aboutsummaryrefslogtreecommitdiff
path: root/Makefile
blob: f1285aa6e9fb2dc2f36cbb011bdc527ff9c25fef (plain) (blame)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
CC          = gcc
# Allow FP reassociation and contraction so reduction loops vectorize, but do
# not assume finite math or approximate libm calls: full -ffast-math produced
# NaN embeddings on GCC/OpenBLAS Linux builds (-ffinite-math-only and
# -fapprox-func are the unsafe parts and stay off).
# -fvisibility=hidden pairs with EMBED_API in the public headers: a new
# public function without the annotation will be missing from the shared lib.
CFLAGS_BASE = -Wall -Wextra -O3 -march=native -fPIC -fvisibility=hidden \
              -fno-math-errno -ffp-contract=fast -fno-trapping-math \
              -fno-signed-zeros -fassociative-math -freciprocal-math
LDFLAGS     = -lm -lpthread
CJSON_CFLAGS ?= $(shell sh -c 'pkg-config --cflags libcjson 2>/dev/null')
CJSON_LDFLAGS ?= $(shell sh -c 'pkg-config --libs libcjson 2>/dev/null')
ifeq ($(strip $(CJSON_LDFLAGS)),)
    CJSON_LDFLAGS = -lcjson
endif

UNAME_S := $(shell uname -s)

# Source files
SRCS = embed.c \
       qwen_kernels.c \
       qwen_kernels_generic.c \
       qwen_kernels_neon.c \
       qwen_kernels_avx.c \
       qwen_tokenizer.c \
       qwen_safetensors.c \
       deps/ae/ae.c \
       deps/ae/anet.c \
       deps/ae/monotonic.c

MLX_SRCS = embed_mlx.c
CUDA_SRCS = embed_cuda.cu

OBJS     = $(SRCS:.c=.o)
MLX_OBJS = $(MLX_SRCS:.c=.o)
TARGET        = embed
SERVER_TARGET = embed-server
LIB           = libembed.a

KERNEL_SRCS = qwen_kernels.c qwen_kernels_generic.c \
              qwen_kernels_neon.c qwen_kernels_avx.c

ifeq ($(shell uname -s),Darwin)
SHARED_LIB   = libembed.dylib
SHARED_FLAGS = -dynamiclib
else
SHARED_LIB   = libembed.so
SHARED_FLAGS = -shared
endif

.PHONY: all cpu mlx cuda gpu test test-bf16 test-safetensors coverage bench \
        bench-model bench-tokenizer debug clean help

all: help

help:
	@echo "Targets:"
	@echo "  make cpu      CPU-only build"
	@echo "  make gpu      Build with GPU backend support (mlx on macOS, cuda on Linux)"
	@echo "  make mlx      Apple Silicon GPU via MLX"
	@echo "  make cuda     NVIDIA GPU via CUDA"
	@echo "  make test     Build and run the test suite (no model files needed)"
	@echo "  make coverage Test-suite line coverage report (clang/llvm-cov)"
	@echo "  make bench    Kernel microbenchmarks; records bench/results/*.json"
	@echo "  make bench-model MODEL_DIR=/path/to/model"
	@echo "                End-to-end embedding benchmark"
	@echo "  make debug    Debug build"
	@echo "  make clean    Remove build artifacts"

# =============================================================================
# CPU build (BLAS: Apple Accelerate on macOS, OpenBLAS on Linux)
# =============================================================================
ifeq ($(UNAME_S),Darwin)
ifeq ($(strip $(CJSON_CFLAGS)),)
    CJSON_CFLAGS = -I/opt/homebrew/include -I/usr/local/include
endif
ifeq ($(strip $(CJSON_LDFLAGS)),-lcjson)
    CJSON_LDFLAGS = -L/opt/homebrew/lib -L/usr/local/lib -lcjson
endif
cpu: CFLAGS  = $(CFLAGS_BASE) -DUSE_BLAS -DACCELERATE_NEW_LAPACK
cpu: LDFLAGS += -framework Accelerate
else
cpu: CFLAGS  = $(CFLAGS_BASE) -DUSE_BLAS -DUSE_OPENBLAS -I/usr/include/openblas
cpu: LDFLAGS += -lopenblas
endif
cpu:
	$(MAKE) clean
	$(MAKE) $(TARGET) $(SERVER_TARGET) $(SHARED_LIB) CFLAGS="$(CFLAGS)" LDFLAGS="$(LDFLAGS)"

# =============================================================================
# MLX build (Apple Silicon GPU via MLX - uses the mlx-c pure C API)
# Requires: brew install mlx mlx-c
# =============================================================================
MLX_PREFIX  := $(shell brew --prefix mlx 2>/dev/null)
MLXC_PREFIX := $(shell brew --prefix mlx-c 2>/dev/null)

mlx: SRCS += embed_mlx.c
mlx: OBJS  = $(SRCS:.c=.o) embed_mlx.o
ifeq ($(UNAME_S),Darwin)
mlx: CFLAGS  = $(CFLAGS_BASE) -DUSE_BLAS -DACCELERATE_NEW_LAPACK -DUSE_MLX \
               -I$(MLXC_PREFIX)/include
mlx: LDFLAGS += -framework Accelerate -framework Metal -framework Foundation \
                -L$(MLXC_PREFIX)/lib -lmlxc \
                -L$(MLX_PREFIX)/lib -lmlx \
                -Wl,-rpath,$(MLX_PREFIX)/lib -Wl,-rpath,$(MLXC_PREFIX)/lib
endif
mlx:
	$(MAKE) clean
	$(MAKE) $(TARGET) $(SERVER_TARGET) $(SHARED_LIB) CFLAGS="$(CFLAGS)" LDFLAGS="$(LDFLAGS)" EXTRA_OBJS="embed_mlx.o"

# =============================================================================
# GPU convenience target (auto-selects mlx on macOS, cuda on Linux)
# =============================================================================
ifeq ($(UNAME_S),Darwin)
gpu: mlx
else
gpu: cuda
endif

# =============================================================================
# CUDA build (Linux NVIDIA GPU - uses CUDA Runtime + cuBLAS)
# =============================================================================
CUDA_HOME ?= /usr/local/cuda
NVCC ?= $(CUDA_HOME)/bin/nvcc

cuda: CFLAGS  = $(CFLAGS_BASE) -DUSE_BLAS -DUSE_OPENBLAS -DUSE_CUDA -I/usr/include/openblas -I$(CUDA_HOME)/include
cuda: LDFLAGS += -L$(CUDA_HOME)/lib64 -lopenblas -lcudart -lcublas
cuda:
	$(MAKE) clean
	$(MAKE) embed_cuda.o
	$(MAKE) $(TARGET) $(SERVER_TARGET) $(SHARED_LIB) CFLAGS="$(CFLAGS)" LDFLAGS="$(LDFLAGS)" EXTRA_OBJS="embed_cuda.o"

# -fPIC so the same object links into both binaries and the shared library.
# -arch=native targets the GPU present at build time (we always build CUDA
# on the machine that runs it), avoiding first-load PTX JIT.
embed_cuda.o: embed_cuda.cu embed_cuda.h embed_internal.h embed.h
	$(NVCC) -O3 -std=c++17 -arch=native -Xcompiler -fPIC,-fvisibility=hidden -DUSE_BLAS -DUSE_OPENBLAS -DUSE_CUDA -I/usr/include/openblas -I$(CUDA_HOME)/include -x cu -c -o $@ $<

# =============================================================================
# Shared library (built by every backend target from the same objects)
# =============================================================================
$(SHARED_LIB): $(OBJS) $(EXTRA_OBJS)
	$(CC) $(CFLAGS) $(SHARED_FLAGS) -o $@ $^ $(LDFLAGS) $(CJSON_LDFLAGS)


# =============================================================================
# Tests (no model files required)
# =============================================================================
ifeq ($(UNAME_S),Darwin)
TEST_BLAS_CFLAGS  = -DUSE_BLAS -DACCELERATE_NEW_LAPACK
TEST_BLAS_LDFLAGS = -framework Accelerate
else
TEST_BLAS_CFLAGS  = -DUSE_BLAS -DUSE_OPENBLAS -I/usr/include/openblas
TEST_BLAS_LDFLAGS = -lopenblas
endif

# Source lists shared by test and coverage targets so the two cannot drift.
TEST_CC_FLAGS         = -Wall -Wextra -O2 -I.
TEST_KERNELS_SRCS     = tests/test_kernels.c $(KERNEL_SRCS)
TEST_TOKENIZER_SRCS   = tests/test_tokenizer.c qwen_tokenizer.c $(KERNEL_SRCS)
TEST_SAFETENSORS_SRCS = tests/test_safetensors.c qwen_safetensors.c
TEST_BF16_SRCS        = tests/test_bf16_model.c embed.c qwen_safetensors.c \
                        $(KERNEL_SRCS)
TEST_QWEN3_SRCS       = tests/test_qwen3.c embed.c qwen_safetensors.c \
                        $(KERNEL_SRCS)
TEST_WORKSPACE_SRCS   = tests/test_workspace.c embed.c qwen_tokenizer.c \
                        qwen_safetensors.c $(KERNEL_SRCS)
TEST_LATE_SRCS        = tests/test_late.c embed.c qwen_tokenizer.c \
                        qwen_safetensors.c $(KERNEL_SRCS)
TEST_SERVER_SRCS      = tests/test_server.c embed.c \
                        qwen_tokenizer.c qwen_safetensors.c $(KERNEL_SRCS) \
                        deps/ae/ae.c deps/ae/anet.c deps/ae/monotonic.c
# The CLI check builds the real CLI (CPU backend) plus a driver that runs it.
TEST_CLI_BIN_SRCS     = embed_cli.c embed.c \
                        qwen_tokenizer.c qwen_safetensors.c $(KERNEL_SRCS)

test:
	$(CC) $(TEST_CC_FLAGS) -o tests/test_kernels_generic \
	    $(TEST_KERNELS_SRCS) -lm -lpthread
	./tests/test_kernels_generic
	$(CC) $(TEST_CC_FLAGS) $(TEST_BLAS_CFLAGS) -o tests/test_kernels_blas \
	    $(TEST_KERNELS_SRCS) -lm -lpthread $(TEST_BLAS_LDFLAGS)
	./tests/test_kernels_blas
	$(CC) $(TEST_CC_FLAGS) -o tests/test_tokenizer \
	    $(TEST_TOKENIZER_SRCS) -lm -lpthread
	./tests/test_tokenizer
	$(CC) $(TEST_CC_FLAGS) -o tests/test_safetensors $(TEST_SAFETENSORS_SRCS)
	./tests/test_safetensors
	$(CC) $(TEST_CC_FLAGS) -o tests/test_bf16_model \
	    $(TEST_BF16_SRCS) -lm -lpthread
	./tests/test_bf16_model
	$(CC) $(TEST_CC_FLAGS) $(TEST_BLAS_CFLAGS) -o tests/test_qwen3 \
	    $(TEST_QWEN3_SRCS) -lm -lpthread $(TEST_BLAS_LDFLAGS)
	./tests/test_qwen3
	$(CC) $(TEST_CC_FLAGS) -o tests/test_workspace \
	    $(TEST_WORKSPACE_SRCS) -lm -lpthread
	./tests/test_workspace
	$(CC) $(TEST_CC_FLAGS) $(TEST_BLAS_CFLAGS) -o tests/test_late \
	    $(TEST_LATE_SRCS) -lm -lpthread $(TEST_BLAS_LDFLAGS)
	./tests/test_late
	$(CC) $(TEST_CC_FLAGS) $(TEST_BLAS_CFLAGS) -o tests/cli_under_test \
	    $(TEST_CLI_BIN_SRCS) -lm -lpthread $(TEST_BLAS_LDFLAGS)
	$(CC) $(TEST_CC_FLAGS) -o tests/test_cli tests/test_cli.c -lm
	./tests/test_cli ./tests/cli_under_test
	$(CC) $(TEST_CC_FLAGS) $(TEST_BLAS_CFLAGS) $(CJSON_CFLAGS) -Ideps/ae \
	    -o tests/test_server $(TEST_SERVER_SRCS) \
	    -lm -lpthread $(TEST_BLAS_LDFLAGS) $(CJSON_LDFLAGS)
	./tests/test_server

test-bf16:
	$(CC) $(TEST_CC_FLAGS) -o tests/test_bf16_model \
	    $(TEST_BF16_SRCS) -lm -lpthread
	./tests/test_bf16_model

test-safetensors:
	$(CC) $(TEST_CC_FLAGS) -o tests/test_safetensors \
	    $(TEST_SAFETENSORS_SRCS) -lm -lpthread
	./tests/test_safetensors


# =============================================================================
# Test-suite line coverage (requires clang + llvm-cov/llvm-profdata. Writes a
# per-file text summary to stdout and an HTML report to coverage/html/index.html.)
# =============================================================================
COV_DIR   = coverage
# Source-based coverage; -O0 so no line is folded away by the optimizer.
# The "." compilation dir keeps report paths relative to the repo root;
# run llvm-cov from the repo root so sources resolve.
COV_FLAGS = -fprofile-instr-generate -fcoverage-mapping -O0 \
            -fcoverage-compilation-dir=.
COV_BINS  = tests/test_kernels_generic tests/test_kernels_blas \
            tests/test_tokenizer tests/test_safetensors \
            tests/test_bf16_model tests/test_qwen3 tests/test_workspace tests/test_late \
            tests/cli_under_test tests/test_server
# Report on project sources only - not the harnesses, vendored deps, or
# system headers (cJSON lands under /opt or /usr otherwise).
COV_IGNORE = -ignore-filename-regex='deps/|tests/|/opt/|/usr/'

ifeq ($(UNAME_S),Darwin)
LLVM_PROFDATA = xcrun llvm-profdata
LLVM_COV      = xcrun llvm-cov
else
LLVM_PROFDATA = llvm-profdata
LLVM_COV      = llvm-cov
endif

coverage:
	rm -rf $(COV_DIR) && mkdir -p $(COV_DIR)
	$(CC) $(TEST_CC_FLAGS) $(COV_FLAGS) -o tests/test_kernels_generic \
	    $(TEST_KERNELS_SRCS) -lm -lpthread
	LLVM_PROFILE_FILE=$(COV_DIR)/kernels_generic.profraw ./tests/test_kernels_generic
	$(CC) $(TEST_CC_FLAGS) $(COV_FLAGS) $(TEST_BLAS_CFLAGS) -o tests/test_kernels_blas \
	    $(TEST_KERNELS_SRCS) -lm -lpthread $(TEST_BLAS_LDFLAGS)
	LLVM_PROFILE_FILE=$(COV_DIR)/kernels_blas.profraw ./tests/test_kernels_blas
	$(CC) $(TEST_CC_FLAGS) $(COV_FLAGS) -o tests/test_tokenizer \
	    $(TEST_TOKENIZER_SRCS) -lm -lpthread
	LLVM_PROFILE_FILE=$(COV_DIR)/tokenizer.profraw ./tests/test_tokenizer
	$(CC) $(TEST_CC_FLAGS) $(COV_FLAGS) -o tests/test_safetensors $(TEST_SAFETENSORS_SRCS)
	LLVM_PROFILE_FILE=$(COV_DIR)/safetensors.profraw ./tests/test_safetensors
	$(CC) $(TEST_CC_FLAGS) $(COV_FLAGS) -o tests/test_bf16_model \
	    $(TEST_BF16_SRCS) -lm -lpthread
	LLVM_PROFILE_FILE=$(COV_DIR)/bf16_model.profraw ./tests/test_bf16_model
	$(CC) $(TEST_CC_FLAGS) $(COV_FLAGS) $(TEST_BLAS_CFLAGS) -o tests/test_qwen3 \
	    $(TEST_QWEN3_SRCS) -lm -lpthread $(TEST_BLAS_LDFLAGS)
	LLVM_PROFILE_FILE=$(COV_DIR)/qwen3.profraw ./tests/test_qwen3
	$(CC) $(TEST_CC_FLAGS) $(COV_FLAGS) -o tests/test_workspace \
	    $(TEST_WORKSPACE_SRCS) -lm -lpthread
	LLVM_PROFILE_FILE=$(COV_DIR)/workspace.profraw ./tests/test_workspace
	$(CC) $(TEST_CC_FLAGS) $(COV_FLAGS) $(TEST_BLAS_CFLAGS) -o tests/test_late \
	    $(TEST_LATE_SRCS) -lm -lpthread $(TEST_BLAS_LDFLAGS)
	LLVM_PROFILE_FILE=$(COV_DIR)/late.profraw ./tests/test_late
	$(CC) $(TEST_CC_FLAGS) $(COV_FLAGS) $(TEST_BLAS_CFLAGS) -o tests/cli_under_test \
	    $(TEST_CLI_BIN_SRCS) -lm -lpthread $(TEST_BLAS_LDFLAGS)
	$(CC) $(TEST_CC_FLAGS) -o tests/test_cli tests/test_cli.c -lm
	LLVM_PROFILE_FILE=$(COV_DIR)/cli_%p.profraw ./tests/test_cli ./tests/cli_under_test
	$(CC) $(TEST_CC_FLAGS) $(COV_FLAGS) $(TEST_BLAS_CFLAGS) $(CJSON_CFLAGS) -Ideps/ae \
	    -o tests/test_server $(TEST_SERVER_SRCS) \
	    -lm -lpthread $(TEST_BLAS_LDFLAGS) $(CJSON_LDFLAGS)
	LLVM_PROFILE_FILE=$(COV_DIR)/server.profraw ./tests/test_server
	$(LLVM_PROFDATA) merge -sparse $(COV_DIR)/*.profraw -o $(COV_DIR)/tests.profdata
	$(LLVM_COV) report -instr-profile=$(COV_DIR)/tests.profdata $(COV_IGNORE) \
	    $(addprefix -object ,$(COV_BINS))
	$(LLVM_COV) show \
			-format=html \
			-output-dir=$(COV_DIR)/html \
			-instr-profile=$(COV_DIR)/tests.profdata \
			-show-line-counts-or-regions \
			-show-branches=count \
			-show-expansions \
			-show-instantiations \
			$(COV_IGNORE) \
			$(addprefix -object ,$(COV_BINS))
	@echo "HTML report: $(COV_DIR)/html/index.html"

bench-tokenizer:
	$(CC) -Wall -Wextra -O3 -march=native -I. -o bench/bench_tokenizer \
	    bench/bench_tokenizer.c qwen_tokenizer.c $(KERNEL_SRCS) -lm -lpthread
	@echo "run: bench/bench_tokenizer <model_dir> [runs]"

# =============================================================================
# Regression microbenchmarks (see bench/bench.h). Each run records a JSON
# sample set under bench/results/, keyed by commit; compare two records with
# tools/benchstat.py to see if a change helped or hurt.
# =============================================================================
# Dirtiness ignores .gitignore so local-only ignore entries do not mark
# every record as -dirty.
BENCH_STAMP = $(shell git rev-parse --short HEAD 2>/dev/null || echo nogit)$(shell git diff --quiet -- ':!.gitignore' 2>/dev/null || echo -dirty)-$(shell date +%Y%m%d-%H%M%S)
BENCH_CC_FLAGS = -Wall -Wextra $(CFLAGS_BASE) $(TEST_BLAS_CFLAGS) -I.

bench:
	$(CC) $(BENCH_CC_FLAGS) -o bench/bench_kernels \
	    bench/bench_kernels.c $(KERNEL_SRCS) -lm -lpthread $(TEST_BLAS_LDFLAGS)
	@mkdir -p bench/results
	./bench/bench_kernels --json bench/results/kernels-$(BENCH_STAMP).json
	@echo "compare: ./tools/benchstat.py bench/results/OLD.json bench/results/NEW.json"

bench-model:
	@test -n "$(MODEL_DIR)" || { echo "usage: make bench-model MODEL_DIR=/path/to/model-dir"; exit 1; }
	$(CC) $(BENCH_CC_FLAGS) -o bench/bench_model \
	    bench/bench_model.c embed.c qwen_safetensors.c $(KERNEL_SRCS) \
	    -lm -lpthread $(TEST_BLAS_LDFLAGS)
	@mkdir -p bench/results
	./bench/bench_model "$(MODEL_DIR)" --json bench/results/model-$(BENCH_STAMP).json

# =============================================================================
# Debug build
# =============================================================================
SANITIZE ?= address,undefined

debug: CFLAGS  = -Wall -Wextra -g -O0 -DDEBUG -fsanitize=$(SANITIZE)
debug: LDFLAGS += -fsanitize=$(SANITIZE)
debug:
	$(MAKE) clean
	$(MAKE) $(TARGET) CFLAGS="$(CFLAGS)" LDFLAGS="$(LDFLAGS)"

# =============================================================================
# Static library and binary link
# =============================================================================
$(LIB): $(OBJS) $(EXTRA_OBJS)
	ar rcs $@ $^

$(TARGET): $(LIB) embed_cli.o
	$(CC) $(CFLAGS) -o $@ embed_cli.o $(LIB) $(LDFLAGS) $(CJSON_LDFLAGS)

$(SERVER_TARGET): $(LIB) embed_server.o
	$(CC) $(CFLAGS) -o $@ embed_server.o $(LIB) $(LDFLAGS) $(CJSON_LDFLAGS)

# =============================================================================
# Compile rules
# =============================================================================
embed.o: embed.c embed.h qwen_kernels.h qwen_safetensors.h
	$(CC) $(CFLAGS) $(CJSON_CFLAGS) -Ideps/ae -c -o $@ $<

embed_server.o: embed_server.c embed_server.h embed.h embed_mlx.h qwen_tokenizer.h deps/ae/ae.h deps/ae/anet.h
	$(CC) $(CFLAGS) $(CJSON_CFLAGS) -Ideps/ae -c -o $@ $<

embed_mlx.o: embed_mlx.c embed_mlx.h embed.h qwen_safetensors.h
	$(CC) $(CFLAGS) -c -o $@ $<

embed_cli.o: embed_cli.c embed.h qwen_kernels.h qwen_tokenizer.h
	$(CC) $(CFLAGS) $(CJSON_CFLAGS) -Ideps/ae -c -o $@ $<


qwen_kernels.o: qwen_kernels.c qwen_kernels.h qwen_kernels_impl.h
	$(CC) $(CFLAGS) -c -o $@ $<

qwen_kernels_generic.o: qwen_kernels_generic.c qwen_kernels_impl.h
	$(CC) $(CFLAGS) -c -o $@ $<

qwen_kernels_neon.o: qwen_kernels_neon.c qwen_kernels_impl.h
	$(CC) $(CFLAGS) -c -o $@ $<

qwen_kernels_avx.o: qwen_kernels_avx.c qwen_kernels_impl.h
	$(CC) $(CFLAGS) -c -o $@ $<

qwen_tokenizer.o: qwen_tokenizer.c qwen_tokenizer.h qwen_kernels.h
	$(CC) $(CFLAGS) -c -o $@ $<

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

deps/ae/%.o: deps/ae/%.c deps/ae/ae.h deps/ae/anet.h deps/ae/monotonic.h
	$(CC) $(CFLAGS) -Ideps/ae -c -o $@ $<

# =============================================================================
clean:
	rm -f $(OBJS) embed_mlx.o embed_cuda.o embed_cli.o embed_server.o \
	      $(TARGET) $(SERVER_TARGET) $(LIB) libembed.dylib libembed.so \
	      tests/test_kernels_generic tests/test_kernels_blas \
	      tests/test_safetensors tests/test_bf16_model tests/test_server \
	      tests/test_qwen3 \
	      tests/test_tokenizer tests/test_workspace tests/test_cli tests/test_late \
	      tests/cli_under_test bench/bench_tokenizer bench/bench_kernels bench/bench_model
	rm -rf $(COV_DIR)