From 3bd51ec7db64a8f75710e3257fca2d2e44f93a5c Mon Sep 17 00:00:00 2001
From: Christian Medina <37550954+cmedinasoriano@users.noreply.github.com>
Date: Wed, 1 Jul 2026 14:09:34 -0400
Subject: [PATCH] feat: add screen support to bypass cgroup limits

---
 train-on-this-server.sh | 12 +++++++++++-
 1 file changed, 11 insertions(+), 1 deletion(-)

diff --git a/train-on-this-server.sh b/train-on-this-server.sh
index 815d187..fd9b0b1 100755
--- a/train-on-this-server.sh
+++ b/train-on-this-server.sh
@@ -80,7 +80,17 @@ echo ""
 # Use torchrun for distributed training (2 GPUs)
 # Add --check-only to validate without training:
 # torchrun --nproc_per_node=2 training/scripts/train.py --config training/configs/ornith-35b-lora.yaml --check-only
-torchrun --nproc_per_node=2 training/scripts/train.py --config training/configs/ornith-35b-lora.yaml
+
+# Run in screen if available (bypasses cgroup limits)
+if command -v screen &> /dev/null; then
+    echo "Running in screen session..."
+    screen -dmS lora_training bash -c "torchrun --nproc_per_node=2 training/scripts/train.py --config training/configs/ornith-35b-lora.yaml"
+    echo "Training started in screen session 'lora_training'"
+    echo "Watch with: screen -r lora_training"
+else
+    echo "Screen not available, running directly..."
+    torchrun --nproc_per_node=2 training/scripts/train.py --config training/configs/ornith-35b-lora.yaml
+fi
 
 echo ""
 echo "=== Training Complete ==="