| # Remote Downloaded Models |
|
|
| ## DMax-16B |
|
|
| - Repository: `Zigeng/DMax-16B` |
| - Revision: `225df0ce2818396fd3f18996c8fbd59b0666e4fe` |
| - Remote path: `/mnt/scratch/l/liyj/hku_diffusion_dllm/models/dmax-16b` |
| - Remote cache: `/mnt/scratch/l/liyj/hku_diffusion_dllm/.cache/huggingface` |
| - Expected size: `32,529,274,876` bytes across seven weight shards |
| - Status: downloaded via Slurm job `680834`; official single-sample smoke completed via `682109` and `682112`. |
| - Consumer jobs: old `680837` failed before torch import because project venv import was broken on the GPU node. Scratch runtime fix `682107` completed. Smoke results: `682109` on H100 NVL 96G and `682112` on A100 80G PCIe. |
| - Note: first attempt on `shiying` completed 4/7 shards but hit project quota during retry; the scratch job copies reusable shards first and resumes there. |
|
|
| ## Historical note |
|
|
| Before the DMax run, no full checkpoint from the original manifest had been materialized. |
|
|
| The original reason was that the core model list contains multi-GB 7B/8B checkpoints and should not |
| be pulled on an xlogin node without a concrete GPU job plan and storage budget. |
| The model manifest is ready for remote materialization via: |
|
|
| ```bash |
| cd /home/l/liyj/shiying/hku_diffusion_dllm |
| . .venv/bin/activate |
| export HF_HOME=$PWD/.cache/huggingface |
| export HF_HUB_CACHE=$PWD/.cache/huggingface/hub |
| export HF_HUB_DISABLE_XET=1 |
| python scripts/download_assets.py --kind models --id llada2-mini --max-workers 1 |
| ``` |
|
|
