| code is not the problem. |
|
|
| I compared our LlmChatViewModel + ChatViewModel to Google’s line‑by‑line. They are identical |
| (only package names differ). So the “empty text blob” is not caused by those files. The only |
| remaining difference is the 270m model itself or the downloaded file. |
|
|
| Why 270m can produce an empty bubble even with Google’s code |
|
|
| - The app does not validate the downloaded file. If HuggingFace returns HTML/redirect |
| content (or a partial file), it still saves it and tries to load it. |
| - 270m uses commitHash: "main" and a fixed sizeInBytes. If the file changes or a resume |
| happens, you can end up with a corrupted file. |
| - Result: the model loads but returns empty tokens, which shows as an empty message bubble. |
|
|
| The 1B model working proves the pipeline is fine. |
|
|
| What I recommend checking (no code change) |
|
|
| 1. On device, check the actual file size: |
| - Path pattern: |
| Android/data/com.hawkfranklin.aura/files/Gemma3_270M_IT/main/gemma3-270m-it- |
| q8.litertlm |
| - If it’s not ~304 MB, the file is wrong/corrupt. |
| 2. Delete that file and re‑download (with an HF token if needed). |
|
|
| What will actually fix it (you choose) |
|
|
| 1. Disable or remove the 270m entry (so only working models show). |
| 2. Use a pinned commit hash from HuggingFace (avoid "main"). |
| 3. If you want 270m specifically, I can update the allowlist with the exact commit hash and |
| size you choose. |
|
|
| If you want me to fix it now, tell me which option you want: |
|
|
| 1. Disable 270m |
| 2. Pin to a specific commit (send the commit hash) |
| 3. Keep 270m but replace with a different file from that repo (tell me which file) |
|
|
| Also: the title line change to “Abliterated Uncensored / RAW Artificial Intelligence” is |
| already in the code. |
|
|
