| --- |
| license: apache-2.0 |
| tags: |
| - multimodal |
| - omni |
| - vision |
| - audio |
| - 3d |
| - qwen3-vl |
| - aether |
| --- |
| # Aether Phase-1 — assembled + wired multimodal omni-model |
|
|
| Qwen3-VL-8B base grafted with **InternViT-6B** (vision), **MiMo-Audio** (audio), and **TRELLIS-SLAT** (3D geometry) via 5 projectors, expanded vocab (+4627 tokens), and a **factorized / 3D-spatial M-RoPE**. Verified end-to-end on ROCm (MI300). |
|
|
| **15.16B params - 32GB VRAM (bf16) - full multimodal forward verified** (seq 5542, logits (1,5542,156296) finite). |
|
|
| This repo is a **resumable assembly package**, not a full weight dump - it pins the base + encoders in MANIFEST.json so the exact model re-assembles on any box (6900XT / Kaggle / Colab / rental) without duplicating the 16GB base. |
|
|
| ## Contents |
| - MANIFEST.json - base + 3 encoder repos/revisions/dims, freeze scheme, M-RoPE routing, verification log |
| - tokenizer/ - expanded tokenizer (151669 to 156296) |
| - new_modules.safetensors - 5 projectors + cam_pose (98.6M params, fresh init pre-alignment) |
| - scripts/ - assemble.py, forward.py (routing + M-RoPE), surgery.py, curate_slivers.py, save_package.py |
| - BUILD.md - full build spec, graft verifications, ROCm recipes, model-building knowledge |
|
|
| ## Resume |
| Clone, pull base+encoders per MANIFEST, run assemble.py then forward.py, load new_modules.safetensors, begin sliver alignment training. |
| |
