feat(infra): split training image into RL + supervised, rename ci-compile → ci-rl

RL models (DQN/PPO) only need basic CUDA runtime (~2GB), while supervised models need cuDNN (~4GB). Splitting saves ~2GB pull time per RL job. Rename the L4 GPU pool from ci-compile to ci-rl to reflect its actual use. Add DaemonSet image pre-puller to cache training images on GPU nodes. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
2026-02-28 10:16:42 +01:00
parent f9b57cdecb
commit c731bef759
13 changed files with 199 additions and 71 deletions
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -9,7 +9,7 @@
 #   kapsule,rust → CPU runner → ci-compile-cpu pool (POP2-32C-128G, 32 vCPU)
 #   kapsule,gpu  → GPU runner → ci-training (L40S)
 # Override nodeSelector: KUBERNETES_NODE_SELECTOR_k8s.scaleway.com/pool-name: "<name>" (k=v format)
-# Pools: ci-compile-cpu (POP2-32C-128G), ci-training (L40S), services, gitlab
+# Pools: ci-compile-cpu (POP2-32C-128G), ci-rl (L4), ci-training (L40S), services, gitlab
 # All compilation (services + training) runs on ci-compile-cpu. CUDA stubs in ci-builder image.

 # Ensure pipeline is always created — individual job rules handle filtering.
@@ -538,31 +538,41 @@ build-web-gateway:
      --destination "${REGISTRY}/web-gateway:${CI_COMMIT_SHA}"
      --destination "${REGISTRY}/web-gateway:latest"

-# training: pre-built CUDA binaries → CUDA runtime image (no recompilation)
-build-training:
+# training-rl: RL binaries → CUDA runtime image (no cuDNN, ~2GB)
+build-training-rl:
  extends: .kaniko-training-base
  script:
    - /kaniko/executor
      --context dir://${CI_PROJECT_DIR}/build-out/training
-      --dockerfile "${CI_PROJECT_DIR}/infra/docker/Dockerfile.training-runtime"
-      --destination "${REGISTRY}/training:${CI_COMMIT_SHA}"
-      --destination "${REGISTRY}/training:latest"
+      --dockerfile "${CI_PROJECT_DIR}/infra/docker/Dockerfile.training-rl"
+      --destination "${REGISTRY}/training-rl:${CI_COMMIT_SHA}"
+      --destination "${REGISTRY}/training-rl:latest"
+
+# training-supervised: supervised binaries → CUDA+cuDNN runtime image (~4GB)
+build-training-supervised:
+  extends: .kaniko-training-base
+  script:
+    - /kaniko/executor
+      --context dir://${CI_PROJECT_DIR}/build-out/training
+      --dockerfile "${CI_PROJECT_DIR}/infra/docker/Dockerfile.training-supervised"
+      --destination "${REGISTRY}/training-supervised:${CI_COMMIT_SHA}"
+      --destination "${REGISTRY}/training-supervised:latest"

 # --------------------------------------------------------------------------
 # Stage 3c: Training — RL on L4 (CPU-heavy), Supervised on L40S (GPU-heavy)
 # --------------------------------------------------------------------------

 # --------------------------------------------------------------------------
-# RL training base: DQN + PPO → ci-compile pool (L4-1-24G)
+# RL training base: DQN + PPO → ci-rl pool (L4-1-24G)
 # RL hyperopt runs parallel PSO trials — CPU+GPU bound, low VRAM.
 # --------------------------------------------------------------------------
 .train-rl-base:
  stage: train
-  image: ${REGISTRY}/training:${CI_COMMIT_SHA}
+  image: ${REGISTRY}/training-rl:${CI_COMMIT_SHA}
  tags:
    - kapsule
    - gpu
-  # Runner default node_selector: k8s.scaleway.com/pool-name=ci-compile-cpu
+  # Runner default node_selector: k8s.scaleway.com/pool-name=ci-rl
  variables:
    # nvidia runtime provided by runner default (runtime_class_name = "nvidia")
    # RL hyperopt is CPU-bound (parallel PSO trials). Use full L4 (8 vCPU).
@@ -589,7 +599,7 @@ build-training:
 # --------------------------------------------------------------------------
 .train-validate-base:
  stage: train
-  image: ${REGISTRY}/training:${CI_COMMIT_SHA}
+  image: ${REGISTRY}/training-supervised:${CI_COMMIT_SHA}
  tags:
    - kapsule
    - gpu
@@ -619,7 +629,7 @@ build-training:
 train-validate-rl:
  extends: .train-rl-base
  needs:
-    - job: build-training
+    - job: build-training-rl
      optional: true
  script:
    - |
@@ -651,7 +661,7 @@ train-validate-rl:
 train-validate-tft:
  extends: .train-validate-base
  needs:
-    - job: build-training
+    - job: build-training-supervised
      optional: true
  script:
    - /usr/local/bin/train_baseline_supervised
@@ -678,7 +688,7 @@ train-validate-tft:
 hyperopt-ppo:
  extends: .train-rl-base
  needs:
-    - job: build-training
+    - job: build-training-rl
      optional: true
  timeout: 4h
  script:
@@ -725,7 +735,7 @@ hyperopt-ppo:
 train-validate-mamba2:
  extends: .train-validate-base
  needs:
-    - job: build-training
+    - job: build-training-supervised
      optional: true
  script:
    - /usr/local/bin/train_baseline_supervised
@@ -752,7 +762,7 @@ train-validate-mamba2:
 train-validate-liquid:
  extends: .train-validate-base
  needs:
-    - job: build-training
+    - job: build-training-supervised
      optional: true
  script:
    - /usr/local/bin/train_baseline_supervised
@@ -779,7 +789,7 @@ train-validate-liquid:
 train-validate-tggn:
  extends: .train-validate-base
  needs:
-    - job: build-training
+    - job: build-training-supervised
      optional: true
  script:
    - /usr/local/bin/train_baseline_supervised
@@ -806,7 +816,7 @@ train-validate-tggn:
 train-validate-tlob:
  extends: .train-validate-base
  needs:
-    - job: build-training
+    - job: build-training-supervised
      optional: true
  script:
    - /usr/local/bin/train_baseline_supervised
@@ -833,7 +843,7 @@ train-validate-tlob:
 train-validate-kan:
  extends: .train-validate-base
  needs:
-    - job: build-training
+    - job: build-training-supervised
      optional: true
  script:
    - /usr/local/bin/train_baseline_supervised
@@ -860,7 +870,7 @@ train-validate-kan:
 train-validate-xlstm:
  extends: .train-validate-base
  needs:
-    - job: build-training
+    - job: build-training-supervised
      optional: true
  script:
    - /usr/local/bin/train_baseline_supervised
@@ -887,7 +897,7 @@ train-validate-xlstm:
 train-validate-diffusion:
  extends: .train-validate-base
  needs:
-    - job: build-training
+    - job: build-training-supervised
      optional: true
  script:
    - /usr/local/bin/train_baseline_supervised
@@ -916,7 +926,7 @@ train-validate-diffusion:
 hyperopt-dqn:
  extends: .train-rl-base
  needs:
-    - job: build-training
+    - job: build-training-rl
      optional: true
  timeout: 4h
  script:
@@ -968,7 +978,7 @@ hyperopt-dqn:
 hyperopt-tft:
  extends: .train-validate-base
  needs:
-    - job: build-training
+    - job: build-training-supervised
      optional: true
  timeout: 4h
  script:
@@ -1007,7 +1017,7 @@ hyperopt-tft:
 hyperopt-mamba2:
  extends: .train-validate-base
  needs:
-    - job: build-training
+    - job: build-training-supervised
      optional: true
  timeout: 4h
  script:
@@ -1046,7 +1056,7 @@ hyperopt-mamba2:
 hyperopt-liquid:
  extends: .train-validate-base
  needs:
-    - job: build-training
+    - job: build-training-supervised
      optional: true
  timeout: 4h
  script:
@@ -1085,7 +1095,7 @@ hyperopt-liquid:
 hyperopt-tggn:
  extends: .train-validate-base
  needs:
-    - job: build-training
+    - job: build-training-supervised
      optional: true
  timeout: 4h
  script:
@@ -1124,7 +1134,7 @@ hyperopt-tggn:
 hyperopt-tlob:
  extends: .train-validate-base
  needs:
-    - job: build-training
+    - job: build-training-supervised
      optional: true
  timeout: 4h
  script:
@@ -1163,7 +1173,7 @@ hyperopt-tlob:
 hyperopt-kan:
  extends: .train-validate-base
  needs:
-    - job: build-training
+    - job: build-training-supervised
      optional: true
  timeout: 4h
  script:
@@ -1202,7 +1212,7 @@ hyperopt-kan:
 hyperopt-xlstm:
  extends: .train-validate-base
  needs:
-    - job: build-training
+    - job: build-training-supervised
      optional: true
  timeout: 4h
  script:
@@ -1241,7 +1251,7 @@ hyperopt-xlstm:
 hyperopt-diffusion:
  extends: .train-validate-base
  needs:
-    - job: build-training
+    - job: build-training-supervised
      optional: true
  timeout: 4h
  script:
@@ -1412,7 +1422,9 @@ deploy:
      optional: true
    - job: build-web-gateway
      optional: true
-    - job: build-training
+    - job: build-training-rl
+      optional: true
+    - job: build-training-supervised
      optional: true
  rules:
    # Only deploy when source code changed (build produced new images)
--- a/infra/docker/Dockerfile.training-runtime
+++ b/infra/docker/Dockerfile.training-runtime
@@ -1,9 +1,9 @@
-# GPU training runtime image — packages pre-built CUDA binaries (no compilation)
-# Binaries are compiled in CI compile-services job with --features ml/cuda
-# Usage: kaniko --context dir://build-out --dockerfile Dockerfile.training-runtime
-# Run:   docker run --gpus all training train_baseline_supervised --model kan [args...]
+# RL training runtime image — DQN + PPO (no cuDNN needed, ~2GB)
+# Binaries are compiled in CI compile-training job with --features ml/cuda
+# Usage: kaniko --context dir://build-out/training --dockerfile Dockerfile.training-rl
+# Run:   docker run --gpus all training-rl train_baseline_rl --model dqn [args...]

-FROM nvidia/cuda:12.4.1-cudnn-runtime-ubuntu22.04
+FROM nvidia/cuda:12.4.1-runtime-ubuntu22.04

 ENV DEBIAN_FRONTEND=noninteractive

@@ -23,20 +23,14 @@ RUN groupadd -g 1000 foxhunt \
    && useradd -u 1000 -g foxhunt -m -s /bin/false foxhunt

 COPY train_baseline_rl \
-     train_baseline_supervised \
     evaluate_baseline \
-     evaluate_supervised \
     hyperopt_baseline_rl \
-     hyperopt_baseline_supervised \
     training_uploader \
     /usr/local/bin/

 RUN chmod +x /usr/local/bin/train_baseline_rl \
-             /usr/local/bin/train_baseline_supervised \
             /usr/local/bin/evaluate_baseline \
-             /usr/local/bin/evaluate_supervised \
             /usr/local/bin/hyperopt_baseline_rl \
-             /usr/local/bin/hyperopt_baseline_supervised \
             /usr/local/bin/training_uploader

 # CUDA runtime environment
--- a/infra/docker/Dockerfile.training-supervised
+++ b/infra/docker/Dockerfile.training-supervised
@@ -0,0 +1,45 @@
+# Supervised training runtime image — TFT, Mamba2, TGGN, TLOB, Liquid, KAN, xLSTM, Diffusion (~4GB)
+# Includes cuDNN for deep neural network operations (convolutions, attention)
+# Binaries are compiled in CI compile-training job with --features ml/cuda
+# Usage: kaniko --context dir://build-out/training --dockerfile Dockerfile.training-supervised
+# Run:   docker run --gpus all training-supervised train_baseline_supervised --model tft [args...]
+
+FROM nvidia/cuda:12.4.1-cudnn-runtime-ubuntu22.04
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN apt-get update && apt-get install -y --no-install-recommends \
+    ca-certificates \
+    libssl3 \
+    curl \
+    unzip \
+    && curl -fsSL https://downloads.rclone.org/v1.69.1/rclone-v1.69.1-linux-amd64.zip -o /tmp/rclone.zip \
+    && unzip -j /tmp/rclone.zip '*/rclone' -d /usr/local/bin/ \
+    && chmod +x /usr/local/bin/rclone \
+    && rm /tmp/rclone.zip \
+    && apt-get purge -y unzip \
+    && rm -rf /var/lib/apt/lists/*
+
+RUN groupadd -g 1000 foxhunt \
+    && useradd -u 1000 -g foxhunt -m -s /bin/false foxhunt
+
+COPY train_baseline_supervised \
+     evaluate_supervised \
+     hyperopt_baseline_supervised \
+     training_uploader \
+     /usr/local/bin/
+
+RUN chmod +x /usr/local/bin/train_baseline_supervised \
+             /usr/local/bin/evaluate_supervised \
+             /usr/local/bin/hyperopt_baseline_supervised \
+             /usr/local/bin/training_uploader
+
+# CUDA runtime environment
+ENV NVIDIA_VISIBLE_DEVICES=all
+ENV NVIDIA_DRIVER_CAPABILITIES=compute,utility
+
+USER foxhunt
+
+WORKDIR /data
+
+CMD ["echo", "Specify training command via K8s Job args"]
--- a/infra/k8s/gitlab/runner-rl-values.yaml
+++ b/infra/k8s/gitlab/runner-rl-values.yaml
@@ -1,5 +1,5 @@
 # GitLab Runner — RL training workloads (DQN, PPO)
-# Runs on ci-compile pool (L4: 24GB VRAM, cheaper, CPU-heavy for env simulation)
+# Runs on ci-rl pool (L4: 24GB VRAM, cheaper, CPU-heavy for env simulation)
 # Mounts separate PVCs (training-data-l4-pvc, sccache-l4-pvc) to avoid RWO conflicts

 gitlabUrl: http://gitlab-webservice-default.foxhunt.svc.cluster.local:8181
@@ -51,7 +51,7 @@ runners:
        helper_memory_limit = "512Mi"
        image_pull_secrets = ["scw-registry", "gitlab-registry"]
      [runners.kubernetes.node_selector]
-        "k8s.scaleway.com/pool-name" = "ci-compile"
+        "k8s.scaleway.com/pool-name" = "ci-rl"
      [runners.kubernetes.node_tolerations]
        "nvidia.com/gpu" = "NoSchedule"
        "node.cilium.io/agent-not-ready" = "NoSchedule"
--- a/infra/k8s/gitlab/runner-values.yaml
+++ b/infra/k8s/gitlab/runner-values.yaml
@@ -1,6 +1,6 @@
 # GitLab Runner — Kubernetes executor
 # Runner manager pod lives on gitlab node pool
-# Default: build pods spawn on ci-compile pool (L4: 24GB VRAM, cheaper)
+# Default: build pods spawn on ci-rl pool (L4: 24GB VRAM, cheaper)
 # Training jobs override to ci-training pool (L40S: 48GB VRAM) via node_selector

 gitlabUrl: http://gitlab-webservice-default.foxhunt.svc.cluster.local:8181
@@ -38,11 +38,11 @@ runners:
        memory_limit_overwrite_max_allowed = "80Gi"
        # L40S scale-to-zero needs ~3-5 min to provision; default 180s times out
        poll_timeout = 600
-        # All GPU node pools (ci-compile L4, ci-training L40S) have nvidia GPU
+        # All GPU node pools (ci-rl L4, ci-training L40S) have nvidia GPU
        # Setting runtimeClassName=nvidia on all pods so CUDA works everywhere
        runtime_class_name = "nvidia"
        # Default resource limits — balanced for both Rust compile and Kaniko builds
-        # L4 node: 8 vCPU / 30Gi RAM, ci-compile pool max_size=2
+        # L4 node: 8 vCPU / 30Gi RAM, ci-rl pool max_size=2
        # Compile/test: runs 2 in parallel (1 per node), needs ~20Gi for ml+CUDA linking
        # Kaniko: runs after compile, needs ~200m/512Mi but gets these defaults
        # With 3500m/12Gi request: 2 pods fit per L4 node = 4 concurrent Kaniko builds
@@ -57,7 +57,7 @@ runners:
        image_pull_secrets = ["scw-registry", "gitlab-registry"]
      # Sub-tables must come AFTER all scalar values (TOML rule)
      [runners.kubernetes.node_selector]
-        "k8s.scaleway.com/pool-name" = "ci-compile"
+        "k8s.scaleway.com/pool-name" = "ci-rl"
      [runners.kubernetes.node_tolerations]
        "nvidia.com/gpu" = "NoSchedule"
        # Cilium CNI agent takes ~30s to initialize on fresh scale-from-zero nodes
--- a/infra/k8s/monitoring/APPLY-ORDER.md
+++ b/infra/k8s/monitoring/APPLY-ORDER.md
@@ -94,7 +94,7 @@

 4. Repeat for gitlab pool (if not autoscaling, same process).

-5. Autoscaling pools (ci-training, ci-compile, gpu-dev):
+5. Autoscaling pools (ci-training, ci-rl, gpu-dev):
   - No action needed — next scale-up automatically uses private IP.

 6. Verify Tailscale still works:
--- a/infra/k8s/monitoring/dcgm-exporter.yaml
+++ b/infra/k8s/monitoring/dcgm-exporter.yaml
@@ -1,5 +1,5 @@
 # NVIDIA DCGM Exporter — GPU metrics for Cockpit
-# Runs on all GPU node pools (ci-training, ci-compile, gpu-dev)
+# Runs on all GPU node pools (ci-training, ci-rl, gpu-dev)
 # Alloy auto-discovers and scrapes these pods via annotations
 ---
 apiVersion: v1
--- a/infra/k8s/training/image-prepuller.yaml
+++ b/infra/k8s/training/image-prepuller.yaml
@@ -0,0 +1,71 @@
+# DaemonSet image pre-puller — keeps training images cached on GPU nodes
+# Runs on ci-rl (L4) and ci-training (L40S) pools so training jobs skip the pull.
+# Init containers pull :latest tags, then the main container sleeps forever.
+---
+apiVersion: apps/v1
+kind: DaemonSet
+metadata:
+  name: image-prepuller
+  namespace: foxhunt
+  labels:
+    app: image-prepuller
+    app.kubernetes.io/part-of: foxhunt
+spec:
+  selector:
+    matchLabels:
+      app: image-prepuller
+  template:
+    metadata:
+      labels:
+        app: image-prepuller
+    spec:
+      affinity:
+        nodeAffinity:
+          requiredDuringSchedulingIgnoredDuringExecution:
+            nodeSelectorTerms:
+              - matchExpressions:
+                  - key: k8s.scaleway.com/pool-name
+                    operator: In
+                    values:
+                      - ci-rl
+                      - ci-training
+      tolerations:
+        - key: nvidia.com/gpu
+          operator: Exists
+          effect: NoSchedule
+        - key: node.cilium.io/agent-not-ready
+          operator: Exists
+          effect: NoSchedule
+      imagePullSecrets:
+        - name: scw-registry
+      initContainers:
+        - name: pull-training-rl
+          image: rg.fr-par.scw.cloud/foxhunt-ci/training-rl:latest
+          command: ["echo", "training-rl image pulled"]
+          resources:
+            requests:
+              cpu: 10m
+              memory: 16Mi
+            limits:
+              cpu: 10m
+              memory: 16Mi
+        - name: pull-training-supervised
+          image: rg.fr-par.scw.cloud/foxhunt-ci/training-supervised:latest
+          command: ["echo", "training-supervised image pulled"]
+          resources:
+            requests:
+              cpu: 10m
+              memory: 16Mi
+            limits:
+              cpu: 10m
+              memory: 16Mi
+      containers:
+        - name: pause
+          image: registry.k8s.io/pause:3.10
+          resources:
+            requests:
+              cpu: 10m
+              memory: 16Mi
+            limits:
+              cpu: 10m
+              memory: 16Mi
--- a/infra/k8s/training/job-template.yaml
+++ b/infra/k8s/training/job-template.yaml
@@ -31,9 +31,10 @@ spec:
        - name: scw-registry
      restartPolicy: Never
      # Native sidecar (K8s 1.28+) — uploads artifacts on training completion
+      # Use training-rl:latest for DQN/PPO, training-supervised:latest for others
      initContainers:
        - name: uploader
-          image: rg.fr-par.scw.cloud/foxhunt-ci/training:latest
+          image: rg.fr-par.scw.cloud/foxhunt-ci/training-supervised:latest
          restartPolicy: Always
          command: ["/usr/local/bin/training_uploader"]
          env:
@@ -74,7 +75,7 @@ spec:
              memory: 256Mi
      containers:
        - name: training
-          image: rg.fr-par.scw.cloud/foxhunt-ci/training:latest
+          image: rg.fr-par.scw.cloud/foxhunt-ci/training-supervised:latest
          # Available binaries in /usr/local/bin/:
          #   train_baseline_rl (for dqn, ppo)
          #   train_baseline_supervised (for tft, mamba2, tggn, tlob, liquid, kan, xlstm, diffusion)
--- a/infra/live/production/kapsule/terragrunt.hcl
+++ b/infra/live/production/kapsule/terragrunt.hcl
@@ -16,10 +16,10 @@ inputs = {
  enable_gitlab_pool = true
  gitlab_type        = "GP1-XS"

-  # CI compile pool (L4 for cargo check/test/build — CUDA CC 89, 24GB VRAM)
-  enable_ci_compile_pool = true
-  ci_compile_type        = "L4-1-24G"
-  ci_compile_max_size    = 1
+  # CI RL pool (L4 for DQN/PPO training + hyperopt — CUDA CC 89, 24GB VRAM)
+  enable_ci_rl_pool = true
+  ci_rl_type        = "L4-1-24G"
+  ci_rl_max_size    = 1

  # CPU compile pool (POP2-32C-128G for cargo check/build — 32 vCPU, 128GB RAM, no GPU)
  enable_ci_compile_cpu_pool = true
--- a/infra/modules/kapsule/main.tf
+++ b/infra/modules/kapsule/main.tf
@@ -11,6 +11,11 @@ moved {
  to   = scaleway_k8s_pool.ci_training
 }

+moved {
+  from = scaleway_k8s_pool.ci_compile
+  to   = scaleway_k8s_pool.ci_rl
+}
+
 resource "scaleway_vpc_private_network" "foxhunt" {
  name   = "${var.cluster_name}-pn"
  region = var.region
@@ -71,16 +76,16 @@ resource "scaleway_k8s_pool" "gitlab" {
  region             = var.region
 }

-# CI compile pool (L4 — 24GB VRAM, CUDA CC 89, for RL training + hyperopt)
-# Nodes auto-labeled: k8s.scaleway.com/pool-name=ci-compile
-resource "scaleway_k8s_pool" "ci_compile" {
-  count       = var.enable_ci_compile_pool ? 1 : 0
+# CI RL pool (L4 — 24GB VRAM, CUDA CC 89, for RL training + hyperopt)
+# Nodes auto-labeled: k8s.scaleway.com/pool-name=ci-rl
+resource "scaleway_k8s_pool" "ci_rl" {
+  count       = var.enable_ci_rl_pool ? 1 : 0
  cluster_id  = scaleway_k8s_cluster.foxhunt.id
-  name        = "ci-compile"
-  node_type   = var.ci_compile_type
+  name        = "ci-rl"
+  node_type   = var.ci_rl_type
  size        = 1
  min_size    = 0
-  max_size    = var.ci_compile_max_size
+  max_size    = var.ci_rl_max_size
  autoscaling        = true
  autohealing        = true
  public_ip_disabled = var.public_ip_disabled
--- a/infra/modules/kapsule/outputs.tf
+++ b/infra/modules/kapsule/outputs.tf
@@ -24,9 +24,9 @@ output "gitlab_pool_id" {
  value       = var.enable_gitlab_pool ? scaleway_k8s_pool.gitlab[0].id : ""
 }

-output "ci_compile_pool_id" {
-  description = "ID of the CI compile node pool (L4)"
-  value       = var.enable_ci_compile_pool ? scaleway_k8s_pool.ci_compile[0].id : ""
+output "ci_rl_pool_id" {
+  description = "ID of the CI RL node pool (L4)"
+  value       = var.enable_ci_rl_pool ? scaleway_k8s_pool.ci_rl[0].id : ""
 }

 output "ci_compile_cpu_pool_id" {
--- a/infra/modules/kapsule/variables.tf
+++ b/infra/modules/kapsule/variables.tf
@@ -39,20 +39,20 @@ variable "gitlab_type" {
  default     = "DEV1-L"
 }

-variable "enable_ci_compile_pool" {
-  description = "Create L4 GPU node pool for CI compilation + testing"
+variable "enable_ci_rl_pool" {
+  description = "Create L4 GPU node pool for RL training + hyperopt"
  type        = bool
  default     = false
 }

-variable "ci_compile_type" {
-  description = "Instance type for the CI compile pool (L4 for CUDA compilation)"
+variable "ci_rl_type" {
+  description = "Instance type for the CI RL pool (L4 for DQN/PPO training)"
  type        = string
  default     = "L4-1-24G"
 }

-variable "ci_compile_max_size" {
-  description = "Maximum number of nodes in the CI compile pool"
+variable "ci_rl_max_size" {
+  description = "Maximum number of nodes in the CI RL pool"
  type        = number
  default     = 1
 }