Skip to content

Distribute and run LLMs with a single file.

Improve build configuration

Justine Tunney555fb801bf45
diff --git a/build/config.mk b/build/config.mk
index dff6a2dd..e786c27c 100644
--- a/build/config.mk
+++ b/build/config.mk
@@ -14,7 +14,7 @@ INSTALL = install
ARFLAGS = rcsD
CXXFLAGS = -frtti -std=gnu++23
-CCFLAGS = -g -ggdb -O3 -fexceptions -fsignaling-nans -ffunction-sections -fdata-sections
+CCFLAGS = -O -fexceptions -fsignaling-nans -ffunction-sections -fdata-sections
CPPFLAGS_ = -iquote. -mcosmo -DGGML_MULTIPLATFORM -Wno-attributes -DLLAMAFILE_DEBUG
TARGET_ARCH = -Xx86_64-mavx -Xx86_64-mtune=znver4
diff --git a/llama.cpp/BUILD.mk b/llama.cpp/BUILD.mk
index 78b05d31..0ef3fafd 100644
--- a/llama.cpp/BUILD.mk
+++ b/llama.cpp/BUILD.mk
@@ -32,20 +32,26 @@ $(LLAMA_CPP_OBJS): private \
-DGGML_MULTIPLATFORM \
-DGGML_USE_LLAMAFILE
-o/$(MODE)/llama.cpp/common.o \
-o/$(MODE)/llama.cpp/llama.o: private \
- CCFLAGS += -O
+o/$(MODE)/llama.cpp/ggml-vector-amd-avx2.o \
+o/$(MODE)/llama.cpp/ggml-vector-amd-avx512bf16.o \
+o/$(MODE)/llama.cpp/ggml-vector-amd-avx512.o \
+o/$(MODE)/llama.cpp/ggml-vector-amd-avx.o \
+o/$(MODE)/llama.cpp/ggml-vector-amd-f16c.o \
+o/$(MODE)/llama.cpp/ggml-vector-amd-fma.o \
+o/$(MODE)/llama.cpp/ggml-vector-arm80.o \
+o/$(MODE)/llama.cpp/ggml-vector-arm82.o: \
+ private CCFLAGS += -O3
o/$(MODE)/llama.cpp/ggml-alloc.o \
o/$(MODE)/llama.cpp/ggml-backend.o \
o/$(MODE)/llama.cpp/grammar-parser.o \
o/$(MODE)/llama.cpp/json-schema-to-grammar.o \
-o/$(MODE)/llama.cpp/llama.o \
+o/$(MODE)/llama.cpp/vector.o \
o/$(MODE)/llama.cpp/unicode.o \
o/$(MODE)/llama.cpp/sampling.o \
o/$(MODE)/llama.cpp/ggml-alloc.o \
-o/$(MODE)/llama.cpp/common.o: private \
- CCFLAGS += -Os
+o/$(MODE)/llama.cpp/common.o: \
+ private CCFLAGS += -Os
o/$(MODE)/llama.cpp/ggml-quants.o: private CXXFLAGS += -Os
o/$(MODE)/llama.cpp/ggml-quants-amd-avx.o: private TARGET_ARCH += -Xx86_64-mtune=sandybridge
diff --git a/llama.cpp/llava/BUILD.mk b/llama.cpp/llava/BUILD.mk
index 155e1f89..42e9b88b 100644
--- a/llama.cpp/llava/BUILD.mk
+++ b/llama.cpp/llava/BUILD.mk
@@ -20,12 +20,6 @@ o/$(MODE)/llama.cpp/llava/llava-quantize: \
o/$(MODE)/llama.cpp/llama.cpp.a \
o/$(MODE)/stb/stb.a
-o/$(MODE)/llama.cpp/llava/clip.o \
-o/$(MODE)/llama.cpp/llava/llava.o \
-o/$(MODE)/llama.cpp/llava/llava-cli.o \
-o/$(MODE)/llama.cpp/llava/llava-quantize.o: private \
- CCFLAGS += -Os
-
$(LLAMA_CPP_LLAVA_OBJS): llama.cpp/llava/BUILD.mk
.PHONY: o/$(MODE)/llama.cpp/llava
diff --git a/llama.cpp/main/BUILD.mk b/llama.cpp/main/BUILD.mk
index aaa94f8d..5985629e 100644
--- a/llama.cpp/main/BUILD.mk
+++ b/llama.cpp/main/BUILD.mk
@@ -18,9 +18,6 @@ o/$(MODE)/llama.cpp/main/main: \
o/$(MODE)/llama.cpp/main/main.1.asc.zip.o \
$(LLAMA_CPP_SERVER_ASSETS:%=o/$(MODE)/%.zip.o)
-o/$(MODE)/llama.cpp/main/main.o: private \
- CCFLAGS += -Os
-
$(LLAMA_CPP_MAIN_OBJS): llama.cpp/main/BUILD.mk
.PHONY: o/$(MODE)/llama.cpp/main
diff --git a/llamafile/BUILD.mk b/llamafile/BUILD.mk
index 895d774f..5a946d45 100644
--- a/llamafile/BUILD.mk
+++ b/llamafile/BUILD.mk
@@ -94,7 +94,6 @@ o/$(MODE)/llamafile: \
# - HWCAP_ASIMDDP +dotprod (e.g. m1, rpi5) __ARM_FEATURE_DOTPROD
#
-o/$(MODE)/llamafile/sgemm.o: private CXXFLAGS += -Os
o/$(MODE)/llamafile/iqk_mul_mat_amd_avx2.o: private TARGET_ARCH += -Xx86_64-mtune=skylake -Xx86_64-mavx2 -Xx86_64-mfma -Xx86_64-mf16c
o/$(MODE)/llamafile/iqk_mul_mat_amd_zen4.o: private TARGET_ARCH += -Xx86_64-mtune=skylake -Xx86_64-mavx2 -Xx86_64-mfma -Xx86_64-mf16c -Xx86_64-mavx512f -Xx86_64-mavx512vl -Xx86_64-mavx512vnni -Xx86_64-mavx512bw -Xx86_64-mavx512dq
o/$(MODE)/llamafile/iqk_mul_mat_arm82.o: private TARGET_ARCH += -Xaarch64-march=armv8.2-a+dotprod+fp16
@@ -113,6 +112,32 @@ o/$(MODE)/llamafile/tinyblas_cpu_mixmul_amd_zen4.o: private TARGET_ARCH += -Xx86
o/$(MODE)/llamafile/tinyblas_cpu_sgemm_arm82.o: private TARGET_ARCH += -Xaarch64-march=armv8.2-a+dotprod+fp16
o/$(MODE)/llamafile/tinyblas_cpu_mixmul_arm82.o: private TARGET_ARCH += -Xaarch64-march=armv8.2-a+dotprod+fp16
+o/$(MODE)/llamafile/sgemm.o: private CXXFLAGS += -Os
+
+o/$(MODE)/llamafile/sgemm_matmul_test.o \
+o/$(MODE)/llamafile/sgemm_sss_test.o \
+o/$(MODE)/llamafile/sgemm_vecdot_test.o \
+o/$(MODE)/llamafile/iqk_mul_mat_amd_avx2.o \
+o/$(MODE)/llamafile/iqk_mul_mat_amd_zen4.o \
+o/$(MODE)/llamafile/iqk_mul_mat_arm82.o \
+o/$(MODE)/llamafile/tinyblas_cpu_mixmul_amd_avx2.o \
+o/$(MODE)/llamafile/tinyblas_cpu_mixmul_amd_avx512f.o \
+o/$(MODE)/llamafile/tinyblas_cpu_mixmul_amd_avx.o \
+o/$(MODE)/llamafile/tinyblas_cpu_mixmul_amd_avxvnni.o \
+o/$(MODE)/llamafile/tinyblas_cpu_mixmul_amd_fma.o \
+o/$(MODE)/llamafile/tinyblas_cpu_mixmul_amd_zen4.o \
+o/$(MODE)/llamafile/tinyblas_cpu_mixmul_arm80.o \
+o/$(MODE)/llamafile/tinyblas_cpu_mixmul_arm82.o \
+o/$(MODE)/llamafile/tinyblas_cpu_sgemm_amd_avx2.o \
+o/$(MODE)/llamafile/tinyblas_cpu_sgemm_amd_avx512f.o \
+o/$(MODE)/llamafile/tinyblas_cpu_sgemm_amd_avx.o \
+o/$(MODE)/llamafile/tinyblas_cpu_sgemm_amd_avxvnni.o \
+o/$(MODE)/llamafile/tinyblas_cpu_sgemm_amd_fma.o \
+o/$(MODE)/llamafile/tinyblas_cpu_sgemm_amd_zen4.o \
+o/$(MODE)/llamafile/tinyblas_cpu_sgemm_arm80.o \
+o/$(MODE)/llamafile/tinyblas_cpu_sgemm_arm82.o: \
+ private CCFLAGS += -O3
+
################################################################################
# testing
diff --git a/llamafile/server/BUILD.mk b/llamafile/server/BUILD.mk
index 9809182b..c763de4f 100644
--- a/llamafile/server/BUILD.mk
+++ b/llamafile/server/BUILD.mk
@@ -21,7 +21,6 @@ o/$(MODE)/llamafile/server/main: \
o/$(MODE)/stb/stb.a \
$(LLAMAFILE_SERVER_OBJS): llamafile/server/BUILD.mk
-$(LLAMAFILE_SERVER_OBJS): private CCFLAGS += -O
o/$(MODE)/llamafile/server/fastjson_test: \
o/$(MODE)/llamafile/server/fastjson_test.o \
diff --git a/stable-diffusion.cpp/BUILD.mk b/stable-diffusion.cpp/BUILD.mk
index e0e4f9a2..e98148c8 100644
--- a/stable-diffusion.cpp/BUILD.mk
+++ b/stable-diffusion.cpp/BUILD.mk
@@ -35,7 +35,6 @@ o/$(MODE)/stable-diffusion.cpp/main: \
o/$(MODE)/stb/stb.a
$(STABLE_DIFFUSION_CPP_OBJS): stable-diffusion.cpp/BUILD.mk
-$(STABLE_DIFFUSION_CPP_OBJS): private CFLAGS += -O
.PHONY: o/$(MODE)/stable-diffusion.cpp
o/$(MODE)/stable-diffusion.cpp: \