Instructions to use tencent/Hy3-FP8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use tencent/Hy3-FP8 with Transformers:

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

pipe = pipeline("text-generation", model="tencent/Hy3-FP8")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("tencent/Hy3-FP8")
model = AutoModelForCausalLM.from_pretrained("tencent/Hy3-FP8")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use tencent/Hy3-FP8 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "tencent/Hy3-FP8"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "tencent/Hy3-FP8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/tencent/Hy3-FP8

SGLang

How to use tencent/Hy3-FP8 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 "tencent/Hy3-FP8" \
    --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": "tencent/Hy3-FP8",
		"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 "tencent/Hy3-FP8" \
        --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": "tencent/Hy3-FP8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use tencent/Hy3-FP8 with Docker Model Runner:
```
docker model run hf.co/tencent/Hy3-FP8
```

yiqichen01 commited on about 5 hours ago

Commit

3b6aef9

verified ·

1 Parent(s): 73b9d31

Upload folder using huggingface_hub

Browse files

Files changed (2) hide show

finetune/README.md +1 -1
finetune/README_CN.md +1 -1

finetune/README.md CHANGED Viewed

@@ -194,7 +194,7 @@ The key parameters in the script are as follows:
 - When resuming from a checkpoint, there may be minor differences in loss due to the randomness of some non-deterministic algorithms. This is normal. See: [HuggingFace Transformers Trainer Randomness](https://huggingface.co/docs/transformers/main/en/main_classes/trainer#randomness)
 - When `--model_name_or_path` is specified, all model-related parameters will be ignored.
 - Samples within a batch are padded to the length of the longest sample in the batch, but the maximum length of each sample is `max_seq_length`. Any excess will be truncated.
-- If you see a warning about bias weights not being loaded, you can ignore it. Hunyuan-Large does not use bias.
 ##### What if GPU Memory is Insufficient?

 - When resuming from a checkpoint, there may be minor differences in loss due to the randomness of some non-deterministic algorithms. This is normal. See: [HuggingFace Transformers Trainer Randomness](https://huggingface.co/docs/transformers/main/en/main_classes/trainer#randomness)
 - When `--model_name_or_path` is specified, all model-related parameters will be ignored.
 - Samples within a batch are padded to the length of the longest sample in the batch, but the maximum length of each sample is `max_seq_length`. Any excess will be truncated.
+- If you see a warning about **linear layer** bias weights not being loaded, you can ignore it; Hy3's linear layers (q_proj / k_proj / v_proj / o_proj, etc.) do not use bias. Note: the MoE router's `e_score_correction_bias` is a buffer and is auto-loaded by the training script, so please do not ignore its loading failure.
 ##### What if GPU Memory is Insufficient?

finetune/README_CN.md CHANGED Viewed

@@ -194,7 +194,7 @@ Are you sure you want to continue connecting (yes/no)?
 - 从 ckpt 继续训练时，loss 可能会有微小的偏差，这是由一些非确定性算法带来的随机性，是正常现象。参考：[HuggingFace Transformers Trainer Randomness](https://huggingface.co/docs/transformers/main/en/main_classes/trainer#randomness)
 - 当 `--model_name_or_path` 有效时，所有模型相关的参数都会被忽略
 - 一个 batch 内的样本会通过 padding 对齐 batch 内最长的样本，而每条样本的长度最长为 max_seq_length，超出的部分会被裁剪
-- 如果报出 bias 权重没有 load 的 warning，忽略即可，Hunyuan-Large 中不会用到 bias
 ##### 显存不足怎么办？

 - 从 ckpt 继续训练时，loss 可能会有微小的偏差，这是由一些非确定性算法带来的随机性，是正常现象。参考：[HuggingFace Transformers Trainer Randomness](https://huggingface.co/docs/transformers/main/en/main_classes/trainer#randomness)
 - 当 `--model_name_or_path` 有效时，所有模型相关的参数都会被忽略
 - 一个 batch 内的样本会通过 padding 对齐 batch 内最长的样本，而每条样本的长度最长为 max_seq_length，超出的部分会被裁剪
+- 如果报出**线性层** bias 权重没有 load 的 warning，忽略即可，Hy3 的线性层（q_proj / k_proj / v_proj / o_proj 等）不使用 bias。注意：MoE 路由的 `e_score_correction_bias` 属于 buffer，已由训练脚本自动加载，如果加载失败请不要忽略。
 ##### 显存不足怎么办？