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

# Quantize Reka yasa2 mmproj F16 -> mostly Q8_0. Some tensors (patch embed, depthwise, down conv)
# cannot use Q8_0 when ncols is not divisible by 32; those stay F16 (same approach as a full Q8_0 run).
#
# Usage: ./scripts/quantize_reka_mmproj_q8.sh [INPUT_F16_GGUF] [OUTPUT_GGUF]
# Optional: QUANTIZE_BIN (default: build_linux/bin or build/bin llama-quantize)

SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
REPO_ROOT="$(cd "$SCRIPT_DIR/.." && pwd)"
cd "$REPO_ROOT"

INPUT_GGUF="${1:-}"
OUTPUT_GGUF="${2:-}"
THREADS="${THREADS:-16}"

if [[ -z "$INPUT_GGUF" || -z "$OUTPUT_GGUF" ]]; then
  echo "Usage: $0 <INPUT_mmproj_F16.gguf> <OUTPUT_mmproj.gguf>" >&2
  exit 1
fi

if [[ -z "${QUANTIZE_BIN:-}" ]]; then
  if [[ -x "$REPO_ROOT/build_linux/bin/llama-quantize" ]]; then
    QUANTIZE_BIN="$REPO_ROOT/build_linux/bin/llama-quantize"
  else
    QUANTIZE_BIN="$REPO_ROOT/build/bin/llama-quantize"
  fi
fi

"$QUANTIZE_BIN" \
  --tensor-type 'patch_embd=f16' \
  --tensor-type '.*\.dw\.weight=f16' \
  --tensor-type '.*down\.conv\.weight=f16' \
  "$INPUT_GGUF" \
  "$OUTPUT_GGUF" \
  Q8_0 \
  "$THREADS"