llama : fix Qwen model type strings

This change upstreams llamafile's cpu matrix
multiplication kernels for ppc64le using MMA
builtins for FP32 datatype.

This change results in a consistent 90%
improvement in input processing time, and 20%
to 80% improvement in output processing time,
across various batch sizes.

The patch is tested with Meta-Lllama-3-8B,
Mistral-7B, Llama-2-7B-chat-hf models on a
IBM POWER10 machine.

Signed-off-by: Amrita H S <amritahs@linux.vnet.ibm.com>

.editorconfig

@@ -24,6 +24,16 @@ insert_final_newline = unset
 [examples/server/public/*]
 indent_size = 2
 
+[examples/server/public/deps_*]
+trim_trailing_whitespace = unset
+indent_style = unset
+indent_size = unset
+
+[examples/server/deps_*]
+trim_trailing_whitespace = unset
+indent_style = unset
+indent_size = unset
+
 [examples/llama.swiftui/llama.swiftui.xcodeproj/*]
 indent_style = tab
 

.github/workflows/build.yml

@@ -55,7 +55,13 @@ jobs:
           sysctl -a
           mkdir build
           cd build
-          cmake -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL_EMBED_LIBRARY=ON -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF ..
+          cmake .. \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_CURL=ON \
+            -DGGML_METAL_USE_BF16=ON \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DGGML_RPC=ON \
+            -DBUILD_SHARED_LIBS=OFF
           cmake --build . --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -92,7 +98,7 @@ jobs:
           name: llama-bin-macos-arm64.zip
 
   macOS-latest-cmake-x64:
-    runs-on: macos-12
+    runs-on: macos-13
 
     steps:
       - name: Clone
@@ -113,7 +119,12 @@ jobs:
           sysctl -a
           # Metal is disabled due to intermittent failures with Github runners not having a GPU:
           # https://github.com/ggerganov/llama.cpp/actions/runs/8635935781/job/23674807267#step:5:2313
-          cmake -B build -DLLAMA_FATAL_WARNINGS=ON -DGGML_METAL=OFF -DLLAMA_CURL=ON -DGGML_RPC=ON -DBUILD_SHARED_LIBS=OFF
+          cmake -B build \
+            -DLLAMA_FATAL_WARNINGS=ON \
+            -DLLAMA_CURL=ON \
+            -DGGML_METAL=OFF \
+            -DGGML_RPC=ON \
+            -DBUILD_SHARED_LIBS=OFF
           cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
 
       - name: Test
@@ -569,6 +580,7 @@ jobs:
           mkdir build
           cd build
           cmake -G Xcode .. \
+            -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TESTS=OFF \
@@ -599,6 +611,7 @@ jobs:
           mkdir build
           cd build
           cmake -G Xcode .. \
+            -DGGML_METAL_USE_BF16=ON \
             -DGGML_METAL_EMBED_LIBRARY=ON \
             -DLLAMA_BUILD_EXAMPLES=OFF \
             -DLLAMA_BUILD_TESTS=OFF \

CMakeLists.txt

@@ -88,6 +88,10 @@ if (NOT DEFINED GGML_LLAMAFILE)
     set(GGML_LLAMAFILE_DEFAULT ON)
 endif()
 
+if (NOT DEFINED GGML_AMX)
+    set(GGML_AMX ON)
+endif()
+
 if (NOT DEFINED GGML_CUDA_GRAPHS)
     set(GGML_CUDA_GRAPHS_DEFAULT ON)
 endif()

CMakePresets.json

@@ -48,10 +48,23 @@
         }
     },
 
+    {
+        "name": "arm64-apple-clang", "hidden": true,
+        "architecture": { "value": "arm64",    "strategy": "external" },
+        "toolset":      { "value": "host=x64", "strategy": "external" },
+        "cacheVariables": {
+            "CMAKE_TOOLCHAIN_FILE": "${sourceDir}/cmake/arm64-apple-clang.cmake"
+        }
+    },
+
     { "name": "arm64-windows-llvm-debug"  , "inherits": [ "base", "arm64-windows-llvm",  "debug"   ] },
     { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm",  "reldbg" ] },
     { "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm",  "reldbg", "static" ] },
 
+    { "name": "arm64-apple-clang-debug"  , "inherits": [ "base", "arm64-apple-clang",  "debug"   ] },
+    { "name": "arm64-apple-clang-release"  , "inherits": [ "base", "arm64-apple-clang",  "reldbg"   ] },
+    { "name": "arm64-apple-clang+static-release"  , "inherits": [ "base", "arm64-apple-clang",  "reldbg", "static" ] },
+
     { "name": "arm64-windows-msvc-debug"  , "inherits": [ "base", "arm64-windows-msvc",  "debug"   ] },
     { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg" ] },
     { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg", "static" ] },

Makefile

@@ -1,7 +1,6 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
 	libllava.a \
-	llama-baby-llama \
 	llama-batched \
 	llama-batched-bench \
 	llama-bench \
@@ -34,6 +33,7 @@ BUILD_TARGETS = \
 	llama-save-load-state \
 	llama-server \
 	llama-simple \
+	llama-simple-chat \
 	llama-speculative \
 	llama-tokenize \
 	llama-vdot \
@@ -55,7 +55,6 @@ TEST_TARGETS = \
 	tests/test-llama-grammar \
 	tests/test-log \
 	tests/test-model-load-cancel \
-	tests/test-opt \
 	tests/test-quantize-fns \
 	tests/test-quantize-perf \
 	tests/test-rope \
@@ -63,6 +62,7 @@ TEST_TARGETS = \
 	tests/test-tokenizer-0 \
 	tests/test-tokenizer-1-bpe \
 	tests/test-tokenizer-1-spm
+#	tests/test-opt \
 
 # Legacy build targets that were renamed in #7809, but should still be removed when the project is cleaned
 LEGACY_TARGETS_CLEAN = main quantize quantize-stats perplexity imatrix embedding vdot q8dot convert-llama2c-to-ggml \
@@ -93,11 +93,6 @@ GGML_METAL := 1
 DEPRECATE_WARNING := 1
 endif
 
-ifdef LLAMA_OPENMP
-GGML_OPENMP := 1
-DEPRECATE_WARNING := 1
-endif
-
 ifdef LLAMA_RPC
 GGML_RPC := 1
 DEPRECATE_WARNING := 1
@@ -584,6 +579,11 @@ ifndef GGML_NO_LLAMAFILE
 	OBJ_GGML    += ggml/src/llamafile/sgemm.o
 endif
 
+ifndef GGML_NO_AMX
+	MK_CPPFLAGS += -DGGML_USE_AMX
+	OBJ_GGML    += ggml/src/ggml-amx.o ggml/src/ggml-amx/mmq.o
+endif
+
 ifdef GGML_RPC
 	MK_CPPFLAGS += -DGGML_USE_RPC
 	OBJ_GGML    += ggml/src/ggml-rpc.o
@@ -878,6 +878,10 @@ ifdef GGML_METAL
 	MK_CPPFLAGS += -DGGML_USE_METAL
 	MK_LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
 	OBJ_GGML	+= ggml/src/ggml-metal.o
+
+ifdef GGML_METAL_USE_BF16
+	MK_CPPFLAGS += -DGGML_METAL_USE_BF16
+endif # GGML_METAL_USE_BF16
 ifdef GGML_METAL_NDEBUG
 	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
 endif
@@ -915,6 +919,7 @@ endif # GGML_METAL
 
 OBJ_GGML += \
 	ggml/src/ggml.o \
+	ggml/src/ggml-cpu.o \
 	ggml/src/ggml-alloc.o \
 	ggml/src/ggml-backend.o \
 	ggml/src/ggml-quants.o \
@@ -935,7 +940,6 @@ OBJ_COMMON = \
 	common/console.o \
 	common/ngram-cache.o \
 	common/sampling.o \
-	common/train.o \
 	common/build-info.o \
 	common/json-schema-to-grammar.o
 
@@ -1047,6 +1051,12 @@ ggml/src/ggml.o: \
 	ggml/include/ggml.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
+ggml/src/ggml-cpu.o: \
+	ggml/src/ggml-cpu.c \
+	ggml/include/ggml.h \
+	ggml/src/ggml-common.h
+	$(CC)  $(CFLAGS)   -c $< -o $@
+
 ggml/src/ggml-alloc.o: \
 	ggml/src/ggml-alloc.c \
 	ggml/include/ggml.h \
@@ -1087,6 +1097,19 @@ ggml/src/llamafile/sgemm.o: \
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 endif # GGML_NO_LLAMAFILE
 
+ifndef GGML_NO_AMX
+ggml/src/ggml-amx.o: \
+	ggml/src/ggml-amx.cpp \
+	ggml/include/ggml-amx.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+
+ggml/src/ggml-amx/mmq.o: \
+	ggml/src/ggml-amx/mmq.cpp \
+	ggml/src/ggml-amx/mmq.h \
+	ggml/include/ggml.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
+endif
+
 ifdef GGML_RPC
 ggml/src/ggml-rpc.o: \
 	ggml/src/ggml-rpc.cpp \
@@ -1199,11 +1222,6 @@ common/json-schema-to-grammar.o: \
 	common/json-schema-to-grammar.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-common/train.o: \
-	common/train.cpp \
-	common/train.h
-	$(CXX) $(CXXFLAGS) -c $< -o $@
-
 common/ngram-cache.o: \
 	common/ngram-cache.cpp \
 	common/ngram-cache.h
@@ -1238,6 +1256,7 @@ clean:
 	rm -vrf ggml/src/ggml-metal-embed.metal
 	rm -vrf ggml/src/ggml-cuda/*.o
 	rm -vrf ggml/src/ggml-cuda/template-instances/*.o
+	rm -vrf ggml/src/ggml-amx/*.o
 	rm -rvf $(BUILD_TARGETS)
 	rm -rvf $(TEST_TARGETS)
 	rm -f vulkan-shaders-gen ggml/src/ggml-vulkan-shaders.hpp ggml/src/ggml-vulkan-shaders.cpp
@@ -1273,6 +1292,11 @@ llama-simple: examples/simple/simple.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
+llama-simple-chat: examples/simple-chat/simple-chat.cpp \
+	$(OBJ_ALL)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
 llama-tokenize: examples/tokenize/tokenize.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
@@ -1370,11 +1394,6 @@ llama-bench: examples/llama-bench/llama-bench.cpp \
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-llama-baby-llama: examples/baby-llama/baby-llama.cpp \
-	$(OBJ_ALL)
-	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
-	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
-
 llama-export-lora: examples/export-lora/export-lora.cpp \
 	$(OBJ_ALL)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
@@ -1440,22 +1459,13 @@ llama-server: \
 	examples/server/server.cpp \
 	examples/server/utils.hpp \
 	examples/server/httplib.h \
-	examples/server/colorthemes.css.hpp \
-	examples/server/style.css.hpp \
-	examples/server/theme-beeninorder.css.hpp \
-	examples/server/theme-ketivah.css.hpp \
-	examples/server/theme-mangotango.css.hpp \
-	examples/server/theme-playground.css.hpp \
-	examples/server/theme-polarnight.css.hpp \
-	examples/server/theme-snowstorm.css.hpp \
 	examples/server/index.html.hpp \
-	examples/server/index-new.html.hpp \
-	examples/server/index.js.hpp \
 	examples/server/completion.js.hpp \
-	examples/server/system-prompts.js.hpp \
-	examples/server/prompt-formats.js.hpp \
-	examples/server/json-schema-to-grammar.mjs.hpp \
 	examples/server/loading.html.hpp \
+	examples/server/deps_daisyui.min.css.hpp \
+	examples/server/deps_markdown-it.js.hpp \
+	examples/server/deps_tailwindcss.js.hpp \
+	examples/server/deps_vue.esm-browser.js.hpp \
 	common/json.hpp \
 	common/stb_image.h \
 	$(OBJ_ALL)

Package.swift

@@ -10,6 +10,7 @@ var sources = [
     "src/unicode.cpp",
     "src/unicode-data.cpp",
     "ggml/src/ggml.c",
+    "ggml/src/ggml-cpu.c",
     "ggml/src/ggml-alloc.c",
     "ggml/src/ggml-backend.cpp",
     "ggml/src/ggml-quants.c",
@@ -60,13 +61,15 @@ let package = Package(
             name: "llama",
             path: ".",
             exclude: [
+               "build",
                "cmake",
                "examples",
                "scripts",
                "models",
                "tests",
                "CMakeLists.txt",
-               "Makefile"
+               "Makefile",
+               "ggml/src/ggml-metal-embed.metal"
             ],
             sources: sources,
             resources: resources,

README.md

@@ -17,7 +17,8 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
 
 ## Hot topics
 
-- **Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669**
+- **Introducing GGUF-my-LoRA** https://github.com/ggerganov/llama.cpp/discussions/10123
+- Hugging Face Inference Endpoints now support GGUF out of the box! https://github.com/ggerganov/llama.cpp/discussions/9669
 - Hugging Face GGUF editor: [discussion](https://github.com/ggerganov/llama.cpp/discussions/9268) | [tool](https://huggingface.co/spaces/CISCai/gguf-editor)
 
 ----
@@ -29,7 +30,7 @@ variety of hardware - locally and in the cloud.
 
 - Plain C/C++ implementation without any dependencies
 - Apple silicon is a first-class citizen - optimized via ARM NEON, Accelerate and Metal frameworks
-- AVX, AVX2 and AVX512 support for x86 architectures
+- AVX, AVX2, AVX512 and AMX support for x86 architectures
 - 1.5-bit, 2-bit, 3-bit, 4-bit, 5-bit, 6-bit, and 8-bit integer quantization for faster inference and reduced memory use
 - Custom CUDA kernels for running LLMs on NVIDIA GPUs (support for AMD GPUs via HIP and Moore Threads MTT GPUs via MUSA)
 - Vulkan and SYCL backend support
@@ -93,6 +94,7 @@ Typically finetunes of the base models below are supported as well.
 - [x] [FalconMamba Models](https://huggingface.co/collections/tiiuae/falconmamba-7b-66b9a580324dd1598b0f6d4a)
 - [x] [Jais](https://huggingface.co/inceptionai/jais-13b-chat)
 - [x] [Bielik-11B-v2.3](https://huggingface.co/collections/speakleash/bielik-11b-v23-66ee813238d9b526a072408a)
+- [x] [RWKV-6](https://github.com/BlinkDL/RWKV-LM)
 
 (instructions for supporting more models: [HOWTO-add-model.md](./docs/development/HOWTO-add-model.md))
 
@@ -122,6 +124,7 @@ Typically finetunes of the base models below are supported as well.
 - Rust (nicer API): [mdrokz/rust-llama.cpp](https://github.com/mdrokz/rust-llama.cpp)
 - Rust (more direct bindings): [utilityai/llama-cpp-rs](https://github.com/utilityai/llama-cpp-rs)
 - C#/.NET: [SciSharp/LLamaSharp](https://github.com/SciSharp/LLamaSharp)
+- C#/VB.NET (more features - community license): [LM-Kit.NET](https://docs.lm-kit.com/lm-kit-net/index.html)
 - Scala 3: [donderom/llm4s](https://github.com/donderom/llm4s)
 - Clojure: [phronmophobic/llama.clj](https://github.com/phronmophobic/llama.clj)
 - React Native: [mybigday/llama.rn](https://github.com/mybigday/llama.rn)
@@ -130,6 +133,8 @@ Typically finetunes of the base models below are supported as well.
 - Flutter/Dart: [netdur/llama_cpp_dart](https://github.com/netdur/llama_cpp_dart)
 - PHP (API bindings and features built on top of llama.cpp): [distantmagic/resonance](https://github.com/distantmagic/resonance) [(more info)](https://github.com/ggerganov/llama.cpp/pull/6326)
 - Guile Scheme: [guile_llama_cpp](https://savannah.nongnu.org/projects/guile-llama-cpp)
+- Swift [srgtuszy/llama-cpp-swift](https://github.com/srgtuszy/llama-cpp-swift)
+- Swift [ShenghaiWang/SwiftLlama](https://github.com/ShenghaiWang/SwiftLlama)
 
 **UI:**
 
@@ -170,6 +175,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 - [LARS - The LLM & Advanced Referencing Solution](https://github.com/abgulati/LARS) (AGPL)
 - [LLMUnity](https://github.com/undreamai/LLMUnity) (MIT)
 - [Llama Assistant](https://github.com/vietanhdev/llama-assistant) (GPL)
+- [PocketPal AI - An iOS and Android App](https://github.com/a-ghorbani/pocketpal-ai) (MIT)
 
 *(to have a project listed here, it should clearly state that it depends on `llama.cpp`)*
 
@@ -185,6 +191,7 @@ Unless otherwise noted these projects are open-source with permissive licensing:
 
 - [Paddler](https://github.com/distantmagic/paddler) - Stateful load balancer custom-tailored for llama.cpp
 - [GPUStack](https://github.com/gpustack/gpustack) - Manage GPU clusters for running LLMs
+- [llama_cpp_canister](https://github.com/onicai/llama_cpp_canister) - llama.cpp as a smart contract on the Internet Computer, using WebAssembly
 
 **Games:**
 - [Lucy's Labyrinth](https://github.com/MorganRO8/Lucys_Labyrinth) - A simple maze game where agents controlled by an AI model will try to trick you.

ci/run.sh

@@ -39,7 +39,7 @@ SRC=`pwd`
 CMAKE_EXTRA="-DLLAMA_FATAL_WARNINGS=ON"
 
 if [ ! -z ${GG_BUILD_METAL} ]; then
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_METAL=ON -DGGML_METAL_USE_BF16=ON"
 fi
 
 if [ ! -z ${GG_BUILD_CUDA} ]; then
@@ -53,7 +53,7 @@ if [ ! -z ${GG_BUILD_SYCL} ]; then
         exit 1
     fi
 
-    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
+    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_SYCL=1 -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON"
 fi
 
 if [ ! -z ${GG_BUILD_VULKAN} ]; then
@@ -326,36 +326,36 @@ function gg_run_open_llama_7b_v2 {
     ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
     ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
 
-    (time ./bin/llama-cli --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-cli --model ${model_f16}  -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 10 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 99 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state--model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -460,34 +460,34 @@ function gg_run_pythia_1_4b {
     ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
     ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
 
-    (time ./bin/llama-cli --model ${model_f16}  -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-cli --model ${model_q8_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-cli --model ${model_q4_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-cli --model ${model_q4_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-cli --model ${model_q5_0} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-cli --model ${model_q5_1} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-cli --model ${model_q2_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-cli --model ${model_q3_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-cli --model ${model_q4_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-cli --model ${model_q5_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-cli --model ${model_q6_k} -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-cli --model ${model_f16}  -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-cli --model ${model_q8_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-cli --model ${model_q4_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-cli --model ${model_q4_1} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-cli --model ${model_q5_0} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-cli --model ${model_q5_1} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-cli --model ${model_q2_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-cli --model ${model_q3_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-cli --model ${model_q4_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-cli --model ${model_q5_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-cli --model ${model_q6_k} -ngl 99 -c 0 -s 1234 -n 64 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test_60} -ngl 99 -c 128 -b 128 --chunks 1 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state     --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state -fa --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -591,36 +591,36 @@ function gg_run_pythia_2_8b {
     ./bin/llama-quantize ${model_f16} ${model_q5_k} q5_k
     ./bin/llama-quantize ${model_f16} ${model_q6_k} q6_k
 
-    (time ./bin/llama-cli --model ${model_f16}  -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 999 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-cli --model ${model_f16}  -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-cli --model ${model_q8_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-cli --model ${model_q4_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-cli --model ${model_q4_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-cli --model ${model_q5_0} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-cli --model ${model_q5_1} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-cli --model ${model_q2_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-cli --model ${model_q3_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-cli --model ${model_q4_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-cli --model ${model_q5_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-cli --model ${model_q6_k} -t 1 -ngl 99 -c 0 -s 1234 -n 256 --ignore-eos -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
-    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
-    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
-    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
-    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
-    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
-    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
-    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
-    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
+    (time ./bin/llama-perplexity --model ${model_f16}  -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-perplexity --model ${model_q8_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_0.log
+    (time ./bin/llama-perplexity --model ${model_q4_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_1.log
+    (time ./bin/llama-perplexity --model ${model_q5_0} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_0.log
+    (time ./bin/llama-perplexity --model ${model_q5_1} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_1.log
+    (time ./bin/llama-perplexity --model ${model_q2_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q2_k.log
+    (time ./bin/llama-perplexity --model ${model_q3_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q3_k.log
+    (time ./bin/llama-perplexity --model ${model_q4_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q4_k.log
+    (time ./bin/llama-perplexity --model ${model_q5_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q5_k.log
+    (time ./bin/llama-perplexity --model ${model_q6_k} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-tg-q6_k.log
 
-    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 999 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
+    (time ./bin/llama-imatrix --model ${model_f16} -f ${wiki_test} -t 1 -ngl 99 -c 2048 -b 512 --chunks 4 ) 2>&1 | tee -a $OUT/${ci}-imatrix.log
 
-    (time ./bin/llama-save-load-state     -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state -fa -ngl 10 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state     -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
-    (time ./bin/llama-save-load-state -fa -ngl 99 --model ${model_q4_0} ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 10 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0     ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
+    (time ./bin/llama-save-load-state --model ${model_q4_0} -ngl 99 -c 0 -fa ) 2>&1 | tee -a $OUT/${ci}-save-load-state.log
 
     function check_ppl {
         qnt="$1"
@@ -706,8 +706,8 @@ function gg_run_embd_bge_small {
 
     ./bin/llama-quantize ${model_f16} ${model_q8_0} q8_0
 
-    (time ./bin/llama-embedding --model ${model_f16}  -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
-    (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
+    (time ./bin/llama-embedding --model ${model_f16}  -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-f16.log
+    (time ./bin/llama-embedding --model ${model_q8_0} -p "I believe the meaning of life is" -ngl 99 -c 0 ) 2>&1 | tee -a $OUT/${ci}-tg-q8_0.log
 
     set +e
 }
@@ -752,7 +752,7 @@ function gg_run_rerank_tiny {
     model_f16="${path_models}/ggml-model-f16.gguf"
 
     # for this model, the SEP token is "</s>"
-    (time ./bin/llama-embedding --model ${model_f16}  -p "what is panda?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
+    (time ./bin/llama-embedding --model ${model_f16} -p "what is panda?</s></s>hi\nwhat is panda?</s></s>it's a bear\nwhat is panda?</s></s>The giant panda (Ailuropoda melanoleuca), sometimes called a panda bear or simply panda, is a bear species endemic to China." -ngl 99 -c 0 --pooling rank --embd-normalize -1 --verbose-prompt) 2>&1 | tee -a $OUT/${ci}-rk-f16.log
 
     # sample output
     # rerank score 0:    0.029

cmake/arm64-apple-clang.cmake

@@ -0,0 +1,16 @@
+set( CMAKE_SYSTEM_NAME Darwin )
+set( CMAKE_SYSTEM_PROCESSOR arm64 )
+
+set( target arm64-apple-darwin-macho )
+
+set( CMAKE_C_COMPILER    clang )
+set( CMAKE_CXX_COMPILER  clang++ )
+
+set( CMAKE_C_COMPILER_TARGET   ${target} )
+set( CMAKE_CXX_COMPILER_TARGET ${target} )
+
+set( arch_c_flags "-march=armv8.4-a -fvectorize -ffp-model=fast -fno-finite-math-only" )
+set( warn_c_flags "-Wno-format -Wno-unused-variable -Wno-unused-function" )
+
+set( CMAKE_C_FLAGS_INIT   "${arch_c_flags} ${warn_c_flags}" )
+set( CMAKE_CXX_FLAGS_INIT "${arch_c_flags} ${warn_c_flags}" )

common/CMakeLists.txt

@@ -66,8 +66,6 @@ add_library(${TARGET} STATIC
     ngram-cache.h
     sampling.cpp
     sampling.h
-    train.cpp
-    train.h
     )
 
 if (BUILD_SHARED_LIBS)

common/arg.cpp

@@ -128,13 +128,13 @@ static void common_params_handle_model_default(common_params & params) {
             }
             params.hf_file = params.model;
         } else if (params.model.empty()) {
-            params.model = fs_get_cache_file(string_split(params.hf_file, '/').back());
+            params.model = fs_get_cache_file(string_split<std::string>(params.hf_file, '/').back());
         }
     } else if (!params.model_url.empty()) {
         if (params.model.empty()) {
-            auto f = string_split(params.model_url, '#').front();
-            f = string_split(f, '?').front();
-            params.model = fs_get_cache_file(string_split(f, '/').back());
+            auto f = string_split<std::string>(params.model_url, '#').front();
+            f = string_split<std::string>(f, '?').front();
+            params.model = fs_get_cache_file(string_split<std::string>(f, '/').back());
         }
     } else if (params.model.empty()) {
         params.model = DEFAULT_MODEL_PATH;
@@ -251,6 +251,9 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
         for (auto & antiprompt : params.antiprompt) {
             string_process_escapes(antiprompt);
         }
+        for (auto & seq_breaker : params.sparams.dry_sequence_breakers) {
+            string_process_escapes(seq_breaker);
+        }
     }
 
     if (!params.kv_overrides.empty()) {
@@ -879,7 +882,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         {"--samplers"}, "SAMPLERS",
         string_format("samplers that will be used for generation in the order, separated by \';\'\n(default: %s)", sampler_type_names.c_str()),
         [](common_params & params, const std::string & value) {
-            const auto sampler_names = string_split(value, ';');
+            const auto sampler_names = string_split<std::string>(value, ';');
             params.sparams.samplers = common_sampler_types_from_names(sampler_names, true);
         }
     ).set_sparam());
@@ -941,10 +944,17 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_sparam());
     add_opt(common_arg(
-        {"--tfs"}, "N",
-        string_format("tail free sampling, parameter z (default: %.1f, 1.0 = disabled)", (double)params.sparams.tfs_z),
+        {"--xtc-probability"}, "N",
+        string_format("xtc probability (default: %.1f, 0.0 = disabled)", (double)params.sparams.xtc_probability),
         [](common_params & params, const std::string & value) {
-            params.sparams.tfs_z = std::stof(value);
+            params.sparams.xtc_probability = std::stof(value);
+        }
+    ).set_sparam());
+    add_opt(common_arg(
+        {"--xtc-threshold"}, "N",
+        string_format("xtc threshold (default: %.1f, 1.0 = disabled)", (double)params.sparams.xtc_threshold),
+        [](common_params & params, const std::string & value) {
+            params.sparams.xtc_threshold = std::stof(value);
         }
     ).set_sparam());
     add_opt(common_arg(
@@ -983,6 +993,64 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
             params.sparams.penalty_freq = std::stof(value);
         }
     ).set_sparam());
+    add_opt(common_arg(
+        {"--dry-multiplier"}, "N",
+        string_format("set DRY sampling multiplier (default: %.1f, 0.0 = disabled)", (double)params.sparams.dry_multiplier),
+        [](common_params & params, const std::string & value) {
+            params.sparams.dry_multiplier = std::stof(value);
+        }
+    ).set_sparam());
+    add_opt(common_arg(
+        {"--dry-base"}, "N",
+        string_format("set DRY sampling base value (default: %.2f)", (double)params.sparams.dry_base),
+        [](common_params & params, const std::string & value) {
+            float potential_base = std::stof(value);
+            if (potential_base >= 1.0f)
+            {
+                params.sparams.dry_base = potential_base;
+            }
+        }
+    ).set_sparam());
+    add_opt(common_arg(
+        {"--dry-allowed-length"}, "N",
+        string_format("set allowed length for DRY sampling (default: %d)", params.sparams.dry_allowed_length),
+        [](common_params & params, int value) {
+            params.sparams.dry_allowed_length = value;
+        }
+    ).set_sparam());
+    add_opt(common_arg(
+        {"--dry-penalty-last-n"}, "N",
+        string_format("set DRY penalty for the last n tokens (default: %d, 0 = disable, -1 = context size)", params.sparams.dry_penalty_last_n),
+        [](common_params & params, int value) {
+            params.sparams.dry_penalty_last_n = value;
+        }
+    ).set_sparam());
+    add_opt(common_arg(
+        {"--dry-sequence-breaker"}, "STRING",
+        string_format("add sequence breaker for DRY sampling, clearing out default breakers (%s) in the process; use \"none\" to not use any sequence breakers\n",
+            params.sparams.dry_sequence_breakers.empty() ? "none" :
+            std::accumulate(std::next(params.sparams.dry_sequence_breakers.begin()),
+                params.sparams.dry_sequence_breakers.end(),
+                std::string("'") + (params.sparams.dry_sequence_breakers[0] == "\n" ? "\\n" : params.sparams.dry_sequence_breakers[0]) + "'",
+                [](const std::string& a, const std::string& b) {
+                    std::string formatted_b = (b == "\n") ? "\\n" : b;
+                    return a + ", '" + formatted_b + "'";
+                }).c_str()),
+        [](common_params & params, const std::string & value) {
+            static bool defaults_cleared = false;
+
+            if (!defaults_cleared) {
+                params.sparams.dry_sequence_breakers.clear();
+                defaults_cleared = true;
+            }
+
+            if (value == "none") {
+                params.sparams.dry_sequence_breakers.clear();
+            } else {
+                params.sparams.dry_sequence_breakers.emplace_back(value);
+            }
+        }
+    ).set_sparam());
     add_opt(common_arg(
         {"--dynatemp-range"}, "N",
         string_format("dynamic temperature range (default: %.1f, 0.0 = disabled)", (double)params.sparams.dynatemp_range),
@@ -999,7 +1067,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_sparam());
     add_opt(common_arg(
         {"--mirostat"}, "N",
-        string_format("use Mirostat sampling.\nTop K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.\n"
+        string_format("use Mirostat sampling.\nTop K, Nucleus and Locally Typical samplers are ignored if used.\n"
         "(default: %d, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0)", params.sparams.mirostat),
         [](common_params & params, int value) {
             params.sparams.mirostat = value;
@@ -1083,7 +1151,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples({LLAMA_EXAMPLE_EMBEDDING, LLAMA_EXAMPLE_RETRIEVAL, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_POOLING"));
     add_opt(common_arg(
-        {"--attention"}, "{causal,non,causal}",
+        {"--attention"}, "{causal,non-causal}",
         "attention type for embeddings, use model default if unspecified",
         [](common_params & params, const std::string & value) {
             /**/ if (value == "causal") { params.attention_type = LLAMA_ATTENTION_TYPE_CAUSAL; }
@@ -1163,14 +1231,14 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         [](common_params & params, int value) {
             params.grp_attn_n = value;
         }
-    ).set_env("LLAMA_ARG_GRP_ATTN_N"));
+    ).set_env("LLAMA_ARG_GRP_ATTN_N").set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_PASSKEY}));
     add_opt(common_arg(
         {"-gaw", "--grp-attn-w"}, "N",
-        string_format("group-attention width (default: %.1f)", (double)params.grp_attn_w),
+        string_format("group-attention width (default: %d)", params.grp_attn_w),
         [](common_params & params, int value) {
             params.grp_attn_w = value;
         }
-    ).set_env("LLAMA_ARG_GRP_ATTN_W"));
+    ).set_env("LLAMA_ARG_GRP_ATTN_W").set_examples({LLAMA_EXAMPLE_MAIN}));
     add_opt(common_arg(
         {"-dkvc", "--dump-kv-cache"},
         "verbose print of the KV cache",
@@ -1681,7 +1749,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_BENCH}));
     add_opt(common_arg(
         {"--embd-normalize"}, "N",
-        string_format("normalisation for embendings (default: %d) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)", params.embd_normalize),
+        string_format("normalisation for embeddings (default: %d) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)", params.embd_normalize),
         [](common_params & params, int value) {
             params.embd_normalize = value;
         }
@@ -1695,7 +1763,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
     ).set_examples({LLAMA_EXAMPLE_EMBEDDING}));
     add_opt(common_arg(
         {"--embd-separator"}, "STRING",
-        "separator of embendings (default \\n) for example \"<#sep#>\"",
+        "separator of embeddings (default \\n) for example \"<#sep#>\"",
         [](common_params & params, const std::string & value) {
             params.embd_sep = value;
         }
@@ -1789,22 +1857,12 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
         }
     ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_THREADS_HTTP"));
     add_opt(common_arg(
-        {"-spf", "--system-prompt-file"}, "FNAME",
-        "set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications",
-        [](common_params & params, const std::string & value) {
-            std::ifstream file(value);
-            if (!file) {
-                throw std::runtime_error(string_format("error: failed to open file '%s'\n", value.c_str()));
-            }
-            std::string system_prompt;
-            std::copy(
-                        std::istreambuf_iterator<char>(file),
-                        std::istreambuf_iterator<char>(),
-                        std::back_inserter(system_prompt)
-                        );
-            params.system_prompt = system_prompt;
+        {"--cache-reuse"}, "N",
+        string_format("min chunk size to attempt reusing from the cache via KV shifting (default: %d)", params.n_cache_reuse),
+        [](common_params & params, int value) {
+            params.n_cache_reuse = value;
         }
-    ).set_examples({LLAMA_EXAMPLE_SERVER}));
+    ).set_examples({LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CACHE_REUSE"));
     add_opt(common_arg(
         {"--metrics"},
         string_format("enable prometheus compatible metrics endpoint (default: %s)", params.endpoint_metrics ? "enabled" : "disabled"),

common/common.cpp

@@ -416,19 +416,6 @@ std::string string_format(const char * fmt, ...) {
     return std::string(buf.data(), size);
 }
 
-std::vector<std::string> string_split(std::string input, char separator) {
-    std::vector<std::string> parts;
-    size_t separator_pos = input.find(separator);
-    while (separator_pos != std::string::npos) {
-        std::string part = input.substr(0, separator_pos);
-        parts.emplace_back(part);
-        input = input.substr(separator_pos + 1);
-        separator_pos = input.find(separator);
-    }
-    parts.emplace_back(input);
-    return parts;
-}
-
 std::string string_strip(const std::string & str) {
     size_t start = 0;
     size_t end = str.size();
@@ -955,7 +942,7 @@ struct common_init_result common_init_from_params(common_params & params) {
         }
 
         if (llama_model_has_encoder(model)) {
-            llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size(), 0, 0));
+            llama_encode(lctx, llama_batch_get_one(tmp.data(), tmp.size()));
             llama_token decoder_start_token_id = llama_model_decoder_start_token(model);
             if (decoder_start_token_id == -1) {
                 decoder_start_token_id = bos;
@@ -964,7 +951,7 @@ struct common_init_result common_init_from_params(common_params & params) {
             tmp.push_back(decoder_start_token_id);
         }
         if (llama_model_has_decoder(model)) {
-            llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch), 0, 0));
+            llama_decode(lctx, llama_batch_get_one(tmp.data(), std::min(tmp.size(), (size_t) params.n_batch)));
         }
         llama_kv_cache_clear(lctx);
         llama_synchronize(lctx);
@@ -1016,6 +1003,9 @@ static ggml_type kv_cache_type_from_str(const std::string & s) {
     if (s == "f16") {
         return GGML_TYPE_F16;
     }
+    if (s == "bf16") {
+        return GGML_TYPE_BF16;
+    }
     if (s == "q8_0") {
         return GGML_TYPE_Q8_0;
     }
@@ -1035,7 +1025,7 @@ static ggml_type kv_cache_type_from_str(const std::string & s) {
         return GGML_TYPE_Q5_1;
     }
 
-    throw std::runtime_error("Invalid cache type: " + s);
+    throw std::runtime_error("Unsupported cache type: " + s);
 }
 
 struct llama_context_params common_context_params_to_llama(const common_params & params) {
@@ -1047,7 +1037,7 @@ struct llama_context_params common_context_params_to_llama(const common_params &
     cparams.n_ubatch          = params.n_ubatch;
     cparams.n_threads         = params.cpuparams.n_threads;
     cparams.n_threads_batch   = params.cpuparams_batch.n_threads == -1 ?
-                                    params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
+                                params.cpuparams.n_threads : params.cpuparams_batch.n_threads;
     cparams.logits_all        = params.logits_all;
     cparams.embeddings        = params.embedding;
     cparams.rope_scaling_type = params.rope_scaling_type;
@@ -1964,6 +1954,8 @@ void yaml_dump_string_multiline(FILE * stream, const char * prop_name, const cha
 
 void yaml_dump_non_result_info(FILE * stream, const common_params & params, const llama_context * lctx,
                                const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc) {
+    ggml_cpu_init(); // some ARM features are detected at runtime
+
     const auto & sparams = params.sparams;
 
     fprintf(stream, "build_commit: %s\n",        LLAMA_COMMIT);
@@ -2019,6 +2011,10 @@ void yaml_dump_non_result_info(FILE * stream, const common_params & params, cons
     fprintf(stream, "chunks: %d # default: -1 (unlimited)\n", params.n_chunks);
     fprintf(stream, "color: %s # default: false\n", params.use_color ? "true" : "false");
     fprintf(stream, "ctx_size: %d # default: 512\n", params.n_ctx);
+    fprintf(stream, "dry_allowed_length: %d # default: 2\n", sparams.dry_allowed_length);
+    fprintf(stream, "dry_base: %.2f # default: 1.75\n", sparams.dry_base);
+    fprintf(stream, "dry_multiplier: %.1f # default: 0.0\n", sparams.dry_multiplier);
+    fprintf(stream, "dry_penalty_last_n: %d # default: -1 (0 = disable, -1 = context size)\n", sparams.dry_penalty_last_n);
     fprintf(stream, "escape: %s # default: false\n", params.escape ? "true" : "false");
     fprintf(stream, "file: # never logged, see prompt instead. Can still be specified for input.\n");
     fprintf(stream, "frequency_penalty: %f # default: 0.0 \n", sparams.penalty_freq);
@@ -2099,11 +2095,12 @@ void yaml_dump_non_result_info(FILE * stream, const common_params & params, cons
     const std::vector<float> tensor_split_vector(params.tensor_split, params.tensor_split + llama_max_devices());
     yaml_dump_vector_float(stream, "tensor_split", tensor_split_vector);
 
-    fprintf(stream, "tfs: %f # default: 1.0\n", sparams.tfs_z);
     fprintf(stream, "threads: %d # default: %u\n", params.cpuparams.n_threads, std::thread::hardware_concurrency());
     fprintf(stream, "top_k: %d # default: 40\n", sparams.top_k);
     fprintf(stream, "top_p: %f # default: 0.95\n", sparams.top_p);
     fprintf(stream, "min_p: %f # default: 0.0\n", sparams.min_p);
+    fprintf(stream, "xtc_probability: %f # default: 0.0\n", sparams.xtc_probability);
+    fprintf(stream, "xtc_threshold: %f # default: 0.1\n", sparams.xtc_threshold);
     fprintf(stream, "typ_p: %f # default: 1.0\n", sparams.typ_p);
     fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
     fprintf(stream, "display_prompt: %s # default: true\n", params.display_prompt ? "true" : "false");

common/common.h

@@ -84,12 +84,15 @@ enum llama_example {
 
 enum common_sampler_type {
     COMMON_SAMPLER_TYPE_NONE        = 0,
-    COMMON_SAMPLER_TYPE_TOP_K       = 1,
-    COMMON_SAMPLER_TYPE_TOP_P       = 2,
-    COMMON_SAMPLER_TYPE_MIN_P       = 3,
-    COMMON_SAMPLER_TYPE_TFS_Z       = 4,
-    COMMON_SAMPLER_TYPE_TYPICAL_P   = 5,
-    COMMON_SAMPLER_TYPE_TEMPERATURE = 6,
+    COMMON_SAMPLER_TYPE_DRY         = 1,
+    COMMON_SAMPLER_TYPE_TOP_K       = 2,
+    COMMON_SAMPLER_TYPE_TOP_P       = 3,
+    COMMON_SAMPLER_TYPE_MIN_P       = 4,
+  //COMMON_SAMPLER_TYPE_TFS_Z       = 5,
+    COMMON_SAMPLER_TYPE_TYPICAL_P   = 6,
+    COMMON_SAMPLER_TYPE_TEMPERATURE = 7,
+    COMMON_SAMPLER_TYPE_XTC         = 8,
+    COMMON_SAMPLER_TYPE_INFILL      = 9,
 };
 
 // dimensionality reduction methods, used by cvector-generator
@@ -102,35 +105,44 @@ enum dimre_method {
 struct common_sampler_params {
     uint32_t seed = LLAMA_DEFAULT_SEED; // the seed used to initialize llama_sampler
 
-    int32_t n_prev            = 64;    // number of previous tokens to remember
-    int32_t n_probs           = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
-    int32_t min_keep          = 0;     // 0 = disabled, otherwise samplers should return at least min_keep tokens
-    int32_t top_k             = 40;    // <= 0 to use vocab size
-    float   top_p             = 0.95f; // 1.0 = disabled
-    float   min_p             = 0.05f; // 0.0 = disabled
-    float   tfs_z             = 1.00f; // 1.0 = disabled
-    float   typ_p             = 1.00f; // typical_p, 1.0 = disabled
-    float   temp              = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
-    float   dynatemp_range    = 0.00f; // 0.0 = disabled
-    float   dynatemp_exponent = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
-    int32_t penalty_last_n    = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
-    float   penalty_repeat    = 1.00f; // 1.0 = disabled
-    float   penalty_freq      = 0.00f; // 0.0 = disabled
-    float   penalty_present   = 0.00f; // 0.0 = disabled
-    int32_t mirostat          = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
-    float   mirostat_tau      = 5.00f; // target entropy
-    float   mirostat_eta      = 0.10f; // learning rate
-    bool    penalize_nl       = false; // consider newlines as a repeatable token
-    bool    ignore_eos        = false;
-    bool    no_perf           = false; // disable performance metrics
+    int32_t n_prev             = 64;    // number of previous tokens to remember
+    int32_t n_probs            = 0;     // if greater than 0, output the probabilities of top n_probs tokens.
+    int32_t min_keep           = 0;     // 0 = disabled, otherwise samplers should return at least min_keep tokens
+    int32_t top_k              = 40;    // <= 0 to use vocab size
+    float   top_p              = 0.95f; // 1.0 = disabled
+    float   min_p              = 0.05f; // 0.0 = disabled
+    float   xtc_probability    = 0.00f; // 0.0 = disabled
+    float   xtc_threshold      = 0.10f; // > 0.5 disables XTC
+    float   typ_p              = 1.00f; // typical_p, 1.0 = disabled
+    float   temp               = 0.80f; // <= 0.0 to sample greedily, 0.0 to not output probabilities
+    float   dynatemp_range     = 0.00f; // 0.0 = disabled
+    float   dynatemp_exponent  = 1.00f; // controls how entropy maps to temperature in dynamic temperature sampler
+    int32_t penalty_last_n     = 64;    // last n tokens to penalize (0 = disable penalty, -1 = context size)
+    float   penalty_repeat     = 1.00f; // 1.0 = disabled
+    float   penalty_freq       = 0.00f; // 0.0 = disabled
+    float   penalty_present    = 0.00f; // 0.0 = disabled
+    float   dry_multiplier     = 0.0f;  // 0.0 = disabled;      DRY repetition penalty for tokens extending repetition:
+    float   dry_base           = 1.75f; // 0.0 = disabled;      multiplier * base ^ (length of sequence before token - allowed length)
+    int32_t dry_allowed_length = 2;     // tokens extending repetitions beyond this receive penalty
+    int32_t dry_penalty_last_n = -1;    // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
+    int32_t mirostat           = 0;     // 0 = disabled, 1 = mirostat, 2 = mirostat 2.0
+    float   mirostat_tau       = 5.00f; // target entropy
+    float   mirostat_eta       = 0.10f; // learning rate
+    bool    penalize_nl        = false; // consider newlines as a repeatable token
+    bool    ignore_eos         = false;
+    bool    no_perf            = false; // disable performance metrics
+
+    std::vector<std::string> dry_sequence_breakers = {"\n", ":", "\"", "*"};     // default sequence breakers for DRY
+
 
     std::vector<enum common_sampler_type> samplers = {
+        COMMON_SAMPLER_TYPE_DRY,
         COMMON_SAMPLER_TYPE_TOP_K,
-        COMMON_SAMPLER_TYPE_TFS_Z,
         COMMON_SAMPLER_TYPE_TYPICAL_P,
         COMMON_SAMPLER_TYPE_TOP_P,
         COMMON_SAMPLER_TYPE_MIN_P,
-        COMMON_SAMPLER_TYPE_TEMPERATURE
+        COMMON_SAMPLER_TYPE_XTC,
+        COMMON_SAMPLER_TYPE_TEMPERATURE,
     };
 
     std::string grammar; // optional BNF-like grammar to constrain sampling
@@ -143,7 +155,7 @@ struct common_sampler_params {
 
 struct common_params {
     int32_t n_predict             =    -1; // new tokens to predict
-    int32_t n_ctx                 =     0; // context size
+    int32_t n_ctx                 =  4096; // context size
     int32_t n_batch               =  2048; // logical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_ubatch              =   512; // physical batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                =     0; // number of tokens to keep from initial prompt
@@ -268,21 +280,21 @@ struct common_params {
 
     // embedding
     bool embedding         = false; // get only sentence embedding
-    int32_t embd_normalize = 2;     // normalisation for embendings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
+    int32_t embd_normalize = 2;     // normalisation for embeddings (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
     std::string embd_out   = "";    // empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
-    std::string embd_sep   = "\n";  // separator of embendings
+    std::string embd_sep   = "\n";  // separator of embeddings
     bool reranking         = false; // enable reranking support on server
 
     // server params
     int32_t port           = 8080;         // server listens on this network port
     int32_t timeout_read   = 600;          // http read timeout in seconds
     int32_t timeout_write  = timeout_read; // http write timeout in seconds
-    int     n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
+    int32_t n_threads_http = -1;           // number of threads to process HTTP requests (TODO: support threadpool)
+    int32_t n_cache_reuse  = 0;            // min chunk size to reuse from the cache via KV shifting
 
     std::string hostname      = "127.0.0.1";
     std::string public_path   = "";                                                                         // NOLINT
     std::string chat_template = "";                                                                         // NOLINT
-    std::string system_prompt = "";                                                                         // NOLINT
     bool enable_chat_template = true;
 
     std::vector<std::string> api_keys;
@@ -374,8 +386,6 @@ bool set_process_priority(enum ggml_sched_priority prio);
 LLAMA_COMMON_ATTRIBUTE_FORMAT(1, 2)
 std::string string_format(const char * fmt, ...);
 
-std::vector<std::string> string_split(std::string input, char separator);
-
 std::string string_strip(const std::string & str);
 std::string string_get_sortable_timestamp();
 
@@ -383,6 +393,7 @@ void string_replace_all(std::string & s, const std::string & search, const std::
 
 template<class T>
 static std::vector<T> string_split(const std::string & str, char delim) {
+    static_assert(!std::is_same<T, std::string>::value, "Please use the specialized version for std::string");
     std::vector<T> values;
     std::istringstream str_stream(str);
     std::string token;
@@ -395,6 +406,22 @@ static std::vector<T> string_split(const std::string & str, char delim) {
     return values;
 }
 
+template<>
+std::vector<std::string> string_split<std::string>(const std::string & input, char separator)
+{
+    std::vector<std::string> parts;
+    size_t begin_pos = 0;
+    size_t separator_pos = input.find(separator);
+    while (separator_pos != std::string::npos) {
+        std::string part = input.substr(begin_pos, separator_pos - begin_pos);
+        parts.emplace_back(part);
+        begin_pos = separator_pos + 1;
+        separator_pos = input.find(separator, begin_pos);
+    }
+    parts.emplace_back(input.substr(begin_pos, separator_pos - begin_pos));
+    return parts;
+}
+
 bool string_parse_kv_override(const char * data, std::vector<llama_model_kv_override> & overrides);
 void string_process_escapes(std::string & input);
 

common/json-schema-to-grammar.cpp

@@ -611,7 +611,7 @@ private:
             }
             return join_seq();
         };
-        return _add_rule(name, "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space");
+        return _add_rule(name, "\"\\\"\" (" + to_rule(transform()) + ") \"\\\"\" space");
     }
 
     /*

common/sampling.cpp

@@ -130,10 +130,12 @@ std::string common_sampler_params::print() const {
 
     snprintf(result, sizeof(result),
             "\trepeat_last_n = %d, repeat_penalty = %.3f, frequency_penalty = %.3f, presence_penalty = %.3f\n"
-            "\ttop_k = %d, tfs_z = %.3f, top_p = %.3f, min_p = %.3f, typical_p = %.3f, temp = %.3f\n"
+            "\tdry_multiplier = %.3f, dry_base = %.3f, dry_allowed_length = %d, dry_penalty_last_n = %d\n"
+            "\ttop_k = %d, top_p = %.3f, min_p = %.3f, xtc_probability = %.3f, xtc_threshold = %.3f, typical_p = %.3f, temp = %.3f\n"
             "\tmirostat = %d, mirostat_lr = %.3f, mirostat_ent = %.3f",
             penalty_last_n, penalty_repeat, penalty_freq, penalty_present,
-            top_k, tfs_z, top_p, min_p, typ_p, temp,
+            dry_multiplier, dry_base, dry_allowed_length, dry_penalty_last_n,
+            top_k, top_p, min_p, xtc_probability, xtc_threshold, typ_p, temp,
             mirostat, mirostat_eta, mirostat_tau);
 
     return std::string(result);
@@ -171,54 +173,54 @@ struct common_sampler * common_sampler_init(const struct llama_model * model, co
                 params.penalize_nl,
                 params.ignore_eos));
 
-    if (params.temp > 0.0f) {
-        if (params.mirostat == 0) {
-            for (const auto & cnstr : params.samplers) {
-                switch (cnstr) {
-                    case COMMON_SAMPLER_TYPE_TOP_K:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
+    if (params.mirostat == 0) {
+        for (const auto & cnstr : params.samplers) {
+            switch (cnstr) {
+                    case COMMON_SAMPLER_TYPE_DRY:
+                    {
+                        std::vector<const char*> c_breakers;
+                        c_breakers.reserve(params.dry_sequence_breakers.size());
+                        for (const auto& str : params.dry_sequence_breakers) {
+                            c_breakers.push_back(str.c_str());
+                        }
+
+                        llama_sampler_chain_add(result->chain, llama_sampler_init_dry      (model, params.dry_multiplier, params.dry_base, params.dry_allowed_length, params.dry_penalty_last_n, c_breakers.data(), c_breakers.size()));
+                    }
                         break;
-                    case COMMON_SAMPLER_TYPE_TOP_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_MIN_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_TFS_Z:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_tail_free(params.tfs_z, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_TYPICAL_P:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
-                        break;
-                    case COMMON_SAMPLER_TYPE_TEMPERATURE:
-                        llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
-                        break;
-                    default:
-                        GGML_ASSERT(false && "unknown sampler type");
-                }
+                case COMMON_SAMPLER_TYPE_TOP_K:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_k    (params.top_k));
+                    break;
+                case COMMON_SAMPLER_TYPE_TOP_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_top_p    (params.top_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_MIN_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_min_p    (params.min_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_XTC:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_xtc      (params.xtc_probability, params.xtc_threshold, params.min_keep, params.seed));
+                    break;
+                case COMMON_SAMPLER_TYPE_TYPICAL_P:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_typical  (params.typ_p, params.min_keep));
+                    break;
+                case COMMON_SAMPLER_TYPE_TEMPERATURE:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_temp_ext (params.temp, params.dynatemp_range, params.dynatemp_exponent));
+                    break;
+                case COMMON_SAMPLER_TYPE_INFILL:
+                    llama_sampler_chain_add(result->chain, llama_sampler_init_infill   (model));
+                    break;
+                default:
+                    GGML_ASSERT(false && "unknown sampler type");
             }
-            llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
-            llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
-        } else if (params.mirostat == 1) {
-            llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
-            llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_n_vocab(model), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
-        } else if (params.mirostat == 2) {
-            llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
-            llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));
-        } else {
-            GGML_ASSERT(false && "unknown mirostat version");
         }
+        llama_sampler_chain_add(result->chain, llama_sampler_init_dist(params.seed));
+    } else if (params.mirostat == 1) {
+        llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
+        llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat(llama_n_vocab(model), params.seed, params.mirostat_tau, params.mirostat_eta, 100));
+    } else if (params.mirostat == 2) {
+        llama_sampler_chain_add(result->chain, llama_sampler_init_temp(params.temp));
+        llama_sampler_chain_add(result->chain, llama_sampler_init_mirostat_v2(params.seed, params.mirostat_tau, params.mirostat_eta));
     } else {
-        if (params.n_probs > 0) {
-            // some use cases require to sample greedily, but still obtain the probabilities of the top tokens
-            // ref: https://github.com/ggerganov/llama.cpp/pull/9605
-            //
-            // the following will not produce exactly the same probs as applyging softmax to the full vocabulary, but
-            // it is much faster, since we avoid sorting all tokens and should give a good approximation
-            llama_sampler_chain_add(result->chain, llama_sampler_init_top_k(params.n_probs));
-            llama_sampler_chain_add(result->chain, llama_sampler_init_softmax());
-        }
-        llama_sampler_chain_add(result->chain, llama_sampler_init_greedy());
+        GGML_ASSERT(false && "unknown mirostat version");
     }
 
     return result;
@@ -366,36 +368,42 @@ std::string common_sampler_prev_str(common_sampler * gsmpl, llama_context * ctx_
 
 char common_sampler_type_to_chr(enum common_sampler_type cnstr) {
     switch (cnstr) {
+        case COMMON_SAMPLER_TYPE_DRY:         return 'd';
         case COMMON_SAMPLER_TYPE_TOP_K:       return 'k';
-        case COMMON_SAMPLER_TYPE_TFS_Z:       return 'f';
         case COMMON_SAMPLER_TYPE_TYPICAL_P:   return 'y';
         case COMMON_SAMPLER_TYPE_TOP_P:       return 'p';
         case COMMON_SAMPLER_TYPE_MIN_P:       return 'm';
         case COMMON_SAMPLER_TYPE_TEMPERATURE: return 't';
+        case COMMON_SAMPLER_TYPE_XTC:         return 'x';
+        case COMMON_SAMPLER_TYPE_INFILL:      return 'i';
         default : return '?';
     }
 }
 
 std::string common_sampler_type_to_str(enum common_sampler_type cnstr) {
     switch (cnstr) {
+        case COMMON_SAMPLER_TYPE_DRY:         return "dry";
         case COMMON_SAMPLER_TYPE_TOP_K:       return "top_k";
-        case COMMON_SAMPLER_TYPE_TFS_Z:       return "tfs_z";
         case COMMON_SAMPLER_TYPE_TYPICAL_P:   return "typ_p";
         case COMMON_SAMPLER_TYPE_TOP_P:       return "top_p";
         case COMMON_SAMPLER_TYPE_MIN_P:       return "min_p";
         case COMMON_SAMPLER_TYPE_TEMPERATURE: return "temperature";
+        case COMMON_SAMPLER_TYPE_XTC:         return "xtc";
+        case COMMON_SAMPLER_TYPE_INFILL:      return "infill";
         default : return "";
     }
 }
 
 std::vector<common_sampler_type> common_sampler_types_from_names(const std::vector<std::string> & names, bool allow_alt_names) {
     std::unordered_map<std::string, common_sampler_type> sampler_canonical_name_map {
+        { "dry",         COMMON_SAMPLER_TYPE_DRY },
         { "top_k",       COMMON_SAMPLER_TYPE_TOP_K },
         { "top_p",       COMMON_SAMPLER_TYPE_TOP_P },
         { "typ_p",       COMMON_SAMPLER_TYPE_TYPICAL_P },
         { "min_p",       COMMON_SAMPLER_TYPE_MIN_P },
-        { "tfs_z",       COMMON_SAMPLER_TYPE_TFS_Z },
         { "temperature", COMMON_SAMPLER_TYPE_TEMPERATURE },
+        { "xtc",         COMMON_SAMPLER_TYPE_XTC },
+        { "infill",      COMMON_SAMPLER_TYPE_INFILL },
     };
 
     // since samplers names are written multiple ways
@@ -409,8 +417,6 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
         { "typ-p",       COMMON_SAMPLER_TYPE_TYPICAL_P },
         { "typ",         COMMON_SAMPLER_TYPE_TYPICAL_P },
         { "min-p",       COMMON_SAMPLER_TYPE_MIN_P },
-        { "tfs-z",       COMMON_SAMPLER_TYPE_TFS_Z },
-        { "tfs",         COMMON_SAMPLER_TYPE_TFS_Z },
         { "temp",        COMMON_SAMPLER_TYPE_TEMPERATURE },
     };
 
@@ -436,12 +442,14 @@ std::vector<common_sampler_type> common_sampler_types_from_names(const std::vect
 
 std::vector<common_sampler_type> common_sampler_types_from_chars(const std::string & chars) {
     std::unordered_map<char, common_sampler_type> sampler_name_map = {
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_DRY),         COMMON_SAMPLER_TYPE_DRY },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_K),       COMMON_SAMPLER_TYPE_TOP_K },
-        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TFS_Z),       COMMON_SAMPLER_TYPE_TFS_Z },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TYPICAL_P),   COMMON_SAMPLER_TYPE_TYPICAL_P },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TOP_P),       COMMON_SAMPLER_TYPE_TOP_P },
         { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_MIN_P),       COMMON_SAMPLER_TYPE_MIN_P },
-        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE }
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_TEMPERATURE), COMMON_SAMPLER_TYPE_TEMPERATURE },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_XTC),         COMMON_SAMPLER_TYPE_XTC },
+        { common_sampler_type_to_chr(COMMON_SAMPLER_TYPE_INFILL),      COMMON_SAMPLER_TYPE_INFILL },
     };
 
     std::vector<common_sampler_type> samplers;

common/train.h

@@ -1,233 +0,0 @@
-// Various helper functions and utilities for training
-
-#pragma once
-
-#include <string>
-#include <random>
-#include <vector>
-
-#include "ggml.h"
-#include "llama.h"
-
-#define LLAMA_TRAIN_MAX_NODES 16384
-
-typedef std::string mt19937_state;
-
-struct train_state {
-    struct ggml_opt_context * opt;
-
-    uint64_t train_its;
-    uint64_t train_samples;
-    uint64_t train_tokens;
-    uint64_t train_epochs;
-
-    size_t        shuffle_samples_hash; // fn, sample_count, *zip(sample_begins, sample_sizes)
-    mt19937_state shuffle_rng_state_current;
-    mt19937_state shuffle_rng_state_next;
-    size_t        shuffle_sample_count;
-    size_t        shuffle_next_sample;
-};
-
-struct train_params_common {
-    const char * fn_train_data;
-    const char * fn_checkpoint_in;
-    const char * fn_checkpoint_out;
-    const char * pattern_fn_it;
-    const char * fn_latest;
-
-    bool print_usage;
-
-    int save_every;
-
-    uint32_t seed;
-
-    int n_ctx;
-    int n_threads;
-    int n_batch;
-    int n_gradient_accumulation;
-    int n_epochs;
-    int n_gpu_layers;
-
-    bool custom_n_ctx;
-
-    bool use_flash;
-    bool use_checkpointing;
-
-    std::string sample_start;
-    bool include_sample_start;
-    bool escape;
-    bool overlapping_samples;
-    bool fill_with_next_samples;
-    bool separate_with_eos;
-    bool separate_with_bos;
-    bool sample_random_offsets;
-
-    bool force_reshuffle;
-
-    int   warmup;
-    int   cos_decay_steps;
-    float cos_decay_restart;
-    float cos_decay_min;
-    bool  enable_restart;
-
-    int   opt_past;
-    float opt_delta;
-    int   opt_max_no_improvement;
-
-    int   adam_n_iter;
-    float adam_alpha;
-    float adam_min_alpha;
-    float adam_decay;
-    int   adam_decay_min_ndim;
-    float adam_beta1;
-    float adam_beta2;
-    float adam_gclip;
-    float adam_eps_f;
-};
-
-typedef void (*save_train_files_callback)(void * data, struct train_state * train);
-
-struct train_opt_callback_data {
-    struct train_params_common * params;
-    struct train_state         * train;
-    save_train_files_callback    save_cb;
-    void                       * save_data;
-    struct llama_context       * lctx;
-    int                          last_save_iter;
-    llama_token                * tokens_data;
-    size_t                       tokens_size;
-    size_t                     * samples_begin;
-    size_t                     * samples_size;
-    size_t                     * shuffled_samples_offs;
-    size_t                     * shuffled_samples_begin;
-    size_t                     * shuffled_samples_size;
-    size_t                       samples_count;
-    struct ggml_tensor         * tokens_input;
-    struct ggml_tensor         * target_probs;
-    int                          first_iter;
-    int                          first_epoch;
-    int                          iter_at_last_epoch;
-    int64_t                      last_time;
-    double                       millis_per_iter;
-};
-
-struct train_state * init_train_state();
-void free_train_state(struct train_state  * state);
-
-struct train_params_common get_default_train_params_common();
-void print_common_train_usage(int /*argc*/, char ** argv, const struct train_params_common * params);
-
-bool consume_common_train_arg(int argc, char ** argv, int * idx, struct train_params_common * params, bool * invalid_param);
-void finish_processing_train_args(struct train_params_common * params);
-
-struct random_normal_distribution;
-struct random_uniform_distribution;
-
-struct random_normal_distribution  * init_random_normal_distribution (int seed, float mean, float std, float min, float max);
-struct random_uniform_distribution * init_random_uniform_distribution(int seed, float min, float max);
-
-void free_random_normal_distribution (struct random_normal_distribution  * rnd);
-void free_random_uniform_distribution(struct random_uniform_distribution * rnd);
-
-struct ggml_tensor * randomize_tensor_normal (struct ggml_tensor * tensor, struct random_normal_distribution * rnd);
-struct ggml_tensor * randomize_tensor_uniform(struct ggml_tensor * tensor, struct random_uniform_distribution * rnd);
-
-// generate random float in interval [0,1)
-float frand();
-float frand_normal (struct random_normal_distribution * rnd);
-float frand_uniform(struct random_uniform_distribution * rnd);
-
-int   clamp (const int v, const int min, const int max);
-float fclamp(const float v, const float min, const float max);
-
-void assert_shape_1d(struct ggml_tensor * tensor, int64_t ne0);
-void assert_shape_2d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1);
-void assert_shape_3d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2);
-void assert_shape_4d(struct ggml_tensor * tensor, int64_t ne0, int64_t ne1, int64_t ne2, int64_t ne3);
-
-size_t tokenize_file(
-        struct llama_context     * lctx,
-        const char               * filename,
-        const std::string        & sample_start,
-        bool                       include_sample_start,
-        bool                       overlapping_samples,
-        unsigned                   context_length,
-        std::vector<llama_token> & out_tokens,
-        std::vector<size_t>      & out_samples_begin,
-        std::vector<size_t>      & out_samples_size);
-
-int64_t get_example_targets_batch(
-        struct llama_context * lctx,
-        struct ggml_tensor   * tokens_input,
-        struct ggml_tensor   * target_probs,
-        int64_t                example_id,
-        const size_t         * samples_offs,
-        const size_t         * samples_begin,
-        const size_t         * samples_size,
-              size_t           samples_count,
-        const llama_token    * train_data,
-        size_t                 n_train_data,
-        bool                   separate_with_eos,
-        bool                   separate_with_bos,
-        bool                   fill_with_next_samples,
-        bool                   sample_random_offsets);
-
-
-void          mt19937_set_state(std::mt19937& rng, const mt19937_state& rng_state);
-mt19937_state mt19937_get_state(const std::mt19937& rng);
-mt19937_state mt19937_seed_to_state(unsigned seed);
-
-mt19937_state shuffle_samples(
-        const mt19937_state & rng_state,
-        size_t              * shuffled_offs,
-        size_t              * shuffled_begins,
-        size_t              * shuffled_sizes,
-        const size_t        * begins,
-        const size_t        * sizes,
-        size_t                count);
-
-size_t hash_combine(size_t h1, size_t h2);
-
-size_t compute_samples_hash(
-    const char* fn,
-    const size_t* samples_begin,
-    const size_t* samples_size,
-    size_t sample_count);
-
-
-std::string replace_str(const char * s, const char * needle, const char * replacement);
-
-void print_duration(double milliseconds);
-
-float cosine_decay(
-    int64_t step,
-    int64_t decay_steps,
-    float   minimum);
-
-float cosine_decay_restart(
-    int64_t step,
-    int64_t decay_steps,
-    float   minimum,
-    float   restart_step_mult);
-
-float learning_schedule(
-    int64_t step,
-    int64_t warmup_steps,
-    int64_t decay_steps,
-    float   learning_rate,
-    float   overall_minimum,
-    float   cos_decay_minimum,
-    float   cos_decay_restart_step_mult,
-    bool    enable_restart);
-
-void copy_tensor_by_name(struct ggml_tensor * dst, struct ggml_context * ctx, const char * name);
-
-void load_opt_context_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct ggml_opt_context * opt);
-void save_opt_context_gguf(struct gguf_context * fctx, struct ggml_opt_context * opt);
-
-bool load_train_state_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct train_state * train);
-void save_train_state_gguf(struct gguf_context * fctx, struct train_state * train);
-
-std::string get_train_filename(const char * filename, const char * pattern_it, const char * latest, int64_t iteration);
-
-void train_opt_callback(void * vdata, int accum_step, float * sched, bool * cancel);

convert_hf_to_gguf.py

@@ -72,7 +72,8 @@ class Model:
     def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path, is_big_endian: bool = False,
                  use_temp_file: bool = False, eager: bool = False,
                  metadata_override: Path | None = None, model_name: str | None = None,
-                 split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False, small_first_shard: bool = False):
+                 split_max_tensors: int = 0, split_max_size: int = 0, dry_run: bool = False,
+                 small_first_shard: bool = False, hparams: dict[str, Any] | None = None):
         if type(self) is Model:
             raise TypeError(f"{type(self).__name__!r} should not be directly instantiated")
 
@@ -87,7 +88,7 @@ class Model:
         self.is_safetensors = len(self.part_names) > 0
         if not self.is_safetensors:
             self.part_names = Model.get_model_part_names(self.dir_model, "pytorch_model", ".bin")
-        self.hparams = Model.load_hparams(self.dir_model)
+        self.hparams = Model.load_hparams(self.dir_model) if hparams is None else hparams
         self.block_count = self.find_hparam(["n_layers", "num_hidden_layers", "n_layer", "num_layers"])
         self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
         self.tensor_names = None
@@ -573,6 +574,9 @@ class Model:
         if chkhsh == "0876d13b50744004aa9aeae05e7b0647eac9d801b5ba4668afc01e709c15e19f":
             # ref: https://huggingface.co/BAAI/bge-small-en-v1.5
             res = "bert-bge"
+        if chkhsh == "8e62295832751ca1e8f92f2226f403dea30dc5165e448b5bfa05af5340c64ec7":
+            # ref: https://huggingface.co/BAAI/bge-large-zh-v1.5
+            res = "bert-bge-large"
         if chkhsh == "b6dc8df998e1cfbdc4eac8243701a65afe638679230920b50d6f17d81c098166":
             # ref: https://huggingface.co/mosaicml/mpt-7b
             res = "mpt"
@@ -1538,6 +1542,17 @@ class LlamaModel(Model):
             special_vocab._set_special_token("eot",    32010)
             special_vocab.add_to_gguf(self.gguf_writer)
 
+        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
+        if tokenizer_config_file.is_file():
+            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
+                tokenizer_config_json = json.load(f)
+                if "add_prefix_space" in tokenizer_config_json:
+                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
+
+        # Apply to granite small models only
+        if self.hparams.get("vocab_size", 32000) == 49152:
+            self.gguf_writer.add_add_bos_token(False)
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         hparams = self.hparams
@@ -1554,17 +1569,6 @@ class LlamaModel(Model):
                 self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR)
                 self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"])
 
-        tokenizer_config_file = self.dir_model / 'tokenizer_config.json'
-        if tokenizer_config_file.is_file():
-            with open(tokenizer_config_file, "r", encoding="utf-8") as f:
-                tokenizer_config_json = json.load(f)
-                if "add_prefix_space" in tokenizer_config_json:
-                    self.gguf_writer.add_add_space_prefix(tokenizer_config_json["add_prefix_space"])
-
-        # Apply to granite small models only
-        if self.hparams.get("vocab_size", 32000) == 49152:
-            self.gguf_writer.add_add_bos_token(False)
-
     @staticmethod
     def permute(weights: Tensor, n_head: int, n_head_kv: int | None):
         if n_head_kv is not None and n_head != n_head_kv:
@@ -2864,6 +2868,9 @@ class Rwkv6Model(Model):
         self.gguf_writer.add_token_list(tokens)
         self.gguf_writer.add_token_types(toktypes)
         special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)
+        special_vocab.chat_template = "rwkv-world"
+        # hack: Add '\n\n' as the EOT token to make it chat normally
+        special_vocab._set_special_token("eot", 261)
         special_vocab.add_to_gguf(self.gguf_writer)
 
     def set_gguf_parameters(self):
@@ -3741,10 +3748,7 @@ class JaisModel(Model):
 
         # Embeddings scale
         self.embeddings_scale = 1.0
-        # note: For some JAIS flavors, output is tied to (same as) wte in original model
-        self.output_is_wte = False
         if 'mup_embeddings_scale' in self.hparams:
-            self.output_is_wte = True   # Hack (?)
             self.embeddings_scale = self.hparams['mup_embeddings_scale']
         elif 'embeddings_scale' in self.hparams:
             self.embeddings_scale = self.hparams['embeddings_scale']
@@ -3801,10 +3805,7 @@ class JaisModel(Model):
 
         if new_name == self.format_tensor_name(gguf.MODEL_TENSOR.TOKEN_EMBD):
             tensors.append((new_name, data_torch * self.embeddings_scale))
-            if self.output_is_wte:
-                tensors.append((self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT), data_torch * self.width_scale))
         elif new_name == self.format_tensor_name(gguf.MODEL_TENSOR.OUTPUT):
-            assert not self.output_is_wte
             tensors.append((new_name, data_torch * self.width_scale))
         else:
             tensors.append((new_name, data_torch))

convert_hf_to_gguf_update.py

@@ -72,6 +72,7 @@ models = [
     {"name": "deepseek-coder", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-base", },
     {"name": "falcon",         "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/tiiuae/falcon-7b", },
     {"name": "bert-bge",       "tokt": TOKENIZER_TYPE.WPM, "repo": "https://huggingface.co/BAAI/bge-small-en-v1.5", },
+    {"name": "bert-bge-large", "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/BAAI/bge-large-zh-v1.5", },
     {"name": "mpt",            "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/mosaicml/mpt-7b", },
     {"name": "starcoder",      "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/bigcode/starcoder2-3b", },
     {"name": "gpt-2",          "tokt": TOKENIZER_TYPE.BPE, "repo": "https://huggingface.co/openai-community/gpt2", },

convert_lora_to_gguf.py

@@ -12,6 +12,7 @@ import json
 from math import prod
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, Callable, Iterable, Iterator, Sequence, SupportsIndex, cast
+from transformers import AutoConfig
 
 import torch
 
@@ -230,7 +231,7 @@ def get_base_tensor_name(lora_tensor_name: str) -> str:
 
 def parse_args() -> argparse.Namespace:
     parser = argparse.ArgumentParser(
-        description="Convert a huggingface PEFT LoRA adapter to a GGML compatible file")
+        description="Convert a Hugging Face PEFT LoRA adapter to a GGUF file")
     parser.add_argument(
         "--outfile", type=Path,
         help="path to write to; default: based on input. {ftype} will be replaced by the outtype.",
@@ -256,17 +257,23 @@ def parse_args() -> argparse.Namespace:
         help="only print out what will be done, without writing any new files",
     )
     parser.add_argument(
-        "--base", type=Path, required=True,
-        help="directory containing base model file",
+        "--base", type=Path,
+        help="directory containing Hugging Face model config files (config.json, tokenizer.json) for the base model that the adapter is based on - only config is needed, actual model weights are not required. If base model is unspecified, it will be loaded from Hugging Face hub based on the adapter config",
     )
     parser.add_argument(
         "lora_path", type=Path,
-        help="directory containing LoRA adapter file",
+        help="directory containing Hugging Face PEFT LoRA config (adapter_model.json) and weights (adapter_model.safetensors or adapter_model.bin)",
     )
 
     return parser.parse_args()
 
 
+def load_hparams_from_hf(hf_model_id: str) -> dict[str, Any]:
+    # normally, adapter does not come with base model config, we need to load it from AutoConfig
+    config = AutoConfig.from_pretrained(hf_model_id)
+    return config.to_dict()
+
+
 if __name__ == '__main__':
     args = parse_args()
     logging.basicConfig(level=logging.DEBUG if args.verbose else logging.INFO)
@@ -281,7 +288,7 @@ if __name__ == '__main__':
 
     ftype = ftype_map[args.outtype]
 
-    dir_base_model: Path = args.base
+    dir_base_model: Path | None = args.base
     dir_lora: Path = args.lora_path
     lora_config = dir_lora / "adapter_config.json"
     input_model = dir_lora / "adapter_model.safetensors"
@@ -301,9 +308,29 @@ if __name__ == '__main__':
         input_model = os.path.join(dir_lora, "adapter_model.bin")
         lora_model = torch.load(input_model, map_location="cpu", weights_only=True)
 
+    # load LoRA config
+    with open(lora_config, "r") as f:
+        lparams: dict[str, Any] = json.load(f)
+
     # load base model
-    logger.info(f"Loading base model: {dir_base_model.name}")
-    hparams = Model.load_hparams(dir_base_model)
+    if dir_base_model is None:
+        if "base_model_name_or_path" in lparams:
+            model_id = lparams["base_model_name_or_path"]
+            logger.info(f"Loading base model from Hugging Face: {model_id}")
+            try:
+                hparams = load_hparams_from_hf(model_id)
+            except OSError as e:
+                logger.error(f"Failed to load base model config: {e}")
+                logger.error("Please try downloading the base model and add its path to --base")
+                sys.exit(1)
+        else:
+            logger.error("'base_model_name_or_path' is not found in adapter_config.json")
+            logger.error("Base model config is required. Please download the base model and add its path to --base")
+            sys.exit(1)
+    else:
+        logger.info(f"Loading base model: {dir_base_model.name}")
+        hparams = Model.load_hparams(dir_base_model)
+
     with torch.inference_mode():
         try:
             model_class = Model.from_model_architecture(hparams["architectures"][0])
@@ -323,13 +350,15 @@ if __name__ == '__main__':
                 self.dir_model_card = dir_lora_model
                 self.lora_alpha = float(lora_alpha)
 
+            def set_vocab(self):
+                pass
+
             def set_type(self):
                 self.gguf_writer.add_type(gguf.GGUFType.ADAPTER)
                 self.gguf_writer.add_string(gguf.Keys.Adapter.TYPE, "lora")
 
             def set_gguf_parameters(self):
                 self.gguf_writer.add_float32(gguf.Keys.Adapter.LORA_ALPHA, self.lora_alpha)
-                super().set_gguf_parameters()
 
             def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:
                 # Never add extra tensors (e.g. rope_freqs) for LoRA adapters
@@ -348,6 +377,9 @@ if __name__ == '__main__':
                         if ".base_layer.weight" in name:
                             continue
                         logger.error(f"Unexpected name '{name}': Not a lora_A or lora_B tensor")
+                        if ".embed_tokens.weight" in name or ".lm_head.weight" in name:
+                            logger.error("Embeddings is present in the adapter. This can be due to new tokens added during fine tuning")
+                            logger.error("Please refer to https://github.com/ggerganov/llama.cpp/pull/9948")
                         sys.exit(1)
 
                     if base_name in tensor_map:
@@ -381,9 +413,6 @@ if __name__ == '__main__':
                     yield (dest_name + ".lora_a", lora_a)
                     yield (dest_name + ".lora_b", lora_b)
 
-        with open(lora_config, "r") as f:
-            lparams: dict[str, Any] = json.load(f)
-
         alpha: float = lparams["lora_alpha"]
 
         model_instance = LoraModel(
@@ -396,6 +425,7 @@ if __name__ == '__main__':
             dry_run=args.dry_run,
             dir_lora_model=dir_lora,
             lora_alpha=alpha,
+            hparams=hparams,
         )
 
         logger.info("Exporting model...")

docs/backend/SYCL.md

@@ -377,7 +377,7 @@ found 2 SYCL devices:
 
 |Chosen Device ID|Setting|
 |-|-|
-|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"` or no action|
+|0|`export ONEAPI_DEVICE_SELECTOR="level_zero:0"` or no action|
 |1|`export ONEAPI_DEVICE_SELECTOR="level_zero:1"`|
 |0 & 1|`export ONEAPI_DEVICE_SELECTOR="level_zero:0;level_zero:1"`|
 

examples/CMakeLists.txt

@@ -13,7 +13,6 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
 if (EMSCRIPTEN)
 else()
     add_subdirectory(cvector-generator)
-    add_subdirectory(baby-llama)
     add_subdirectory(batched-bench)
     add_subdirectory(batched)
     add_subdirectory(convert-llama2c-to-ggml)
@@ -49,6 +48,7 @@ else()
     endif()
     add_subdirectory(save-load-state)
     add_subdirectory(simple)
+    add_subdirectory(simple-chat)
     add_subdirectory(speculative)
     add_subdirectory(tokenize)
 endif()

examples/baby-llama/CMakeLists.txt

@@ -1,5 +0,0 @@
-set(TARGET llama-baby-llama)
-add_executable(${TARGET} baby-llama.cpp)
-install(TARGETS ${TARGET} RUNTIME)
-target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
-target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/batched-bench/batched-bench.cpp

@@ -74,7 +74,6 @@ int main(int argc, char ** argv) {
                 batch.n_seq_id + i,
                 batch.seq_id   + i,
                 batch.logits   + i,
-                0, 0, 0, // unused
             };
 
             const int ret = llama_decode(ctx, batch_view);

examples/chat-persistent.sh

@@ -23,8 +23,9 @@ CUR_PROMPT_CACHE="${CHAT_SAVE_DIR}/current-cache.bin"
 NEXT_PROMPT_FILE="${CHAT_SAVE_DIR}/next-prompt.txt"
 NEXT_PROMPT_CACHE="${CHAT_SAVE_DIR}/next-cache.bin"
 
-SESSION_SIZE_MSG_PATTERN='main: session file matches [[:digit:]]+ / [[:digit:]]+'
-SAMPLE_TIME_MSG_PATTERN='sample time =[[:space:]]+[[:digit:]]+.[[:digit:]]+ ms /[[:space:]]+[[:digit:]]+'
+SESSION_AND_SAMPLE_PATTERN='main: session file matches [[:digit:]]+ / [[:digit:]]+'\
+'|'\
+'sampling time =[[:space:]]+[[:digit:]]+.[[:digit:]]+ ms /[[:space:]]+[[:digit:]]+'
 SED_DELETE_MESSAGES="/^(${USER_NAME}:|${AI_NAME}:|\\.\\.\\.)/,\$d"
 
 CTX_SIZE=2048
@@ -129,15 +130,12 @@ while read -e line; do
 
     printf ' '
 
-    # HACK get num tokens from debug message
-    # TODO get both messages in one go
-    if  ! session_size_msg="$(tail -n30 "$LOG" | grep -oE "$SESSION_SIZE_MSG_PATTERN")" ||
-        ! sample_time_msg="$(tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then
+    if ! session_and_sample_msg=$(tail -n30 "$LOG" | grep -oE "$SESSION_AND_SAMPLE_PATTERN"); then
         echo >&2 "Couldn't get number of tokens from ./llama-cli output!"
         exit 1
     fi
 
-    n_tokens=$(($(cut -d/ -f2 <<<"$session_size_msg") + $(cut -d/ -f2 <<<"$sample_time_msg")))
+    n_tokens=$(awk '{sum+=$1} END {print sum}' <<< "$(cut -d/ -f2 <<< "$session_and_sample_msg")")
 
     if ((n_tokens > CTX_ROTATE_POINT)); then
         tail -c+$((n_prompt_len_pre + 1)) "$CUR_PROMPT_FILE" >>"$NEXT_PROMPT_FILE"

examples/cvector-generator/cvector-generator.cpp

@@ -339,7 +339,7 @@ static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) {
 
 static bool get_hidden_layers(llama_context * ctx, std::vector<llama_token> & tokens) {
     llama_kv_cache_clear(ctx);
-    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
+    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
         fprintf(stderr, "%s : failed to eval\n", __func__);
         return false;
     }

examples/eval-callback/eval-callback.cpp

@@ -131,7 +131,7 @@ static bool run(llama_context * ctx, const common_params & params) {
 
     std::vector<llama_token> tokens = common_tokenize(ctx, params.prompt, add_bos);
 
-    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) {
+    if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size()))) {
         LOG_ERR("%s : failed to eval\n", __func__);
         return false;
     }

examples/imatrix/imatrix.cpp

@@ -496,6 +496,8 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
         // clear the KV cache
         llama_kv_cache_clear(ctx);
 
+        llama_batch batch = llama_batch_init(n_batch, 0, 1);
+
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
@@ -508,9 +510,14 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
                 tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
-            // TODO: use batch.logits to save computations instead of relying on logits_all == true
-            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
+            common_batch_clear(batch);
+            for (int i = 0; i < batch_size; i++) {
+                common_batch_add(batch, tokens[batch_start + i], j*n_batch + i, {0}, true);
+            }
+
+            if (llama_decode(ctx, batch)) {
                 LOG_ERR("%s : failed to eval\n", __func__);
+                llama_batch_free(batch);
                 return false;
             }
 
@@ -523,6 +530,8 @@ static bool compute_imatrix(llama_context * ctx, const common_params & params) {
             }
         }
 
+        llama_batch_free(batch);
+
         const auto t_end = std::chrono::high_resolution_clock::now();
 
         if (i == 0) {

examples/infill/infill.cpp

@@ -396,7 +396,7 @@ int main(int argc, char ** argv) {
 
                 LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
 
-                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
                     LOG_ERR("%s : failed to eval\n", __func__);
                     return 1;
                 }

examples/json_schema_to_grammar.py

@@ -540,7 +540,7 @@ class SchemaConverter:
         return self._add_rule(
             name,
             to_rule(transform()) if self._raw_pattern \
-                else "\"\\\"\" " + to_rule(transform()) + " \"\\\"\" space")
+                else "\"\\\"\" (" + to_rule(transform()) + ") \"\\\"\" space")
 
 
     def _resolve_ref(self, ref):

examples/llama-bench/llama-bench.cpp

@@ -21,12 +21,6 @@
 #include "ggml.h"
 #include "llama.h"
 #include "common.h"
-#include "ggml-cuda.h"
-#include "ggml-sycl.h"
-
-#ifdef GGML_USE_CANN
-#include "ggml-cann.h"
-#endif
 
 #ifdef _WIN32
 #define WIN32_LEAN_AND_MEAN
@@ -82,95 +76,27 @@ static T stdev(const std::vector<T> & v) {
 }
 
 static std::string get_cpu_info() {
-    std::string id;
-#ifdef __linux__
-    FILE * f = fopen("/proc/cpuinfo", "r");
-    if (f) {
-        char buf[1024];
-        while (fgets(buf, sizeof(buf), f)) {
-            if (strncmp(buf, "model name", 10) == 0) {
-                char * p = strchr(buf, ':');
-                if (p) {
-                    p++;
-                    while (std::isspace(*p)) {
-                        p++;
-                    }
-                    while (std::isspace(p[strlen(p) - 1])) {
-                        p[strlen(p) - 1] = '\0';
-                    }
-                    id = p;
-                    break;
-                }
-            }
-        }
-        fclose(f);
-    }
-#elif defined(_WIN32)
-    HKEY hKey;
-    if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
-                     TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
-                     0,
-                     KEY_READ,
-                     &hKey) != ERROR_SUCCESS) {
-        // fail to open registry key
-        return "";
-    }
-    char cpu_brand[256];
-    DWORD cpu_brand_size = sizeof(cpu_brand);
-    if (RegQueryValueExA(hKey,
-                        TEXT("ProcessorNameString"),
-                        NULL,
-                        NULL,
-                        (LPBYTE)cpu_brand,
-                        &cpu_brand_size) == ERROR_SUCCESS) {
-        id.assign(cpu_brand, cpu_brand_size);
-        if (id.find('\0') != std::string::npos) {
-            id.resize(id.find('\0'));
+    std::vector<std::string> cpu_list;
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        auto * dev = ggml_backend_dev_get(i);
+        auto dev_type = ggml_backend_dev_type(dev);
+        if (dev_type == GGML_BACKEND_DEVICE_TYPE_CPU || dev_type == GGML_BACKEND_DEVICE_TYPE_ACCEL) {
+            cpu_list.push_back(ggml_backend_dev_description(dev));
         }
     }
-    RegCloseKey(hKey);
-#endif
-    // TODO: other platforms
-    return id;
+    return join(cpu_list, ", ");
 }
 
 static std::string get_gpu_info() {
-    std::string id;
-#ifdef GGML_USE_CUDA
-    int count = ggml_backend_cuda_get_device_count();
-    for (int i = 0; i < count; i++) {
-        char buf[128];
-        ggml_backend_cuda_get_device_description(i, buf, sizeof(buf));
-        id += buf;
-        if (i < count - 1) {
-            id += "/";
+    std::vector<std::string> gpu_list;
+    for (size_t i = 0; i < ggml_backend_dev_count(); i++) {
+        auto * dev = ggml_backend_dev_get(i);
+        auto dev_type = ggml_backend_dev_type(dev);
+        if (dev_type == GGML_BACKEND_DEVICE_TYPE_GPU) {
+            gpu_list.push_back(ggml_backend_dev_description(dev));
         }
     }
-#endif
-#ifdef GGML_USE_SYCL
-    int count = ggml_backend_sycl_get_device_count();
-    for (int i = 0; i < count; i++) {
-        char buf[128];
-        ggml_sycl_get_device_description(i, buf, sizeof(buf));
-        id += buf;
-        if (i < count - 1) {
-            id += "/";
-        }
-    }
-#endif
-#ifdef GGML_USE_CANN
-    uint32_t count = ggml_backend_cann_get_device_count();
-    for (uint32_t i = 0; i < count; i++) {
-        char buf[128];
-        ggml_backend_cann_get_device_description(i, buf, sizeof(buf));
-        id += buf;
-        if (i < count - 1) {
-            id += "/";
-        }
-    }
-#endif
-    // TODO: other backends
-    return id;
+    return join(gpu_list, ", ");
 }
 
 // command line params
@@ -330,6 +256,9 @@ static ggml_type ggml_type_from_name(const std::string & s) {
     if (s == "f16") {
         return GGML_TYPE_F16;
     }
+    if (s == "bf16") {
+        return GGML_TYPE_BF16;
+    }
     if (s == "q8_0") {
         return GGML_TYPE_Q8_0;
     }
@@ -938,29 +867,15 @@ struct test {
     }
 
     static std::string get_backend() {
-        if (cuda) {
-            return GGML_CUDA_NAME;
+        std::vector<std::string> backends;
+        for (size_t i = 0; i < ggml_backend_reg_count(); i++) {
+            auto * reg = ggml_backend_reg_get(i);
+            std::string name = ggml_backend_reg_name(reg);
+            if (name != "CPU") {
+                backends.push_back(ggml_backend_reg_name(reg));
+            }
         }
-        if (vulkan) {
-            return "Vulkan";
-        }
-        if (kompute) {
-            return "Kompute";
-        }
-        if (metal) {
-            return "Metal";
-        }
-        if (sycl) {
-            return GGML_SYCL_NAME;
-        }
-        if (gpu_blas) {
-            return "GPU BLAS";
-        }
-        if (blas) {
-            return "BLAS";
-        }
-
-        return "CPU";
+        return backends.empty() ? "CPU" : join(backends, ",");
     }
 
     static const std::vector<std::string> & get_fields() {
@@ -1428,7 +1343,7 @@ struct sql_printer : public printer {
     }
 };
 
-static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_batch, int n_threads) {
+static void test_prompt(llama_context * ctx, int n_prompt, int n_batch, int n_threads) {
     llama_set_n_threads(ctx, n_threads, n_threads);
 
     const llama_model * model = llama_get_model(ctx);
@@ -1444,14 +1359,14 @@ static void test_prompt(llama_context * ctx, int n_prompt, int n_past, int n_bat
         for (int i = 1; i < n_tokens; i++) {
             tokens[i] = std::rand() % n_vocab;
         }
-        llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens, n_past + n_processed, 0));
+        llama_decode(ctx, llama_batch_get_one(tokens.data(), n_tokens));
         n_processed += n_tokens;
     }
 
     llama_synchronize(ctx);
 }
 
-static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads) {
+static void test_gen(llama_context * ctx, int n_gen, int n_threads) {
     llama_set_n_threads(ctx, n_threads, n_threads);
 
     const llama_model * model = llama_get_model(ctx);
@@ -1460,7 +1375,7 @@ static void test_gen(llama_context * ctx, int n_gen, int n_past, int n_threads)
     llama_token token = llama_add_bos_token(model) ? llama_token_bos(model) : std::rand() % n_vocab;
 
     for (int i = 0; i < n_gen; i++) {
-        llama_decode(ctx, llama_batch_get_one(&token, 1, n_past + i, 0));
+        llama_decode(ctx, llama_batch_get_one(&token, 1));
         llama_synchronize(ctx);
         token = std::rand() % n_vocab;
     }
@@ -1596,13 +1511,13 @@ int main(int argc, char ** argv) {
                 fprintf(stderr, "llama-bench: benchmark %d/%ld: warmup prompt run\n", params_idx, params_count);
             }
             //test_prompt(ctx, std::min(t.n_batch, std::min(t.n_prompt, 32)), 0, t.n_batch, t.n_threads);
-            test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
+            test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
         }
         if (t.n_gen > 0) {
             if (params.progress) {
                 fprintf(stderr, "llama-bench: benchmark %d/%ld: warmup generation run\n", params_idx, params_count);
             }
-            test_gen(ctx, 1, 0, t.n_threads);
+            test_gen(ctx, 1, t.n_threads);
         }
 
         for (int i = 0; i < params.reps; i++) {
@@ -1614,13 +1529,13 @@ int main(int argc, char ** argv) {
                 if (params.progress) {
                     fprintf(stderr, "llama-bench: benchmark %d/%ld: prompt run %d/%d\n", params_idx, params_count, i + 1, params.reps);
                 }
-                test_prompt(ctx, t.n_prompt, 0, t.n_batch, t.n_threads);
+                test_prompt(ctx, t.n_prompt, t.n_batch, t.n_threads);
             }
             if (t.n_gen > 0) {
                 if (params.progress) {
                     fprintf(stderr, "llama-bench: benchmark %d/%ld: generation run %d/%d\n", params_idx, params_count, i + 1, params.reps);
                 }
-                test_gen(ctx, t.n_gen, t.n_prompt, t.n_threads);
+                test_gen(ctx, t.n_gen, t.n_threads);
             }
 
             uint64_t t_ns = get_time_ns() - t_start;

examples/llama.android/llama/src/main/cpp/llama-android.cpp

@@ -283,9 +283,6 @@ Java_android_llama_cpp_LLamaAndroid_new_1batch(JNIEnv *, jobject, jint n_tokens,
         nullptr,
         nullptr,
         nullptr,
-        0,
-        0,
-        0,
     };
 
     if (embd) {

examples/llama.swiftui/llama.cpp.swift/LibLlama.swift

@@ -46,7 +46,6 @@ actor LlamaContext {
         let sparams = llama_sampler_chain_default_params()
         self.sampling = llama_sampler_chain_init(sparams)
         llama_sampler_chain_add(self.sampling, llama_sampler_init_temp(0.4))
-        llama_sampler_chain_add(self.sampling, llama_sampler_init_softmax())
         llama_sampler_chain_add(self.sampling, llama_sampler_init_dist(1234))
     }
 

examples/llama.vim

@@ -0,0 +1,783 @@
+" LLM-based text completion using llama.cpp
+"
+" requires:
+"
+"   - neovim or vim
+"   - curl
+"   - llama.cpp server instance
+"   - FIM-compatible model
+"
+" sample config:
+"
+"   - Tab       - accept the current suggestion
+"   - Shift+Tab - accept just the first line of the suggestion
+"   - Ctrl+F    - toggle FIM completion manually
+"
+" make symlink or copy this file to ~/.config/nvim/autoload/llama.vim
+"
+" start the llama.cpp server with a FIM-compatible model. for example:
+"
+"   $ llama-server -m {model.gguf} --port 8012 -ngl 99 -fa -dt 0.1 --ubatch-size 512 --batch-size 1024 --cache-reuse 256
+"
+"   --batch-size [512, model max context]
+"
+"     adjust the batch size to control how much of the provided local context will be used during the inference
+"     lower values will use smaller part of the context around the cursor, which will result in faster processing
+"
+"   --ubatch-size [64, 2048]
+"
+"     chunks the batch into smaller chunks for faster processing
+"     depends on the specific hardware. use llama-bench to profile and determine the best size
+"
+"   --cache-reuse (ge:llama_config.n_predict, 1024]
+"
+"     this should be either 0 (disabled) or strictly larger than g:llama_config.n_predict
+"     using non-zero value enables context reuse on the server side which dramatically improves the performance at
+"     large contexts. a value of 256 should be good for all cases
+"
+" run this once to initialise llama.vim:
+"
+"   :call llama#init()
+"
+" more info: https://github.com/ggerganov/llama.cpp/pull/9787
+"
+
+" colors (adjust to your liking)
+highlight llama_hl_hint guifg=#ff772f ctermfg=202
+highlight llama_hl_info guifg=#77ff2f ctermfg=119
+
+" general parameters:
+"
+"   endpoint:         llama.cpp server endpoint
+"   n_prefix:         number of lines before the cursor location to include in the local prefix
+"   n_suffix:         number of lines after  the cursor location to include in the local suffix
+"   n_predict:        max number of tokens to predict
+"   t_max_prompt_ms:  max alloted time for the prompt processing (TODO: not yet supported)
+"   t_max_predict_ms: max alloted time for the prediction
+"   show_info:        show extra info about the inference (0 - disabled, 1 - statusline, 2 - inline)
+"   auto_fim:         trigger FIM completion automatically on cursor movement
+"   max_line_suffix:  do not auto-trigger FIM completion if there are more than this number of characters to the right of the cursor
+"
+" ring buffer of chunks, accumulated with time upon:
+"
+"  - completion request
+"  - yank
+"  - entering a buffer
+"  - leaving a buffer
+"  - writing a file
+"
+" parameters for the ring-buffer with extra context:
+"
+"   ring_n_chunks:    max number of chunks to pass as extra context to the server (0 to disable)
+"   ring_chunk_size:  max size of the chunks (in number of lines)
+"                     note: adjust these numbers so that you don't overrun your context
+"                           at ring_n_chunks = 64 and ring_chunk_size = 64 you need ~32k context
+"   ring_scope:       the range around the cursor position (in number of lines) for gathering chunks after FIM
+"   ring_update_ms:   how often to process queued chunks in normal mode
+"
+let s:default_config = {
+    \ 'endpoint':         'http://127.0.0.1:8012/infill',
+    \ 'n_prefix':         256,
+    \ 'n_suffix':         64,
+    \ 'n_predict':        128,
+    \ 't_max_prompt_ms':  500,
+    \ 't_max_predict_ms': 3000,
+    \ 'show_info':        2,
+    \ 'auto_fim':         v:true,
+    \ 'max_line_suffix':  8,
+    \ 'ring_n_chunks':    64,
+    \ 'ring_chunk_size':  64,
+    \ 'ring_scope':       1024,
+    \ 'ring_update_ms':   1000,
+    \ }
+
+let g:llama_config = get(g:, 'llama_config', s:default_config)
+
+function! s:get_indent(str)
+    let l:count = 0
+    for i in range(len(a:str))
+        if a:str[i] == "\t"
+            let l:count += &tabstop - 1
+        else
+            break
+        endif
+    endfor
+    return l:count
+endfunction
+
+function! s:rand(i0, i1) abort
+    return a:i0 + rand() % (a:i1 - a:i0 + 1)
+endfunction
+
+function! llama#init()
+    if !executable('curl')
+        echohl WarningMsg
+        echo 'llama.vim requires the "curl" command to be available'
+        echohl None
+        return
+    endif
+
+    let s:pos_x = 0 " cursor position upon start of completion
+    let s:pos_y = 0
+
+    let s:line_cur = ''
+
+    let s:line_cur_prefix = ''
+    let s:line_cur_suffix = ''
+
+    let s:ring_chunks = [] " current set of chunks used as extra context
+    let s:ring_queued = [] " chunks that are queued to be sent for processing
+    let s:ring_n_evict = 0
+
+    let s:hint_shown = v:false
+    let s:pos_y_pick = -9999 " last y where we picked a chunk
+    let s:pos_dx = 0
+    let s:content = []
+    let s:can_accept = v:false
+
+    let s:timer_fim = -1
+    let s:t_fim_start = reltime() " used to measure total FIM time
+    let s:t_last_move = reltime() " last time the cursor moved
+
+    let s:current_job = v:null
+
+    let s:ghost_text_nvim = exists('*nvim_buf_get_mark')
+    let s:ghost_text_vim = has('textprop')
+
+    if s:ghost_text_vim
+        let s:hlgroup_hint = 'llama_hl_hint'
+        let s:hlgroup_info = 'llama_hl_info'
+
+        if empty(prop_type_get(s:hlgroup_hint))
+            call prop_type_add(s:hlgroup_hint, {'highlight': s:hlgroup_hint})
+        endif
+        if empty(prop_type_get(s:hlgroup_info))
+            call prop_type_add(s:hlgroup_info, {'highlight': s:hlgroup_info})
+        endif
+    endif
+
+    augroup llama
+        autocmd!
+        autocmd InsertEnter     * inoremap <expr> <silent> <C-F> llama#fim_inline(v:false)
+        autocmd InsertLeavePre  * call llama#fim_cancel()
+
+        autocmd CursorMoved     * call s:on_move()
+        autocmd CursorMovedI    * call s:on_move()
+        autocmd CompleteChanged * call llama#fim_cancel()
+
+        if g:llama_config.auto_fim
+            autocmd CursorMovedI * call llama#fim(v:true)
+        endif
+
+        " gather chunks upon yanking
+        autocmd TextYankPost    * if v:event.operator ==# 'y' | call s:pick_chunk(v:event.regcontents, v:false, v:true) | endif
+
+        " gather chunks upon entering/leaving a buffer
+        autocmd BufEnter        * call timer_start(100, {-> s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)})
+        autocmd BufLeave        * call                      s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)
+
+        " gather chunk upon saving the file
+        autocmd BufWritePost    * call s:pick_chunk(getline(max([1, line('.') - g:llama_config.ring_chunk_size/2]), min([line('.') + g:llama_config.ring_chunk_size/2, line('$')])), v:true, v:true)
+    augroup END
+
+    silent! call llama#fim_cancel()
+
+    " init background update of the ring buffer
+    if g:llama_config.ring_n_chunks > 0
+        call s:ring_update()
+    endif
+endfunction
+
+" compute how similar two chunks of text are
+" 0 - no similarity, 1 - high similarity
+" TODO: figure out something better
+function! s:chunk_sim(c0, c1)
+    let l:lines0 = len(a:c0)
+    let l:lines1 = len(a:c1)
+
+    let l:common = 0
+
+    for l:line0 in a:c0
+        for l:line1 in a:c1
+            if l:line0 == l:line1
+                let l:common += 1
+                break
+            endif
+        endfor
+    endfor
+
+    return 2.0 * l:common / (l:lines0 + l:lines1)
+endfunction
+
+" pick a random chunk of size g:llama_config.ring_chunk_size from the provided text and queue it for processing
+"
+" no_mod   - do not pick chunks from buffers with pending changes
+" do_evict - evict chunks that are very similar to the new one
+"
+function! s:pick_chunk(text, no_mod, do_evict)
+    " do not pick chunks from buffers with pending changes or buffers that are not files
+    if a:no_mod && (getbufvar(bufnr('%'), '&modified') || !buflisted(bufnr('%')) || !filereadable(expand('%')))
+        return
+    endif
+
+    " if the extra context option is disabled - do nothing
+    if g:llama_config.ring_n_chunks <= 0
+        return
+    endif
+
+    " don't pick very small chunks
+    if len(a:text) < 3
+        return
+    endif
+
+    if len(a:text) + 1 < g:llama_config.ring_chunk_size
+        let l:chunk = a:text
+    else
+        let l:l0 = s:rand(0, max([0, len(a:text) - g:llama_config.ring_chunk_size/2]))
+        let l:l1 = min([l:l0 + g:llama_config.ring_chunk_size/2, len(a:text)])
+
+        let l:chunk = a:text[l:l0:l:l1]
+    endif
+
+    let l:chunk_str = join(l:chunk, "\n") . "\n"
+
+    " check if this chunk is already added
+    let l:exist = v:false
+
+    for i in range(len(s:ring_chunks))
+        if s:ring_chunks[i].data == l:chunk
+            let l:exist = v:true
+            break
+        endif
+    endfor
+
+    for i in range(len(s:ring_queued))
+        if s:ring_queued[i].data == l:chunk
+            let l:exist = v:true
+            break
+        endif
+    endfor
+
+    if l:exist
+        return
+    endif
+
+    " evict queued chunks that are very similar to the new one
+    for i in range(len(s:ring_queued) - 1, 0, -1)
+        if s:chunk_sim(s:ring_queued[i].data, l:chunk) > 0.9
+            if a:do_evict
+                call remove(s:ring_queued, i)
+                let s:ring_n_evict += 1
+            else
+                return
+            endif
+        endif
+    endfor
+
+    " also from s:ring_chunks
+    for i in range(len(s:ring_chunks) - 1, 0, -1)
+        if s:chunk_sim(s:ring_chunks[i].data, l:chunk) > 0.9
+            if a:do_evict
+                call remove(s:ring_chunks, i)
+                let s:ring_n_evict += 1
+            else
+                return
+            endif
+        endif
+    endfor
+
+    " TODO: become parameter ?
+    if len(s:ring_queued) == 16
+        call remove(s:ring_queued, 0)
+    endif
+
+    call add(s:ring_queued, {'data': l:chunk, 'str': l:chunk_str, 'time': reltime(), 'filename': expand('%')})
+
+    "let &statusline = 'extra context: ' . len(s:ring_chunks) . ' / ' . len(s:ring_queued)
+endfunction
+
+" picks a queued chunk, sends it for processing and adds it to s:ring_chunks
+" called every g:llama_config.ring_update_ms
+function! s:ring_update()
+    call timer_start(g:llama_config.ring_update_ms, {-> s:ring_update()})
+
+    " update only if in normal mode or if the cursor hasn't moved for a while
+    if mode() !=# 'n' && reltimefloat(reltime(s:t_last_move)) < 3.0
+        return
+    endif
+
+    if len(s:ring_queued) == 0
+        return
+    endif
+
+    " move the first queued chunk to the ring buffer
+    if len(s:ring_chunks) == g:llama_config.ring_n_chunks
+        call remove(s:ring_chunks, 0)
+    endif
+
+    call add(s:ring_chunks, remove(s:ring_queued, 0))
+
+    "let &statusline = 'updated context: ' . len(s:ring_chunks) . ' / ' . len(s:ring_queued)
+
+    " send asynchronous job with the new extra context so that it is ready for the next FIM
+    let l:extra_context = []
+    for l:chunk in s:ring_chunks
+        call add(l:extra_context, {
+            \ 'text':     l:chunk.str,
+            \ 'time':     l:chunk.time,
+            \ 'filename': l:chunk.filename
+            \ })
+    endfor
+
+    " no samplers needed here
+    let l:request = json_encode({
+        \ 'input_prefix':     "",
+        \ 'input_suffix':     "",
+        \ 'input_extra':      l:extra_context,
+        \ 'prompt':           "",
+        \ 'n_predict':        1,
+        \ 'temperature':      0.0,
+        \ 'stream':           v:false,
+        \ 'samplers':         ["temperature"],
+        \ 'cache_prompt':     v:true,
+        \ 't_max_prompt_ms':  1,
+        \ 't_max_predict_ms': 1
+        \ })
+
+    let l:curl_command = [
+        \ "curl",
+        \ "--silent",
+        \ "--no-buffer",
+        \ "--request", "POST",
+        \ "--url", g:llama_config.endpoint,
+        \ "--header", "Content-Type: application/json",
+        \ "--data", l:request
+        \ ]
+
+    " no callbacks because we don't need to process the response
+    if s:ghost_text_nvim
+        call jobstart(l:curl_command, {})
+    elseif s:ghost_text_vim
+        call job_start(l:curl_command, {})
+    endif
+endfunction
+
+" necessary for 'inoremap <expr>'
+function! llama#fim_inline(is_auto) abort
+    call llama#fim(a:is_auto)
+    return ''
+endfunction
+
+" the main FIM call
+" takes local context around the cursor and sends it together with the extra context to the server for completion
+function! llama#fim(is_auto) abort
+    " we already have a suggestion for the current cursor position
+    if s:hint_shown && !a:is_auto
+        call llama#fim_cancel()
+        return
+    endif
+
+    call llama#fim_cancel()
+
+    " avoid sending repeated requests too fast
+    if reltimefloat(reltime(s:t_fim_start)) < 0.6
+        if s:timer_fim != -1
+            call timer_stop(s:timer_fim)
+            let s:timer_fim = -1
+        endif
+
+        let s:t_fim_start = reltime()
+        let s:timer_fim = timer_start(600, {-> llama#fim(v:true)})
+        return
+    endif
+
+    let s:t_fim_start = reltime()
+
+    let s:content = []
+    let s:can_accept = v:false
+
+    let s:pos_x = col('.') - 1
+    let s:pos_y = line('.')
+    let l:max_y = line('$')
+
+    let l:lines_prefix = getline(max([1, s:pos_y - g:llama_config.n_prefix]), s:pos_y - 1)
+    let l:lines_suffix = getline(s:pos_y + 1, min([l:max_y, s:pos_y + g:llama_config.n_suffix]))
+
+    let s:line_cur = getline('.')
+
+    let s:line_cur_prefix = strpart(s:line_cur, 0, s:pos_x)
+    let s:line_cur_suffix = strpart(s:line_cur, s:pos_x)
+
+    if a:is_auto && len(s:line_cur_suffix) > g:llama_config.max_line_suffix
+        return
+    endif
+
+    let l:prefix = ""
+        \ . join(l:lines_prefix, "\n")
+        \ . "\n"
+
+    let l:prompt = ""
+        \ . s:line_cur_prefix
+
+    let l:suffix = ""
+        \ . s:line_cur_suffix
+        \ . "\n"
+        \ . join(l:lines_suffix, "\n")
+        \ . "\n"
+
+    " prepare the extra context data
+    let l:extra_context = []
+    for l:chunk in s:ring_chunks
+        call add(l:extra_context, {
+            \ 'text':     l:chunk.str,
+            \ 'time':     l:chunk.time,
+            \ 'filename': l:chunk.filename
+            \ })
+    endfor
+
+    " the indentation of the current line
+    let l:indent = strlen(matchstr(s:line_cur_prefix, '^\s*'))
+
+    let l:request = json_encode({
+        \ 'input_prefix':     l:prefix,
+        \ 'input_suffix':     l:suffix,
+        \ 'input_extra':      l:extra_context,
+        \ 'prompt':           l:prompt,
+        \ 'n_predict':        g:llama_config.n_predict,
+        \ 'n_indent':         l:indent,
+        \ 'top_k':            40,
+        \ 'top_p':            0.99,
+        \ 'stream':           v:false,
+        \ 'samplers':         ["top_k", "top_p", "infill"],
+        \ 'cache_prompt':     v:true,
+        \ 't_max_prompt_ms':  g:llama_config.t_max_prompt_ms,
+        \ 't_max_predict_ms': g:llama_config.t_max_predict_ms
+        \ })
+
+    let l:curl_command = [
+        \ "curl",
+        \ "--silent",
+        \ "--no-buffer",
+        \ "--request", "POST",
+        \ "--url", g:llama_config.endpoint,
+        \ "--header", "Content-Type: application/json",
+        \ "--data", l:request
+        \ ]
+
+    if s:current_job != v:null
+        if s:ghost_text_nvim
+            call jobstop(s:current_job)
+        elseif s:ghost_text_vim
+            call job_stop(s:current_job)
+        endif
+    endif
+
+    " send the request asynchronously
+    if s:ghost_text_nvim
+        let s:current_job = jobstart(l:curl_command, {
+            \ 'on_stdout': function('s:fim_on_stdout', [s:pos_x, s:pos_y, a:is_auto]),
+            \ 'on_exit':   function('s:fim_on_exit'),
+            \ 'stdout_buffered': v:true
+            \ })
+    elseif s:ghost_text_vim
+        let s:current_job = job_start(l:curl_command, {
+            \ 'out_cb': function('s:fim_on_stdout', [s:pos_x, s:pos_y, a:is_auto]),
+            \ 'exit_cb':   function('s:fim_on_exit')
+            \ })
+    endif
+
+    " TODO: per-file location
+    let l:delta_y = abs(s:pos_y - s:pos_y_pick)
+
+    " gather some extra context nearby and process it in the background
+    " only gather chunks if the cursor has moved a lot
+    " TODO: something more clever? reranking?
+    if a:is_auto && l:delta_y > 32
+        " expand the prefix even further
+        call s:pick_chunk(getline(max([1,       s:pos_y - g:llama_config.ring_scope]), max([1,       s:pos_y - g:llama_config.n_prefix])), v:false, v:false)
+
+        " pick a suffix chunk
+        call s:pick_chunk(getline(min([l:max_y, s:pos_y + g:llama_config.n_suffix]),   min([l:max_y, s:pos_y + g:llama_config.n_suffix + g:llama_config.ring_chunk_size])), v:false, v:false)
+
+        let s:pos_y_pick = s:pos_y
+    endif
+endfunction
+
+" if first_line == v:true accept only the first line of the response
+function! llama#fim_accept(first_line)
+    " insert the suggestion at the cursor location
+    if s:can_accept && len(s:content) > 0
+        call setline(s:pos_y, s:line_cur[:(s:pos_x - 1)] . s:content[0])
+        if len(s:content) > 1
+            if !a:first_line
+                call append(s:pos_y, s:content[1:-1])
+            endif
+        endif
+
+        " move the cursor to the end of the accepted text
+        if !a:first_line && len(s:content) > 1
+            call cursor(s:pos_y + len(s:content) - 1, s:pos_x + s:pos_dx + 1)
+        else
+            call cursor(s:pos_y, s:pos_x + len(s:content[0]))
+        endif
+    endif
+
+    call llama#fim_cancel()
+endfunction
+
+function! llama#fim_cancel()
+    let s:hint_shown = v:false
+
+    " clear the virtual text
+    let l:bufnr = bufnr('%')
+
+    if s:ghost_text_nvim
+        let l:id_vt_fim = nvim_create_namespace('vt_fim')
+        call nvim_buf_clear_namespace(l:bufnr, l:id_vt_fim,  0, -1)
+    elseif s:ghost_text_vim
+        call prop_remove({'type': s:hlgroup_hint, 'all': v:true})
+        call prop_remove({'type': s:hlgroup_info, 'all': v:true})
+    endif
+
+    " remove the mappings
+    silent! iunmap <buffer> <Tab>
+    silent! iunmap <buffer> <S-Tab>
+    silent! iunmap <buffer> <Esc>
+endfunction
+
+function! s:on_move()
+    let s:t_last_move = reltime()
+
+    call llama#fim_cancel()
+endfunction
+
+" callback that processes the FIM result from the server and displays the suggestion
+function! s:fim_on_stdout(pos_x, pos_y, is_auto, job_id, data, event = v:null)
+    if s:ghost_text_nvim
+        let l:raw = join(a:data, "\n")
+    elseif s:ghost_text_vim
+        let l:raw = a:data
+    endif
+
+    if len(l:raw) == 0
+        return
+    endif
+
+    if a:pos_x != col('.') - 1 || a:pos_y != line('.')
+        return
+    endif
+
+    " show the suggestion only in insert mode
+    if mode() !=# 'i'
+        return
+    endif
+
+    let s:pos_x = a:pos_x
+    let s:pos_y = a:pos_y
+
+    let s:can_accept = v:true
+    let l:has_info   = v:false
+
+    if s:can_accept && v:shell_error
+        if !a:is_auto
+            call add(s:content, "<| curl error: is the server on? |>")
+        endif
+        let s:can_accept = v:false
+    endif
+
+    let l:n_prompt    = 0
+    let l:t_prompt_ms = 1.0
+    let l:s_prompt    = 0
+
+    let l:n_predict    = 0
+    let l:t_predict_ms = 1.0
+    let l:s_predict    = 0
+
+    " get the generated suggestion
+    if s:can_accept
+        let l:response = json_decode(l:raw)
+
+        for l:part in split(get(l:response, 'content', ''), "\n", 1)
+            call add(s:content, l:part)
+        endfor
+
+        " remove trailing new lines
+        while len(s:content) > 0 && s:content[-1] == ""
+            call remove(s:content, -1)
+        endwhile
+
+        let l:generation_settings = get(l:response, 'generation_settings', {})
+        let l:n_ctx = get(l:generation_settings, 'n_ctx', 0)
+
+        let l:n_cached  = get(l:response, 'tokens_cached', 0)
+        let l:truncated = get(l:response, 'truncated', v:false)
+
+        " if response.timings is available
+        if len(get(l:response, 'timings', {})) > 0
+            let l:has_info = v:true
+            let l:timings  = get(l:response, 'timings', {})
+
+            let l:n_prompt    = get(l:timings, 'prompt_n', 0)
+            let l:t_prompt_ms = get(l:timings, 'prompt_ms', 1)
+            let l:s_prompt    = get(l:timings, 'prompt_per_second', 0)
+
+            let l:n_predict    = get(l:timings, 'predicted_n', 0)
+            let l:t_predict_ms = get(l:timings, 'predicted_ms', 1)
+            let l:s_predict    = get(l:timings, 'predicted_per_second', 0)
+        endif
+    endif
+
+    if len(s:content) == 0
+        call add(s:content, "")
+        let s:can_accept = v:false
+    endif
+
+    if len(s:content) == 0
+        return
+    endif
+
+    " NOTE: the following is logic for discarding predictions that repeat existing text
+    "       the code is quite ugly and there is very likely a simpler and more canonical way to implement this
+    "
+    "       still, I wonder if there is some better way that avoids having to do these special hacks?
+    "       on one hand, the LLM 'sees' the contents of the file before we start editing, so it is normal that it would
+    "       start generating whatever we have given it via the extra context. but on the other hand, it's not very
+    "       helpful to re-generate the same code that is already there
+
+    " truncate the suggestion if the first line is empty
+    if len(s:content) == 1 && s:content[0] == ""
+        let s:content = [""]
+    endif
+
+    " ... and the next lines are repeated
+    if len(s:content) > 1 && s:content[0] == "" && s:content[1:] == getline(s:pos_y + 1, s:pos_y + len(s:content) - 1)
+        let s:content = [""]
+    endif
+
+    " truncate the suggestion if it repeats the suffix
+    if len(s:content) == 1 && s:content[0] == s:line_cur_suffix
+        let s:content = [""]
+    endif
+
+    " find the first non-empty line (strip whitespace)
+    let l:cmp_y = s:pos_y + 1
+    while l:cmp_y < line('$') && getline(l:cmp_y) =~? '^\s*$'
+        let l:cmp_y += 1
+    endwhile
+
+    if (s:line_cur_prefix . s:content[0]) == getline(l:cmp_y)
+        " truncate the suggestion if it repeats the next line
+        if len(s:content) == 1
+            let s:content = [""]
+        endif
+
+        " ... or if the second line of the suggestion is the prefix of line l:cmp_y + 1
+        if len(s:content) == 2 && s:content[-1] == getline(l:cmp_y + 1)[:len(s:content[-1]) - 1]
+            let s:content = [""]
+        endif
+
+        " ... or if the middle chunk of lines of the suggestion is the same as [l:cmp_y + 1, l:cmp_y + len(s:content) - 1)
+        if len(s:content) > 2 && join(s:content[1:-1], "\n") == join(getline(l:cmp_y + 1, l:cmp_y + len(s:content) - 1), "\n")
+            let s:content = [""]
+        endif
+    endif
+
+    " keep only lines that have the same or larger whitespace prefix as s:line_cur_prefix
+    "let l:indent = strlen(matchstr(s:line_cur_prefix, '^\s*'))
+    "for i in range(1, len(s:content) - 1)
+    "    if strlen(matchstr(s:content[i], '^\s*')) < l:indent
+    "        let s:content = s:content[:i - 1]
+    "        break
+    "    endif
+    "endfor
+
+    let s:pos_dx = len(s:content[-1])
+
+    let s:content[-1] .= s:line_cur_suffix
+
+    call llama#fim_cancel()
+
+    " display virtual text with the suggestion
+    let l:bufnr = bufnr('%')
+
+    if s:ghost_text_nvim
+        let l:id_vt_fim = nvim_create_namespace('vt_fim')
+    endif
+
+    " construct the info message
+    if g:llama_config.show_info > 0 && l:has_info
+        let l:prefix = '   '
+
+        if l:truncated
+            let l:info = printf("%s | WARNING: the context is full: %d / %d, increase the server context size or reduce g:llama_config.ring_n_chunks",
+                \ g:llama_config.show_info == 2 ? l:prefix : 'llama.vim',
+                \ l:n_cached, l:n_ctx
+                \ )
+        else
+            let l:info = printf("%s | c: %d / %d, r: %d / %d, e: %d, q: %d / 16 | p: %d (%.2f ms, %.2f t/s) | g: %d (%.2f ms, %.2f t/s) | t: %.2f ms",
+                \ g:llama_config.show_info == 2 ? l:prefix : 'llama.vim',
+                \ l:n_cached,  l:n_ctx, len(s:ring_chunks), g:llama_config.ring_n_chunks, s:ring_n_evict, len(s:ring_queued),
+                \ l:n_prompt,  l:t_prompt_ms,  l:s_prompt,
+                \ l:n_predict, l:t_predict_ms, l:s_predict,
+                \ 1000.0 * reltimefloat(reltime(s:t_fim_start))
+                \ )
+        endif
+
+        if g:llama_config.show_info == 1
+            " display the info in the statusline
+            let &statusline = l:info
+            let l:info = ''
+        endif
+    endif
+
+    " display the suggestion and append the info to the end of the first line
+    if s:ghost_text_nvim
+        call nvim_buf_set_extmark(l:bufnr, l:id_vt_fim, s:pos_y - 1, s:pos_x - 1, {
+            \ 'virt_text': [[s:content[0], 'llama_hl_hint'], [l:info, 'llama_hl_info']],
+            \ 'virt_text_win_col': virtcol('.') - 1
+            \ })
+
+        call nvim_buf_set_extmark(l:bufnr, l:id_vt_fim, s:pos_y - 1, 0, {
+            \ 'virt_lines': map(s:content[1:], {idx, val -> [[val, 'llama_hl_hint']]}),
+            \ 'virt_text_win_col': virtcol('.')
+            \ })
+    elseif s:ghost_text_vim
+        let l:new_suffix = s:content[0]
+        if !empty(l:new_suffix)
+            call prop_add(s:pos_y, s:pos_x + 1, {
+                        \ 'type': s:hlgroup_hint,
+                        \ 'text': l:new_suffix
+                        \ })
+        endif
+        for line in s:content[1:]
+            call prop_add(s:pos_y, 0, {
+                        \ 'type': s:hlgroup_hint,
+                        \ 'text': line,
+                        \ 'text_padding_left': s:get_indent(line),
+                        \ 'text_align': 'below'
+                        \ })
+        endfor
+        if !empty(l:info)
+            call prop_add(s:pos_y, 0, {
+                        \ 'type': s:hlgroup_info,
+                        \ 'text': l:info,
+                        \ 'text_padding_left': col('$'),
+                        \ 'text_wrap': 'truncate'
+                        \ })
+        endif
+    endif
+
+    " setup accept shortcuts
+    inoremap <buffer> <Tab>   <C-O>:call llama#fim_accept(v:false)<CR>
+    inoremap <buffer> <S-Tab> <C-O>:call llama#fim_accept(v:true)<CR>
+
+    let s:hint_shown = v:true
+endfunction
+
+function! s:fim_on_exit(job_id, exit_code, event = v:null)
+    if a:exit_code != 0
+        echom "Job failed with exit code: " . a:exit_code
+    endif
+
+    let s:current_job = v:null
+endfunction

examples/llava/clip.cpp

@@ -4,6 +4,7 @@
 // Note: Even when using identical normalized image inputs (see normalize_image_u8_to_f32()) we have a significant difference in resulting embeddings compared to pytorch
 #include "clip.h"
 #include "ggml.h"
+#include "ggml-cpu.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 

examples/llava/llava-cli.cpp

@@ -20,7 +20,7 @@ static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_toke
         if (n_eval > n_batch) {
             n_eval = n_batch;
         }
-        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
+        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
             LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
             return false;
         }

examples/llava/llava.cpp

@@ -401,6 +401,39 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co
     return true;
 }
 
+struct llava_embd_batch {
+    std::vector<llama_pos>      pos;
+    std::vector<int32_t>        n_seq_id;
+    std::vector<llama_seq_id>   seq_id_0;
+    std::vector<llama_seq_id *> seq_ids;
+    std::vector<int8_t>         logits;
+    llama_batch batch;
+    llava_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
+        pos     .resize(n_tokens);
+        n_seq_id.resize(n_tokens);
+        seq_ids .resize(n_tokens + 1);
+        logits  .resize(n_tokens);
+        seq_id_0.resize(1);
+        seq_id_0[0] = seq_id;
+        seq_ids [n_tokens] = nullptr;
+        batch = {
+            /*n_tokens       =*/ n_tokens,
+            /*tokens         =*/ nullptr,
+            /*embd           =*/ embd,
+            /*pos            =*/ pos.data(),
+            /*n_seq_id       =*/ n_seq_id.data(),
+            /*seq_id         =*/ seq_ids.data(),
+            /*logits         =*/ logits.data(),
+        };
+        for (int i = 0; i < n_tokens; i++) {
+            batch.pos     [i] = pos_0 + i;
+            batch.n_seq_id[i] = 1;
+            batch.seq_id  [i] = seq_id_0.data();
+            batch.logits  [i] = false;
+        }
+    }
+};
+
 bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed, int n_batch, int * n_past) {
     int n_embd  = llama_n_embd(llama_get_model(ctx_llama));
 
@@ -409,8 +442,9 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
         if (n_eval > n_batch) {
             n_eval = n_batch;
         }
-        llama_batch batch = {int32_t(n_eval), nullptr, (image_embed->embed+i*n_embd), nullptr, nullptr, nullptr, nullptr, *n_past, 1, 0, };
-        if (llama_decode(ctx_llama, batch)) {
+        float * embd = image_embed->embed+i*n_embd;
+        llava_embd_batch llava_batch = llava_embd_batch(embd, n_eval, *n_past, 0);
+        if (llama_decode(ctx_llama, llava_batch.batch)) {
             LOG_ERR("%s : failed to eval\n", __func__);
             return false;
         }
@@ -432,7 +466,7 @@ struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * c
     bool image_embed_result = llava_image_embed_make_with_clip_img(ctx_clip, n_threads, img, &image_embed, &n_image_pos);
     if (!image_embed_result) {
         clip_image_u8_free(img);
-        LOG_ERR("%s: coulnd't embed the image\n", __func__);
+        LOG_ERR("%s: couldn't embed the image\n", __func__);
         return NULL;
     }
 

examples/llava/minicpmv-cli.cpp

@@ -97,7 +97,7 @@ static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_toke
         if (n_eval > n_batch) {
             n_eval = n_batch;
         }
-        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
+        if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval))) {
             LOG_ERR("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
             return false;
         }

examples/lookahead/lookahead.cpp

@@ -89,8 +89,8 @@ int main(int argc, char ** argv) {
     const auto t_enc_start = ggml_time_us();
 
     // eval the prompt
-    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0,           0));
-    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0));
+    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1));
+    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1));
 
     for (int s = 1; s < W + G + 1; ++s) {
         llama_kv_cache_seq_cp(ctx, 0, s, -1, -1);

examples/lookup/lookup.cpp

@@ -89,8 +89,8 @@ int main(int argc, char ** argv){
 
     const auto t_enc_start = ggml_time_us();
 
-    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1, 0,           0));
-    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0));
+    llama_decode(ctx, llama_batch_get_one( inp.data(), n_input - 1));
+    llama_decode(ctx, llama_batch_get_one(&inp.back(),           1));
 
     const auto t_enc_end = ggml_time_us();
 

examples/main/README.md

@@ -187,6 +187,30 @@ Use the `--no-penalize-nl` option to disable newline penalization when applying
 
 Example usage: `--repeat-penalty 1.15 --repeat-last-n 128 --no-penalize-nl`
 
+### DRY Repetition Penalty
+
+DRY (Don't Repeat Yourself) sampling is an effective technique for reducing repetition in generated text even across long contexts by penalizing tokens based on their recent usage patterns (original [PR link](https://github.com/oobabooga/text-generation-webui/pull/5677)).
+
+- `--dry-multiplier N`: Set the DRY sampling multiplier (default: 0.0, 0.0 = disabled).
+- `--dry-base N`: Set the DRY sampling base value (default: 1.75).
+- `--dry-allowed-length N`: Set the allowed length for DRY sampling (default: 2).
+- `--dry-penalty-last-n N`: Set DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size).
+- `--dry-sequence-breaker STRING`: Add a sequence breaker for DRY sampling. Can be used more than once to add multiple sequence breakers. Using this clears out the default breakers, which consist of: `['\n', ':', '"', '*']`. If the string `"none"` is supplied, no sequence breakers are used.
+
+The `dry-multiplier` option controls the strength of the DRY sampling effect. A value of 0.0 disables DRY sampling, while higher values increase its influence. A typical recommended value is 0.8.
+
+The `dry-base` option sets the base value for the exponential penalty calculation in DRY sampling. Higher values lead to more aggressive penalization of repetitions.
+
+The `dry-allowed-length` option sets the maximum length of repeated sequences that will not be penalized. Repetitions shorter than or equal to this length are not penalized, allowing for natural repetitions of short phrases or common words.
+
+The `dry-penalty-last-n` option controls how many recent tokens to consider when applying the DRY penalty. A value of -1 considers the entire context. Use a positive value to limit the consideration to a specific number of recent tokens.
+
+The `dry-sequence-breaker` option adds a single sequence breaker and can be used more than once to specify multiple sequence breakers. Sequence breakers interrupt sequence matching and break the input into parts where matching can be applied.
+
+DRY sampling provides more nuanced control over text generation, particularly for reducing long-range repetitions and maintaining global coherence.
+
+Example usage: `--dry-multiplier 0.8 --dry-base 1.75 --dry-allowed-length 2 --dry-penalty-last-n -1 --dry-sequence-breaker "—" --dry-sequence-breaker "##"`
+
 ### Top-K Sampling
 
 -   `--top-k N`: Limit the next token selection to the K most probable tokens (default: 40).
@@ -211,14 +235,6 @@ The Min-P sampling method was designed as an alternative to Top-P, and aims to e
 
 Example usage: `--min-p 0.05`
 
-### Tail-Free Sampling (TFS)
-
--   `--tfs N`: Enable tail free sampling with parameter z (default: 1.0, 1.0 = disabled).
-
-Tail-free sampling (TFS) is a text generation technique that aims to reduce the impact of less likely tokens, which may be less relevant, less coherent, or nonsensical, on the output. Similar to Top-P it tries to determine the bulk of the most likely tokens dynamically. But TFS filters out logits based on the second derivative of their probabilities. Adding tokens is stopped after the sum of the second derivatives reaches the parameter z. In short: TFS looks at how quickly the probabilities of the tokens decrease and cuts off the tail of unlikely tokens using the parameter z. Typical values for z are in the range of 0.9 to 0.95. A value of 1.0 would include all tokens and thus disables the effect of TFS.
-
-Example usage: `--tfs 0.95`
-
 ### Locally Typical Sampling
 
 -   `--typical N`: Enable locally typical sampling with parameter p (default: 1.0, 1.0 = disabled).
@@ -241,6 +257,19 @@ The `--mirostat-ent` option sets the Mirostat target entropy (tau), which repres
 
 Example usage: `--mirostat 2 --mirostat-lr 0.05 --mirostat-ent 3.0`
 
+### XTC Sampling
+
+-   `--xtc-probability N`: Sets the chance for token removal (checked once on sampler start) (default: 0.0).
+-   `--xtc-threshold N`: Sets a minimum probability threshold for tokens to be removed (default: 0.1).
+
+Exclude Top Choices (XTC) is a unique sampler that is designed to remove top tokens from consideration and avoid more obvious and repetitive outputs. With a chance of `xtc-probability` it searches for tokens with probabilities of `xtc-threshold` and above, then removes all such tokens except the least probable one.
+
+By removing top tokens XTC can improve the variety of answers, break writing clichés and inhibit repition, since clichés and repeated phrases are usually more likely to appear. By keeping the last token above the threshold, XTC ensures that the answer is still coherent. XTC is meant to be used for creative tasks, but feel free to experiment with different settings for different models.
+
+Being experimental and unique, XTC is disabled by default. The recommended combination of samplers is Min-P followed by XTC on its default settings: `--sampling-seq mx --min-p 0.02 --xtc-probability 0.5`.
+
+Example usage: `--xtc-probability 0.5 --xtc-threshold 0.1`
+
 ### Logit Bias
 
 -   `-l TOKEN_ID(+/-)BIAS, --logit-bias TOKEN_ID(+/-)BIAS`: Modify the likelihood of a token appearing in the generated text completion.
@@ -284,10 +313,6 @@ These options help improve the performance and memory usage of the LLaMA models.
 
  These flags attempt optimizations that help on some systems with non-uniform memory access. This currently consists of one of the above strategies, and disabling prefetch and readahead for mmap. The latter causes mapped pages to be faulted in on first access instead of all at once, and in combination with pinning threads to NUMA nodes, more of the pages end up on the NUMA node where they are used. Note that if the model is already in the system page cache, for example because of a previous run without this option, this will have little effect unless you drop the page cache first. This can be done by rebooting the system or on Linux by writing '3' to '/proc/sys/vm/drop_caches' as root.
 
-### Memory Float 32
-
--   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. This doubles the context memory requirement and cached prompt file size but does not appear to increase generation quality in a measurable way. Not recommended.
-
 ### Batch Size
 
 -   `-b N, --batch-size N`: Set the batch size for prompt processing (default: `2048`). This large batch size benefits users who have BLAS installed and enabled it during the build. If you don't have BLAS enabled ("BLAS=0"), you can use a smaller number, such as 8, to see the prompt progress as it's evaluated in some situations.
@@ -308,6 +333,15 @@ These options help improve the performance and memory usage of the LLaMA models.
 
 For information about 4-bit quantization, which can significantly improve performance and reduce memory usage, please refer to llama.cpp's primary [README](../../README.md#prepare-and-quantize).
 
+## LoRA (Low-Rank Adaptation) adapters
+
+-   `--lora FNAME`: Optional path to a LoRA adapter to use with scaling of 1.0. Can be mixed with `--lora-scaled` and can be repeated to use multiple adapters.
+-   `--lora-scaled FNAME`: Optional path to a LoRA adapter with user-defined scaling. Can be mixed with `--lora` and can repeated to use multiple adapters.
+
+You can add LoRA adapters using `--lora` or `--lora-scaled`. For example: `--lora my_adapter_1.gguf --lora my_adapter_2.gguf ...` or `--lora-scaled lora_task_A.gguf 0.5 --lora-scaled lora_task_B.gguf 0.5`.
+
+LoRA adapters should be in GGUF format. To convert from Hugging Face format use the `convert-lora-to-gguf.py` script. LoRA adapters are loaded separately and applied during inference - they are not merged with the main model. This means that mmap model loading is fully supported when using LoRA adapters. The old `--lora-base` flag has been removed now that merging is no longer performed.
+
 ## Additional Options
 
 These options provide extra functionality and customization when running the LLaMA models:
@@ -316,6 +350,4 @@ These options provide extra functionality and customization when running the LLa
 -   `--verbose-prompt`: Print the prompt before generating text.
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance.
--   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
--   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
 -   `-hfr URL --hf-repo URL`: The url to the Hugging Face model repository. Used in conjunction with `--hf-file` or `-hff`. The model is downloaded and stored in the file provided by `-m` or `--model`. If `-m` is not provided, the model is auto-stored in the path specified by the `LLAMA_CACHE` environment variable  or in an OS-specific local cache.

examples/main/main.cpp

@@ -528,7 +528,7 @@ int main(int argc, char ** argv) {
         int enc_input_size = embd_inp.size();
         llama_token * enc_input_buf = embd_inp.data();
 
-        if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size, 0, 0))) {
+        if (llama_encode(ctx, llama_batch_get_one(enc_input_buf, enc_input_size))) {
             LOG_ERR("%s : failed to eval\n", __func__);
             return 1;
         }
@@ -569,30 +569,30 @@ int main(int argc, char ** argv) {
                     if (!params.ctx_shift){
                         LOG_DBG("\n\n%s: context full and context shift is disabled => stopping\n", __func__);
                         break;
-                    } else {
-                        if (params.n_predict == -2) {
-                            LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
-                            break;
-                        }
-
-                        const int n_left    = n_past - params.n_keep;
-                        const int n_discard = n_left/2;
-
-                        LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
-                                n_past, n_left, n_ctx, params.n_keep, n_discard);
-
-                        llama_kv_cache_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
-                        llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
-
-                        n_past -= n_discard;
-
-                        LOG_DBG("after swap: n_past = %d\n", n_past);
-
-                        LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
-
-                        LOG_DBG("clear session path\n");
-                        path_session.clear();
                     }
+
+                    if (params.n_predict == -2) {
+                        LOG_DBG("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
+                        break;
+                    }
+
+                    const int n_left    = n_past - params.n_keep;
+                    const int n_discard = n_left/2;
+
+                    LOG_DBG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
+                            n_past, n_left, n_ctx, params.n_keep, n_discard);
+
+                    llama_kv_cache_seq_rm (ctx, 0, params.n_keep            , params.n_keep + n_discard);
+                    llama_kv_cache_seq_add(ctx, 0, params.n_keep + n_discard, n_past, -n_discard);
+
+                    n_past -= n_discard;
+
+                    LOG_DBG("after swap: n_past = %d\n", n_past);
+
+                    LOG_DBG("embd: %s\n", string_from(ctx, embd).c_str());
+
+                    LOG_DBG("clear session path\n");
+                    path_session.clear();
                 }
             } else {
                 // context extension via Self-Extend
@@ -648,7 +648,7 @@ int main(int argc, char ** argv) {
 
                 LOG_DBG("eval: %s\n", string_from(ctx, embd).c_str());
 
-                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval, n_past, 0))) {
+                if (llama_decode(ctx, llama_batch_get_one(&embd[i], n_eval))) {
                     LOG_ERR("%s : failed to eval\n", __func__);
                     return 1;
                 }

examples/parallel/parallel.cpp

@@ -308,7 +308,6 @@ int main(int argc, char ** argv) {
                 batch.n_seq_id + i,
                 batch.seq_id   + i,
                 batch.logits   + i,
-                0, 0, 0, // unused
             };
 
             const int ret = llama_decode(ctx, batch_view);

examples/perplexity/perplexity.cpp

@@ -408,14 +408,21 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
         // clear the KV cache
         llama_kv_cache_clear(ctx);
 
+        llama_batch batch = llama_batch_init(n_batch, 0, 1);
+
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
 
+            common_batch_clear(batch);
+            for (int i = 0; i < batch_size; i++) {
+                common_batch_add(batch, tokens[batch_start + i], j*n_batch + i, {0}, true);
+            }
+
             //LOG_DBG("    Batch %d: starts at %d, size is %d, n_past is %d\n",j,batch_start,batch_size,j * n_batch);
-            // TODO: use llama_batch.logits instead of relying on logits_all == true
-            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
+            if (llama_decode(ctx, batch)) {
                 //LOG_ERR("%s : failed to eval\n", __func__);
+                llama_batch_free(batch);
                 return {tokens, -1, logit_history, prob_history};
             }
 
@@ -435,6 +442,8 @@ static results_perplexity perplexity_v2(llama_context * ctx, const common_params
             }
         }
 
+        llama_batch_free(batch);
+
         const auto t_end = std::chrono::high_resolution_clock::now();
 
         if (i == 0) {
@@ -704,7 +713,6 @@ static bool decode_helper(llama_context * ctx, llama_batch & batch, std::vector<
             batch.n_seq_id + i,
             batch.seq_id   + i,
             batch.logits   + i,
-            0, 0, 0, // unused
         };
 
         const int ret = llama_decode(ctx, batch_view);
@@ -1791,6 +1799,8 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
         // clear the KV cache
         llama_kv_cache_clear(ctx);
 
+        llama_batch batch = llama_batch_init(n_batch, 0, 1);
+
         for (int j = 0; j < num_batches; ++j) {
             const int batch_start = start + j * n_batch;
             const int batch_size  = std::min(end - batch_start, n_batch);
@@ -1803,9 +1813,14 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
                 tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
             }
 
-            // TODO: use llama_batch.logits instead of relying on logits_all == true
-            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
+            common_batch_clear(batch);
+            for (int i = 0; i < batch_size; i++) {
+                common_batch_add(batch, tokens[batch_start + i], j*n_batch + i, {0}, true);
+            }
+
+            if (llama_decode(ctx, batch)) {
                 LOG_ERR("%s : failed to eval\n", __func__);
+                llama_batch_free(batch);
                 return;
             }
 
@@ -1818,6 +1833,8 @@ static void kl_divergence(llama_context * ctx, const common_params & params) {
             }
         }
 
+        llama_batch_free(batch);
+
         const auto t_end = std::chrono::high_resolution_clock::now();
 
         if (i == 0) {

examples/rpc/rpc-server.cpp

@@ -1,3 +1,5 @@
+#include "ggml-cpu.h"
+
 #ifdef GGML_USE_CUDA
 #include "ggml-cuda.h"
 #endif

examples/save-load-state/save-load-state.cpp

@@ -42,15 +42,21 @@ int main(int argc, char ** argv) {
 
     llama_sampler * smpl = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl, llama_sampler_init_softmax());
     llama_sampler_chain_add(smpl, llama_sampler_init_dist(params.sparams.seed));
 
     // tokenize prompt
     auto tokens = common_tokenize(ctx, params.prompt, true);
 
+    // prepare the batch
+    llama_batch batch = llama_batch_init(tokens.size(), 0, 1);
+    for (size_t i = 0; i < tokens.size(); i++) {
+        common_batch_add(batch, tokens[i], i, {0}, false);
+    }
+    batch.logits[batch.n_tokens - 1] = true; // generate next token
+
     // evaluate prompt
-    llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), n_past, 0));
-    n_past += tokens.size();
+    llama_decode(ctx, batch);
+    n_past += batch.n_tokens;
 
     // save state (rng, logits, embedding and kv_cache) to file
     {
@@ -77,8 +83,12 @@ int main(int argc, char ** argv) {
         printf("%s", next_token_str.c_str());
         result0 += next_token_str;
 
-        if (llama_decode(ctx, llama_batch_get_one(&next_token, 1, n_past, 0))) {
+        common_batch_clear(batch);
+        common_batch_add(batch, next_token, n_past, {0}, true);
+
+        if (llama_decode(ctx, batch)) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
+            llama_batch_free(batch);
             llama_free(ctx);
             llama_free_model(model);
             return 1;
@@ -96,7 +106,6 @@ int main(int argc, char ** argv) {
 
     llama_sampler * smpl2 = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl2, llama_sampler_init_softmax());
     llama_sampler_chain_add(smpl2, llama_sampler_init_dist(params.sparams.seed));
 
     printf("\nsecond run: %s", params.prompt.c_str());
@@ -133,8 +142,12 @@ int main(int argc, char ** argv) {
         printf("%s", next_token_str.c_str());
         result1 += next_token_str;
 
-        if (llama_decode(ctx2, llama_batch_get_one(&next_token, 1, n_past, 0))) {
+        common_batch_clear(batch);
+        common_batch_add(batch, next_token, n_past, {0}, true);
+
+        if (llama_decode(ctx2, batch)) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
+            llama_batch_free(batch);
             llama_free(ctx2);
             llama_free_model(model);
             return 1;
@@ -156,7 +169,6 @@ int main(int argc, char ** argv) {
 
     llama_sampler * smpl3 = llama_sampler_chain_init(sparams);
 
-    llama_sampler_chain_add(smpl3, llama_sampler_init_softmax());
     llama_sampler_chain_add(smpl3, llama_sampler_init_dist(params.sparams.seed));
 
     printf("\nsingle seq run: %s", params.prompt.c_str());
@@ -221,8 +233,12 @@ int main(int argc, char ** argv) {
         printf("%s", next_token_str.c_str());
         result2 += next_token_str;
 
-        if (llama_decode(ctx3, llama_batch_get_one(&next_token, 1, n_past, 1))) {
+        common_batch_clear(batch);
+        common_batch_add(batch, next_token, n_past, {1}, true);
+
+        if (llama_decode(ctx3, batch)) {
             fprintf(stderr, "\n%s : failed to evaluate\n", __func__);
+            llama_batch_free(batch);
             llama_free(ctx3);
             llama_free_model(model);
             return 1;
@@ -236,6 +252,7 @@ int main(int argc, char ** argv) {
     llama_sampler_free(smpl2);
     llama_sampler_free(smpl3);
 
+    llama_batch_free(batch);
     llama_free(ctx3);
     llama_free_model(model);
 

examples/server/CMakeLists.txt

@@ -15,22 +15,13 @@ set(TARGET_SRCS
     httplib.h
 )
 set(PUBLIC_ASSETS
-    colorthemes.css
-    style.css
-    theme-beeninorder.css
-    theme-ketivah.css
-    theme-mangotango.css
-    theme-playground.css
-    theme-polarnight.css
-    theme-snowstorm.css
     index.html
-    index-new.html
-    index.js
     completion.js
-    system-prompts.js
-    prompt-formats.js
-    json-schema-to-grammar.mjs
     loading.html
+    deps_daisyui.min.css
+    deps_markdown-it.js
+    deps_tailwindcss.js
+    deps_vue.esm-browser.js
 )
 
 foreach(asset ${PUBLIC_ASSETS})

examples/server/README.md

@@ -60,8 +60,6 @@ The project is under active development, and we are [looking for feedback and co
 | `--yarn-attn-factor N` | YaRN: scale sqrt(t) or attention magnitude (default: 1.0)<br/>(env: LLAMA_ARG_YARN_ATTN_FACTOR) |
 | `--yarn-beta-slow N` | YaRN: high correction dim or alpha (default: 1.0)<br/>(env: LLAMA_ARG_YARN_BETA_SLOW) |
 | `--yarn-beta-fast N` | YaRN: low correction dim or beta (default: 32.0)<br/>(env: LLAMA_ARG_YARN_BETA_FAST) |
-| `-gan, --grp-attn-n N` | group-attention factor (default: 1)<br/>(env: LLAMA_ARG_GRP_ATTN_N) |
-| `-gaw, --grp-attn-w N` | group-attention width (default: 512.0)<br/>(env: LLAMA_ARG_GRP_ATTN_W) |
 | `-dkvc, --dump-kv-cache` | verbose print of the KV cache |
 | `-nkvo, --no-kv-offload` | disable KV offload<br/>(env: LLAMA_ARG_NO_KV_OFFLOAD) |
 | `-ctk, --cache-type-k TYPE` | KV cache data type for K (default: f16)<br/>(env: LLAMA_ARG_CACHE_TYPE_K) |
@@ -101,7 +99,7 @@ The project is under active development, and we are [looking for feedback and co
 
 | Argument | Explanation |
 | -------- | ----------- |
-| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: top_k;tfs_z;typ_p;top_p;min_p;temperature) |
+| `--samplers SAMPLERS` | samplers that will be used for generation in the order, separated by ';'<br/>(default: top_k;typ_p;top_p;min_p;temperature) |
 | `-s, --seed SEED` | RNG seed (default: -1, use random seed for -1) |
 | `--sampling-seq SEQUENCE` | simplified sequence for samplers that will be used (default: kfypmt) |
 | `--ignore-eos` | ignore end of stream token and continue generating (implies --logit-bias EOS-inf) |
@@ -110,15 +108,19 @@ The project is under active development, and we are [looking for feedback and co
 | `--top-k N` | top-k sampling (default: 40, 0 = disabled) |
 | `--top-p N` | top-p sampling (default: 0.9, 1.0 = disabled) |
 | `--min-p N` | min-p sampling (default: 0.1, 0.0 = disabled) |
-| `--tfs N` | tail free sampling, parameter z (default: 1.0, 1.0 = disabled) |
 | `--typical N` | locally typical sampling, parameter p (default: 1.0, 1.0 = disabled) |
 | `--repeat-last-n N` | last n tokens to consider for penalize (default: 64, 0 = disabled, -1 = ctx_size) |
 | `--repeat-penalty N` | penalize repeat sequence of tokens (default: 1.0, 1.0 = disabled) |
 | `--presence-penalty N` | repeat alpha presence penalty (default: 0.0, 0.0 = disabled) |
 | `--frequency-penalty N` | repeat alpha frequency penalty (default: 0.0, 0.0 = disabled) |
+| `--dry-multiplier N` | DRY sampling multiplier (default: 0.0, 0.0 = disabled) |
+| `--dry-base N` | DRY sampling base value (default: 1.75) |
+| `--dry-allowed-length N` | allowed length for DRY sampling (default: 2) |
+| `--dry-penalty-last-n N` | DRY penalty for the last n tokens (default: -1, 0 = disable, -1 = context size) |
+| `--dry-sequence-breaker STRING` | add sequence breaker for DRY sampling, clearing out default breakers (`['\n', ':', '"', '*']`) in the process; use `"none"` to not use any sequence breakers
 | `--dynatemp-range N` | dynamic temperature range (default: 0.0, 0.0 = disabled) |
 | `--dynatemp-exp N` | dynamic temperature exponent (default: 1.0) |
-| `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus, Tail Free and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
+| `--mirostat N` | use Mirostat sampling.<br/>Top K, Nucleus and Locally Typical samplers are ignored if used.<br/>(default: 0, 0 = disabled, 1 = Mirostat, 2 = Mirostat 2.0) |
 | `--mirostat-lr N` | Mirostat learning rate, parameter eta (default: 0.1) |
 | `--mirostat-ent N` | Mirostat target entropy, parameter tau (default: 5.0) |
 | `-l, --logit-bias TOKEN_ID(+/-)BIAS` | modifies the likelihood of token appearing in the completion,<br/>i.e. `--logit-bias 15043+1` to increase likelihood of token ' Hello',<br/>or `--logit-bias 15043-1` to decrease likelihood of token ' Hello' |
@@ -149,7 +151,7 @@ The project is under active development, and we are [looking for feedback and co
 | `--ssl-cert-file FNAME` | path to file a PEM-encoded SSL certificate<br/>(env: LLAMA_ARG_SSL_CERT_FILE) |
 | `-to, --timeout N` | server read/write timeout in seconds (default: 600)<br/>(env: LLAMA_ARG_TIMEOUT) |
 | `--threads-http N` | number of threads used to process HTTP requests (default: -1)<br/>(env: LLAMA_ARG_THREADS_HTTP) |
-| `-spf, --system-prompt-file FNAME` | set a file to load a system prompt (initial prompt of all slots), this is useful for chat applications |
+| `--cache-reuse N` | min chunk size to attempt reusing from the cache via KV shifting (default: 0)<br/>(env: LLAMA_ARG_CACHE_REUSE) |
 | `--metrics` | enable prometheus compatible metrics endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_METRICS) |
 | `--slots` | enable slots monitoring endpoint (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_SLOTS) |
 | `--props` | enable changing global properties via POST /props (default: disabled)<br/>(env: LLAMA_ARG_ENDPOINT_PROPS) |
@@ -320,7 +322,18 @@ node index.js
 
       - The prompt is a string or an array with the first element given as a string
       - The model's `tokenizer.ggml.add_bos_token` metadata is `true`
-      - The system prompt is empty
+
+    These input shapes and data type are allowed for `prompt`:
+
+      - Single string: `"string"`
+      - Single sequence of tokens: `[12, 34, 56]`
+      - Mixed tokens and strings: `[12, 34, "string", 56, 78]`
+
+    Multiple prompts are also supported. In this case, the completion result will be an array.
+
+      - Only strings: `["string1", "string2"]`
+      - Strings and sequences of tokens: `["string1", [12, 34, 56]]`
+      - Mixed types: `[[12, 34, "string", 56, 78], [12, 34, 56], "string"]`
 
     `temperature`: Adjust the randomness of the generated text. Default: `0.8`
 
@@ -336,6 +349,8 @@ node index.js
 
     `n_predict`: Set the maximum number of tokens to predict when generating text. **Note:** May exceed the set limit slightly if the last token is a partial multibyte character. When 0, no tokens will be generated but the prompt is evaluated into the cache. Default: `-1`, where `-1` is infinity.
 
+    `n_indent`: Specify the minimum line indentation for the generated text in number of whitespace characters. Useful for code completion tasks. Default: `0`
+
     `n_keep`: Specify the number of tokens from the prompt to retain when the context size is exceeded and tokens need to be discarded. The number excludes the BOS token.
     By default, this value is set to `0`, meaning no tokens are kept. Use `-1` to retain all tokens from the prompt.
 
@@ -344,8 +359,6 @@ node index.js
     `stop`: Specify a JSON array of stopping strings.
     These words will not be included in the completion, so make sure to add them to the prompt for the next iteration. Default: `[]`
 
-    `tfs_z`: Enable tail free sampling with parameter z. Default: `1.0`, which is disabled.
-
     `typical_p`: Enable locally typical sampling with parameter p. Default: `1.0`, which is disabled.
 
     `repeat_penalty`: Control the repetition of token sequences in the generated text. Default: `1.1`
@@ -358,6 +371,16 @@ node index.js
 
     `frequency_penalty`: Repeat alpha frequency penalty. Default: `0.0`, which is disabled.
 
+    `dry_multiplier`: Set the DRY (Don't Repeat Yourself) repetition penalty multiplier. Default: `0.0`, which is disabled.
+
+    `dry_base`: Set the DRY repetition penalty base value. Default: `1.75`
+
+    `dry_allowed_length`: Tokens that extend repetition beyond this receive exponentially increasing penalty: multiplier * base ^ (length of repeating sequence before token - allowed length). Default: `2`
+
+    `dry_penalty_last_n`: How many tokens to scan for repetitions. Default: `-1`, where `0` is disabled and `-1` is context size.
+
+    `dry_sequence_breakers`: Specify an array of sequence breakers for DRY sampling. Only a JSON array of strings is accepted. Default: `['\n', ':', '"', '*']`
+
     `mirostat`: Enable Mirostat sampling, controlling perplexity during text generation. Default: `0`, where `0` is disabled, `1` is Mirostat, and `2` is Mirostat 2.0.
 
     `mirostat_tau`: Set the Mirostat target entropy, parameter tau. Default: `5.0`
@@ -378,13 +401,15 @@ node index.js
 
     `min_keep`: If greater than 0, force samplers to return N possible tokens at minimum. Default: `0`
 
+    `t_max_predict_ms`: Set a time limit in milliseconds for the prediction (a.k.a. text-generation) phase. The timeout will trigger if the generation takes more than the specified time (measured since the first token was generated) and if a new-line character has already been generated. Useful for FIM applications. Default: `0`, which is disabled.
+
     `image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `prompt`. You can determine the place of the image in the prompt as in the following: `USER:[img-12]Describe the image in detail.\nASSISTANT:`. In this case, `[img-12]` will be replaced by the embeddings of the image with id `12` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 12}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
 
     `id_slot`: Assign the completion task to an specific slot. If is -1 the task will be assigned to a Idle slot.  Default: `-1`
 
     `cache_prompt`: Re-use KV cache from a previous request if possible. This way the common prefix does not have to be re-processed, only the suffix that differs between the requests. Because (depending on the backend) the logits are **not** guaranteed to be bit-for-bit identical for different batch sizes (prompt processing vs. token generation) enabling this option can cause nondeterministic results. Default: `false`
 
-    `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values.
+    `samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values.
 
 **Response format**
 
@@ -525,8 +550,31 @@ Takes a prefix and a suffix and returns the predicted completion as stream.
 
 - `input_prefix`: Set the prefix of the code to infill.
 - `input_suffix`: Set the suffix of the code to infill.
+- `input_extra`:  Additional context inserted before the FIM prefix.
+- `prompt`:       Added after the `FIM_MID` token
 
-It also accepts all the options of `/completion`.
+`input_extra` is array of `{"filename": string, "text": string}` objects.
+
+The endpoint also accepts all the options of `/completion`.
+
+If the model has `FIM_REPO` and `FIM_FILE_SEP` tokens, the [repo-level pattern](https://arxiv.org/pdf/2409.12186) is used:
+
+```txt
+<FIM_REP>myproject
+<FIM_SEP>{chunk 0 filename}
+{chunk 0 text}
+<FIM_SEP>{chunk 1 filename}
+{chunk 1 text}
+...
+<FIM_SEP>filename
+<FIM_PRE>[input_prefix]<FIM_SUF>[input_suffix]<FIM_MID>[prompt]
+```
+
+If the tokens are missing, then the extra context is simply prefixed at the start:
+
+```txt
+[input_extra]<FIM_PRE>[input_prefix]<FIM_SUF>[input_suffix]<FIM_MID>[prompt]
+```
 
 ### **GET** `/props`: Get server global properties.
 
@@ -536,14 +584,12 @@ This endpoint is public (no API key check). By default, it is read-only. To make
 
 ```json
 {
-  "system_prompt": "",
   "default_generation_settings": { ... },
   "total_slots": 1,
   "chat_template": ""
 }
 ```
 
-- `system_prompt` - the system prompt (initial prompt of all slots). Please note that this does not take into account the chat template. It will append the prompt at the beginning of formatted prompt.
 - `default_generation_settings` - the default generation settings for the `/completion` endpoint, which has the same fields as the `generation_settings` response object from the `/completion` endpoint.
 - `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
 - `chat_template` - the model's original Jinja2 prompt template
@@ -554,7 +600,7 @@ To use this endpoint with POST method, you need to start server with `--props`
 
 *Options:*
 
-- `system_prompt`: Change the system prompt (initial prompt of all slots). Please note that this does not take into account the chat template. It will append the prompt at the beginning of formatted prompt.
+- None yet
 
 ### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API
 
@@ -646,7 +692,10 @@ Given a ChatML-formatted json description in `messages`, it returns the predicte
 
 ### GET `/slots`: Returns the current slots processing state
 
-This endpoint can be disabled with `--no-slots`
+> [!WARNING]
+> This endpoint is intended for debugging and may be modified in future versions. For security reasons, we strongly advise against enabling it in production environments.
+
+This endpoint is disabled by default and can be enabled with `--slots`
 
 If query param `?fail_on_no_slot=1` is set, this endpoint will respond with status code 503 if there is no available slots.
 
@@ -663,6 +712,7 @@ Example:
         "grammar": "",
         "id": 0,
         "ignore_eos": false,
+        "is_processing": false,
         "logit_bias": [],
         "min_p": 0.05000000074505806,
         "mirostat": 0,
@@ -689,21 +739,18 @@ Example:
         "repeat_penalty": 1.100000023841858,
         "samplers": [
             "top_k",
-            "tfs_z",
             "typical_p",
             "top_p",
             "min_p",
             "temperature"
         ],
         "seed": 42,
-        "state": 1,
         "stop": [
             "\n"
         ],
         "stream": false,
         "task_id": 0,
         "temperature": 0.0,
-        "tfs_z": 1.0,
         "top_k": 40,
         "top_p": 0.949999988079071,
         "typical_p": 1.0
@@ -711,10 +758,6 @@ Example:
 ]
 ```
 
-Possible values for `slot[i].state` are:
-- `0`: SLOT_STATE_IDLE
-- `1`: SLOT_STATE_PROCESSING
-
 ### GET `/metrics`: Prometheus compatible metrics exporter
 
 This endpoint is only accessible if `--metrics` is set.
@@ -885,6 +928,16 @@ Apart from error types supported by OAI, we also have custom types that are spec
 }
 ```
 
+### Legacy completion web UI
+
+A new chat-based UI has replaced the old completion-based since [this PR](https://github.com/ggerganov/llama.cpp/pull/10175). If you want to use the old completion, start the server with `--path ./examples/server/public_legacy`
+
+For example:
+
+```sh
+./llama-server -m my_model.gguf -c 8192 --path ./examples/server/public_legacy
+```
+
 ### Extending or building alternative Web Front End
 
 You can extend the front end by running the server binary with `--path` set to `./your-directory` and importing `/completion.js` to get access to the llamaComplete() method.

examples/server/chat.mjs

@@ -1,7 +1,7 @@
 import * as readline from 'node:readline'
 import { stdin, stdout } from 'node:process'
 import { readFileSync } from 'node:fs'
-import { SchemaConverter }  from './public/json-schema-to-grammar.mjs'
+import { SchemaConverter }  from './public_legacy/json-schema-to-grammar.mjs'
 
 const args = process.argv.slice(2);
 const grammarJsonSchemaFile = args.find(

examples/server/deps.sh

@@ -6,5 +6,20 @@ DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
 PUBLIC=$DIR/public
 
 echo "download js bundle files"
-curl https://npm.reversehttp.com/@preact/signals-core,@preact/signals,htm/preact,preact,preact/hooks > $PUBLIC/index.js
-echo >> $PUBLIC/index.js # add newline
+
+# Note for contributors: Always pin to a specific version "maj.min.patch" to avoid breaking the CI
+
+curl -L https://cdn.tailwindcss.com/3.4.14 > $PUBLIC/deps_tailwindcss.js
+echo >> $PUBLIC/deps_tailwindcss.js # add newline
+
+curl -L https://cdnjs.cloudflare.com/ajax/libs/daisyui/4.12.14/styled.min.css > $PUBLIC/deps_daisyui.min.css
+curl -L https://cdnjs.cloudflare.com/ajax/libs/daisyui/4.12.14/themes.min.css >> $PUBLIC/deps_daisyui.min.css
+echo >> $PUBLIC/deps_daisyui.min.css # add newline
+
+curl -L https://unpkg.com/vue@3.5.12/dist/vue.esm-browser.js > $PUBLIC/deps_vue.esm-browser.js
+echo >> $PUBLIC/deps_vue.esm-browser.js # add newline
+
+curl -L https://cdnjs.cloudflare.com/ajax/libs/markdown-it/13.0.2/markdown-it.js > $PUBLIC/deps_markdown-it.js
+echo >> $PUBLIC/deps_markdown-it.js # add newline
+
+ls -lah $PUBLIC

examples/server/public/completion.js

@@ -1,12 +1,16 @@
 const paramDefaults = {
   stream: true,
-  n_predict: 500,
   temperature: 0.2,
-  stop: ["</s>"]
 };
 
 let generation_settings = null;
 
+export class CompletionError extends Error {
+  constructor(message, name, data) {
+    super(message);
+    this.name = name;
+  }
+};
 
 // Completes the prompt as a generator. Recommended for most use cases.
 //
@@ -29,7 +33,7 @@ export async function* llama(prompt, params = {}, config = {}) {
 
   const completionParams = { ...paramDefaults, ...params, prompt };
 
-  const response = await fetch(`${api_url}/completion`, {
+  const response = await fetch(`${api_url}${config.endpoint || '/completion'}`, {
     method: 'POST',
     body: JSON.stringify(completionParams),
     headers: {
@@ -41,6 +45,18 @@ export async function* llama(prompt, params = {}, config = {}) {
     signal: controller.signal,
   });
 
+  const status = response.status;
+  if (status !== 200) {
+    try {
+      const body = await response.json();
+      if (body && body.error && body.error.message) {
+        throw new CompletionError(body.error.message, 'ServerError');
+      }
+    } catch (err) {
+      throw new CompletionError(err.message, 'ServerError');
+    }
+  }
+
   const reader = response.body.getReader();
   const decoder = new TextDecoder();
 
@@ -78,7 +94,12 @@ export async function* llama(prompt, params = {}, config = {}) {
       for (const line of lines) {
         const match = regex.exec(line);
         if (match) {
-          result[match[1]] = match[2]
+          result[match[1]] = match[2];
+          if (result.data === '[DONE]') {
+            cont = false;
+            break;
+          }
+
           // since we know this is llama.cpp, let's just decode the json in data
           if (result.data) {
             result.data = JSON.parse(result.data);

examples/server/public_legacy/completion.js

@@ -0,0 +1,209 @@
+const paramDefaults = {
+  stream: true,
+  n_predict: 500,
+  temperature: 0.2,
+  stop: ["</s>"]
+};
+
+let generation_settings = null;
+
+
+// Completes the prompt as a generator. Recommended for most use cases.
+//
+// Example:
+//
+//    import { llama } from '/completion.js'
+//
+//    const request = llama("Tell me a joke", {n_predict: 800})
+//    for await (const chunk of request) {
+//      document.write(chunk.data.content)
+//    }
+//
+export async function* llama(prompt, params = {}, config = {}) {
+  let controller = config.controller;
+  const api_url = config.api_url?.replace(/\/+$/, '') || "";
+
+  if (!controller) {
+    controller = new AbortController();
+  }
+
+  const completionParams = { ...paramDefaults, ...params, prompt };
+
+  const response = await fetch(`${api_url}${config.endpoint || '/completion'}`, {
+    method: 'POST',
+    body: JSON.stringify(completionParams),
+    headers: {
+      'Connection': 'keep-alive',
+      'Content-Type': 'application/json',
+      'Accept': 'text/event-stream',
+      ...(params.api_key ? {'Authorization': `Bearer ${params.api_key}`} : {})
+    },
+    signal: controller.signal,
+  });
+
+  const reader = response.body.getReader();
+  const decoder = new TextDecoder();
+
+  let content = "";
+  let leftover = ""; // Buffer for partially read lines
+
+  try {
+    let cont = true;
+
+    while (cont) {
+      const result = await reader.read();
+      if (result.done) {
+        break;
+      }
+
+      // Add any leftover data to the current chunk of data
+      const text = leftover + decoder.decode(result.value);
+
+      // Check if the last character is a line break
+      const endsWithLineBreak = text.endsWith('\n');
+
+      // Split the text into lines
+      let lines = text.split('\n');
+
+      // If the text doesn't end with a line break, then the last line is incomplete
+      // Store it in leftover to be added to the next chunk of data
+      if (!endsWithLineBreak) {
+        leftover = lines.pop();
+      } else {
+        leftover = ""; // Reset leftover if we have a line break at the end
+      }
+
+      // Parse all sse events and add them to result
+      const regex = /^(\S+):\s(.*)$/gm;
+      for (const line of lines) {
+        const match = regex.exec(line);
+        if (match) {
+          result[match[1]] = match[2];
+          if (result.data === '[DONE]') {
+            cont = false;
+            break;
+          }
+
+          // since we know this is llama.cpp, let's just decode the json in data
+          if (result.data) {
+            result.data = JSON.parse(result.data);
+            content += result.data.content;
+
+            // yield
+            yield result;
+
+            // if we got a stop token from server, we will break here
+            if (result.data.stop) {
+              if (result.data.generation_settings) {
+                generation_settings = result.data.generation_settings;
+              }
+              cont = false;
+              break;
+            }
+          }
+          if (result.error) {
+            try {
+              result.error = JSON.parse(result.error);
+              if (result.error.message.includes('slot unavailable')) {
+                // Throw an error to be caught by upstream callers
+                throw new Error('slot unavailable');
+              } else {
+                console.error(`llama.cpp error [${result.error.code} - ${result.error.type}]: ${result.error.message}`);
+              }
+            } catch(e) {
+              console.error(`llama.cpp error ${result.error}`)
+            }
+          }
+        }
+      }
+    }
+  } catch (e) {
+    if (e.name !== 'AbortError') {
+      console.error("llama error: ", e);
+    }
+    throw e;
+  }
+  finally {
+    controller.abort();
+  }
+
+  return content;
+}
+
+// Call llama, return an event target that you can subscribe to
+//
+// Example:
+//
+//    import { llamaEventTarget } from '/completion.js'
+//
+//    const conn = llamaEventTarget(prompt)
+//    conn.addEventListener("message", (chunk) => {
+//      document.write(chunk.detail.content)
+//    })
+//
+export const llamaEventTarget = (prompt, params = {}, config = {}) => {
+  const eventTarget = new EventTarget();
+  (async () => {
+    let content = "";
+    for await (const chunk of llama(prompt, params, config)) {
+      if (chunk.data) {
+        content += chunk.data.content;
+        eventTarget.dispatchEvent(new CustomEvent("message", { detail: chunk.data }));
+      }
+      if (chunk.data.generation_settings) {
+        eventTarget.dispatchEvent(new CustomEvent("generation_settings", { detail: chunk.data.generation_settings }));
+      }
+      if (chunk.data.timings) {
+        eventTarget.dispatchEvent(new CustomEvent("timings", { detail: chunk.data.timings }));
+      }
+    }
+    eventTarget.dispatchEvent(new CustomEvent("done", { detail: { content } }));
+  })();
+  return eventTarget;
+}
+
+// Call llama, return a promise that resolves to the completed text. This does not support streaming
+//
+// Example:
+//
+//     llamaPromise(prompt).then((content) => {
+//       document.write(content)
+//     })
+//
+//     or
+//
+//     const content = await llamaPromise(prompt)
+//     document.write(content)
+//
+export const llamaPromise = (prompt, params = {}, config = {}) => {
+  return new Promise(async (resolve, reject) => {
+    let content = "";
+    try {
+      for await (const chunk of llama(prompt, params, config)) {
+        content += chunk.data.content;
+      }
+      resolve(content);
+    } catch (error) {
+      reject(error);
+    }
+  });
+};
+
+/**
+ * (deprecated)
+ */
+export const llamaComplete = async (params, controller, callback) => {
+  for await (const chunk of llama(params.prompt, params, { controller })) {
+    callback(chunk);
+  }
+}
+
+// Get the model info from the server. This is useful for getting the context window and so on.
+export const llamaModelInfo = async (config = {}) => {
+  if (!generation_settings) {
+    const api_url = config.api_url?.replace(/\/+$/, '') || "";
+    const props = await fetch(`${api_url}/props`).then(r => r.json());
+    generation_settings = props.default_generation_settings;
+  }
+  return generation_settings;
+}

examples/server/public_legacy/index-new.html

@@ -40,10 +40,15 @@
       repeat_last_n: 0, // 0 = disable penalty, -1 = context size
       repeat_penalty: 1.0, // 1.0 = disabled
       penalize_nl: false, // true only useful for infinite completion
+      dry_multiplier: 0.0, // 0.0 = disabled, 0.8 works well
+      dry_base: 1.75,     // 0.0 = disabled
+      dry_allowed_length: 2, // tokens extending repetitions beyond this receive penalty, 2 works well
+      dry_penalty_last_n: -1, // how many tokens to scan for repetitions (0 = disable penalty, -1 = context size)
       top_k: 0, // <= 0 to use vocab size
       top_p: 1.0, // 1.0 = disabled
       min_p: 0.05, // 0 = disabled; recommended for non-english: ~ 0.4
-      tfs_z: 1.0, // 1.0 = disabled
+      xtc_probability: 0.0, // 0 = disabled;
+      xtc_threshold: 0.1, // > 0.5 disables XTC;
       typical_p: 1.0, // 1.0 = disabled
       presence_penalty: 0.0, // 0.0 = disabled
       frequency_penalty: 0.0, // 0.0 = disabled
@@ -831,11 +836,16 @@ return html`
         <fieldset class="params">
           ${IntField({ label: "Top-K", title: "Limits the selection of the next token to the K most probable tokens. 1 means no randomness = greedy sampling. If set to 0, it means the entire vocabulary size is considered.", max: 100, min: 0, step: 1, name: "top_k", value: params.value.top_k })}
           ${IntField({ label: "Penalize Last N", title: "The last n tokens that are taken into account to penalise repetitions. A value of 0 means that this function is deactivated and -1 means that the entire size of the context is taken into account.", max: 2048, min: 0, step: 16, name: "repeat_last_n", value: params.value.repeat_last_n })}
-          ${FloatField({ label: "Top-P", title: "Limits the selection of the next token to a subset of tokens whose combined probability reaches a threshold value P = top-P. If set to 1, it means the entire vocabulary size is considered.", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })}
           ${FloatField({ label: "Presence Penalty", title: "A penalty that is applied if certain tokens appear repeatedly in the generated text. A higher value leads to fewer repetitions.", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
-          ${FloatField({ label: "TFS-Z", title: "Activates tail-free sampling, a method used to limit the prediction of tokens that are too frequent. The parameter z controls the strength of this limitation. A value of 1.0 means that this function is deactivated.", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
           ${FloatField({ label: "Frequency Penalty", title: "A penalty that is applied based on the frequency with which certain tokens occur in the training data set. A higher value results in rare tokens being favoured.", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}
+          ${FloatField({ label: "Top-P", title: "Limits the selection of the next token to a subset of tokens whose combined probability reaches a threshold value P = top-P. If set to 1, it means the entire vocabulary size is considered.", max: 1.0, min: 0.0, name: "top_p", step: 0.01, value: params.value.top_p })}
           ${FloatField({ label: "Typical-P", title: "Activates local typical sampling, a method used to limit the prediction of tokens that are atypical in the current context. The parameter p controls the strength of this limitation. A value of 1.0 means that this function is deactivated.", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
+          ${FloatField({ label: "XTC probability", title: "Sets the chance for token removal (checked once on sampler start)", max: 1.0, min: 0.0, name: "xtc_probability", step: 0.01, value: params.value.xtc_probability })}
+          ${FloatField({ label: "XTC threshold", title: "Sets a minimum probability threshold for tokens to be removed", max: 0.5, min: 0.0, name: "xtc_threshold", step: 0.01, value: params.value.xtc_threshold })}
+          ${FloatField({ label: "DRY Penalty Multiplier", title: "Set the DRY repetition penalty multiplier. Default is 0.0, which disables DRY.", max: 5.0, min: 0.0, name: "dry_multiplier", step: 0.01, value: params.value.dry_multiplier })}
+          ${FloatField({ label: "DRY Base", title: "Set the DRY repetition penalty base value. Default is 1.75", max: 3.0, min: 1.0, name: "dry_base", step: 0.01, value: params.value.dry_base })}
+          ${IntField({ label: "DRY Allowed Length", title: "Tokens that extend repetition beyond this receive exponentially increasing penalty. Default is 2", max: 10, min: 1, step: 1, name: "dry_allowed_length", value: params.value.dry_allowed_length })}
+          ${IntField({ label: "DRY Penalty Last N", title: "How many tokens to scan for repetitions. Default is -1, where 0 is disabled and -1 is context size", max: 2048, min: -1, step: 16, name: "dry_penalty_last_n", value: params.value.dry_penalty_last_n })}
           ${IntField({ label: "Min Keep", title: "If greater than 0, samplers are forced to return N possible tokens at minimum. Default is 0", max: 10, min: 0, name: "min_keep", value: params.value.min_keep })}
         </fieldset>
 
@@ -1132,12 +1142,15 @@ document.addEventListener('DOMContentLoaded', (event) => {
   const snapSettings = {
     temperature: { snapValue: 1.0, snapRangeMultiplier: 6 },
     min_p: { snapValue: 0.05, snapRangeMultiplier: 2 },
+    xtc_probability: { snapValue: 0.0, snapRangeMultiplier: 4 },
+    xtc_threshold: { snapValue: 0.5, snapRangeMultiplier: 4 },
     top_p: { snapValue: 1.0, snapRangeMultiplier: 4 },
-    tfs_z: { snapValue: 1.0, snapRangeMultiplier: 4 },
     typical_p: { snapValue: 1.0, snapRangeMultiplier: 4 },
     repeat_penalty: { snapValue: 1.0, snapRangeMultiplier: 4 },
     presence_penalty: { snapValue: 0.0, snapRangeMultiplier: 4 },
     frequency_penalty: { snapValue: 0.0, snapRangeMultiplier: 4 },
+    dry_multiplier: { snapValue: 0.0, snapRangeMultiplier: 4 },
+    dry_base: { snapValue: 1.75, snapRangeMultiplier: 4 },
   };
   // add an event listener for each slider
   Object.keys(snapSettings).forEach(sliderName => {

examples/server/public_legacy/json-schema-to-grammar.mjs

@@ -529,7 +529,7 @@ export class SchemaConverter {
       return joinSeq();
     };
 
-    return this._addRule(name, "\"\\\"\" " + toRule(transform()) + " \"\\\"\" space")
+    return this._addRule(name, "\"\\\"\" (" + toRule(transform()) + ") \"\\\"\" space")
   }
 
   _notStrings(strings) {

examples/server/public_legacy/loading.html

@@ -0,0 +1,12 @@
+<!DOCTYPE html>
+<html>
+    <head>
+        <meta http-equiv="refresh" content="5">
+    </head>
+    <body>
+        <div id="loading">
+            The model is loading. Please wait.<br/>
+            The user interface will appear soon.
+        </div>
+    </body>
+</html>

examples/server/tests/features/ctx_shift.feature

@@ -13,6 +13,10 @@ Feature: llama.cpp server
     And   32 as batch size
     And   2 slots
 
+    # the prompt is 301 tokens
+    # the slot context is 256/2 = 128 tokens
+    # the prompt is truncated to keep the last 109 tokens
+    # 64 tokens are generated thanks to shifting the context when it gets full
   Scenario: Inference with context shift
     And   64 server max tokens to predict
     Then  the server is starting

examples/server/tests/features/infill.feature

@@ -0,0 +1,36 @@
+@llama.cpp
+@infill
+Feature: llama.cpp server
+
+  # The current model is made by adding FIM tokens to the existing stories260K
+  # We may want to use a better model in the future, maybe something like SmolLM 360M
+
+  Background: Server startup
+    Given a server listening on localhost:8080
+    And   a model file tinyllamas/stories260K-infill.gguf from HF repo ggml-org/models
+    And   a model file test-model-infill.gguf
+    And   a model alias tinyllama-infill
+    And   42 as server seed
+    And   1024 as batch size
+    And   1024 as ubatch size
+    And   2048 KV cache size
+    And   64 max tokens to predict
+    And   0.0 temperature
+    Then  the server is starting
+    Then  the server is healthy
+
+  Scenario: Infill without input_extra
+    Given a prompt "Complete this"
+    And   an infill input extra none none
+    And   an infill input prefix "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n    int n_threads = llama_"
+    And   an infill input suffix "}\n"
+    And   an infill request with no api error
+    Then  64 tokens are predicted matching One|day|she|saw|big|scary|bird
+
+  Scenario: Infill with input_extra
+    Given a prompt "Complete this"
+    And   an infill input extra "llama.h" "LLAMA_API int32_t llama_n_threads();\n"
+    And   an infill input prefix "#include <cstdio>\n#include \"llama.h\"\n\nint main() {\n    int n_threads = llama_"
+    And   an infill input suffix "}\n"
+    And   an infill request with no api error
+    Then  64 tokens are predicted matching cuts|Jimmy|mom|came|into|the|room"

examples/server/tests/features/security.feature

@@ -64,5 +64,5 @@ Feature: Security
       | localhost       | Access-Control-Allow-Origin      | localhost         |
       | web.mydomain.fr | Access-Control-Allow-Origin      | web.mydomain.fr   |
       | origin          | Access-Control-Allow-Credentials | true              |
-      | web.mydomain.fr | Access-Control-Allow-Methods     | POST              |
+      | web.mydomain.fr | Access-Control-Allow-Methods     | GET, POST         |
       | web.mydomain.fr | Access-Control-Allow-Headers     | *                 |

examples/server/tests/features/steps/steps.py

@@ -80,6 +80,11 @@ def step_server_config(context, server_fqdn: str, server_port: str):
     context.lora_file = None
     context.disable_ctx_shift = False
 
+    # infill
+    context.infill_input_extra = None
+    context.infill_input_suffix = ''
+    context.infill_input_prefix = ''
+
     context.tasks_result = []
     context.concurrent_tasks = []
     context.prompts = []
@@ -255,13 +260,13 @@ async def step_wait_for_server_status(context, expecting_status: Literal['health
 async def step_all_slots_status(context, expected_slot_status_string: Literal['idle', 'busy'] | str):
     match expected_slot_status_string:
         case 'idle':
-            expected_slot_status = 0
+            expected_slot_status = False
         case 'busy':
-            expected_slot_status = 1
+            expected_slot_status = True
         case _:
             assert False, "unknown status"
 
-    expected_slots = [{'id': slot_id, 'state': expected_slot_status}
+    expected_slots = [{'id': slot_id, 'is_processing': expected_slot_status}
                       for slot_id in range(context.n_slots)]
     await request_slots_status(context, expected_slots)
 
@@ -291,6 +296,28 @@ async def step_request_completion(context, api_error: Literal['raised'] | str):
         assert completion == api_error_code, f"completion must be an {api_error_code} status code: {completion}"
 
 
+@step('an infill request with {api_error} api error')
+@async_run_until_complete
+async def step_request_completion(context, api_error: Literal['raised'] | str):
+    if api_error != 'no':
+        raise ValueError(f'api_error={api_error} is not yet implemented')
+    payload = {
+        "prompt": context.prompts[0],
+        "input_suffix": context.infill_input_suffix,
+        "input_prefix": context.infill_input_prefix,
+        "n_predict": context.n_predict,
+        "seed": context.seed,
+        "temperature": context.temperature,
+    }
+    if context.infill_input_extra is not None:
+        payload['input_extra'] = context.infill_input_extra
+    async with aiohttp.ClientSession(timeout=DEFAULT_TIMEOUT_SECONDS) as session:
+        async with session.post(f'{context.base_url}/infill',
+                                json=payload) as response:
+            assert response.status == 200
+            context.tasks_result = [await response.json()]
+
+
 @step('{predicted_n:d} tokens are predicted matching {re_content}')
 def step_n_tokens_predicted_with_content(context, predicted_n, re_content):
     context.completion = context.tasks_result.pop()
@@ -539,6 +566,25 @@ def step_a_prompt_prompt(context, prompt):
     context.n_prompts = len(context.prompts)
 
 
+# TODO: allow this to be repeated
+@step('an infill input extra {filename} {text}')
+def step_infill_input_extra(context, filename, text):
+    if filename == 'none':
+        context.infill_input_extra = None
+    else:
+        context.infill_input_extra = [{'filename': filename, 'text': text}]
+
+
+@step('an infill input suffix {text}')
+def step_infill_input_suffix(context, text):
+    context.infill_input_suffix = text
+
+
+@step('an infill input prefix {text}')
+def step_infill_input_prefix(context, text):
+    context.infill_input_prefix = text
+
+
 @step('{num_prompts:d} prompts {prompt} with seed {seed:d}')
 def step_many_prompts(context, num_prompts, prompt, seed):
     if context.seed is None:
@@ -1308,8 +1354,8 @@ async def wait_for_slots_status(context,
                 if status_code == 503 and status_code == expected_http_status_code:
                     return
                 if status_code == 200 and status_code == expected_http_status_code:
-                    n_slots_idle = sum(1 if slot["state"] == 0 else 0 for slot in slots)
-                    n_slots_processing = sum(1 if slot["state"] != 0 else 0 for slot in slots)
+                    n_slots_idle = sum(1 if not slot["is_processing"] else 0 for slot in slots)
+                    n_slots_processing = sum(1 if slot["is_processing"] else 0 for slot in slots)
                     if ((slots_idle is None or slots_idle == n_slots_idle)
                         and (slots_processing is None or slots_processing == n_slots_processing)):
                         return

examples/server/themes/buttons-top/index.html

@@ -226,7 +226,6 @@
       top_k: 40, // <= 0 to use vocab size
       top_p: 0.95, // 1.0 = disabled
       min_p: 0.05, // 0 = disabled
-      tfs_z: 1.0, // 1.0 = disabled
       typical_p: 1.0, // 1.0 = disabled
       presence_penalty: 0.0, // 0.0 = disabled
       frequency_penalty: 0.0, // 0.0 = disabled
@@ -788,7 +787,6 @@
           <details>
             <summary>More options</summary>
             <fieldset class="two">
-              ${FloatField({ label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
               ${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
               ${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
               ${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}

examples/server/themes/wild/index.html

@@ -229,7 +229,6 @@
       top_k: 40, // <= 0 to use vocab size
       top_p: 0.95, // 1.0 = disabled
       min_p: 0.05, // 0 = disabled
-      tfs_z: 1.0, // 1.0 = disabled
       typical_p: 1.0, // 1.0 = disabled
       presence_penalty: 0.0, // 0.0 = disabled
       frequency_penalty: 0.0, // 0.0 = disabled
@@ -791,7 +790,6 @@
           <details>
             <summary>More options</summary>
             <fieldset class="two">
-              ${FloatField({ label: "TFS-Z", max: 1.0, min: 0.0, name: "tfs_z", step: 0.01, value: params.value.tfs_z })}
               ${FloatField({ label: "Typical P", max: 1.0, min: 0.0, name: "typical_p", step: 0.01, value: params.value.typical_p })}
               ${FloatField({ label: "Presence penalty", max: 1.0, min: 0.0, name: "presence_penalty", step: 0.01, value: params.value.presence_penalty })}
               ${FloatField({ label: "Frequency penalty", max: 1.0, min: 0.0, name: "frequency_penalty", step: 0.01, value: params.value.frequency_penalty })}

examples/server/utils.hpp

@@ -24,6 +24,22 @@
 #define DEFAULT_OAICOMPAT_MODEL "gpt-3.5-turbo-0613"
 
 using json = nlohmann::ordered_json;
+using llama_tokens = std::vector<llama_token>;
+
+#define SLT_INF(slot, fmt, ...) LOG_INF("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
+#define SLT_WRN(slot, fmt, ...) LOG_WRN("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
+#define SLT_ERR(slot, fmt, ...) LOG_ERR("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
+#define SLT_DBG(slot, fmt, ...) LOG_DBG("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, (slot).id_task, __VA_ARGS__)
+
+#define SRV_INF(fmt, ...) LOG_INF("srv  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
+#define SRV_WRN(fmt, ...) LOG_WRN("srv  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
+#define SRV_ERR(fmt, ...) LOG_ERR("srv  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
+#define SRV_DBG(fmt, ...) LOG_DBG("srv  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
+
+#define QUE_INF(fmt, ...) LOG_INF("que  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
+#define QUE_WRN(fmt, ...) LOG_WRN("que  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
+#define QUE_ERR(fmt, ...) LOG_ERR("que  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
+#define QUE_DBG(fmt, ...) LOG_DBG("que  %12.*s: " fmt, 12, __func__, __VA_ARGS__)
 
 // https://community.openai.com/t/openai-chat-list-of-error-codes-and-types/357791/11
 enum error_type {
@@ -52,9 +68,237 @@ static T json_value(const json & body, const std::string & key, const T & defaul
 }
 
 //
-// chat template utils
+// tokenizer and input processing utils
 //
 
+static bool json_is_array_of_numbers(const json & data) {
+    if (data.is_array()) {
+        for (const auto & e : data) {
+            if (!e.is_number_integer()) {
+                return false;
+            }
+        }
+        return true;
+    }
+    return false;
+}
+
+// is array having BOTH numbers & strings?
+static bool json_is_array_of_mixed_numbers_strings(const json & data) {
+    bool seen_string = false;
+    bool seen_number = false;
+    if (data.is_array()) {
+        for (const auto & e : data) {
+            seen_string |= e.is_string();
+            seen_number |= e.is_number_integer();
+            if (seen_number && seen_string) {
+                return true;
+            }
+        }
+    }
+    return false;
+}
+
+/**
+ * this handles 2 cases:
+ * - only string, example: "string"
+ * - mixed string and tokens, example: [12, 34, "string", 56, 78]
+ */
+static llama_tokens tokenize_mixed(const llama_context * ctx, const json & json_prompt, bool add_special, bool parse_special) {
+    // If `add_bos` is true, we only add BOS, when json_prompt is a string,
+    // or the first element of the json_prompt array is a string.
+    llama_tokens prompt_tokens;
+
+    if (json_prompt.is_array()) {
+        bool first = true;
+        for (const auto & p : json_prompt) {
+            if (p.is_string()) {
+                auto s = p.template get<std::string>();
+
+                llama_tokens p;
+                if (first) {
+                    p = common_tokenize(ctx, s, add_special, parse_special);
+                    first = false;
+                } else {
+                    p = common_tokenize(ctx, s, false, parse_special);
+                }
+
+                prompt_tokens.insert(prompt_tokens.end(), p.begin(), p.end());
+            } else {
+                if (first) {
+                    first = false;
+                }
+
+                prompt_tokens.push_back(p.template get<llama_token>());
+            }
+        }
+    } else {
+        auto s = json_prompt.template get<std::string>();
+        prompt_tokens = common_tokenize(ctx, s, add_special, parse_special);
+    }
+
+    return prompt_tokens;
+}
+
+/**
+ * break the input "prompt" object into multiple prompt if needed, then tokenize them
+ * this supports these cases:
+ * - "prompt": "string"
+ * - "prompt": [12, 34, 56]
+ * - "prompt": [12, 34, "string", 56, 78]
+ * and multiple prompts (multi-tasks):
+ * - "prompt": ["string1", "string2"]
+ * - "prompt": ["string1", [12, 34, 56]]
+ * - "prompt": [[12, 34, "string", 56, 78], [12, 34, 56]]
+ */
+static std::vector<llama_tokens> tokenize_input_prompts(llama_context * ctx, const json & json_prompt, bool add_special, bool parse_special) {
+    std::vector<llama_tokens> result;
+    if (json_prompt.is_string() || json_is_array_of_mixed_numbers_strings(json_prompt)) {
+        // string or mixed
+        result.push_back(tokenize_mixed(ctx, json_prompt, add_special, parse_special));
+    } else if (json_is_array_of_numbers(json_prompt)) {
+        // array of tokens
+        result.push_back(json_prompt.get<llama_tokens>());
+    } else if (json_prompt.is_array()) {
+        // array of prompts
+        result.reserve(json_prompt.size());
+        for (const auto & p : json_prompt) {
+            if (p.is_string() || json_is_array_of_mixed_numbers_strings(p)) {
+                result.push_back(tokenize_mixed(ctx, p, add_special, parse_special));
+            } else if (json_is_array_of_numbers(p)) {
+                // array of tokens
+                result.push_back(p.get<llama_tokens>());
+            } else {
+                throw std::runtime_error("element of \"prompt\" must be a string, an list of tokens, or a list of mixed strings & tokens");
+            }
+        }
+    } else {
+        throw std::runtime_error("\"prompt\" must be a string, an list of tokens, a list of mixed strings & tokens, or a list of prompts");
+    }
+    return result;
+}
+
+//
+// template utils
+//
+
+// format rerank task: [BOS]query[EOS][SEP]doc[EOS]
+static llama_tokens format_rerank(const struct llama_model * model, const llama_tokens & query, const llama_tokens & doc) {
+    llama_tokens result;
+    result.reserve(doc.size() + query.size() + 4);
+    result.push_back(llama_token_bos(model));
+    result.insert(result.end(), query.begin(), query.end());
+    result.push_back(llama_token_eos(model));
+    result.push_back(llama_token_sep(model));
+    result.insert(result.end(), doc.begin(), doc.end());
+    result.push_back(llama_token_eos(model));
+    return result;
+}
+
+// format infill task
+static llama_tokens format_infill(
+        const llama_context * ctx,
+        const json & input_prefix,
+        const json & input_suffix,
+        const json & input_extra,
+        const int n_batch,
+        const int n_predict,
+        const int n_ctx,
+        const bool spm_infill,
+        const llama_tokens & tokens_prompt
+    ) {
+    // TODO: optimize this block by reducing memory allocations and movement
+
+    // use FIM repo-level pattern:
+    // ref: https://arxiv.org/pdf/2409.12186
+    //
+    // [FIM_REP]myproject
+    // [FIM_SEP]filename0
+    // extra chunk 0
+    // [FIM_SEP]filename1
+    // extra chunk 1
+    // ...
+    // [FIM_SEP]filename
+    // [FIM_PRE]prefix[FIM_SUF]suffix[FIM_MID]prompt
+    //
+    llama_tokens extra_tokens;
+    extra_tokens.reserve(n_ctx);
+
+    auto model = llama_get_model(ctx);
+    auto tokens_prefix = tokenize_mixed(ctx, input_prefix, false, false);
+    auto tokens_suffix = tokenize_mixed(ctx, input_suffix, false, false);
+
+    if (llama_token_fim_rep(model) != LLAMA_TOKEN_NULL) {
+        // TODO: make project name an input
+        static const auto k_fim_repo = common_tokenize(ctx, "myproject\n", false, false);
+
+        extra_tokens.push_back(llama_token_fim_rep(model));
+        extra_tokens.insert(extra_tokens.end(), k_fim_repo.begin(), k_fim_repo.end());
+    }
+    for (const auto & chunk : input_extra) {
+        // { "text": string, "filename": string }
+        const std::string text     = json_value(chunk, "text",     std::string());
+        const std::string filename = json_value(chunk, "filename", std::string("tmp"));
+
+        if (llama_token_fim_sep(model) != LLAMA_TOKEN_NULL) {
+            const auto k_fim_file = common_tokenize(ctx, filename + "\n", false, false);
+
+            extra_tokens.insert(extra_tokens.end(), llama_token_fim_sep(model));
+            extra_tokens.insert(extra_tokens.end(), k_fim_file.begin(), k_fim_file.end());
+        } else {
+            // chunk separator in binary form to avoid confusing the AI
+            static const char k_chunk_prefix_str[] = {0x0a, 0x0a, 0x2d, 0x2d, 0x2d, 0x20, 0x73, 0x6e, 0x69, 0x70, 0x70, 0x65, 0x74, 0x20, 0x2d, 0x2d, 0x2d, 0x0a, 0x0a, 0x00};
+            static const auto k_chunk_prefix_tokens = common_tokenize(ctx, k_chunk_prefix_str, false, false);
+
+            extra_tokens.insert(extra_tokens.end(), k_chunk_prefix_tokens.begin(), k_chunk_prefix_tokens.end());
+        }
+
+        const auto chunk_tokens = common_tokenize(ctx, text, false, false);
+        extra_tokens.insert(extra_tokens.end(), chunk_tokens.begin(), chunk_tokens.end());
+    }
+
+    if (llama_token_fim_sep(model) != LLAMA_TOKEN_NULL) {
+        // TODO: current filename
+        static const auto k_fim_file = common_tokenize(ctx, "filename\n", false, false);
+
+        extra_tokens.insert(extra_tokens.end(), llama_token_fim_sep(model));
+        extra_tokens.insert(extra_tokens.end(), k_fim_file.begin(), k_fim_file.end());
+    }
+
+    // for now pick FIM context to fit in a batch (ratio prefix:suffix = 3:1, TODO: configurable?)
+    const int n_prefix_take = std::min<int>(tokens_prefix.size(),                3*(n_batch/4));
+    const int n_suffix_take = std::min<int>(tokens_suffix.size(), std::max<int>(0, (n_batch/4) - (2 + tokens_prompt.size())));
+
+    SRV_DBG("n_prefix_take = %d, n_suffix_take = %d, total = %d\n", n_prefix_take, n_suffix_take, (n_prefix_take + n_suffix_take));
+
+    // fill the rest of the context with extra chunks
+    const int n_extra_take = std::min<int>(std::max<int>(0, n_ctx - (n_batch) - 2*n_predict), extra_tokens.size());
+
+    tokens_prefix.erase(tokens_prefix.begin(), tokens_prefix.begin() + tokens_prefix.size() - n_prefix_take);
+    tokens_suffix.resize(n_suffix_take);
+
+    tokens_prefix.insert(tokens_prefix.begin(), llama_token_fim_pre(model));
+    tokens_prefix.insert(tokens_prefix.end(),   tokens_prompt.begin(), tokens_prompt.end());
+    tokens_suffix.insert(tokens_suffix.begin(), llama_token_fim_suf(model));
+
+    auto embd_inp = spm_infill ? tokens_suffix : tokens_prefix;
+    auto embd_end = spm_infill ? tokens_prefix : tokens_suffix;
+
+    if (llama_add_bos_token(model)) {
+        embd_inp.insert(embd_inp.begin(), llama_token_bos(model));
+    }
+
+    SRV_DBG("extra: n_ctx = %d, n_extra_take = %d, n_extra = %d\n", n_ctx, n_extra_take, (int) extra_tokens.size());
+
+    // put the extra context before the FIM prefix
+    embd_inp.insert(embd_inp.begin(), extra_tokens.end() - n_extra_take, extra_tokens.end());
+
+    embd_inp.insert(embd_inp.end(), embd_end.begin(), embd_end.end());
+    embd_inp.push_back(llama_token_fim_mid(model));
+
+    return embd_inp;
+}
+
 // Format given chat. If tmpl is empty, we take the template from model metadata
 inline std::string format_chat(const struct llama_model * model, const std::string & tmpl, const std::vector<json> & messages) {
     std::vector<common_chat_msg> chat;
@@ -195,18 +439,60 @@ static std::string gen_chatcmplid() {
 // other common utils
 //
 
-static size_t common_part(const std::vector<llama_token> & a, const std::vector<llama_token> & b) {
+static size_t longest_common_prefix(const llama_tokens & a, const llama_tokens & b) {
     size_t i;
     for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
 
     return i;
 }
 
-static size_t common_part(const std::string & a, const std::string & b) {
-    size_t i;
-    for (i = 0; i < a.size() && i < b.size() && a[i] == b[i]; i++) {}
+static size_t longest_common_subsequence(const llama_tokens & a, const llama_tokens & b) {
+    // check for empty sequences
+    if (a.empty() || b.empty()) {
+        return 0;
+    }
 
-    return i;
+    // get the lengths of the input sequences
+    size_t a_len = a.size();
+    size_t b_len = b.size();
+
+    // initialize the maximum length of the longest common subsequence (LCS)
+    size_t max_length = 0;
+
+    // use two rows instead of a 2D matrix to optimize space
+    std::vector<size_t> prev_row(b_len + 1, 0);
+    std::vector<size_t> curr_row(b_len + 1, 0);
+
+    // iterate through the elements of a
+    for (size_t i = 1; i <= a_len; i++) {
+        // iterate through the elements of b
+        for (size_t j = 1; j <= b_len; j++) {
+            // if elements at the current positions match
+            if (a[i - 1] == b[j - 1]) {
+                // if it's the first element of either sequences, set LCS length to 1
+                if (i == 1 || j == 1) {
+                    curr_row[j] = 1;
+                } else {
+                    // increment LCS length by 1 compared to the previous element
+                    curr_row[j] = prev_row[j - 1] + 1;
+                }
+
+                // update max_length if necessary
+                if (curr_row[j] > max_length) {
+                    max_length = curr_row[j];
+                }
+            } else {
+                // reset LCS length if elements don't match
+                curr_row[j] = 0;
+            }
+        }
+
+        // update the previous row for the next iteration
+        prev_row = curr_row;
+    }
+
+    // return the maximum length of the LCS
+    return max_length;
 }
 
 static bool ends_with(const std::string & str, const std::string & suffix) {
@@ -229,18 +515,6 @@ static size_t find_partial_stop_string(const std::string &stop, const std::strin
     return std::string::npos;
 }
 
-static bool json_is_array_of_numbers(const json & data) {
-    if (data.is_array()) {
-        for (const auto & e : data) {
-            if (!e.is_number()) {
-                return false;
-            }
-        }
-        return true;
-    }
-    return false;
-}
-
 // TODO: reuse llama_detokenize
 template <class Iter>
 static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) {
@@ -360,9 +634,9 @@ static json oaicompat_completion_params_parse(
 
     // Handle "logprobs" field
     // TODO: The response format of this option is not yet OAI-compatible, but seems like no one really using it; We may need to fix it in the future
-    if (body.contains("logprobs")) {
+    if (json_value(body, "logprobs", false)) {
         llama_params["n_probs"] = json_value(body, "top_logprobs", 20);
-    } else if (body.contains("top_logprobs")) {
+    } else if (body.contains("top_logprobs") && !body.at("top_logprobs").is_null()) {
         throw std::runtime_error("top_logprobs requires logprobs to be set to true");
     }
 
@@ -375,7 +649,7 @@ static json oaicompat_completion_params_parse(
     }
 
     // Copy remaining properties to llama_params
-    // This allows user to use llama.cpp-specific params like "mirostat", "tfs_z",... via OAI endpoint.
+    // This allows user to use llama.cpp-specific params like "mirostat", ... via OAI endpoint.
     // See "launch_slot_with_task()" for a complete list of params supported by llama.cpp
     for (const auto & item : body.items()) {
         // Exception: if "n_predict" is present, we overwrite the value specified earlier by "max_tokens"

examples/simple-chat/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET llama-simple-chat)
+add_executable(${TARGET} simple-chat.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/simple-chat/README.md

@@ -0,0 +1,7 @@
+# llama.cpp/example/simple-chat
+
+The purpose of this example is to demonstrate a minimal usage of llama.cpp to create a simple chat program using the chat template from the GGUF file.
+
+```bash
+./llama-simple-chat -m Meta-Llama-3.1-8B-Instruct.gguf -c 2048
+...

examples/simple-chat/simple-chat.cpp

@@ -0,0 +1,197 @@
+#include "llama.h"
+#include <cstdio>
+#include <cstring>
+#include <iostream>
+#include <string>
+#include <vector>
+
+static void print_usage(int, char ** argv) {
+    printf("\nexample usage:\n");
+    printf("\n    %s -m model.gguf [-c context_size] [-ngl n_gpu_layers]\n", argv[0]);
+    printf("\n");
+}
+
+int main(int argc, char ** argv) {
+    std::string model_path;
+    int ngl = 99;
+    int n_ctx = 2048;
+
+    // parse command line arguments
+    for (int i = 1; i < argc; i++) {
+        try {
+            if (strcmp(argv[i], "-m") == 0) {
+                if (i + 1 < argc) {
+                    model_path = argv[++i];
+                } else {
+                    print_usage(argc, argv);
+                    return 1;
+                }
+            } else if (strcmp(argv[i], "-c") == 0) {
+                if (i + 1 < argc) {
+                    n_ctx = std::stoi(argv[++i]);
+                } else {
+                    print_usage(argc, argv);
+                    return 1;
+                }
+            } else if (strcmp(argv[i], "-ngl") == 0) {
+                if (i + 1 < argc) {
+                    ngl = std::stoi(argv[++i]);
+                } else {
+                    print_usage(argc, argv);
+                    return 1;
+                }
+            } else {
+                print_usage(argc, argv);
+                return 1;
+            }
+        } catch (std::exception & e) {
+            fprintf(stderr, "error: %s\n", e.what());
+            print_usage(argc, argv);
+            return 1;
+        }
+    }
+    if (model_path.empty()) {
+        print_usage(argc, argv);
+        return 1;
+    }
+
+    // only print errors
+    llama_log_set([](enum ggml_log_level level, const char * text, void * /* user_data */) {
+        if (level >= GGML_LOG_LEVEL_ERROR) {
+            fprintf(stderr, "%s", text);
+        }
+    }, nullptr);
+
+    // initialize the model
+    llama_model_params model_params = llama_model_default_params();
+    model_params.n_gpu_layers = ngl;
+
+    llama_model * model = llama_load_model_from_file(model_path.c_str(), model_params);
+    if (!model) {
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        return 1;
+    }
+
+    // initialize the context
+    llama_context_params ctx_params = llama_context_default_params();
+    ctx_params.n_ctx = n_ctx;
+    ctx_params.n_batch = n_ctx;
+
+    llama_context * ctx = llama_new_context_with_model(model, ctx_params);
+    if (!ctx) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
+    // initialize the sampler
+    llama_sampler * smpl = llama_sampler_chain_init(llama_sampler_chain_default_params());
+    llama_sampler_chain_add(smpl, llama_sampler_init_min_p(0.05f, 1));
+    llama_sampler_chain_add(smpl, llama_sampler_init_temp(0.8f));
+    llama_sampler_chain_add(smpl, llama_sampler_init_dist(LLAMA_DEFAULT_SEED));
+
+    // helper function to evaluate a prompt and generate a response
+    auto generate = [&](const std::string & prompt) {
+        std::string response;
+
+        // tokenize the prompt
+        const int n_prompt_tokens = -llama_tokenize(model, prompt.c_str(), prompt.size(), NULL, 0, true, true);
+        std::vector<llama_token> prompt_tokens(n_prompt_tokens);
+        if (llama_tokenize(model, prompt.c_str(), prompt.size(), prompt_tokens.data(), prompt_tokens.size(), llama_get_kv_cache_used_cells(ctx) == 0, true) < 0) {
+            GGML_ABORT("failed to tokenize the prompt\n");
+        }
+
+        // prepare a batch for the prompt
+        llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
+        llama_token new_token_id;
+        while (true) {
+            // check if we have enough space in the context to evaluate this batch
+            int n_ctx = llama_n_ctx(ctx);
+            int n_ctx_used = llama_get_kv_cache_used_cells(ctx);
+            if (n_ctx_used + batch.n_tokens > n_ctx) {
+                printf("\033[0m\n");
+                fprintf(stderr, "context size exceeded\n");
+                exit(0);
+            }
+
+            if (llama_decode(ctx, batch)) {
+                GGML_ABORT("failed to decode\n");
+            }
+
+            // sample the next token
+            new_token_id = llama_sampler_sample(smpl, ctx, -1);
+
+            // is it an end of generation?
+            if (llama_token_is_eog(model, new_token_id)) {
+                break;
+            }
+
+            // convert the token to a string, print it and add it to the response
+            char buf[256];
+            int n = llama_token_to_piece(model, new_token_id, buf, sizeof(buf), 0, true);
+            if (n < 0) {
+                GGML_ABORT("failed to convert token to piece\n");
+            }
+            std::string piece(buf, n);
+            printf("%s", piece.c_str());
+            fflush(stdout);
+            response += piece;
+
+            // prepare the next batch with the sampled token
+            batch = llama_batch_get_one(&new_token_id, 1);
+        }
+
+        return response;
+    };
+
+    std::vector<llama_chat_message> messages;
+    std::vector<char> formatted(llama_n_ctx(ctx));
+    int prev_len = 0;
+    while (true) {
+        // get user input
+        printf("\033[32m> \033[0m");
+        std::string user;
+        std::getline(std::cin, user);
+
+        if (user.empty()) {
+            break;
+        }
+
+        // add the user input to the message list and format it
+        messages.push_back({"user", strdup(user.c_str())});
+        int new_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), true, formatted.data(), formatted.size());
+        if (new_len > (int)formatted.size()) {
+            formatted.resize(new_len);
+            new_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), true, formatted.data(), formatted.size());
+        }
+        if (new_len < 0) {
+            fprintf(stderr, "failed to apply the chat template\n");
+            return 1;
+        }
+
+        // remove previous messages to obtain the prompt to generate the response
+        std::string prompt(formatted.begin() + prev_len, formatted.begin() + new_len);
+
+        // generate a response
+        printf("\033[33m");
+        std::string response = generate(prompt);
+        printf("\n\033[0m");
+
+        // add the response to the messages
+        messages.push_back({"assistant", strdup(response.c_str())});
+        prev_len = llama_chat_apply_template(model, nullptr, messages.data(), messages.size(), false, nullptr, 0);
+        if (prev_len < 0) {
+            fprintf(stderr, "failed to apply the chat template\n");
+            return 1;
+        }
+    }
+
+    // free resources
+    for (auto & msg : messages) {
+        free(const_cast<char *>(msg.content));
+    }
+    llama_sampler_free(smpl);
+    llama_free(ctx);
+    llama_free_model(model);
+
+    return 0;
+}

examples/simple/simple.cpp

@@ -138,7 +138,7 @@ int main(int argc, char ** argv) {
 
     // prepare a batch for the prompt
 
-    llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size(), 0, 0);
+    llama_batch batch = llama_batch_get_one(prompt_tokens.data(), prompt_tokens.size());
 
     // main loop
 
@@ -175,7 +175,7 @@ int main(int argc, char ** argv) {
             fflush(stdout);
 
             // prepare the next batch with the sampled token
-            batch = llama_batch_get_one(&new_token_id, 1, n_pos, 0);
+            batch = llama_batch_get_one(&new_token_id, 1);
 
             n_decode += 1;
         }

examples/speculative/speculative.cpp

@@ -39,6 +39,11 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
+    if (params.n_predict < -1) {
+        LOG_ERR("%s: --n-predict must be >= -1\n", __func__);
+        return 1;
+    }
+
     common_init();
 
     if (params.model_draft.empty()) {
@@ -155,9 +160,9 @@ int main(int argc, char ** argv) {
     const auto t_enc_start = ggml_time_us();
 
     // eval the prompt with both models
-    llama_decode(ctx_tgt, llama_batch_get_one( inp.data(), n_input - 1, 0,           0));
-    llama_decode(ctx_tgt, llama_batch_get_one(&inp.back(),           1, n_input - 1, 0));
-    llama_decode(ctx_dft, llama_batch_get_one( inp.data(), n_input,     0,           0));
+    llama_decode(ctx_tgt, llama_batch_get_one( inp.data(), n_input - 1));
+    llama_decode(ctx_tgt, llama_batch_get_one(&inp.back(),           1));
+    llama_decode(ctx_dft, llama_batch_get_one( inp.data(), n_input));
 
     const auto t_enc_end = ggml_time_us();
 
@@ -180,8 +185,6 @@ int main(int argc, char ** argv) {
     // target model sampling context (reuse the llama_context's sampling instance)
     struct common_sampler * smpl = common_sampler_init(model_tgt, params.sparams);
 
-    struct llama_sampler * softmax = llama_sampler_init_softmax();
-
     // draft sequence data
     std::vector<seq_draft> drafts(n_seq_dft);
 
@@ -190,8 +193,8 @@ int main(int argc, char ** argv) {
         drafts[s].smpl = common_sampler_init(model_dft, params.sparams);
     }
 
-    llama_batch batch_dft = llama_batch_init(params.n_ctx, 0, 1);
-    llama_batch batch_tgt = llama_batch_init(params.n_ctx, 0, n_seq_dft);
+    llama_batch batch_dft = llama_batch_init(llama_n_batch(ctx_dft), 0, 1);
+    llama_batch batch_tgt = llama_batch_init(llama_n_batch(ctx_tgt), 0, n_seq_dft);
 
     const auto t_dec_start = ggml_time_us();
 
@@ -441,7 +444,7 @@ int main(int argc, char ** argv) {
             ++n_past_dft;
         }
 
-        if (n_predict > params.n_predict || has_eos) {
+        if ((params.n_predict >= 0 && n_predict > params.n_predict) || has_eos) {
             break;
         }
 
@@ -624,7 +627,6 @@ int main(int argc, char ** argv) {
         common_sampler_free(drafts[s].smpl);
     }
 
-    llama_sampler_free(softmax);
     llama_batch_free(batch_dft);
 
     llama_free(ctx_tgt);

flake.lock

@@ -5,11 +5,11 @@
         "nixpkgs-lib": "nixpkgs-lib"
       },
       "locked": {
-        "lastModified": 1727826117,
-        "narHash": "sha256-K5ZLCyfO/Zj9mPFldf3iwS6oZStJcU4tSpiXTMYaaL0=",
+        "lastModified": 1730504689,
+        "narHash": "sha256-hgmguH29K2fvs9szpq2r3pz2/8cJd2LPS+b4tfNFCwE=",
         "owner": "hercules-ci",
         "repo": "flake-parts",
-        "rev": "3d04084d54bedc3d6b8b736c70ef449225c361b1",
+        "rev": "506278e768c2a08bec68eb62932193e341f55c90",
         "type": "github"
       },
       "original": {
@@ -20,11 +20,11 @@
     },
     "nixpkgs": {
       "locked": {
-        "lastModified": 1728018373,
-        "narHash": "sha256-NOiTvBbRLIOe5F6RbHaAh6++BNjsb149fGZd1T4+KBg=",
+        "lastModified": 1730200266,
+        "narHash": "sha256-l253w0XMT8nWHGXuXqyiIC/bMvh1VRszGXgdpQlfhvU=",
         "owner": "NixOS",
         "repo": "nixpkgs",
-        "rev": "bc947f541ae55e999ffdb4013441347d83b00feb",
+        "rev": "807e9154dcb16384b1b765ebe9cd2bba2ac287fd",
         "type": "github"
       },
       "original": {
@@ -36,14 +36,14 @@
     },
     "nixpkgs-lib": {
       "locked": {
-        "lastModified": 1727825735,
-        "narHash": "sha256-0xHYkMkeLVQAMa7gvkddbPqpxph+hDzdu1XdGPJR+Os=",
+        "lastModified": 1730504152,
+        "narHash": "sha256-lXvH/vOfb4aGYyvFmZK/HlsNsr/0CVWlwYvo2rxJk3s=",
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/fb192fec7cc7a4c26d51779e9bab07ce6fa5597a.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/cc2f28000298e1269cea6612cd06ec9979dd5d7f.tar.gz"
       },
       "original": {
         "type": "tarball",
-        "url": "https://github.com/NixOS/nixpkgs/archive/fb192fec7cc7a4c26d51779e9bab07ce6fa5597a.tar.gz"
+        "url": "https://github.com/NixOS/nixpkgs/archive/cc2f28000298e1269cea6612cd06ec9979dd5d7f.tar.gz"
       }
     },
     "root": {

ggml/CMakeLists.txt

@@ -99,6 +99,9 @@ option(GGML_AVX512      "ggml: enable AVX512"           OFF)
 option(GGML_AVX512_VBMI "ggml: enable AVX512-VBMI"      OFF)
 option(GGML_AVX512_VNNI "ggml: enable AVX512-VNNI"      OFF)
 option(GGML_AVX512_BF16 "ggml: enable AVX512-BF16"      OFF)
+option(GGML_AMX_TILE    "ggml: enable AMX-TILE"         OFF)
+option(GGML_AMX_INT8    "ggml: enable AMX-INT8"         OFF)
+option(GGML_AMX_BF16    "ggml: enable AMX-BF16"         OFF)
 option(GGML_FMA         "ggml: enable FMA"              ${INS_ENB})
 if (NOT MSVC)
     option(GGML_F16C    "ggml: enable F16C"             ${INS_ENB}) # in MSVC F16C is implied with AVX2/AVX512
@@ -150,6 +153,7 @@ option(GGML_VULKAN_VALIDATE                 "ggml: enable Vulkan validation"
 option(GGML_VULKAN_RUN_TESTS                "ggml: run Vulkan tests"                          OFF)
 option(GGML_KOMPUTE                         "ggml: use Kompute"                               OFF)
 option(GGML_METAL                           "ggml: use Metal"                                 ${GGML_METAL_DEFAULT})
+option(GGML_METAL_USE_BF16                  "ggml: use bfloat if available"                   OFF)
 option(GGML_METAL_NDEBUG                    "ggml: disable Metal debugging"                   OFF)
 option(GGML_METAL_SHADER_DEBUG              "ggml: compile Metal with -fno-fast-math"         OFF)
 option(GGML_METAL_EMBED_LIBRARY             "ggml: embed Metal library"                       ${GGML_METAL})
@@ -158,6 +162,7 @@ set   (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING
 set   (GGML_METAL_STD "" CACHE STRING       "ggml: metal standard version (-std flag)")
 option(GGML_OPENMP                          "ggml: use OpenMP"                                ON)
 option(GGML_RPC                             "ggml: use RPC"                                   OFF)
+option(GGML_AMX                             "ggml: use AMX"                                   OFF)
 option(GGML_SYCL                            "ggml: use SYCL"                                  OFF)
 option(GGML_SYCL_F16                        "ggml: use 16 bit floats for sycl calculations"   OFF)
 set   (GGML_SYCL_TARGET "INTEL" CACHE STRING
@@ -214,12 +219,12 @@ include(CMakePackageConfigHelpers)
 # all public headers
 set(GGML_PUBLIC_HEADERS
     include/ggml.h
+    include/ggml-cpu.h
     include/ggml-alloc.h
     include/ggml-backend.h
     include/ggml-blas.h
     include/ggml-cann.h
     include/ggml-cuda.h
-    include/ggml.h
     include/ggml-kompute.h
     include/ggml-metal.h
     include/ggml-rpc.h

ggml/include/ggml-amx.h

@@ -0,0 +1,25 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+// buffer_type API
+GGML_API ggml_backend_buffer_type_t ggml_backend_amx_buffer_type(void);
+
+GGML_API bool ggml_backend_is_amx(ggml_backend_t backend);
+
+// backend API
+GGML_API ggml_backend_t ggml_backend_amx_init(void);
+
+GGML_API void ggml_backend_amx_set_n_threads(ggml_backend_t backend_amx, int n_threads);
+
+GGML_API ggml_backend_reg_t ggml_backend_amx_reg(void);
+
+#ifdef  __cplusplus
+}
+#endif

ggml/include/ggml-backend.h

@@ -114,11 +114,12 @@ extern "C" {
     //
 
     enum ggml_backend_dev_type {
+        // CPU device using system memory
         GGML_BACKEND_DEVICE_TYPE_CPU,
+        // GPU device using dedicated memory
         GGML_BACKEND_DEVICE_TYPE_GPU,
-        // devices with full capabilities (excludes backends such as BLAS that only support matrix multiplication)
-        GGML_BACKEND_DEVICE_TYPE_CPU_FULL,
-        GGML_BACKEND_DEVICE_TYPE_GPU_FULL
+        // accelerator devices intended to be used together with the CPU backend (e.g. BLAS or AMX)
+        GGML_BACKEND_DEVICE_TYPE_ACCEL
     };
 
     // functionality supported by the device
@@ -167,10 +168,14 @@ extern "C" {
     GGML_API ggml_backend_dev_t ggml_backend_reg_dev_get(ggml_backend_reg_t reg, size_t index);
     GGML_API void *             ggml_backend_reg_get_proc_address(ggml_backend_reg_t reg, const char * name);
 
+    // Common functions that may be obtained using ggml_backend_reg_get_proc_address
 
-    // Functions that may be obtained using ggml_backend_reg_get_proc_address
-    typedef ggml_backend_buffer_type_t (*ggml_backend_split_buffer_type_t)(const float *);
-    typedef void (*ggml_backend_set_n_threads_t)(ggml_backend_t, int);
+    // Split buffer type for tensor parallelism
+    typedef ggml_backend_buffer_type_t   (*ggml_backend_split_buffer_type_t)(int main_device, const float * tensor_split);
+    // Set the number of threads for the backend
+    typedef void                         (*ggml_backend_set_n_threads_t)(ggml_backend_t backend, int n_threads);
+    // Get additional buffer types provided by the device (returns a NULL-terminated array)
+    typedef ggml_backend_buffer_type_t * (*ggml_backend_dev_get_extra_bufts_t)(ggml_backend_dev_t device);
 
     //
     // Backend registry
@@ -192,7 +197,7 @@ extern "C" {
     GGML_API ggml_backend_t ggml_backend_init_by_name(const char * name, const char * params);
     // = ggml_backend_dev_init(ggml_backend_dev_by_type(type), params)
     GGML_API ggml_backend_t ggml_backend_init_by_type(enum ggml_backend_dev_type type, const char * params);
-    // = ggml_backend_dev_init(ggml_backend_dev_by_type(GPU_FULL) OR ggml_backend_dev_by_type(CPU_FULL), NULL)
+    // = ggml_backend_dev_init(ggml_backend_dev_by_type(GPU) OR ggml_backend_dev_by_type(CPU), NULL)
     GGML_API ggml_backend_t ggml_backend_init_best(void);
 
     //
@@ -300,27 +305,10 @@ extern "C" {
     GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
     GGML_API void ggml_backend_view_init(struct ggml_tensor * tensor);
 
-    //
-    // CPU backend
-    //
-
-    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
-
-    GGML_API bool ggml_backend_is_cpu                (ggml_backend_t backend);
-    GGML_API void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
-    GGML_API void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
-    GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
-
-    // Create a backend buffer from an existing pointer
+    // CPU buffer types are always available
     GGML_API ggml_backend_buffer_t      ggml_backend_cpu_buffer_from_ptr(void * ptr, size_t size);
     GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
 
-    GGML_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
-
-#ifdef GGML_USE_CPU_HBM
-    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
-#endif
-
 #ifdef  __cplusplus
 }
 #endif

ggml/include/ggml-cann.h

@@ -34,6 +34,8 @@ extern "C" {
  */
 #define GGML_CANN_MAX_DEVICES 16
 
+GGML_API ggml_backend_reg_t ggml_backend_cann_reg(void);
+
 /**
  * @brief Initializes the CANN backend for a specified device.
  *

ggml/include/ggml-cpp.h

@@ -0,0 +1,38 @@
+#pragma once
+
+#ifndef __cplusplus
+#error "This header is for C++ only"
+#endif
+
+#include "ggml.h"
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+#include <memory>
+
+// Smart pointers for ggml types
+
+// ggml
+
+struct ggml_context_deleter { void operator()(ggml_context * ctx) { ggml_free(ctx); } };
+struct gguf_context_deleter { void operator()(gguf_context * ctx) { gguf_free(ctx); } };
+
+typedef std::unique_ptr<ggml_context, ggml_context_deleter> ggml_context_ptr;
+typedef std::unique_ptr<gguf_context, gguf_context_deleter> gguf_context_ptr;
+
+// ggml-alloc
+
+struct ggml_gallocr_deleter { void operator()(ggml_gallocr_t galloc) { ggml_gallocr_free(galloc); } };
+
+typedef std::unique_ptr<ggml_gallocr_t, ggml_gallocr_deleter> ggml_gallocr_ptr;
+
+// ggml-backend
+
+struct ggml_backend_deleter        { void operator()(ggml_backend_t backend)       { ggml_backend_free(backend); } };
+struct ggml_backend_buffer_deleter { void operator()(ggml_backend_buffer_t buffer) { ggml_backend_buffer_free(buffer); } };
+struct ggml_backend_event_deleter  { void operator()(ggml_backend_event_t event)   { ggml_backend_event_free(event); } };
+struct ggml_backend_sched_deleter  { void operator()(ggml_backend_sched_t sched)   { ggml_backend_sched_free(sched); } };
+
+typedef std::unique_ptr<ggml_backend,        ggml_backend_deleter>        ggml_backend_ptr;
+typedef std::unique_ptr<ggml_backend_buffer, ggml_backend_buffer_deleter> ggml_backend_buffer_ptr;
+typedef std::unique_ptr<ggml_backend_event,  ggml_backend_event_deleter>  ggml_backend_event_ptr;
+typedef std::unique_ptr<ggml_backend_sched,  ggml_backend_sched_deleter>  ggml_backend_sched_ptr;

ggml/include/ggml-cpu.h

@@ -0,0 +1,150 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+    // Scheduling priorities
+    enum ggml_sched_priority {
+        GGML_SCHED_PRIO_NORMAL,
+        GGML_SCHED_PRIO_MEDIUM,
+        GGML_SCHED_PRIO_HIGH,
+        GGML_SCHED_PRIO_REALTIME
+    };
+
+    // Threadpool params
+    // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
+    struct ggml_threadpool_params {
+        bool                cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
+        int                 n_threads;                   // number of threads
+        enum ggml_sched_priority prio;                   // thread priority
+        uint32_t            poll;                        // polling level (0 - no polling, 100 - aggressive polling)
+        bool                strict_cpu;                  // strict cpu placement
+        bool                paused;                      // start in paused state
+    };
+
+    struct ggml_threadpool;     // forward declaration, see ggml.c
+
+    typedef struct ggml_threadpool * ggml_threadpool_t;
+
+    // the compute plan that needs to be prepared for ggml_graph_compute()
+    // since https://github.com/ggerganov/ggml/issues/287
+    struct ggml_cplan {
+        size_t    work_size; // size of work buffer, calculated by `ggml_graph_plan()`
+        uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
+
+        int n_threads;
+        struct ggml_threadpool * threadpool;
+
+        // abort ggml_graph_compute when true
+        ggml_abort_callback abort_callback;
+        void *              abort_callback_data;
+    };
+
+    // numa strategies
+    enum ggml_numa_strategy {
+        GGML_NUMA_STRATEGY_DISABLED   = 0,
+        GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
+        GGML_NUMA_STRATEGY_ISOLATE    = 2,
+        GGML_NUMA_STRATEGY_NUMACTL    = 3,
+        GGML_NUMA_STRATEGY_MIRROR     = 4,
+        GGML_NUMA_STRATEGY_COUNT
+    };
+
+    GGML_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
+    GGML_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
+
+    GGML_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
+    GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
+
+    GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
+    GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
+
+    GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
+    GGML_API void    ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
+
+    GGML_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_API void    ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
+
+    GGML_API float   ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
+    GGML_API void    ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
+
+    GGML_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+    GGML_API void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
+
+    GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
+    GGML_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
+    GGML_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
+    GGML_API struct ggml_threadpool *      ggml_threadpool_new          (struct ggml_threadpool_params  * params);
+    GGML_API void                          ggml_threadpool_free         (struct ggml_threadpool * threadpool);
+    GGML_API int                           ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
+    GGML_API void                          ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
+    GGML_API void                          ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
+
+    // ggml_graph_plan() has to be called before ggml_graph_compute()
+    // when plan.work_size > 0, caller must allocate memory for plan.work_data
+    GGML_API struct ggml_cplan ggml_graph_plan(
+                  const struct ggml_cgraph * cgraph,
+                                       int   n_threads, /* = GGML_DEFAULT_N_THREADS */
+                    struct ggml_threadpool * threadpool /* = NULL */ );
+    GGML_API enum ggml_status  ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+
+    // same as ggml_graph_compute() but the work data is allocated as a part of the context
+    // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
+    GGML_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
+
+    // TODO: move to backend interface
+    GGML_API int ggml_cpu_has_neon       (void);
+    GGML_API int ggml_cpu_has_sve        (void);
+    GGML_API int ggml_cpu_has_matmul_int8(void);
+    // get the sve vector length in bytes
+    GGML_API int ggml_cpu_get_sve_cnt(void);
+
+    // Internal types and functions exposed for tests and benchmarks
+
+    typedef void (*ggml_from_float_to_mat_t)
+                                     (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, int64_t k, int64_t bs);
+    typedef void (*ggml_vec_dot_t)  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
+                                       const void * GGML_RESTRICT y, size_t by, int nrc);
+    typedef void (*ggml_gemv_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
+                                       const void * GGML_RESTRICT y, int nr, int nc);
+    typedef void (*ggml_gemm_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
+                                       const void * GGML_RESTRICT y, int nr, int nc);
+
+    struct ggml_type_traits_cpu {
+        ggml_from_float_to_mat_t from_float_to_mat;
+        ggml_vec_dot_t           vec_dot;
+        enum ggml_type           vec_dot_type;
+        int64_t                  nrows; // number of rows to process simultaneously
+        int64_t                  ncols; // number of columns to process simultaneously
+        ggml_gemv_t              gemv;
+        ggml_gemm_t              gemm;
+    };
+
+    GGML_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
+
+    GGML_API void ggml_cpu_init(void);
+
+    //
+    // CPU backend
+    //
+
+    GGML_API ggml_backend_t ggml_backend_cpu_init(void);
+
+    GGML_API bool ggml_backend_is_cpu                (ggml_backend_t backend);
+    GGML_API void ggml_backend_cpu_set_n_threads     (ggml_backend_t backend_cpu, int n_threads);
+    GGML_API void ggml_backend_cpu_set_threadpool    (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
+    GGML_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
+
+    GGML_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
+
+#ifdef GGML_USE_CPU_HBM
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
+#endif
+
+#ifdef __cplusplus
+}
+#endif

ggml/include/ggml-cuda.h

@@ -28,7 +28,7 @@ GGML_API bool ggml_backend_is_cuda(ggml_backend_t backend);
 GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_buffer_type(int device);
 
 // split tensor buffer that splits matrices by rows across multiple devices
-GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(const float * tensor_split);
+GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_split_buffer_type(int main_device, const float * tensor_split);
 
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
 GGML_API ggml_backend_buffer_type_t ggml_backend_cuda_host_buffer_type(void);

ggml/include/ggml-kompute.h

@@ -11,6 +11,8 @@
 extern "C" {
 #endif
 
+#define GGML_KOMPUTE_MAX_DEVICES 16
+
 struct ggml_vk_device {
     int index;
     int type; // same as VkPhysicalDeviceType
@@ -41,6 +43,8 @@ GGML_API bool ggml_backend_is_kompute(ggml_backend_t backend);
 
 GGML_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
 
+GGML_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
+
 #ifdef __cplusplus
 }
 #endif

ggml/include/ggml-sycl.h

@@ -19,6 +19,8 @@ extern "C" {
 // backend API
 GGML_API ggml_backend_t ggml_backend_sycl_init(int device);
 
+GGML_API bool ggml_backend_is_sycl(ggml_backend_t backend);
+
 // devide buffer
 GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_buffer_type(int device);
 
@@ -29,14 +31,19 @@ GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_split_buffer_type(const fl
 GGML_API ggml_backend_buffer_type_t ggml_backend_sycl_host_buffer_type(void);
 
 GGML_API void ggml_backend_sycl_print_sycl_devices(void);
-GGML_API void ggml_sycl_get_gpu_list(int *id_list, int max_len);
-GGML_API void ggml_sycl_get_device_description(int device, char *description, size_t description_size);
+GGML_API void ggml_backend_sycl_get_gpu_list(int *id_list, int max_len);
+GGML_API void ggml_backend_sycl_get_device_description(int device,
+                                                       char *description,
+                                                       size_t description_size);
 GGML_API int  ggml_backend_sycl_get_device_count();
 GGML_API void ggml_backend_sycl_get_device_memory(int device, size_t *free, size_t *total);
 
 // SYCL doesn't support registering host memory, keep here for reference
 // GGML_API bool ggml_backend_sycl_register_host_buffer(void * buffer, size_t size);
 // GGML_API void ggml_backend_sycl_unregister_host_buffer(void * buffer);
+
+GGML_API ggml_backend_reg_t ggml_backend_sycl_reg(void);
+
 #ifdef  __cplusplus
 }
 #endif

ggml/include/ggml-vulkan.h

@@ -24,6 +24,8 @@ GGML_API ggml_backend_buffer_type_t ggml_backend_vk_buffer_type(size_t dev_num);
 // pinned host buffer for use with the CPU backend for faster copies between CPU and GPU
 GGML_API ggml_backend_buffer_type_t ggml_backend_vk_host_buffer_type(void);
 
+GGML_API ggml_backend_reg_t ggml_backend_vk_reg(void);
+
 #ifdef  __cplusplus
 }
 #endif

ggml/include/ggml.h

@@ -217,7 +217,6 @@
 
 #define GGML_MAX_DIMS           4
 #define GGML_MAX_PARAMS         2048
-#define GGML_MAX_CONTEXTS       64
 #define GGML_MAX_SRC            10
 #define GGML_MAX_N_THREADS      512
 #define GGML_MAX_OP_PARAMS      64
@@ -510,7 +509,7 @@ extern "C" {
         GGML_OP_WIN_UNPART,
         GGML_OP_GET_REL_POS,
         GGML_OP_ADD_REL_POS,
-        GGML_OP_RWKV_WKV,
+        GGML_OP_RWKV_WKV6,
 
         GGML_OP_UNARY,
 
@@ -559,10 +558,10 @@ extern "C" {
 
     enum ggml_log_level {
         GGML_LOG_LEVEL_NONE  = 0,
-        GGML_LOG_LEVEL_INFO  = 1,
-        GGML_LOG_LEVEL_WARN  = 2,
-        GGML_LOG_LEVEL_ERROR = 3,
-        GGML_LOG_LEVEL_DEBUG = 4,
+        GGML_LOG_LEVEL_DEBUG = 1,
+        GGML_LOG_LEVEL_INFO  = 2,
+        GGML_LOG_LEVEL_WARN  = 3,
+        GGML_LOG_LEVEL_ERROR = 4,
         GGML_LOG_LEVEL_CONT  = 5, // continue previous log
     };
 
@@ -574,6 +573,13 @@ extern "C" {
         GGML_TENSOR_FLAG_LOSS   =  8, // ...defines loss for numerical optimization (multiple loss tensors add up)
     };
 
+    struct ggml_init_params {
+        // memory pool
+        size_t mem_size;   // bytes
+        void * mem_buffer; // if NULL, memory will be allocated internally
+        bool   no_alloc;   // don't allocate memory for the tensor data
+    };
+
     // n-dimensional tensor
     struct ggml_tensor {
         enum ggml_type type;
@@ -619,66 +625,6 @@ extern "C" {
     // If it returns true, the computation is aborted
     typedef bool (*ggml_abort_callback)(void * data);
 
-    // Scheduling priorities
-    enum ggml_sched_priority {
-        GGML_SCHED_PRIO_NORMAL,
-        GGML_SCHED_PRIO_MEDIUM,
-        GGML_SCHED_PRIO_HIGH,
-        GGML_SCHED_PRIO_REALTIME
-    };
-
-    // Threadpool params
-    // Use ggml_threadpool_params_default() or ggml_threadpool_params_init() to populate the defaults
-    struct ggml_threadpool_params {
-        bool                cpumask[GGML_MAX_N_THREADS]; // mask of cpu cores (all-zeros means use default affinity settings)
-        int                 n_threads;                   // number of threads
-        enum ggml_sched_priority prio;                   // thread priority
-        uint32_t            poll;                        // polling level (0 - no polling, 100 - aggressive polling)
-        bool                strict_cpu;                  // strict cpu placement
-        bool                paused;                      // start in paused state
-    };
-
-    struct ggml_threadpool;     // forward declaration, see ggml.c
-
-    typedef struct ggml_threadpool * ggml_threadpool_t;
-
-    // the compute plan that needs to be prepared for ggml_graph_compute()
-    // since https://github.com/ggerganov/ggml/issues/287
-    struct ggml_cplan {
-        size_t    work_size; // size of work buffer, calculated by `ggml_graph_plan()`
-        uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
-
-        int n_threads;
-        struct ggml_threadpool * threadpool;
-
-        // abort ggml_graph_compute when true
-        ggml_abort_callback abort_callback;
-        void *              abort_callback_data;
-    };
-
-    // scratch buffer
-    struct ggml_scratch {
-        size_t offs;
-        size_t size;
-        void * data;
-    };
-
-    struct ggml_init_params {
-        // memory pool
-        size_t mem_size;   // bytes
-        void * mem_buffer; // if NULL, memory will be allocated internally
-        bool   no_alloc;   // don't allocate memory for the tensor data
-    };
-
-    // numa strategies
-    enum ggml_numa_strategy {
-        GGML_NUMA_STRATEGY_DISABLED   = 0,
-        GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
-        GGML_NUMA_STRATEGY_ISOLATE    = 2,
-        GGML_NUMA_STRATEGY_NUMACTL    = 3,
-        GGML_NUMA_STRATEGY_MIRROR     = 4,
-        GGML_NUMA_STRATEGY_COUNT
-    };
 
     //
     // GUID
@@ -701,9 +647,6 @@ extern "C" {
     // accepts a UTF-8 path, even on Windows
     GGML_API FILE *  ggml_fopen(const char * fname, const char * mode);
 
-    GGML_API void    ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
-    GGML_API bool    ggml_is_numa(void); // true if init detected that system has >1 NUMA node
-
     GGML_API void    ggml_print_object (const struct ggml_object * obj);
     GGML_API void    ggml_print_objects(const struct ggml_context * ctx);
 
@@ -760,12 +703,12 @@ extern "C" {
 
     // main
 
-    GGML_API struct ggml_context * ggml_init(struct ggml_init_params params);
-    GGML_API void                  ggml_free(struct ggml_context * ctx);
+    GGML_API struct ggml_context * ggml_init (struct ggml_init_params params);
+    GGML_API void                  ggml_reset(struct ggml_context * ctx);
+    GGML_API void                  ggml_free (struct ggml_context * ctx);
 
     GGML_API size_t  ggml_used_mem(const struct ggml_context * ctx);
 
-    GGML_API size_t  ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch);
     GGML_API bool    ggml_get_no_alloc(struct ggml_context * ctx);
     GGML_API void    ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc);
 
@@ -805,8 +748,7 @@ extern "C" {
             int64_t ne2,
             int64_t ne3);
 
-    GGML_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
-    GGML_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
+    GGML_API void * ggml_new_buffer(struct ggml_context * ctx, size_t nbytes);
 
     GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
     GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src);
@@ -816,35 +758,25 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_get_next_tensor (const struct ggml_context * ctx, struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
 
-    GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
-    GGML_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
-    GGML_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
-
     // Converts a flat index into coordinates
-    GGML_API void    ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
+    GGML_API void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
 
-    GGML_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
-    GGML_API void    ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
-
-    GGML_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
-    GGML_API void    ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
-
-    GGML_API float   ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
-    GGML_API void    ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
-
-    GGML_API float   ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
-    GGML_API void    ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
+    GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
 
     GGML_API void *  ggml_get_data    (const struct ggml_tensor * tensor);
     GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
 
-    GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
-
     GGML_API const char *         ggml_get_name   (const struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_set_name   (      struct ggml_tensor * tensor, const char * name);
     GGML_ATTRIBUTE_FORMAT(2, 3)
     GGML_API struct ggml_tensor * ggml_format_name(      struct ggml_tensor * tensor, const char * fmt, ...);
 
+    // Tensor flags
+    GGML_API void ggml_set_input(struct ggml_tensor * tensor);
+    GGML_API void ggml_set_output(struct ggml_tensor * tensor);
+    GGML_API void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor);
+    GGML_API void ggml_set_loss(struct ggml_tensor * tensor);
+
     //
     // operations on tensors with backpropagation
     //
@@ -1814,6 +1746,9 @@ extern "C" {
             struct ggml_tensor * a,
             enum ggml_prec       prec);
 
+    GGML_API enum ggml_prec ggml_flash_attn_ext_get_prec(
+            const struct ggml_tensor * a);
+
     // TODO: needs to be adapted to ggml_flash_attn_ext
     GGML_API struct ggml_tensor * ggml_flash_attn_back(
            struct ggml_context * ctx,
@@ -1887,7 +1822,7 @@ extern "C" {
             struct ggml_tensor  * pw,
             struct ggml_tensor  * ph);
 
-    GGML_API struct ggml_tensor * ggml_rwkv_wkv(
+    GGML_API struct ggml_tensor * ggml_rwkv_wkv6(
             struct ggml_context * ctx,
             struct ggml_tensor  * k,
             struct ggml_tensor  * v,
@@ -2060,9 +1995,6 @@ extern "C" {
     // automatic differentiation
     //
 
-    GGML_API void ggml_set_param(struct ggml_context * ctx, struct ggml_tensor * tensor);
-    GGML_API void ggml_set_loss(struct ggml_tensor * tensor);
-
     GGML_API void ggml_build_forward_expand (struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
     GGML_API void ggml_build_backward_expand(struct ggml_context * ctx, struct ggml_cgraph * gf, struct ggml_cgraph * gb, bool accumulate);
 
@@ -2094,27 +2026,6 @@ extern "C" {
     GGML_API size_t ggml_graph_overhead(void);
     GGML_API size_t ggml_graph_overhead_custom(size_t size, bool grads);
 
-    GGML_API struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads);
-    GGML_API void                          ggml_threadpool_params_init   (struct ggml_threadpool_params * p, int n_threads);
-    GGML_API bool                          ggml_threadpool_params_match  (const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1);
-    GGML_API struct ggml_threadpool *      ggml_threadpool_new          (struct ggml_threadpool_params  * params);
-    GGML_API void                          ggml_threadpool_free         (struct ggml_threadpool * threadpool);
-    GGML_API int                           ggml_threadpool_get_n_threads(struct ggml_threadpool * threadpool);
-    GGML_API void                          ggml_threadpool_pause        (struct ggml_threadpool * threadpool);
-    GGML_API void                          ggml_threadpool_resume       (struct ggml_threadpool * threadpool);
-
-    // ggml_graph_plan() has to be called before ggml_graph_compute()
-    // when plan.work_size > 0, caller must allocate memory for plan.work_data
-    GGML_API struct ggml_cplan ggml_graph_plan(
-                  const struct ggml_cgraph * cgraph,
-                                       int   n_threads, /* = GGML_DEFAULT_N_THREADS */
-                    struct ggml_threadpool * threadpool /* = NULL */ );
-    GGML_API enum ggml_status  ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
-
-    // same as ggml_graph_compute() but the work data is allocated as a part of the context
-    // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
-    GGML_API enum ggml_status  ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
-
     GGML_API struct ggml_tensor * ggml_graph_get_tensor(struct ggml_cgraph * cgraph, const char * name);
 
     GGML_API void                 ggml_graph_export(const struct ggml_cgraph * cgraph, const char * fname);
@@ -2285,6 +2196,8 @@ extern "C" {
         } lbfgs;
     };
 
+    GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
+
     GGML_API struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);
 
     // optimize the function defined by the tensor f
@@ -2316,12 +2229,6 @@ extern "C" {
             ggml_opt_callback callback,
             void * callback_data);
 
-    //
-    // tensor flags
-    //
-    GGML_API void ggml_set_input(struct ggml_tensor * tensor);
-    GGML_API void ggml_set_output(struct ggml_tensor * tensor);
-
     //
     // quantization
     //
@@ -2488,9 +2395,8 @@ extern "C" {
     GGML_API int ggml_cpu_has_avx512_vbmi(void);
     GGML_API int ggml_cpu_has_avx512_vnni(void);
     GGML_API int ggml_cpu_has_avx512_bf16(void);
+    GGML_API int ggml_cpu_has_amx_int8   (void);
     GGML_API int ggml_cpu_has_fma        (void);
-    GGML_API int ggml_cpu_has_neon       (void);
-    GGML_API int ggml_cpu_has_sve        (void);
     GGML_API int ggml_cpu_has_arm_fma    (void);
     GGML_API int ggml_cpu_has_metal      (void);
     GGML_API int ggml_cpu_has_f16c       (void);
@@ -2507,17 +2413,9 @@ extern "C" {
     GGML_API int ggml_cpu_has_sycl       (void);
     GGML_API int ggml_cpu_has_rpc        (void);
     GGML_API int ggml_cpu_has_vsx        (void);
-    GGML_API int ggml_cpu_has_matmul_int8(void);
     GGML_API int ggml_cpu_has_cann       (void);
     GGML_API int ggml_cpu_has_llamafile  (void);
 
-    // get the sve vector length in bytes
-    GGML_API int ggml_cpu_get_sve_cnt(void);
-
-    //
-    // Internal types and functions exposed for tests and benchmarks
-    //
-
 #ifdef  __cplusplus
 // restrict not standard in C++
 #define GGML_RESTRICT
@@ -2526,14 +2424,6 @@ extern "C" {
 #endif
     typedef void (*ggml_to_float_t)  (const void  * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
     typedef void (*ggml_from_float_t)(const float * GGML_RESTRICT x, void  * GGML_RESTRICT y, int64_t k);
-    typedef void (*ggml_from_float_to_mat_t)
-                                     (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t nr, int64_t k, int64_t bs);
-    typedef void (*ggml_vec_dot_t)  (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
-                                       const void * GGML_RESTRICT y, size_t by, int nrc);
-    typedef void (*ggml_gemv_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
-                                       const void * GGML_RESTRICT y, int nr, int nc);
-    typedef void (*ggml_gemm_t)     (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x,
-                                       const void * GGML_RESTRICT y, int nr, int nc);
 
     struct ggml_type_traits {
         const char             * type_name;
@@ -2544,13 +2434,6 @@ extern "C" {
         ggml_to_float_t          to_float;
         ggml_from_float_t        from_float;
         ggml_from_float_t        from_float_ref;
-        ggml_from_float_to_mat_t from_float_to_mat;
-        ggml_vec_dot_t           vec_dot;
-        enum ggml_type           vec_dot_type;
-        int64_t                  nrows; // number of rows to process simultaneously
-        int64_t                  ncols; // number of columns to process simultaneously
-        ggml_gemv_t              gemv;
-        ggml_gemm_t              gemm;
     };
 
     GGML_API const struct ggml_type_traits * ggml_get_type_traits(enum ggml_type type);