#!/bin/bash
#SBATCH --job-name=smhier-canary
#SBATCH --account=torch_pr_375_cds
#SBATCH --partition=h200_cds
#SBATCH --gres=gpu:h200:1
#SBATCH --cpus-per-task=8
#SBATCH --mem=64G
#SBATCH --time=01:30:00
#SBATCH --output=/scratch/ajc10180/repos/smooth-hierarchy/logs/canary-%j.out
#SBATCH --error=/scratch/ajc10180/repos/smooth-hierarchy/logs/canary-%j.err

set -euo pipefail

REPO=/scratch/ajc10180/repos/smooth-hierarchy
DATA=/scratch/ajc10180/data
SHIM_DIR=$REPO/.shim
SCRATCH_BASE=/scratch/ajc10180
mkdir -p "$REPO/logs" "$REPO/checkpoints"

echo "[canary] $(date -u +%FT%TZ) starting on $(hostname)"
nvidia-smi || echo "[canary] WARN: nvidia-smi unavailable"

# ---- Scratch caches (match setup.sbatch) ----
export UV_CACHE_DIR="$SCRATCH_BASE/.uv-cache"
export UV_PYTHON_INSTALL_DIR="$SCRATCH_BASE/.uv-python"
export UV_TOOL_DIR="$SCRATCH_BASE/.uv-tool"
export PIP_CACHE_DIR="$SCRATCH_BASE/.pip-cache"
export HF_HOME="$SCRATCH_BASE/.hf-home"
export TRANSFORMERS_CACHE="$HF_HOME"
export TORCH_HOME="$SCRATCH_BASE/.torch-home"

# ---- FFmpeg shared libs (for torchcodec) ----
FFMPEG_DIR="$SCRATCH_BASE/ffmpeg-shared"
export LD_LIBRARY_PATH="$FFMPEG_DIR/lib:${LD_LIBRARY_PATH:-}"
export PATH="$FFMPEG_DIR/bin:$PATH"

# ---- NVIDIA lib dirs from torch wheel (for torchcodec CUDA NPP etc.) ----
NVLIBS=""
for d in "$REPO"/.venv/lib/python3.10/site-packages/nvidia/*/lib; do
  [[ -d "$d" ]] && NVLIBS="$d:$NVLIBS"
done
export LD_LIBRARY_PATH="${NVLIBS}$LD_LIBRARY_PATH"
echo "[canary] NVIDIA libs: $(echo $NVLIBS | tr ':' '\n' | wc -l) dirs"

# ---- uv on PATH + reuse existing venv ----
export PATH="$HOME/.local/bin:$PATH"
cd "$REPO"
command -v uv
uv --version

# ---- W&B + HF keys via KeyHandler ----
WORKSPACE=/scratch/ajc10180/repos/smooth-hierarchy
# KeyHandler lives in the workspace .tools-venv; source env by running a short python snippet locally.
# But on cluster we don't have workspace .tools-venv — export via cluster ~/.raca if present, else expect env.
if [[ -f "$HOME/.raca/keys.env" ]]; then
  # shellcheck disable=SC1091
  . "$HOME/.raca/keys.env"
fi
: "${WANDB_API_KEY:?WANDB_API_KEY not set — export before sbatch or put in ~/.raca/keys.env}"
echo "[canary] WANDB_API_KEY: ${WANDB_API_KEY:0:10}..."

# ---- Verify torchcodec now loads on GPU node (the thing setup couldn't verify on CPU) ----
./.venv/bin/python -c "
import torch
print('torch', torch.__version__, 'cuda avail:', torch.cuda.is_available(), 'device:', torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'cpu')
import torchcodec
print('torchcodec', torchcodec.__version__)
from torchcodec.decoders import VideoDecoder
import glob
vids = sorted(glob.glob('/scratch/ajc10180/data/ucf101/*/*.avi'))[:1]
d = VideoDecoder(vids[0])
f = d[0]
print('decoded shape:', tuple(f.shape), 'dtype:', f.dtype, 'n_frames:', d.metadata.num_frames)
print('GPU SMOKE TEST OK')
"

# ---- Canary training run ----
# Per EXPERIMENT_README.md canary spec:
#   1 GPU, ~1 hr, --limit_train_batches 200 for 2 epochs
#   image_size 64, batch_size 128 (README validated)
#   Verify: checkpoints save, W&B logs 4 loss components, recon panels at beta in {0, 0.1, 0.5, 0.9, 1.0}
echo "[canary] $(date -u +%FT%TZ) launching train.py"
./.venv/bin/python -u train.py \
  --data_root "$DATA/ucf101" \
  --annotation_path "$DATA/ucf101_annotations" \
  --dataset ucf101 \
  --image_size 64 \
  --batch_size 128 \
  --num_workers 8 \
  --max_epochs 2 \
  --limit_train_batches 200 \
  --checkpoint_dir "$REPO/checkpoints/canary-${SLURM_JOB_ID}" \
  --checkpoint_every_n_train_steps 100 \
  --wandb_project smooth-hierarchy \
  --wandb_run_name "canary-${SLURM_JOB_ID}" \
  --wandb_tags canary,ucf101,image64 \
  --wandb_notes "smooth-hierarchy-ucf101 canary: 2 epochs x 200 batches to verify pipeline" \
  --wandb_log_images_every 50 \
  --accelerator gpu \
  --devices 1

echo "[canary] $(date -u +%FT%TZ) DONE"
ls -la "$REPO/checkpoints/canary-${SLURM_JOB_ID}/" || true