How to use from the
Use from the
Diffusers library
pip install -U diffusers transformers accelerate
import torch
from diffusers import DiffusionPipeline

# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("TeamAIQ/SoulX-FlashHead-1_3B-INT8-ConvRot", dtype=torch.bfloat16, device_map="cuda")

prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]

SoulX-FlashHead-1_3B-INT8-ConvRot

INT8 ConvRot version of Soul-AILab/SoulX-FlashHead-1_3B, made for Comfy Kitchen's INT8 tensor-core kernels.

File From Size
Model_Lite/diffusion_pytorch_model_int8_convrot.safetensors Model_Lite/diffusion_pytorch_model.safetensors 1.9 GB
Model_Pro/diffusion_pytorch_model_int8_convrot.safetensors Model_Pro/diffusion_pytorch_model.safetensors 1.9 GB
VAE_LTX/diffusion_pytorch_model.safetensors copied as is 1.7 GB

Format

  • Linear weights of the transformer blocks: INT8 with one FP32 scale per output channel, after a Hadamard (ConvRot) rotation of the input channels in groups of 256. The INT8 kernels rotate the activations the same way, so each layer keeps its original result while outliers stay small.
  • Embeddings, norms, biases, AdaLN projections and the input / output layers keep their original precision.
  • Per layer: <name>.weight (I8), <name>.weight_scale (F32) and <name>.comfy_quant (layer config). File metadata: quantization_format: int8_convrot, convrot_format: comfy_quant_v1.
  • Loads directly in Wan2GP and runs on Comfy Kitchen's INT8 kernels.

The license and usage terms of the original model apply.


Quantized by AIQUEST 路 YouTube 路 X 路 aiquest.site

Downloads last month
25
Inference Providers NEW
This model isn't deployed by any Inference Provider. 馃檵 Ask for provider support

Model tree for TeamAIQ/SoulX-FlashHead-1_3B-INT8-ConvRot

Quantized
(2)
this model