Warn that the fused attention layout needs a splitting loader
Browse files
README.md
CHANGED
|
@@ -11,6 +11,15 @@ tags:
|
|
| 11 |
- dit
|
| 12 |
- ideogram
|
| 13 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
> Ungated mirror of [ideogram-ai/ideogram-4-fp8](https://huggingface.co/ideogram-ai/ideogram-4-fp8), republished by Unsloth so it
|
| 15 |
> can be downloaded without a Hub gate. The weights are unmodified. See NOTICE and LICENSE
|
| 16 |
> in this repo for the terms, which are the upstream terms and are unchanged.
|
|
|
|
| 11 |
- dit
|
| 12 |
- ideogram
|
| 13 |
---
|
| 14 |
+
> **Loading note.** This repo keeps ideogram's original FUSED attention layout
|
| 15 |
+
> (`attention.qkv` / `attention.o`). Released diffusers expects the split
|
| 16 |
+
> `to_q`/`to_k`/`to_v`/`to_out.0`, so a plain `from_pretrained` here silently
|
| 17 |
+
> random-initialises the attention weights and produces garbage images rather than
|
| 18 |
+
> raising. Use [unsloth/ideogram-4-nf4-diffusers](https://huggingface.co/unsloth/ideogram-4-nf4-diffusers)
|
| 19 |
+
> for the diffusers-native layout, or a loader that splits the fused qkv itself
|
| 20 |
+
> (Unsloth Studio does this). The text encoder config also needs transformers 5.x.
|
| 21 |
+
> Both behaviours are inherited from upstream, not introduced by this mirror.
|
| 22 |
+
|
| 23 |
> Ungated mirror of [ideogram-ai/ideogram-4-fp8](https://huggingface.co/ideogram-ai/ideogram-4-fp8), republished by Unsloth so it
|
| 24 |
> can be downloaded without a Hub gate. The weights are unmodified. See NOTICE and LICENSE
|
| 25 |
> in this repo for the terms, which are the upstream terms and are unchanged.
|