llama : define magic numbers as integer constants (#1518) (#1520)

The underlying representation of multibyte character literals is
implementation-defined. This could, at least in principle, cause
cross-build data export/import issues independent of endianness.

Define magic numbers as integer literals to be on the safe side.

Signed-off-by: Juuso Alasuutari <juuso.alasuutari@gmail.com>

Makefile

@@ -74,6 +74,15 @@ ifeq ($(UNAME_S),Haiku)
 	CXXFLAGS += -pthread
 endif
 
+ifdef LLAMA_GPROF
+	CFLAGS   += -pg
+	CXXFLAGS += -pg
+endif
+ifdef LLAMA_PERF
+	CFLAGS   += -DGGML_PERF
+	CXXFLAGS += -DGGML_PERF
+endif
+
 # Architecture specific
 # 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
@@ -106,7 +115,7 @@ ifndef LLAMA_NO_ACCELERATE
 	endif
 endif
 ifdef LLAMA_OPENBLAS
-	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas
+	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas -I/usr/include/openblas
 	ifneq ($(shell grep -e "Arch Linux" -e "ID_LIKE=arch" /etc/os-release 2>/dev/null),)
 		LDFLAGS += -lopenblas -lcblas
 	else
@@ -135,14 +144,6 @@ ifdef LLAMA_CLBLAST
 ggml-opencl.o: ggml-opencl.c ggml-opencl.h
 	$(CC) $(CFLAGS) -c $< -o $@
 endif
-ifdef LLAMA_GPROF
-	CFLAGS   += -pg
-	CXXFLAGS += -pg
-endif
-ifdef LLAMA_PERF
-	CFLAGS   += -DGGML_PERF
-	CXXFLAGS += -DGGML_PERF
-endif
 ifneq ($(filter aarch64%,$(UNAME_M)),)
 	# Apple M1, M2, etc.
 	# Raspberry Pi 3, 4, Zero 2 (64-bit)

README.md

@@ -9,6 +9,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 **Hot topics:**
 
+- Quantization formats `Q4` and `Q8` have changed again (19 May) - [(info)](https://github.com/ggerganov/llama.cpp/pull/1508)
 - Quantization formats `Q4` and `Q5` have changed - requantize any old models [(info)](https://github.com/ggerganov/llama.cpp/pull/1405)
 - [Roadmap May 2023](https://github.com/ggerganov/llama.cpp/discussions/1220)
 
@@ -80,6 +81,7 @@ as the main playground for developing new features for the [ggml](https://github
 - [X] [Koala](https://bair.berkeley.edu/blog/2023/04/03/koala/)
 - [X] [OpenBuddy 🐶 (Multilingual)](https://github.com/OpenBuddy/OpenBuddy)
 - [X] [Pygmalion 7B / Metharme 7B](#using-pygmalion-7b--metharme-7b)
+- [X] [WizardLM](https://github.com/nlpxucan/WizardLM)
 
 **Bindings:**
 
@@ -333,16 +335,16 @@ Several quantization methods are supported. They differ in the resulting model d
 
 | Model | Measure      | F16    | Q4_0   | Q4_1   | Q5_0   | Q5_1   | Q8_0   |
 |------:|--------------|-------:|-------:|-------:|-------:|-------:|-------:|
-|    7B | perplexity   | 5.9066 | 6.1565 | 6.0910 | 5.9862 | 5.9481 | 5.9069 |
-|    7B | file size    |  13.0G |   4.0G |   4.8G |   4.4G |   4.8G |   7.1G |
-|    7B | ms/tok @ 4th |    128 |     50 |     54 |     75 |     83 |     75 |
-|    7B | ms/tok @ 8th |    123 |     44 |     52 |     53 |     58 |     72 |
-|    7B | bits/weight  |   16.0 |    5.0 |    6.0 |    5.5 |    6.0 |    9.0 |
-|   13B | perplexity   | 5.2543 | 5.3860 | 5.3607 | 5.2856 | 5.2706 | 5.2548 |
-|   13B | file size    |  25.0G |   7.6G |   9.1G |   8.4G |   9.1G |    14G |
-|   13B | ms/tok @ 4th |    239 |     93 |    101 |    150 |    164 |    141 |
-|   13B | ms/tok @ 8th |    240 |     81 |     96 |     96 |    104 |    136 |
-|   13B | bits/weight  |   16.0 |    5.0 |    6.0 |    5.5 |    6.0 |    9.0 |
+|    7B | perplexity   | 5.9066 | 6.1565 | 6.0912 | 5.9862 | 5.9481 | 5.9070 |
+|    7B | file size    |  13.0G |   3.5G |   3.9G |   4.3G |   4.7G |   6.7G |
+|    7B | ms/tok @ 4th |    127 |     55 |     54 |     76 |     83 |     72 |
+|    7B | ms/tok @ 8th |    122 |     43 |     45 |     52 |     56 |     67 |
+|    7B | bits/weight  |   16.0 |    4.5 |    5.0 |    5.5 |    6.0 |    8.5 |
+|   13B | perplexity   | 5.2543 | 5.3860 | 5.3608 | 5.2856 | 5.2706 | 5.2548 |
+|   13B | file size    |  25.0G |   6.8G |   7.6G |   8.3G |   9.1G |    13G |
+|   13B | ms/tok @ 4th |      - |    103 |    105 |    148 |    160 |    131 |
+|   13B | ms/tok @ 8th |      - |     73 |     82 |     98 |    105 |    128 |
+|   13B | bits/weight  |   16.0 |    4.5 |    5.0 |    5.5 |    6.0 |    8.5 |
 
 ### Perplexity (measuring model quality)
 

convert.py

@@ -121,7 +121,6 @@ def make_tensors_list() -> List[str]:
             f'layers.{i}.feed_forward.w1.weight',
             f'layers.{i}.feed_forward.w2.weight',
             f'layers.{i}.feed_forward.w3.weight',
-            f'layers.{i}.atttention_norm.weight',
             f'layers.{i}.ffn_norm.weight',
         ]
     return ret
@@ -1055,7 +1054,7 @@ def load_some_model(path: Path) -> ModelPlus:
         files = list(path.glob("model-00001-of-*.safetensors"))
         if not files:
             # Try the PyTorch patterns too, with lower priority
-            globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt"]
+            globs = ["consolidated.00.pth", "pytorch_model-00001-of-*.bin", "*.pt", "pytorch_model.bin" ]
             files = [file for glob in globs for file in path.glob(glob)]
         if not files:
             # Try GGML too, but with lower priority, since if both a non-GGML

examples/CMakeLists.txt

@@ -36,4 +36,5 @@ else()
     add_subdirectory(embedding)
     add_subdirectory(save-load-state)
     add_subdirectory(benchmark)
+    add_subdirectory(baby-llama)
 endif()

examples/baby-llama/CMakeLists.txt

@@ -0,0 +1,4 @@
+set(TARGET baby-llama)
+add_executable(${TARGET} baby-llama.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/benchmark/benchmark-matmult.cpp

@@ -1,6 +1,7 @@
-#include <locale.h>
 #include "ggml.h"
 #include "build-info.h"
+
+#include <locale.h>
 #include <assert.h>
 #include <math.h>
 #include <cstring>
@@ -15,7 +16,7 @@
 #include <iterator>
 #include <algorithm>
 
-float tensor_sum_elements(struct ggml_tensor * tensor) {
+float tensor_sum_elements(const ggml_tensor * tensor) {
     float sum = 0;
     if (tensor->type==GGML_TYPE_F32) {
         for (int j = 0; j < tensor->ne[1]; j++) {
@@ -27,21 +28,15 @@ float tensor_sum_elements(struct ggml_tensor * tensor) {
     return sum;
 }
 
+void tensor_dump(const ggml_tensor * tensor, const char * name) {
+    printf("%15s: type = %i (%5s) ne = %5d x %5d x %5d, nb = (%5li, %5li, %5li) - ", name,
+        tensor->type, ggml_type_name(tensor->type),
+        (int) tensor->ne[0], (int) tensor->ne[1], (int) tensor->ne[2], tensor->nb[0], tensor->nb[1], tensor->nb[2]);
+    float sum = tensor_sum_elements(tensor);
+    printf("Sum of tensor %s is %6.2f\n", name, sum);
+}
 
-/*
-    These are mapping to unknown
-    GGML_TYPE_I8,
-    GGML_TYPE_I16,
-    GGML_TYPE_I32,
-    GGML_TYPE_COUNT,
-*/
-
-#define TENSOR_TYPE_AS_STR(TYPE) TYPE == GGML_TYPE_F32 ? "FP32" : TYPE == GGML_TYPE_F16 ? "FP16" : TYPE == GGML_TYPE_Q4_0 ? "Q4_0" : TYPE == GGML_TYPE_Q4_1 ? "Q4_1" : "UNKNOWN"
-
-#define TENSOR_DUMP(TENSOR) printf("%15s: type = %i (%5s) ne = %5d x %5d x %5d, nb = (%5li, %5li, %5li) - ", #TENSOR, \
-        TENSOR->type,TENSOR_TYPE_AS_STR(TENSOR->type),\
-        (int) TENSOR->ne[0], (int) TENSOR->ne[1], (int) TENSOR->ne[2], TENSOR->nb[0], TENSOR->nb[1], TENSOR->nb[2]); \
-    { float sum = tensor_sum_elements(TENSOR); printf("Sum of tensor %s is %6.2f\n",#TENSOR, sum); }
+#define TENSOR_DUMP(tensor) tensor_dump(tensor, #tensor)
 
 struct benchmark_params_struct {
     int32_t n_threads     = 1;
@@ -59,8 +54,6 @@ void print_usage(int /*argc*/, char ** argv, struct benchmark_params_struct para
 }
 
 int main(int argc, char ** argv)  {
-
-
     struct benchmark_params_struct benchmark_params;
 
     bool invalid_param = false;
@@ -84,11 +77,11 @@ int main(int argc, char ** argv)  {
             print_usage(argc, argv, benchmark_params);
             exit(0);
         }
-        if (invalid_param) {
-            fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
-            print_usage(argc, argv, benchmark_params);
-            exit(1);
-        }
+    }
+    if (invalid_param) {
+        fprintf(stderr, "error: invalid parameter for argument: %s\n", arg.c_str());
+        print_usage(argc, argv, benchmark_params);
+        exit(1);
     }
 
     fprintf(stderr, "%s: build = %d (%s)\n", __func__, BUILD_NUMBER, BUILD_COMMIT);
@@ -216,10 +209,10 @@ int main(int argc, char ** argv)  {
     // Let's use the F32 result from above as a reference for the q4_0 multiplication
     float sum_of_F32_reference = tensor_sum_elements(gf.nodes[0]);
 
+    printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; gigaFLOPS\n");
+    printf("=====================================================================================\n");
 
-    printf("Iteration;NThreads; SizeX; SizeY; SizeZ; Required_FLOPS; Elapsed_u_Seconds; FLOPS_per_u_Second\n");
-    printf("==============================================================================================\n");
-
+    double  gflops_sum = 0;
     for (int i=0;i<benchmark_params.n_iterations ;i++) {
 
         long long int start = ggml_time_us();
@@ -227,12 +220,13 @@ int main(int argc, char ** argv)  {
         ggml_graph_compute(ctx, &gf31);
         long long int stop = ggml_time_us();
         long long int usec = stop-start;
-        float flops_per_usec = (1.0f*flops_per_matrix)/usec;
-        printf("%9i;%8i;%6i;%6i;%6i;%15lli;%18lli;%19.2f\n",
+        double gflops = (double)(flops_per_matrix)/usec/1000.0;
+        gflops_sum += gflops;
+        printf("%9i;%8i;%6i;%6i;%6i;%15lli;%18lli;%10.2f\n",
             i,
             gf31.n_threads,
             sizex, sizey, sizez, flops_per_matrix,
-            usec,flops_per_usec);
+            usec,gflops);
 
 #ifdef VERBOSE_DEBUGGING
         TENSOR_DUMP("res",gf31.nodes[0])
@@ -256,7 +250,8 @@ int main(int argc, char ** argv)  {
 
         // Running a different graph computation to make sure we override the CPU cache lines
         ggml_graph_compute(ctx, &gf32);
-
     }
-
+    printf("\n");
+    printf("Average%78.2f\n",gflops_sum/((double)benchmark_params.n_iterations));
+    printf("=====================================================================================\n");
 }

examples/chat-persistent.sh

@@ -0,0 +1,151 @@
+#!/bin/bash
+
+set -euo pipefail
+
+cd "$(dirname "$0")/.." || exit
+
+if [[ -z "${PROMPT_CACHE_FILE+x}" || -z "${CHAT_SAVE_DIR+x}" ]]; then
+    echo >&2 "error: PROMPT_CACHE_FILE and CHAT_SAVE_DIR must be provided"
+    exit 1
+fi
+
+MODEL="${MODEL:-./models/13B/ggml-model-q4_0.bin}"
+PROMPT_TEMPLATE="${PROMPT_TEMPLATE:-./prompts/chat.txt}"
+USER_NAME="${USER_NAME:-User}"
+AI_NAME="${AI_NAME:-ChatLLaMa}"
+DATE_TIME="$(date +%H:%M)"
+DATE_YEAR="$(date +%Y)"
+
+LOG="${CHAT_SAVE_DIR}/main.log"
+LOG_BG="${CHAT_SAVE_DIR}/main-bg.log"
+CUR_PROMPT_FILE="${CHAT_SAVE_DIR}/current-prompt.txt"
+CUR_PROMPT_CACHE="${CHAT_SAVE_DIR}/current-cache.bin"
+NEXT_PROMPT_FILE="${CHAT_SAVE_DIR}/next-prompt.txt"
+NEXT_PROMPT_CACHE="${CHAT_SAVE_DIR}/next-cache.bin"
+
+SESSION_SIZE_MSG_PATTERN='main: session file matches \d+ / \d+'
+SAMPLE_TIME_MSG_PATTERN='sample time =\s+\d+.\d+ ms /\s+\d+'
+SED_DELETE_MESSAGES="/^(${USER_NAME}:|${AI_NAME}:|\\.\\.\\.)/,\$d"
+
+CTX_SIZE=2048
+CTX_ROTATE_POINT=$((CTX_SIZE * 3 / 5)) # REVIEW
+OPTS=(--model "$MODEL" --ctx_size "$CTX_SIZE" --repeat_last_n 256 "$@")
+
+# An unbuffered `tail -c+N`
+skip_bytes() {
+    LANG=C IFS= read -r -n "$1" -d '' c
+    while LANG=C IFS= read -r -n 1 -d '' c; do
+        printf '%s' "$c"
+    done
+}
+
+mkdir -p "$CHAT_SAVE_DIR"
+echo >"$LOG"
+trap "tail -n100 ${LOG}" EXIT
+
+if [[ ! -e "$CUR_PROMPT_FILE" ]]; then
+    sed -e "s/\[\[USER_NAME\]\]/${USER_NAME}/g" \
+        -e "s/\[\[AI_NAME\]\]/${AI_NAME}/g" \
+        -e "s/\[\[DATE_TIME\]\]/${DATE_TIME}/g" \
+        -e "s/\[\[DATE_YEAR\]\]/${DATE_YEAR}/g" \
+        "$PROMPT_TEMPLATE" >"$CUR_PROMPT_FILE"
+fi
+
+if [[ ! -e "$NEXT_PROMPT_FILE" ]]; then
+    sed -r "$SED_DELETE_MESSAGES" "$CUR_PROMPT_FILE" >"$NEXT_PROMPT_FILE"
+fi
+
+if [[ "$(tail -c4 "$NEXT_PROMPT_FILE")" != "..." ]]; then
+    echo '...' >>"$NEXT_PROMPT_FILE"
+fi
+
+if [[ ! -e "$PROMPT_CACHE_FILE" ]]; then
+    echo 'Prompt cache does not exist, building...'
+    # Default batch_size to 8 here for better user feedback during initial prompt processing
+    ./main 2>>"$LOG" \
+        --batch_size 8 \
+        "${OPTS[@]}" \
+        --prompt-cache "$PROMPT_CACHE_FILE" \
+        --file "$CUR_PROMPT_FILE" \
+        --n_predict 1
+    echo
+    echo 'Done!'
+fi
+
+if [[ ! -e "$CUR_PROMPT_CACHE" ]]; then
+    cp "$PROMPT_CACHE_FILE" "$CUR_PROMPT_CACHE"
+fi
+if [[ ! -e "$NEXT_PROMPT_CACHE" ]]; then
+    cp "$PROMPT_CACHE_FILE" "$NEXT_PROMPT_CACHE"
+fi
+
+printf '%s ' "$(< "$CUR_PROMPT_FILE")"
+n_tokens=0
+
+while read -e line; do
+    # Limit generation to remaining context, with a buffer and estimating 2 chars/token for input
+    n_predict=$((CTX_SIZE - n_tokens - ${#line} / 2 - 32))
+
+    # Swap prompts when we're about to run out of context
+    if ((n_predict <= 0)); then
+        wait # for background main (below) to finish with next prompt
+        mv "$NEXT_PROMPT_FILE"  "$CUR_PROMPT_FILE"
+        mv "$NEXT_PROMPT_CACHE" "$CUR_PROMPT_CACHE"
+
+        sed -r "$SED_DELETE_MESSAGES" "$CUR_PROMPT_FILE" >"$NEXT_PROMPT_FILE"
+        echo '...' >>"$NEXT_PROMPT_FILE"
+        cp "$PROMPT_CACHE_FILE" "$NEXT_PROMPT_CACHE"
+
+        n_tokens=0
+        n_predict=$((CTX_SIZE / 2))
+    fi
+
+    echo " ${line}" >>"$CUR_PROMPT_FILE"
+    if ((n_tokens > CTX_ROTATE_POINT)); then
+        echo " ${line}" >>"$NEXT_PROMPT_FILE"
+    fi
+
+    n_prompt_len_pre=$(($(wc -c <"$CUR_PROMPT_FILE")))
+
+    printf '%s: ' "$AI_NAME" >>"$CUR_PROMPT_FILE"
+
+    ./main 2>>"$LOG" "${OPTS[@]}" \
+            --prompt-cache "$CUR_PROMPT_CACHE" \
+            --prompt-cache-all \
+            --file "$CUR_PROMPT_FILE" \
+            --reverse-prompt "${USER_NAME}:" \
+            --n_predict "$n_predict" |
+        skip_bytes 1 |                  # skip BOS token added by ./main
+        tee "$CUR_PROMPT_FILE.tmp" |    # save prompt + generation to tmp file
+        skip_bytes "$n_prompt_len_pre"  # print generation
+
+    mv "$CUR_PROMPT_FILE.tmp" "$CUR_PROMPT_FILE"
+
+    # if we hit n_predict instead of reverse-prompt, we need to add the prompt
+    if [[ "$(tail -n1 "$CUR_PROMPT_FILE")" != "${USER_NAME}:" ]]; then
+        printf '\n%s:' "$USER_NAME"
+        printf '\n%s:' "$USER_NAME" >> "$CUR_PROMPT_FILE"
+    fi
+
+    printf ' '
+
+    # HACK get num tokens from debug message
+    # TODO get both messages in one go
+    if  ! session_size_msg="$(tail -n30 "$LOG" | grep -oE "$SESSION_SIZE_MSG_PATTERN")" ||
+        ! sample_time_msg="$( tail -n10 "$LOG" | grep -oE "$SAMPLE_TIME_MSG_PATTERN")"; then
+        echo >&2 "Couldn't get number of tokens from ./main output!"
+        exit 1
+    fi
+
+    n_tokens=$(($(cut -d/ -f2 <<<"$session_size_msg") + $(cut -d/ -f2 <<<"$sample_time_msg")))
+
+    if ((n_tokens > CTX_ROTATE_POINT)); then
+        tail -c+$((n_prompt_len_pre + 1)) "$CUR_PROMPT_FILE" >>"$NEXT_PROMPT_FILE"
+    fi
+
+    # Update cache for next prompt in background, ideally during user input
+    ./main >>"$LOG_BG" 2>&1 "${OPTS[@]}" \
+          --prompt-cache "$NEXT_PROMPT_CACHE" \
+          --file "$NEXT_PROMPT_FILE" \
+          --n_predict 1 &
+done

examples/common.cpp

@@ -8,6 +8,7 @@
 #include <iterator>
 #include <algorithm>
 #include <sstream>
+#include <unordered_set>
 
 #if defined(__APPLE__) && defined(__MACH__)
 #include <sys/types.h>
@@ -28,21 +29,21 @@
 
 int32_t get_num_physical_cores() {
 #ifdef __linux__
-    std::ifstream cpuinfo("/proc/cpuinfo");
-    std::string line;
-    while (std::getline(cpuinfo, line)) {
-        std::size_t pos = line.find("cpu cores");
-        if (pos != std::string::npos) {
-            pos = line.find(": ", pos);
-            if (pos != std::string::npos) {
-                try {
-                    // Extract the number and return it
-                    return static_cast<int32_t>(std::stoul(line.substr(pos + 2)));
-                } catch (const std::invalid_argument &) {
-                    // Ignore if we could not parse
-                }
-            }
+    // enumerate the set of thread siblings, num entries is num cores
+    std::unordered_set<std::string> siblings;
+    for (uint32_t cpu=0; cpu < UINT32_MAX; ++cpu) {
+        std::ifstream thread_siblings("/sys/devices/system/cpu"
+            + std::to_string(cpu) + "/topology/thread_siblings");
+        if (!thread_siblings.is_open()) {
+            break; // no more cpus
         }
+        std::string line;
+        if (std::getline(thread_siblings, line)) {
+            siblings.insert(line);
+        }
+    }
+    if (siblings.size() > 0) {
+        return static_cast<int32_t>(siblings.size());
     }
 #elif defined(__APPLE__) && defined(__MACH__)
     int32_t num_physical_cores;
@@ -277,6 +278,12 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
             params.use_color = true;
         } else if (arg == "--mlock") {
             params.use_mlock = true;
+        } else if (arg == "--gpu-layers" || arg == "-ngl" || arg == "--n-gpu-layers") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_gpu_layers = std::stoi(argv[i]);
         } else if (arg == "--no-mmap") {
             params.use_mmap = false;
         } else if (arg == "--mtest") {
@@ -314,12 +321,6 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 invalid_param = true;
                 break;
             }
-        } else if (arg == "--n-parts") {
-            if (++i >= argc) {
-                invalid_param = true;
-                break;
-            }
-            params.n_parts = std::stoi(argv[i]);
         } else if (arg == "-h" || arg == "--help") {
             gpt_print_usage(argc, argv, default_params);
             exit(0);
@@ -350,7 +351,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
     }
     if (params.prompt_cache_all &&
             (params.interactive || params.interactive_first ||
-             params.instruct || params.antiprompt.size())) {
+             params.instruct)) {
         fprintf(stderr, "error: --prompt-cache-all not supported in interactive mode yet\n");
         gpt_print_usage(argc, argv, default_params);
         exit(1);
@@ -372,8 +373,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
     fprintf(stderr, "  --multiline-input     allows you to write or paste multiple lines without ending each in '\\'\n");
     fprintf(stderr, "  -r PROMPT, --reverse-prompt PROMPT\n");
-    fprintf(stderr, "                        run in interactive mode and poll user input upon seeing PROMPT (can be\n");
-    fprintf(stderr, "                        specified more than once for multiple prompts).\n");
+    fprintf(stderr, "                        halt generation at PROMPT, return control in interactive mode\n");
+    fprintf(stderr, "                        (can be specified more than once for multiple prompts).\n");
     fprintf(stderr, "  --color               colorise output to distinguish prompt and user input from generations\n");
     fprintf(stderr, "  -s SEED, --seed SEED  RNG seed (default: -1, use random seed for < 0)\n");
     fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
@@ -411,7 +412,6 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  --no-penalize-nl      do not penalize newline token\n");
     fprintf(stderr, "  --memory-f32          use f32 instead of f16 for memory key+value\n");
     fprintf(stderr, "  --temp N              temperature (default: %.1f)\n", (double)params.temp);
-    fprintf(stderr, "  --n-parts N           number of model parts (default: -1 = determine from dimensions)\n");
     fprintf(stderr, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
     fprintf(stderr, "  --perplexity          compute perplexity over the prompt\n");
     fprintf(stderr, "  --keep                number of tokens to keep from the initial prompt (default: %d, -1 = all)\n", params.n_keep);
@@ -421,6 +421,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     if (llama_mmap_supported()) {
         fprintf(stderr, "  --no-mmap             do not memory-map model (slower load but may reduce pageouts if not using mlock)\n");
     }
+    fprintf(stderr, "  -ngl N, --n-gpu-layers N\n");
+    fprintf(stderr, "                        number of layers to store in VRAM\n");
     fprintf(stderr, "  --mtest               compute maximum memory usage\n");
     fprintf(stderr, "  --verbose-prompt      print prompt before generation\n");
     fprintf(stderr, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
@@ -463,14 +465,14 @@ std::vector<llama_token> llama_tokenize(struct llama_context * ctx, const std::s
 struct llama_context * llama_init_from_gpt_params(const gpt_params & params) {
     auto lparams = llama_context_default_params();
 
-    lparams.n_ctx      = params.n_ctx;
-    lparams.n_parts    = params.n_parts;
-    lparams.seed       = params.seed;
-    lparams.f16_kv     = params.memory_f16;
-    lparams.use_mmap   = params.use_mmap;
-    lparams.use_mlock  = params.use_mlock;
-    lparams.logits_all = params.perplexity;
-    lparams.embedding  = params.embedding;
+    lparams.n_ctx        = params.n_ctx;
+    lparams.n_gpu_layers = params.n_gpu_layers;
+    lparams.seed         = params.seed;
+    lparams.f16_kv       = params.memory_f16;
+    lparams.use_mmap     = params.use_mmap;
+    lparams.use_mlock    = params.use_mlock;
+    lparams.logits_all   = params.perplexity;
+    lparams.embedding    = params.embedding;
 
     llama_context * lctx = llama_init_from_file(params.model.c_str(), lparams);
 
@@ -576,6 +578,37 @@ void console_set_color(console_state & con_st, console_color_t color) {
 }
 
 char32_t getchar32() {
+#if defined(_WIN32)
+    HANDLE hConsole = GetStdHandle(STD_INPUT_HANDLE);
+    wchar_t high_surrogate = 0;
+
+    while (true) {
+        INPUT_RECORD record;
+        DWORD count;
+        if (!ReadConsoleInputW(hConsole, &record, 1, &count) || count == 0) {
+            return WEOF;
+        }
+
+        if (record.EventType == KEY_EVENT && record.Event.KeyEvent.bKeyDown) {
+            wchar_t wc = record.Event.KeyEvent.uChar.UnicodeChar;
+            if (wc == 0) {
+                continue;
+            }
+
+            if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
+                high_surrogate = wc;
+                continue;
+            } else if ((wc >= 0xDC00) && (wc <= 0xDFFF)) { // Check if wc is a low surrogate
+                if (high_surrogate != 0) { // Check if we have a high surrogate
+                    return ((high_surrogate - 0xD800) << 10) + (wc - 0xDC00) + 0x10000;
+                }
+            }
+
+            high_surrogate = 0; // Reset the high surrogate
+            return static_cast<char32_t>(wc);
+        }
+    }
+#else
     wchar_t wc = getwchar();
     if (static_cast<wint_t>(wc) == WEOF) {
         return WEOF;
@@ -594,6 +627,7 @@ char32_t getchar32() {
 #endif
 
     return static_cast<char32_t>(wc);
+#endif
 }
 
 void pop_cursor(console_state & con_st) {
@@ -747,7 +781,7 @@ bool console_readline(console_state & con_st, std::string & line) {
             break;
         }
 
-        if (input_char == WEOF || input_char == 0x04 /* Ctrl+D*/) {
+        if (input_char == (char32_t) WEOF || input_char == 0x04 /* Ctrl+D*/) {
             end_of_stream = true;
             break;
         }
@@ -762,7 +796,7 @@ bool console_readline(console_state & con_st, std::string & line) {
             char32_t code = getchar32();
             if (code == '[' || code == 0x1B) {
                 // Discard the rest of the escape sequence
-                while ((code = getchar32()) != WEOF) {
+                while ((code = getchar32()) != (char32_t) WEOF) {
                     if ((code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z') || code == '~') {
                         break;
                     }

examples/common.h

@@ -21,13 +21,13 @@
 int32_t get_num_physical_cores();
 
 struct gpt_params {
-    int32_t seed          = -1;   // RNG seed
+    int32_t seed          = -1;  // RNG seed
     int32_t n_threads     = get_num_physical_cores();
     int32_t n_predict     = -1;  // new tokens to predict
-    int32_t n_parts       = -1;   // amount of model parts (-1 = determine from model dimensions)
-    int32_t n_ctx         = 512;  // context size
-    int32_t n_batch       = 512;  // batch size for prompt processing (must be >=32 to use BLAS)
-    int32_t n_keep        = 0;    // number of tokens to keep from initial prompt
+    int32_t n_ctx         = 512; // context size
+    int32_t n_batch       = 512; // batch size for prompt processing (must be >=32 to use BLAS)
+    int32_t n_keep        = 0;   // number of tokens to keep from initial prompt
+    int32_t n_gpu_layers  = 0;   // number of layers to store in VRAM
 
     // sampling parameters
     std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens
@@ -44,15 +44,15 @@ struct gpt_params {
     float   mirostat_tau      = 5.00f; // target entropy
     float   mirostat_eta      = 0.10f; // learning rate
 
-    std::string model  = "models/lamma-7B/ggml-model.bin"; // model path
-    std::string prompt = "";
+    std::string model             = "models/7B/ggml-model.bin"; // model path
+    std::string prompt            = "";
     std::string path_prompt_cache = "";  // path to file for saving/loading prompt eval state
     std::string input_prefix      = "";  // string to prefix user inputs with
     std::string input_suffix      = "";  // string to suffix user inputs with
     std::vector<std::string> antiprompt; // string upon seeing which more user input is prompted
 
     std::string lora_adapter = "";  // lora adapter path
-    std::string lora_base = "";     // base model path for the lora adapter
+    std::string lora_base    = "";  // base model path for the lora adapter
 
     bool memory_f16        = true;  // use f16 instead of f32 for memory kv
     bool random_prompt     = false; // do not randomize prompt if none provided

examples/embedding/embedding.cpp

@@ -6,7 +6,6 @@
 
 int main(int argc, char ** argv) {
     gpt_params params;
-    params.model = "models/llama-7B/ggml-model.bin";
 
     if (gpt_params_parse(argc, argv, params) == false) {
         return 1;
@@ -32,6 +31,8 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
+    llama_init_backend();
+
     llama_context * ctx;
 
     // load the model

examples/main/main.cpp

@@ -50,7 +50,6 @@ void sigint_handler(int signo) {
 
 int main(int argc, char ** argv) {
     gpt_params params;
-    params.model = "models/llama-7B/ggml-model.bin";
 
     if (gpt_params_parse(argc, argv, params) == false) {
         return 1;
@@ -97,8 +96,7 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
-//    params.prompt = R"(// this function checks if the number n is prime
-//bool is_prime(int n) {)";
+    llama_init_backend();
 
     llama_context * ctx;
     g_ctx = &ctx;
@@ -209,8 +207,8 @@ int main(int argc, char ** argv) {
         params.antiprompt.push_back("### Instruction:\n\n");
     }
 
-    // enable interactive mode if reverse prompt or interactive start is specified
-    if (params.antiprompt.size() != 0 || params.interactive_first) {
+    // enable interactive mode if interactive start is specified
+    if (params.interactive_first) {
         params.interactive = true;
     }
 
@@ -242,7 +240,7 @@ int main(int argc, char ** argv) {
         sigint_action.sa_flags = 0;
         sigaction(SIGINT, &sigint_action, NULL);
 #elif defined (_WIN32)
-        auto console_ctrl_handler = [](DWORD ctrl_type) -> BOOL {
+        auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
             return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
         };
         SetConsoleCtrlHandler(static_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
@@ -306,7 +304,7 @@ int main(int argc, char ** argv) {
 
     std::vector<llama_token> embd;
 
-    while (n_remain != 0 || params.interactive) {
+    while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
         // predict
         if (embd.size() > 0) {
             // infinite text generation via context swapping
@@ -504,9 +502,8 @@ int main(int argc, char ** argv) {
             console_set_color(con_st, CONSOLE_COLOR_DEFAULT);
         }
 
-        // in interactive mode, and not currently processing queued inputs;
-        // check if we should prompt the user for more
-        if (params.interactive && (int) embd_inp.size() <= n_consumed) {
+        // if not currently processing queued inputs;
+        if ((int) embd_inp.size() <= n_consumed) {
 
             // check for reverse prompt
             if (params.antiprompt.size()) {
@@ -517,10 +514,21 @@ int main(int argc, char ** argv) {
 
                 is_antiprompt = false;
                 // Check if each of the reverse prompts appears at the end of the output.
+                // If we're not running interactively, the reverse prompt might be tokenized with some following characters
+                // so we'll compensate for that by widening the search window a bit.
                 for (std::string & antiprompt : params.antiprompt) {
-                    if (last_output.find(antiprompt.c_str(), last_output.length() - antiprompt.length(), antiprompt.length()) != std::string::npos) {
-                        is_interacting = true;
+                    size_t extra_padding = params.interactive ? 0 : 2;
+                    size_t search_start_pos = last_output.length() > static_cast<size_t>(antiprompt.length() + extra_padding)
+                        ? last_output.length() - static_cast<size_t>(antiprompt.length() + extra_padding)
+                        : 0;
+
+                    if (last_output.find(antiprompt.c_str(), search_start_pos) != std::string::npos) {
+                        if (params.interactive) {
+                            is_interacting = true;
+                            console_set_color(con_st, CONSOLE_COLOR_USER_INPUT);
+                        }
                         is_antiprompt = true;
+                        fflush(stdout);
                         break;
                     }
                 }

examples/perplexity/perplexity.cpp

@@ -116,7 +116,6 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
 
 int main(int argc, char ** argv) {
     gpt_params params;
-    params.model = "models/llama-7B/ggml-model.bin";
 
     params.n_batch = 512;
     if (gpt_params_parse(argc, argv, params) == false) {
@@ -144,6 +143,8 @@ int main(int argc, char ** argv) {
         params.prompt = gpt_random_prompt(rng);
     }
 
+    llama_init_backend();
+
     llama_context * ctx;
 
     // load the model and apply lora adapter, if any

examples/quantize-stats/quantize-stats.cpp

@@ -321,7 +321,6 @@ int main(int argc, char ** argv) {
         auto lparams = llama_context_default_params();
 
         lparams.n_ctx      = 256;
-        lparams.n_parts    = 1;
         lparams.seed       = 1;
         lparams.f16_kv     = false;
         lparams.use_mlock  = false;

examples/quantize/quantize.cpp

@@ -1,7 +1,7 @@
-#include "ggml.h"
-#include "llama.h"
 #include "build-info.h"
 
+#include "llama.h"
+
 #include <cstdio>
 #include <map>
 #include <string>
@@ -42,8 +42,6 @@ bool try_parse_ftype(const std::string & ftype_str, llama_ftype & ftype, std::st
 //  ./quantize models/llama/ggml-model.bin [models/llama/ggml-model-quant.bin] type [nthreads]
 //
 int main(int argc, char ** argv) {
-    ggml_time_init();
-
     if (argc < 3) {
         fprintf(stderr, "usage: %s model-f32.bin [model-quant.bin] type [nthreads]\n", argv[0]);
         for (auto it = LLAMA_FTYPE_MAP.begin(); it != LLAMA_FTYPE_MAP.end(); it++) {
@@ -52,12 +50,7 @@ int main(int argc, char ** argv) {
         return 1;
     }
 
-    // needed to initialize f16 tables
-    {
-        struct ggml_init_params params = { 0, NULL, false };
-        struct ggml_context * ctx = ggml_init(params);
-        ggml_free(ctx);
-    }
+    llama_init_backend();
 
     // parse command line arguments
     const std::string fname_inp = argv[1];
@@ -116,25 +109,25 @@ int main(int argc, char ** argv) {
     }
     fprintf(stderr, "\n");
 
-    const int64_t t_main_start_us = ggml_time_us();
+    const int64_t t_main_start_us = llama_time_us();
 
     int64_t t_quantize_us = 0;
 
     // load the model
     {
-        const int64_t t_start_us = ggml_time_us();
+        const int64_t t_start_us = llama_time_us();
 
         if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), ftype, nthread)) {
             fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
             return 1;
         }
 
-        t_quantize_us = ggml_time_us() - t_start_us;
+        t_quantize_us = llama_time_us() - t_start_us;
     }
 
     // report timing
     {
-        const int64_t t_main_end_us = ggml_time_us();
+        const int64_t t_main_end_us = llama_time_us();
 
         printf("\n");
         printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0);

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

@@ -8,7 +8,6 @@
 
 int main(int argc, char ** argv) {
     gpt_params params;
-    params.model = "models/llama-7B/ggml-model.bin";
     params.seed = 42;
     params.n_threads = 4;
     params.repeat_last_n = 64;
@@ -27,7 +26,6 @@ int main(int argc, char ** argv) {
     auto lparams = llama_context_default_params();
 
     lparams.n_ctx     = params.n_ctx;
-    lparams.n_parts   = params.n_parts;
     lparams.seed      = params.seed;
     lparams.f16_kv    = params.memory_f16;
     lparams.use_mmap  = params.use_mmap;

ggml-cuda.cu

@@ -32,24 +32,32 @@ static_assert(sizeof(half) == sizeof(ggml_fp16_t), "wrong fp16 size");
         }                                                                               \
     } while (0)
 
+typedef void (*dequantize_kernel_t)(const void * vx, const int ib, const int iqs, float & v0, float & v1);
 typedef void (*to_fp32_cuda_t)(const void * x, float * y, int k, cudaStream_t stream);
+typedef void (*dequantize_mul_mat_vec_cuda_t)(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream);
+
+// QK = number of values after dequantization
+// QR = QK / number of values before dequantization
 
 #define QK4_0 32
+#define QR4_0 2
 typedef struct {
-    float   d;              // delta
+    half    d;              // delta
     uint8_t qs[QK4_0 / 2];  // nibbles / quants
 } block_q4_0;
-static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
+static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
 
 #define QK4_1 32
+#define QR4_1 2
 typedef struct {
-    float   d;              // delta
-    float   m;              // min
+    half    d;              // delta
+    half    m;              // min
     uint8_t qs[QK4_1 / 2];  // nibbles / quants
 } block_q4_1;
-static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
+static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
 
 #define QK5_0 32
+#define QR5_0 2
 typedef struct {
     half d;                 // delta
     uint8_t qh[4];          // 5-th bit of quants
@@ -58,6 +66,7 @@ typedef struct {
 static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
 
 #define QK5_1 32
+#define QR5_1 2
 typedef struct {
     half d;                 // delta
     half m;                 // min
@@ -67,148 +76,237 @@ typedef struct {
 static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
 
 #define QK8_0 32
+#define QR8_0 1
 typedef struct {
-    float   d;              // delta
+    half    d;              // delta
     int8_t  qs[QK8_0];      // quants
 } block_q8_0;
-static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
+static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
 
-static __global__ void dequantize_block_q4_0(const void * vx, float * y) {
-    static const int qk = QK4_0;
+#define CUDA_MUL_BLOCK_SIZE 256
+#define CUDA_DEQUANTIZE_BLOCK_SIZE 256
+#define CUDA_DMMV_BLOCK_SIZE 32 // dmmv = dequantize_mul_mat_vec
 
+static __global__ void mul_f32(const float * x, const float * y, float * dst, const int kx, const int ky) {
+    const int i = blockDim.x*blockIdx.x + threadIdx.x;
+
+    if (i >= kx) {
+        return;
+    }
+    dst[i] = x[i] * y[i%ky];
+}
+
+static __device__ void dequantize_q4_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){
     const block_q4_0 * x = (const block_q4_0 *) vx;
 
-    const int i = blockIdx.x;
+    const float d = x[ib].d;
 
-    const float d = x[i].d;
+    const uint8_t vui = x[ib].qs[iqs];
 
-    for (int j = 0; j < qk/2; ++j) {
-        const int x0 = (x[i].qs[j] & 0xf) - 8;
-        const int x1 = (x[i].qs[j] >>  4) - 8;
+    const int8_t vi0 = vui & 0xF;
+    const int8_t vi1 = vui >> 4;
 
-        y[i*qk + j + 0   ] = x0*d;
-        y[i*qk + j + qk/2] = x1*d;
-    }
+    v0 = (vi0 - 8)*d;
+    v1 = (vi1 - 8)*d;
 }
 
-static __global__ void dequantize_block_q4_1(const void * vx, float * y) {
-    static const int qk = QK4_1;
-
+static __device__ void dequantize_q4_1(const void * vx, const int ib, const int iqs, float & v0, float & v1){
     const block_q4_1 * x = (const block_q4_1 *) vx;
 
-    const int i = blockIdx.x;
+    const float d = x[ib].d;
+    const float m = x[ib].m;
 
-    const float d = x[i].d;
-    const float m = x[i].m;
+    const uint8_t vui = x[ib].qs[iqs];
 
-    for (int j = 0; j < qk/2; ++j) {
-        const int x0 = (x[i].qs[j] & 0xf);
-        const int x1 = (x[i].qs[j] >>  4);
+    const int8_t vi0 = vui & 0xF;
+    const int8_t vi1 = vui >> 4;
 
-        y[i*qk + j + 0   ] = x0*d + m;
-        y[i*qk + j + qk/2] = x1*d + m;
-    }
+    v0 = vi0*d + m;
+    v1 = vi1*d + m;
 }
 
-static __global__ void dequantize_block_q5_0(const void * vx, float * y) {
-    static const int qk = QK5_0;
-
+static __device__ void dequantize_q5_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){
     const block_q5_0 * x = (const block_q5_0 *) vx;
 
-    const int i = blockIdx.x;
-
-    const float d = x[i].d;
+    const float d = x[ib].d;
 
     uint32_t qh;
-    memcpy(&qh, x[i].qh, sizeof(qh));
+    memcpy(&qh, x[ib].qh, sizeof(qh));
 
-    for (int j = 0; j < qk/2; ++j) {
-        const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-        const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+    const uint8_t xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
+    const uint8_t xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
 
-        const int32_t x0 = ((x[i].qs[j] & 0xf) | xh_0) - 16;
-        const int32_t x1 = ((x[i].qs[j] >>  4) | xh_1) - 16;
+    const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0) - 16;
+    const int32_t x1 = ((x[ib].qs[iqs] >>  4) | xh_1) - 16;
 
-        y[i*qk + j + 0   ] = x0*d;
-        y[i*qk + j + qk/2] = x1*d;
-    }
+    v0 = x0*d;
+    v1 = x1*d;
 }
 
-static __global__ void dequantize_block_q5_1(const void * vx, float * y) {
-    static const int qk = QK5_1;
-
+static __device__ void dequantize_q5_1(const void * vx, const int ib, const int iqs, float & v0, float & v1){
     const block_q5_1 * x = (const block_q5_1 *) vx;
 
-    const int i = blockIdx.x;
-
-    const float d = x[i].d;
-    const float m = x[i].m;
+    const float d = x[ib].d;
+    const float m = x[ib].m;
 
     uint32_t qh;
-    memcpy(&qh, x[i].qh, sizeof(qh));
+    memcpy(&qh, x[ib].qh, sizeof(qh));
 
-    for (int j = 0; j < qk/2; ++j) {
-        const uint8_t xh_0 = ((qh >> (j +  0)) << 4) & 0x10;
-        const uint8_t xh_1 = ((qh >> (j + 12))     ) & 0x10;
+    const uint8_t xh_0 = ((qh >> (iqs +  0)) << 4) & 0x10;
+    const uint8_t xh_1 = ((qh >> (iqs + 12))     ) & 0x10;
 
-        const int x0 = (x[i].qs[j] & 0xf) | xh_0;
-        const int x1 = (x[i].qs[j] >>  4) | xh_1;
+    const int32_t x0 = ((x[ib].qs[iqs] & 0xf) | xh_0);
+    const int32_t x1 = ((x[ib].qs[iqs] >>  4) | xh_1);
 
-        y[i*qk + j + 0   ] = x0*d + m;
-        y[i*qk + j + qk/2] = x1*d + m;
-    }
+    v0 = x0*d + m;
+    v1 = x1*d + m;
 }
 
-static __global__ void dequantize_block_q8_0(const void * vx, float * y) {
-    static const int qk = QK8_0;
-
+static __device__ void dequantize_q8_0(const void * vx, const int ib, const int iqs, float & v0, float & v1){
     const block_q8_0 * x = (const block_q8_0 *) vx;
 
-    const int i = blockIdx.x;
+    const float d = x[ib].d;
 
-    const float d = x[i].d;
+    const int8_t vi0 = x[ib].qs[iqs + 0];
+    const int8_t vi1 = x[ib].qs[iqs + 1];
 
-    for (int j = 0; j < qk; ++j) {
-        y[i*qk + j] = x[i].qs[j]*d;
+    v0 = vi0*d;
+    v1 = vi1*d;
+}
+
+static __device__ void convert_f16(const void * vx, const int ib, const int iqs, float & v0, float & v1){
+    const half * x = (const half *) vx;
+
+    v0 = __half2float(x[ib + 0]);
+    v1 = __half2float(x[ib + 1]);
+}
+
+template <int qk, int qr, dequantize_kernel_t dequantize_kernel>
+static __global__ void dequantize_block(const void * vx, float * y, const int k) {
+    const int i = blockDim.x*blockIdx.x + 2*threadIdx.x;
+
+    if (i >= k) {
+        return;
+    }
+
+    const int ib = i/qk; // block index
+    const int iqs = (i%qk)/qr; // quant index
+    const int iybs = i - i%qk; // y block start index
+    const int y_offset = qr == 1 ? 1 : qk/2;
+
+    // dequantize
+    float & v0 = y[iybs + iqs + 0];
+    float & v1 = y[iybs + iqs + y_offset];
+    dequantize_kernel(vx, ib, iqs, v0, v1);
+}
+
+template <int block_size, int qk, int qr, dequantize_kernel_t dequantize_kernel>
+static __global__ void dequantize_mul_mat_vec(const void * vx, const float * y, float * dst, const int ncols) {
+    const int row = blockIdx.x;
+    const int tid = threadIdx.x;
+
+    const int y_offset = qr == 1 ? 1 : qk/2;
+
+    __shared__ float tmp[block_size]; // separate sum for each thread
+    tmp[tid] = 0;
+
+    for (int i = 0; i < ncols/block_size; i += 2) {
+        const int col = i*block_size + 2*tid;
+        const int ib = (row*ncols + col)/qk; // block index
+        const int iqs = (col%qk)/qr; // quant index
+        const int iybs = col - col%qk; // y block start index
+
+        // dequantize
+        float v0, v1;
+        dequantize_kernel(vx, ib, iqs, v0, v1);
+
+        // matrix multiplication
+        tmp[tid] += v0 * y[iybs + iqs + 0];
+        tmp[tid] += v1 * y[iybs + iqs + y_offset];
+    }
+
+    // sum up partial sums and write back result
+    __syncthreads();
+    for (int s=block_size/2; s>0; s>>=1) {
+        if (tid < s) {
+            tmp[tid] += tmp[tid + s];
+        }
+        __syncthreads();
+    }
+    if (tid == 0) {
+        dst[row] = tmp[0];
     }
 }
 
-static void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
-    const int nb = k / QK4_0;
-    dequantize_block_q4_0<<<nb, 1, 0, stream>>>(vx, y);
+static void mul_f32_cuda(const float * x, const float * y, float * dst, const int kx, const int ky, cudaStream_t stream) {
+    const int num_blocks = (kx + CUDA_MUL_BLOCK_SIZE - 1) / CUDA_MUL_BLOCK_SIZE;
+    mul_f32<<<num_blocks, CUDA_MUL_BLOCK_SIZE, 0, stream>>>(x, y, dst, kx, ky);
 }
 
-static void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
-    const int nb = k / QK4_1;
-    dequantize_block_q4_1<<<nb, 1, 0, stream>>>(vx, y);
+static void dequantize_row_q4_0_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
+    dequantize_block<QK4_0, QR4_0, dequantize_q4_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
 }
 
-static void dequantize_row_q5_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
-    const int nb = k / QK5_0;
-    dequantize_block_q5_0<<<nb, 1, 0, stream>>>(vx, y);
+static void dequantize_row_q4_1_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
+    dequantize_block<QK4_1, QR4_1, dequantize_q4_1><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
 }
 
-static void dequantize_row_q5_1_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
-    const int nb = k / QK5_1;
-    dequantize_block_q5_1<<<nb, 1, 0, stream>>>(vx, y);
+static void dequantize_row_q5_0_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
+    dequantize_block<QK5_0, QR5_0, dequantize_q5_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
 }
 
-static void dequantize_row_q8_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
-    const int nb = k / QK8_0;
-    dequantize_block_q8_0<<<nb, 1, 0, stream>>>(vx, y);
+static void dequantize_row_q5_1_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
+    dequantize_block<QK5_1, QR5_1, dequantize_q5_1><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
 }
 
-// TODO: optimize
-static __global__ void convert_fp16_to_fp32(const void * vx, float * y) {
-    const half * x = (const half *) vx;
-
-    const int i = blockIdx.x;
-
-    y[i] = __half2float(x[i]);
+static void dequantize_row_q8_0_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
+    dequantize_block<QK8_0, QR8_0, dequantize_q8_0><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
 }
 
-static void convert_fp16_to_fp32_cuda(const void * x, float * y, int k, cudaStream_t stream) {
-    convert_fp16_to_fp32<<<k, 1, 0, stream>>>(x, y);
+static void dequantize_mul_mat_vec_q4_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
+    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, QK4_0, QR4_0, dequantize_q4_0>
+        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+}
+
+static void dequantize_mul_mat_vec_q4_1_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
+    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, QK4_1, QR4_1, dequantize_q4_1>
+        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+}
+
+static void dequantize_mul_mat_vec_q5_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
+    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, QK5_0, QR5_0, dequantize_q5_0>
+        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+}
+
+static void dequantize_mul_mat_vec_q5_1_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
+    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, QK5_1, QR5_1, dequantize_q5_1>
+        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+}
+
+static void dequantize_mul_mat_vec_q8_0_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
+    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, QK8_0, QR8_0, dequantize_q8_0>
+        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
+}
+
+static void convert_fp16_to_fp32_cuda(const void * vx, float * y, const int k, cudaStream_t stream) {
+    const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
+    dequantize_block<32, 1, convert_f16><<<num_blocks, CUDA_DEQUANTIZE_BLOCK_SIZE, 0, stream>>>(vx, y, k);
+}
+
+static void convert_mul_mat_vec_f16_cuda(const void * vx, const float * y, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % CUDA_DMMV_BLOCK_SIZE == 0);
+    dequantize_mul_mat_vec<CUDA_DMMV_BLOCK_SIZE, 32, 1, convert_f16>
+        <<<nrows, CUDA_DMMV_BLOCK_SIZE, 0, stream>>>(vx, y, dst, ncols);
 }
 
 static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
@@ -230,8 +328,27 @@ static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
     }
 }
 
+static dequantize_mul_mat_vec_cuda_t ggml_get_dequantize_mul_mat_vec_cuda(ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            return dequantize_mul_mat_vec_q4_0_cuda;
+        case GGML_TYPE_Q4_1:
+            return dequantize_mul_mat_vec_q4_1_cuda;
+        case GGML_TYPE_Q5_0:
+            return dequantize_mul_mat_vec_q5_0_cuda;
+        case GGML_TYPE_Q5_1:
+            return dequantize_mul_mat_vec_q5_1_cuda;
+        case GGML_TYPE_Q8_0:
+            return dequantize_mul_mat_vec_q8_0_cuda;
+        case GGML_TYPE_F16:
+            return convert_mul_mat_vec_f16_cuda;
+        default:
+            return nullptr;
+    }
+}
+
 // buffer pool for cuda
-#define MAX_CUDA_BUFFERS 16
+#define MAX_CUDA_BUFFERS 256
 
 struct scoped_spin_lock {
     std::atomic_flag& lock;
@@ -365,6 +482,67 @@ static cudaError_t ggml_cuda_h2d_tensor_2d(void * dst, const struct ggml_tensor
     }
 }
 
+static void ggml_cuda_mul_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
+    GGML_ASSERT(src1->backend == GGML_BACKEND_CUDA);
+    const int64_t ne00 = src0->ne[0];
+    const int64_t ne01 = src0->ne[1];
+    const int64_t ne02 = src0->ne[2];
+    const int64_t ne03 = src0->ne[2];
+    const int64_t ne0 = ne00 * ne01 * ne02 * ne03;
+    const int64_t ne10 = src1->ne[0];
+    const int64_t ne11 = src1->ne[1];
+    const int64_t ne12 = src1->ne[2];
+    const int64_t ne13 = src1->ne[3];
+    const int nb2  = dst->nb[2];
+    const int nb3  = dst->nb[3];
+    size_t x_size, d_size;
+
+    float * d_X = (float *) ggml_cuda_pool_malloc(ne0 * sizeof(float), &x_size); // src0
+    float * d_Y = (float *) src1->data; // src1 is already on device, broadcasted.
+    float * d_D = (float *) ggml_cuda_pool_malloc(ne0 * sizeof(float), &d_size); // dst
+
+    for (int64_t i03 = 0; i03 < ne03; i03++) {
+        for (int64_t i02 = 0; i02 < ne02; i02++) {
+            const int i0 = i03*ne02 + i02;
+            float * c_X2 = d_X + i0*ne01*ne00;
+            float * c_D2 = d_D + i0*ne01*ne00;
+
+            cudaStream_t cudaStream = g_cudaStreams[i0 % GGML_CUDA_MAX_STREAMS];
+            cudaStream_t cudaStream2 = g_cudaStreams2[i0 % GGML_CUDA_MAX_STREAMS];
+            cudaEvent_t  cudaEvent = g_cudaEvents[i0 % GGML_CUDA_MAX_EVENTS];
+
+            // copy src0 to device
+            CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_X2, src0, i03, i02, cudaStream2));
+            CUDA_CHECK(cudaEventRecord(cudaEvent, cudaStream2));
+
+            // wait for data
+            CUDA_CHECK(cudaStreamWaitEvent(cudaStream, cudaEvent, 0));
+
+            for (int64_t i01 = 0; i01 < ne01; i01++) {
+                const int64_t i13 = i03%ne13;
+                const int64_t i12 = i02%ne12;
+                const int64_t i11 = i01%ne11;
+                const int i1 = i13*ne12*ne11 + i12*ne11 + i11;
+
+                float * c_X1 = c_X2 + i01*ne00;
+                float * c_Y = d_Y + i1*ne10;
+                float * c_D1 = c_D2 + i01*ne00;
+
+                // compute
+                mul_f32_cuda(c_X1, c_Y, c_D1, ne00, ne10, cudaStream);
+                CUDA_CHECK(cudaGetLastError());
+            }
+
+            // copy dst to host
+            float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
+            CUDA_CHECK(cudaMemcpyAsync(d, c_D2, sizeof(float)*ne00*ne01, cudaMemcpyDeviceToHost, cudaStream));
+        }
+    }
+    CUDA_CHECK(cudaDeviceSynchronize());
+    ggml_cuda_pool_free(d_X, x_size);
+    ggml_cuda_pool_free(d_D, d_size);
+}
+
 static void ggml_cuda_mul_mat_f32(const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst) {
     const int64_t ne00 = src0->ne[0];
     const int64_t ne01 = src0->ne[1];
@@ -528,6 +706,7 @@ static void ggml_cuda_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor
     const int nb2  = dst->nb[2];
     const int nb3  = dst->nb[3];
     const ggml_type type = src0->type;
+    const bool mul_mat_vec = ne11 == 1;
 
     const float alpha = 1.0f;
     const float beta = 0.0f;
@@ -538,12 +717,16 @@ static void ggml_cuda_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor
     const size_t q_sz = ggml_type_size(type) * x_ne / ggml_blck_size(type);
 
     size_t x_size, y_size, d_size, q_size;
-    float * d_X = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * x_ne, &x_size);
+    float * d_X = nullptr;
+    if (!mul_mat_vec) {
+        d_X = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * x_ne, &x_size);
+    }
     float * d_Y = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * y_ne, &y_size);
     float * d_D = (float *) ggml_cuda_pool_malloc(n_mm * sizeof(float) * d_ne, &d_size);
     char  * d_Q = (char  *) ggml_cuda_pool_malloc(n_mm * q_sz, &q_size);
 
     const to_fp32_cuda_t to_fp32_cuda = ggml_get_to_fp32_cuda(type);
+    dequantize_mul_mat_vec_cuda_t dmmv = ggml_get_dequantize_mul_mat_vec_cuda(type);
     GGML_ASSERT(to_fp32_cuda != nullptr);
 
     for (int64_t i03 = 0; i03 < ne03; i03++) {
@@ -553,31 +736,54 @@ static void ggml_cuda_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor
             cudaStream_t cudaStream2 = g_cudaStreams2[i % GGML_CUDA_MAX_STREAMS];
             cudaEvent_t  cudaEvent = g_cudaEvents[i % GGML_CUDA_MAX_EVENTS];
 
-            float * c_X = d_X + i * x_ne;
             float * c_Y = d_Y + i * y_ne;
             float * c_D = d_D + i * d_ne;
             char  * c_Q = d_Q + i * q_sz;
 
-            // copy src0 and convert to fp32 on device
-            CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Q, src0, i03, i02, cudaStream2));
-            to_fp32_cuda(c_Q, c_X, x_ne, cudaStream2);
-            CUDA_CHECK(cudaGetLastError());
-            CUDA_CHECK(cudaEventRecord(cudaEvent, cudaStream2));
+            // copy src0 to device if necessary
+            if (src0->backend == GGML_BACKEND_CPU) {
+                CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Q, src0, i03, i02, cudaStream2));
+            } else if (src0->backend == GGML_BACKEND_CUDA) {
+                c_Q = ((char *) src0->data) + i * q_sz;
+            } else {
+                GGML_ASSERT(false);
+            }
+            if (mul_mat_vec) { // specialized dequantize_mul_mat_vec kernel
+                CUDA_CHECK(cudaEventRecord(cudaEvent, cudaStream2));
 
-            // copy src1 to device
-            CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Y, src1, i03, i02, cudaStream));
+                // copy src1 to device
+                CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Y, src1, i03, i02, cudaStream));
 
-            // wait for conversion
-            CUDA_CHECK(cudaStreamWaitEvent(cudaStream, cudaEvent, 0));
+                // wait for data
+                CUDA_CHECK(cudaStreamWaitEvent(cudaStream, cudaEvent, 0));
 
-            // compute
-            CUBLAS_CHECK(cublasSetStream(g_cublasH, cudaStream));
-            CUBLAS_CHECK(
-                cublasSgemm(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N,
-                        ne01, ne11, ne10,
-                        &alpha, c_X, ne00,
-                                c_Y, ne10,
-                        &beta,  c_D, ne01));
+                // compute
+                dmmv(c_Q, c_Y, c_D, ne00, ne01, cudaStream);
+                CUDA_CHECK(cudaGetLastError());
+
+            } else { // general dequantization kernel + cuBLAS matrix matrix multiplication
+                float * c_X = d_X + i * x_ne;
+
+                // convert src0 to fp32 on device
+                to_fp32_cuda(c_Q, c_X, x_ne, cudaStream2);
+                CUDA_CHECK(cudaGetLastError());
+                CUDA_CHECK(cudaEventRecord(cudaEvent, cudaStream2));
+
+                // copy src1 to device
+                CUDA_CHECK(ggml_cuda_h2d_tensor_2d(c_Y, src1, i03, i02, cudaStream));
+
+                // wait for conversion
+                CUDA_CHECK(cudaStreamWaitEvent(cudaStream, cudaEvent, 0));
+
+                // compute
+                CUBLAS_CHECK(cublasSetStream(g_cublasH, cudaStream));
+                CUBLAS_CHECK(
+                    cublasSgemm(g_cublasH, CUBLAS_OP_T, CUBLAS_OP_N,
+                            ne01, ne11, ne10,
+                            &alpha, c_X, ne00,
+                                    c_Y, ne10,
+                            &beta,  c_D, ne01));
+            }
 
             // copy dst to host
             float * d = (float *) ((char *) dst->data + i02*nb2 + i03*nb3);
@@ -586,12 +792,19 @@ static void ggml_cuda_mul_mat_q_f32(const ggml_tensor * src0, const ggml_tensor
     }
 
     CUDA_CHECK(cudaDeviceSynchronize());
-    ggml_cuda_pool_free(d_X, x_size);
+    if (!mul_mat_vec) {
+        ggml_cuda_pool_free(d_X, x_size);
+    }
     ggml_cuda_pool_free(d_Y, y_size);
     ggml_cuda_pool_free(d_D, d_size);
     ggml_cuda_pool_free(d_Q, q_size);
 }
 
+void ggml_cuda_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
+    GGML_ASSERT(src0->type == GGML_TYPE_F32 && src1->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32);
+    ggml_cuda_mul_f32(src0, src1, dst);
+}
+
 bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst) {
     const int64_t ne10 = src1->ne[0];
 
@@ -602,8 +815,7 @@ bool ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_te
     if ((src0->type == GGML_TYPE_F32 || src0->type == GGML_TYPE_F16 || ggml_is_quantized(src0->type)) &&
         src1->type == GGML_TYPE_F32 &&
         dst->type == GGML_TYPE_F32 &&
-        (ne0 >= 32 && ne1 >= 32 && ne10 >= 32)) {
-
+        ((ne0 >= 32 && ne1 >= 32 && ne10 >= 32) || src0->backend == GGML_BACKEND_CUDA)) {
         return true;
     }
 
@@ -655,3 +867,59 @@ size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct
         return 0;
     }
 }
+
+void ggml_cuda_transform_tensor(ggml_tensor * tensor) {
+    const int64_t ne0 = tensor->ne[0];
+    const int64_t ne1 = tensor->ne[1];
+    const int64_t ne2 = tensor->ne[2];
+    const int64_t ne3 = tensor->ne[3];
+
+    const ggml_type type = tensor->type;
+    const size_t q_sz = ggml_type_size(type) * ne0 * ne1 * ne2 * ne3 / ggml_blck_size(type);
+
+    size_t q_size;
+    char * dst = (char *) ggml_cuda_pool_malloc(q_sz, &q_size);
+
+    cudaStream_t cudaStream2 = g_cudaStreams2[0];
+
+    // copy tensor to device
+    for (int64_t i3 = 0; i3 < ne3; i3++) {
+        for (int64_t i2 = 0; i2 < ne2; i2++) {
+            int i = i3*ne2 + i2;
+            CUDA_CHECK(ggml_cuda_h2d_tensor_2d(dst + i*ne0*ne1, tensor, i3, i2, cudaStream2));
+        }
+    }
+
+    tensor->data = dst;
+    tensor->backend = GGML_BACKEND_CUDA;
+}
+
+void ggml_cuda_load_data(const char * fname, struct ggml_tensor * tensor, const size_t offset) {
+    FILE * fp = fopen(fname, "rb");
+
+    const size_t size = ggml_nbytes(tensor);
+
+    void * buf;
+    CUDA_CHECK(cudaMalloc(&buf, size));
+    void * buf_host = malloc(size);
+
+#ifdef _WIN32
+    int ret = _fseeki64(fp, (__int64) offset, SEEK_SET);
+#else
+    int ret = fseek(fp, (long) offset, SEEK_SET);
+#endif
+    GGML_ASSERT(ret == 0); // same
+
+    size_t ret2 = fread(buf_host, size, 1, fp);
+    if (ret2 != 1) {
+        fprintf(stderr, "unexpectedly reached end of file");
+        exit(1);
+    }
+
+    cudaMemcpy(buf, buf_host, size, cudaMemcpyHostToDevice);
+    cudaDeviceSynchronize();
+
+    tensor->data = buf;
+    free(buf_host);
+    fclose(fp);
+}

ggml-cuda.h

@@ -6,6 +6,7 @@ extern "C" {
 
 void   ggml_init_cublas(void);
 
+void   ggml_cuda_mul(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
 bool   ggml_cuda_can_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
 size_t ggml_cuda_mul_mat_get_wsize(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst);
 void   ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tensor * src1, struct ggml_tensor * dst, void * wdata, size_t wsize);
@@ -14,6 +15,9 @@ void   ggml_cuda_mul_mat(const struct ggml_tensor * src0, const struct ggml_tens
 void * ggml_cuda_host_malloc(size_t size);
 void   ggml_cuda_host_free(void * ptr);
 
+void ggml_cuda_transform_tensor(struct ggml_tensor * tensor);
+void ggml_cuda_load_data(const char * fname, struct ggml_tensor * tensors, size_t offset);
+
 #ifdef  __cplusplus
 }
 #endif

ggml.h

@@ -190,9 +190,12 @@
 #define GGML_FILE_MAGIC   0x67676d6c // "ggml"
 #define GGML_FILE_VERSION 1
 
+#define GGML_QNT_VERSION        2    // bump this on quantization format changes
+#define GGML_QNT_VERSION_FACTOR 1000 // do not change this
+
 #define GGML_MAX_DIMS          4
 #define GGML_MAX_NODES         4096
-#define GGML_MAX_PARAMS        16
+#define GGML_MAX_PARAMS        256
 #define GGML_MAX_CONTEXTS      64
 #define GGML_MAX_OPT           4
 #define GGML_DEFAULT_N_THREADS 4
@@ -243,6 +246,11 @@ extern "C" {
         GGML_TYPE_COUNT,
     };
 
+    enum ggml_backend {
+        GGML_BACKEND_CPU = 0,
+        GGML_BACKEND_CUDA = 1,
+    };
+
     // model file types
     enum ggml_ftype {
         GGML_FTYPE_UNKNOWN     = -1,
@@ -262,12 +270,16 @@ extern "C" {
 
         GGML_OP_DUP,
         GGML_OP_ADD,
+        GGML_OP_ADD1,
+        GGML_OP_ACC,
         GGML_OP_SUB,
         GGML_OP_MUL,
         GGML_OP_DIV,
         GGML_OP_SQR,
         GGML_OP_SQRT,
+        GGML_OP_LOG,
         GGML_OP_SUM,
+        GGML_OP_SUM_ROWS,
         GGML_OP_MEAN,
         GGML_OP_REPEAT,
         GGML_OP_ABS,
@@ -277,12 +289,15 @@ extern "C" {
         GGML_OP_RELU,
         GGML_OP_GELU,
         GGML_OP_SILU,
+        GGML_OP_SILU_BACK,
         GGML_OP_NORM, // normalize
         GGML_OP_RMS_NORM,
+        GGML_OP_RMS_NORM_BACK,
 
         GGML_OP_MUL_MAT,
 
         GGML_OP_SCALE,
+        GGML_OP_SET,
         GGML_OP_CPY,
         GGML_OP_CONT,
         GGML_OP_RESHAPE,
@@ -290,10 +305,15 @@ extern "C" {
         GGML_OP_PERMUTE,
         GGML_OP_TRANSPOSE,
         GGML_OP_GET_ROWS,
+        GGML_OP_GET_ROWS_BACK,
+        GGML_OP_DIAG,
         GGML_OP_DIAG_MASK_INF,
+        GGML_OP_DIAG_MASK_ZERO,
         GGML_OP_SOFT_MAX,
         GGML_OP_ROPE,
+        GGML_OP_ROPE_BACK,
         GGML_OP_ALIBI,
+        GGML_OP_CLAMP,
         GGML_OP_CONV_1D_1S,
         GGML_OP_CONV_1D_2S,
 
@@ -321,7 +341,8 @@ extern "C" {
 
     // n-dimensional tensor
     struct ggml_tensor {
-        enum ggml_type type;
+        enum ggml_type    type;
+        enum ggml_backend backend;
 
         int     n_dims;
         int64_t ne[GGML_MAX_DIMS]; // number of elements
@@ -352,7 +373,7 @@ extern "C" {
 
         char name[32];
 
-        char padding[8]; // TODO: remove and add padding to name?
+        char padding[16];
     };
 
     // computation graph
@@ -496,6 +517,29 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_add1(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    GGML_API struct ggml_tensor * ggml_acc(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                nb2,
+            size_t                nb3,
+            size_t                offset);
+
+    GGML_API struct ggml_tensor * ggml_acc_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                nb2,
+            size_t                nb3,
+            size_t                offset);
+
     GGML_API struct ggml_tensor * ggml_sub(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -519,12 +563,24 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    GGML_API struct ggml_tensor * ggml_log(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    GGML_API struct ggml_tensor * ggml_log_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // return scalar
-    // TODO: compute sum along rows
     GGML_API struct ggml_tensor * ggml_sum(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // sums along rows, with input shape [a,b,c,d] return shape [1,b,c,d]
+    GGML_API struct ggml_tensor * ggml_sum_rows(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
     // mean along rows
     GGML_API struct ggml_tensor * ggml_mean(
             struct ggml_context * ctx,
@@ -566,6 +622,13 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // a - x
+    // b - dy
+    GGML_API struct ggml_tensor * ggml_silu_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // normalize along rows
     // TODO: eps is hardcoded to 1e-5 for now
     GGML_API struct ggml_tensor * ggml_norm(
@@ -576,6 +639,13 @@ extern "C" {
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
+    // a - x
+    // b - dy
+    GGML_API struct ggml_tensor * ggml_rms_norm_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
     // A: m rows, n columns
     // B: p rows, n columns (i.e. we transpose it internally)
     // result is m columns, p rows
@@ -588,12 +658,66 @@ extern "C" {
     // operations on tensors without backpropagation
     //
 
-    // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_scale(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_scale_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b);
+
+    // b -> view(a,offset,nb1,nb2,3), return modified a
+    GGML_API struct ggml_tensor * ggml_set(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                nb2,
+            size_t                nb3,
+            size_t                offset);
+
+    // b -> view(a,offset,nb1,nb2,3), return view(a)
+    GGML_API struct ggml_tensor * ggml_set_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                nb2,
+            size_t                nb3,
+            size_t                offset);
+
+    GGML_API struct ggml_tensor * ggml_set_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                offset);
+
+    GGML_API struct ggml_tensor * ggml_set_1d_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                offset);
+
+    // b -> view(a,offset,nb1,nb2,3), return modified a
+    GGML_API struct ggml_tensor * ggml_set_2d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                offset);
+
+    // b -> view(a,offset,nb1,nb2,3), return view(a)
+    GGML_API struct ggml_tensor * ggml_set_2d_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            size_t                nb1,
+            size_t                offset);
+
+
     // a -> b, return view(b)
     GGML_API struct ggml_tensor * ggml_cpy(
             struct ggml_context * ctx,
@@ -614,6 +738,11 @@ extern "C" {
 
     // return view(a)
     // TODO: when we start computing gradient, make a copy instead of view
+    GGML_API struct ggml_tensor * ggml_reshape_1d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0);
+
     GGML_API struct ggml_tensor * ggml_reshape_2d(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -629,6 +758,14 @@ extern "C" {
             int64_t               ne1,
             int64_t               ne2);
 
+    GGML_API struct ggml_tensor * ggml_reshape_4d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3);
+
     // offset in bytes
     GGML_API struct ggml_tensor * ggml_view_1d(
             struct ggml_context * ctx,
@@ -654,6 +791,18 @@ extern "C" {
             size_t                nb2, // slice stride in bytes
             size_t                offset);
 
+    GGML_API struct ggml_tensor * ggml_view_4d(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int64_t               ne0,
+            int64_t               ne1,
+            int64_t               ne2,
+            int64_t               ne3,
+            size_t                nb1, // row   stride in bytes
+            size_t                nb2, // slice stride in bytes
+            size_t                nb3,
+            size_t                offset);
+
     GGML_API struct ggml_tensor * ggml_permute(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
@@ -672,20 +821,50 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
+    GGML_API struct ggml_tensor * ggml_get_rows_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            struct ggml_tensor  * b,
+            struct ggml_tensor  * c);
+
+    GGML_API struct ggml_tensor * ggml_diag(
+        struct ggml_context     * ctx,
+        struct ggml_tensor      * a);
+
     // set elements above the diagonal to -INF
-    // in-place, returns view(a)
     GGML_API struct ggml_tensor * ggml_diag_mask_inf(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   n_past);
 
     // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_diag_mask_inf_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past);
+
+    // set elements above the diagonal to 0
+    GGML_API struct ggml_tensor * ggml_diag_mask_zero(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past);
+
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_diag_mask_zero_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past);
+
     GGML_API struct ggml_tensor * ggml_soft_max(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // rotary position embedding
     // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_soft_max_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a);
+
+    // rotary position embedding
     // if mode & 1 == 1, skip n_past elements
     // if mode & 2 == 1, GPT-NeoX style
     // TODO: avoid creating a new tensor every time
@@ -696,13 +875,39 @@ extern "C" {
             int                   n_dims,
             int                   mode);
 
+    // in-place, returns view(a)
+    GGML_API struct ggml_tensor * ggml_rope_inplace(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past,
+            int                   n_dims,
+            int                   mode);
+
+    // rotary position embedding backward, i.e compute dx from dy
+    // a - dy
+    GGML_API struct ggml_tensor * ggml_rope_back(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            int                   n_past,
+            int                   n_dims,
+            int                   mode);
+
     // alibi position embedding
     // in-place, returns view(a)
     struct ggml_tensor * ggml_alibi(
             struct ggml_context * ctx,
             struct ggml_tensor  * a,
             int                   n_past,
-            int                   n_head);
+            int                   n_head,
+            float                 bias_max);
+
+    // clamp
+    // in-place, returns view(a)
+    struct ggml_tensor * ggml_clamp(
+            struct ggml_context * ctx,
+            struct ggml_tensor  * a,
+            float                 min,
+            float                 max);
 
     // padding = 1
     // TODO: we don't support extra parameters for now
@@ -740,13 +945,13 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_map_unary_f32(
             struct ggml_context        * ctx,
             struct ggml_tensor         * a,
-            const  ggml_unary_op_f32_t fun);
+                   ggml_unary_op_f32_t   fun);
 
     GGML_API struct ggml_tensor * ggml_map_binary_f32(
             struct ggml_context         * ctx,
             struct ggml_tensor          * a,
             struct ggml_tensor          * b,
-            const  ggml_binary_op_f32_t fun);
+                   ggml_binary_op_f32_t   fun);
 
     //
     // automatic differentiation

llama-util.h

@@ -101,12 +101,12 @@ struct llama_file {
         LLAMA_ASSERT(ret == 0); // same
     }
 
-    void read_raw(void * ptr, size_t size) {
-        if (size == 0) {
+    void read_raw(void * ptr, size_t len) const {
+        if (len == 0) {
             return;
         }
         errno = 0;
-        std::size_t ret = std::fread(ptr, size, 1, fp);
+        std::size_t ret = std::fread(ptr, len, 1, fp);
         if (ferror(fp)) {
             throw std::runtime_error(format("read error: %s", strerror(errno)));
         }
@@ -127,12 +127,12 @@ struct llama_file {
         return std::string(chars.data(), len);
     }
 
-    void write_raw(const void * ptr, size_t size) {
-        if (size == 0) {
+    void write_raw(const void * ptr, size_t len) const {
+        if (len == 0) {
             return;
         }
         errno = 0;
-        size_t ret = std::fwrite(ptr, size, 1, fp);
+        size_t ret = std::fwrite(ptr, len, 1, fp);
         if (ret != 1) {
             throw std::runtime_error(format("write error: %s", strerror(errno)));
         }
@@ -172,7 +172,7 @@ struct llama_mmap {
 #ifdef _POSIX_MAPPED_FILES
     static constexpr bool SUPPORTED = true;
 
-    llama_mmap(struct llama_file * file, bool prefetch = true) {
+    llama_mmap(struct llama_file * file, size_t prefetch = (size_t) -1 /* -1 = max value */) {
         size = file->size;
         int fd = fileno(file->fp);
         int flags = MAP_SHARED;
@@ -184,9 +184,9 @@ struct llama_mmap {
             throw std::runtime_error(format("mmap failed: %s", strerror(errno)));
         }
 
-        if (prefetch) {
+        if (prefetch > 0) {
             // Advise the kernel to preload the mapped memory
-            if (madvise(addr, file->size, MADV_WILLNEED)) {
+            if (madvise(addr, std::min(file->size, prefetch), MADV_WILLNEED)) {
                 fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
                         strerror(errno));
             }
@@ -267,9 +267,9 @@ struct llama_mlock {
         }
     }
 
-    void init(void * addr) {
-        LLAMA_ASSERT(this->addr == NULL && this->size == 0);
-        this->addr = addr;
+    void init(void * ptr) {
+        LLAMA_ASSERT(addr == NULL && size == 0);
+        addr = ptr;
     }
 
     void grow_to(size_t target_size) {
@@ -340,14 +340,14 @@ struct llama_mlock {
         return (size_t) si.dwPageSize;
     }
 
-    bool raw_lock(void * addr, size_t size) {
+    bool raw_lock(void * ptr, size_t len) {
         for (int tries = 1; ; tries++) {
-            if (VirtualLock(addr, size)) {
+            if (VirtualLock(ptr, len)) {
                 return true;
             }
             if (tries == 2) {
                 fprintf(stderr, "warning: failed to VirtualLock %zu-byte buffer (after previously locking %zu bytes): %s\n",
-                        size, this->size, llama_format_win_err(GetLastError()).c_str());
+                    len, size, llama_format_win_err(GetLastError()).c_str());
                 return false;
             }
 
@@ -363,7 +363,7 @@ struct llama_mlock {
             // is equal to the number of pages in its minimum working set minus
             // a small overhead."
             // Hopefully a megabyte is enough overhead:
-            size_t increment = size + 1048576;
+            size_t increment = len + 1048576;
             // The minimum must be <= the maximum, so we need to increase both:
             min_ws_size += increment;
             max_ws_size += increment;
@@ -375,8 +375,8 @@ struct llama_mlock {
         }
     }
 
-    void raw_unlock(void * addr, size_t size) {
-        if (!VirtualUnlock(addr, size)) {
+    void raw_unlock(void * ptr, size_t len) {
+        if (!VirtualUnlock(ptr, len)) {
             fprintf(stderr, "warning: failed to VirtualUnlock buffer: %s\n",
                     llama_format_win_err(GetLastError()).c_str());
         }
@@ -388,12 +388,12 @@ struct llama_mlock {
         return (size_t) 65536;
     }
 
-    bool raw_lock(const void * addr, size_t size) {
+    bool raw_lock(const void * addr, size_t len) {
         fprintf(stderr, "warning: mlock not supported on this system\n");
         return false;
     }
 
-    void raw_unlock(const void * addr, size_t size) {}
+    void raw_unlock(const void * addr, size_t len) {}
 #endif
 };
 
@@ -404,10 +404,10 @@ struct llama_buffer {
 
     llama_buffer() = default;
 
-    void resize(size_t size) {
+    void resize(size_t len) {
         delete[] addr;
-        addr = new uint8_t[size];
-        this->size = size;
+        addr = new uint8_t[len];
+        size = len;
     }
 
     ~llama_buffer() {

llama.cpp

@@ -1,6 +1,7 @@
 // Defines fileno on msys:
 #ifndef _GNU_SOURCE
 #define _GNU_SOURCE
+#include <cstddef>
 #include <cstdint>
 #include <cstdio>
 #endif
@@ -9,6 +10,9 @@
 #include "llama.h"
 
 #include "ggml.h"
+#ifdef GGML_USE_CUBLAS
+#include "ggml-cuda.h"
+#endif
 
 #include <array>
 #include <ctime>
@@ -42,6 +46,7 @@ enum e_model {
     MODEL_65B,
 };
 
+
 static const size_t MB = 1024*1024;
 
 // computed for n_ctx == 2048
@@ -107,7 +112,7 @@ struct llama_hparams {
     enum llama_ftype ftype = LLAMA_FTYPE_MOSTLY_F16;
 
     bool operator!=(const llama_hparams & other) const {
-        return memcmp(this, &other, sizeof(llama_hparams));
+        return static_cast<bool>(memcmp(this, &other, sizeof(llama_hparams)));
     }
 };
 
@@ -403,6 +408,7 @@ enum llama_file_version {
     LLAMA_FILE_VERSION_GGMF_V1, // added version field and scores in vocab
     LLAMA_FILE_VERSION_GGJT_V1, // added padding
     LLAMA_FILE_VERSION_GGJT_V2, // changed quantization format
+    LLAMA_FILE_VERSION_GGJT_V3, // changed Q4 and Q8 quantization format
 };
 
 struct llama_file_loader {
@@ -421,24 +427,30 @@ struct llama_file_loader {
     }
     void read_magic() {
         uint32_t magic = file.read_u32();
-        uint32_t version = 0;
 
-        if (magic != 'ggml') {
-            version = file.read_u32();
-        }
-
-        if (magic == 'ggml' && version == 0) {
+        if (magic == LLAMA_FILE_MAGIC_GGML) {
             file_version = LLAMA_FILE_VERSION_GGML;
-        } else if (magic == 'ggmf' && version == 1) {
-            file_version = LLAMA_FILE_VERSION_GGMF_V1;
-        } else if (magic == 'ggjt' && version == 1) {
-            file_version = LLAMA_FILE_VERSION_GGJT_V1;
-        } else if (magic == 'ggjt' && version == 2) {
-            file_version = LLAMA_FILE_VERSION_GGJT_V2;
-        } else {
-            throw format("unknown (magic, version) combination: %08x, %08x; is this really a GGML file?",
-                         magic, version);
+            return;
         }
+
+        uint32_t version = file.read_u32();
+
+        switch (magic) {
+            case LLAMA_FILE_MAGIC_GGMF:
+                switch (version) {
+                    case 1: file_version = LLAMA_FILE_VERSION_GGMF_V1; return;
+                }
+                break;
+            case LLAMA_FILE_MAGIC_GGJT:
+                switch (version) {
+                    case 1: file_version = LLAMA_FILE_VERSION_GGJT_V1; return;
+                    case 2: file_version = LLAMA_FILE_VERSION_GGJT_V2; return;
+                    case 3: file_version = LLAMA_FILE_VERSION_GGJT_V3; return;
+                }
+        }
+
+        throw format("unknown (magic, version) combination: %08x, %08x; is this really a GGML file?",
+                     magic, version);
     }
     void read_hparams() {
         hparams.n_vocab = file.read_u32();
@@ -496,7 +508,7 @@ struct llama_file_loader {
 
             if (file_version >= LLAMA_FILE_VERSION_GGJT_V1) {
                 // skip to the next multiple of 32 bytes
-                file.seek(-file.tell() & 31, SEEK_CUR);
+                file.seek(-static_cast<ptrdiff_t>(file.tell()) & 31, SEEK_CUR);
             }
             shard.file_idx = file_idx;
             shard.file_off = file.tell();
@@ -571,7 +583,7 @@ struct llama_file_saver {
         file.write_u32(new_type);
         file.write_raw(tensor.ne.data(), sizeof(tensor.ne[0]) * tensor.ne.size());
         file.write_raw(tensor.name.data(), tensor.name.size());
-        file.seek(-file.tell() & 31, SEEK_CUR);
+        file.seek(-static_cast<ptrdiff_t>(file.tell()) & 31, SEEK_CUR);
         LLAMA_ASSERT(new_size == llama_calc_tensor_size(tensor.ne, new_type));
         file.write_raw(new_data, new_size);
     }
@@ -638,7 +650,7 @@ struct llama_model_loader {
         }
     }
 
-    struct ggml_tensor * get_tensor(const std::string & name, const std::vector<uint32_t> & ne) {
+    struct ggml_tensor * get_tensor(const std::string & name, const std::vector<uint32_t> & ne, ggml_backend backend) {
         auto it = tensors_map.name_to_idx.find(name);
         if (it == tensors_map.name_to_idx.end()) {
             throw format("llama.cpp: tensor '%s' is missing from model", name.c_str());
@@ -649,10 +661,10 @@ struct llama_model_loader {
                          name.c_str(), llama_format_tensor_shape(ne).c_str(), llama_format_tensor_shape(lt.ne).c_str());
         }
 
-        return get_tensor_for(lt);
+        return get_tensor_for(lt, backend);
     }
 
-    struct ggml_tensor * get_tensor_for(llama_load_tensor & lt) {
+    struct ggml_tensor * get_tensor_for(llama_load_tensor & lt, ggml_backend backend) {
         struct ggml_tensor * tensor;
         if (lt.ne.size() == 2) {
             tensor = ggml_new_tensor_2d(ggml_ctx, lt.type, lt.ne.at(0), lt.ne.at(1));
@@ -662,6 +674,7 @@ struct llama_model_loader {
         }
         ggml_set_name(tensor, lt.name.c_str());
         LLAMA_ASSERT(lt.ggml_tensor == NULL); // if this fails, we called get_tensor twice on the same tensor
+        tensor->backend = backend;
         lt.ggml_tensor = tensor;
         num_ggml_tensors_created++;
         return tensor;
@@ -675,12 +688,16 @@ struct llama_model_loader {
 
     void load_all_data(llama_progress_callback progress_callback, void *  progress_callback_user_data, llama_mlock * lmlock) {
         size_t data_size = 0;
+        size_t prefetch_size = 0;
         for (const llama_load_tensor & lt : tensors_map.tensors) {
             data_size += lt.size;
+            if (lt.ggml_tensor->backend == GGML_BACKEND_CPU) {
+                prefetch_size += lt.size;
+            }
         }
 
         if (use_mmap) {
-            mapping.reset(new llama_mmap(&file_loaders.at(0)->file));
+            mapping.reset(new llama_mmap(&file_loaders.at(0)->file, prefetch_size));
             if (!lmlock) {
                 // Don't call the callback since the actual loading will be lazy
                 // and we can't measure it.
@@ -693,6 +710,9 @@ struct llama_model_loader {
 
         size_t done_size = 0;
         for (llama_load_tensor & lt : tensors_map.tensors) {
+            if (lt.ggml_tensor->backend != GGML_BACKEND_CPU) {
+                continue;
+            }
             if (progress_callback) {
                 progress_callback((float) done_size / data_size, progress_callback_user_data);
             }
@@ -705,9 +725,6 @@ struct llama_model_loader {
                 lmlock->grow_to(done_size);
             }
         }
-        if (progress_callback) {
-            progress_callback(1.0f, progress_callback_user_data);
-        }
     }
 
     void load_data_for(llama_load_tensor & lt) {
@@ -809,9 +826,9 @@ static bool kv_cache_init(
 struct llama_context_params llama_context_default_params() {
     struct llama_context_params result = {
         /*.n_ctx                       =*/ 512,
-        /*.n_parts                     =*/ -1,
+        /*.gpu_layers                  =*/ 0,
         /*.seed                        =*/ -1,
-        /*.f16_kv                      =*/ false,
+        /*.f16_kv                      =*/ true,
         /*.logits_all                  =*/ false,
         /*.vocab_only                  =*/ false,
         /*.use_mmap                    =*/ true,
@@ -832,6 +849,21 @@ bool llama_mlock_supported() {
     return llama_mlock::SUPPORTED;
 }
 
+void llama_init_backend() {
+    ggml_time_init();
+
+    // needed to initialize f16 tables
+    {
+        struct ggml_init_params params = { 0, NULL, false };
+        struct ggml_context * ctx = ggml_init(params);
+        ggml_free(ctx);
+    }
+}
+
+int64_t llama_time_us() {
+    return ggml_time_us();
+}
+
 //
 // model loading
 //
@@ -841,7 +873,8 @@ static const char *llama_file_version_name(llama_file_version version) {
         case LLAMA_FILE_VERSION_GGML: return "'ggml' (old version with low tokenizer quality and no mmap support)";
         case LLAMA_FILE_VERSION_GGMF_V1: return "ggmf v1 (old version with no mmap support)";
         case LLAMA_FILE_VERSION_GGJT_V1: return "ggjt v1 (pre #1405)";
-        case LLAMA_FILE_VERSION_GGJT_V2: return "ggjt v2 (latest)";
+        case LLAMA_FILE_VERSION_GGJT_V2: return "ggjt v2 (pre #1508)";
+        case LLAMA_FILE_VERSION_GGJT_V3: return "ggjt v3 (latest)";
     }
 
     return "unknown";
@@ -876,6 +909,7 @@ static void llama_model_load_internal(
         const std::string & fname,
         llama_context & lctx,
         int n_ctx,
+        int n_gpu_layers,
         ggml_type memory_type,
         bool use_mmap,
         bool use_mlock,
@@ -920,11 +954,19 @@ static void llama_model_load_internal(
         fprintf(stderr, "%s: model size = %s\n",  __func__, llama_model_type_name(model.type));
     }
 
-    if (file_version != LLAMA_FILE_VERSION_GGJT_V2) {
+    if (file_version < LLAMA_FILE_VERSION_GGJT_V2) {
         if (hparams.ftype != LLAMA_FTYPE_ALL_F32     &&
             hparams.ftype != LLAMA_FTYPE_MOSTLY_F16  &&
             hparams.ftype != LLAMA_FTYPE_MOSTLY_Q8_0) {
-            throw format("this format is no longer supported (see https://github.com/ggerganov/llama.cpp/pull/1305)");
+            throw format("this format is no longer supported (see https://github.com/ggerganov/llama.cpp/pull/1405)");
+        }
+    }
+
+    if (file_version < LLAMA_FILE_VERSION_GGJT_V3) {
+        if (hparams.ftype == LLAMA_FTYPE_MOSTLY_Q4_0 ||
+            hparams.ftype == LLAMA_FTYPE_MOSTLY_Q4_1 ||
+            hparams.ftype == LLAMA_FTYPE_MOSTLY_Q8_0) {
+            throw format("this format is no longer supported (see https://github.com/ggerganov/llama.cpp/pull/1508)");
         }
     }
 
@@ -937,27 +979,7 @@ static void llama_model_load_internal(
     size_t ctx_size;
     size_t mmapped_size;
     ml->calc_sizes(&ctx_size, &mmapped_size);
-    fprintf(stderr, "%s: ggml ctx size = %6.2f KB\n", __func__, ctx_size/1024.0);
-
-    // print memory requirements
-    {
-        const size_t scale = memory_type == GGML_TYPE_F32 ? 2 : 1;
-
-        // this is the total memory required to run the inference
-        const size_t mem_required =
-            ctx_size +
-            mmapped_size +
-            MEM_REQ_SCRATCH0().at(model.type) +
-            MEM_REQ_SCRATCH1().at(model.type) +
-            MEM_REQ_EVAL().at(model.type);
-
-        // this is the memory required by one llama_state
-        const size_t mem_required_state =
-            scale*MEM_REQ_KV_SELF().at(model.type);
-
-        fprintf(stderr, "%s: mem required  = %7.2f MB (+ %7.2f MB per state)\n", __func__,
-                mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0);
-    }
+    fprintf(stderr, "%s: ggml ctx size = %7.2f MB\n", __func__, ctx_size/1024.0/1024.0);
 
     // create the ggml context
     {
@@ -979,7 +1001,14 @@ static void llama_model_load_internal(
         }
     }
 
+#ifdef GGML_USE_CUBLAS
+#define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_CUDA
+#else
+#define LLAMA_BACKEND_OFFLOAD GGML_BACKEND_CPU
+#endif
+
     // prepare memory for the weights
+    size_t vram_total = 0;
     {
         const uint32_t n_embd  = hparams.n_embd;
         const uint32_t n_layer = hparams.n_layer;
@@ -987,33 +1016,87 @@ static void llama_model_load_internal(
 
         ml->ggml_ctx = ctx;
 
-        model.tok_embeddings = ml->get_tensor("tok_embeddings.weight", {n_embd, n_vocab});
-        model.norm           = ml->get_tensor("norm.weight",           {n_embd});
-        model.output         = ml->get_tensor("output.weight",         {n_embd, n_vocab});
+        model.tok_embeddings = ml->get_tensor("tok_embeddings.weight", {n_embd, n_vocab}, GGML_BACKEND_CPU);
+        model.norm           = ml->get_tensor("norm.weight",           {n_embd},          GGML_BACKEND_CPU);
+
+        // "output" tensor
+        {
+            ggml_backend backend_output;
+            if (n_gpu_layers > int(n_layer)) { // NOLINT
+                backend_output = LLAMA_BACKEND_OFFLOAD;
+            } else {
+                backend_output = GGML_BACKEND_CPU;
+            }
+
+            model.output = ml->get_tensor("output.weight", {n_embd, n_vocab}, backend_output);
+        }
+
+        const int i_gpu_start = n_layer - n_gpu_layers;
 
         model.layers.resize(n_layer);
         for (uint32_t i = 0; i < n_layer; ++i) {
+            const ggml_backend backend = int(i) < i_gpu_start ? GGML_BACKEND_CPU : LLAMA_BACKEND_OFFLOAD;
+
             auto & layer = model.layers[i];
 
             std::string layers_i = "layers." + std::to_string(i);
 
-            layer.attention_norm = ml->get_tensor(layers_i + ".attention_norm.weight", {n_embd});
+            layer.attention_norm = ml->get_tensor(layers_i + ".attention_norm.weight", {n_embd}, backend);
 
-            layer.wq = ml->get_tensor(layers_i + ".attention.wq.weight", {n_embd, n_embd});
-            layer.wk = ml->get_tensor(layers_i + ".attention.wk.weight", {n_embd, n_embd});
-            layer.wv = ml->get_tensor(layers_i + ".attention.wv.weight", {n_embd, n_embd});
-            layer.wo = ml->get_tensor(layers_i + ".attention.wo.weight", {n_embd, n_embd});
+            layer.wq = ml->get_tensor(layers_i + ".attention.wq.weight", {n_embd, n_embd}, backend);
+            layer.wk = ml->get_tensor(layers_i + ".attention.wk.weight", {n_embd, n_embd}, backend);
+            layer.wv = ml->get_tensor(layers_i + ".attention.wv.weight", {n_embd, n_embd}, backend);
+            layer.wo = ml->get_tensor(layers_i + ".attention.wo.weight", {n_embd, n_embd}, backend);
 
-            layer.ffn_norm = ml->get_tensor(layers_i + ".ffn_norm.weight", {n_embd});
+            layer.ffn_norm = ml->get_tensor(layers_i + ".ffn_norm.weight", {n_embd}, backend);
 
-            layer.w1 = ml->get_tensor(layers_i + ".feed_forward.w1.weight", {n_embd,   n_ff});
-            layer.w2 = ml->get_tensor(layers_i + ".feed_forward.w2.weight", {  n_ff,   n_embd});
-            layer.w3 = ml->get_tensor(layers_i + ".feed_forward.w3.weight", {n_embd,   n_ff});
+            layer.w1 = ml->get_tensor(layers_i + ".feed_forward.w1.weight", {n_embd,   n_ff},   backend);
+            layer.w2 = ml->get_tensor(layers_i + ".feed_forward.w2.weight", {  n_ff,   n_embd}, backend);
+            layer.w3 = ml->get_tensor(layers_i + ".feed_forward.w3.weight", {n_embd,   n_ff},   backend);
+
+            if (backend == GGML_BACKEND_CUDA) {
+                vram_total +=
+                    ggml_nbytes(layer.attention_norm) + ggml_nbytes(layer.wq) + ggml_nbytes(layer.wk)             +
+                    ggml_nbytes(layer.wv)             + ggml_nbytes(layer.wo) + ggml_nbytes(layer.attention_norm) +
+                    ggml_nbytes(layer.w1)             + ggml_nbytes(layer.w2) + ggml_nbytes(layer.w3);
+            }
         }
     }
 
     ml->done_getting_tensors();
 
+    // print memory requirements
+    {
+        const size_t scale = memory_type == GGML_TYPE_F32 ? 2 : 1;
+
+        // this is the total memory required to run the inference
+        const size_t mem_required =
+            ctx_size +
+            mmapped_size - vram_total + // weights in VRAM not in memory
+            MEM_REQ_SCRATCH0().at(model.type) +
+            MEM_REQ_SCRATCH1().at(model.type) +
+            MEM_REQ_EVAL().at(model.type);
+
+        // this is the memory required by one llama_state
+        const size_t mem_required_state =
+            scale*MEM_REQ_KV_SELF().at(model.type);
+
+        fprintf(stderr, "%s: mem required  = %7.2f MB (+ %7.2f MB per state)\n", __func__,
+                mem_required / 1024.0 / 1024.0, mem_required_state / 1024.0 / 1024.0);
+
+#ifdef GGML_USE_CUBLAS
+        const int n_gpu = std::min(n_gpu_layers, int(hparams.n_layer));
+
+        fprintf(stderr, "%s: [cublas] offloading %d layers to GPU\n", __func__, n_gpu);
+        if (n_gpu_layers > (int) hparams.n_layer) {
+            fprintf(stderr, "%s: [cublas] offloading output layer to GPU\n", __func__);
+        }
+        fprintf(stderr, "%s: [cublas] total VRAM used: %zu MB\n", __func__, vram_total / 1024 / 1024);
+#else
+        (void) n_gpu_layers;
+#endif
+    }
+
     // populate `tensors_by_name`
     for (llama_load_tensor & lt : ml->tensors_map.tensors) {
         model.tensors_by_name.emplace_back(lt.name, lt.ggml_tensor);
@@ -1021,6 +1104,33 @@ static void llama_model_load_internal(
 
     ml->load_all_data(progress_callback, progress_callback_user_data, use_mlock ? &lctx.model.mlock_mmap : NULL);
 
+#ifdef GGML_USE_CUBLAS
+    {
+        size_t done_size = 0;
+        size_t data_size = 0;
+        for (llama_load_tensor & lt : ml->tensors_map.tensors) {
+            data_size += lt.size;
+            if (lt.ggml_tensor->backend == GGML_BACKEND_CPU) {
+                done_size += lt.size;
+            }
+        }
+        for (llama_load_tensor & lt : ml->tensors_map.tensors) {
+            if (lt.ggml_tensor->backend != GGML_BACKEND_CUDA) {
+                continue;
+            }
+            if (progress_callback) {
+                progress_callback((float) done_size / data_size, progress_callback_user_data);
+            }
+            ggml_cuda_load_data(fname.c_str(), lt.ggml_tensor, lt.shards.at(0).file_off);
+            done_size += lt.size;
+        }
+    }
+#endif // GGML_USE_CUBLAS
+
+    if (progress_callback) {
+        progress_callback(1.0f, progress_callback_user_data);
+    }
+
     model.mapping = std::move(ml->mapping);
 
     // loading time will be recalculate after the first eval, so
@@ -1032,6 +1142,7 @@ static bool llama_model_load(
         const std::string & fname,
         llama_context & lctx,
         int n_ctx,
+        int n_gpu_layers,
         ggml_type memory_type,
         bool use_mmap,
         bool use_mlock,
@@ -1039,7 +1150,7 @@ static bool llama_model_load(
         llama_progress_callback progress_callback,
         void *progress_callback_user_data) {
     try {
-        llama_model_load_internal(fname, lctx, n_ctx, memory_type, use_mmap, use_mlock,
+        llama_model_load_internal(fname, lctx, n_ctx, n_gpu_layers, memory_type, use_mmap, use_mlock,
                                   vocab_only, progress_callback, progress_callback_user_data);
         return true;
     } catch (const std::string & err) {
@@ -1119,17 +1230,15 @@ static bool llama_eval_internal(
         {
             cur = ggml_rms_norm(ctx0, inpL);
 
-            // cur = attention_norm*cur
-            cur = ggml_mul(ctx0,
-                        ggml_repeat(ctx0, model.layers[il].attention_norm, cur),
-                        cur);
+            // cur = cur*attention_norm(broadcasted)
+            cur = ggml_mul(ctx0, cur, model.layers[il].attention_norm);
         }
 
         // self-attention
         {
             // 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);
+            struct ggml_tensor * Qcur = ggml_rope_inplace(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_inplace(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);
             ggml_set_name(Qcur, "Qcur");
             ggml_set_name(Kcur, "Kcur");
 
@@ -1170,17 +1279,19 @@ static bool llama_eval_internal(
             struct ggml_tensor * KQ_scale = ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head));
             ggml_set_name(KQ_scale, "1/sqrt(n_embd/n_head)");
 
-            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, KQ_scale);
+            // KQ_scaled shape [n_past + N, N, n_head, 1]
+            struct ggml_tensor * KQ_scaled = ggml_scale_inplace(ctx0, KQ, KQ_scale);
             ggml_set_name(KQ_scaled, "KQ_scaled");
 
             // KQ_masked = mask_past(KQ_scaled)
-            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf_inplace(ctx0, KQ_scaled, n_past);
             ggml_set_name(KQ_masked, "KQ_masked");
 
             // KQ = soft_max(KQ_masked)
-            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max_inplace(ctx0, KQ_masked);
             ggml_set_name(KQ_soft_max, "KQ_soft_max");
 
+
             // split cached V into n_head heads
             struct ggml_tensor * V =
                 ggml_view_3d(ctx0, kv_self.v,
@@ -1227,10 +1338,8 @@ static bool llama_eval_internal(
             {
                 cur = ggml_rms_norm(ctx0, inpFF);
 
-                // cur = ffn_norm*cur
-                cur = ggml_mul(ctx0,
-                        ggml_repeat(ctx0, model.layers[il].ffn_norm, cur),
-                        cur);
+                // cur = cur*ffn_norm(broadcasted)
+                cur = ggml_mul(ctx0, cur, model.layers[il].ffn_norm);
             }
 
             struct ggml_tensor * tmp = ggml_mul_mat(ctx0,
@@ -1267,10 +1376,8 @@ static bool llama_eval_internal(
 
         inpL = ggml_rms_norm(ctx0, inpL);
 
-        // inpL = norm*inpL
-        inpL = ggml_mul(ctx0,
-                    ggml_repeat(ctx0, model.norm, inpL),
-                    inpL);
+        // inpL = inpL*norm(broadcasted)
+        inpL = ggml_mul(ctx0, inpL, model.norm);
 
         embeddings = inpL;
     }
@@ -1281,7 +1388,7 @@ static bool llama_eval_internal(
     lctx.use_buf(ctx0, -1);
 
     // logits -> probs
-    //inpL = ggml_soft_max(ctx0, inpL);
+    //inpL = ggml_soft_max_inplace(ctx0, inpL);
 
     // run the computation
     ggml_build_forward_expand(&gf, inpL);
@@ -2094,7 +2201,7 @@ struct llama_context * llama_init_from_file(
             unsigned * cur_percentage_p = (unsigned *) ctx;
             unsigned percentage = (unsigned) (100 * progress);
             while (percentage > *cur_percentage_p) {
-                ++*cur_percentage_p;
+                *cur_percentage_p = percentage;
                 fprintf(stderr, ".");
                 fflush(stderr);
                 if (percentage >= 100) {
@@ -2109,7 +2216,7 @@ struct llama_context * llama_init_from_file(
 
     ggml_type memory_type = params.f16_kv ? GGML_TYPE_F16 : GGML_TYPE_F32;
 
-    if (!llama_model_load(path_model, *ctx, params.n_ctx, memory_type,
+    if (!llama_model_load(path_model, *ctx, params.n_ctx, params.n_gpu_layers, memory_type,
                           params.use_mmap, params.use_mlock, params.vocab_only,
                           params.progress_callback, params.progress_callback_user_data)) {
         fprintf(stderr, "%s: failed to load model\n", __func__);
@@ -2187,7 +2294,7 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
     {
         uint32_t magic;
         fin.read((char *) &magic, sizeof(magic));
-        if (magic != 'ggla') {
+        if (magic != LLAMA_FILE_MAGIC_GGLA) {
             fprintf(stderr, "%s: bad file magic\n", __func__);
             return 1;
         }
@@ -2251,7 +2358,7 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
 
         // maybe this should in llama_model_loader
         if (model_loader->use_mmap) {
-            model_loader->mapping.reset(new llama_mmap(&model_loader->file_loaders.at(0)->file, /* prefetch */ false));
+            model_loader->mapping.reset(new llama_mmap(&model_loader->file_loaders.at(0)->file, /* prefetch */ 0));
         }
     }
 
@@ -2344,7 +2451,7 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
                 }
                 size_t idx = model_loader->tensors_map.name_to_idx[base_name];
                 llama_load_tensor & lt = model_loader->tensors_map.tensors[idx];
-                base_t = model_loader->get_tensor(base_name, { (uint32_t)dest_t->ne[0], (uint32_t)dest_t->ne[1] });
+                base_t = model_loader->get_tensor(base_name, { (uint32_t)dest_t->ne[0], (uint32_t)dest_t->ne[1] }, GGML_BACKEND_CPU);
                 lt.data = (uint8_t *) lt.ggml_tensor->data;
                 model_loader->load_data_for(lt);
                 lt.ggml_tensor->data = lt.data;
@@ -2375,7 +2482,7 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
 
             if (scaling != 1.0f) {
                 ggml_tensor * scale_tensor = ggml_new_f32(lora_ctx, scaling);
-                BA = ggml_scale(lora_ctx, BA, scale_tensor);
+                BA = ggml_scale_inplace(lora_ctx, BA, scale_tensor);
             }
 
             ggml_tensor * r;
@@ -2570,8 +2677,8 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
 }
 
 // Sets the state reading from the specified source address
-size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) {
-    const uint8_t * inp = src;
+size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
+    uint8_t * inp = src;
 
     // set rng
     {

llama.h

@@ -19,10 +19,16 @@
 #    define LLAMA_API
 #endif
 
-#define LLAMA_FILE_VERSION           2
-#define LLAMA_FILE_MAGIC             'ggjt'
-#define LLAMA_FILE_MAGIC_UNVERSIONED 'ggml'
-#define LLAMA_SESSION_MAGIC          'ggsn'
+#define LLAMA_FILE_MAGIC_GGJT        0x67676a74u // 'ggjt'
+#define LLAMA_FILE_MAGIC_GGLA        0x67676c61u // 'ggla'
+#define LLAMA_FILE_MAGIC_GGMF        0x67676d66u // 'ggmf'
+#define LLAMA_FILE_MAGIC_GGML        0x67676d6cu // 'ggml'
+#define LLAMA_FILE_MAGIC_GGSN        0x6767736eu // 'ggsn'
+
+#define LLAMA_FILE_VERSION           3
+#define LLAMA_FILE_MAGIC             LLAMA_FILE_MAGIC_GGJT
+#define LLAMA_FILE_MAGIC_UNVERSIONED LLAMA_FILE_MAGIC_GGML
+#define LLAMA_SESSION_MAGIC          LLAMA_FILE_MAGIC_GGSN
 #define LLAMA_SESSION_VERSION        1
 
 #ifdef __cplusplus
@@ -40,9 +46,9 @@ extern "C" {
     typedef int llama_token;
 
     typedef struct llama_token_data {
-        llama_token id;  // token id
-        float logit; // log-odds of the token
-        float p;     // probability of the token
+        llama_token id; // token id
+        float logit;    // log-odds of the token
+        float p;        // probability of the token
     } llama_token_data;
 
     typedef struct llama_token_data_array {
@@ -54,9 +60,9 @@ extern "C" {
     typedef void (*llama_progress_callback)(float progress, void *ctx);
 
     struct llama_context_params {
-        int n_ctx;   // text context
-        int n_parts; // -1 for default
-        int seed;    // RNG seed, -1 for random
+        int n_ctx;        // text context
+        int n_gpu_layers; // number of layers to store in VRAM
+        int seed;         // RNG seed, -1 for random
 
         bool f16_kv;     // use fp16 for KV cache
         bool logits_all; // the llama_eval() call computes all logits, not just the last one
@@ -73,16 +79,16 @@ extern "C" {
 
     // model file types
     enum llama_ftype {
-        LLAMA_FTYPE_ALL_F32     = 0,
-        LLAMA_FTYPE_MOSTLY_F16  = 1,  // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_0 = 2,  // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q4_1 = 3,  // except 1d tensors
+        LLAMA_FTYPE_ALL_F32              = 0,
+        LLAMA_FTYPE_MOSTLY_F16           = 1, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_0          = 2, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_1          = 3, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
-        // LLAMA_FTYPE_MOSTLY_Q4_2 = 5,  // support has been removed
-        // LLAMA_FTYPE_MOSTLY_Q4_3 (6) support has been removed
-        LLAMA_FTYPE_MOSTLY_Q8_0 = 7,  // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_0 = 8,  // except 1d tensors
-        LLAMA_FTYPE_MOSTLY_Q5_1 = 9,  // except 1d tensors
+        // LLAMA_FTYPE_MOSTLY_Q4_2       = 5, // support has been removed
+        // LLAMA_FTYPE_MOSTLY_Q4_3       = 6, // support has been removed
+        LLAMA_FTYPE_MOSTLY_Q8_0          = 7, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_0          = 8, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q5_1          = 9, // except 1d tensors
     };
 
     LLAMA_API struct llama_context_params llama_context_default_params();
@@ -90,6 +96,13 @@ extern "C" {
     LLAMA_API bool llama_mmap_supported();
     LLAMA_API bool llama_mlock_supported();
 
+    // TODO: not great API - very likely to change
+    // Initialize the llama + ggml backend
+    // Call once at the start of the program
+    LLAMA_API void llama_init_backend();
+
+    LLAMA_API int64_t llama_time_us();
+
     // Various functions for loading a ggml llama model.
     // Allocate (almost) all memory needed for the model.
     // Return NULL on failure
@@ -138,7 +151,7 @@ extern "C" {
 
     // Set the state reading from the specified address
     // Returns the number of bytes read
-    LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src);
+    LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src);
 
     // Save/load session file
     LLAMA_API bool llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out);

tests/CMakeLists.txt

@@ -10,3 +10,5 @@ llama_add_test(test-quantize-fns.cpp)
 llama_add_test(test-quantize-perf.cpp)
 llama_add_test(test-sampling.cpp)
 llama_add_test(test-tokenizer-0.cpp ${CMAKE_CURRENT_SOURCE_DIR}/../models/ggml-vocab.bin)
+# llama_add_test(test-grad0.c) # SLOW
+# llama_add_test(test-opt.c) # SLOW

tests/test-opt.c

@@ -0,0 +1,205 @@
+#include "ggml.h"
+
+#include <math.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <assert.h>
+
+#define MAX_NARGS 2
+
+
+//
+// logging
+//
+#define GGML_DEBUG 0
+#if (GGML_DEBUG >= 1)
+#define GGML_PRINT_DEBUG(...) printf(__VA_ARGS__)
+#else
+#define GGML_PRINT_DEBUG(...)
+#endif
+
+#if (GGML_DEBUG >= 5)
+#define GGML_PRINT_DEBUG_5(...) printf(__VA_ARGS__)
+#else
+#define GGML_PRINT_DEBUG_5(...)
+#endif
+
+#if (GGML_DEBUG >= 10)
+#define GGML_PRINT_DEBUG_10(...) printf(__VA_ARGS__)
+#else
+#define GGML_PRINT_DEBUG_10(...)
+#endif
+
+#define GGML_PRINT(...) printf(__VA_ARGS__)
+
+
+float frand() {
+    return (float)rand()/(float)RAND_MAX;
+}
+
+int irand(int n) {
+    return rand()%n;
+}
+
+void get_random_dims(int64_t * dims, int ndims) {
+    dims[0] = dims[1] = dims[2] = dims[3] = 1;
+
+    for (int i = 0; i < ndims; i++) {
+        dims[i] = 1 + irand(4);
+    }
+}
+
+void get_random_dims_minmax(int64_t * dims, int ndims, int min, int max) {
+    dims[0] = dims[1] = dims[2] = dims[3] = 1;
+
+    for (int i = 0; i < ndims; i++) {
+        dims[i] = min + irand(max-min);
+    }
+}
+
+
+struct ggml_tensor * get_random_tensor(
+        struct ggml_context * ctx0,
+        int ndims,
+        int64_t ne[],
+        float fmin,
+        float fmax) {
+    struct ggml_tensor * result = ggml_new_tensor(ctx0, GGML_TYPE_F32, ndims, ne);
+
+    switch (ndims) {
+        case 1:
+            for (int i0 = 0; i0 < ne[0]; i0++) {
+                ((float *)result->data)[i0] = frand()*(fmax - fmin) + fmin;
+            }
+            break;
+        case 2:
+            for (int i1 = 0; i1 < ne[1]; i1++) {
+                for (int i0 = 0; i0 < ne[0]; i0++) {
+                    ((float *)result->data)[i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
+                }
+            }
+            break;
+        case 3:
+            for (int i2 = 0; i2 < ne[2]; i2++) {
+                for (int i1 = 0; i1 < ne[1]; i1++) {
+                    for (int i0 = 0; i0 < ne[0]; i0++) {
+                        ((float *)result->data)[i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
+                    }
+                }
+            }
+            break;
+        case 4:
+            for (int i3 = 0; i3 < ne[3]; i3++) {
+                for (int i2 = 0; i2 < ne[2]; i2++) {
+                    for (int i1 = 0; i1 < ne[1]; i1++) {
+                        for (int i0 = 0; i0 < ne[0]; i0++) {
+                            ((float *)result->data)[i3*ne[2]*ne[1]*ne[0] + i2*ne[1]*ne[0] + i1*ne[0] + i0] = frand()*(fmax - fmin) + fmin;
+                        }
+                    }
+                }
+            }
+            break;
+        default:
+            assert(false);
+    };
+
+    return result;
+}
+
+float get_element(const struct ggml_tensor * t, int idx) {
+    return ((float *)t->data)[idx];
+}
+
+void set_element(struct ggml_tensor * t, int idx, float value) {
+    ((float *)t->data)[idx] = value;
+}
+
+int main(int argc, const char ** argv) {
+    struct ggml_init_params params = {
+        .mem_size   = 1024*1024*1024,
+        .mem_buffer = NULL,
+        .no_alloc   = false,
+    };
+    struct ggml_context * ctx = ggml_init(params);
+
+    int64_t ne1[4] = {4, 1024, 1, 1};
+    int64_t ne2[4] = {4, 2048, 1, 1};;
+    int64_t ne3[4] = {1024, 2048, 1, 1};
+
+    struct ggml_tensor * a = get_random_tensor(ctx, 2, ne1, -1, +1);
+    struct ggml_tensor * b = get_random_tensor(ctx, 2, ne2, -1, +1);
+    ggml_set_param(ctx, a);
+    ggml_set_param(ctx, b);
+
+    struct ggml_tensor * c = get_random_tensor(ctx, 2, ne3, -1, +1);
+
+    struct ggml_tensor * ab = ggml_mul_mat(ctx, a, b);
+    struct ggml_tensor * d  = ggml_sub(ctx, c, ab);
+    struct ggml_tensor * e  = ggml_sum(ctx, ggml_sqr(ctx, d));
+
+
+    struct ggml_cgraph ge = ggml_build_forward(e);
+    ggml_graph_reset  (&ge);
+    ggml_graph_compute(ctx, &ge);
+    const float fe = ggml_get_f32_1d(e, 0);
+    printf("%s: e = %.4f\n", __func__, fe);
+
+    struct ggml_opt_params opt_params = ggml_opt_default_params(GGML_OPT_ADAM);
+
+    ggml_opt(ctx, opt_params, e);
+
+    ggml_graph_reset  (&ge);
+    ggml_graph_compute(ctx, &ge);
+    const float fe_opt = ggml_get_f32_1d(e, 0);
+    printf("%s: original  e = %.4f\n", __func__, fe);
+    printf("%s: optimized e = %.4f\n", __func__, fe_opt);
+
+    const bool success = (fe_opt <= fe);
+    assert(success);
+
+    ggml_free(ctx);
+    return success ? 0 : -1;
+}
+// int64_t ne1[4] = {4, 128, 1, 1};
+// int64_t ne2[4] = {4, 256, 1, 1};;
+// int64_t ne3[4] = {128, 256, 1, 1};
+// main: original  e = 25890.9375
+// main: optimized e = 10094.7031
+
+// int64_t ne1[4] = {8, 128, 1, 1};
+// int64_t ne2[4] = {8, 256, 1, 1};;
+// int64_t ne3[4] = {128, 256, 1, 1};
+// main: original  e = 39429.5078
+// main: optimized e = 9275.8936
+
+// int64_t ne1[4] = {16, 128, 1, 1};
+// int64_t ne2[4] = {16, 256, 1, 1};;
+// int64_t ne3[4] = {128, 256, 1, 1};
+// main: original  e = 68371.1328
+// main: optimized e = 7854.4502
+
+
+// int64_t ne1[4] = {32, 128, 1, 1};
+// int64_t ne2[4] = {32, 256, 1, 1};;
+// int64_t ne3[4] = {128, 256, 1, 1};
+// main: original  e = 126061.1953
+// main: optimized e = 5451.0166
+
+// int64_t ne1[4] = {4, 1024, 1, 1};
+// int64_t ne2[4] = {4, 2048, 1, 1};;
+// int64_t ne3[4] = {1024, 2048, 1, 1};
+// main: original  e = 1620817.8750
+// main: optimized e = 698387.6875
+
+// another run on M1
+// int64_t ne1[4] = {4, 1024, 1, 1};
+// int64_t ne2[4] = {4, 2048, 1, 1};;
+// int64_t ne3[4] = {1024, 2048, 1, 1};
+// main: original  e = 1629595.6250
+// main: optimized e = 698169.1250
+
+// int64_t ne1[4] = {32, 1024, 1, 1};
+// int64_t ne2[4] = {32, 2048, 1, 1};;
+// int64_t ne3[4] = {1024, 2048, 1, 1};
+// main: original  e = 8146770.5000
+// main: optimized e = 651119.1250

tests/test-sampling.cpp

@@ -1,6 +1,10 @@
-#include "llama.h"
 #include "ggml.h"
-#include <cassert>
+#include "llama.h"
+
+#ifdef NDEBUG
+#undef NDEBUG
+#endif
+
 #include <cmath>
 #include <numeric>
 #include <cassert>
@@ -8,7 +12,6 @@
 #include <vector>
 #include <algorithm>
 
-
 void dump(const llama_token_data_array * candidates) {
     for (size_t i = 0; i < candidates->size; i++) {
         printf("%d: %f (%f)\n", candidates->data[i].id, candidates->data[i].p, candidates->data[i].logit);