File size: 1,748 Bytes
5ee43e9 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 | #!/usr/bin/env python3
# MPNet sentence-embedding on Neuron
import argparse
import logging
import time
import torch
from transformers import AutoTokenizer, MPNetModel
import torch_neuronx # ensures Neuron backend
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
def main():
parser = argparse.ArgumentParser(description="Run MPNet encoder on Neuron")
parser.add_argument(
"--model",
type=str,
default="microsoft/mpnet-base",
help="MPNet model name on Hugging Face Hub",
)
args = parser.parse_args()
torch.set_default_dtype(torch.float32)
torch.manual_seed(42)
# load tokenizer & model
tokenizer = AutoTokenizer.from_pretrained(args.model)
model = MPNetModel.from_pretrained(
args.model, torch_dtype=torch.float32, attn_implementation="eager"
).eval()
# tokenize sample sentence
text = "MPNet is a variant of BERT with permutation language modeling."
inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True)
# pre-run to lock shapes
with torch.no_grad():
_ = model(**inputs).pooler_output
# compile
model.forward = torch.compile(model.forward, backend="neuron", fullgraph=True)
# warmup
warmup_start = time.time()
with torch.no_grad():
_ = model(**inputs)
warmup_time = time.time() - warmup_start
# benchmark run
run_start = time.time()
with torch.no_grad():
embeddings = model(**inputs).pooler_output
run_time = time.time() - run_start
logger.info("Warmup: %.2f s, Run: %.4f s", warmup_time, run_time)
logger.info("Output embedding shape: %s", embeddings.shape) # [1, hidden]
if __name__ == "__main__":
main() |