Qwen3 Omni AudioTransformer EU AI Act Compliance Profile

Qwen/Qwen3-Omni-30B-A3B-InstructのAudioEncoderを抽出した重みです。ライセンスは元モデルに従います。

詳細はQwen3-Omniテクニカルレポートを参照。

pip install git+https://github.com/huggingface/transformers

Usage

import soundfile as sf
import torch
from transformers import WhisperFeatureExtractor
from transformers.models.qwen3_omni_moe.modeling_qwen3_omni_moe import Qwen3OmniMoeAudioEncoder

model_id = "Atotti/Qwen3-Omni-AudioTransformer"
feature_extractor = WhisperFeatureExtractor.from_pretrained(model_id)
qwen_encoder = Qwen3OmniMoeAudioEncoder.from_pretrained(model_id, dtype=torch.bfloat16).eval()

wav, sr = sf.read("audio_16k.wav", dtype="float32")

features = feature_extractor(wav, sampling_rate=sr, return_tensors="pt", padding=False)
x = features["input_features"][0].to(qwen_encoder.dtype)
feat_len = torch.tensor([x.shape[-1]], dtype=torch.long)

with torch.no_grad():
    out = qwen_encoder(input_features=x, feature_lens=feat_len)
    hid = out.last_hidden_state # [T', D]
    print(hid.shape) # torch.Size([37, 2048])
    print(hid[0, :10]) # tensor([ 0.0088, -0.0032,  0.0154, -0.0138,  0.0040, -0.0037, -0.0017,  0.0180, -0.0149,  0.0051])

Your obligations as a Qwen3 Omni AudioTransformer deployer

Track your progress

Sign up to track obligations, receive deadline reminders, and generate audit packages.

Tags

Base Model