Stable diffusion-forge-add-on https://github.com/yoikoarmor/sd-forge-llm-prompt-gen-yoiko

yoiko-Qwen3.5-9B-lora

yoiko-Qwen3.5-9B-lora is a LoRA adapter package for Qwen/Qwen3.5-9B.

This package contains the adapter only.

It does not include:

• the Qwen base model
• training data
• internal experiment logs
• Stable Diffusion Forge extension code

Intended use

This adapter is intended for prompt generation / prompt rewriting workflows, including use inside a Stable Diffusion Forge extension.

Base model

Expected base model:

• Qwen/Qwen3.5-9B

Users must obtain the base model separately and must review the base model's own license and usage terms.

License

This LoRA adapter package is released under the Apache License 2.0.

This license applies to the adapter package contents in this directory.

The base model is not included here and may have its own separate license.

Recommended inference conditions

Typical tested setup:

• 4-bit or bf16 base model loading depending on runtime behavior
• NF4 quantization when 4-bit is used
• bfloat16 compute dtype
• device_map="auto"
• adapter tokenizer / adapter chat template when available

Files in this package

• adapter_model.safetensors
• adapter_config.json
• chat_template.jinja
• tokenizer.json
• tokenizer_config.json
• README.md
• LICENSE

Minimal load example

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch

base_model = "Qwen/Qwen3.5-9B"
adapter_path = "yoikoarmor/yoiko-Qwen3.5-9B-lora"

tokenizer = AutoTokenizer.from_pretrained(adapter_path, trust_remote_code=False)

quant_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

model = AutoModelForCausalLM.from_pretrained(
    base_model,
    device_map="auto",
    quantization_config=quant_config,
    torch_dtype=torch.bfloat16,
    trust_remote_code=False,
)

model = PeftModel.from_pretrained(model, adapter_path, is_trainable=False)
model.eval()

Known limitations

• This is an adapter package, not a standalone model
• Qwen 3.5 family runtimes can differ depending on fast-path availability
• Prompt quality still depends on prompting style and generation settings
• Stable Diffusion Forge integration behavior may differ from standalone scripts