File size: 5,542 Bytes
ce001b1 44e6208 ce001b1 6c4249d f8377e9 ccd1ae7 f8377e9 ccd1ae7 f8377e9 8ff3726 f8377e9 8ff3726 f8377e9 8ff3726 f8377e9 8ff3726 f8377e9 8ff3726 f8377e9 ce001b1 f8377e9 ce001b1 f8377e9 180e24a f8377e9 44e6208 f8377e9 180e24a f8377e9 180e24a f8377e9 44e6208 f8377e9 44e6208 f8377e9 44e6208 f8377e9 44e6208 f8377e9 44e6208 f8377e9 44e6208 f8377e9 44e6208 f8377e9 a6b6275 f8377e9 a6b6275 f8377e9 44e6208 f8377e9 180e24a f8377e9 44e6208 f8377e9 180e24a f8377e9 180e24a f8377e9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 | ---
license: other
pipeline_tag: text-to-image
tags:
- unet
- orbitquant
- quantization
- w4a4
- w3a4
- w2a4
- custom-runtime-required
- sdxl
---
# USE CONVROT INT4 and INT8 for now
# OrbitQuant SDXL Exports
Experimental OrbitQuant and ConvRot quantized exports for CreaPrompt Lightning Hyper-SDXL.
> **Work in progress:** The native A4 runtime is not public yet. These models cannot currently be used without the custom runtime. A ComfyUI loader node that outputs a model compatible with the core KSampler is still under development.
## W2A4 Recovery Status
WIP for W2A4 block-level rollout recovery.
## Model Format
These are not standard SDXL checkpoints that can be loaded directly by existing SDXL loaders.
Depending on the export, the files can contain:
* OrbitQuant-packed SDXL UNet weights
* ConvRot-quantized SDXL UNet weights
* A4 activation-quantization metadata
* Quantized CLIP and text-encoder weights
* FP16 VAE weights
* `__orbitquant__` metadata
* Codebooks
* Rotations
* Quantization scales
* Runtime configuration
A custom runtime is required to execute the packed OrbitQuant and ConvRot weights without expanding them back into ordinary FP16 weights.
# Examples
## Full OrbitQuant versus Full ConvRot
<table>
<tr>
<td align="center">
<img
src="https://huggingface.co/ApacheOne/orbitquant_unets/resolve/main/hybrid_quants/orbitW4A4convrotW4A4quant/examples/comparisons/FULL_OQW4A4_vs_FULL_CRW4A4_CLIPINT4_comparison.png"
width="760"
>
<br>
<strong>Full OrbitQuant W4A4 versus Full ConvRot W4A4 — CLIP INT4</strong>
</td>
</tr>
</table>
## Full-UNet Tests
<table>
<tr>
<td align="center" width="50%">
<img
src="https://huggingface.co/ApacheOne/orbitquant_unets/resolve/main/hybrid_quants/orbitW4A4convrotW4A4quant/examples/full_unet/convrot_w4a4_clip_int4/FULLUNET_CRW4A4_CLIPINT4_test.png"
width="420"
>
<br>
<strong>Full ConvRot W4A4 UNet — CLIP INT4</strong>
</td>
<td align="center" width="50%">
<img
src="https://huggingface.co/ApacheOne/orbitquant_unets/resolve/main/hybrid_quants/orbitW4A4convrotW4A4quant/examples/full_unet/orbitquant_w4a4_clip_int4/FULLUNET_OQW4A4_CLIPINT4_test.png"
width="420"
>
<br>
<strong>Full OrbitQuant W4A4 UNet — CLIP INT4</strong>
</td>
</tr>
</table>
## Mixed OrbitQuant and ConvRot UNet — CLIP FP16
<table>
<tr>
<td align="center" width="50%">
<img
src="https://huggingface.co/ApacheOne/orbitquant_unets/resolve/main/hybrid_quants/orbitW4A4convrotW4A4quant/examples/mixed_unet/clip_fp16/creaprompt_SDXL_OQW4A4_CRW4A4_CLIPFP16_test.png"
width="420"
>
<br>
<strong>OrbitQuant W4A4 + ConvRot W4A4 — CLIP FP16</strong>
</td>
<td align="center" width="50%">
<img
src="https://huggingface.co/ApacheOne/orbitquant_unets/resolve/main/hybrid_quants/orbitW4A4convrotW4A4quant/examples/mixed_unet/clip_fp16/creaprompt_SDXL_OQW4A4_CRW8A8_CLIPFP16_test.png"
width="420"
>
<br>
<strong>OrbitQuant W4A4 + ConvRot W8A8 — CLIP FP16</strong>
</td>
</tr>
</table>
## Mixed OrbitQuant and ConvRot UNet — CLIP INT4
<table>
<tr>
<td align="center" width="50%">
<img
src="https://huggingface.co/ApacheOne/orbitquant_unets/resolve/main/hybrid_quants/orbitW4A4convrotW4A4quant/examples/mixed_unet/clip_int4/creaprompt_SDXL_OQW4A4_CRW4A4_CLIPINT4_test.png"
width="420"
>
<br>
<strong>OrbitQuant W4A4 + ConvRot W4A4 — CLIP INT4</strong>
</td>
<td align="center" width="50%">
<img
src="https://huggingface.co/ApacheOne/orbitquant_unets/resolve/main/hybrid_quants/orbitW4A4convrotW4A4quant/examples/mixed_unet/clip_int4/creaprompt_SDXL_OQW4A4_CRW8A8_CLIPINT4_test.png"
width="420"
>
<br>
<strong>OrbitQuant W4A4 + ConvRot W8A8 — CLIP INT4</strong>
</td>
</tr>
</table>
## Mixed OrbitQuant and ConvRot UNet — CLIP INT8
<table>
<tr>
<td align="center" width="50%">
<img
src="https://huggingface.co/ApacheOne/orbitquant_unets/resolve/main/hybrid_quants/orbitW4A4convrotW4A4quant/examples/mixed_unet/clip_int8/creaprompt_OQW4A4_CRW8A8_test.png"
width="420"
>
<br>
<strong>OrbitQuant W4A4 + ConvRot W8A8 — CLIP INT8</strong>
</td>
<td align="center" width="50%">
<img
src="https://huggingface.co/ApacheOne/orbitquant_unets/resolve/main/hybrid_quants/orbitW4A4convrotW4A4quant/examples/mixed_unet/clip_int8/first_model_test.png"
width="420"
>
<br>
<strong>First Mixed-Quant Model Test — CLIP INT8</strong>
</td>
</tr>
</table>
## Root Example
<table>
<tr>
<td align="center">
<img
src="https://huggingface.co/ApacheOne/orbitquant_unets/resolve/main/hybrid_quants/orbitW4A4convrotW4A4quant/examples/creaprompt_SDXL_OQW4A4_CRW8A8_CLIPINT4_test.png"
width="520"
>
<br>
<strong>OrbitQuant W4A4 + ConvRot W8A8 — CLIP INT4</strong>
</td>
</tr>
</table>
## Runtime Availability
The following components are still required for normal use:
* Native packed OrbitQuant weight execution
* Native A4 activation quantization
* Native ConvRot quantized operators
* ComfyUI model loader
* Integration with the core ComfyUI KSampler
* Runtime validation
* Performance tuning
Until the custom runtime is released, these files should be considered experimental development exports. |