EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Actions Projects Releases Wiki Activity

[Docs] Re-arch on doc and make QwQ doc work (#271)

Browse Source 
### What this PR does / why we need it?
Re-arch on tutorials, move singe npu / multi npu / multi node to index.
- Unifiy docker run cmd
- Use dropdown to hide build from source installation doc
- Re-arch tutorials to include Qwen/QwQ/DeepSeek
- Make QwQ doc works

### Does this PR introduce _any_ user-facing change?
No

### How was this patch tested?
CI test



Signed-off-by: Yikun Jiang <yikunkero@gmail.com>
This commit is contained in:
Yikun Jiang
2025-03-10 09:27:48 +08:00
committed by GitHub
parent 18bb8d1f52
commit 38334f5daa
11 changed files with 414 additions and 357 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
docs/source/conf.py
View File

@@ -72,6 +72,8 @@ myst_substitutions = {
# This value should be updated when cut down release.
'pip_vllm_ascend_version': "0.7.3rc1",
'pip_vllm_version': "0.7.3",
# CANN image tag
'cann_image_tag': "8.0.0-910b-ubuntu22.04-py3.10",
}
# Add any paths that contain templates here, relative to this directory.

10
docs/source/developer_guide/contributing.md
View File

@@ -53,21 +53,21 @@ locally. The simplest way to run these integration tests locally is through a co
git clone https://github.com/vllm-project/vllm-ascend.git
cd vllm-ascend
IMAGE=vllm-ascend-dev-image
CONTAINER_NAME=vllm-ascend-dev
DEVICE=/dev/davinci1
export IMAGE=vllm-ascend-dev-image
export CONTAINER_NAME=vllm-ascend-dev
export DEVICE=/dev/davinci1
# The first build will take about 10 mins (10MB/s) to download the base image and packages
docker build -t $IMAGE -f ./Dockerfile .
# You can also specify the mirror repo via setting VLLM_REPO to speedup
# docker build -t $IMAGE -f ./Dockerfile . --build-arg VLLM_REPO=https://gitee.com/mirrors/vllm
docker run --name $CONTAINER_NAME --network host --device $DEVICE \
docker run --rm --name $CONTAINER_NAME --network host --device $DEVICE \
--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/ \
-ti --rm $IMAGE bash
-ti $IMAGE bash
cd vllm-ascend
pip install -r requirements-dev.txt

10
docs/source/developer_guide/contributing.zh.md
View File

@@ -48,21 +48,21 @@ git commit -sm "your commit info"
git clone https://github.com/vllm-project/vllm-ascend.git
cd vllm-ascend
IMAGE=vllm-ascend-dev-image
CONTAINER_NAME=vllm-ascend-dev
DEVICE=/dev/davinci1
export IMAGE=vllm-ascend-dev-image
export CONTAINER_NAME=vllm-ascend-dev
export DEVICE=/dev/davinci1
# 首次构建会花费10分钟（10MB/s）下载基础镜像和包
docker build -t $IMAGE -f ./Dockerfile .
# 您还可以通过设置 VLLM_REPO 来指定镜像仓库以加速
# docker build -t $IMAGE -f ./Dockerfile . --build-arg VLLM_REPO=https://gitee.com/mirrors/vllm
docker run --name $CONTAINER_NAME --network host --device $DEVICE \
docker run --rm --name $CONTAINER_NAME --network host --device $DEVICE \
--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/ \
-ti --rm $IMAGE bash
-ti $IMAGE bash
cd vllm-ascend
pip install -r requirements-dev.txt

2
docs/source/index.md
View File

@@ -35,7 +35,7 @@ By using vLLM Ascend plugin, popular open-source models, including Transformer-l
:maxdepth: 1
quick_start
installation
tutorials
tutorials/index.md
faqs
:::

41
docs/source/installation.md
View File

@@ -44,10 +44,12 @@ Refer to [Ascend Environment Setup Guide](https://ascend.github.io/docs/sources/
The easiest way to prepare your software environment is using CANN image directly:
```bash
```{code-block} bash
:substitutions:
# Update DEVICE according to your device (/dev/davinci[0-7])
export DEVICE=/dev/davinci7
# Update the vllm-ascend image
export IMAGE=quay.io/ascend/cann:|cann_image_tag|
docker run --rm \
--name vllm-ascend-env \
--device $DEVICE \
@@ -59,14 +61,16 @@ docker run --rm \
-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 \
-it quay.io/ascend/cann:8.0.0-910b-ubuntu22.04-py3.10 bash
-it $IMAGE bash
```
:::{dropdown} Click here to see "Install CANN manally"
:animate: fade-in-slide-down
You can also install CANN manually:
:::{note}
```{note}
This guide takes aarch64 as an example. If you run on x86, you need to replace `aarch64` with `x86_64` for the package name shown below.
:::
```
```bash
# Create a virtual environment
@@ -94,6 +98,8 @@ chmod +x. /Ascend-cann-nnal_8.0.0_linux-aarch64.run
source /usr/local/Ascend/nnal/atb/set_env.sh
```
:::
::::
::::{tab-item} Before using docker
@@ -125,6 +131,7 @@ pip install vllm==|pip_vllm_version|
pip install vllm-ascend==|pip_vllm_ascend_version| --extra-index https://download.pytorch.org/whl/cpu/
```
:::{dropdown} Click here to see "Build from source code"
or build from **source code**:
```{code-block} bash
@@ -140,6 +147,7 @@ git clone --depth 1 --branch |vllm_ascend_version| https://github.com/vllm-proj
cd vllm-ascend
pip install -e . --extra-index https://download.pytorch.org/whl/cpu/
```
:::
Current version depends on a unreleased `torch-npu`, you need to install manually:
@@ -167,14 +175,23 @@ pip install ./torch_npu-2.5.1.dev20250308-cp310-cp310-manylinux_2_17_aarch64.man
You can just pull the **prebuilt image** and run it with bash.
:::{dropdown} Click here to see "Build from Dockerfile"
or build IMAGE from **source code**:
```bash
git clone https://github.com/vllm-project/vllm-ascend.git
cd vllm-ascend
docker build -t vllm-ascend-dev-image:latest -f ./Dockerfile .
```
:::
```{code-block} bash
:substitutions:
# Update DEVICE according to your device (/dev/davinci[0-7])
DEVICE=/dev/davinci7
export DEVICE=/dev/davinci7
# Update the vllm-ascend image
IMAGE=quay.io/ascend/vllm-ascend:|vllm_ascend_version|
docker pull $IMAGE
export IMAGE=quay.io/ascend/vllm-ascend:|vllm_ascend_version|
docker run --rm \
--name vllm-ascend-env \
--device $DEVICE \
@@ -189,14 +206,6 @@ docker run --rm \
-it $IMAGE bash
```
or build IMAGE from **source code**:
```bash
git clone https://github.com/vllm-project/vllm-ascend.git
cd vllm-ascend
docker build -t vllm-ascend-dev-image:latest -f ./Dockerfile .
```
::::
:::::

31
docs/source/quick_start.md
View File

@@ -11,12 +11,13 @@
```{code-block} bash
:substitutions:
# You can change version a suitable one base on your requirement, e.g. main
# Update DEVICE according to your device (/dev/davinci[0-7])
export DEVICE=/dev/davinci0
# Update the vllm-ascend image
export IMAGE=quay.io/ascend/vllm-ascend:|vllm_ascend_version|
docker run \
docker run --rm \
--name vllm-ascend \
--device /dev/davinci0 \
--device $DEVICE \
--device /dev/davinci_manager \
--device /dev/devmm_svm \
--device /dev/hisi_hdc \
@@ -32,17 +33,19 @@ docker run \
## Usage
There are two ways to start vLLM on Ascend NPU:
### Offline Batched Inference with vLLM
With vLLM installed, you can start generating texts for list of input prompts (i.e. offline batch inferencing).
You can use Modelscope mirror to speed up download:
```bash
# Use Modelscope mirror to speed up download
export VLLM_USE_MODELSCOPE=true
```
There are two ways to start vLLM on Ascend NPU:
:::::{tab-set}
::::{tab-item} Offline Batched Inference
With vLLM installed, you can start generating texts for list of input prompts (i.e. offline batch inferencing).
Try to run below Python script directly or use `python3` shell to generate texts:
```python
@@ -64,15 +67,15 @@ for output in outputs:
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
```
### OpenAI Completions API with vLLM
::::
::::{tab-item} OpenAI Completions API
vLLM can also be deployed as a server that implements the OpenAI API protocol. Run
the following command to start the vLLM server with the
[Qwen/Qwen2.5-0.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-0.5B-Instruct) model:
```bash
# Use Modelscope mirror to speed up download
export VLLM_USE_MODELSCOPE=true
# Deploy vLLM server (The first run will take about 3-5 mins (10 MB/s) to download models)
vllm serve Qwen/Qwen2.5-0.5B-Instruct &
```
@@ -124,3 +127,5 @@ INFO: Application shutdown complete.
```
Finally, you can exit container by using `ctrl-D`.
::::
:::::

317
docs/source/tutorials.md
View File

@@ -1,317 +0,0 @@
# Tutorials
## Run vllm-ascend on Single NPU
### Offline Inference on Single NPU
Run docker container:
```{code-block} bash
:substitutions:
docker run \
--name vllm-ascend \
--device /dev/davinci0 \
--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 quay.io/ascend/vllm-ascend:|vllm_ascend_version| bash
```
Setup environment variables:
```bash
# Use Modelscope mirror to speed up model download
export VLLM_USE_MODELSCOPE=True
# To avoid NPU out of memory, set `max_split_size_mb` to any value lower than you need to allocate for Qwen2.5-7B-Instruct
export PYTORCH_NPU_ALLOC_CONF=max_split_size_mb:256
```
:::{note}
`max_split_size_mb` prevents the native allocator from splitting blocks larger than this size (in MB). This can reduce fragmentation and may allow some borderline workloads to complete without running out of memory. You can find more details [<u>here</u>](https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/800alpha003/apiref/envref/envref_07_0061.html).
:::
Run the following script to execute offline inference on a single NPU:
```python
from vllm import LLM, SamplingParams
prompts = [
"Hello, my name is",
"The future of AI is",
]
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
llm = LLM(model="Qwen/Qwen2.5-7B-Instruct", max_model_len=26240)
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}")
```
If you run this script successfully, you can see the info shown below:
```bash
Prompt: 'Hello, my name is', Generated text: ' Daniel and I am an 8th grade student at York Middle School. I'
Prompt: 'The future of AI is', Generated text: ' following you. As the technology advances, a new report from the Institute for the'
```
### Online Serving on Single NPU
Run docker container to start the vLLM server on a single NPU:
```{code-block} bash
:substitutions:
docker run \
--name vllm-ascend \
--device /dev/davinci0 \
--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 \
-e VLLM_USE_MODELSCOPE=True \
-e PYTORCH_NPU_ALLOC_CONF=max_split_size_mb:256 \
-it quay.io/ascend/vllm-ascend:|vllm_ascend_version| \
vllm serve Qwen/Qwen2.5-7B-Instruct --max_model_len 26240
```
:::{note}
Add `--max_model_len` option to avoid ValueError that the Qwen2.5-7B model's max seq len (32768) is larger than the maximum number of tokens that can be stored in KV cache (26240). This will differ with different NPU series base on the HBM size. Please modify the value according to a suitable value for your NPU series.
:::
If your service start successfully, you can see the info shown below:
```bash
INFO: Started server process [6873]
INFO: Waiting for application startup.
INFO: Application startup complete.
```
Once your server is started, you can query the model with input prompts:
```bash
curl http://localhost:8000/v1/completions \
-H "Content-Type: application/json" \
-d '{
"model": "Qwen/Qwen2.5-7B-Instruct",
"prompt": "The future of AI is",
"max_tokens": 7,
"temperature": 0
}'
```
If you query the server successfully, you can see the info shown below (client):
```bash
{"id":"cmpl-b25a59a2f985459781ce7098aeddfda7","object":"text_completion","created":1739523925,"model":"Qwen/Qwen2.5-7B-Instruct","choices":[{"index":0,"text":" here. It’s not just a","logprobs":null,"finish_reason":"length","stop_reason":null,"prompt_logprobs":null}],"usage":{"prompt_tokens":5,"total_tokens":12,"completion_tokens":7,"prompt_tokens_details":null}}
```
Logs of the vllm server:
```bash
INFO: 172.17.0.1:49518 - "POST /v1/completions HTTP/1.1" 200 OK
INFO 02-13 08:34:35 logger.py:39] Received request cmpl-574f00e342904692a73fb6c1c986c521-0: prompt: 'San Francisco is a', params: SamplingParams(n=1, presence_penalty=0.0, frequency_penalty=0.0, repetition_penalty=1.0, temperature=0.0, top_p=1.0, top_k=-1, min_p=0.0, seed=None, stop=[], stop_token_ids=[], bad_words=[], include_stop_str_in_output=False, ignore_eos=False, max_tokens=7, min_tokens=0, logprobs=None, prompt_logprobs=None, skip_special_tokens=True, spaces_between_special_tokens=True, truncate_prompt_tokens=None, guided_decoding=None), prompt_token_ids: [23729, 12879, 374, 264], lora_request: None, prompt_adapter_request: None.
```
## Run vllm-ascend on Multi-NPU
### Distributed Inference on Multi-NPU
Run docker container:
```{code-block} bash
:substitutions:
docker run \
--name vllm-ascend \
--device /dev/davinci0 \
--device /dev/davinci1 \
--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 quay.io/ascend/vllm-ascend:|vllm_ascend_version| bash
```
Setup environment variables:
```bash
# Use Modelscope mirror to speed up model download
export VLLM_USE_MODELSCOPE=True
# To avoid NPU out of memory, set `max_split_size_mb` to any value lower than you need to allocate for Qwen2.5-7B-Instruct
export PYTORCH_NPU_ALLOC_CONF=max_split_size_mb:256
```
Run the following script to execute offline inference on multi-NPU:
```python
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",
]
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
llm = LLM(model="Qwen/Qwen2.5-7B-Instruct",
tensor_parallel_size=2,
distributed_executor_backend="mp",
max_model_len=26240)
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()
```
If you run this script successfully, you can see the info shown below:
```bash
Prompt: 'Hello, my name is', Generated text: ' Daniel and I am an 8th grade student at York Middle School. I'
Prompt: 'The future of AI is', Generated text: ' following you. As the technology advances, a new report from the Institute for the'
```
## Online Serving on Multi Machine
Run docker container on each machine:
```{code-block} bash
:substitutions:
docker run \
--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 quay.io/ascend/vllm-ascend:|vllm_ascend_version| bash
```
Choose one machine as head node, the other are worker nodes, then start ray on each machine:
:::{note}
Check out your `nic_name` by command `ip addr`.
:::
```shell
# Head node
export HCCL_IF_IP={local_ip}
export GLOO_SOCKET_IFNAME={nic_name}
export TP_SOCKET_IFNAME={nic_name}
export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES=1
ray start --head --num-gpus=8
# Worker node
export HCCL_IF_IP={local_ip}
export ASCEND_PROCESS_LOG_PATH={plog_save_path}
export GLOO_SOCKET_IFNAME={nic_name}
export TP_SOCKET_IFNAME={nic_name}
export RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES=1
export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
ray start --address='{head_node_ip}:{port_num}' --num-gpus=8 --node-ip-address={local_ip}
```
Start the vLLM server on head node:
```shell
export VLLM_HOST_IP={head_node_ip}
export HCCL_CONNECT_TIMEOUT=120
export ASCEND_PROCESS_LOG_PATH={plog_save_path}
export HCCL_IF_IP={head_node_ip}
if [ -d "{plog_save_path}" ]; then
rm -rf {plog_save_path}
echo ">>> remove {plog_save_path}"
fi
LOG_FILE="multinode_$(date +%Y%m%d_%H%M).log"
VLLM_TORCH_PROFILER_DIR=./vllm_profile
python -m vllm.entrypoints.openai.api_server \
--model="Deepseek/DeepSeek-V2-Lite-Chat" \
--trust-remote-code \
--enforce-eager \
--max-model-len {max_model_len} \
--distributed_executor_backend "ray" \
--tensor-parallel-size 16 \
--disable-log-requests \
--disable-log-stats \
--disable-frontend-multiprocessing \
--port {port_num} \
```
Once your server is started, you can query the model with input prompts:
```shell
curl -X POST http://127.0.0.1:{prot_num}/v1/completions \
-H "Content-Type: application/json" \
-d '{
"model": "Deepseek/DeepSeek-V2-Lite-Chat",
"prompt": "The future of AI is",
"max_tokens": 24
}'
```
If you query the server successfully, you can see the info shown below (client):
```
{"id":"cmpl-6dfb5a8d8be54d748f0783285dd52303","object":"text_completion","created":1739957835,"model":"/home/data/DeepSeek-V2-Lite-Chat/","choices":[{"index":0,"text":" heavily influenced by neuroscience and cognitiveGuionistes. The goalochondria is to combine the efforts of researchers, technologists,","logprobs":null,"finish_reason":"length","stop_reason":null,"prompt_logprobs":null}],"usage":{"prompt_tokens":6,"total_tokens":30,"completion_tokens":24,"prompt_tokens_details":null}}
```
Logs of the vllm server:
```
INFO: 127.0.0.1:59384 - "POST /v1/completions HTTP/1.1" 200 OK
INFO 02-19 17:37:35 metrics.py:453] Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1.9 tokens/s, Running: 0 reqs, Swapped: 0 reqs, Pending: 0 reqs, GPU KV cache usage: 0.0%, CPU KV cache usage: 0.0%.
```

9
docs/source/tutorials/index.md Normal file
View File

@@ -0,0 +1,9 @@
# Tutorials
:::{toctree}
:caption: Deployment
:maxdepth: 1
single_npu
multi_npu
multi_node
:::

109
docs/source/tutorials/multi_node.md Normal file
View File

@@ -0,0 +1,109 @@
# Multi-Node (DeepSeek)
## Online Serving on Multi node
Run docker container on each machine:
```{code-block} bash
:substitutions:
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 quay.io/ascend/vllm-ascend:|vllm_ascend_version| bash
```
Choose one machine as head node, the other are worker nodes, then start ray on each machine:
:::{note}
Check out your `nic_name` by command `ip addr`.
:::
```shell
# Head node
export HCCL_IF_IP={local_ip}
export GLOO_SOCKET_IFNAME={nic_name}
export TP_SOCKET_IFNAME={nic_name}
export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES=1
ray start --head --num-gpus=8
# Worker node
export HCCL_IF_IP={local_ip}
export ASCEND_PROCESS_LOG_PATH={plog_save_path}
export GLOO_SOCKET_IFNAME={nic_name}
export TP_SOCKET_IFNAME={nic_name}
export RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES=1
export ASCEND_RT_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
ray start --address='{head_node_ip}:{port_num}' --num-gpus=8 --node-ip-address={local_ip}
```
Start the vLLM server on head node:
```shell
export VLLM_HOST_IP={head_node_ip}
export HCCL_CONNECT_TIMEOUT=120
export ASCEND_PROCESS_LOG_PATH={plog_save_path}
export HCCL_IF_IP={head_node_ip}
if [ -d "{plog_save_path}" ]; then
rm -rf {plog_save_path}
echo ">>> remove {plog_save_path}"
fi
LOG_FILE="multinode_$(date +%Y%m%d_%H%M).log"
VLLM_TORCH_PROFILER_DIR=./vllm_profile
python -m vllm.entrypoints.openai.api_server \
--model="Deepseek/DeepSeek-V2-Lite-Chat" \
--trust-remote-code \
--enforce-eager \
--max-model-len {max_model_len} \
--distributed_executor_backend "ray" \
--tensor-parallel-size 16 \
--disable-log-requests \
--disable-log-stats \
--disable-frontend-multiprocessing \
--port {port_num} \
```
Once your server is started, you can query the model with input prompts:
```shell
curl -X POST http://127.0.0.1:{prot_num}/v1/completions \
-H "Content-Type: application/json" \
-d '{
"model": "Deepseek/DeepSeek-V2-Lite-Chat",
"prompt": "The future of AI is",
"max_tokens": 24
}'
```
If you query the server successfully, you can see the info shown below (client):
```
{"id":"cmpl-6dfb5a8d8be54d748f0783285dd52303","object":"text_completion","created":1739957835,"model":"/home/data/DeepSeek-V2-Lite-Chat/","choices":[{"index":0,"text":" heavily influenced by neuroscience and cognitiveGuionistes. The goalochondria is to combine the efforts of researchers, technologists,","logprobs":null,"finish_reason":"length","stop_reason":null,"prompt_logprobs":null}],"usage":{"prompt_tokens":6,"total_tokens":30,"completion_tokens":24,"prompt_tokens_details":null}}
```
Logs of the vllm server:
```
INFO: 127.0.0.1:59384 - "POST /v1/completions HTTP/1.1" 200 OK
INFO 02-19 17:37:35 metrics.py:453] Avg prompt throughput: 0.0 tokens/s, Avg generation throughput: 1.9 tokens/s, Running: 0 reqs, Swapped: 0 reqs, Pending: 0 reqs, GPU KV cache usage: 0.0%, CPU KV cache usage: 0.0%.
```

107
docs/source/tutorials/multi_npu.md Normal file
View File

@@ -0,0 +1,107 @@
# Multi-NPU (QwQ 32B)
## Run vllm-ascend on Multi-NPU
Run docker container:
```{code-block} bash
:substitutions:
# Update the vllm-ascend image
export IMAGE=quay.io/ascend/vllm-ascend:|vllm_ascend_version|
docker run --rm \
--name vllm-ascend \
--device /dev/davinci0 \
--device /dev/davinci1 \
--device /dev/davinci2 \
--device /dev/davinci3 \
--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
```
Setup environment variables:
```bash
# Use Modelscope mirror to speed up model download
export VLLM_USE_MODELSCOPE=True
# To avoid NPU out of memory, set `max_split_size_mb` to any value lower than you need to allocate for Qwen2.5-7B-Instruct
export PYTORCH_NPU_ALLOC_CONF=max_split_size_mb:256
```
### Online Inference on Multi-NPU
Run the following script to start the vLLM server on Multi-NPU:
```bash
vllm serve Qwen/QwQ-32B --max-model-len 4096 --port 8000 -tp 4
```
Once your server is started, you can query the model with input prompts
```bash
curl http://localhost:8000/v1/completions \
-H "Content-Type: application/json" \
-d '{
"model": "Qwen/QwQ-32B",
"prompt": "QwQ-32B是什么？",
"max_tokens": "128",
"top_p": "0.95",
"top_k": "40",
"temperature": "0.6"
}'
```
### Offline Inference on Multi-NPU
Run the following script to execute offline inference on multi-NPU:
```python
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",
]
sampling_params = SamplingParams(temperature=0.6, top_p=0.95, top_k=40)
llm = LLM(model="Qwen/QwQ-32B",
tensor_parallel_size=4,
distributed_executor_backend="tp",
max_model_len=4096)
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()
```
If you run this script successfully, you can see the info shown below:
```bash
Prompt: 'Hello, my name is', Generated text: ' Daniel and I am an 8th grade student at York Middle School. I'
Prompt: 'The future of AI is', Generated text: ' following you. As the technology advances, a new report from the Institute for the'
```

133
docs/source/tutorials/single_npu.md Normal file
View File

@@ -0,0 +1,133 @@
# Single NPU (Qwen 7B)
## Run vllm-ascend on Single NPU
### Offline Inference on Single NPU
Run docker container:
```{code-block} bash
:substitutions:
# Update the vllm-ascend image
export IMAGE=quay.io/ascend/vllm-ascend:|vllm_ascend_version|
docker run --rm \
--name vllm-ascend \
--device /dev/davinci0 \
--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
```
Setup environment variables:
```bash
# Use Modelscope mirror to speed up model download
export VLLM_USE_MODELSCOPE=True
# To avoid NPU out of memory, set `max_split_size_mb` to any value lower than you need to allocate for Qwen2.5-7B-Instruct
export PYTORCH_NPU_ALLOC_CONF=max_split_size_mb:256
```
:::{note}
`max_split_size_mb` prevents the native allocator from splitting blocks larger than this size (in MB). This can reduce fragmentation and may allow some borderline workloads to complete without running out of memory. You can find more details [<u>here</u>](https://www.hiascend.com/document/detail/zh/CANNCommunityEdition/800alpha003/apiref/envref/envref_07_0061.html).
:::
Run the following script to execute offline inference on a single NPU:
```python
from vllm import LLM, SamplingParams
prompts = [
"Hello, my name is",
"The future of AI is",
]
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
llm = LLM(model="Qwen/Qwen2.5-7B-Instruct", max_model_len=26240)
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}")
```
If you run this script successfully, you can see the info shown below:
```bash
Prompt: 'Hello, my name is', Generated text: ' Daniel and I am an 8th grade student at York Middle School. I'
Prompt: 'The future of AI is', Generated text: ' following you. As the technology advances, a new report from the Institute for the'
```
### Online Serving on Single NPU
Run docker container to start the vLLM server on a single NPU:
```{code-block} bash
:substitutions:
# Update the vllm-ascend image
export IMAGE=quay.io/ascend/vllm-ascend:|vllm_ascend_version|
docker run --rm \
--name vllm-ascend \
--device /dev/davinci0 \
--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 \
-e VLLM_USE_MODELSCOPE=True \
-e PYTORCH_NPU_ALLOC_CONF=max_split_size_mb:256 \
-it $IMAGE \
vllm serve Qwen/Qwen2.5-7B-Instruct --max_model_len 26240
```
:::{note}
Add `--max_model_len` option to avoid ValueError that the Qwen2.5-7B model's max seq len (32768) is larger than the maximum number of tokens that can be stored in KV cache (26240). This will differ with different NPU series base on the HBM size. Please modify the value according to a suitable value for your NPU series.
:::
If your service start successfully, you can see the info shown below:
```bash
INFO: Started server process [6873]
INFO: Waiting for application startup.
INFO: Application startup complete.
```
Once your server is started, you can query the model with input prompts:
```bash
curl http://localhost:8000/v1/completions \
-H "Content-Type: application/json" \
-d '{
"model": "Qwen/Qwen2.5-7B-Instruct",
"prompt": "The future of AI is",
"max_tokens": 7,
"temperature": 0
}'
```
If you query the server successfully, you can see the info shown below (client):
```bash
{"id":"cmpl-b25a59a2f985459781ce7098aeddfda7","object":"text_completion","created":1739523925,"model":"Qwen/Qwen2.5-7B-Instruct","choices":[{"index":0,"text":" here. It’s not just a","logprobs":null,"finish_reason":"length","stop_reason":null,"prompt_logprobs":null}],"usage":{"prompt_tokens":5,"total_tokens":12,"completion_tokens":7,"prompt_tokens_details":null}}
```
Logs of the vllm server:
```bash
INFO: 172.17.0.1:49518 - "POST /v1/completions HTTP/1.1" 200 OK
INFO 02-13 08:34:35 logger.py:39] Received request cmpl-574f00e342904692a73fb6c1c986c521-0: prompt: 'San Francisco is a', params: SamplingParams(n=1, presence_penalty=0.0, frequency_penalty=0.0, repetition_penalty=1.0, temperature=0.0, top_p=1.0, top_k=-1, min_p=0.0, seed=None, stop=[], stop_token_ids=[], bad_words=[], include_stop_str_in_output=False, ignore_eos=False, max_tokens=7, min_tokens=0, logprobs=None, prompt_logprobs=None, skip_special_tokens=True, spaces_between_special_tokens=True, truncate_prompt_tokens=None, guided_decoding=None), prompt_token_ids: [23729, 12879, 374, 264], lora_request: None, prompt_adapter_request: None.
```