| --- |
| license: other |
| license_name: mixed-third-party-runtime-licenses |
| pretty_name: GOLD B200 portable runtime |
| tags: |
| - runtime |
| - gold |
| - knowledge-distillation |
| --- |
| # GOLD B200 portable runtime |
|
|
| Release `gold-runtime-b200-cu130-20261015-r1`. Linux x86_64, glibc>=2.35; standalone Python3.12.12, |
| TRL1.15.0, Torch2.11.0+cu130, Transformers5.6.0 and vLLM0.24.0. |
| Archive: 4,548,092,673 bytes, SHA256 |
| `4f78b026233a0f3e7bee2bd5d18ea4db01765644dd3b1aeeec2de1abcdebfbf6`. |
| Unpacked regular files: 9,682,853,603 bytes. Reserve additional disk for |
| the downloaded archive, node-local JIT caches, real weights and run outputs. |
| |
| CPU, B200 CUDA/native kernels, two GOLD optimizer steps with HF and colocated |
| vLLM, and clean archive restore/CLI have passed. Tiny random Llama/Qwen2 models |
| are synthetic qualification fixtures. This is runtime implementation validation, |
| not paper reproduction. Production models/data, CUDA toolkit, NVIDIA drivers, |
| credentials and JIT caches are excluded. Dependencies retain their own licenses; |
| this bundle does not relicense third-party software. |
| Full anonymous download with eight HTTP-range workers and every manifest file SHA256 also passed. |
| |
| ## Download and install |
| |
| Immutable payload revision: `1ff8becb57cb28c4d92a09971a6123b210b2e1d0`. All release files are under |
| `releases/gold-runtime-b200-cu130-20261015-r1/`. Anonymous download requires no HF token. Use the company-approved |
| proxy from your existing environment; no proxy endpoint is embedded here. |
| |
| Download [`download_release.py`](https://huggingface.co/datasets/codemaivanngu/gold-b200-runtime/resolve/1ff8becb57cb28c4d92a09971a6123b210b2e1d0/releases/gold-runtime-b200-cu130-20261015-r1/download_release.py) |
| and run it with an existing Python3.10+ interpreter: |
| |
| ```bash |
| python3 download_release.py --revision 1ff8becb57cb28c4d92a09971a6123b210b2e1d0 --output /your/node-local/gold-payload --workers 8 |
| ``` |
| |
| This downloader verifies the manifest and all file SHA256s and uses parallel |
| HTTP ranges for the runtime archive. It never installs or uses GPUs. |
| From the downloaded directory, `sha256sum -c SHA256SUMS` verifies the original |
| payload; `python3 install.py` verifies again, refuses existing runtime paths, |
| restores only the documented /opt paths and runs the CPU gate with GPUs hidden. |
| Creating those /opt paths requires the appropriate node permissions. |
| The installer has Python3.10-compatible syntax; clean restore was executed |
| using existing Python3.12.12. It does not install drivers or start training. |
| |
| Before GPU execution, obtain the actual node/GPU assignment and run the |
| documented `node_gate.py` UUID/occupancy preflight. Empty cards are not leases. |
| Company driver/toolkit/compiler compatibility still needs the same per-node |
| CUDA/HF/vLLM gates. No company node has been qualified by this release. |
| |
| The pinned upstream source is TRL commit |
| `dae7b6adc8d2729cfa8d21bc8480656ede769240`. Its example calls a nonexistent |
| `GOLDTrainer.end()`: the declared `gold_entrypoint.py` cleanup shim closes only |
| its owned generation engine and does not alter loss/data/hyperparameters. |
| See the release README, manifest, uv.lock, qualification.json and |
| restore-receipt.json for details. |
| |
