Makefile

# learn-llm —— 纯 C 实现的 Qwen3-0.6B 推理
#
#   make                    编译 ./qwen3
#   make ref                用 PyTorch 导出对拍所需的参考数值
#   make test               编译并运行四组对拍测试（需先 make ref）
#   make clean              清掉全部生成物
#   make BACKEND=gcd|naive  换 matmul 后端

CC      := cc
CFLAGS  := -std=c11 -O2 -g -Wall -Wextra -Wshadow -Wconversion -Wno-sign-conversion
LDFLAGS := -lm

# matmul 后端: accelerate（Apple BLAS）| gcd（GCD 多线程）| naive（可移植的朴素循环）
BACKEND ?= accelerate
ifeq ($(BACKEND),accelerate)
  CFLAGS  += -DUSE_ACCELERATE -DACCELERATE_NEW_LAPACK
  LDFLAGS += -framework Accelerate
else ifeq ($(BACKEND),gcd)
  CFLAGS  += -DUSE_GCD
endif

MODEL   ?= ../models/Qwen3-0.6B
PYTHON  ?= python

# 除 main.c 外的所有源文件，主程序和测试共用
LIB_SRC := $(filter-out src/main.c,$(wildcard src/*.c))
LIB_OBJ := $(LIB_SRC:.c=.o)
TESTS   := test_ops test_layer test_forward test_tokenizer

.PHONY: all test ref clean

all: qwen3

qwen3: $(LIB_OBJ) src/main.o
	$(CC) $^ -o $@ $(LDFLAGS)

$(TESTS): %: $(LIB_OBJ) tests/%.o
	$(CC) $^ -o $@ $(LDFLAGS)

%.o: %.c
	$(CC) $(CFLAGS) -c $< -o $@

ref:
	@for c in ops layer tok forward; do $(PYTHON) tests/dump_ref.py $$c || exit 1; done

test: $(TESTS)
	@./test_ops       $(MODEL)
	@echo
	@./test_layer     $(MODEL)
	@echo
	@./test_forward   $(MODEL)
	@echo
	@./test_tokenizer $(MODEL) ref/tok_cases.txt

clean:
	rm -rf $(LIB_OBJ) src/main.o tests/*.o qwen3 $(TESTS) ref __pycache__ tests/__pycache__