend to | |
| PYTHONPATH. Without `--no-deps`, pip drags PyPI torch and triton into the | |
| overlay and silently shadows the cluster runtime. | |