#!/usr/bin/env bash

set -euo pipefail

# Step 1: fp4/fp8 -> bf16
INPUT_FP8_HF_PATH="deepseek-ai/DeepSeek-V4-Flash"
OUTPUT_BF16_HF_PATH="deepseek-ai/DeepSeek-V4-Flash-bf16"

python3 convert_weight.py \
  --input-fp8-hf-path "${INPUT_FP8_HF_PATH}" \
  --output-bf16-hf-path "${OUTPUT_BF16_HF_PATH}"

# Step 2: bf16 -> bf16-mp16
MP=8
HF_CKPT_BF16_PATH="${OUTPUT_BF16_HF_PATH}"
SAVE_PATH="deepseek-ai/DeepSeek-V4-Flash-bf16-mp16"
EXPERTS=256

python3 convert.py \
  --hf-ckpt-path "${HF_CKPT_BF16_PATH}" \
  --save-path "${SAVE_PATH}" \
  --n-experts "${EXPERTS}" \
  --model-parallel "${MP}"