benchmark_llama.py

#!/usr/bin/env python3
from transformers import AutoTokenizer, AutoModelForCausalLM
import time
import torch

DEVICE = "cuda:1"

tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
model = AutoModelForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf", torch_dtype=torch.float16, low_cpu_mem_usage=True)
model.to(DEVICE)


# forward
print("Forward benchmarks")
print(50 * "=")

for batch_size in (1, 4, 16):
    for input_seq in (4, 16, 256):
        input_ids = torch.ones((batch_size, input_seq), dtype=torch.long, device=DEVICE)
        attention_mask = torch.ones_like(input_ids)
        attention_mask[0, 3] = 0

        times = []
        for _ in range(3):
            start_time = time.time()
            with torch.no_grad():
                logits = model(input_ids=input_ids, attention_mask=attention_mask).logits
            times.append(time.time() - start_time)

        result = min(times)

        print(f"Forward bsz={batch_size}, input_seq={input_seq}: {result}")


# generate
print("Generate benchmarks")
print(50 * "=")

for batch_size in (1, 16):
    for input_seq in (4, 256):
        input_ids = torch.ones((batch_size, input_seq), dtype=torch.long, device=DEVICE)
        attention_mask = torch.ones_like(input_ids)
        attention_mask[0, 3] = 0

        times = []
        for _ in range(3):
            start_time = time.time()
            out = model.generate(input_ids=input_ids, max_new_tokens=256, do_sample=False)
            times.append(time.time() - start_time)

        result = min(times)

        print(f"Generate bsz={batch_size}, input_seq={input_seq}: {result}")