Instructions to use Ismantic/Summer-0.5B-Chat with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Ismantic/Summer-0.5B-Chat with Transformers:

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

pipe = pipeline("text-generation", model="Ismantic/Summer-0.5B-Chat")

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Ismantic/Summer-0.5B-Chat")
model = AutoModelForCausalLM.from_pretrained("Ismantic/Summer-0.5B-Chat", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Ismantic/Summer-0.5B-Chat with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Ismantic/Summer-0.5B-Chat"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Ismantic/Summer-0.5B-Chat",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/Ismantic/Summer-0.5B-Chat

SGLang

How to use Ismantic/Summer-0.5B-Chat 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 "Ismantic/Summer-0.5B-Chat" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Ismantic/Summer-0.5B-Chat",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "Ismantic/Summer-0.5B-Chat" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Ismantic/Summer-0.5B-Chat",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use Ismantic/Summer-0.5B-Chat with Docker Model Runner:
```
docker model run hf.co/Ismantic/Summer-0.5B-Chat
```

Summer-0.5B-Chat / example_vllm.py

tf-bao

Upload Summer-0.5B-Chat

d05f7e5 verified about 2 months ago

Raw History Blame Contribute Delete

2.13 kB

"""用 vLLM 跑这个 chat 模型,支持多轮。
    pip install vllm
    pip install git+https://github.com/Ismantic/PieceTokenizer
    python example_vllm.py
## 为什么要 skip_tokenizer_init
vLLM 能加载**权重** —— config.json 里 architectures 是 Qwen3ForCausalLM,
它用自己那份 Qwen3 实现,按 state_dict 的 key 名灌进去。
但它**认不了这个词表**:81903 的 piece 词表不是 HF 标准格式,AutoTokenizer
走不通。所以要 skip_tokenizer_init=True,自己用 apply_chat_template 编码好
token id 用 TokensPrompt 喂进去,拿回来的 id 自己解码。
推论:`vllm serve` 那种 OpenAI 兼容服务**开箱用不了** —— 它要把文本转成
token,而它转不了。调用方得自己传 token id。
## repetition_penalty=1.15
**不是可选的美化项。** 纯贪心(temperature=0)容易陷入复读循环,这是这个
项目自己反复测出来的结论,不是随便调的参数——见仓库 docs/POSTTRAIN.md。
想要跑分可比的贪心结果,把 temperature 设成 0、repetition_penalty 设成 1.0。
"""
from vllm import LLM, SamplingParams
from vllm.inputs import TokensPrompt

from tokenizer import PieceTokenizerWrapper

HERE = "."

tok = PieceTokenizerWrapper(HERE)
llm = LLM(model=HERE, skip_tokenizer_init=True, dtype="bfloat16",
          gpu_memory_utilization=0.85, max_model_len=4096,
          trust_remote_code=True)

conversation = [
    {"role": "user", "content": "机器翻译的基本任务是什么？"},
]
ids = tok.apply_chat_template(conversation, tokenize=True,
                              add_generation_prompt=True)

outputs = llm.generate(
    [TokensPrompt(prompt_token_ids=ids)],
    SamplingParams(temperature=0.6, repetition_penalty=1.15, max_tokens=600,
                   stop_token_ids=tok.stop_token_ids),
)

reply = tok.decode(list(outputs[0].outputs[0].token_ids), skip_special_tokens=True)
print(reply)

# 多轮:把回复接回 conversation,再走一次 apply_chat_template。
# conversation.append({"role": "assistant", "content": reply})
# conversation.append({"role": "user", "content": "……下一个问题……"})

	"""用 vLLM 跑这个 chat 模型,支持多轮。

	pip install vllm
	pip install git+https://github.com/Ismantic/PieceTokenizer
	python example_vllm.py

	## 为什么要 skip_tokenizer_init

	vLLM 能加载权重 —— config.json 里 architectures 是 Qwen3ForCausalLM,
	它用自己那份 Qwen3 实现,按 state_dict 的 key 名灌进去。

	但它认不了这个词表:81903 的 piece 词表不是 HF 标准格式,AutoTokenizer
	走不通。所以要 skip_tokenizer_init=True,自己用 apply_chat_template 编码好
	token id 用 TokensPrompt 喂进去,拿回来的 id 自己解码。

	推论:`vllm serve` 那种 OpenAI 兼容服务开箱用不了 —— 它要把文本转成
	token,而它转不了。调用方得自己传 token id。

	## repetition_penalty=1.15

	不是可选的美化项。纯贪心(temperature=0)容易陷入复读循环,这是这个
	项目自己反复测出来的结论,不是随便调的参数——见仓库 docs/POSTTRAIN.md。
	想要跑分可比的贪心结果,把 temperature 设成 0、repetition_penalty 设成 1.0。
	"""
	from vllm import LLM, SamplingParams
	from vllm.inputs import TokensPrompt

	from tokenizer import PieceTokenizerWrapper

	HERE = "."

	tok = PieceTokenizerWrapper(HERE)
	llm = LLM(model=HERE, skip_tokenizer_init=True, dtype="bfloat16",
	gpu_memory_utilization=0.85, max_model_len=4096,
	trust_remote_code=True)

	conversation = [
	{"role": "user", "content": "机器翻译的基本任务是什么？"},
	]
	ids = tok.apply_chat_template(conversation, tokenize=True,
	add_generation_prompt=True)

	outputs = llm.generate(
	[TokensPrompt(prompt_token_ids=ids)],
	SamplingParams(temperature=0.6, repetition_penalty=1.15, max_tokens=600,
	stop_token_ids=tok.stop_token_ids),
	)

	reply = tok.decode(list(outputs[0].outputs[0].token_ids), skip_special_tokens=True)
	print(reply)

	# 多轮:把回复接回 conversation,再走一次 apply_chat_template。
	# conversation.append({"role": "assistant", "content": reply})
	# conversation.append({"role": "user", "content": "……下一个问题……"})