From 20e67d4d64ffb7c46227189131814d27e0429da0 Mon Sep 17 00:00:00 2001
From: jgrusewski <jgrusewski@localhost>
Date: Fri, 27 Feb 2026 18:46:16 +0100
Subject: [PATCH] =?UTF-8?q?feat(ci):=20tune=20RL=20resources=20to=20actual?=
 =?UTF-8?q?=20usage=20=E2=80=94=202000m=20CPU,=20concurrent=3D2?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

RL env simulation is single-threaded (~1 core actual). Lowering from
7000m to 2000m request allows DQN + PPO hyperopt to run simultaneously
on one L4 node.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
---
 .gitlab-ci.yml                         | 11 +++++------
 infra/k8s/gitlab/runner-rl-values.yaml | 21 ++++++++++-----------
 2 files changed, 15 insertions(+), 17 deletions(-)

diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index 6f49f74d7..511391418 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -423,12 +423,11 @@ build-training:
   # Picked up by gitlab-runner-rl which mounts training-data-l4-pvc + sccache-l4-pvc
   # Runner default node_selector = ci-compile (L4-1-24G)
   variables:
-    # RL is CPU-bound — full L4 node (7800m CPU, 45Gi RAM, 1 GPU)
-    # concurrent=1 on RL runner → one job gets everything for fastest iteration
-    KUBERNETES_CPU_REQUEST: "7000m"
-    KUBERNETES_CPU_LIMIT: "7800m"
-    KUBERNETES_MEMORY_REQUEST: "16Gi"
-    KUBERNETES_MEMORY_LIMIT: "40Gi"
+    # RL is single-threaded (~1 core actual), 2000m allows 2 concurrent on L4
+    KUBERNETES_CPU_REQUEST: "2000m"
+    KUBERNETES_CPU_LIMIT: "3800m"
+    KUBERNETES_MEMORY_REQUEST: "4Gi"
+    KUBERNETES_MEMORY_LIMIT: "8Gi"
   rules:
     - if: $CI_COMMIT_BRANCH == "main" && $CI_PIPELINE_SOURCE == "push"
       when: manual
diff --git a/infra/k8s/gitlab/runner-rl-values.yaml b/infra/k8s/gitlab/runner-rl-values.yaml
index cec26abc3..034c2bc6e 100644
--- a/infra/k8s/gitlab/runner-rl-values.yaml
+++ b/infra/k8s/gitlab/runner-rl-values.yaml
@@ -34,18 +34,17 @@ runners:
         image = "rust:1.89-slim"
         privileged = false
         node_selector_overwrite_allowed = ".*"
-        cpu_request_overwrite_max_allowed = "7800m"
-        cpu_limit_overwrite_max_allowed = "7800m"
-        memory_request_overwrite_max_allowed = "40Gi"
-        memory_limit_overwrite_max_allowed = "40Gi"
+        cpu_request_overwrite_max_allowed = "4000m"
+        cpu_limit_overwrite_max_allowed = "4000m"
+        memory_request_overwrite_max_allowed = "16Gi"
+        memory_limit_overwrite_max_allowed = "16Gi"
         poll_timeout = 600
         runtime_class_name = "nvidia"
-        # RL jobs are CPU-bound — give full L4 node (7800m alloc, 45Gi RAM, 1 GPU)
-        # concurrent=1 so one job gets everything for fastest iteration
-        cpu_request = "7000m"
-        cpu_limit = "7800m"
-        memory_request = "16Gi"
-        memory_limit = "40Gi"
+        # RL is single-threaded (~1 core actual), 2000m request allows 2 concurrent jobs
+        cpu_request = "2000m"
+        cpu_limit = "3800m"
+        memory_request = "4Gi"
+        memory_limit = "8Gi"
         helper_cpu_request = "100m"
         helper_cpu_limit = "500m"
         helper_memory_request = "128Mi"
@@ -70,4 +69,4 @@ runners:
 
   tags: "kapsule-rl,gpu"
 
-concurrent: 1
+concurrent: 2