#!/usr/bin/env bash
# Quantize Qwen/Qwen3-32B into 4 schemes, one at a time, upload each, then
# delete the local checkpoint. Disk is 500GB and two 97GB GPUs are available,
# so we shard the model across both GPUs (device_map="auto", no CPU offload)
# and run the full-quality calibration footprint. The base stays cached across
# schemes (no re-download).

MODEL="Qwen/Qwen3-32B"
PY="/home/vllm_env/bin/python"
CALIB_SAMPLES=512
CALIB_SEQLEN=2048
cd /home

# Make both GPUs visible for model sharding.
export CUDA_VISIBLE_DEVICES=0,1

# shellcheck disable=SC1091
source /home/vllm_env/bin/activate

# Reduce CUDA fragmentation during calibration. Auth uses the persisted HF
# credential file (~/.cache/huggingface/token), so no env token is required.
export PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

run() {
  local scheme="$1"; shift
  echo "===== START ${scheme} $(date) ====="
  "${PY}" quantize_my_model.py \
    --scheme "${scheme}" \
    --model-id "${MODEL}" \
    --num-calibration-samples "${CALIB_SAMPLES}" \
    --max-seq-length "${CALIB_SEQLEN}" \
    --upload-to-hub \
    --delete-local-after-upload \
    "$@" \
    && echo "===== OK ${scheme} $(date) =====" \
    || echo "!!!!! FAILED ${scheme} $(date) !!!!!"
  echo "--- disk after ${scheme} ---"; df -h /home
}

run nvfp4
run mxfp4
run fp8
run mxfp8

echo "ALL DONE $(date)"