GPA-0.3B-preview（昇腾NPU适配版）

pip install modelscope transformers torch torch_npu

modelscope download --model AutoArk/GPA --local_dir ./GPA_model

import torch
import torch_npu
from transformers import AutoModelForCausalLM, AutoTokenizer

MODEL_DIR = "./GPA_model"
device = torch.device("npu:0")

model = AutoModelForCausalLM.from_pretrained(
    MODEL_DIR,
    torch_dtype=torch.bfloat16,
    trust_remote_code=True,
).to(device)
model.eval()

tokenizer = AutoTokenizer.from_pretrained(MODEL_DIR, trust_remote_code=True)

inputs = tokenizer("Hello, who are you?", return_tensors="pt").to(device)
outputs = model.generate(**inputs, max_new_tokens=128)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

# Text generation
python inference.py --task text --prompt "Hello!" --max_new_tokens 128

# Accuracy validation (NPU vs CPU)
python inference.py --task accuracy

# Performance benchmark
python inference.py --task benchmark --num_iters 10

# Full evaluation (accuracy + benchmark + demo)
python inference.py --task full --output eval_results.json

.
├── inference.py              # NPU inference script (text/accuracy/benchmark/memory/full)
├── README.md                 # This documentation
├── results_accuracy.json     # Accuracy evaluation results (JSON)
├── results_benchmark.json    # Performance benchmark results (JSON)
├── results_text.json         # Text generation results (JSON)
└── GPA_model/               # Model files (downloaded from ModelScope)
    ├── config.json
    ├── model.safetensors
    ├── modeling_qwen3.py
    ├── configuration_qwen3.py
    ├── tokenizer.json
    ├── tokenizer_config.json
    ├── vocab.json
    ├── merges.txt
    ├── generation_config.json
    ├── chat_template.jinja
    ├── special_tokens_map.json
    ├── added_tokens.json
    ├── BiCodec/
    │   ├── config.json
    │   ├── model.safetensors
    │   └── wav2vec2-large-xlsr-53/
    └── glm-4-voice-tokenizer/
        ├── config.json
        ├── model.safetensors
        └── preprocessor_config.json

@misc{gpa2026,
  title={Unifying Speech Recognition, Synthesis and Conversion with Autoregressive Transformer},
  author={Runyuan Cai, Yu Lin, Yiming Wang, Chunlin Fu and Xiaodong Zeng},
  year={2026},
  howpublished={\url{https://github.com/AutoArk/GPA}},
}