#!/usr/bin/env bash
set -euo pipefail

cd "$(dirname "$0")/.."

PYTHON_BIN="${PYTHON_BIN:-python}"
MODEL_ID="${MODEL_ID:-facebook/mms-1b-all}"
TARGET_LANG="${TARGET_LANG:-rmc-script_latin}"
OUTPUT_DIR="${OUTPUT_DIR:-artifacts/models/mms-1b-all-romani-adapter-gpu}"
EVAL_ROOT="${EVAL_ROOT:-artifacts/evals/mms-1b-all-romani-adapter-gpu}"
TRAIN_MANIFEST="${TRAIN_MANIFEST:-artifacts/manifests/clean/train_clean.csv}"
VALIDATION_MANIFEST="${VALIDATION_MANIFEST:-artifacts/manifests/clean/validation_clean.csv}"
TEST_MANIFEST="${TEST_MANIFEST:-artifacts/manifests/test.csv}"
EPOCHS="${EPOCHS:-4}"
MAX_STEPS="${MAX_STEPS:-0}"
TRAIN_LIMIT="${TRAIN_LIMIT:-0}"
VALIDATION_LIMIT="${VALIDATION_LIMIT:-0}"
BATCH_SIZE="${BATCH_SIZE:-1}"
GRAD_ACCUM="${GRAD_ACCUM:-8}"
LEARNING_RATE="${LEARNING_RATE:-1e-3}"
WARMUP_STEPS="${WARMUP_STEPS:-25}"
EVAL_STEPS="${EVAL_STEPS:-50}"
SAVE_ADAPTER_STEPS="${SAVE_ADAPTER_STEPS:-50}"
LOGGING_STEPS="${LOGGING_STEPS:-10}"
NUM_PROC="${NUM_PROC:-2}"
MIXED_PRECISION="${MIXED_PRECISION:-auto}"
RUN_EVALS="${RUN_EVALS:-1}"

"${PYTHON_BIN}" - <<'PY'
import torch

if not torch.cuda.is_available():
    raise SystemExit("CUDA is not available. Select a GPU runtime first.")
print("CUDA:", torch.cuda.get_device_name(0))
print("CUDA memory GB:", round(torch.cuda.get_device_properties(0).total_memory / 1024**3, 1))
PY

train_args=(
  scripts/train_mms_adapter.py
  --model-id "${MODEL_ID}"
  --target-lang "${TARGET_LANG}"
  --train-manifest "${TRAIN_MANIFEST}"
  --validation-manifest "${VALIDATION_MANIFEST}"
  --output-dir "${OUTPUT_DIR}"
  --num-train-epochs "${EPOCHS}"
  --per-device-train-batch-size "${BATCH_SIZE}"
  --per-device-eval-batch-size 1
  --gradient-accumulation-steps "${GRAD_ACCUM}"
  --learning-rate "${LEARNING_RATE}"
  --warmup-steps "${WARMUP_STEPS}"
  --eval-steps "${EVAL_STEPS}"
  --save-adapter-steps "${SAVE_ADAPTER_STEPS}"
  --logging-steps "${LOGGING_STEPS}"
  --num-proc "${NUM_PROC}"
  --mixed-precision "${MIXED_PRECISION}"
  --gradient-checkpointing
)

if [[ "${MAX_STEPS}" != "0" ]]; then
  train_args+=(--max-steps "${MAX_STEPS}")
fi
if [[ "${TRAIN_LIMIT}" != "0" ]]; then
  train_args+=(--train-limit "${TRAIN_LIMIT}")
fi
if [[ "${VALIDATION_LIMIT}" != "0" ]]; then
  train_args+=(--validation-limit "${VALIDATION_LIMIT}")
fi

"${PYTHON_BIN}" "${train_args[@]}"

ADAPTER_DIR_FOR_EVAL="${OUTPUT_DIR}"
PROCESSOR_DIR_FOR_EVAL="${OUTPUT_DIR}/processor"
if [[ -f "${OUTPUT_DIR}/best_adapter/adapter.${TARGET_LANG}.safetensors" ]]; then
  ADAPTER_DIR_FOR_EVAL="${OUTPUT_DIR}/best_adapter"
  PROCESSOR_DIR_FOR_EVAL="${OUTPUT_DIR}/best_adapter/processor"
fi

if [[ "${RUN_EVALS}" != "0" ]]; then
  "${PYTHON_BIN}" scripts/evaluate_mms_asr.py \
    --model-id "${MODEL_ID}" \
    --target-lang "${TARGET_LANG}" \
    --processor-dir "${PROCESSOR_DIR_FOR_EVAL}" \
    --adapter-dir "${ADAPTER_DIR_FOR_EVAL}" \
    --manifest "${VALIDATION_MANIFEST}" \
    --output-dir "${EVAL_ROOT}/validation-clean"

  "${PYTHON_BIN}" scripts/evaluate_mms_asr.py \
    --model-id "${MODEL_ID}" \
    --target-lang "${TARGET_LANG}" \
    --processor-dir "${PROCESSOR_DIR_FOR_EVAL}" \
    --adapter-dir "${ADAPTER_DIR_FOR_EVAL}" \
    --manifest "${TEST_MANIFEST}" \
    --output-dir "${EVAL_ROOT}/test"
fi

echo
echo "Adapter: ${OUTPUT_DIR}/adapter.${TARGET_LANG}.safetensors"
echo "Best/evaluated adapter: ${ADAPTER_DIR_FOR_EVAL}/adapter.${TARGET_LANG}.safetensors"
echo "Processor: ${OUTPUT_DIR}/processor"
if [[ "${RUN_EVALS}" != "0" ]]; then
  echo "Validation metrics: ${EVAL_ROOT}/validation-clean/metrics.json"
  echo "Test metrics: ${EVAL_ROOT}/test/metrics.json"
fi