Spaces:
Running on Zero
Running on Zero
| title: DiffusionGemma 26B A4B | |
| emoji: 💎 | |
| colorFrom: blue | |
| colorTo: green | |
| sdk: gradio | |
| sdk_version: 6.17.3 | |
| python_version: 3.12.12 | |
| app_file: app.py | |
| short_description: Text and image chat with DiffusionGemma | |
| startup_duration_timeout: 1h | |
| models: | |
| - google/diffusiongemma-26B-A4B-it | |
| tags: | |
| - image-text-to-text | |
| - conversational | |
| - diffusion-gemma | |
| - gemma | |
| - zerogpu | |
| # DiffusionGemma 26B A4B | |
| This Space runs [`google/diffusiongemma-26B-A4B-it`](https://huggingface.co/google/diffusiongemma-26B-A4B-it), Google's 25.2B parameter / 3.8B active parameter discrete-diffusion, multimodal Gemma 4 model for text and image input with text output. | |
| The app follows the model card defaults for DiffusionGemma generation: entropy-bound sampling, 48 maximum denoising steps, a 0.8 to 0.4 temperature schedule, 0.1 entropy bound, 0.005 confidence threshold, and adaptive stopping. Image input is placed before text in each user turn, and assistant thinking content is never fed back into multi-turn history. | |
| Implementation research and deployment notes are captured in [`MODEL_RESEARCH.md`](MODEL_RESEARCH.md). | |
| Live Space: <https://huggingface.co/spaces/Mike0021/diffusiongemma-26b-a4b-it> | |
| ## Runtime | |
| Use Gradio on ZeroGPU with `zero-a10g` hardware. The model's BF16 weights are about 51.7 GB, so the app requests `@spaces.GPU(size="xlarge")` for the full 96 GB ZeroGPU allocation. If deploying on dedicated hardware instead, use an 80 GB class GPU or larger and keep `GRADIO_SSR_MODE=false`. | |
| Recommended deployment flow: | |
| ```bash | |
| hf upload <namespace>/<space-name> . . --type space \ | |
| --exclude '.git/*' --exclude '__pycache__/*' --exclude '.venv/*' --exclude '*.pyc' \ | |
| --commit-message 'Add DiffusionGemma Space' | |
| ``` | |
| Create the Space and set hardware/variables with `huggingface_hub.HfApi` or the Hub UI when the installed `hf` CLI does not expose those controls. This deployment uses `zero-a10g`, the `GRADIO_SSR_MODE=false` variable, and an `HF_TOKEN` Space secret for model downloads. | |
| The live revision was verified with `gradio_client.Client(...).view_api()` followed by text and image `predict()` calls. A default-style text prompt returned a clean 103-token answer in 1.5s model time, and a small OCR image prompt returned `BLUE7 is visible in the image.` | |
| ## Research Sources | |
| - Model card and usage examples: <https://huggingface.co/google/diffusiongemma-26B-A4B-it> | |
| - Google DiffusionGemma docs: <https://ai.google.dev/gemma/docs/diffusiongemma> | |
| - Hugging Face Spaces configuration: <https://huggingface.co/docs/hub/spaces-config-reference> | |
| - Hugging Face ZeroGPU docs: <https://huggingface.co/docs/hub/spaces-zerogpu> | |
| - Space iteration quickstart: <https://gist.githubusercontent.com/gary149/37c955b832558837c40e1c14ff6d955d/raw/ad35807f8466378afd04d7653d53683a847b96c4/hf-spaces-agent-quickstart-compact.md> | |