license: apache-2.0
license_link: https://huggingface.co/Qwen/Qwen3.5-9B/blob/main/LICENSE
pipeline_tag: image-text-to-text
base_model:
- llmfan46/Qwen3.5-9B-ultra-uncensored-heretic
tags: - qwen3_5
- heretic
- uncensored
- decensored
- abliterated
- ara
- int8_convrot
- gptq
- smoothquant
- comfyui
- quantized
Qwen3.5-9B-ultra-uncensored-heretic — INT8 ConvRot (ComfyUI text encoder)
INT8 ConvRot quantisation of llmfan46/Qwen3.5-9B-ultra-uncensored-heretic, packaged as a ComfyUI text encoder for MOSS-Image pipelines.
Every decoder-layer Linear of the language model is stored as int8 with a per-row fp32 scale, using ComfyUI's int8_tensorwise + convrot layout (256-column group Hadamard rotation). It loads directly with the stock CLIPLoader (type: moss_image) — no custom node required.
| Base model | llmfan46/Qwen3.5-9B-ultra-uncensored-heretic (apache-2.0) |
| Quantised | 248 language-model linears (24 GDN × 5, 8 full-attention × 4, 32 MLP × 3) |
| Recipe | GPTQ (block 128, damp 0.01, act-order) + SmoothQuant α=0.5, calibrated on 96 MOSS-Image captions + 16× empty prompt |
| Grid | int8, per-row scale, convrot_groupsize: 256 |
| Size | 11.09 GiB (from 17.53 GiB bf16, −37 %) |
| Visual tower | kept (bf16, plus int8 for its blocks in the alternative file) for future image-conditioned use |
File format
Plain safetensors with ComfyUI's quantised-linear convention. Each quantised linear carries three tensors:
model.language_model.layers.N.<module>.weight int8 (Hadamard-rotated, quantised)
model.language_model.layers.N.<module>.weight_scale float32 (out_features, 1)
model.language_model.layers.N.<module>.comfy_quant uint8 {"format":"int8_tensorwise","convrot":true,"convrot_groupsize":256}
Norms that SmoothQuant folded scales into are stored in bf16. Non-quantised tensors (embedding, conv1d, A_log, dt_bias, visual tower) keep the checkpoint dtype.
Usage (ComfyUI)
ComfyUI/models/text_encoders/qwen3.5_9b_ultra_heretic_int8_gptq.safetensors
CLIPLoader clip_name: qwen3.5_9b_ultra_heretic_int8_gptq.safetensors
type: moss_image
device: default
Works with the stock CLIPLoader and with SmartCLIPLoader — both produce pixel-identical output on this file. The W8A8 path (comfy_kitchen.int8_linear) is selected automatically from comfy_quant.
Evaluation
Same prompt / seed / base model (moss_image_base_bf16) / LoRA, 1024×1536, 50 steps, cfg 6, er_sde + sgm_uniform, comparing against the bf16 encoder:
| Encoder | MAE vs bf16 | PSNR |
|---|---|---|
| RTN int8 (248 linears) | 16.85 / 255 | 18.44 dB |
| GPTQ + SmoothQuant (this file) | 16.38 / 255 | 18.91 dB |
- Composition, style, details and anatomy are indistinguishable from bf16.
- There is a small, systematic +6.3 % exposure lift (+16/255, uniform across R/G/B). This comes from the W8A8 activation quantisation ComfyUI performs at load time, not from the weight grid: GPTQ improved the weight-side output error 10–20× (rtn 1e-5–1e-4 → gptq 1e-6–1e-5) yet moved the pixels by only 2.8 %. Use a slightly darker prompt or a global exposure adjustment if it matters.
Limitations
- This is a text encoder, not a chat model — do not expect generation quality comparable to the bf16 base.
- The activation rounding is not modelled during GPTQ (
act_awarevariants were tested and produced unstable multi-layer drift; not included). - Text-encoding quality is calibrated on anime-style caption traffic; other domains may shift more.
- The base model is an uncensored / abliterated derivative — outputs may include explicit content. Users are responsible for compliance with local law.
Credits
- Base model: llmfan46/Qwen3.5-9B-ultra-uncensored-heretic (apache-2.0)
- Quantisation approach: sorryhyun/anima_lora (
library/moss/int8.py,library/int8/convrot.py,te_gptq.py) and ComfyUI'scomfy_kitchenconvrot kernels - Format reference:
Comfy-Org/MOSS-Imagetext_encoders/qwen3.5_9b_int8_convrot.safetensors
Licensed under apache-2.0, following the base model.