| # Model Weight Snapshots | |
| This directory is reserved for physically bundled checkpoint snapshots. | |
| Expected layout after materialization: | |
| ```text | |
| models/ | |
| zai-org/ | |
| GLM-5.2-FP8/ | |
| config.json | |
| tokenizer.json | |
| *.safetensors | |
| Qwen/ | |
| Qwen3-Omni-30B-A3B-Instruct/ | |
| config.json | |
| tokenizer.json | |
| *.safetensors | |
| ``` | |
| Populate it with: | |
| ```bash | |
| python scripts/materialize_weights.py --profile glm52_plus_qwen_omni | |
| ``` | |
| Then upload the directory to this same repo: | |
| ```bash | |
| hf upload-large-folder wassemgtk/glm-5.2-visual-runtime models --type model | |
| ``` | |
| The one-click Compose profile mounts `../models:/models:ro` into the vLLM containers and can serve from local paths by setting: | |
| ```bash | |
| GLM_MODEL_PATH=/models/zai-org/GLM-5.2-FP8 | |
| QWEN_OMNI_MODEL_PATH=/models/Qwen/Qwen3-Omni-30B-A3B-Instruct | |
| ``` | |