from transformers import AutoModelForCausalLM, AutoTokenizer
device = "cuda" # the device to load the model onto

model = AutoModelForCausalLM.from_pretrained(
    "Hack337/WavGPT-2.5",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("Hack337/WavGPT-2.5")

prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "system", "content": "Вы очень полезный помощник."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)

generated_ids = model.generate(
    model_inputs.input_ids,
    max_new_tokens=512
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

Use the code below to get started with the model using NPU.

python

from transformers import AutoTokenizer, TextStreamer
from intel_npu_acceleration_library import NPUModelForCausalLM
import torch

# Load the NPU-optimized model without LoRA
model = NPUModelForCausalLM.from_pretrained(
    "Hack337/WavGPT-2.5",
    use_cache=True,
    dtype=torch.float16  # Use float16 for the NPU
).eval()

# Load the tokenizer
tokenizer = AutoTokenizer.from_pretrained("Hack337/WavGPT-2.5")
tokenizer.pad_token_id = tokenizer.eos_token_id
streamer = TextStreamer(tokenizer, skip_special_tokens=True)

# Prompt handling
prompt = "Give me a short introduction to large language model."
messages = [
    {"role": "system", "content": "Вы очень полезный помощник."},
    {"role": "user", "content": prompt}
]

# Convert to a text format compatible with the model
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
prefix = tokenizer([text], return_tensors="pt")["input_ids"].to("npu")

# Generation configuration
generation_kwargs = dict(
    input_ids=prefix,
    streamer=streamer,
    do_sample=True,
    top_k=50,
    top_p=0.9,
    max_new_tokens=512,
)

# Run inference on the NPU
print("Run inference")
_ = model.generate(**generation_kwargs)

PEFT 0.11.1