Instructions to use j-llm/Qwen3.5-2B-Longbook2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use j-llm/Qwen3.5-2B-Longbook2 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="j-llm/Qwen3.5-2B-Longbook2")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("j-llm/Qwen3.5-2B-Longbook2")
model = AutoModelForMultimodalLM.from_pretrained("j-llm/Qwen3.5-2B-Longbook2", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use j-llm/Qwen3.5-2B-Longbook2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "j-llm/Qwen3.5-2B-Longbook2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "j-llm/Qwen3.5-2B-Longbook2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/j-llm/Qwen3.5-2B-Longbook2

SGLang

How to use j-llm/Qwen3.5-2B-Longbook2 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "j-llm/Qwen3.5-2B-Longbook2" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "j-llm/Qwen3.5-2B-Longbook2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "j-llm/Qwen3.5-2B-Longbook2" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "j-llm/Qwen3.5-2B-Longbook2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use j-llm/Qwen3.5-2B-Longbook2 with Docker Model Runner:
```
docker model run hf.co/j-llm/Qwen3.5-2B-Longbook2
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Qwen3.5-2B-Longbook2 (accuracy-recovered)

変換元モデルは Qwen3.5-2B の全 full-attention 層を Gated DeltaNet (GDN) に置き換えた all-GDN モデルです。置換で落ちた精度を、元モデル Qwen/Qwen3.5-2B を教師とした蒸留で回復したものがこのリポジトリです。 追加の知識データでの SFT は行っていません(復元のみ)。

全層が線形アテンション → KV キャッシュが増えず、長文でもメモリ・速度が一定
標準の transformers でそのまま読み込めます(trust_remote_code 不要)
flash-linear-attention を入れると高速カーネルが使われます

復元方法

Stage 1 (層ごとの蒸留): 置換した各 GDN 層の出力を元の full-attention 層の出力に合わせる (相対 MSE + コサイン)。層間の入力は detach し、全層を並列に学習。
Stage 2 (end-to-end): 教師 logits への KL + 隠れ状態損失で全体を微調整。
学習した重みを元の safetensors シャードの該当テンソルだけ差し替えて出力。

fp32 マスター重み + bf16 autocast、fused AdamW。

結果

(評価値は未記入)

(同一の評価テキスト・同一条件で測定。教師は Qwen/Qwen3.5-2B)

使い方

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "j-llm/Qwen3.5-2B-Longbook2"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, dtype=torch.bfloat16, device_map="cuda").eval()
ids = tok.apply_chat_template([{"role": "user", "content": "日本の首都はどこ？"}],
        add_generation_prompt=True, return_tensors="pt", return_dict=True, enable_thinking=False).to(model.device)
out = model.generate(**ids, max_new_tokens=128, use_cache=False)   # 下の注意参照
print(tok.decode(out[0, ids["input_ids"].shape[1]:], skip_special_tokens=True))

注意

全層が線形アテンションのため、現行 transformers の標準 Cache は get_seq_length で ValueError を出すことがあります。その場合は use_cache=False(遅いが確実)にするか、 get_seq_length を例外時に自前カウンタへフォールバックさせるパッチを当ててください。
蒸留は教師の挙動への近似であり、元モデルと完全には一致しません。特に推論・数学・コードは元モデルより劣る可能性があります。
ライセンスは元モデル(Qwen/Qwen3.5-2B)に従います。