Scriptwriter dataset sync

restart-train.sh

@@ -3,20 +3,12 @@
 set -euo pipefail
 
 echo "=== Force-stopping old training ==="
-# Kill tmux session hard
 tmux kill-session -t scriptwriter-train 2>/dev/null || true
 tmux kill-server 2>/dev/null || true
-# Kill any leftover trainers / python
 pkill -9 -f 'train_runpod.py' 2>/dev/null || true
 pkill -9 -f 'scripts/train_runpod' 2>/dev/null || true
 sleep 2
-# Confirm gone
-if pgrep -af 'train_runpod' >/dev/null 2>&1; then
-  echo "WARNING: train process still alive:"
-  pgrep -af 'train_runpod' || true
-  pkill -9 -f train_runpod || true
-  sleep 1
-fi
+pgrep -af 'train_runpod' >/dev/null 2>&1 && pkill -9 -f train_runpod || true
 echo "Old session cleared."
 
 DEST=/workspace/scriptwriter-runpod
@@ -33,19 +25,20 @@ export HF_DATASET_REPO="${HF_DATASET_REPO:-datamatters24/scriptwriter-corpus-ia}
 export HF_MODEL_REPO="${HF_MODEL_REPO:-datamatters24/scriptwriter-lora-ia}"
 export BASE_MODEL="${BASE_MODEL:-meta-llama/Llama-3.2-3B-Instruct}"
 export RUNPOD_SKIP_FORCE_PIP=1
-export MAX_SEQ_LENGTH_CAP="${MAX_SEQ_LENGTH_CAP:-4096}"
+export MAX_SEQ_LENGTH=2048
+export MAX_SEQ_LENGTH_CAP=2048
+export PER_DEVICE_TRAIN_BATCH_SIZE=1
+export GRADIENT_ACCUMULATION_STEPS=8
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+export RESUME_FROM_CHECKPOINT=1
 
 if [[ ! -f /workspace/data/processed/train.jsonl ]]; then
   echo "train.jsonl missing — use fix-and-train.sh instead"
   exit 1
 fi
 
-# Skip heavy pip in runpod_train_tmux by marking skip
 cd "$DEST"
-# Patch: run worker directly to avoid re-download delays
 mkdir -p /workspace/logs /workspace/models/lora
-LOG=/workspace/logs/train_$(date +%Y%m%d_%H%M%S).log
-# Ensure ROOT data symlink
 mkdir -p "$DEST/data"
 ln -sfn /workspace/data/processed "$DEST/data/processed"
 
@@ -63,9 +56,16 @@ export CONFIG_PATH=/workspace/scriptwriter-runpod/config/training.yaml
 export BASE_MODEL="${BASE_MODEL:-meta-llama/Llama-3.2-3B-Instruct}"
 export OUTPUT_DIR=/workspace/models/lora
 export HF_MODEL_REPO="${HF_MODEL_REPO:-datamatters24/scriptwriter-lora-ia}"
-export MAX_SEQ_LENGTH_CAP="${MAX_SEQ_LENGTH_CAP:-4096}"
+export MAX_SEQ_LENGTH=2048
+export MAX_SEQ_LENGTH_CAP=2048
+export PER_DEVICE_TRAIN_BATCH_SIZE=1
+export GRADIENT_ACCUMULATION_STEPS=8
+export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True
+export RESUME_FROM_CHECKPOINT=1
 exec > >(tee -a /workspace/logs/train_live.log) 2>&1
 echo "========== TRAIN START $(date -Is) =========="
+echo "OOM-safe: seq=2048 batch=1 accum=8 gpu0-only + checkpointing"
+ls /workspace/models/lora/checkpoint-* 2>/dev/null | tail -5 || echo "no prior checkpoint"
 nvidia-smi || true
 python3 scripts/train_runpod.py
 echo "========== UPLOAD $(date -Is) =========="
@@ -73,11 +73,10 @@ python3 scripts/sync_hf.py upload-model --repo "$HF_MODEL_REPO" --folder /worksp
 echo "========== DONE $(date -Is) =========="
 echo "STOP THE POD"
 INNER
-# rewrite worker with expanded log path - the heredoc above uses train_live.log which is fine
 chmod +x "$WORKER"
 
-tmux has-session -t scriptwriter-train 2>/dev/null && tmux kill-session -t scriptwriter-train || true
 tmux new-session -d -s scriptwriter-train -n train "bash $WORKER"
 tmux new-window -t scriptwriter-train -n monitor
 tmux send-keys -t scriptwriter-train:monitor "watch -n 15 'nvidia-smi --query-gpu=index,memory.used,utilization.gpu --format=csv; echo; tail -25 /workspace/logs/train_live.log'" Enter
 echo "Started tmux scriptwriter-train — attach: tmux attach -t scriptwriter-train"
+echo "OOM fix: 2048 tokens, batch 1, gradient checkpointing; resumes checkpoint-100 if present"

scriptwriter-runpod-ready.tgz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:737bbc057cf99a7fd6f3cb64a481da4e1bd8fa803b4887c9d231c012eeb766c6
-size 9728
+oid sha256:4746d83e090484c84c0ec1d60401c872a0c25a7e6990fae6eec9bf4b836df63c
+size 10052