Skip to content

Latest commit

 

History

History
113 lines (86 loc) · 4.31 KB

README.md

File metadata and controls

113 lines (86 loc) · 4.31 KB
Raw

PaddlePaddle Custom Device Implementation for Tecorigin SDAA

English | 简体中文

Please refer to the following steps to compile, install and verify the custom device implementation for Tecorigin SDAA.

Tecorigin SDAA System Requirements

Module Version
TecoDriver 1.1.0
TecoToolkit 1.1.0

Prepare environment and source code

# 1. pull PaddlePaddle Tecorigin SDAA development docker image
wget http://mirrors.tecorigin.com/repository/teco-docker-tar-repo/release/ubuntu22.04/x86_64/1.1.0/paddle-1.1.0-paddle_sdaa1.1.0.tar
docker load < paddle-1.1.0-paddle_sdaa1.1.0.tar

# 2. refer to the following commands to start docker container and activate conda environment (PaddlePaddle framwork has been installed in the conda environment)
docker run -it --name="paddle_sdaa_dev" --net=host -v $(pwd):/work \
--device=/dev/tcaicard0 --device=/dev/tcaicard1 \
--device=/dev/tcaicard2 --device=/dev/tcaicard3 \
--cap-add SYS_PTRACE --cap-add SYS_ADMIN --shm-size 64g \
jfrog.tecorigin.net/tecotp-docker/release/ubuntu22.04/x86_64/paddle:1.1.0-paddle_sdaa1.1.0 /bin/bash

conda activate paddle_env

# 3. clone the source code
git clone https://github.com/PaddlePaddle/PaddleCustomDevice
cd PaddleCustomDevice

Installation and Verification

Source Code Compile

# 1. update cmake
pip install -U cmake

# 2. checkout branch to `develop`
git checkout develop

# 3. execute the following commands to update submodule
git submodule sync
git submodule update --init --recursive

# 4. go to Tecorigin sdaa directory
cd backends/sdaa

# 5. execute compile script
bash compile.sh

# 6. install the generated whl package, which is under build/dist directory
pip install build/dist/*.whl --force-reinstall

Verification

# 1. using paddle_sdaa utils's `run_check` to check whether paddle-sdaa plugin and PaddlePaddle framework are installed.
python3 -c "import paddle_sdaa; paddle_sdaa.utils.run_check()"
# expected output:
paddle-sdaa and paddlepaddle are installed successfully!


# 2. list available hardware backends
python3 -c "import paddle; print(paddle.device.get_all_custom_device_type())"
# expected ouput:
['sdaa']

# 3. run relu forward
python3 -c "import paddle;paddle.set_device('sdaa');print(paddle.nn.functional.relu(paddle.to_tensor([-2., 1.])))"
# expected output:
Tensor(shape=[2], dtype=float32, place=Place(sdaa:0), stop_gradient=True,
       [0., 1.])

Train and Inference Demo

# demo for training, evaluation and inference
python tests/test_MNIST_model.py

# training output as following
Epoch [1/2], Iter [01/14], reader_cost: 1.41201 s, batch_cost: 1.56096 s, ips: 2624.03256 samples/s, eta: 0:00:43
Epoch [1/2], Iter [02/14], reader_cost: 0.70611 s, batch_cost: 0.84809 s, ips: 4829.67512 samples/s, eta: 0:00:22
... ...
Epoch [2/2], Iter [14/14], reader_cost: 0.11122 s, batch_cost: 0.24438 s, ips: 16760.81762 samples/s, eta: 0:00:00
Epoch ID: 2, Epoch time: 3.50429 s, reader_cost: 1.55708 s, batch_cost: 3.42131 s, avg ips: 16363.92196 samples/s
Eval - Epoch ID: 2, Top1 accurary:: 0.84607, Top5 accurary:: 0.98462

# inference output as following
I0307 05:21:33.673595  6583 interpretercore.cc:237] New Executor is Running.
I0307 05:21:33.703184  6583 analysis_predictor.cc:1503] CustomDevice is enabled
... ...
I0307 05:21:33.707281  6583 analysis_predictor.cc:1660] ======= optimize end =======
I0307 05:21:33.707347  6583 naive_executor.cc:164] ---  skip [feed], feed -> inputs
I0307 05:21:33.707659  6583 naive_executor.cc:164] ---  skip [linear_5.tmp_1], fetch -> fetch
Output data size is 10
Output data shape is (1, 10)

Environment Variables

Subject Variable Name Type Description Default Value
Debug CUSTOM_DEVICE_BLACK_LIST String Ops in back list will fallbacks to CPU ""
Profiling ENABLE_SDPTI String enable sdpti 1
Debug HIGH_PERFORMANCE_CONV String set HIGH_PERFORMANCE_CONV to "1" can enable high performance conv API 0
Debug FLAGS_sdaa_runtime_debug bool print runtime information false
Feature FLAGS_sdaa_reuse_event bool enable event pool true