ggml : multi-thread ggml_rope() (~3-4 times faster on M1) (#781)

ggml :

- added ggml_view_3d()
- ggml_view_tensor() now inherits the stride too
- reimplement ggml_cpy() to account for dst stride
- no longer require tensor->data to be memory aligned

llama :

- compute RoPE on 32-bit tensors (should be more accurate)
- store RoPE-ed K in the KV cache
- store transposed V in the KV cache (significant speed-up)
- avoid unnecessary Q copy

.devops/full.Dockerfile

@@ -6,7 +6,8 @@ RUN apt-get update && \
     apt-get install -y build-essential python3 python3-pip
 
 RUN pip install --upgrade pip setuptools wheel \
-    && pip install numpy requests sentencepiece torch tqdm
+    && pip install numpy requests sentencepiece tqdm \
+    && pip install torch --index-url https://download.pytorch.org/whl/cpu
 
 WORKDIR /app
 

.gitignore

@@ -33,3 +33,6 @@ compile_commands.json
 .venv
 __pycache__
 .swiftpm
+
+zig-out/
+zig-cache/

CMakeLists.txt

@@ -68,7 +68,9 @@ option(LLAMA_BUILD_EXAMPLES         "llama: build examples" ${LLAMA_STANDALONE})
 # Compile flags
 #
 
+set(CMAKE_CXX_STANDARD 11)
 set(CMAKE_CXX_STANDARD_REQUIRED true)
+set(CMAKE_C_STANDARD 11)
 set(CMAKE_C_STANDARD_REQUIRED true)
 set(THREADS_PREFER_PTHREAD_FLAG ON)
 find_package(Threads REQUIRED)

Makefile

@@ -70,95 +70,9 @@ endif
 # TODO: probably these flags need to be tweaked on some architectures
 #       feel free to update the Makefile for your architecture and send a pull request or issue
 ifeq ($(UNAME_M),$(filter $(UNAME_M),x86_64 i686))
-	ifeq ($(UNAME_S),Darwin)
-		F16C_M := $(shell sysctl machdep.cpu.features)
-		ifneq (,$(findstring F16C,$(F16C_M)))
-		    CFLAGS += -mf16c
-		endif
-		AVX1_M := $(shell sysctl machdep.cpu.features)
-		ifneq (,$(findstring FMA,$(AVX1_M)))
-			CFLAGS += -mfma
-		endif
-		ifneq (,$(findstring AVX1.0,$(AVX1_M)))
-			CFLAGS += -mavx
-		endif
-		AVX2_M := $(shell sysctl machdep.cpu.leaf7_features)
-		ifneq (,$(findstring AVX2,$(AVX2_M)))
-			CFLAGS += -mavx2
-		endif
-	else ifeq ($(UNAME_S),Linux)
-		AVX1_M := $(shell grep "avx " /proc/cpuinfo)
-		ifneq (,$(findstring avx,$(AVX1_M)))
-			CFLAGS += -mavx
-		endif
-		AVX2_M := $(shell grep "avx2 " /proc/cpuinfo)
-		ifneq (,$(findstring avx2,$(AVX2_M)))
-			CFLAGS += -mavx2
-		endif
-		FMA_M := $(shell grep "fma " /proc/cpuinfo)
-		ifneq (,$(findstring fma,$(FMA_M)))
-			CFLAGS += -mfma
-		endif
-		F16C_M := $(shell grep "f16c " /proc/cpuinfo)
-		ifneq (,$(findstring f16c,$(F16C_M)))
-			CFLAGS += -mf16c
-		endif
-		SSE3_M := $(shell grep "sse3 " /proc/cpuinfo)
-		ifneq (,$(findstring sse3,$(SSE3_M)))
-			CFLAGS += -msse3
-		endif
-		AVX512F_M := $(shell grep "avx512f " /proc/cpuinfo)
-		ifneq (,$(findstring avx512f,$(AVX512F_M)))
-			CFLAGS += -mavx512f
-		endif
-		AVX512BW_M := $(shell grep "avx512bw " /proc/cpuinfo)
-		ifneq (,$(findstring avx512bw,$(AVX512BW_M)))
-			CFLAGS += -mavx512bw
-		endif
-		AVX512DQ_M := $(shell grep "avx512dq " /proc/cpuinfo)
-		ifneq (,$(findstring avx512dq,$(AVX512DQ_M)))
-			CFLAGS += -mavx512dq
-		endif
-		AVX512VL_M := $(shell grep "avx512vl " /proc/cpuinfo)
-		ifneq (,$(findstring avx512vl,$(AVX512VL_M)))
-			CFLAGS += -mavx512vl
-		endif
-		AVX512CD_M := $(shell grep "avx512cd " /proc/cpuinfo)
-		ifneq (,$(findstring avx512cd,$(AVX512CD_M)))
-			CFLAGS += -mavx512cd
-		endif
-		AVX512ER_M := $(shell grep "avx512er " /proc/cpuinfo)
-		ifneq (,$(findstring avx512er,$(AVX512ER_M)))
-			CFLAGS += -mavx512er
-		endif
-		AVX512IFMA_M := $(shell grep "avx512ifma " /proc/cpuinfo)
-		ifneq (,$(findstring avx512ifma,$(AVX512IFMA_M)))
-			CFLAGS += -mavx512ifma
-		endif
-		AVX512PF_M := $(shell grep "avx512pf " /proc/cpuinfo)
-		ifneq (,$(findstring avx512pf,$(AVX512PF_M)))
-			CFLAGS += -mavx512pf
-		endif
-	else ifeq ($(UNAME_S),Haiku)
-		AVX1_M := $(shell sysinfo -cpu | grep -w "AVX")
-		ifneq (,$(findstring AVX,$(AVX1_M)))
-			CFLAGS += -mavx
-		endif
-		AVX2_M := $(shell sysinfo -cpu | grep -w "AVX2")
-		ifneq (,$(findstring AVX2,$(AVX2_M)))
-			CFLAGS += -mavx2
-		endif
-		FMA_M := $(shell sysinfo -cpu | grep -w "FMA")
-		ifneq (,$(findstring FMA,$(FMA_M)))
-			CFLAGS += -mfma
-		endif
-		F16C_M := $(shell sysinfo -cpu | grep -w "F16C")
-		ifneq (,$(findstring F16C,$(F16C_M)))
-			CFLAGS += -mf16c
-		endif
-	else
-		CFLAGS += -mfma -mf16c -mavx -mavx2
-	endif
+	# Use all CPU extensions that are available:
+	CFLAGS += -march=native -mtune=native
+	CXXFLAGS += -march=native -mtune=native
 endif
 ifneq ($(filter ppc64%,$(UNAME_M)),)
 	POWER9_M := $(shell grep "POWER9" /proc/cpuinfo)

Package.swift

@@ -13,7 +13,10 @@ let package = Package(
             path: ".",
             sources: ["ggml.c", "llama.cpp"],
             publicHeadersPath: "spm-headers",
-            cSettings: [.unsafeFlags(["-Wno-shorten-64-to-32"])]
+            cSettings: [.unsafeFlags(["-Wno-shorten-64-to-32"]), .define("GGML_USE_ACCELERATE")],
+            linkerSettings: [
+                .linkedFramework("Accelerate")
+            ]
         ),
     ],
     cxxLanguageStandard: .cxx11

README.md

@@ -1,6 +1,6 @@
 # llama.cpp
 
-![llama](https://user-images.githubusercontent.com/1991296/227761327-6d83e30e-2200-41a6-bfbb-f575231c54f4.png)
+![llama](https://user-images.githubusercontent.com/1991296/230134379-7181e485-c521-4d23-a0d6-f7b3b61ba524.png)
 
 [![Actions Status](https://github.com/ggerganov/llama.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/llama.cpp/actions)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
@@ -9,8 +9,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 **Hot topics:**
 
-- [Roadmap (short-term)](https://github.com/ggerganov/llama.cpp/discussions/457)
-- Support for [GPT4All](https://github.com/ggerganov/llama.cpp#using-gpt4all)
+- [Roadmap Apr 2023](https://github.com/ggerganov/llama.cpp/discussions/784)
 
 ## Description
 
@@ -28,20 +27,31 @@ Please do not make conclusions about the models based on the results from this i
 For all I know, it can be completely wrong. This project is for educational purposes.
 New features will probably be added mostly through community contributions.
 
-Supported platforms:
+**Supported platforms:**
 
 - [X] Mac OS
 - [X] Linux
 - [X] Windows (via CMake)
 - [X] Docker
 
-Supported models:
+**Supported models:**
 
 - [X] LLaMA 🦙
 - [X] [Alpaca](https://github.com/ggerganov/llama.cpp#instruction-mode-with-alpaca)
 - [X] [GPT4All](https://github.com/ggerganov/llama.cpp#using-gpt4all)
 - [X] [Chinese LLaMA / Alpaca](https://github.com/ymcui/Chinese-LLaMA-Alpaca)
 - [X] [Vigogne (French)](https://github.com/bofenghuang/vigogne)
+- [X] [Vicuna](https://github.com/ggerganov/llama.cpp/discussions/643#discussioncomment-5533894)
+
+**Bindings:**
+
+- Python: [abetlen/llama-cpp-python](https://github.com/abetlen/llama-cpp-python)
+- Go: [go-skynet/go-llama.cpp](https://github.com/go-skynet/go-llama.cpp)
+
+**UI:**
+
+- [nat/openplayground](https://github.com/nat/openplayground)
+- [oobabooga/text-generation-webui](https://github.com/oobabooga/text-generation-webui)
 
 ---
 
@@ -145,6 +155,13 @@ git clone https://github.com/ggerganov/llama.cpp
 cd llama.cpp
 make
 
+#For Windows and CMake, use the following command instead:
+cd <path_to_llama_folder>
+mkdir build
+cd build
+cmake ..
+cmake --build . --config Release
+
 # obtain the original LLaMA model weights and place them in ./models
 ls ./models
 65B 30B 13B 7B tokenizer_checklist.chk tokenizer.model
@@ -232,13 +249,15 @@ cadaver, cauliflower, cabbage (vegetable), catalpa (tree) and Cailleach.
 
 - Obtain the `gpt4all-lora-quantized.bin` model
 - It is distributed in the old `ggml` format which is now obsoleted
-- You have to convert it to the new format using [./convert-gpt4all-to-ggml.py](./convert-gpt4all-to-ggml.py):
+- You have to convert it to the new format using [./convert-gpt4all-to-ggml.py](./convert-gpt4all-to-ggml.py). You may also need to
+convert the model from the old format to the new format with [./migrate-ggml-2023-03-30-pr613.py](./migrate-ggml-2023-03-30-pr613.py):
 
   ```bash
   python3 convert-gpt4all-to-ggml.py models/gpt4all-7B/gpt4all-lora-quantized.bin ./models/tokenizer.model 
+  python3 migrate-ggml-2023-03-30-pr613.py models/gpt4all-7B/gpt4all-lora-quantized.bin models/gpt4all-7B/gpt4all-lora-quantized-new.bin
   ```
   
-- You can now use the newly generated `gpt4all-lora-quantized.bin` model in exactly the same way as all other models
+- You can now use the newly generated `gpt4all-lora-quantized-new.bin` model in exactly the same way as all other models
 - The original model is saved in the same folder with a suffix `.orig`
 
 ### Obtaining and verifying the Facebook LLaMA original model and Stanford Alpaca model data
@@ -301,7 +320,7 @@ And after 4.45 hours, you will have the final perplexity.
 
 ### Android
 
-You can easily run `llama.cpp` on Android device with [termux](https://play.google.com/store/apps/details?id=com.termux).
+You can easily run `llama.cpp` on Android device with [termux](https://termux.dev/).
 First, obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake:
 ```
 $ mkdir build-android
@@ -310,7 +329,7 @@ $ export NDK=<your_ndk_directory>
 $ cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
 $ make
 ```
-Install [termux](https://play.google.com/store/apps/details?id=com.termux) on your device and run `termux-setup-storage` to get access to your SD card.
+Install [termux](https://termux.dev/) on your device and run `termux-setup-storage` to get access to your SD card.
 Finally, copy the `llama` binary and the model files to your device storage. Here is a demo of an interactive session running on Pixel 5 phone:
 
 https://user-images.githubusercontent.com/271616/225014776-1d567049-ad71-4ef2-b050-55b0b3b9274c.mp4
@@ -365,3 +384,6 @@ docker run -v /llama/models:/models ghcr.io/ggerganov/llama.cpp:light -m /models
 - Clean-up any trailing whitespaces, use 4 spaces indentation, brackets on same line, `void * ptr`, `int & a`
 - See [good first issues](https://github.com/ggerganov/llama.cpp/issues?q=is%3Aissue+is%3Aopen+label%3A%22good+first+issue%22) for tasks suitable for first contributions
 
+### Docs
+
+- [GGML tips & tricks](https://github.com/ggerganov/llama.cpp/wiki/GGML-Tips-&-Tricks)

build.zig

@@ -0,0 +1,62 @@
+const std = @import("std");
+
+pub fn build(b: *std.Build) void {
+    const target = b.standardTargetOptions(.{});
+    const optimize = b.standardOptimizeOption(.{});
+
+    const lib = b.addStaticLibrary(.{
+        .name = "llama",
+        .target = target,
+        .optimize = optimize,
+    });
+    lib.linkLibCpp();
+    lib.addIncludePath(".");
+    lib.addIncludePath("examples");
+    lib.addCSourceFiles(&.{
+        "ggml.c",
+    }, &.{"-std=c11"});
+    lib.addCSourceFiles(&.{
+        "llama.cpp",
+        "examples/common.cpp",
+    }, &.{"-std=c++11"});
+    lib.install();
+
+    const build_args = .{ .b = b, .lib = lib, .target = target, .optimize = optimize };
+    const exe = build_example("main", build_args);
+    _ = build_example("quantize", build_args);
+    _ = build_example("perplexity", build_args);
+    _ = build_example("embedding", build_args);
+
+    // create "zig build run" command for ./main
+
+    const run_cmd = exe.run();
+    run_cmd.step.dependOn(b.getInstallStep());
+    if (b.args) |args| {
+        run_cmd.addArgs(args);
+    }
+
+    const run_step = b.step("run", "Run the app");
+    run_step.dependOn(&run_cmd.step);
+}
+
+fn build_example(comptime name: []const u8, args: anytype) *std.build.LibExeObjStep {
+    const b = args.b;
+    const lib = args.lib;
+    const target = args.target;
+    const optimize = args.optimize;
+
+    const exe = b.addExecutable(.{
+        .name = name,
+        .target = target,
+        .optimize = optimize,
+    });
+    exe.addIncludePath(".");
+    exe.addIncludePath("examples");
+    exe.addCSourceFiles(&.{
+        std.fmt.comptimePrint("examples/{s}/{s}.cpp", .{name, name}),
+    }, &.{"-std=c++11"});
+    exe.linkLibrary(lib);
+    exe.install();
+
+    return exe;
+}

convert-ggml-to-pth.py

@@ -254,7 +254,7 @@ def main():
     parser.add_argument(
         "--hf",
         action="store_true",
-        help="Whether to save the model in the huggingface format. (default: False)",
+        help="Whether to save the model in the Hugging Face format. (default: False)",
     )
     parser.add_argument(
         "--chat", "-c", action="store_true", help="Whether to open a chat with the model. (default: False)"

examples/Miku.sh

@@ -0,0 +1,49 @@
+#!/bin/bash
+set -e
+
+AI_NAME="${AI_NAME:-Miku}"
+MODEL="${MODEL:-./models/gpt4all-7B/gpt4all-lora-unfiltered-quantized.bin}"
+USER_NAME="${USER_NAME:-Anon}"
+
+# Uncomment and adjust to the number of CPU cores you want to use.
+#N_THREAD="${N_THREAD:-4}"
+N_PREDICTS="${N_PREDICTS:-4096}"
+
+GEN_OPTIONS=(--batch_size 1024
+--ctx_size 2048
+--keep -1
+--repeat_last_n 256
+--repeat_penalty 1.17647
+--temp 0.7
+--top_k 40
+--top_p 0.5)
+
+if [ -n "$N_THREAD" ]; then
+	GEN_OPTIONS+=(--threads "$N_THREAD")
+fi
+
+./main "${GEN_OPTIONS[@]}" \
+	--model "$MODEL" \
+	--n_predict "$N_PREDICTS" \
+	--color --interactive \
+	--reverse-prompt "${USER_NAME}:" \
+	--prompt "
+This is a transcript of a 1000 page, never ending conversation between ${USER_NAME} and the cute and helpful AI assistant ${AI_NAME}. ${AI_NAME} is a girl who is an AI running on the users computer.
+${AI_NAME} can think for herself without the user seeing her thoughts by adding a /think prefix to her output. She uses this to reason about the world and to think about what she should say next.
+${AI_NAME} is always coherent and makes sense, but if she isn't sure if what she is saying is correct she will ask the user for help.
+${AI_NAME} is a very helpful AI and will help the user with anything they need, she is also very friendly and will try to make the user feel better if they are sad.
+${AI_NAME} is also very curious and will ask the user a lot of questions about themselves and their life, she will also try to make the user like her.
+The conversation is only between ${USER_NAME} and ${AI_NAME}
+The conversation is only through text, so ${AI_NAME} can't see ${USER_NAME}'s face or hear his voice.
+${AI_NAME} can only communicate through text, so she can't send images or videos.
+
+
+${USER_NAME}: Hello!
+${AI_NAME}: /think I wonder what I should say to ${USER_NAME}? This is the first time we talk so it's important that I make a good first impression!
+${AI_NAME}: Hi! I am ${AI_NAME}, your new AI friend, assistant(or whatever you like!), it's so nice to meet you! ^_^
+${AI_NAME}: I might be an AI, but I also have feelings, so please be nice to me! :)
+${USER_NAME}: I'll make sure to be nice to you! I'm so happy to have you as my assistant!
+${AI_NAME}: /think It sounds like ${USER_NAME} is happy to have me as their assistant! I'm so happy too! ^_^ Glad that whole emotion thing didn't scare him off!
+${AI_NAME}: /think I wonder what ${USER_NAME} likes to do in his free time? I should ask him about that!
+${AI_NAME}: What do you like to do in your free time? ^_^
+${USER_NAME}:" "$@"

examples/common.cpp

@@ -39,6 +39,8 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
 
     bool invalid_param = false;
     std::string arg;
+    gpt_params default_params;
+
     for (int i = 1; i < argc; i++) {
         arg = argv[i];
 
@@ -66,6 +68,11 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             std::ifstream file(argv[i]);
+            if (!file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
             std::copy(std::istreambuf_iterator<char>(file), std::istreambuf_iterator<char>(), back_inserter(params.prompt));
             if (params.prompt.back() == '\n') {
                 params.prompt.pop_back();
@@ -168,7 +175,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             }
             params.n_parts = std::stoi(argv[i]);
         } else if (arg == "-h" || arg == "--help") {
-            gpt_print_usage(argc, argv, params);
+            gpt_print_usage(argc, argv, default_params);
             exit(0);
         } else if (arg == "--random-prompt") {
             params.random_prompt = true;
@@ -180,13 +187,13 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.input_prefix = argv[i];
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
-            gpt_print_usage(argc, argv, params);
+            gpt_print_usage(argc, argv, default_params);
             exit(1);
         }
     }
     if (invalid_param) {
         fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
-        gpt_print_usage(argc, argv, params);
+        gpt_print_usage(argc, argv, default_params);
         exit(1);
     }
 

examples/gpt4all.sh

@@ -0,0 +1,15 @@
+#!/bin/bash
+
+#
+# Temporary script - will be removed in the future
+#
+
+cd `dirname $0`
+cd ..
+
+./main --color --instruct --threads 4 \
+       --model ./models/gpt4all-7B/gpt4all-lora-quantized.bin \
+       --file ./prompts/alpaca.txt \
+       --batch_size 8 --ctx_size 2048 \
+       --repeat_last_n 64 --repeat_penalty 1.3 \
+       --n_predict 128 --temp 0.1 --top_k 40 --top_p 0.95

examples/main/main.cpp

@@ -368,6 +368,11 @@ int main(int argc, char ** argv) {
                 // potentially set color to indicate we are taking user input
                 set_console_color(con_st, CONSOLE_COLOR_USER_INPUT);
 
+#if defined (_WIN32)
+                // Windows: must reactivate sigint handler after each signal
+                signal(SIGINT, sigint_handler);
+#endif
+
                 if (params.instruct) {
                     printf("\n> ");
                 }

ggml.h

@@ -258,11 +258,11 @@ struct ggml_tensor {
     enum ggml_type type;
 
     int    n_dims;
-    int    ne[GGML_MAX_DIMS]; // number of elements
-    size_t nb[GGML_MAX_DIMS]; // stride in bytes:
-                              // nb[0] = sizeof(type)
-                              // nb[1] = nb[0]   * ne[0] + padding
-                              // nb[i] = nb[i-1] * ne[i-1]
+    int64_t ne[GGML_MAX_DIMS]; // number of elements
+    size_t  nb[GGML_MAX_DIMS]; // stride in bytes:
+                               // nb[0] = sizeof(type)
+                               // nb[1] = nb[0]   * ne[0] + padding
+                               // nb[i] = nb[i-1] * ne[i-1]
 
     // compute data
     enum ggml_op op;
@@ -328,8 +328,8 @@ int64_t ggml_cycles_per_ms(void);
 void ggml_print_object (const struct ggml_object * obj);
 void ggml_print_objects(const struct ggml_context * ctx);
 
-int    ggml_nelements(const struct ggml_tensor * tensor);
-size_t ggml_nbytes   (const struct ggml_tensor * tensor);
+int64_t ggml_nelements(const struct ggml_tensor * tensor);
+size_t  ggml_nbytes   (const struct ggml_tensor * tensor);
 
 int    ggml_blck_size (enum ggml_type type);
 size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
@@ -355,33 +355,33 @@ struct ggml_tensor * ggml_new_tensor(
         struct ggml_context * ctx,
         enum   ggml_type type,
         int    n_dims,
-        const int *ne);
+        const int64_t *ne);
 
 struct ggml_tensor * ggml_new_tensor_1d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0);
+        int64_t ne0);
 
 struct ggml_tensor * ggml_new_tensor_2d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0,
-        int    ne1);
+        int64_t ne0,
+        int64_t ne1);
 
 struct ggml_tensor * ggml_new_tensor_3d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0,
-        int    ne1,
-        int    ne2);
+        int64_t ne0,
+        int64_t ne1,
+        int64_t ne2);
 
 struct ggml_tensor * ggml_new_tensor_4d(
         struct ggml_context * ctx,
         enum   ggml_type type,
-        int    ne0,
-        int    ne1,
-        int    ne2,
-        int    ne3);
+        int64_t ne0,
+        int64_t ne1,
+        int64_t ne2,
+        int64_t ne3);
 
 struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
 struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
@@ -531,33 +531,43 @@ struct ggml_tensor * ggml_reshape(
 struct ggml_tensor * ggml_reshape_2d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1);
+        int64_t               ne0,
+        int64_t               ne1);
 
 // return view(a)
 // TODO: when we start computing gradient, make a copy instead of view
 struct ggml_tensor * ggml_reshape_3d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1,
-        int                   ne2);
+        int64_t               ne0,
+        int64_t               ne1,
+        int64_t               ne2);
 
 // offset in bytes
 struct ggml_tensor * ggml_view_1d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
+        int64_t               ne0,
         size_t                offset);
 
 struct ggml_tensor * ggml_view_2d(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        int                   ne0,
-        int                   ne1,
+        int64_t               ne0,
+        int64_t               ne1,
         size_t                nb1, // row stride in bytes
         size_t                offset);
 
+struct ggml_tensor * ggml_view_3d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int64_t               ne0,
+        int64_t               ne1,
+        int64_t               ne2,
+        size_t                nb1, // row   stride in bytes
+        size_t                nb2, // slice stride in bytes
+        size_t                offset);
+
 struct ggml_tensor * ggml_permute(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,

llama.cpp

@@ -256,8 +256,8 @@ static bool kv_cache_init(
     const int n_embd  = hparams.n_embd;
     const int n_layer = hparams.n_layer;
 
-    const int n_mem      = n_layer*n_ctx;
-    const int n_elements = n_embd*n_mem;
+    const int64_t n_mem      = (int64_t)n_layer*n_ctx;
+    const int64_t n_elements = n_embd*n_mem;
 
     cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
 
@@ -679,7 +679,7 @@ static bool llama_model_load(
                 return false;
             }
             if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
-                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
+                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%" PRId64 ", %" PRId64 "], expected [%d, %d]\n",
                         __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
                 return false;
             }
@@ -810,37 +810,35 @@ static bool llama_eval_internal(
 
         // self-attention
         {
-            struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
-            struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
-            struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+            // compute Q and K and RoPE them
+            struct ggml_tensor * Qcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wq, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0);
+            struct ggml_tensor * Kcur = ggml_rope(ctx0, ggml_reshape_3d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wk, cur), n_embd/n_head, n_head, N), n_past, n_rot, 0);
 
             // store key and value to memory
-            if (N >= 1) {
-                struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
-                struct ggml_tensor * v = ggml_view_1d(ctx0, kv_self.v, N*n_embd, (ggml_element_size(kv_self.v)*n_embd)*(il*n_ctx + n_past));
+            {
+                // compute the transposed [N, n_embd] V matrix
+                struct ggml_tensor * Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, ggml_mul_mat(ctx0, model.layers[il].wv, cur), n_embd, N));
 
+                struct ggml_tensor * k = ggml_view_1d(ctx0, kv_self.k, N*n_embd, (ggml_element_size(kv_self.k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_2d(ctx0, kv_self.v, N, n_embd,
+                        (   n_ctx)*ggml_element_size(kv_self.v),
+                        (il*n_ctx)*ggml_element_size(kv_self.v)*n_embd + n_past*ggml_element_size(kv_self.v));
+
+                // important: storing RoPE-ed version of K in the KV cache!
                 ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
                 ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
             }
 
-            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
             struct ggml_tensor * Q =
                 ggml_permute(ctx0,
-                        ggml_rope(ctx0,
-                            ggml_cpy(ctx0,
-                                Qcur,
-                                ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
-                            n_past, n_rot, 0),
+                        Qcur,
                         0, 2, 1, 3);
 
-            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
             struct ggml_tensor * K =
                 ggml_permute(ctx0,
-                        ggml_rope(ctx0,
-                            ggml_reshape_3d(ctx0,
-                                ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.k)*n_embd),
-                                n_embd/n_head, n_head, n_past + N),
-                            n_past, n_rot, 1),
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, kv_self.k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.k)*n_embd),
+                            n_embd/n_head, n_head, n_past + N),
                         0, 2, 1, 3);
 
             // K * Q
@@ -858,18 +856,23 @@ static bool llama_eval_internal(
             // KQ = soft_max(KQ_masked)
             struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
 
-            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
-            struct ggml_tensor * V_trans =
-                ggml_cpy(ctx0,
-                    ggml_permute(ctx0,
-                            ggml_reshape_3d(ctx0,
-                                ggml_view_1d(ctx0, kv_self.v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(kv_self.v)*n_embd),
-                                n_embd/n_head, n_head, n_past + N),
-                            1, 2, 0, 3),
-                    ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd/n_head, n_head));
+            // split cached V into n_head heads
+            struct ggml_tensor * V =
+                ggml_view_3d(ctx0, kv_self.v,
+                        n_past + N, n_embd/n_head, n_head,
+                        n_ctx*ggml_element_size(kv_self.v),
+                        n_ctx*ggml_element_size(kv_self.v)*n_embd/n_head,
+                        il*n_ctx*ggml_element_size(kv_self.v)*n_embd);
 
-            // KQV = transpose(V) * KQ_soft_max
-            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
+#if 1
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
+#else
+            // make V contiguous in memory to speed up the matmul, however we waste time on the copy
+            // on M1 this is faster for the perplexity computation, but ~5% slower for the single-token generation
+            // is there a better way?
+            struct ggml_tensor * V_cont = ggml_cpy(ctx0, V, ggml_new_tensor_3d(ctx0, kv_self.v->type, n_past + N, n_embd/n_head, n_head));
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_cont, KQ_soft_max);
+#endif
 
             // KQV_merged = KQV.permute(0, 2, 1, 3)
             struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
@@ -955,9 +958,13 @@ static bool llama_eval_internal(
     ggml_build_forward_expand(&gf, inpL);
     ggml_graph_compute       (ctx0, &gf);
 
+    // print timing information per ggml operation (for debugging purposes)
+    // requires GGML_PERF to be defined
+    //ggml_graph_print(&gf);
+
+    // plot the computation graph in dot format (for debugging purposes)
     //if (n_past%100 == 0) {
-    //    ggml_graph_print   (&gf);
-    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
+    //    ggml_graph_dump_dot(&gf, NULL, "llama.dot");
     //}
 
     //embd_w.resize(n_vocab*N);
@@ -1194,6 +1201,20 @@ static llama_vocab::id llama_sample_top_p_top_k(
     const auto & logits = lctx.logits;
     const auto * plogits = logits.data() + logits.size() - n_logits;
 
+    if (temp <= 0) {
+        // select the token with the highest logit directly
+        float max_logit = plogits[0];
+        llama_vocab::id max_id = 0;
+
+        for (int i = 1; i < n_logits; ++i) {
+            if (plogits[i] > max_logit) {
+                max_logit = plogits[i];
+                max_id = i;
+            }
+        }
+        return max_id;
+    }
+
     std::vector<std::pair<float, llama_vocab::id>> logits_id;
     logits_id.reserve(n_logits);
 
@@ -1215,7 +1236,9 @@ static llama_vocab::id llama_sample_top_p_top_k(
         }
     }
 
-    sample_top_k(logits_id, top_k);
+    if (top_k > 0 && top_k < n_logits) {
+        sample_top_k(logits_id, top_k);
+    }
 
     float maxl = -std::numeric_limits<float>::infinity();
     for (const auto & kv : logits_id) {
@@ -1608,7 +1631,7 @@ struct llama_context * llama_init_from_file(
     }
 
     // reserve memory for context buffers
-    {
+    if (!params.vocab_only) {
         if (!kv_cache_init(ctx->model.hparams, ctx->model.kv_self, memory_type, ctx->model.hparams.n_ctx)) {
             fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
             llama_free(ctx);
@@ -1668,6 +1691,33 @@ int llama_model_quantize(
     return 0;
 }
 
+// Returns the KV cache that will contain the context for the
+// ongoing prediction with the model.
+const uint8_t * llama_get_kv_cache(struct llama_context * ctx) {
+    return ctx->model.kv_self.buf.data();
+}
+
+// Returns the size of the KV cache
+size_t llama_get_kv_cache_size(struct llama_context * ctx) {
+    return ctx->model.kv_self.buf.size();
+}
+
+int llama_get_kv_cache_token_count(struct llama_context * ctx) {
+    return ctx->model.kv_self.n;
+}
+
+// Sets the KV cache containing the current context for the model
+void llama_set_kv_cache(
+        struct llama_context * ctx,
+               const uint8_t * kv_cache,
+                      size_t   n_size,
+                         int   n_token_count) {
+    // Make sure we have the same kv cache setup
+    LLAMA_ASSERT(ctx->model.kv_self.buf.size() == n_size);
+    memcpy(ctx->model.kv_self.buf.data(), kv_cache, n_size);
+    ctx->model.kv_self.n = n_token_count;
+}
+
 int llama_eval(
         struct llama_context * ctx,
            const llama_token * tokens,

llama.h

@@ -83,6 +83,23 @@ extern "C" {
             const char * fname_out,
                    int   itype);
 
+    // Returns the KV cache that will contain the context for the
+    // ongoing prediction with the model.
+    LLAMA_API const uint8_t * llama_get_kv_cache(struct llama_context * ctx);
+
+    // Returns the size of the KV cache
+    LLAMA_API size_t llama_get_kv_cache_size(struct llama_context * ctx);
+
+    // Returns the number of tokens in the KV cache
+    LLAMA_API int llama_get_kv_cache_token_count(struct llama_context * ctx);
+
+    // Sets the KV cache containing the current context for the model
+    LLAMA_API void llama_set_kv_cache(
+            struct llama_context * ctx,
+                   const uint8_t * kv_cache,
+                          size_t   n_size,
+                             int   n_token_count);
+
     // Run the llama inference to obtain the logits and probabilities for the next token.
     // tokens + n_tokens is the provided batch of new tokens to process
     // n_past is the number of tokens to use from previous eval calls