CUDA_PATH ?= /usr/local/cuda ifeq ($(wildcard $(CUDA_PATH)/bin/nvcc),) CUDA_PATH := /usr endif NVCC := $(CUDA_PATH)/bin/nvcc CXX := g++ NVCC_FLAGS := -O3 -c -Xcompiler -fPIC -std=c++17 -I. -I$(CUDA_PATH)/include CXX_FLAGS := -O3 -fPIC -std=c++17 -I. -I$(CUDA_PATH)/include LDFLAGS := -L$(CUDA_PATH)/lib64 -L. -lcublas -lcudart -Wl,-rpath,'$$ORIGIN' -Wl,-rpath,$(CUDA_PATH)/lib64 TARGET_LIB := libparvulus.so TARGET_CLI := parvulus_runtime TARGET_TEST := test_vision_encoder .PHONY: all clean distclean test weights all: $(TARGET_LIB) $(TARGET_CLI) $(TARGET_TEST) weights: @if [ ! -f parvulus_model_weights.bin ] || [ ! -f action_scale.bin ] || [ ! -f task_embeddings.bin ]; then \ python3 export_from_safetensors.py; \ fi parvulus_engine.o: parvulus_engine.cu parvulus_engine.h vision_encoder.cu vision_encoder.cuh vision_encoder.h kernels.cuh $(NVCC) $(NVCC_FLAGS) parvulus_engine.cu -o parvulus_engine.o $(TARGET_LIB): parvulus_engine.o parvulus_c_api.cpp parvulus_c_api.h $(CXX) -shared -fPIC parvulus_engine.o parvulus_c_api.cpp -o $(TARGET_LIB) $(CXX_FLAGS) $(LDFLAGS) $(TARGET_CLI): parvulus_runtime_cli.cpp $(TARGET_LIB) $(CXX) $(CXX_FLAGS) parvulus_runtime_cli.cpp -o $(TARGET_CLI) -lparvulus $(LDFLAGS) $(TARGET_TEST): test_vision_encoder.cpp $(TARGET_LIB) $(CXX) $(CXX_FLAGS) test_vision_encoder.cpp -o $(TARGET_TEST) -lparvulus $(LDFLAGS) test: all weights ./$(TARGET_TEST) ./$(TARGET_CLI) --benchmark 20 clean: rm -f parvulus_engine.o $(TARGET_LIB) $(TARGET_CLI) $(TARGET_TEST) distclean: clean rm -f parvulus_model_weights.bin action_scale.bin task_embeddings.bin *.onnx