Instructions to use Qwen/Qwen3-30B-A3B-Instruct-2507 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Qwen/Qwen3-30B-A3B-Instruct-2507 with Transformers:

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

pipe = pipeline("text-generation", model="Qwen/Qwen3-30B-A3B-Instruct-2507")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-30B-A3B-Instruct-2507")
model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-30B-A3B-Instruct-2507", device_map="auto")
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]:]))

Inference
Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Qwen/Qwen3-30B-A3B-Instruct-2507 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Qwen/Qwen3-30B-A3B-Instruct-2507"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Qwen/Qwen3-30B-A3B-Instruct-2507",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Qwen/Qwen3-30B-A3B-Instruct-2507

SGLang

How to use Qwen/Qwen3-30B-A3B-Instruct-2507 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 "Qwen/Qwen3-30B-A3B-Instruct-2507" \
    --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": "Qwen/Qwen3-30B-A3B-Instruct-2507",
		"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 "Qwen/Qwen3-30B-A3B-Instruct-2507" \
        --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": "Qwen/Qwen3-30B-A3B-Instruct-2507",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Qwen/Qwen3-30B-A3B-Instruct-2507 with Docker Model Runner:
```
docker model run hf.co/Qwen/Qwen3-30B-A3B-Instruct-2507
```

Lora Fine Tning- Several Issues

#25

by aetherforge - opened Nov 30, 2025

Discussion

aetherforge

Nov 30, 2025

Has anyone successfully trained this model? I have tried several ways including unsloth, Lora, etc. Error after error. Any tips or a training script that works on a L40S-180 would be awesome. Free access to L40S-90, 180, and 360(reasonable access, nothing crazy, you want to fine tune a model or 2, I'm good with that) to anyone that can help.

aetherforge

Nov 30, 2025

Quick context:setup 2× NVIDIA L40S on OVH (44 GB each, single node) PyTorch 2.5.1 + cu121 transformers 4.57.2 peft + trl installed from pip this week (can post exact versions if needed)
QLoRA, 4-bit, LoRA rank 32, target_modules:
["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"] Training on 372 train examples + 93 val examples JSONL, already formatted to chat / SFT style. Model and data load fine, tokenization runs, GPUs spin up, but as soon as the training loop starts it dies cleanly: Progress bar sits at 0/72 then exits nvidia-smi goes back to no python processes, No obvious Python traceback, just the usual warnings about
use_cache=True is incompatible with gradient checkpointing and
torch.utils.checkpoint: use_reentrant should be passed explicitly. Same basic behavior whether I try Unsloth’s SFTTrainer or a plain HF QLoRA trainer. If anyone has a minimal LoRA / QLoRA training script that actually runs on this checkpoint on a single 2×L40S box, I’m happy to adapt it and then share back a cleaned-up version here so others don’t have to fight this.

adityachhabra

Jan 26

Seems to be connected to this: https://github.com/pytorch/pytorch/issues/168329

Loads quickly without expert layers - ["up_proj", "down_proj"]

aetherforge

Mar 2

Thanks, I just spun up the latest qwen3.5-122b, its a beast

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment