license: apache-2.0
base_model:

• llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic
  base_model_relation: quantized
  tags:
• comfyui
• diffusion-single-file

ComfyUI clip model of llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic

How to use

How to reproduce

Use https://github.com/bedovyy/comfy-dit-quantizer with the below config json.

python quantize.py configs/te_gemma-4-fp8.json -q -n 131072 \
  gemma-4-E4B-it-ultra-uncensored-heretic.safetensors \
  gemma-4-E4B-it-ultra-uncensored-heretic-fp8.safetensors

{
  "format": "comfy_quant",
  "replace_names": {
    "model.audio_tower": "audio_model",
    "model.embed_audio": "audio_projector",
    "model.language_model": "model",
    "model.embed_vision": "multi_modal_projector",
    "model.vision_tower": "vision_model"
  },
  "block_names": ["language_model"],
  "rules": [
    { "policy": "keep", "match": [
            "k_proj", "o_proj", "q_proj", "v_proj",
            "per_layer_input_gate", "per_layer_projection"
    ] },
    { "policy": "float8_e4m3fn",  "match": [
            "embed_tokens", "embed_tokens_per_layer",
            "per_layer_model_projection",
            "down_proj", "gate_proj", "up_proj"
    ] }
  ]
}

Then, copy tokenizer_json from https://huggingface.co/Comfy-Org/gemma-4/blob/main/text_encoders/gemma4_e4b_it_fp8_scaled.safetensors

#!/usr/bin/env python3
import sys
from safetensors import safe_open
from safetensors.torch import save_file

src, dst = sys.argv[1], sys.argv[2]

with safe_open(src, framework="pt", device="cpu") as f:
    tokenizer = f.get_tensor("tokenizer_json")

with safe_open(dst, framework="pt", device="cpu") as f:
    meta = f.metadata() or {}
    tensors = {k: f.get_tensor(k) for k in f.keys()}

tensors["tokenizer_json"] = tokenizer
save_file(tensors, dst, metadata=meta)