add inductor xpu #16
add inductor xpu #16
Conversation
| else | ||
| wget -q -e "https_proxy=http://proxy-us.intel.com:912" https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh | ||
| bash Miniconda3-latest-Linux-x86_64.sh -b | ||
| source ${HOME}/miniconda3/etc/profile.d/conda.sh 2>&1 >> /dev/null |
There was a problem hiding this comment.
this line can be replaced by source ${HOME}/miniconda3/bin/activate
| fi | ||
| else | ||
| wget -q -e "https_proxy=http://proxy-us.intel.com:912" https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh | ||
| bash Miniconda3-latest-Linux-x86_64.sh -b |
There was a problem hiding this comment.
Do you have met the miniconda package issue like conda/conda#13225 recently? Seems the package https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh (same as Miniconda3-py311_23.9.0-0-Linux-x86_64.sh) has issue now. suggest to use other python versions, eg. Miniconda3-py39_23.9.0-0-Linux-x86_64.sh
There was a problem hiding this comment.
Never met this issue. We usually install the Miniconda3-latest-Linux-x86_64.sh on our machine firstly
There was a problem hiding this comment.
Ok, please aware it, when we meet such issue can use similar workaround
| fi | ||
|
|
||
| # set gpu governor | ||
| if [[ -z "${USER_PASS}" ]];then USER_PASS="gta";fi |
There was a problem hiding this comment.
Don't hardcode passwd in this public repo, if we really need such info, let's pass them by Jenkins credential
There was a problem hiding this comment.
set it as jenkins parameter
There was a problem hiding this comment.
if the cpu mode setting moved, do we still need this one?
| TRITON_CODEGEN_INTEL_XPU_BACKEND=1 python setup.py bdist_wheel | ||
| pip install dist/*.whl | ||
| source ${HOME}/env.sh | ||
| python -c "import triton" |
There was a problem hiding this comment.
maybe need move to parent dir to do the import check to avoid some path issue, because under there is a triton folder python
| wget -q -e use_proxy=no ${ipex_whl} | ||
| python -m pip install --force-reinstall $(basename ${ipex_whl}) | ||
| else | ||
| bash ${WORKSPACE}/inductor-tools/scripts/env_prepare.sh |
There was a problem hiding this comment.
I don't see any parameter in the beginning of this env_prepare.sh script, please add them. By the way, please also add oneapi version as a parameter for this script and env.sh to add the basekit version control. You can refer latest ci/nightly workflows in xpu backend repo
| @@ -0,0 +1,57 @@ | |||
| installed_torch_git_version=$(python -c "import torch;print(torch.version.git_version)"|| true) | |||
| ''' | ||
| }//retry | ||
| }//stage | ||
| stage('Accuracy-Test') { |
There was a problem hiding this comment.
Add a multi-choice parameter SCENARIO in this groovy and jenkins job, the value can be accuracy, performance, if the accuracy in the SCENARIO do the accuracy test, else bypass this stage. The performance stage similar with accuracy part. Default value can be choose accuracy & performance
| pip install styleFrame scipy pandas | ||
| pushd ${WORKSPACE}/pytorch | ||
| rm -rf inductor_log | ||
| bash inductor_xpu_test.sh huggingface amp_bf16 inference accuracy xpu 0 & \ |
There was a problem hiding this comment.
Also keep SUITE, DT, MODE as this groovy and jenkins job's parameters, all those parameters use multi-choice parameter, and combine those parameter values, do test one-by-one. Default value can be set as current hardcode cmd. For example, if the SUITE={huggingface,timm_models}, DT={float32,amp_bf16}, MODE={inference,training}, then it will be have 8 combinations need to test.
| conda activate ${conda_env} | ||
| source ${HOME}/env.sh ${oneapi_ver} | ||
|
|
||
| cd ${WORKSPACE}/pytorch/inductor_log/huggingface |
There was a problem hiding this comment.
Extract this log summary part as a standalone scripts, and keep the flexibility for different suite/dt/mode/scenario(low priority). A simple way is that log summary scripts receive those parameters and maintain the criteria for different tests (by using a configure file or dict like structure). E.g. {"huggingface-amp_bf16-inference-accuracy":44, "huggingface-amp_bf16-training-accuracy":42}
| }//retry | ||
| }//stage | ||
| stage('Performance-Test') { | ||
| println('================================================================') |
There was a problem hiding this comment.
Similar request with accuracy test stage
| bash inductor_xpu_test.sh huggingface amp_bf16 inference accuracy xpu 0 & \ | ||
| bash inductor_xpu_test.sh huggingface amp_bf16 training accuracy xpu 1 & \ | ||
| bash inductor_xpu_test.sh huggingface amp_fp16 inference accuracy xpu 2 & \ | ||
| bash inductor_xpu_test.sh huggingface amp_fp16 training accuracy xpu 3 & wait |
There was a problem hiding this comment.
The right cmd is:
bash inductor_xpu_test.sh ${SUITE} ${DT} ${MODE} accuracy xpu 0 static 4 0 & \
bash inductor_xpu_test.sh ${SUITE} ${DT} ${MODE} accuracy xpu 1 static 4 1 & \
bash inductor_xpu_test.sh ${SUITE} ${DT} ${MODE} accuracy xpu 2 static 4 2 & \
bash inductor_xpu_test.sh ${SUITE} ${DT} ${MODE} accuracy xpu 3 static 4 3 & waitIt means the one combination will be split as 4 sub-test run on 4 card at the same time. Please modify the cmd in performance part also. Your previously cmd run 2 combination at same time and each combination run 2 times on 2 card, not split them as sub-test.
|
Hi @RUIJIEZHONG66166 any update for those comments? |
No description provided.