Makefile

SHELL := /bin/bash

# Configuration variables
NATIVE_ANDROID = $(abspath ../Native-LLM-for-Android)
QWEN_VL_DIR = $(NATIVE_ANDROID)/Export_ONNX/QwenVL
ONNX_SRC_DIR = $(QWEN_VL_DIR)/onnx
ONNX_DEST_DIR = $(QWEN_VL_DIR)/onnx-dist
STAGING_DIR = /tmp/transformers.js/staging
TRANSFORMERS_JS_PATH = ../transformers.js
ONNX_TOOLS_PATH = $(NATIVE_ANDROID)/ONNX_Tools

# Python paths from venvs
NATIVE_PYTHON = $(NATIVE_ANDROID)/.venv/bin/python3
TRANSFORMERS_PYTHON = $(TRANSFORMERS_JS_PATH)/.venv/bin/python3

# Model parts
PARTS = A B C D E

define progress_bar
	printf "\r   Progress: \033[1;32m["; \
	_done=$$(($1 * 20 / $2)); \
	for ((i=0; i<_done; i++)); do printf "="; done; \
	printf "\033[0m"; \
	_left=$$((20 - _done)); \
	for ((i=0; i<_left; i++)); do printf " "; done; \
	printf "\033[1;32m]\033[0m $1/$2  Processing: \033[1;34m%s\033[K\033[0m\r" "$3"
endef

# See https://github.com/pytorch/pytorch/issues/94280#issuecomment-2089196400
# Original export scripts export a bunch of tensor files, so we merge into one / two files instead.
export-merged-source-models: export-merged-source-models-first-pass export-merged-source-models-second-pass
	@echo "✅ Exporting merged source models complete"

export-merged-source-models-first-pass:
	@echo "💾 First pass: Export all models with merged tensors..."
	@mkdir -p $(ONNX_DEST_DIR)
	@files=`find $(ONNX_SRC_DIR) -name "*.onnx"`; \
	total=`echo "$$files" | wc -w | tr -d ' '`; \
	echo "Files found (first pass): $$total"; \
	current=0; \
	for item in $$files; do \
	current=$$((current + 1)); \
	$(call progress_bar,$$current,$$total,$$item); \
	$(NATIVE_PYTHON) -u -c "import onnx, os, sys; src='$$item'; dest_dir='$(ONNX_DEST_DIR)'; \
		m = onnx.load(src); \
		d = os.path.join(dest_dir, os.path.basename(src)); \
		onnx.save_model(m, d, all_tensors_to_one_file=True, save_as_external_data=True, location=os.path.basename(d)+'.data')" || exit 1; \
	done; \
	echo "✅ Done first pass"

export-merged-source-models-second-pass:
	@echo "💾 Second pass: Converting large models to external data format..."
	@files=`find $(ONNX_DEST_DIR) -name "*.onnx"`; \
	total=`echo "$$files" | wc -w | tr -d ' '`; \
	echo "Files found (second pass): $$total"; \
	current=0; \
	for item in $$files; do \
		current=$$((current + 1)); \
		$(call progress_bar,$$current,$$total,$$item); \
		$(NATIVE_PYTHON) -c 'import onnx, os, sys; \
			src = """'"$$item"'"""; \
			total_size = os.path.getsize(src); \
			d = os.path.join(dest_dir, os.path.basename(src)); \
			total_size += os.path.getsize(src + ".data") if os.path.exists(src + ".data") else 0; \
			needs_external = total_size > 2e9; \
			onnx.save_model( \
				onnx.load(src), \
				d, \
				save_as_external_data=needs_external, \
				all_tensors_to_one_file=True, \
				location=(os.path.basename(src) + ".data") if needs_external else None \
			); \
			not needs_external and os.path.exists(src + ".data") and os.remove(src + ".data") \
			' || exit 1; \
	done; \
	echo "✅ Done second models"


all-in-one: export quantize clean-large-files fix-gpu-buffers export-merged-source-models
	@echo "✨ All done! ONNX models exported, slimmed, quantized and fixed"

export: export-abcd export-e
	@echo "✅ Export complete"

export-abcd:
	@echo "🚀 Exporting parts A, B, C, D..."
	cd ../Native-LLM-for-Android/Export_ONNX/QwenVL && \
	$(NATIVE_PYTHON) QwenVL_Export_ABCD.py "Qwen/Qwen2-VL-2B-Instruct"

export-e:
	@echo "🚀 Exporting part E..."
	cd ../Native-LLM-for-Android/Export_ONNX/QwenVL && \
	$(NATIVE_PYTHON) QwenVL_Export_E.py "Qwen/Qwen2-VL-2B-Instruct"

slim:
	@echo "🗜️  Slimming ONNX models..."
	@files=`find $(ONNX_SRC_DIR) -name "*.onnx" -type f ! -name "QwenVL_E.onnx"`; \
	total=`echo "$$files" | wc -w | tr -d ' '`; \
	echo "Files found: $$total"; \
	current=0; \
	for item in $$files; do \
		current=$$((current + 1)); \
		$(call progress_bar,$$current,$$total,$$item); \
		onnxslim --verbose "$$item" "$$item" || exit 1; \
	done; \
	echo "✅ Slimming complete"

quantize:
	@echo "⚡ Starting quantization..."
	for part in $(PARTS); do \
		$(MAKE) quantize-$$part || exit 1; \
	done
	@echo "✅ Quantization complete"

quantize-%:
	@echo "⚡ Quantizing part $*..."
	mkdir -p $(ONNX_DEST_DIR)
	cd $(TRANSFORMERS_JS_PATH) && \
	mkdir -p $(STAGING_DIR) && \
	rm -f $(STAGING_DIR)/* && \
	ln -sf $$(realpath $(ONNX_SRC_DIR))/* $(STAGING_DIR)/ && \
	find $(STAGING_DIR) -name "*_*_*.onnx_data" -delete && \
	find $(STAGING_DIR) -name "*_*_*.onnx" -delete && \
	find $(STAGING_DIR) -name "*.onnx" ! -name "QwenVL_$**.onnx" -delete && \
	EXTRA_FLAGS=""; \
	if [ "$*" = "A" ]; then EXTRA_FLAGS="--op_block_list Conv DynamicQuantizeLinear DequantizeLinear Resize"; fi; \
	echo "Extra Flags for part $*: $$EXTRA_FLAGS" && \
	PYTHONPATH=$(TRANSFORMERS_JS_PATH) .venv/bin/python3 -m scripts.quantize \
		--input_folder '$(STAGING_DIR)' \
		--output_folder '$(ONNX_DEST_DIR)' \
		--mode q4f16 $$EXTRA_FLAGS

clean-large-files:
	@echo "🧹 Removing ONNX files over 2GB..."
	cd $(ONNX_DEST_DIR) && \
	for f in $$(find . -name "*.onnx" -type f); do \
		total_size=0; \
		if [ -f "$$f"".data" ]; then \
			total_size=$$(( $$(stat -f %z "$$f") + $$(stat -f %z "$$f"".data") )); \
		elif [ -f "$$f""_data" ]; then \
			total_size=$$(( $$(stat -f %z "$$f") + $$(stat -f %z "$$f""_data") )); \
		else \
			total_size=$$(stat -f %z "$$f"); \
		fi; \
		size_mb=$$(( total_size / 1048576 )); \
		if [ $$total_size -ge 2147483648 ]; then \
			echo "   Removing $$f (size: $$size_mb MB)..."; \
			rm -f "$$f" "$$f"".data" "$$f""_data"; \
		fi \
	done
	@echo "✅ Large file cleanup complete"

fix-gpu-buffers:
	@echo "🔧 Fixing GPU buffers for E models..."
	@files=`find $(ONNX_DEST_DIR) -name "QwenVL_E_*.onnx" -type f`; \
	total=`echo "$$files" | wc -w | tr -d ' '`; \
	echo "Files found: $$total"; \
	current=0; \
	for item in $$files; do \
		current=$$((current + 1)); \
		$(call progress_bar,$$current,$$total,$$item); \
		cd $(NATIVE_ANDROID) && .venv/bin/python3 ONNX_Tools/clamp_for_gpu_buffers.py --overwrite "$$item" || exit 1; \
	done; \
	echo "✅ GPU buffer fixes complete"