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

cd /e2e-data/evad-tech-vla/wanghan58/workspace/LTA_openwebtext_dualt
export PYTHONPATH="$(pwd)${PYTHONPATH:+:$PYTHONPATH}"
export TOKENIZERS_PARALLELISM=false
export PYTHONUNBUFFERED=1

RUN_NAME="${RUN_NAME:-smoke_lta_openwebtext_dirichlet_dualt_len1024_1gpu}"
CHECKPOINT="${CHECKPOINT:-runs/${RUN_NAME}/latest.pt}"
DATA_PATH="${DATA_PATH:-/e2e-data/evad-tech-vla/wanghan58/data/small_benchmarks/langflow_2604_11748/openwebtext}"
TOKENIZER_PATH="${TOKENIZER_PATH:-/e2e-data/evad-tech-vla/wanghan58/models/flowtext_scorers/gpt2-standard/tokenizer.json}"
GPU="${GPU:-0}"

CUDA_VISIBLE_DEVICES="${GPU}" python eval.py \
  --checkpoint "${CHECKPOINT}" \
  --data_path "${DATA_PATH}" \
  --text_column text \
  --tokenizer_path "${TOKENIZER_PATH}" \
  --max_records 8 \
  --max_len 1024 \
  --batch_size 1 \
  --mask_ratios 0.1,1.0 \
  --fill_t 0.0 \
  --infer_steps 8 \
  --decode_solver flowmap \
  --noise_init logistic_normal \
  --pure_noise_samples 2