WIP: try to use d2g for train-backwards

taskcluster/config.yml

@@ -110,6 +110,36 @@ workers:
             implementation: generic-worker
             os: linux
             worker-type: '{alias}'
+        b-linux-v100-gpu-d2g:
+            provisioner: '{trust-domain}-{level}'
+            implementation: docker-worker
+            os: linux
+            worker-type: 'b-linux-v100-gpu'
+        b-linux-v100-gpu-d2g-4:
+            provisioner: '{trust-domain}-{level}'
+            implementation: docker-worker
+            os: linux
+            worker-type: 'b-linux-v100-gpu-4'
+        b-linux-v100-gpu-d2g-4-300gb:
+            provisioner: '{trust-domain}-{level}'
+            implementation: docker-worker
+            os: linux
+            worker-type: 'b-linux-v100-gpu-4-300gb'
+        b-linux-v100-gpu-d2g-4-300gb-standard:
+            provisioner: '{trust-domain}-{level}'
+            implementation: docker-worker
+            os: linux
+            worker-type: 'b-linux-v100-gpu-4-300gb-standard'
+        b-linux-v100-gpu-d2g-4-1tb:
+            provisioner: '{trust-domain}-{level}'
+            implementation: docker-worker
+            os: linux
+            worker-type: 'b-linux-v100-gpu-4-1tb'
+        b-linux-v100-gpu-d2g-4-1tb-standard:
+            provisioner: '{trust-domain}-{level}'
+            implementation: docker-worker
+            os: linux
+            worker-type: 'b-linux-v100-gpu-4-1tb-standard'
         images:
             provisioner: '{trust-domain}-{level}'
             implementation: docker-worker
@@ -158,6 +188,29 @@ local-worker-aliases:
         by-worker-class:
             gcp-standard: 'b-linux-v100-gpu-4-1tb-standard'
             default: 'b-linux-v100-gpu-4-1tb'
+    # Use for quick tasks that need a GPU, eg: evaluate
+    b-gpu-d2g:
+        by-worker-class:
+            gcp-standard: 'b-linux-v100-gpu-d2g'
+            default: 'b-linux-v100-gpu-d2g'
+    # Use for tasks that need lots of GPU power, but not lots of disk space
+    # eg: translation & scoring
+    b-largegpu-d2g:
+        by-worker-class:
+            gcp-standard: 'b-linux-v100-gpu-d2g-4'
+            default: 'b-linux-v100-gpu-d2g-4'
+    # Use for tasks that needs lots of GPU power and increased disk space
+    # eg: bicleaner
+    b-largegpu-d2g-largedisk:
+        by-worker-class:
+            gcp-standard: 'b-linux-v100-gpu-d2g-4-300gb-standard'
+            default: 'b-linux-v100-gpu-d2g-4-300gb'
+    # Use for tasks that need lots of GPU power and immensive amounts of disk space
+    # eg: training
+    b-largegpu-d2g-xlargedisk:
+        by-worker-class:
+            gcp-standard: 'b-linux-v100-gpu-d2g-4-1tb-standard'
+            default: 'b-linux-v100-gpu-d2g-4-1tb'
 
 # Keys are worker type, and align with the `worker-type` entries in the
 # `worker.aliases` above.
@@ -191,3 +244,32 @@ worker-configuration:
         env:
             GPUS: "0 1 2 3"
             WORKSPACE: "12000"
+    b-linux-v100-gpu-d2g:
+        env:
+            GPUS: "0"
+            WORKSPACE: "12000"
+
+    b-linux-v100-gpu-d2g-4:
+        env:
+            GPUS: "0 1 2 3"
+            WORKSPACE: "12000"
+
+    b-linux-v100-gpu-d2g-4-300gb:
+        env:
+            GPUS: "0 1 2 3"
+            WORKSPACE: "12000"
+
+    b-linux-v100-gpu-d2g-4-1tb:
+        env:
+            GPUS: "0 1 2 3"
+            WORKSPACE: "12000"
+
+    b-linux-v100-gpu-d2g-4-300gb-standard:
+        env:
+            GPUS: "0 1 2 3"
+            WORKSPACE: "12000"
+
+    b-linux-v100-gpu-d2g-4-1tb-standard:
+        env:
+            GPUS: "0 1 2 3"
+            WORKSPACE: "12000"

taskcluster/kinds/train-backwards/kind.yml

@@ -57,12 +57,15 @@ tasks:
                 - worker.env
         worker-type:
             by-tasks-for:
-                github-pull-request: b-largegpu
-                default: b-largegpu-largedisk
+                github-pull-request: b-largegpu-d2g
+                default: b-largegpu-d2g-largedisk
+
         worker:
+            docker-image: {"in-tree": "train"}
             max-run-time: 2592000
             # train_taskcluster.py exits with 17 if a request to Taskcluster fails
             retry-exit-status: [17]
+
             env:
                 ARTIFACT_EXT: zst
                 COMPRESSION_CMD: zstdmt
@@ -72,6 +75,7 @@ tasks:
 
                 # Weight & Biases publication token is stored in that secret
                 TASKCLUSTER_SECRET: project/translations/level-1/weights-and-biases
+
             artifacts:
                 - name: public/build
                   path: artifacts