| # Hugging Face migration notes |
|
|
| This repository was pulled from: |
|
|
| - source Space: `https://huggingface.co/spaces/ChemFM/MaterialInverseDesignDemo` |
|
|
| Target destination: |
|
|
| - destination org: `https://huggingface.co/aim4composites` |
|
|
| ## What changed for GPU hosting |
|
|
| - README front matter now recommends `suggested_hardware: t4-small` |
| - `spaces` is kept in `requirements.txt` so ZeroGPU can allocate GPU time through `@spaces.GPU` |
| - the app text now describes the Space as compatible with both ZeroGPU and dedicated GPU hardware |
| - runtime knobs were added so the thermoforming step can be tuned without code edits: |
| - `MG_DEFAULT_N_GENERATE` |
| - `MG_MAX_GENERATE` |
| - `MG_THERMO_RESTARTS` |
| - `MG_THERMO_EPOCHS` |
|
|
| ## Recommended Hugging Face Space settings |
|
|
| 1. Create a new Gradio Space under `aim4composites`. |
| 2. Push this repository to that new Space. |
| 3. In the Space hardware settings, choose either ZeroGPU or `Nvidia T4 small`. |
| 4. Leave storage at the default unless larger artifacts are added later. |
| 5. Restart the Space and verify generation completes. On `Nvidia T4 small`, logs should show CUDA is available. On ZeroGPU, CUDA is allocated when the decorated generation function runs. |
|
|
| ## Suggested first validation pass |
|
|
| 1. Open the Space and load the default test condition. |
| 2. Run material inverse design with the default `3` candidates. |
| 3. Select the top-ranked design. |
| 4. Run thermoforming once with a small target example. |
| 5. Confirm: |
| - the app starts successfully |
| - candidate generation completes |
| - thermoforming completes |
| - no out-of-memory or timeout errors appear in the logs |
|
|
| ## Optional runtime tuning if the Space feels slow |
|
|
| Set these Space variables: |
|
|
| - `MG_DEFAULT_N_GENERATE=2` |
| - `MG_MAX_GENERATE=4` |
| - `MG_THERMO_RESTARTS=6` |
| - `MG_THERMO_EPOCHS=300` |
|
|
| If runtime remains comfortably fast, increase these gradually. |
|
|
