ggml : disable fast-math for Metal (cmake build only)

Author: ggerganov

CMakeLists.txt

@@ -95,6 +95,7 @@ option(LLAMA_HIP_UMA                         "llama: use HIP unified memory arch
 option(LLAMA_CLBLAST                         "llama: use CLBlast"                               OFF)
 option(LLAMA_METAL                           "llama: use Metal"                                 ${LLAMA_METAL_DEFAULT})
 option(LLAMA_METAL_NDEBUG                    "llama: disable Metal debugging"                   OFF)
+option(LLAMA_METAL_NO_FAST_MATH              "llama: compile Metal with -fno-fast-math"         OFF)
 option(LLAMA_MPI                             "llama: use MPI"                                   OFF)
 option(LLAMA_QKK_64                          "llama: use super-block size of 64 for k-quants"   OFF)
 
@@ -173,6 +174,27 @@ if (LLAMA_METAL)
     # copy ggml-metal.metal to bin directory
     configure_file(ggml-metal.metal ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal COPYONLY)
 
+    if (LLAMA_METAL_NO_FAST_MATH)
+        # custom command to do the following:
+        #   xcrun -sdk macosx metal    -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
+        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o ggml.metallib
+        #
+        # note: this is the only way I found to disable fast-math in Metal. it's ugly, but at least it works
+        #       disabling fast math is needed in order to pass tests/test-backend-ops
+        add_custom_command(
+            OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml.metallib
+            COMMAND xcrun -sdk macosx metal -fno-fast-math -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.metal -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air
+            COMMAND xcrun -sdk macosx metallib ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml.metallib
+            DEPENDS ggml-metal.metal
+            COMMENT "Compiling Metal kernels"
+        )
+
+        add_custom_target(
+            ggml-metal ALL
+            DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml.metallib
+        )
+    endif()
+
     set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS}
         ${FOUNDATION_LIBRARY}
         ${METAL_FRAMEWORK}

ci/run.sh

@@ -30,6 +30,12 @@ sd=`dirname $0`
 cd $sd/../
 SRC=`pwd`
 
+CMAKE_EXTRA=""
+
+if [ ! -z ${GG_BUILD_METAL} ]; then
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DLLAMA_METAL_NO_FAST_MATH=ON"
+fi
+
 ## helpers
 
 # download a file if it does not exist or if it is outdated
@@ -81,8 +87,8 @@ function gg_run_ctest_debug {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Debug ..     ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                               ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time cmake -DCMAKE_BUILD_TYPE=Debug ${CMAKE_EXTRA} .. ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                          ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     (time ctest --output-on-failure -E test-opt ) 2>&1 | tee -a $OUT/${ci}-ctest.log
 
@@ -109,8 +115,8 @@ function gg_run_ctest_release {
 
     set -e
 
-    (time cmake -DCMAKE_BUILD_TYPE=Release ..   ) 2>&1 | tee -a $OUT/${ci}-cmake.log
-    (time make -j                               ) 2>&1 | tee -a $OUT/${ci}-make.log
+    (time cmake -DCMAKE_BUILD_TYPE=Release .. ${CMAKE_EXTRA} ) 2>&1 | tee -a $OUT/${ci}-cmake.log
+    (time make -j                                            ) 2>&1 | tee -a $OUT/${ci}-make.log
 
     if [ -z ${GG_BUILD_LOW_PERF} ]; then
         (time ctest --output-on-failure ) 2>&1 | tee -a $OUT/${ci}-ctest.log

ggml-metal.m

@@ -257,13 +257,14 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
         bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
 #endif
         NSError * error = nil;
-        NSString * libPath = [bundle pathForResource:@"default" ofType:@"metallib"];
+        NSString * libPath = [bundle pathForResource:@"ggml" ofType:@"metallib"];
         if (libPath != nil) {
+            // pre-compiled library found
             NSURL * libURL = [NSURL fileURLWithPath:libPath];
             GGML_METAL_LOG_INFO("%s: loading '%s'\n", __func__, [libPath UTF8String]);
             ctx->library = [ctx->device newLibraryWithURL:libURL error:&error];
         } else {
-            GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
+            GGML_METAL_LOG_INFO("%s: ggml.metallib not found, loading from source\n", __func__);
 
             NSString * sourcePath;
             NSString * ggmlMetalPathResources = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
@@ -291,6 +292,13 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
             options = [MTLCompileOptions new];
             options.preprocessorMacros = @{ @"QK_K" : @(64) };
 #endif
+            // try to disable fast-math
+            // NOTE: this seems to have no effect whatsoever
+            //       instead, in order to disable fast-math, we have to build ggml.metallib from the command line
+            //       using xcrun -sdk macosx metal -fno-fast-math -c ggml-metal.metal -o ggml-metal.air
+            //       and go through the "pre-compiled library found" path above
+            //[options setFastMathEnabled:false];
+
             ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error];
         }