llm : prevent from 1-D tensors being GPU split (#3697 )

build : link against build info instead of compiling against it (#3879 )
* cmake : fix build when .git does not exist * cmake : simplify BUILD_INFO target * cmake : add missing dependencies on BUILD_INFO * build : link against build info instead of compiling against it * zig : make build info a .cpp source instead of a header Co-authored-by: Matheus C. França <matheus-catarino@hotmail.com> * cmake : revert change to CMP0115 --------- Co-authored-by: Matheus C. França <matheus-catarino@hotmail.com>
2026-08-01 17:18:03 +02:00 · 2023-11-02 09:54:44 +02:00 · 2023-11-02 08:50:16 +02:00 · 2023-11-02 08:35:10 +02:00 · 2023-11-02 08:33:37 +02:00
40 changed files with 201 additions and 236 deletions
@@ -65,7 +65,7 @@ models-mnt
 /parallel
 /train-text-from-scratch
 /vdot
-build-info.h
+/common/build-info.cpp
 arm_neon.h
 compile_commands.json
 CMakeSettings.json
@@ -100,39 +100,6 @@ option(LLAMA_BUILD_TESTS                "llama: build tests"    ${LLAMA_STANDALO
 option(LLAMA_BUILD_EXAMPLES             "llama: build examples" ${LLAMA_STANDALONE})
 option(LLAMA_BUILD_SERVER               "llama: build server example"                           ON)

-#
-# Build info header
-#
-
-# Generate initial build-info.h
-include(${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake)
-
-if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/.git")
-    set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/.git")
-
-    # Is git submodule
-    if(NOT IS_DIRECTORY "${GIT_DIR}")
-        file(READ ${GIT_DIR} REAL_GIT_DIR_LINK)
-        string(REGEX REPLACE "gitdir: (.*)\n$" "\\1" REAL_GIT_DIR ${REAL_GIT_DIR_LINK})
-        set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/${REAL_GIT_DIR}")
-    endif()
-
-    # Add a custom target for build-info.h
-    add_custom_target(BUILD_INFO ALL DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h")
-
-    # Add a custom command to rebuild build-info.h when .git/index changes
-    add_custom_command(
-        OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h"
-        COMMENT "Generating build details from Git"
-        COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION} -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME} -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.cmake"
-        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
-        DEPENDS "${GIT_DIR}/index"
-        VERBATIM
-    )
-else()
-    message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
-endif()
-
 #
 # Compile flags
 #
@@ -542,9 +542,9 @@ llama.o: llama.cpp ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h l
 	$(CXX) $(CXXFLAGS) -c $< -o $@

 COMMON_H_DEPS = common/common.h common/sampling.h common/log.h
-COMMON_DEPS   = common.o sampling.o grammar-parser.o
+COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o

-common.o: common/common.cpp build-info.h $(COMMON_H_DEPS)
+common.o: common/common.cpp $(COMMON_H_DEPS)
 	$(CXX) $(CXXFLAGS) -c $< -o $@

 sampling.o: common/sampling.cpp $(COMMON_H_DEPS)
@@ -563,46 +563,46 @@ libllama.so: llama.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -shared -fPIC -o $@ $^ $(LDFLAGS)

 clean:
-	rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult build-info.h *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
+	rm -vrf *.o tests/*.o *.so *.dll benchmark-matmult common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)

 #
 # Examples
 #

-main: examples/main/main.cpp                                  build-info.h ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
+main: examples/main/main.cpp                                  ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 	@echo
 	@echo '====  Run ./main -h for help.  ===='
 	@echo

-infill: examples/infill/infill.cpp                            build-info.h ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
+infill: examples/infill/infill.cpp                            ggml.o llama.o $(COMMON_DEPS) console.o grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-simple: examples/simple/simple.cpp                            build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+simple: examples/simple/simple.cpp                            ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-batched: examples/batched/batched.cpp                         build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+batched: examples/batched/batched.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-batched-bench: examples/batched-bench/batched-bench.cpp       build-info.h ggml.o llama.o common.o $(OBJS)
+batched-bench: examples/batched-bench/batched-bench.cpp       build-info.o ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-quantize: examples/quantize/quantize.cpp                      build-info.h ggml.o llama.o $(OBJS)
+quantize: examples/quantize/quantize.cpp                      build-info.o ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.h ggml.o llama.o $(OBJS)
+quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.o ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-perplexity: examples/perplexity/perplexity.cpp                build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+perplexity: examples/perplexity/perplexity.cpp                ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-embedding: examples/embedding/embedding.cpp                   build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-save-load-state: examples/save-load-state/save-load-state.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h,$(filter-out %.hpp,$^)) -o $@ $(LDFLAGS) $(LWINSOCK2) -Wno-cast-qual

 gguf: examples/gguf/gguf.cpp ggml.o llama.o $(OBJS)
@@ -614,7 +614,7 @@ train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratc
 convert-llama2c-to-ggml: examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp ggml.o llama.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-llama-bench: examples/llama-bench/llama-bench.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+llama-bench: examples/llama-bench/llama-bench.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

 llava: examples/llava/llava.cpp examples/llava/llava-utils.h examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
@@ -623,19 +623,19 @@ llava: examples/llava/llava.cpp examples/llava/llava-utils.h examples/llava/clip
 baby-llama: examples/baby-llama/baby-llama.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-beam-search: examples/beam-search/beam-search.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+beam-search: examples/beam-search/beam-search.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-finetune: examples/finetune/finetune.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
+finetune: examples/finetune/finetune.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-export-lora: examples/export-lora/export-lora.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+export-lora: examples/export-lora/export-lora.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-speculative: examples/speculative/speculative.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+speculative: examples/speculative/speculative.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-parallel: examples/parallel/parallel.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+parallel: examples/parallel/parallel.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

 ifdef LLAMA_METAL
@@ -648,7 +648,7 @@ swift: examples/batched.swift
 	(cd examples/batched.swift; make build)
 endif

-build-info.h: $(wildcard .git/index) scripts/build-info.sh
+common/build-info.cpp: $(wildcard .git/index) scripts/build-info.sh
 	@sh scripts/build-info.sh $(CC) > $@.tmp
 	@if ! cmp -s $@.tmp $@; then \
 		mv $@.tmp $@; \
@@ -656,13 +656,16 @@ build-info.h: $(wildcard .git/index) scripts/build-info.sh
 		rm $@.tmp; \
 	fi

+build-info.o: common/build-info.cpp
+	$(CXX) $(CXXFLAGS) -c $(filter-out %.h,$^) -o $@
+
 #
 # Tests
 #

 tests: $(TEST_TARGETS)

-benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.h ggml.o $(OBJS)
+benchmark-matmult: examples/benchmark/benchmark-matmult.cpp build-info.o ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

 run-benchmark-matmult: benchmark-matmult
@@ -676,40 +679,40 @@ vdot: pocs/vdot/vdot.cpp ggml.o $(OBJS)
 q8dot: pocs/vdot/q8dot.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)

-tests/test-llama-grammar: tests/test-llama-grammar.cpp build-info.h ggml.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+tests/test-llama-grammar: tests/test-llama-grammar.cpp ggml.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-tests/test-grammar-parser: tests/test-grammar-parser.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+tests/test-grammar-parser: tests/test-grammar-parser.cpp ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-tests/test-double-float: tests/test-double-float.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-double-float: tests/test-double-float.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-tests/test-grad0: tests/test-grad0.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-grad0: tests/test-grad0.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-tests/test-opt: tests/test-opt.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-opt: tests/test-opt.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-tests/test-quantize-fns: tests/test-quantize-fns.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-quantize-fns: tests/test-quantize-fns.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-tests/test-quantize-perf: tests/test-quantize-perf.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-quantize-perf: tests/test-quantize-perf.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-tests/test-sampling: tests/test-sampling.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-sampling: tests/test-sampling.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-0-falcon: tests/test-tokenizer-0-falcon.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-0-llama: tests/test-tokenizer-0-llama.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-1-bpe: tests/test-tokenizer-1-bpe.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

-tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp build-info.h ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+tests/test-tokenizer-1-llama: tests/test-tokenizer-1-llama.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)

 tests/test-c.o: tests/test-c.c llama.h
@@ -10,7 +10,6 @@ const Maker = struct {
    builder: *std.build.Builder,
    target: CrossTarget,
    optimize: Mode,
-    config_header: *ConfigHeader,
    enable_lto: bool,

    include_dirs: ArrayList([]const u8),
@@ -41,26 +40,24 @@ const Maker = struct {
        const commit_hash = try std.ChildProcess.exec(
            .{ .allocator = builder.allocator, .argv = &.{ "git", "rev-parse", "HEAD" } },
        );
-        const config_header = builder.addConfigHeader(
-            .{ .style = .blank, .include_path = "build-info.h" },
-            .{
-                .BUILD_NUMBER = 0,
-                .BUILD_COMMIT = commit_hash.stdout[0 .. commit_hash.stdout.len - 1], // omit newline
-                .BUILD_COMPILER = builder.fmt("Zig {s}", .{zig_version}),
-                .BUILD_TARGET = try target.allocDescription(builder.allocator),
-            },
-        );
+        try std.fs.cwd().writeFile("common/build-info.cpp", builder.fmt(
+            \\int LLAMA_BUILD_NUMBER = {};
+            \\char const *LLAMA_COMMIT = "{s}";
+            \\char const *LLAMA_COMPILER = "Zig {s}";
+            \\char const *LLAMA_BUILD_TARGET = "{s}";
+            \\
+        , .{ 0, commit_hash.stdout[0 .. commit_hash.stdout.len - 1], zig_version, try target.allocDescription(builder.allocator) }));
        var m = Maker{
            .builder = builder,
            .target = target,
            .optimize = builder.standardOptimizeOption(.{}),
-            .config_header = config_header,
            .enable_lto = false,
            .include_dirs = ArrayList([]const u8).init(builder.allocator),
            .cflags = ArrayList([]const u8).init(builder.allocator),
            .cxxflags = ArrayList([]const u8).init(builder.allocator),
            .objs = ArrayList(*Compile).init(builder.allocator),
        };
+
        try m.addCFlag("-std=c11");
        try m.addCxxFlag("-std=c++11");
        try m.addProjectInclude(&.{});
@@ -72,7 +69,7 @@ const Maker = struct {
        const o = m.builder.addObject(.{ .name = name, .target = m.target, .optimize = m.optimize });
        if (o.target.getAbi() != .msvc)
            o.defineCMacro("_GNU_SOURCE", null);
-        o.addConfigHeader(m.config_header);
+
        if (std.mem.endsWith(u8, src, ".c")) {
            o.addCSourceFiles(&.{src}, m.cflags.items);
            o.linkLibC();
@@ -85,7 +82,6 @@ const Maker = struct {
                o.linkLibCpp();
            }
        }
-        o.addConfigHeader(m.config_header);
        for (m.include_dirs.items) |i| o.addIncludePath(.{ .path = i });
        o.want_lto = m.enable_lto;
        return o;
@@ -105,7 +101,6 @@ const Maker = struct {
            // linkLibCpp already add (libc++ + libunwind + libc)
            e.linkLibCpp();
        }
-        e.addConfigHeader(m.config_header);
        m.builder.installArtifact(e);
        e.want_lto = m.enable_lto;
        return e;
@@ -121,6 +116,7 @@ pub fn build(b: *std.build.Builder) !void {
    const ggml_backend = make.obj("ggml-backend", "ggml-backend.c");
    const ggml_quants = make.obj("ggml-quants", "ggml-quants.c");
    const llama = make.obj("llama", "llama.cpp");
+    const buildinfo = make.obj("common", "common/build-info.cpp");
    const common = make.obj("common", "common/common.cpp");
    const console = make.obj("console", "common/console.cpp");
    const sampling = make.obj("sampling", "common/sampling.cpp");
@@ -128,14 +124,14 @@ pub fn build(b: *std.build.Builder) !void {
    const train = make.obj("train", "common/train.cpp");
    const clip = make.obj("clip", "examples/llava/clip.cpp");

-    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, sampling, console, grammar_parser });
-    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common });
-    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common });
-    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common });
-    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, train });
-    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, train });
+    _ = make.exe("main", "examples/main/main.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, console, grammar_parser });
+    _ = make.exe("quantize", "examples/quantize/quantize.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
+    _ = make.exe("perplexity", "examples/perplexity/perplexity.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
+    _ = make.exe("embedding", "examples/embedding/embedding.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo });
+    _ = make.exe("finetune", "examples/finetune/finetune.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });
+    _ = make.exe("train-text-from-scratch", "examples/train-text-from-scratch/train-text-from-scratch.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, train });

-    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, sampling, grammar_parser, clip });
+    const server = make.exe("server", "examples/server/server.cpp", &.{ ggml, ggml_alloc, ggml_backend, ggml_quants, llama, common, buildinfo, sampling, grammar_parser, clip });
    if (server.target.isWindows()) {
        server.linkSystemLibrary("ws2_32");
    }
@@ -1,8 +1,46 @@
 # common

+
+# Build info header
+#
+
+if(EXISTS "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
+    set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/../.git")
+
+    # Is git submodule
+    if(NOT IS_DIRECTORY "${GIT_DIR}")
+        file(READ ${GIT_DIR} REAL_GIT_DIR_LINK)
+        string(REGEX REPLACE "gitdir: (.*)\n$" "\\1" REAL_GIT_DIR ${REAL_GIT_DIR_LINK})
+        set(GIT_DIR "${CMAKE_CURRENT_SOURCE_DIR}/${REAL_GIT_DIR}")
+    endif()
+
+    set(GIT_INDEX "${GIT_DIR}/index")
+else()
+    message(WARNING "Git repository not found; to enable automatic generation of build info, make sure Git is installed and the project is a Git repository.")
+    set(GIT_INDEX "")
+endif()
+
+# Add a custom command to rebuild build-info.cpp when .git/index changes
+add_custom_command(
+    OUTPUT "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp"
+    COMMENT "Generating build details from Git"
+    COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
+            -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
+            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/../scripts/build-info.cmake"
+    WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
+    DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
+    VERBATIM
+)
+set(TARGET build_info)
+add_library(${TARGET} OBJECT build-info.cpp)
+if (BUILD_SHARED_LIBS)
+    set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
+endif()
+
+
 set(TARGET common)

-add_library(${TARGET} OBJECT
+add_library(${TARGET} STATIC
    common.h
    common.cpp
    sampling.h
@@ -21,4 +59,4 @@ endif()

 target_include_directories(${TARGET} PUBLIC .)
 target_compile_features(${TARGET} PUBLIC cxx_std_11)
-target_link_libraries(${TARGET} PRIVATE llama)
+target_link_libraries(${TARGET} PRIVATE llama build_info)
@@ -0,0 +1,4 @@
+int LLAMA_BUILD_NUMBER = @BUILD_NUMBER@;
+char const *LLAMA_COMMIT = "@BUILD_COMMIT@";
+char const *LLAMA_COMPILER = "@BUILD_COMPILER@";
+char const *LLAMA_BUILD_TARGET = "@BUILD_TARGET@";
@@ -1,5 +1,4 @@
 #include "common.h"
-#include "build-info.h"
 #include "llama.h"

 #include <algorithm>
@@ -1199,8 +1198,8 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
                               const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
    const llama_sampling_params & sparams = params.sparams;

-    fprintf(stream, "build_commit: %s\n", BUILD_COMMIT);
-    fprintf(stream, "build_number: %d\n", BUILD_NUMBER);
+    fprintf(stream, "build_commit: %s\n",        LLAMA_COMMIT);
+    fprintf(stream, "build_number: %d\n",        LLAMA_BUILD_NUMBER);
    fprintf(stream, "cpu_has_arm_fma: %s\n",     ggml_cpu_has_arm_fma()     ? "true" : "false");
    fprintf(stream, "cpu_has_avx: %s\n",         ggml_cpu_has_avx()         ? "true" : "false");
    fprintf(stream, "cpu_has_avx2: %s\n",        ggml_cpu_has_avx2()        ? "true" : "false");
@@ -26,11 +26,17 @@
 #define die(msg)          do { fputs("error: " msg "\n", stderr);                exit(1); } while (0)
 #define die_fmt(fmt, ...) do { fprintf(stderr, "error: " fmt "\n", __VA_ARGS__); exit(1); } while (0)

-#define print_build_info() do {                                                             \
-    fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);         \
-    fprintf(stderr, "%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);  \
+#define print_build_info() do {                                                                     \
+    fprintf(stderr, "%s: build = %d (%s)\n", __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);           \
+    fprintf(stderr, "%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);    \
 } while(0)

+// build info
+extern int LLAMA_BUILD_NUMBER;
+extern char const *LLAMA_COMMIT;
+extern char const *LLAMA_COMPILER;
+extern char const *LLAMA_BUILD_TARGET;
+
 //
 // CLI argument parsing
 //
@@ -1,9 +1,6 @@
 set(TARGET benchmark)
 add_executable(${TARGET} benchmark-matmult.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
 target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()
@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "ggml.h"

@@ -3,6 +3,3 @@ add_executable(${TARGET} embedding.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()
@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "llama.h"

@@ -3,6 +3,3 @@ add_executable(${TARGET} infill.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-    add_dependencies(${TARGET} BUILD_INFO)
-endif()
@@ -2,7 +2,6 @@

 #include "console.h"
 #include "llama.h"
-#include "build-info.h"
 #include "grammar-parser.h"

 #include <cassert>
@@ -184,8 +183,8 @@ int main(int argc, char ** argv) {
        LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
    }

-    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
-    LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);
+    LOG_TEE("%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
+    LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);

    if (params.seed == LLAMA_DEFAULT_SEED) {
        params.seed = time(NULL);
@@ -3,6 +3,3 @@ add_executable(${TARGET} llama-bench.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()
@@ -19,7 +19,6 @@
 #include "ggml.h"
 #include "llama.h"
 #include "common.h"
-#include "build-info.h"
 #include "ggml-cuda.h"

 // utils
@@ -641,8 +640,8 @@ struct test {
    }
 };

-const std::string test::build_commit = BUILD_COMMIT;
-const int         test::build_number = BUILD_NUMBER;
+const std::string test::build_commit = LLAMA_COMMIT;
+const int         test::build_number = LLAMA_BUILD_NUMBER;
 const bool        test::cuda         = !!ggml_cpu_has_cublas();
 const bool        test::opencl       = !!ggml_cpu_has_clblast();
 const bool        test::metal        = !!ggml_cpu_has_metal();
@@ -5,9 +5,6 @@ target_link_libraries(${TARGET} PRIVATE common ggml ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
 if (NOT MSVC)
    target_compile_options(${TARGET} PRIVATE -Wno-cast-qual) # stb_image.h
-    endif()
-if(TARGET BUILD_INFO)
-    add_dependencies(${TARGET} BUILD_INFO)
 endif()

 set(TARGET llava)
@@ -15,6 +12,3 @@ add_executable(${TARGET} llava.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama clip ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-    add_dependencies(${TARGET} BUILD_INFO)
-endif()
@@ -3,6 +3,3 @@ add_executable(${TARGET} main.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()
@@ -2,7 +2,6 @@

 #include "console.h"
 #include "llama.h"
-#include "build-info.h"

 #include <cassert>
 #include <cinttypes>
@@ -153,8 +152,8 @@ int main(int argc, char ** argv) {
        LOG_TEE("%s: warning: scaling RoPE frequency by %g.\n", __func__, params.rope_freq_scale);
    }

-    LOG_TEE("%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
-    LOG_TEE("%s: built with %s for %s\n", __func__, BUILD_COMPILER, BUILD_TARGET);
+    LOG_TEE("%s: build = %d (%s)\n",      __func__, LLAMA_BUILD_NUMBER, LLAMA_COMMIT);
+    LOG_TEE("%s: built with %s for %s\n", __func__, LLAMA_COMPILER, LLAMA_BUILD_TARGET);

    if (params.seed == LLAMA_DEFAULT_SEED) {
        params.seed = time(NULL);
@@ -3,6 +3,3 @@ add_executable(${TARGET} parallel.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()
@@ -1,8 +1,6 @@
 // A basic application simulating a server with multiple clients.
 // The clients submite requests to the server and they are processed in parallel.

-#include "build-info.h"
-
 #include "common.h"
 #include "llama.h"

@@ -3,6 +3,3 @@ add_executable(${TARGET} perplexity.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()
@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "llama.h"

@@ -1,6 +1,6 @@
 set(TARGET quantize-stats)
 add_executable(${TARGET} quantize-stats.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
 target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
@@ -1,5 +1,4 @@
 #define LLAMA_API_INTERNAL
-#include "build-info.h"
 #include "common.h"
 #include "ggml.h"
 #include "llama.h"
@@ -1,9 +1,6 @@
 set(TARGET quantize)
 add_executable(${TARGET} quantize.cpp)
 install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_link_libraries(${TARGET} PRIVATE llama build_info ${CMAKE_THREAD_LIBS_INIT})
 target_include_directories(${TARGET} PRIVATE ../../common)
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()
@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "llama.h"

@@ -3,6 +3,3 @@ add_executable(${TARGET} save-load-state.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()
@@ -1,4 +1,3 @@
-#include "build-info.h"
 #include "common.h"
 #include "llama.h"

@@ -11,6 +11,3 @@ if (WIN32)
    TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
 endif()
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()
@@ -1,6 +1,5 @@
 #include "common.h"
 #include "llama.h"
-#include "build-info.h"
 #include "grammar-parser.h"

 #include "../llava/clip.h"
@@ -2264,8 +2263,8 @@ int main(int argc, char **argv)

    llama_backend_init(params.numa);

-    LOG_INFO("build info", {{"build", BUILD_NUMBER},
-                            {"commit", BUILD_COMMIT}});
+    LOG_INFO("build info", {{"build", LLAMA_BUILD_NUMBER},
+                            {"commit", LLAMA_COMMIT}});

    LOG_INFO("system info", {
                                {"n_threads", params.n_threads},
@@ -3,6 +3,3 @@ add_executable(${TARGET} speculative.cpp)
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
-if(TARGET BUILD_INFO)
-  add_dependencies(${TARGET} BUILD_INFO)
-endif()
@@ -1,5 +1,3 @@
-#include "build-info.h"
-
 #include "common.h"
 #include "llama.h"

@@ -7420,7 +7420,7 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
    } else if (all_on_device && !use_tensor_cores && src0->type == GGML_TYPE_F16 && !ggml_is_contiguous(src0) && !ggml_is_transposed(src1) && src1->ne[1] == 1) {
        // KQV single-batch
        ggml_cuda_mul_mat_vec_nc(src0, src1, dst);
-    } else if (all_on_device && src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) {
+    } else if (all_on_device && use_tensor_cores && src0->type == GGML_TYPE_F16 && src1->type == GGML_TYPE_F32 && !ggml_is_transposed(src0) && !ggml_is_transposed(src1)) {
        // KQ + KQV multi-batch
        ggml_cuda_mul_mat_mat_batched_cublas(src0, src1, dst);
    } else if (src0->type == GGML_TYPE_F32) {
@@ -1419,34 +1419,35 @@ void ggml_metal_graph_compute(
                                default: GGML_ASSERT(false);
                            };

-                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
-                            [encoder setBuffer:id_src1 offset:offs_src1        atIndex:1];
-                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:2];
-                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:3];
-                            [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:4];
-                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:5];
-                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:6];
-                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:7];
-                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:8];
-                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:9];
-                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:10];
-                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:11];
-                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:12];
-                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:13];
-                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:14];
-                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:15];
-                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:16];
-                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:17];
-                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:18];
-                            [encoder setBytes:&n_past  length:sizeof(     int) atIndex:19];
-                            [encoder setBytes:&n_dims  length:sizeof(     int) atIndex:20];
-                            [encoder setBytes:&mode    length:sizeof(     int) atIndex:21];
-                            [encoder setBytes:&freq_base  length:sizeof(float) atIndex:22];
-                            [encoder setBytes:&freq_scale length:sizeof(float) atIndex:23];
-                            [encoder setBytes:&ext_factor  length:sizeof(float) atIndex:24];
-                            [encoder setBytes:&attn_factor length:sizeof(float) atIndex:25];
-                            [encoder setBytes:&beta_fast   length:sizeof(float) atIndex:26];
-                            [encoder setBytes:&beta_slow   length:sizeof(float) atIndex:27];
+                            [encoder setBuffer:id_src0     offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_src1     offset:offs_src1        atIndex:1];
+                            [encoder setBuffer:id_dst      offset:offs_dst         atIndex:2];
+                            [encoder setBytes:&ne00        length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne01        length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne02        length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&ne03        length:sizeof( int64_t) atIndex:6];
+                            [encoder setBytes:&nb00        length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb01        length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb02        length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&nb03        length:sizeof(uint64_t) atIndex:10];
+                            [encoder setBytes:&ne0         length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne1         length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne2         length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&ne3         length:sizeof( int64_t) atIndex:14];
+                            [encoder setBytes:&nb0         length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb1         length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb2         length:sizeof(uint64_t) atIndex:17];
+                            [encoder setBytes:&nb3         length:sizeof(uint64_t) atIndex:18];
+                            [encoder setBytes:&n_past      length:sizeof(     int) atIndex:19];
+                            [encoder setBytes:&n_dims      length:sizeof(     int) atIndex:20];
+                            [encoder setBytes:&mode        length:sizeof(     int) atIndex:21];
+                            [encoder setBytes:&n_orig_ctx  length:sizeof(     int) atIndex:22];
+                            [encoder setBytes:&freq_base   length:sizeof(   float) atIndex:23];
+                            [encoder setBytes:&freq_scale  length:sizeof(   float) atIndex:24];
+                            [encoder setBytes:&ext_factor  length:sizeof(   float) atIndex:25];
+                            [encoder setBytes:&attn_factor length:sizeof(   float) atIndex:26];
+                            [encoder setBytes:&beta_fast   length:sizeof(   float) atIndex:27];
+                            [encoder setBytes:&beta_slow   length:sizeof(   float) atIndex:28];

                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
                        } break;
@@ -1070,20 +1070,20 @@ static float rope_yarn_ramp(const float low, const float high, const int i0) {
 // MIT licensed. Copyright (c) 2023 Jeffrey Quesnelle and Bowen Peng.
 static void rope_yarn(
    float theta_extrap, float freq_scale, float corr_dims[2], int64_t i0, float ext_factor, float mscale,
-    float * cos_theta, float * sin_theta
+    thread float * cos_theta, thread float * sin_theta
 ) {
    // Get n-d rotational scaling corrected for extrapolation
    float theta_interp = freq_scale * theta_extrap;
    float theta = theta_interp;
    if (ext_factor != 0.0f) {
-        ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
+        float ramp_mix = rope_yarn_ramp(corr_dims[0], corr_dims[1], i0) * ext_factor;
        theta = theta_interp * (1 - ramp_mix) + theta_extrap * ramp_mix;

        // Get n-d magnitude scaling corrected for interpolation
-        mscale *= 1.0f + 0.1f * logf(1.0f / freq_scale);
+        mscale *= 1.0f + 0.1f * log(1.0f / freq_scale);
    }
-    *cos_theta = cosf(theta) * mscale;
-    *sin_theta = sinf(theta) * mscale;
+    *cos_theta = cos(theta) * mscale;
+    *sin_theta = sin(theta) * mscale;
 }

 // Apparently solving `n_rot = 2pi * x * base^((2 * max_pos_emb) / n_dims)` for x, we get
@@ -1123,8 +1123,13 @@ typedef void (rope_t)(
        constant         int & n_past,
        constant         int & n_dims,
        constant         int & mode,
+        constant         int & n_orig_ctx,
        constant       float & freq_base,
        constant       float & freq_scale,
+        constant       float & ext_factor,
+        constant       float & attn_factor,
+        constant       float & beta_fast,
+        constant       float & beta_slow,
        uint  tiitg[[thread_index_in_threadgroup]],
        uint3 tptg[[threads_per_threadgroup]],
        uint3 tgpig[[threadgroup_position_in_grid]]);
@@ -1153,6 +1158,7 @@ kernel void kernel_rope(
        constant         int & n_past,
        constant         int & n_dims,
        constant         int & mode,
+        constant         int & n_orig_ctx,
        constant       float & freq_base,
        constant       float & freq_scale,
        constant       float & ext_factor,
@@ -1837,6 +1837,12 @@ struct llama_model_loader {
            throw std::runtime_error(format("%s: tensor '%s' not found", __func__, name.c_str()));
        }

+        if (backend == GGML_BACKEND_GPU_SPLIT) {
+            if (ne.size() == 1) {
+                throw std::runtime_error(format("%s: 1-dimensional tensor '%s' cannot be split on the GPU", __func__, name.c_str()));
+            }
+        }
+
        {
            bool is_ok = true;
            for (size_t i = 0; i < ne.size(); ++i) {
@@ -2817,8 +2823,8 @@ static void llm_load_tensors(
                        layer.ffn_down   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split);
                        layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd},       backend);

-                        layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
-                        layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff},           backend);
+                        layer.ffn_up   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, backend_split);
+                        layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "bias", i),           {n_ff}, backend);

                        if (backend == GGML_BACKEND_GPU) {
                            vram_weights +=
@@ -2877,13 +2883,13 @@ static void llm_load_tensors(
                        layer.attn_norm     = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM,   "weight", i), {n_embd}, backend);
                        layer.attn_norm_b   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM,   "bias",   i), {n_embd}, backend);
                        layer.wqkv          = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV,    "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split);
-                        layer.bqkv          = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV,    "bias",   i), {n_embd + 2*n_embd_gqa},         backend_split);
+                        layer.bqkv          = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV,    "bias",   i), {n_embd + 2*n_embd_gqa},         backend);
                        layer.wo            = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT,    "weight", i), {n_embd, n_embd},   backend_split);
-                        layer.bo            = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT,    "bias",   i), {n_embd},           backend_split);
+                        layer.bo            = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT,    "bias",   i), {n_embd},           backend);
                        layer.ffn_down      = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN,    "weight", i), {n_ff, n_embd}, backend_split);
-                        layer.ffn_down_b    = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN,    "bias",   i), {n_embd},       backend_split);
+                        layer.ffn_down_b    = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN,    "bias",   i), {n_embd},       backend);
                        layer.ffn_up        = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,      "weight", i), {n_embd,   n_ff}, backend_split);
-                        layer.ffn_up_b      = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,      "bias",   i), {n_ff},           backend_split);
+                        layer.ffn_up_b      = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,      "bias",   i), {n_ff},           backend);
                        layer.ffn_norm      = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM,    "weight", i), {n_embd}, backend);
                        layer.ffn_norm_b    = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM,    "bias",   i), {n_embd}, backend);
                        layer.attn_q_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {64}, backend);
@@ -2949,19 +2955,19 @@ static void llm_load_tensors(
                        layer.attn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_NORM,   "bias", i),   {n_embd}, backend);

                        layer.wqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, backend_split);
-                        layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa},         backend_split);
+                        layer.bqkv = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_QKV, "bias", i),   {n_embd + 2*n_embd_gqa},         backend);

                        layer.wo   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd},                backend_split);
-                        layer.bo   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd},                        backend_split);
+                        layer.bo   = ml.create_tensor(ctx, tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd},                        backend);

                        layer.ffn_norm   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, backend);
                        layer.ffn_norm_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, backend);

                        layer.ffn_down   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, backend_split);
-                        layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd},       backend_split);
+                        layer.ffn_down_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd},       backend);

-                        layer.ffn_up = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
-                        layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff},           backend_split);
+                        layer.ffn_up   = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, backend_split);
+                        layer.ffn_up_b = ml.create_tensor(ctx, tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff},           backend);

                        if (backend == GGML_BACKEND_GPU) {
                            vram_weights +=
@@ -1,5 +1,5 @@
-set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/scripts/build-info.h.in")
-set(HEADER_FILE "${CMAKE_CURRENT_SOURCE_DIR}/build-info.h")
+set(TEMPLATE_FILE "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp.in")
+set(OUTPUT_FILE "${CMAKE_CURRENT_SOURCE_DIR}/common/build-info.cpp")
 set(BUILD_NUMBER 0)
 set(BUILD_COMMIT "unknown")
 set(BUILD_COMPILER "unknown")
@@ -24,15 +24,21 @@ if(Git_FOUND)
        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
        OUTPUT_VARIABLE HEAD
        OUTPUT_STRIP_TRAILING_WHITESPACE
+        RESULT_VARIABLE RES
    )
+    if (RES EQUAL 0)
+        set(BUILD_COMMIT ${HEAD})
+    endif()
    execute_process(
        COMMAND ${GIT_EXECUTABLE} rev-list --count HEAD
        WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}
        OUTPUT_VARIABLE COUNT
        OUTPUT_STRIP_TRAILING_WHITESPACE
+        RESULT_VARIABLE RES
    )
-    set(BUILD_COMMIT ${HEAD})
-    set(BUILD_NUMBER ${COUNT})
+    if (RES EQUAL 0)
+        set(BUILD_NUMBER ${COUNT})
+    endif()
 endif()

 if(MSVC)
@@ -53,22 +59,22 @@ else()
    set(BUILD_TARGET ${OUT})
 endif()

-# Only write the header if it's changed to prevent unnecessary recompilation
-if(EXISTS ${HEADER_FILE})
-    file(READ ${HEADER_FILE} CONTENTS)
-    string(REGEX MATCH "BUILD_COMMIT \"([^\"]*)\"" _ ${CONTENTS})
+# Only write the build info if it changed
+if(EXISTS ${OUTPUT_FILE})
+    file(READ ${OUTPUT_FILE} CONTENTS)
+    string(REGEX MATCH "LLAMA_COMMIT = \"([^\"]*)\";" _ ${CONTENTS})
    set(OLD_COMMIT ${CMAKE_MATCH_1})
-    string(REGEX MATCH "BUILD_COMPILER \"([^\"]*)\"" _ ${CONTENTS})
+    string(REGEX MATCH "LLAMA_COMPILER = \"([^\"]*)\";" _ ${CONTENTS})
    set(OLD_COMPILER ${CMAKE_MATCH_1})
-    string(REGEX MATCH "BUILD_TARGET \"([^\"]*)\"" _ ${CONTENTS})
+    string(REGEX MATCH "LLAMA_BUILD_TARGET = \"([^\"]*)\";" _ ${CONTENTS})
    set(OLD_TARGET ${CMAKE_MATCH_1})
    if (
        NOT OLD_COMMIT   STREQUAL BUILD_COMMIT   OR
        NOT OLD_COMPILER STREQUAL BUILD_COMPILER OR
        NOT OLD_TARGET   STREQUAL BUILD_TARGET
    )
-        configure_file(${TEMPLATE_FILE} ${HEADER_FILE})
+        configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
    endif()
 else()
-    configure_file(${TEMPLATE_FILE} ${HEADER_FILE})
+    configure_file(${TEMPLATE_FILE} ${OUTPUT_FILE})
 endif()
@@ -1,9 +0,0 @@
-#ifndef BUILD_INFO_H
-#define BUILD_INFO_H
-
-#define BUILD_NUMBER @BUILD_NUMBER@
-#define BUILD_COMMIT "@BUILD_COMMIT@"
-#define BUILD_COMPILER "@BUILD_COMPILER@"
-#define BUILD_TARGET "@BUILD_TARGET@"
-
-#endif // BUILD_INFO_H
@@ -24,12 +24,7 @@ if out=$($CC -dumpmachine); then
  build_target=$out
 fi

-echo "#ifndef BUILD_INFO_H"
-echo "#define BUILD_INFO_H"
-echo
-echo "#define BUILD_NUMBER $build_number"
-echo "#define BUILD_COMMIT \"$build_commit\""
-echo "#define BUILD_COMPILER \"$build_compiler\""
-echo "#define BUILD_TARGET \"$build_target\""
-echo
-echo "#endif // BUILD_INFO_H"
+echo "int LLAMA_BUILD_NUMBER = ${build_number};"
+echo "char const *LLAMA_COMMIT = \"${build_commit}\";"
+echo "char const *LLAMA_COMPILER = \"${build_compiler}\";"
+echo "char const *LLAMA_BUILD_TARGET = \"${build_target}\";"
Author	SHA1	Message	Date
Georgi Gerganov	1efae9b7dc	llm : prevent from 1-D tensors being GPU split (#3697 )	2023-11-02 09:54:44 +02:00
cebtenzzre GitHub Matheus C. França	b12fa0d1c1	build : link against build info instead of compiling against it (#3879 ) * cmake : fix build when .git does not exist * cmake : simplify BUILD_INFO target * cmake : add missing dependencies on BUILD_INFO * build : link against build info instead of compiling against it * zig : make build info a .cpp source instead of a header Co-authored-by: Matheus C. França <matheus-catarino@hotmail.com> * cmake : revert change to CMP0115 --------- Co-authored-by: Matheus C. França <matheus-catarino@hotmail.com>	2023-11-02 08:50:16 +02:00
Georgi GerganovandGitHub	4d719a6d4e	cuda : check if this fixes Pascal card regression (#3882 )	2023-11-02 08:35:10 +02:00
Georgi GerganovandGitHub	183b3fac6c	metal : fix build errors and kernel sig after #2268 (#3898 )	2023-11-02 08:33:37 +02:00