Views
No views yet
1python quantize.py configs/te_gemma-4-fp8.json -q -n 131072 \
2 gemma-4-E4B-it-ultra-uncensored-heretic.safetensors \
3 gemma-4-E4B-it-ultra-uncensored-heretic-fp8.safetensors1{
2 "format": "comfy_quant",
3 "replace_names": {
4 "model.audio_tower": "audio_model",
5 "model.embed_audio": "audio_projector",
6 "model.language_model": "model",
7 "model.embed_vision": "multi_modal_projector",
8 "model.vision_tower": "vision_model"
9 },
10 "block_names": ["language_model"],
11 "rules": [
12 { "policy": "keep", "match": [
13 "k_proj", "o_proj", "q_proj", "v_proj",
14 "per_layer_input_gate", "per_layer_projection"
15 ] },
16 { "policy": "float8_e4m3fn", "match": [
17 "embed_tokens", "embed_tokens_per_layer",
18 "per_layer_model_projection",
19 "down_proj", "gate_proj", "up_proj"
20 ] }
21 ]
22}tokenizer_json from https://huggingface.co/Comfy-Org/gemma-4/blob/main/text_encoders/gemma4_e4b_it_fp8_scaled.safetensors1#!/usr/bin/env python3
2import sys
3from safetensors import safe_open
4from safetensors.torch import save_file
5
6src, dst = sys.argv[1], sys.argv[2]
7
8with safe_open(src, framework="pt", device="cpu") as f:
9 tokenizer = f.get_tensor("tokenizer_json")
10
11with safe_open(dst, framework="pt", device="cpu") as f:
12 meta = f.metadata() or {}
13 tensors = {k: f.get_tensor(k) for k in f.keys()}
14
15tensors["tokenizer_json"] = tokenizer
16save_file(tensors, dst, metadata=meta)