feat: update ray resources to match the newer/cleaner torchx resources form

starpit · starpit · commit 35260c23c269 · 2023-03-18T20:00:34.000-04:00
this allows fixing: codeflare logs when late-attaching may not stream out gpu utilization

BREAKING CHANGE: this changes the structure of the ray form; tests may need updates. Also, any automated -y runs will require an update.
diff --git a/package-lock.json b/package-lock.json
diff --git a/plugins/plugin-codeflare/package.json b/plugins/plugin-codeflare/package.json
@@ -30,13 +30,13 @@
     "@types/split2": "^3.2.1"
   },
   "dependencies": {
-    "@guidebooks/store": "^6.1.9",
+    "@guidebooks/store": "^6.2.1",
     "@logdna/tail-file": "^3.0.1",
     "@patternfly/react-charts": "^6.94.18",
     "@patternfly/react-core": "^4.276.6",
     "asciinema-player": "^3.1.0",
     "chokidar": "^3.5.3",
-    "madwizard": "^6.4.1",
+    "madwizard": "^6.5.3",
     "needle": "^3.2.0",
     "open": "^8.4.2",
     "pretty-bytes": "^6.1.0",
diff --git a/tests/kind/profiles/non-gpu1/keep-it-simple b/tests/kind/profiles/non-gpu1/keep-it-simple
@@ -14,7 +14,7 @@
     "ml/codeflare/training/byoc/form": "{\"Path to source\":\"$PWD/tests/kind/inputs/qiskit\",\"Base image\":\"rayproject/ray:2.1.0\",\"Command line prefix\":\"python3 main.py\"}",
     "kubernetes/choose/secret/image-pull": "No secret needed, since my image is public",
     "s3/choose/bucket/maybe": "My data is not stored in S3",
-    "ml/ray/start/resources": "{\"Number of CPUs\":\"500m\",\"Number of GPUs\":\"0\",\"Minimum Workers\":\"1\",\"Maximum Workers\":\"1\",\"Worker Memory\":\"500Mi\",\"Head Memory\":\"3Gi\",\"Ephemeral Storage\":\"5Gi\"}",
+    "ml/ray/start/resources": "{\"Number of Workers\":1,\"CPUs per worker\":\"500m\",\"GPUs per worker\":0,\"Memory per worker\":\"1.5Gi\",\"Ephemeral Storage per worker\":\"5Gi\"}",
     "kubernetes/context": "kind-codeflare-test",
     "kubernetes/choose/ns": "default",
     "ml/ray/storage/s3/maybe": "My code does not use Ray Workflows",
diff --git a/tests/kind/profiles/non-gpu1/mcad-coscheduler b/tests/kind/profiles/non-gpu1/mcad-coscheduler
@@ -14,7 +14,13 @@
     "ml/codeflare/training/byoc/form": "{\"Path to source\":\"$PWD/tests/kind/inputs/qiskit\",\"Base image\":\"rayproject/ray:2.1.0\",\"Command line prefix\":\"python3 main.py\"}",
     "kubernetes/choose/secret/image-pull": "No secret needed, since my image is public",
     "s3/choose/bucket/maybe": "My data is not stored in S3",
-    "ml/ray/start/resources": "{\"Number of CPUs\":\"200m\",\"Number of GPUs\":\"0\",\"Minimum Workers\":\"1\",\"Maximum Workers\":\"1\",\"Worker Memory\":\"500Mi\",\"Head Memory\":\"3Gi\",\"Ephemeral Storage\":\"5Gi\"}",
+    "ml/ray/start/resources": {
+      "Number of Workers": 1,
+      "CPUs per worker": "200m",
+      "GPUs per worker": 0,
+      "Memory per worker": "1.25Gi",
+      "Ephemeral Storage per worker": "5Gi"
+    },
     "kubernetes/context": "kind-codeflare-test",
     "kubernetes/choose/ns": "default",
     "ml/ray/storage/s3/maybe": "My code does not use Ray Workflows",
diff --git a/tests/kind/profiles/non-gpu1/mcad-default b/tests/kind/profiles/non-gpu1/mcad-default
@@ -14,7 +14,7 @@
     "ml/codeflare/training/byoc/form": "{\"Path to source\":\"$PWD/tests/kind/inputs/qiskit\",\"Base image\":\"rayproject/ray:2.1.0\",\"Command line prefix\":\"python3 main.py\"}",
     "kubernetes/choose/secret/image-pull": "No secret needed, since my image is public",
     "s3/choose/bucket/maybe": "My data is not stored in S3",
-    "ml/ray/start/resources": "{\"Number of CPUs\":\"200m\",\"Number of GPUs\":\"0\",\"Minimum Workers\":\"1\",\"Maximum Workers\":\"1\",\"Worker Memory\":\"500Mi\",\"Head Memory\":\"3Gi\",\"Ephemeral Storage\":\"5Gi\"}",
+    "ml/ray/start/resources": "{\"Number of Workers\":1,\"CPUs per worker\":\"200m\",\"GPUs per worker\":0,\"Memory per worker\":\"1.25Gi\",\"Ephemeral Storage per worker\":\"5Gi\"}",
     "kubernetes/context": "kind-codeflare-test",
     "kubernetes/choose/ns": "default",
     "ml/ray/storage/s3/maybe": "My code does not use Ray Workflows",
diff --git a/tests/kind/profiles/non-gpu1/mcad-preinstalled b/tests/kind/profiles/non-gpu1/mcad-preinstalled
@@ -14,7 +14,13 @@
     "ml/codeflare/training/byoc/form": "{\"Path to source\":\"$PWD/tests/kind/inputs/qiskit\",\"Base image\":\"rayproject/ray:2.1.0\",\"Command line prefix\":\"python3 main.py\"}",
     "kubernetes/choose/secret/image-pull": "No secret needed, since my image is public",
     "s3/choose/bucket/maybe": "My data is not stored in S3",
-    "ml/ray/start/resources": "{\"Number of CPUs\":\"200m\",\"Number of GPUs\":\"0\",\"Minimum Workers\":\"1\",\"Maximum Workers\":\"1\",\"Worker Memory\":\"500Mi\",\"Head Memory\":\"3Gi\",\"Ephemeral Storage\":\"5Gi\"}",
+    "ml/ray/start/resources": {
+      "Number of Workers": 1,
+      "CPUs per worker": "200m",
+      "GPUs per worker": 0,
+      "Memory per worker": "1.25Gi",
+      "Ephemeral Storage per worker": "5Gi"
+    },
     "kubernetes/context": "kind-codeflare-test",
     "kubernetes/choose/ns": "default",
     "ml/ray/storage/s3/maybe": "My code does not use Ray Workflows",
diff --git a/tests/kind/profiles/non-gpu1/ray-autoscaler b/tests/kind/profiles/non-gpu1/ray-autoscaler
@@ -14,7 +14,13 @@
     "ml/codeflare/training/byoc/form": "{\"Path to source\":\"$PWD/tests/kind/inputs/qiskit\",\"Base image\":\"rayproject/ray:2.1.0\",\"Command line prefix\":\"python3 main.py\"}",
     "kubernetes/choose/secret/image-pull": "No secret needed, since my image is public",
     "s3/choose/bucket/maybe": "My data is not stored in S3",
-    "ml/ray/start/resources": "{\"Number of CPUs\":\"200m\",\"Number of GPUs\":\"0\",\"Minimum Workers\":\"0\",\"Maximum Workers\":\"0\",\"Worker Memory\":\"500Mi\",\"Head Memory\":\"2.5Gi\",\"Ephemeral Storage\":\"5Gi\"}",
+    "ml/ray/start/resources": {
+      "Number of Workers": 1,
+      "CPUs per worker": "200m",
+      "GPUs per worker": 0,
+      "Memory per worker": "1.25Gi",
+      "Ephemeral Storage per worker": "5Gi"
+    },
     "kubernetes/context": "kind-codeflare-test",
     "kubernetes/choose/ns": "default",
     "ml/ray/storage/s3/maybe": "My code does not use Ray Workflows",
diff --git a/tests/kind/profiles/non-gpu2/keep-it-simple b/tests/kind/profiles/non-gpu2/keep-it-simple
@@ -14,12 +14,18 @@
     "ml/codeflare/training/byoc/form": "{\"Path to source\":\"$PWD/tests/kind/inputs/ray-basic\",\"Base image\":\"rayproject/ray:2.1.0\",\"Command line prefix\":\"python3 main.py\"}",
     "kubernetes/choose/secret/image-pull": "No secret needed, since my image is public",
     "s3/choose/bucket/maybe": "My data is not stored in S3",
-    "ml/ray/start/resources": "{\"Number of CPUs\":\"500m\",\"Number of GPUs\":\"0\",\"Minimum Workers\":\"1\",\"Maximum Workers\":\"1\",\"Worker Memory\":\"500Mi\",\"Head Memory\":\"3Gi\",\"Ephemeral Storage\":\"5Gi\"}",
+    "ml/ray/start/resources": {
+      "Number of Workers": 1,
+      "CPUs per worker": "500m",
+      "GPUs per worker": 0,
+      "Memory per worker": "1.5Gi",
+      "Ephemeral Storage per worker": "5Gi"
+    },
     "kubernetes/context": "kind-codeflare-test",
     "kubernetes/choose/ns": "default",
     "ml/ray/storage/s3/maybe": "My code does not use Ray Workflows",
     "ml/ray/cluster/choose": "codeflare-test-ray-cluster",
     "ml/ray/cluster/choose/kubernetes": "codeflare-test-ray-cluster",
     "ml/ray/cluster/kubernetes/choose-pod-scheduler": "Keep It Simple"
   }
-}
+}
diff --git a/tests/kind/profiles/non-gpu3/keep-it-simple b/tests/kind/profiles/non-gpu3/keep-it-simple
@@ -14,7 +14,13 @@
     "ml/codeflare/training/byoc/form": "{\"Path to source\":\"$PWD/tests/kind/inputs/qiskit\",\"Base image\":\"rayproject/ray:2.1.0\",\"Command line prefix\":\"python3 main.py\"}",
     "kubernetes/choose/secret/image-pull": "No secret needed, since my image is public",
     "s3/choose/bucket/maybe": "My data is not stored in S3",
-    "ml/ray/start/resources": "{\"Number of CPUs\":\"500m\",\"Number of GPUs\":\"0\",\"Minimum Workers\":\"1\",\"Maximum Workers\":\"1\",\"Worker Memory\":\"500Mi\",\"Head Memory\":\"3Gi\",\"Ephemeral Storage\":\"5Gi\"}",
+    "ml/ray/start/resources": {
+      "Number of Workers": 1,
+      "CPUs per worker": "500m",
+      "GPUs per worker": 0,
+      "Memory per worker": "1.5Gi",
+      "Ephemeral Storage per worker": "5Gi"
+    },
     "kubernetes/context": "kind-codeflare-test",
     "kubernetes/choose/ns": "default",
     "ml/ray/storage/s3/maybe": "My code does not use Ray Workflows",
diff --git a/tests/kind/profiles/non-gpu4/keep-it-simple b/tests/kind/profiles/non-gpu4/keep-it-simple
@@ -14,7 +14,13 @@
     "ml/codeflare/training/byoc/form": "{\"Path to source\":\"$PWD/tests/kind/inputs/qiskit\",\"Base image\":\"rayproject/ray:2.1.0\",\"Command line prefix\":\"python3 main.py\"}",
     "kubernetes/choose/secret/image-pull": "No secret needed, since my image is public",
     "s3/choose/bucket/maybe": "My data is not stored in S3",
-    "ml/ray/start/resources": "{\"Number of CPUs\":\"500m\",\"Number of GPUs\":\"0\",\"Minimum Workers\":\"1\",\"Maximum Workers\":\"1\",\"Worker Memory\":\"500Mi\",\"Head Memory\":\"3Gi\",\"Ephemeral Storage\":\"5Gi\"}",
+    "ml/ray/start/resources": {
+      "Number of Workers": 1,
+      "CPUs per worker": "500m",
+      "GPUs per worker": 0,
+      "Memory per worker": "1.5Gi",
+      "Ephemeral Storage per worker": "5Gi"
+    },
     "kubernetes/context": "kind-codeflare-test",
     "kubernetes/choose/ns": "default",
     "ml/ray/storage/s3/maybe": "My code does not use Ray Workflows",
diff --git a/tests/kind/profiles/non-gpu5/keep-it-simple b/tests/kind/profiles/non-gpu5/keep-it-simple
@@ -14,7 +14,7 @@
     "ml/codeflare/training/byoc/form": "{\"Path to source\":\"$PWD/tests/kind/inputs/qiskit-with-dashdash\",\"Base image\":\"rayproject/ray:2.1.0\",\"Command line prefix\":\"python3 intentionally-not-main.py\"}",
     "kubernetes/choose/secret/image-pull": "No secret needed, since my image is public",
     "s3/choose/bucket/maybe": "My data is not stored in S3",
-    "ml/ray/start/resources": "{\"Number of CPUs\":\"500m\",\"Number of GPUs\":\"0\",\"Minimum Workers\":\"1\",\"Maximum Workers\":\"1\",\"Worker Memory\":\"500Mi\",\"Head Memory\":\"3Gi\",\"Ephemeral Storage\":\"5Gi\"}",
+    "ml/ray/start/resources": "{\"Number of Workers\":1,\"CPUs per worker\":\"500m\",\"GPUs per worker\":0,\"Memory per worker\":\"1.5Gi\",\"Ephemeral Storage per worker\":\"5Gi\"}",
     "kubernetes/context": "kind-codeflare-test",
     "kubernetes/choose/ns": "default",
     "ml/ray/storage/s3/maybe": "My code does not use Ray Workflows",
diff --git a/tests/kind/profiles/non-gpu6/keep-it-simple b/tests/kind/profiles/non-gpu6/keep-it-simple
@@ -14,7 +14,7 @@
     "ml/codeflare/training/byoc/form": "{\"Path to source\":\"$PWD/tests/kind/inputs/torchx\",\"Base image\":\"ghcr.io/pytorch/torchx:0.5.0dev0\",\"Command line prefix\":\"python3 compute_world_size/main.py\"}",
     "kubernetes/choose/secret/image-pull": "No secret needed, since my image is public",
     "s3/choose/bucket/maybe": "My data is not stored in S3",
-    "ml/torchx/run/resources": "{\"Number of Workers\":\"1\",\"CPUs per worker\":\"500m\",\"GPUs per worker\":\"0\",\"Memory per worker\":\"500Mi\",\"Ephemeral Storage per worker\":\"5Gi\"}",
+    "ml/ray/start/resources": "{\"Number of Workers\":1,\"CPUs per worker\":\"500m\",\"GPUs per worker\":0,\"Memory per worker\":\"1.5Gi\",\"Ephemeral Storage per worker\":\"5Gi\"}",
     "kubernetes/context": "kind-codeflare-test",
     "kubernetes/choose/ns": "default",
     "ml/ray/storage/s3/maybe": "My code does not use Ray Workflows",
diff --git a/tests/kind/profiles/non-gpu6/mcad-default b/tests/kind/profiles/non-gpu6/mcad-default
@@ -14,7 +14,7 @@
     "ml/codeflare/training/byoc/form": "{\"Path to source\":\"$PWD/tests/kind/inputs/torchx\",\"Base image\":\"bitnami/pytorch:1.13.1\",\"Command line prefix\":\"python3 compute_world_size/main.py\"}",
     "kubernetes/choose/secret/image-pull": "No secret needed, since my image is public",
     "s3/choose/bucket/maybe": "My data is not stored in S3",
-    "ml/torchx/run/resources": "{\"Number of Workers\":\"1\",\"CPUs per worker\":\"500m\",\"GPUs per worker\":\"0\",\"Memory per worker\":\"500Mi\",\"Ephemeral Storage per worker\":\"5Gi\"}",
+    "ml/ray/start/resources": "{\"Number of Workers\":1,\"CPUs per worker\":\"500m\",\"GPUs per worker\":0,\"Memory per worker\":\"1.5Gi\",\"Ephemeral Storage per worker\":\"5Gi\"}",
     "kubernetes/context": "kind-codeflare-test",
     "kubernetes/choose/ns": "default",
     "ml/ray/storage/s3/maybe": "My code does not use Ray Workflows",
@@ -24,4 +24,4 @@
     "kubernetes/mcad/choose/job-priority": "Default Priority",
     "kubernetes/mcad/choose/scheduler": "MCAD with the Default Kubernetes Scheduler"
   }
-}
+}
diff --git a/tests/kind/run.sh b/tests/kind/run.sh
@@ -27,9 +27,10 @@ export RAY_KUBE_CLUSTER_NAME=codeflare-test-ray-cluster
 export NODE=node
 export CODEFLARE_HEADLESS_HOME=${CODEFLARE_HEADLESS_HOME-$ROOT/dist/headless}
 
-while getopts "ab:f:is:" opt
+while getopts "Vab:f:is:" opt
 do
     case $opt in
+        V) VERBOSE=true; continue;;
         a) FORCE_ALL=true; continue;;
         f) FORCE=$OPTARG; continue;;
         s) export GUIDEBOOK_STORE=$OPTARG; echo "[Test] Using store=$GUIDEBOOK_STORE"; continue;;
@@ -79,7 +80,10 @@ function run {
     fi
 
     local guidebook=${2-$GUIDEBOOK}
-    local yes=$([ -z "$FORCE_ALL" ] && [ "$FORCE" != "$profileFull" ] && [ -f "$MWPROFILES_PATH/$profile" ] && echo "--yes" || echo "")
+    local yes=${YES-$([ -z "$FORCE_ALL" ] && [ "$FORCE" != "$profileFull" ] && [ -f "$MWPROFILES_PATH/$profile" ] && echo "--yes" || echo "")}
+    if [[ -n "$VERBOSE" ]]; then 
+        local verbose="-V"
+    fi
 
     local dashdashFile="$MWPROFILES_PATH_BASE"/$variant/dashdash.txt
     if [ -f "$dashdashFile" ]; then
@@ -98,7 +102,7 @@ function run {
     fi
 
     echo "[Test] Running with variant=$variant profile=$profile yes=$yes"
-    GUIDEBOOK_NAME="main-job-run" "$ROOT"/bin/codeflare -p $profile $yes $guidebook -- $DASHDASH | tee $OUTPUT
+    GUIDEBOOK_NAME="main-job-run" "$ROOT"/bin/codeflare -p $profile $verbose $yes $guidebook -- $DASHDASH | tee $OUTPUT
 }
 
 #
