@@ -28,12 +28,52 @@ jobs:
2828 steps :
2929 - name : Checkout XLA
3030 uses : actions/checkout@b4ffde65f46336ab88eb53be808477a3936bae11 # v4.1.1
31+ with :
32+ path : " openxla/xla"
33+
34+ - name : Print machine specs
35+ run : |
36+ nvidia-smi
37+ free -h # Memory information
38+ df -h # Disk space information
39+ uname -a # Kernel information
3140
3241 - name : Create results directory
42+ run : mkdir results
43+
44+ - name : Configure XLA for GPU backend
45+ run : |
46+ ./configure.py --backend CUDA --nccl
47+
48+ - name : Set TF_CPP_MAX_VLOG_LEVEL
49+ env :
50+ TF_CPP_MAX_VLOG_LEVEL : 1
51+ run : |
52+ echo "TF_CPP_MAX_VLOG_LEVEL is: $TF_CPP_MAX_VLOG_LEVEL"
53+
54+ - name : Build hlo_runner_main
55+ run : bazel build -c opt --config=cuda --dynamic_mode=off //xla/tools/multihost_hlo_runner:hlo_runner_main
56+ # TODO(juliagmt): Add more performance-criticalHLOs to benchmark.
57+ - name : Run xla/tools/hlo_opt/tests/gpu_hlo_pass.hlo
3358 run : |
34- mkdir results
35- ls
36- cd results
59+ ./bazel-bin/xla/tools/multihost_hlo_runner/hlo_runner_main --device_type=gpu --log_output=True --use_spmd_partitioning --xla_gpu_dump_xspace_to=xla/results/xspace.pbtxt xla/tools/hlo_opt/tests/gpu_hlo_pass.hlo
60+
61+ - name : Compute the cost of gpu_hlo_pass.hlo
62+ run : |
63+ PWD=$(pwd)
64+ bazel run //xla/tools:compute_cost -- --input=$PWD/xla/tools/hlo_opt/tests/gpu_hlo_pass.hlo --format=hlo --gpu
65+
66+ - name : Checkout juliagmt-google/xla
67+ uses : actions/checkout@b4ffde65f46336ab88eb53be808477a3936bae11 # Replace with the current SHA of v4.1.1 or later
68+ with :
69+ repository : juliagmt-google/xla
70+ path : juliagmt-google-xla
71+
72+ - name : Run get_device_stats_main
73+ run : |
74+ bazel run //xla/tools:get_device_stats_main -- --input=$PWD/xla/results/xspace.pb
75+ working-directory : juliagmt-google-xla
76+
3777 - name : Upload Artifacts
3878 uses : actions/upload-artifact@65c4c4a1ddee5b72f698fdd19549f0f0fb45cf08
3979 with :
0 commit comments