Single Node (Atlas 300I Series)

Single Node (Atlas 300I Series)#

备注

  1. This Atlas 300I series is currently experimental. In future versions, there may be behavioral changes related to model coverage and performance improvement.

  2. Currently, the 310I series only supports eager mode and the float16 data type.

  3. There are some known issues for running vLLM on 310p series, you can refer to vllm-ascend #3316 and #2795. You can use v0.10.0rc1 version first.

Run vLLM on Atlas 300I Series#

运行 docker 容器：

# Update the vllm-ascend image
export IMAGE=quay.io/ascend/vllm-ascend:v0.10.0rc1-310p
docker run --rm \
--name vllm-ascend \
--device /dev/davinci0 \
--device /dev/davinci1 \
--device /dev/davinci2 \
--device /dev/davinci3 \
--device /dev/davinci4 \
--device /dev/davinci5 \
--device /dev/davinci6 \
--device /dev/davinci7 \
--device /dev/davinci_manager \
--device /dev/devmm_svm \
--device /dev/hisi_hdc \
-v /usr/local/dcmi:/usr/local/dcmi \
-v /usr/local/bin/npu-smi:/usr/local/bin/npu-smi \
-v /usr/local/Ascend/driver/lib64/:/usr/local/Ascend/driver/lib64/ \
-v /usr/local/Ascend/driver/version.info:/usr/local/Ascend/driver/version.info \
-v /etc/ascend_install.info:/etc/ascend_install.info \
-v /root/.cache:/root/.cache \
-p 8000:8000 \
-it $IMAGE bash

Set up environment variables:

# Load model from ModelScope to speed up download
export VLLM_USE_MODELSCOPE=True

# Set `max_split_size_mb` to reduce memory fragmentation and avoid out of memory
export PYTORCH_NPU_ALLOC_CONF=max_split_size_mb:256

在NPU上进行在线推理#

Run the following script to start the vLLM server on NPU (Qwen3-0.6B:1 card, Qwen2.5-7B-Instruct:2 cards, Pangu-Pro-MoE-72B: 8 cards):

运行以下命令以启动 vLLM 服务器：

vllm serve Qwen/Qwen3-0.6B \
    --tensor-parallel-size 1 \
    --enforce-eager \
    --dtype float16 \
    --compilation-config '{"custom_ops":["none", "+rms_norm", "+rotary_embedding"]}'

一旦服务器启动，就可以通过输入提示词来查询模型。

curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "The future of AI is",
    "max_tokens": 64,
    "top_p": 0.95,
    "top_k": 50,
    "temperature": 0.6
  }'

运行以下命令以启动 vLLM 服务器：

vllm serve Qwen/Qwen2.5-7B-Instruct \
    --tensor-parallel-size 2 \
    --enforce-eager \
    --dtype float16 \
    --compilation-config '{"custom_ops":["none", "+rms_norm", "+rotary_embedding"]}'

一旦服务器启动，就可以通过输入提示词来查询模型。

curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "The future of AI is",
    "max_tokens": 64,
    "top_p": 0.95,
    "top_k": 50,
    "temperature": 0.6
  }'

运行以下命令以启动 vLLM 服务器：

vllm serve Qwen/Qwen2.5-VL-3B-Instruct \
    --tensor-parallel-size 1 \
    --enforce-eager \
    --dtype float16 \
    --compilation-config '{"custom_ops":["none", "+rms_norm", "+rotary_embedding"]}'

一旦服务器启动，就可以通过输入提示词来查询模型。

curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "The future of AI is",
    "max_tokens": 64,
    "top_p": 0.95,
    "top_k": 50,
    "temperature": 0.6
  }'

下载该模型：

git lfs install
git clone https://gitcode.com/ascend-tribe/pangu-pro-moe-model.git

运行以下命令以启动 vLLM 服务器：

vllm serve /home/pangu-pro-moe-mode/ \
--tensor-parallel-size 4 \
--enable-expert-parallel \
--dtype "float16" \
--trust-remote-code \
--enforce-eager

Once your server is started, you can query the model with input prompts.

export question="你是谁？"
curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "prompt": "[unused9]系统：[unused10][unused9]用户：'${question}'[unused10][unused9]助手：",
    "max_tokens": 64,
    "top_p": 0.95,
    "top_k": 50,
    "temperature": 0.6
  }'

如果你成功运行此脚本，你就可以看到结果。

离线推理#

运行以下脚本（example.py）以在 NPU 上执行离线推理：

from vllm import LLM, SamplingParams
import gc
import torch
from vllm import LLM, SamplingParams
from vllm.distributed.parallel_state import (destroy_distributed_environment,
                                             destroy_model_parallel)

def clean_up():
    destroy_model_parallel()
    destroy_distributed_environment()
    gc.collect()
    torch.npu.empty_cache()
prompts = [
    "Hello, my name is",
    "The future of AI is",
]
# Create a sampling params object.
sampling_params = SamplingParams(max_tokens=100, temperature=0.0)
# Create an LLM.
llm = LLM(
    model="Qwen/Qwen3-0.6B",
    tensor_parallel_size=1,
    enforce_eager=True, # For 300I series, only eager mode is supported.
    dtype="float16", # IMPORTANT cause some ATB ops cannot support bf16 on 300I series
    compilation_config={"custom_ops":["none", "+rms_norm", "+rotary_embedding"]}, # High performance for 300I series
)
# Generate texts from the prompts.
outputs = llm.generate(prompts, sampling_params)
for output in outputs:
    prompt = output.prompt
    generated_text = output.outputs[0].text
    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
del llm
clean_up()

from vllm import LLM, SamplingParams
import gc
import torch
from vllm import LLM, SamplingParams
from vllm.distributed.parallel_state import (destroy_distributed_environment,
                                             destroy_model_parallel)

def clean_up():
    destroy_model_parallel()
    destroy_distributed_environment()
    gc.collect()
    torch.npu.empty_cache()
prompts = [
    "Hello, my name is",
    "The future of AI is",
]
# Create a sampling params object.
sampling_params = SamplingParams(max_tokens=100, temperature=0.0)
# Create an LLM.
llm = LLM(
    model="Qwen/Qwen2.5-7B-Instruct",
    tensor_parallel_size=2,
    enforce_eager=True, # For 300I series, only eager mode is supported.
    dtype="float16", # IMPORTANT cause some ATB ops cannot support bf16 on 300I series
    compilation_config={"custom_ops":["none", "+rms_norm", "+rotary_embedding"]}, # High performance for 300I series
)
# Generate texts from the prompts.
outputs = llm.generate(prompts, sampling_params)
for output in outputs:
    prompt = output.prompt
    generated_text = output.outputs[0].text
    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
del llm
clean_up()

from vllm import LLM, SamplingParams
import gc
import torch
from vllm import LLM, SamplingParams
from vllm.distributed.parallel_state import (destroy_distributed_environment,
                                             destroy_model_parallel)

def clean_up():
    destroy_model_parallel()
    destroy_distributed_environment()
    gc.collect()
    torch.npu.empty_cache()
prompts = [
    "Hello, my name is",
    "The future of AI is",
]
# Create a sampling params object.
sampling_params = SamplingParams(max_tokens=100, top_p=0.95, top_k=50, temperature=0.6)
# Create an LLM.
llm = LLM(
    model="Qwen/Qwen2.5-VL-3B-Instruct",
    tensor_parallel_size=1,
    enforce_eager=True, # For 300I series, only eager mode is supported.
    dtype="float16", # IMPORTANT cause some ATB ops cannot support bf16 on 300I series
    compilation_config={"custom_ops":["none", "+rms_norm", "+rotary_embedding"]}, # High performance for 300I series
)
# Generate texts from the prompts.
outputs = llm.generate(prompts, sampling_params)
for output in outputs:
    prompt = output.prompt
    generated_text = output.outputs[0].text
    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
del llm
clean_up()

下载该模型：

git lfs install
git clone https://gitcode.com/ascend-tribe/pangu-pro-moe-model.git

import gc
from transformers import AutoTokenizer
import torch

from vllm import LLM, SamplingParams
from vllm.distributed.parallel_state import (destroy_distributed_environment,
                                             destroy_model_parallel)

def clean_up():
    destroy_model_parallel()
    destroy_distributed_environment()
    gc.collect()
    torch.npu.empty_cache()


if __name__ == "__main__":

    tokenizer = AutoTokenizer.from_pretrained("/home/pangu-pro-moe-mode/", trust_remote_code=True)
    tests = [
        "Hello, my name is",
        "The future of AI is",
    ]
    prompts = []
    for text in tests:
        messages = [
        {"role": "system", "content": ""},    # Optionally customize system content
        {"role": "user", "content": text}
    ]
        prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)        # 推荐使用官方的template
        prompts.append(prompt)
    sampling_params = SamplingParams(temperature=0.6, top_p=0.95, top_k=40)

    llm = LLM(model="/home/pangu-pro-moe-mode/",
            tensor_parallel_size=8,
            distributed_executor_backend="mp",
            enable_expert_parallel=True,
            dtype="float16",
            max_model_len=1024,
            trust_remote_code=True,
            enforce_eager=True)

    outputs = llm.generate(prompts, sampling_params)
    for output in outputs:
        prompt = output.prompt
        generated_text = output.outputs[0].text
        print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")

    del llm
    clean_up()

运行脚本：

python example.py

如果你成功运行此脚本，你可以看到如下所示的信息：

Prompt: 'Hello, my name is', Generated text: " Lina. I'm a 22-year-old student from China. I'm interested in studying in the US. I'm looking for a job in the US. I want to know if there are any opportunities in the US for me to work. I'm also interested in the culture and lifestyle in the US. I want to know if there are any opportunities for me to work in the US. I'm also interested in the culture and lifestyle in the US. I'm interested in the culture"
Prompt: 'The future of AI is', Generated text: " not just about the technology itself, but about how we use it to solve real-world problems. As AI continues to evolve, it's important to consider the ethical implications of its use. AI has the potential to bring about significant changes in society, but it also has the power to create new challenges. Therefore, it's crucial to develop a comprehensive approach to AI that takes into account both the benefits and the risks associated with its use. This includes addressing issues such as bias, privacy, and accountability."
目录