Migrate CUDA CI workflows (#156)

huggingface · Mar 13, 2024 · d719e6e · d719e6e
1 parent fd32ad5
commit d719e6e
Show file tree

Hide file tree

Showing 10 changed files with 127 additions and 118 deletions.
diff --git a/.github/workflows/test_api_cuda.yaml b/.github/workflows/test_api_cuda.yaml
@@ -22,7 +22,8 @@ jobs:
             { torch_cuda: cu121, torch_pre_release: 1, cuda_version: 12.1.1 },
           ]
 
-    runs-on: nvidia-gpu
+    runs-on: [multi-gpu, nvidia-gpu, 4-a10, ci]
+
     steps:
       - name: Checkout
         uses: actions/checkout@v3
@@ -37,18 +38,13 @@ jobs:
           --tag opt-bench-cuda:${{ matrix.image.cuda_version }}
           .
 
-      - name: Get GPUs with most free memory
-        id: get_devices
-        run: |
-          echo "::set-output name=devices::$(nvidia-smi --query-gpu=memory.free,index --format=csv,noheader,nounits | sort -n -k1 | tail -n 2 | awk -F', ' '{print $2}' | xargs echo -n | sed 's/ /,/g' | awk '{print $0}')"
-
       - name: Run tests
         run: docker run
           --rm
-          --pid host
+          --gpus all
           --shm-size 64G
           --env USE_CUDA="1"
-          --gpus '"device=${{ steps.get_devices.outputs.devices }}"'
+          --env PROCESS_SPECIFIC_VRAM="0"
           --volume $(pwd):/workspace/optimum-benchmark
           --workdir /workspace/optimum-benchmark
           --entrypoint /bin/bash

diff --git a/.github/workflows/test_cli_cuda_onnxruntime.yaml b/.github/workflows/test_cli_cuda_onnxruntime.yaml
@@ -13,7 +13,8 @@ concurrency:
 
 jobs:
   build_image_and_run_cli_cuda_onnxruntime_tests:
-    runs-on: nvidia-gpu
+    runs-on: [single-gpu, nvidia-gpu, a10, ci]
+
     steps:
       - name: Checkout
         uses: actions/checkout@v3
@@ -28,20 +29,15 @@ jobs:
           --tag opt-bench-cuda:11.8.0
           .
 
-      - name: Get GPUs with most free memory
-        id: get_devices
-        run: |
-          echo "::set-output name=devices::$(nvidia-smi --query-gpu=memory.free,index --format=csv,noheader,nounits | sort -n -k1 | tail -n 2 | awk -F', ' '{print $2}' | xargs echo -n | sed 's/ /,/g' | awk '{print $0}')"
-
       - name: Run tests
         run: docker run
           --rm
-          --pid host
+          --gpus all
           --shm-size 64G
           --env USE_CUDA="1"
-          --entrypoint /bin/bash
-          --gpus '"device=${{ steps.get_devices.outputs.devices }}"'
+          --env PROCESS_SPECIFIC_VRAM="0"
           --volume $(pwd):/workspace/optimum-benchmark
           --workdir /workspace/optimum-benchmark
+          --entrypoint /bin/bash
           opt-bench-cuda:11.8.0
           -c "pip install -e .[testing,onnxruntime-gpu,diffusers,timm] && pytest -k 'cli and cuda and onnxruntime' -x"
diff --git a/.github/workflows/test_cli_cuda_pytorch.yaml b/.github/workflows/test_cli_cuda_pytorch.yaml
@@ -22,7 +22,8 @@ jobs:
             { torch_cuda: cu121, torch_pre_release: 1, cuda_version: 12.1.1 },
           ]
 
-    runs-on: nvidia-gpu
+    runs-on: [multi-gpu, nvidia-gpu, 4-a10, ci]
+
     steps:
       - name: Checkout
         uses: actions/checkout@v3
@@ -37,18 +38,13 @@ jobs:
           --tag opt-bench-cuda:${{ matrix.image.cuda_version }}
           .
 
-      - name: Get GPUs with most free memory
-        id: get_devices
-        run: |
-          echo "::set-output name=devices::$(nvidia-smi --query-gpu=memory.free,index --format=csv,noheader,nounits | sort -n -k1 | tail -n 2 | awk -F', ' '{print $2}' | xargs echo -n | sed 's/ /,/g' | awk '{print $0}')"
-
       - name: Run tests
         run: docker run
           --rm
-          --pid host
+          --gpus all
           --shm-size 64G
           --env USE_CUDA="1"
-          --gpus '"device=${{ steps.get_devices.outputs.devices }}"'
+          --env PROCESS_SPECIFIC_VRAM="0"
           --volume $(pwd):/workspace/optimum-benchmark
           --workdir /workspace/optimum-benchmark
           --entrypoint /bin/bash

diff --git a/.github/workflows/test_cli_cuda_torch_ort.yaml b/.github/workflows/test_cli_cuda_torch_ort.yaml
@@ -13,7 +13,8 @@ concurrency:
 
 jobs:
   build_image_and_run_cli_cuda_torch_ort_tests:
-    runs-on: nvidia-gpu
+    runs-on: [multi-gpu, nvidia-gpu, 4-a10, ci]
+
     steps:
       - name: Checkout
         uses: actions/checkout@v3
@@ -28,20 +29,15 @@ jobs:
           --tag opt-bench-cuda:11.8.0
           .
 
-      - name: Get GPUs with most free memory
-        id: get_devices
-        run: |
-          echo "::set-output name=devices::$(nvidia-smi --query-gpu=memory.free,index --format=csv,noheader,nounits | sort -n -k1 | tail -n 2 | awk -F', ' '{print $2}' | xargs echo -n | sed 's/ /,/g' | awk '{print $0}')"
-
       - name: Run tests
         run: docker run
           --rm
-          --pid host
+          --gpus all
           --shm-size 64G
           --env USE_CUDA="1"
-          --entrypoint /bin/bash
-          --gpus '"device=${{ steps.get_devices.outputs.devices }}"'
+          --env PROCESS_SPECIFIC_VRAM="0"
           --volume $(pwd):/workspace/optimum-benchmark
           --workdir /workspace/optimum-benchmark
+          --entrypoint /bin/bash
           opt-bench-cuda:11.8.0
           -c "pip install -e .[testing,torch-ort,peft] && python -m torch_ort.configure && pytest -k 'cli and cuda and torch_ort' -x"
diff --git a/.github/workflows/test_cli_tensorrt_llm.yaml b/.github/workflows/test_cli_tensorrt_llm.yaml
@@ -13,7 +13,8 @@ concurrency:
 
 jobs:
   pull_image_and_run_cli_tensorrt_llm_tests:
-    runs-on: nvidia-gpu
+    runs-on: [single-gpu, nvidia-gpu, a10, ci]
+
     steps:
       - name: Checkout
         uses: actions/checkout@v3
@@ -26,18 +27,14 @@ jobs:
           --tag opt-bench-tensorrt-llm:latest
           .
 
-      - name: Get GPUs with most free memory
-        id: get_devices
-        run: |
-          echo "::set-output name=devices::$(nvidia-smi --query-gpu=memory.free,index --format=csv,noheader,nounits | sort -n -k1 | tail -n 2 | awk -F', ' '{print $2}' | xargs echo -n | sed 's/ /,/g' | awk '{print $0}')"
-
       - name: Run tests
         run: docker run
           --rm
+          --gpus all
           --pid host
           --shm-size 64G
           --env USE_CUDA="1"
-          --gpus '"device=${{ steps.get_devices.outputs.devices }}"'
+          --env PROCESS_SPECIFIC_VRAM="0"
           --volume $(pwd):/workspace/optimum-benchmark
           --workdir /workspace/optimum-benchmark
           --entrypoint /bin/bash

diff --git a/.github/workflows/test_cli_tensorrt_onnxruntime.yaml b/.github/workflows/test_cli_tensorrt_onnxruntime.yaml
@@ -13,7 +13,8 @@ concurrency:
 
 jobs:
   build_image_and_run_cli_tensorrt_onnxruntime_tests:
-    runs-on: nvidia-gpu
+    runs-on: [single-gpu, nvidia-gpu, a10, ci]
+
     steps:
       - name: Checkout
         uses: actions/checkout@v3
@@ -26,18 +27,14 @@ jobs:
           --tag opt-bench-tensorrt:latest
           .
 
-      - name: Get GPUs with most free memory
-        id: get_devices
-        run: |
-          echo "::set-output name=devices::$(nvidia-smi --query-gpu=memory.free,index --format=csv,noheader,nounits | sort -n -k1 | tail -n 2 | awk -F', ' '{print $2}' | xargs echo -n | sed 's/ /,/g' | awk '{print $0}')"
-
       - name: Run tests
         run: docker run
           --rm
+          --gpus all
           --pid host
           --shm-size 64G
           --env USE_CUDA="1"
-          --gpus '"device=${{ steps.get_devices.outputs.devices }}"'
+          --env PROCESS_SPECIFIC_VRAM="0"
           --volume $(pwd):/workspace/optimum-benchmark
           --workdir /workspace/optimum-benchmark
           --entrypoint /bin/bash

diff --git a/docker/cpu.dockerfile b/docker/cpu.dockerfile
@@ -1,4 +1,6 @@
-FROM ubuntu:latest
+ARG UBUNTU_VERSION=20.04
+
+FROM ubuntu:${UBUNTU_VERSION}
 
 # Ignore interactive questions during `docker build`
 ENV DEBIAN_FRONTEND noninteractive

diff --git a/examples/pytorch_timm.yaml b/examples/pytorch_timm.yaml
@@ -18,6 +18,7 @@ launcher:
   device_isolation: true
 
 benchmark:
+  memory: true
   input_shapes:
     batch_size: 1