Add model card metadata: base_model, license, pipeline_tag, tags
Browse files
README.md
CHANGED
|
@@ -1,3 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
# Ovi FusionModel - FP8 Quantized
|
| 2 |
|
| 3 |
This is the Ovi FusionModel quantized with FP8 (e4m3_e4m3_dynamic_per_tensor) for faster inference.
|
|
|
|
| 1 |
+
---
|
| 2 |
+
base_model: chetwinlow1/Ovi
|
| 3 |
+
license: apache-2.0
|
| 4 |
+
pipeline_tag: image-to-video
|
| 5 |
+
tags:
|
| 6 |
+
- ovi
|
| 7 |
+
- image-to-video
|
| 8 |
+
- text-to-video
|
| 9 |
+
- video-generation
|
| 10 |
+
- audio-generation
|
| 11 |
+
- quantized
|
| 12 |
+
- fp8
|
| 13 |
+
- e4m3
|
| 14 |
+
base_model_relation: quantized
|
| 15 |
+
---
|
| 16 |
+
## About this repository
|
| 17 |
+
|
| 18 |
+
FP8 (e4m3) quantized [Ovi](https://huggingface.co/chetwinlow1/Ovi) fusion model,
|
| 19 |
+
published by WaveSpeed AI. Ovi generates video and its matching audio in one
|
| 20 |
+
pass; both towers are quantized here — 30 video blocks and 30 audio blocks —
|
| 21 |
+
with `e4m3_e4m3_dynamic_per_tensor` on attention and FFN layers and
|
| 22 |
+
`e4m3_weightonly` elsewhere. Compute stays in bf16. FP8 matmul needs Hopper
|
| 23 |
+
(H100/H200) or newer to be faster than bf16.
|
| 24 |
+
|
| 25 |
+
The original model is [Ovi](https://github.com/character-ai/Ovi) by Character.AI,
|
| 26 |
+
released under Apache-2.0; that license carries over to these weights.
|
| 27 |
+
|
| 28 |
+
---
|
| 29 |
+
|
| 30 |
# Ovi FusionModel - FP8 Quantized
|
| 31 |
|
| 32 |
This is the Ovi FusionModel quantized with FP8 (e4m3_e4m3_dynamic_per_tensor) for faster inference.
|