Upload GRANT_DEPLOY.md
Browse files- GRANT_DEPLOY.md +18 -0
GRANT_DEPLOY.md
ADDED
|
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# ZeroGPU activation
|
| 2 |
+
|
| 3 |
+
The full Gradio implementation is already in `app.py`. Once ZeroGPU is attached to this Space:
|
| 4 |
+
|
| 5 |
+
1. Change the README frontmatter from `sdk: static` to `sdk: gradio`.
|
| 6 |
+
2. Set `sdk_version: 6.26.0`, `app_file: app.py`, `python_version: "3.12"`, and `startup_duration_timeout: 1h`.
|
| 7 |
+
3. Keep the hardware flavor set to `zero-a10g`. The handler requests `size="xlarge"` because the BF16 model is well over 56 GB.
|
| 8 |
+
4. Upload the changed README and inspect build/runtime logs before calling the API.
|
| 9 |
+
|
| 10 |
+
Suggested community grant discussion title:
|
| 11 |
+
|
| 12 |
+
`Apply for a GPU community grant: Personal project`
|
| 13 |
+
|
| 14 |
+
Suggested description:
|
| 15 |
+
|
| 16 |
+
> This public open-source research demo lets visitors explore the tool-use and stopping behavior of vcruz305/Muse-Glimmer-30B-Hermes-Agentic. It safely displays proposed tool calls without executing commands, file operations, web requests, or destructive actions, and exposes a documented Gradio API/MCP endpoint for reproducible evaluation.
|
| 17 |
+
>
|
| 18 |
+
> The 30B BF16 checkpoint needs more than 56 GB of device memory, so the app requires a ZeroGPU xlarge allocation. The demo is educational and non-commercial, and I cannot currently cover dedicated GPU hosting costs. I am happy to provide more context if helpful.
|