llama.swiftui : use llama.cpp as SPM package (#4804 )

llama : print tensor meta for debugging
llama.swiftui : add visionOS target (#4805 )
2026-07-11 15:05:54 +02:00 · 2024-01-07 10:20:50 +02:00 · 2024-01-07 09:51:12 +02:00 · 2024-01-07 09:46:55 +02:00 · 2024-01-07 08:52:42 +02:00 · 2024-01-07 08:45:26 +02:00
20 changed files with 147 additions and 175 deletions
@@ -9,7 +9,7 @@
  git,
  python3,
  mpi,
-  openblas, # TODO: Use the generic `blas` so users could switch betwen alternative implementations
+  openblas, # TODO: Use the generic `blas` so users could switch between alternative implementations
  cudaPackages,
  darwin,
  rocmPackages,
@@ -177,27 +177,29 @@ if (LLAMA_METAL)
    if (LLAMA_METAL_SHADER_DEBUG)
        # 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
+        #   xcrun -sdk macosx metallib                   ggml-metal.air   -o default.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
        # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
+        # note: unfortunately, we have to call it default.metallib instead of ggml.metallib
+        #       ref: https://github.com/ggerganov/whisper.cpp/issues/1720
        set(XC_FLAGS -fno-fast-math -fno-inline -g)
        if (LLAMA_QKK_64)
            set(XC_FLAGS ${XC_FLAGS} -DQK_K=64)
        endif()

        add_custom_command(
-            OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml.metallib
+            OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
            COMMAND xcrun -sdk macosx metal    ${XC_FLAGS} -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
+            COMMAND xcrun -sdk macosx metallib                ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal.air   -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
            DEPENDS ggml-metal.metal
            COMMENT "Compiling Metal kernels"
        )

        add_custom_target(
            ggml-metal ALL
-            DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml.metallib
+            DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
        )
    endif()

@@ -228,7 +230,11 @@ if (LLAMA_BLAS)
            if (${LLAMA_BLAS_VENDOR} MATCHES "Generic")
                pkg_check_modules(DepBLAS REQUIRED blas)
            elseif (${LLAMA_BLAS_VENDOR} MATCHES "OpenBLAS")
-                pkg_check_modules(DepBLAS REQUIRED openblas)
+                # As of openblas v0.3.22, the 64-bit is named openblas64.pc
+                pkg_check_modules(DepBLAS openblas64)
+                if (NOT DepBLAS_FOUND)
+                    pkg_check_modules(DepBLAS REQUIRED openblas)
+                endif()
            elseif (${LLAMA_BLAS_VENDOR} MATCHES "FLAME")
                pkg_check_modules(DepBLAS REQUIRED blis)
            elseif (${LLAMA_BLAS_VENDOR} MATCHES "ATLAS")
@@ -0,0 +1,61 @@
+#!/bin/bash
+#
+# Few-shot translation example.
+# Requires a base model (i.e. no fine-tuned or instruct models).
+#
+# Usage:
+#
+#   cd llama.cpp
+#   make -j
+#
+#   ./examples/base-translate.sh <model-base> "<text>" [extra-main-args]
+#
+
+if [ $# -lt 2 ]; then
+  echo "Usage: ./base-translate.sh <model-base> \"<text>\" [extra-main-args]"
+  exit 1
+fi
+
+eargs=""
+if [ $# -gt 2 ]; then
+  eargs="${@:3}"
+fi
+
+ftmp="__llama.cpp_example_tmp__.txt"
+trap "rm -f $ftmp" EXIT
+
+echo "Translate from English to French:
+
+===
+
+sea otter, peppermint, plush girafe:
+
+sea otter => loutre de mer
+peppermint => menthe poivrée
+plush girafe => girafe peluche
+
+===
+
+violin
+
+violin => violon
+
+===
+
+phone, computer, mouse, keyboard:
+
+phone => téléphone
+computer => ordinateur
+mouse => souris
+keyboard => clavier
+
+===
+" > $ftmp
+
+echo "$2
+" >> $ftmp
+
+model=$1
+
+# generate the most likely continuation until the string "===" is found
+./main -m $model -f $ftmp -n 64 --temp 0 --repeat-penalty 1.0 --no-penalize-nl -r "===" $eargs
@@ -3,15 +3,9 @@
 #include "llama.h"
 #include "common.h"
 #include "train.h"
-#include <unordered_map>
 #include <vector>
-#include <cassert>
-#include <climits>
 #include <cstring>
-#include <cstdarg>
 #include <ctime>
-#include <random>
-#include <stdexcept>
 #include <algorithm>
 #include <string>

@@ -1,8 +1,5 @@
 import Foundation
-
-// To use this in your own project, add llama.cpp as a swift package dependency
-// and uncomment this import line.
-// import llama
+import llama

 enum LlamaError: Error {
    case couldNotInitializeContext
@@ -161,7 +158,7 @@ actor LlamaContext {
            new_token_id = llama_sample_token_greedy(context, &candidates_p)
        }

-        if new_token_id == llama_token_eos(context) || n_cur == n_len {
+        if new_token_id == llama_token_eos(model) || n_cur == n_len {
            print("\n")
            let new_token_str = String(cString: temporary_invalid_cchars + [0])
            temporary_invalid_cchars.removeAll()
@@ -1,5 +0,0 @@
-//
-//  Use this file to import your target's public headers that you would like to expose to Swift.
-//
-
-#import "llama.h"
@@ -7,68 +7,31 @@
 	objects = {

 /* Begin PBXBuildFile section */
-		542376082B0D9BFB008E6A1C /* ggml-quants.c in Sources */ = {isa = PBXBuildFile; fileRef = 542376072B0D9BFB008E6A1C /* ggml-quants.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
-		5423760B2B0D9C4B008E6A1C /* ggml-backend.c in Sources */ = {isa = PBXBuildFile; fileRef = 5423760A2B0D9C4B008E6A1C /* ggml-backend.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
-		542EA09D2AC8723900A8AEE9 /* ggml.c in Sources */ = {isa = PBXBuildFile; fileRef = 542EA09B2AC8723900A8AEE9 /* ggml.c */; settings = {COMPILER_FLAGS = "-DGGML_USE_ACCELERATE -DGGML_USE_METAL -DGGML_USE_K_QUANTS -O3"; }; };
-		542EA0A02AC8725700A8AEE9 /* ggml-alloc.c in Sources */ = {isa = PBXBuildFile; fileRef = 542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */; settings = {COMPILER_FLAGS = "-O3"; }; };
-		542EA0A32AC8729100A8AEE9 /* llama.cpp in Sources */ = {isa = PBXBuildFile; fileRef = 542EA0A12AC8729100A8AEE9 /* llama.cpp */; settings = {COMPILER_FLAGS = "-DGGML_USE_K_QUANTS -DGGML_USE_METAL -O3"; }; };
 		549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 549479CA2AC9E16000E0F78B /* Metal.framework */; };
-		549479CD2AC9E42A00E0F78B /* ggml-metal.m in Sources */ = {isa = PBXBuildFile; fileRef = 549479C52AC9E0F200E0F78B /* ggml-metal.m */; settings = {COMPILER_FLAGS = "-fno-objc-arc -DGGML_SWIFT -DGGML_USE_METAL -O3"; }; };
 		7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = 7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */; };
 		8A1C83772AC328BD0096AF73 /* llama_swiftuiApp.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */; };
 		8A1C83792AC328BD0096AF73 /* ContentView.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A1C83782AC328BD0096AF73 /* ContentView.swift */; };
 		8A1C837B2AC328BE0096AF73 /* Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 8A1C837A2AC328BE0096AF73 /* Assets.xcassets */; };
-		8A1C837E2AC328BE0096AF73 /* Preview Assets.xcassets in Resources */ = {isa = PBXBuildFile; fileRef = 8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */; };
 		8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */ = {isa = PBXBuildFile; fileRef = 8A39BE092AC7601000BFEB40 /* Accelerate.framework */; };
 		8A3F84242AC4C891005E2EE8 /* models in Resources */ = {isa = PBXBuildFile; fileRef = 8A3F84232AC4C891005E2EE8 /* models */; };
 		8A907F332AC7138A006146EA /* LibLlama.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A907F322AC7134E006146EA /* LibLlama.swift */; };
 		8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */ = {isa = PBXBuildFile; fileRef = 8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */; };
+		DF810E132B4A5BA200301144 /* llama in Frameworks */ = {isa = PBXBuildFile; productRef = DF810E122B4A5BA200301144 /* llama */; };
 		F1FE20E22B465ECA00B45541 /* LoadCustomButton.swift in Sources */ = {isa = PBXBuildFile; fileRef = F1FE20E12B465EC900B45541 /* LoadCustomButton.swift */; };
-		F1FE20DC2B465C4500B45541 /* ggml-metal.metal in Resources */ = {isa = PBXBuildFile; fileRef = 549479C82AC9E10B00E0F78B /* ggml-metal.metal */; };
 /* End PBXBuildFile section */

-/* Begin PBXBuildRule section */
-		F1FE20DB2B465C2100B45541 /* PBXBuildRule */ = {
-			isa = PBXBuildRule;
-			compilerSpec = com.apple.compilers.proxy.script;
-			fileType = sourcecode.metal;
-			inputFiles = (
-			);
-			isEditable = 1;
-			outputFiles = (
-				"${DERIVED_FILES_DIR}/ggml-metal.air",
-				"${DERIVED_FILES_DIR}/ggml.metallib",
-			);
-			script = "# metal\nxcrun metal    -c \"${INPUT_FILE_PATH}\" -o \"${DERIVED_FILES_DIR}/${INPUT_FILE_BASE}.air\"\nxcrun metallib -o \"${DERIVED_FILES_DIR}/${INPUT_FILE_BASE%-metal}.metallib\" \"${DERIVED_FILES_DIR}/${INPUT_FILE_BASE}.air\"\n";
-		};
-/* End PBXBuildRule section */
-
 /* Begin PBXFileReference section */
-		542376062B0D9BEA008E6A1C /* ggml-quants.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-quants.h"; path = "../../ggml-quants.h"; sourceTree = "<group>"; };
-		542376072B0D9BFB008E6A1C /* ggml-quants.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-quants.c"; path = "../../ggml-quants.c"; sourceTree = "<group>"; };
-		542376092B0D9C40008E6A1C /* ggml-backend.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; name = "ggml-backend.h"; path = "../../ggml-backend.h"; sourceTree = "<group>"; };
-		5423760A2B0D9C4B008E6A1C /* ggml-backend.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-backend.c"; path = "../../ggml-backend.c"; sourceTree = "<group>"; };
-		542EA09B2AC8723900A8AEE9 /* ggml.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = ggml.c; path = ../../ggml.c; sourceTree = "<group>"; };
-		542EA09C2AC8723900A8AEE9 /* ggml.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = ggml.h; path = ../../ggml.h; sourceTree = "<group>"; };
-		542EA09E2AC8725700A8AEE9 /* ggml-alloc.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-alloc.h"; path = "../../ggml-alloc.h"; sourceTree = "<group>"; };
-		542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.c; name = "ggml-alloc.c"; path = "../../ggml-alloc.c"; sourceTree = "<group>"; };
-		542EA0A12AC8729100A8AEE9 /* llama.cpp */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.cpp.cpp; name = llama.cpp; path = ../../llama.cpp; sourceTree = "<group>"; };
-		542EA0A22AC8729100A8AEE9 /* llama.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = llama.h; path = ../../llama.h; sourceTree = "<group>"; };
-		549479C52AC9E0F200E0F78B /* ggml-metal.m */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.objc; name = "ggml-metal.m"; path = "../../ggml-metal.m"; sourceTree = "<group>"; };
-		549479C62AC9E0F200E0F78B /* ggml-metal.h */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.c.h; name = "ggml-metal.h"; path = "../../ggml-metal.h"; sourceTree = "<group>"; };
-		549479C82AC9E10B00E0F78B /* ggml-metal.metal */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.metal; name = "ggml-metal.metal"; path = "../../ggml-metal.metal"; sourceTree = "<group>"; };
 		549479CA2AC9E16000E0F78B /* Metal.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Metal.framework; path = System/Library/Frameworks/Metal.framework; sourceTree = SDKROOT; };
 		7FA3D2B22B2EA2F600543F92 /* DownloadButton.swift */ = {isa = PBXFileReference; fileEncoding = 4; lastKnownFileType = sourcecode.swift; path = DownloadButton.swift; sourceTree = "<group>"; };
-		8A08D20A2AC73B1500FE6CD4 /* bridging-header.h */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.c.h; path = "bridging-header.h"; sourceTree = "<group>"; };
 		8A1C83732AC328BD0096AF73 /* llama.swiftui.app */ = {isa = PBXFileReference; explicitFileType = wrapper.application; includeInIndex = 0; path = llama.swiftui.app; sourceTree = BUILT_PRODUCTS_DIR; };
 		8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = llama_swiftuiApp.swift; sourceTree = "<group>"; };
 		8A1C83782AC328BD0096AF73 /* ContentView.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = ContentView.swift; sourceTree = "<group>"; };
 		8A1C837A2AC328BE0096AF73 /* Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = Assets.xcassets; sourceTree = "<group>"; };
-		8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */ = {isa = PBXFileReference; lastKnownFileType = folder.assetcatalog; path = "Preview Assets.xcassets"; sourceTree = "<group>"; };
 		8A39BE092AC7601000BFEB40 /* Accelerate.framework */ = {isa = PBXFileReference; lastKnownFileType = wrapper.framework; name = Accelerate.framework; path = System/Library/Frameworks/Accelerate.framework; sourceTree = SDKROOT; };
 		8A3F84232AC4C891005E2EE8 /* models */ = {isa = PBXFileReference; lastKnownFileType = folder; name = models; path = llama.swiftui/Resources/models; sourceTree = "<group>"; };
 		8A907F322AC7134E006146EA /* LibLlama.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LibLlama.swift; sourceTree = "<group>"; };
 		8A9F7C4C2AC332EE008AE1EA /* LlamaState.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LlamaState.swift; sourceTree = "<group>"; };
+		DF2D2FE72B4A59BE00FCB72D /* llama.cpp */ = {isa = PBXFileReference; lastKnownFileType = wrapper; name = llama.cpp; path = ../..; sourceTree = "<group>"; };
 		F1FE20E12B465EC900B45541 /* LoadCustomButton.swift */ = {isa = PBXFileReference; lastKnownFileType = sourcecode.swift; path = LoadCustomButton.swift; sourceTree = "<group>"; };
 /* End PBXFileReference section */

@@ -77,6 +40,7 @@
 			isa = PBXFrameworksBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
+				DF810E132B4A5BA200301144 /* llama in Frameworks */,
 				549479CB2AC9E16000E0F78B /* Metal.framework in Frameworks */,
 				8A39BE0A2AC7601100BFEB40 /* Accelerate.framework in Frameworks */,
 			);
@@ -85,30 +49,10 @@
 /* End PBXFrameworksBuildPhase section */

 /* Begin PBXGroup section */
-		8A08D1F62AC7383900FE6CD4 /* llama.cpp */ = {
-			isa = PBXGroup;
-			children = (
-				5423760A2B0D9C4B008E6A1C /* ggml-backend.c */,
-				542376092B0D9C40008E6A1C /* ggml-backend.h */,
-				542376062B0D9BEA008E6A1C /* ggml-quants.h */,
-				542376072B0D9BFB008E6A1C /* ggml-quants.c */,
-				549479C82AC9E10B00E0F78B /* ggml-metal.metal */,
-				549479C62AC9E0F200E0F78B /* ggml-metal.h */,
-				549479C52AC9E0F200E0F78B /* ggml-metal.m */,
-				542EA09B2AC8723900A8AEE9 /* ggml.c */,
-				542EA09C2AC8723900A8AEE9 /* ggml.h */,
-				542EA09F2AC8725700A8AEE9 /* ggml-alloc.c */,
-				542EA09E2AC8725700A8AEE9 /* ggml-alloc.h */,
-				542EA0A12AC8729100A8AEE9 /* llama.cpp */,
-				542EA0A22AC8729100A8AEE9 /* llama.h */,
-			);
-			name = llama.cpp;
-			sourceTree = "<group>";
-		};
 		8A1C836A2AC328BD0096AF73 = {
 			isa = PBXGroup;
 			children = (
-				8A08D1F62AC7383900FE6CD4 /* llama.cpp */,
+				DF2D2FE72B4A59BE00FCB72D /* llama.cpp */,
 				8A907F312AC7134E006146EA /* llama.cpp.swift */,
 				8A3F84232AC4C891005E2EE8 /* models */,
 				8A1C83752AC328BD0096AF73 /* llama.swiftui */,
@@ -133,19 +77,10 @@
 				8A9F7C4A2AC332BF008AE1EA /* UI */,
 				8A1C83762AC328BD0096AF73 /* llama_swiftuiApp.swift */,
 				8A1C837A2AC328BE0096AF73 /* Assets.xcassets */,
-				8A1C837C2AC328BE0096AF73 /* Preview Content */,
 			);
 			path = llama.swiftui;
 			sourceTree = "<group>";
 		};
-		8A1C837C2AC328BE0096AF73 /* Preview Content */ = {
-			isa = PBXGroup;
-			children = (
-				8A1C837D2AC328BE0096AF73 /* Preview Assets.xcassets */,
-			);
-			path = "Preview Content";
-			sourceTree = "<group>";
-		};
 		8A39BE082AC7601000BFEB40 /* Frameworks */ = {
 			isa = PBXGroup;
 			children = (
@@ -173,7 +108,6 @@
 		8A907F312AC7134E006146EA /* llama.cpp.swift */ = {
 			isa = PBXGroup;
 			children = (
-				8A08D20A2AC73B1500FE6CD4 /* bridging-header.h */,
 				8A907F322AC7134E006146EA /* LibLlama.swift */,
 			);
 			path = llama.cpp.swift;
@@ -209,12 +143,12 @@
 				8A1C83712AC328BD0096AF73 /* Resources */,
 			);
 			buildRules = (
-				F1FE20DB2B465C2100B45541 /* PBXBuildRule */,
 			);
 			dependencies = (
 			);
 			name = llama.swiftui;
 			packageProductDependencies = (
+				DF810E122B4A5BA200301144 /* llama */,
 			);
 			productName = llama.swiftui;
 			productReference = 8A1C83732AC328BD0096AF73 /* llama.swiftui.app */;
@@ -261,9 +195,7 @@
 			isa = PBXResourcesBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
-				F1FE20DC2B465C4500B45541 /* ggml-metal.metal in Resources */,
 				8A3F84242AC4C891005E2EE8 /* models in Resources */,
-				8A1C837E2AC328BE0096AF73 /* Preview Assets.xcassets in Resources */,
 				8A1C837B2AC328BE0096AF73 /* Assets.xcassets in Resources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
@@ -275,18 +207,12 @@
 			isa = PBXSourcesBuildPhase;
 			buildActionMask = 2147483647;
 			files = (
-				542376082B0D9BFB008E6A1C /* ggml-quants.c in Sources */,
-				549479CD2AC9E42A00E0F78B /* ggml-metal.m in Sources */,
 				F1FE20E22B465ECA00B45541 /* LoadCustomButton.swift in Sources */,
-				542EA09D2AC8723900A8AEE9 /* ggml.c in Sources */,
 				8A907F332AC7138A006146EA /* LibLlama.swift in Sources */,
-				542EA0A32AC8729100A8AEE9 /* llama.cpp in Sources */,
 				8A9F7C4D2AC332EE008AE1EA /* LlamaState.swift in Sources */,
 				8A1C83792AC328BD0096AF73 /* ContentView.swift in Sources */,
 				8A1C83772AC328BD0096AF73 /* llama_swiftuiApp.swift in Sources */,
 				7FA3D2B32B2EA2F600543F92 /* DownloadButton.swift in Sources */,
-				542EA0A02AC8725700A8AEE9 /* ggml-alloc.c in Sources */,
-				5423760B2B0D9C4B008E6A1C /* ggml-backend.c in Sources */,
 			);
 			runOnlyForDeploymentPostprocessing = 0;
 		};
@@ -416,11 +342,9 @@
 			isa = XCBuildConfiguration;
 			buildSettings = {
 				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
-				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
 				CLANG_ENABLE_MODULES = YES;
 				CODE_SIGN_STYLE = Automatic;
 				CURRENT_PROJECT_VERSION = 1;
-				DEVELOPMENT_ASSET_PATHS = "\"llama.swiftui/Preview Content\"";
 				DEVELOPMENT_TEAM = STLSG3FG8Q;
 				ENABLE_PREVIEWS = YES;
 				GENERATE_INFOPLIST_FILE = YES;
@@ -437,11 +361,12 @@
 				MARKETING_VERSION = 1.0;
 				PRODUCT_BUNDLE_IDENTIFIER = "com.bachittle.llama-swift";
 				PRODUCT_NAME = "$(TARGET_NAME)";
+				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator xros xrsimulator";
+				SUPPORTS_XR_DESIGNED_FOR_IPHONE_IPAD = NO;
 				SWIFT_EMIT_LOC_STRINGS = YES;
-				SWIFT_OBJC_BRIDGING_HEADER = "llama.cpp.swift/bridging-header.h";
 				SWIFT_OPTIMIZATION_LEVEL = "-Onone";
 				SWIFT_VERSION = 5.0;
-				TARGETED_DEVICE_FAMILY = "1,2";
+				TARGETED_DEVICE_FAMILY = "1,2,7";
 			};
 			name = Debug;
 		};
@@ -449,11 +374,9 @@
 			isa = XCBuildConfiguration;
 			buildSettings = {
 				ASSETCATALOG_COMPILER_APPICON_NAME = AppIcon;
-				ASSETCATALOG_COMPILER_GLOBAL_ACCENT_COLOR_NAME = AccentColor;
 				CLANG_ENABLE_MODULES = YES;
 				CODE_SIGN_STYLE = Automatic;
 				CURRENT_PROJECT_VERSION = 1;
-				DEVELOPMENT_ASSET_PATHS = "\"llama.swiftui/Preview Content\"";
 				DEVELOPMENT_TEAM = STLSG3FG8Q;
 				ENABLE_PREVIEWS = YES;
 				GENERATE_INFOPLIST_FILE = YES;
@@ -470,10 +393,11 @@
 				MARKETING_VERSION = 1.0;
 				PRODUCT_BUNDLE_IDENTIFIER = "com.bachittle.llama-swift";
 				PRODUCT_NAME = "$(TARGET_NAME)";
+				SUPPORTED_PLATFORMS = "iphoneos iphonesimulator xros xrsimulator";
+				SUPPORTS_XR_DESIGNED_FOR_IPHONE_IPAD = NO;
 				SWIFT_EMIT_LOC_STRINGS = YES;
-				SWIFT_OBJC_BRIDGING_HEADER = "llama.cpp.swift/bridging-header.h";
 				SWIFT_VERSION = 5.0;
-				TARGETED_DEVICE_FAMILY = "1,2";
+				TARGETED_DEVICE_FAMILY = "1,2,7";
 			};
 			name = Release;
 		};
@@ -499,6 +423,13 @@
 			defaultConfigurationName = Release;
 		};
 /* End XCConfigurationList section */
+
+/* Begin XCSwiftPackageProductDependency section */
+		DF810E122B4A5BA200301144 /* llama */ = {
+			isa = XCSwiftPackageProductDependency;
+			productName = llama;
+		};
+/* End XCSwiftPackageProductDependency section */
 	};
 	rootObject = 8A1C836B2AC328BD0096AF73 /* Project object */;
 }
@@ -1,11 +0,0 @@
-{
-  "colors" : [
-    {
-      "idiom" : "universal"
-    }
-  ],
-  "info" : {
-    "author" : "xcode",
-    "version" : 1
-  }
-}
@@ -1,6 +0,0 @@
-{
-  "info" : {
-    "author" : "xcode",
-    "version" : 1
-  }
-}
@@ -447,8 +447,14 @@ struct llama_client_slot
    }

    bool has_budget(gpt_params &global_params) {
+        if (params.n_predict == -1 && global_params.n_predict == -1)
+        {
+            return true; // limitless
+        }
+
        n_remaining = -1;
-        if(params.n_predict != -1)
+
+        if (params.n_predict != -1)
        {
            n_remaining = params.n_predict - n_decoded;
        }
@@ -456,7 +462,8 @@ struct llama_client_slot
        {
            n_remaining = global_params.n_predict - n_decoded;
        }
-        return n_remaining > 0 || n_remaining == -1; // no budget || limitless
+
+        return n_remaining > 0; // no budget
    }

    bool available() const {
@@ -1102,7 +1109,7 @@ struct llama_server_context
        }

        // check the limits
-        if (slot.n_decoded > 2 && slot.has_next_token && !slot.has_budget(params))
+        if (slot.n_decoded > 0 && slot.has_next_token && !slot.has_budget(params))
        {
            slot.stopped_limit = true;
            slot.has_next_token = false;
@@ -1703,7 +1710,6 @@ struct llama_server_context

            llama_batch_add(batch, slot.sampled, system_tokens.size() + slot.n_past, { slot.id }, true);

-            slot.n_decoded += 1;
            slot.n_past += 1;
        }

@@ -1921,6 +1927,7 @@ struct llama_server_context

                llama_sampling_accept(slot.ctx_sampling, ctx, id, true);

+                slot.n_decoded += 1;
                if (slot.n_decoded == 1)
                {
                    slot.t_start_genereration = ggml_time_us();
@@ -90,7 +90,7 @@ extern "C" {
        void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);

        // compute graph without a plan
-        void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        bool (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);

        // check if the backend supports an operation
        bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
@@ -195,11 +195,14 @@ void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_
    ggml_backend_synchronize(backend);
 }

-void ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    backend->iface.graph_compute(backend, cgraph);
+bool ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    if (!backend->iface.graph_compute(backend, cgraph)) {
+        return false;
+    }

    // TODO: optional sync
    ggml_backend_synchronize(backend);
+    return true;
 }

 bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
@@ -597,7 +600,7 @@ static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_bac
    GGML_UNUSED(backend);
 }

-static void ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
    struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;

    struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
@@ -611,6 +614,7 @@ static void ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_c
    cplan.work_data = cpu_ctx->work_data;

    ggml_graph_compute(cgraph, &cplan);
+    return true;
 }

 static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
@@ -58,7 +58,7 @@ extern "C" {

    GGML_API void ggml_backend_graph_plan_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
    GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API void ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API bool ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
    GGML_API bool ggml_backend_supports_op       (ggml_backend_t backend, const struct ggml_tensor * op);

    // tensor copy between different backends
@@ -183,7 +183,7 @@ static __device__ __forceinline__ int __vsubss4(const int a, const int b) {
 static __device__ __forceinline__ int __dp4a(const int a, const int b, int c) {
 #if defined(__gfx906__) || defined(__gfx908__) || defined(__gfx90a__) || defined(__gfx1030__)
    c = __builtin_amdgcn_sdot4(a, b, c, false);
-#elif defined(__gfx1100__)
+#elif defined(RDNA3)
    c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
 #elif defined(__gfx1010__) || defined(__gfx900__)
    int tmp1;
@@ -9910,7 +9910,7 @@ static void ggml_backend_cuda_graph_plan_compute(ggml_backend_t backend, ggml_ba
    UNUSED(plan);
 }

-static void ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
+static bool ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph * cgraph) {
    ggml_backend_context_cuda * cuda_ctx = (ggml_backend_context_cuda *)backend->context;

    ggml_cuda_set_main_device(cuda_ctx->device);
@@ -9967,6 +9967,8 @@ static void ggml_backend_cuda_graph_compute(ggml_backend_t backend, ggml_cgraph
    }

    UNUSED(backend);
+
+    return true;
 }

 static bool ggml_backend_cuda_supports_op(ggml_backend_t backend, const ggml_tensor * op) {
@@ -87,7 +87,7 @@ int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx);

 // same as ggml_graph_compute but uses Metal
 // creates gf->n_threads command buffers in parallel
-void ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);
+bool ggml_metal_graph_compute(struct ggml_metal_context * ctx, struct ggml_cgraph * gf);

 //
 // backend API
@@ -258,14 +258,14 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
        bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
 #endif
        NSError * error = nil;
-        NSString * libPath = [bundle pathForResource:@"ggml" ofType:@"metallib"];
+        NSString * libPath = [bundle pathForResource:@"default" 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: ggml.metallib not found, loading from source\n", __func__);
+            GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);

            NSString * sourcePath;
            NSString * ggmlMetalPathResources = [[NSProcessInfo processInfo].environment objectForKey:@"GGML_METAL_PATH_RESOURCES"];
@@ -295,7 +295,7 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
 #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
+            //       instead, in order to disable fast-math, we have to build default.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];
@@ -977,7 +977,7 @@ static bool ggml_metal_supports_op(const struct ggml_tensor * op) {
            return false;
    }
 }
-void ggml_metal_graph_compute(
+bool ggml_metal_graph_compute(
        struct ggml_metal_context * ctx,
               struct ggml_cgraph * gf) {
    @autoreleasepool {
@@ -2405,10 +2405,11 @@ void ggml_metal_graph_compute(
        MTLCommandBufferStatus status = (MTLCommandBufferStatus) [ctx->command_buffers[i] status];
        if (status != MTLCommandBufferStatusCompleted) {
            GGML_METAL_LOG_INFO("%s: command buffer %d failed with status %lu\n", __func__, i, status);
-            GGML_ASSERT(false);
+            return false;
        }
    }

+    return true;
    }
 }

@@ -2688,10 +2689,10 @@ static ggml_backend_buffer_type_t ggml_backend_metal_get_default_buffer_type(ggm
    UNUSED(backend);
 }

-static void ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+static bool ggml_backend_metal_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
    struct ggml_metal_context * metal_ctx = (struct ggml_metal_context *)backend->context;

-    ggml_metal_graph_compute(metal_ctx, cgraph);
+    return ggml_metal_graph_compute(metal_ctx, cgraph);
 }

 static bool ggml_backend_metal_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
@@ -70,7 +70,7 @@ static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block s
 // 2-bit quantization
 // weight is represented as x = a * q + b
 // 16 blocks of 16 elements each
-// Effectively 2.5625 bits per weight
+// Effectively 2.625 bits per weight
 typedef struct {
    uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
    uint8_t qs[QK_K/4];      // quants
@@ -9704,10 +9704,10 @@ static void ggml_compute_forward_group_norm(
 #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
 // helper function to determine if it is better to use BLAS or not
 // for large matrices, BLAS is faster
-static bool ggml_compute_forward_mul_mat_use_blas(
-        const struct ggml_tensor * src0,
-        const struct ggml_tensor * src1,
-              struct ggml_tensor * dst) {
+static bool ggml_compute_forward_mul_mat_use_blas(struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+
    //const int64_t ne00 = src0->ne[0];
    //const int64_t ne01 = src0->ne[1];

@@ -9787,7 +9787,7 @@ static void ggml_compute_forward_mul_mat(
 #endif

 #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
-    if (ggml_compute_forward_mul_mat_use_blas(src0, src1, dst)) {
+    if (ggml_compute_forward_mul_mat_use_blas(dst)) {
        if (params->ith != 0) {
            return;
        }
@@ -16301,24 +16301,6 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {

                //n_tasks = MIN(n_threads, MAX(1, nr0/128));
                //printf("nr0 = %8d, nr1 = %8d, nr0*nr1 = %8d, n_tasks%d\n", nr0, nr1, nr0*nr1, n_tasks);
-
-#if defined(GGML_USE_CUBLAS)
-                if (ggml_cuda_can_mul_mat(node->src[0], node->src[1], node)) {
-                    n_tasks = 1; // TODO: this actually is doing nothing
-                                 //       the threads are still spinning
-                }
-#elif defined(GGML_USE_CLBLAST)
-                if (ggml_cl_can_mul_mat(node->src[0], node->src[1], node)) {
-                    n_tasks = 1; // TODO: this actually is doing nothing
-                                 //       the threads are still spinning
-                }
-#endif
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
-                if (ggml_compute_forward_mul_mat_use_blas(node->src[0], node->src[1], node)) {
-                    n_tasks = 1; // TODO: this actually is doing nothing
-                                 //       the threads are still spinning
-                }
-#endif
            } break;
        case GGML_OP_MUL_MAT_ID:
            {
@@ -16491,6 +16473,7 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
            state->shared->node_n += 1;
            return (thread_ret_t) GGML_EXIT_ABORTED;
        }
+
        if (atomic_fetch_sub(&state->shared->n_active, 1) == 1) {
            // all other threads are finished and spinning
            // do finalize and init here so we don't have synchronize again
@@ -16556,14 +16539,18 @@ static thread_ret_t ggml_graph_compute_thread(void * data) {
        } else {
            // wait for other threads to finish
            const int last = node_n;
+
+            const bool do_yield = last < 0 || cgraph->nodes[last]->op == GGML_OP_MUL_MAT;
+
            while (true) {
                // TODO: this sched_yield can have significant impact on the performance - either positive or negative
                //       depending on the workload and the operating system.
                //       since it is not clear what is the best approach, it should potentially become user-configurable
                //       ref: https://github.com/ggerganov/ggml/issues/291
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
-                sched_yield();
-#endif
+                // UPD:  adding the do_yield flag seems to resolve the issue universally
+                if (do_yield) {
+                    sched_yield();
+                }

                node_n = atomic_load(&state->shared->node_n);
                if (node_n != last) break;
@@ -16642,7 +16629,7 @@ struct ggml_cplan ggml_graph_plan(struct ggml_cgraph * cgraph, int n_threads) {
                    } else
 #endif
 #if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS)
-                    if (ggml_compute_forward_mul_mat_use_blas(node->src[0], node->src[1], node)) {
+                    if (ggml_compute_forward_mul_mat_use_blas(node)) {
                        if (node->src[0]->type != GGML_TYPE_F32) {
                            // here we need memory just for single 2D matrix from src0
                            cur = ggml_type_size(GGML_TYPE_F32)*(node->src[0]->ne[0]*node->src[0]->ne[1]);
@@ -2180,7 +2180,11 @@ struct llama_model_loader {
                    type_max   = type;
                }

-                // LLAMA_LOG_INFO("%s: - tensor %4d: %32s %-8s [ %s ]\n", __func__, i, name, ggml_type_name(meta->type), llama_format_tensor_shape(meta).c_str());
+                // TODO: make runtime configurable
+#if 0
+                struct ggml_tensor * meta = ggml_get_tensor(ctx_meta, gguf_get_tensor_name(ctx_gguf, i));
+                LLAMA_LOG_INFO("%s: - tensor %4d: %32s %-8s [ %s ]\n", __func__, i, ggml_get_name(meta), ggml_type_name(type), llama_format_tensor_shape(meta).c_str());
+#endif
            }

            switch (type_max) {
@@ -1 +1 @@
-3fd01e00e40583ccd4b393a7c6502d6a4455a1d5
+f96711108d55bdbbd277e6be07204dce6a94fb93
Author	SHA1	Message	Date
Alex Azarov	72d8407b36	llama.swiftui : use llama.cpp as SPM package (#4804 )	2024-01-07 10:20:50 +02:00
Georgi Gerganov	d117d4dc5d	llama : print tensor meta for debugging	2024-01-07 09:51:12 +02:00
Alex Azarov	3418c03ecc	llama.swiftui : add visionOS target (#4805 )	2024-01-07 09:46:55 +02:00
Konstantin Zhuravlyov	63ee677efd	ggml : use __builtin_amdgcn_sudot4 in __dp4a for gfx11 (#4787 )	2024-01-07 08:52:42 +02:00
Georgi Gerganov	67984921a7	server : fix n_predict check (#4798 )	2024-01-07 08:45:26 +02:00
Daniel Illescas Romero	c75ca5d96f	llama.swiftui : use correct pointer for llama_token_eos (#4797 )	2024-01-06 17:12:59 +02:00
Georgi Gerganov	96e80dabc6	examples : improve base-translate.sh script (#4783 )	2024-01-06 11:40:24 +02:00
a-n-n-a-l-e-e	eec22a1c63	cmake : check for openblas64 (#4134 ) openblas v0.3.22 64-bit pkg-config file is named openblas64.pc https://github.com/OpenMathLib/OpenBLAS/issues/3790	2024-01-05 18:04:40 +02:00
Ikko Eltociear Ashimine	be36bb946a	flake.nix : fix typo (#4700 ) betwen -> between	2024-01-05 18:02:44 +02:00
Georgi Gerganov	91d38876df	metal : switch back to default.metallib (ggml/681) ggml-ci	2024-01-05 18:02:06 +02:00
Georgi Gerganov	d061bf9405	ggml : fix q2_k bpw in comments (ggml/680)	2024-01-05 18:02:06 +02:00
Finn Voorhees	1bf681f90e	ggml : add error handling to graph_compute (whisper/1714)	2024-01-05 18:02:06 +02:00
Georgi Gerganov	c1d7cb28d3	ggml : do not sched_yield when calling BLAS (#4761 ) * ggml : do not sched_yield when calling BLAS ggml-ci * ggml : fix do_yield logic ggml-ci * ggml : simplify do_yield logic ggml-ci	2024-01-05 15:18:21 +02:00
Georgi Gerganov	3681f22443	examples : add few-shot translation example (#4783 )	2024-01-05 15:11:10 +02:00
Daniel Bevenius	b3a7c20b5c	finetune : remove unused includes (#4756 ) This commit removes unused includes from finetune.cpp. Signed-off-by: Daniel Bevenius <daniel.bevenius@gmail.com>	2024-01-04 21:45:37 +02:00