Llama 3-8B-Instruct 在昇腾 NPU 上的 SGLang 性能实测

确认模型加载无误后，接下来配置 SGLang Engine。该模式支持直接在脚本中调用昇腾 NPU 执行推理。

创建 sglang_engine_setup.py 进行初始化：

import os
import sglang as sgl

os.environ['MAX_JOBS'] = '1'
os.environ['SGLANG_TARGET_BACKEND'] = 'ascend'
MODEL_PATH = os.path.expanduser("~/models/Llama-3-8B")

print("Initializing SGLang Engine (Backend: Ascend)...")
try:
    engine = sgl.Engine(
        model_path=MODEL_PATH,
        tp_size=1,              # 单卡即可
        trust_remote_code=True,
        backend="ascend",       # 指定使用昇腾 NPU
        dtype="float16"         # FP16 精度节省显存
    )
    print("✅ Engine initialized successfully!")
except Exception as e:
    print(f"❌ Initialization failed: {e}")
    raise

为了方便批量测试，可以封装一个推理函数：

BATCH_SIZE = 4
MAX_NEW_TOKENS = 50

def run_inference(prompts):
    outputs = []
    for prompt in prompts:
        out = engine.generate(prompt, max_new_tokens=MAX_NEW_TOKENS)
        outputs.append(out)
    return outputs

test_prompts = ["Hello world!"] * BATCH_SIZE
sample_output = run_inference(test_prompts)
print("Sample output:", sample_output[0])

性能基准测试

推理吞吐量测试

吞吐量是评估大模型推理性能的核心指标，通常关注 tokens/sec（每秒生成 token 数）和 samples/sec（每秒处理样本数）。吞吐量越高，越适合多用户并发或大批量离线生成场景。

import torch
import torch_npu
from transformers import AutoTokenizer, AutoModelForCausalLM
import time

model_name = "/path/to/your/model"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name, 
    torch_dtype=torch.float16, 
    device_map="npu"
)
model.eval()

prompt = "Describe the architecture of Ascend NPU."
inputs = tokenizer(prompt, return_tensors="pt").to("npu")

# Warmup
for _ in range(5):
    model.generate(**inputs, max_new_tokens=32)

num_iters = 20
total_tokens = 0
start = time.time()

for _ in range(num_iters):
    out = model.generate(**inputs, max_new_tokens=128)
    gen_tokens = out.shape[-1] - inputs["input_ids"].shape[-1]
    total_tokens += gen_tokens

end = time.time()
throughput = total_tokens / (end - start)
print(f"Throughput: {throughput:.2f} tokens/sec")

推理时延测试

时延（Latency）关注单个请求的响应速度，包括 E2E Latency（端到端）和 Per-token Latency（单 token 解码时间）。

import torch
import torch_npu
import time
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "/path/to/model"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name, 
    torch_dtype=torch.float16, 
    device_map="npu"
)
model.eval()

inputs = tokenizer("Hello, explain NPU.", return_tensors="pt").to("npu")

# Warmup
for _ in range(5):
    model.generate(**inputs, max_new_tokens=16)

# E2E Latency
start = time.time()
output = model.generate(**inputs, max_new_tokens=64)
end = time.time()
latency_ms = (end - start) * 1000
print(f"E2E Latency: {latency_ms:.2f} ms")

# Per-token Latency
input_len = inputs["input_ids"].shape[-1]
output_len = output.shape[-1]
gen_token_count = output_len - input_len
print(f"Per-Token Latency: {latency_ms/gen_token_count:.2f} ms/token")

显存占用监控

显存是部署大模型的关键限制因素。Ascend 提供 npu-smi 查看 HBM 使用情况，PyTorch 层面也可统计。

import torch_npu

allocated = torch_npu.memory.npu_memory_reserved()
cached = torch_npu.memory.npu_memory_allocated()
print(f"Reserved HBM: {allocated/1024/1024:.2f} MB")
print(f"Allocated HBM: {cached/1024/1024:.2f} MB")

系统命令同样有效：

npu-smi info

批量吞吐与自动化测试

不同 Batch Size 下的性能表现能反映 NPU 的并行利用率。以下脚本可自动测试不同配置下的延迟与吞吐量。

import torch
import torch_npu
import time
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "/path/to/your/model"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name, 
    torch_dtype=torch.float16, 
    device_map="npu"
)
model.eval()

def measure(bs=1, seq=128):
    text = "Ascend NPU performance test. " * (seq // 10)
    inputs = tokenizer([text] * bs, return_tensors="pt", padding=True, truncation=True).to("npu")
    
    # warmup
    for _ in range(3):
        model.generate(**inputs, max_new_tokens=32)
    
    start = time.time()
    out = model.generate(**inputs, max_new_tokens=seq)
    end = time.time()
    
    input_len = inputs["input_ids"].shape[-1]
    output_len = out.shape[-1]
    gen_tokens = (output_len - input_len) * bs
    latency = end - start
    throughput = gen_tokens / latency
    return latency, throughput, gen_tokens

print("batch_size, seq_len, latency(s), throughput(tokens/s)")
for bs in [1, 2, 4, 8, 16]:
    lat, th, tk = measure(bs=bs, seq=128)
    print(f"{bs}, 128, {lat:.3f}, {th:.2f}")

测试结果分析：

随着 batch size 增大，总吞吐量显著提升，虽然总延迟略有增加，但单 token 平均延迟下降，体现了 Ascend NPU 在并发推理中的强大能力。

压力测试

为了全面评估稳定性，我们进行了多维度的压力测试，涵盖批量吞吐量、延迟、长序列生成及多轮迭代。

import torch
import torch_npu
import time
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "/path/to/your/model"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name, 
    torch_dtype=torch.float16, 
    device_map="npu"
)
model.eval()

batch_sizes = [1, 2, 4, 8, 16]
seq_lengths = [64, 128, 256]
num_iters = 10
prompt = "Describe the architecture and optimization of Ascend NPU."

def stress_test(batch_size, seq_len):
    texts = [prompt] * batch_size
    inputs = tokenizer(texts, return_tensors="pt", padding=True, truncation=True).to("npu")
    
    # warmup
    for _ in range(3):
        model.generate(**inputs, max_new_tokens=32)
    
    total_tokens = 0
    total_latency = 0.0
    
    for _ in range(num_iters):
        start = time.time()
        output = model.generate(**inputs, max_new_tokens=seq_len)
        end = time.time()
        
        gen_tokens = (output.shape[-1] - inputs["input_ids"].shape[-1]) * batch_size
        total_tokens += gen_tokens
        total_latency += (end - start)
    
    avg_latency = total_latency / num_iters
    avg_throughput = total_tokens / total_latency
    return avg_latency, avg_throughput

print("Batch, SeqLen, AvgLatency(s), AvgThroughput(tokens/s)")
for seq_len in seq_lengths:
    for bs in batch_sizes:
        avg_lat, avg_th = stress_test(bs, seq_len)
        print(f"{bs}, {seq_len}, {avg_lat:.3f}, {avg_th:.2f}")

# 单 token 延迟分析
bs_test, seq_test = 4, 128
inputs = tokenizer([prompt]*bs_test, return_tensors="pt", padding=True, truncation=True).to("npu")
output = model.generate(**inputs, max_new_tokens=seq_test)
total_tokens = (output.shape[-1] - inputs["input_ids"].shape[-1]) * bs_test

start = time.time()
_ = model.generate(**inputs, max_new_tokens=seq_test)
end = time.time()

e2e_latency = end - start
per_token_latency = e2e_latency / total_tokens
print(f"\nE2E Latency for batch {bs_test}, seq {seq_test}: {e2e_latency:.3f}s")
print(f"Per-token Latency: {per_token_latency*1000:.2f} ms/token")

压力测试总结：

从结果来看，Llama 3-8B-Instruct 在 SGLang 调度下，Ascend NPU 能够在大批量、高并发和长序列生成场景中保持高吞吐和低延迟。即使在极端配置下（如 Batch 16 + Seq 256），吞吐量依然稳定，证明其完全胜任实际生产环境的部署需求。

总结

本次实测表明，Llama 3-8B-Instruct 在 Ascend NPU 上配合 SGLang 框架，展现出极高的推理效率。无论是吞吐量还是延迟控制，都在不同并发规模下表现优异，且显存占用得到有效管理。开发者可以直接基于此方案快速完成模型部署与性能调优，无需过度依赖特定云平台，重点在于环境驱动的正确配置与参数调优。