llama : update stubs for systems without mmap and mlock (#1266)

Co-authored-by: John Doe <john.doe@example.com>

.gitignore

@@ -28,7 +28,7 @@ models/*
 /result
 /perplexity
 /embedding
-/benchmark-q4_0-matmult
+/benchmark-matmult
 /vdot
 /Pipfile
 

CMakeLists.txt

@@ -258,9 +258,22 @@ if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "arm" OR ${CMAKE_SYSTEM_PROCESSOR} MATCHES
         # TODO: arm msvc?
     else()
         if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "aarch64")
+            # Apple M1, M2, etc.
+            # Raspberry Pi 3, 4, Zero 2 (64-bit)
             add_compile_options(-mcpu=native)
         endif()
-        # TODO: armv6,7,8 version specific flags
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv6")
+            # Raspberry Pi 1, Zero
+            add_compile_options(-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access)
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv7")
+            # Raspberry Pi 2
+            add_compile_options(-mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations)
+        endif()
+        if (${CMAKE_SYSTEM_PROCESSOR} MATCHES "armv8")
+            # Raspberry Pi 3, 4, Zero 2 (32-bit)
+            add_compile_options(-mfp16-format=ieee -mno-unaligned-access)
+        endif()
     endif()
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "^(x86_64|i686|AMD64)$")
     message(STATUS "x86 detected")

Makefile

@@ -34,10 +34,15 @@ endif
 #
 
 # keep standard at C11 and C++11
-CFLAGS   = -I.              -O3 -DNDEBUG -std=c11   -fPIC
-CXXFLAGS = -I. -I./examples -O3 -DNDEBUG -std=c++11 -fPIC
+CFLAGS   = -I.              -O3 -std=c11   -fPIC
+CXXFLAGS = -I. -I./examples -O3 -std=c++11 -fPIC
 LDFLAGS  =
 
+ifndef LLAMA_DEBUG
+	CFLAGS   += -DNDEBUG
+	CXXFLAGS += -DNDEBUG
+endif
+
 # warnings
 CFLAGS   += -Wall -Wextra -Wpedantic -Wcast-qual -Wdouble-promotion -Wshadow -Wstrict-prototypes -Wpointer-arith
 CXXFLAGS += -Wall -Wextra -Wpedantic -Wcast-qual -Wno-unused-function -Wno-multichar
@@ -130,19 +135,21 @@ ifdef LLAMA_PERF
 	CXXFLAGS += -DGGML_PERF
 endif
 ifneq ($(filter aarch64%,$(UNAME_M)),)
+	# Apple M1, M2, etc.
+	# Raspberry Pi 3, 4, Zero 2 (64-bit)
 	CFLAGS   += -mcpu=native
 	CXXFLAGS += -mcpu=native
 endif
 ifneq ($(filter armv6%,$(UNAME_M)),)
-	# Raspberry Pi 1, 2, 3
+	# Raspberry Pi 1, Zero
 	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access
 endif
 ifneq ($(filter armv7%,$(UNAME_M)),)
-	# Raspberry Pi 4
+	# Raspberry Pi 2
 	CFLAGS += -mfpu=neon-fp-armv8 -mfp16-format=ieee -mno-unaligned-access -funsafe-math-optimizations
 endif
 ifneq ($(filter armv8%,$(UNAME_M)),)
-	# Raspberry Pi 4
+	# Raspberry Pi 3, 4, Zero 2 (32-bit)
 	CFLAGS += -mfp16-format=ieee -mno-unaligned-access
 endif
 
@@ -175,7 +182,7 @@ common.o: examples/common.cpp examples/common.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 clean:
-	rm -vf *.o main quantize quantize-stats perplexity embedding benchmark-q4_0-matmult
+	rm -vf *.o main quantize quantize-stats perplexity embedding benchmark-matmult
 
 main: examples/main/main.cpp ggml.o llama.o common.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
@@ -205,9 +212,9 @@ libllama.so: llama.o ggml.o $(OBJS)
 # Tests
 #
 
-benchmark: examples/benchmark/benchmark-q4_0-matmult.c ggml.o $(OBJS)
-	$(CXX) $(CXXFLAGS) $^ -o benchmark-q4_0-matmult $(LDFLAGS)
-	./benchmark-q4_0-matmult
+benchmark-matmult: examples/benchmark/benchmark-matmult.cpp ggml.o $(OBJS)
+	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)
+	./$@
 
 .PHONY: tests
 tests:

examples/CMakeLists.txt

@@ -35,4 +35,5 @@ else()
     add_subdirectory(perplexity)
     add_subdirectory(embedding)
     add_subdirectory(save-load-state)
+    add_subdirectory(benchmark)
 endif()

examples/benchmark/CMakeLists.txt

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

examples/benchmark/benchmark-matmult.cpp

@@ -1,11 +1,3 @@
-/*
-    License: MIT License
-
-    Changelog:
-    - 2023-03-31 Initial version by Sebastian Apel (https://github.com/SebastianApel)
-
-*/
-
 #include <locale.h>
 #include "ggml.h"
 #include <assert.h>
@@ -45,7 +37,7 @@ float tensor_sum_elements(struct ggml_tensor * tensor) {
 
 #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, \
+#define TENSOR_DUMP(TENSOR) printf("%15s: type = %i (%5s) ne = %5ld x %5ld x %5ld, nb = (%5li, %5li, %5li) - ", #TENSOR, \
         TENSOR->type,TENSOR_TYPE_AS_STR(TENSOR->type),\
         TENSOR->ne[0], TENSOR->ne[1], 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); }
@@ -98,12 +90,9 @@ int main(int argc, char ** argv)  {
         }
     }
 
-
     // create the ggml context
     printf("Starting Test\n");
 
-
-
     struct ggml_context * ctx;
     //const int sizex = 4096;
     //const int sizey = 11008;
@@ -125,16 +114,18 @@ int main(int argc, char ** argv)  {
 #endif
 
     //printf("Memsize required = %i\n", sizex*sizex);
-    ggml_type wtype = GGML_TYPE_F32;
 
     size_t ctx_size = 0;
-    ctx_size += sizex*sizey*ggml_type_sizef(wtype);
-    ctx_size += sizex*sizey*ggml_type_sizef(wtype);
     ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
-    ctx_size += sizex*sizeof(float);
-    ctx_size += 1024*1024*100;
+    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32);
+    ctx_size += sizex*sizez*ggml_type_sizef(GGML_TYPE_F32);
+    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_Q4_0);
+    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_Q4_0);
+    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
+    ctx_size += sizex*sizey*ggml_type_sizef(GGML_TYPE_F32); // BLAS
+    ctx_size += 1024*1024*16;
 
-    printf("Allocating Memory of size %li byes, %li MB\n",ctx_size, (ctx_size/1024/1024));
+    printf("Allocating Memory of size %li bytes, %li MB\n",ctx_size, (ctx_size/1024/1024));
 
     struct ggml_init_params params = {
         /*.mem_size   =*/ ctx_size,
@@ -217,7 +208,7 @@ int main(int argc, char ** argv)  {
     const int dimz = sizez;
     long long int flops_per_dot_product = dimy + dimy;
     long long int flops_per_matrix = flops_per_dot_product * dimx * dimz; ;
-    printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - aboout %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000);
+    printf("Matrix Multiplication of (%i,%i,%i) x (%i,%i,%i) - about %6.2f gFLOPS\n\n", sizex, sizey, 1, sizex, sizez, 1, 1.0f*flops_per_matrix / 1000 / 1000 / 1000);
 
 
     // Let's use the F32 result from above as a reference for the q4_0 multiplication
@@ -234,7 +225,6 @@ 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 sec = usec/1000000;
         float flops_per_usec = (1.0f*flops_per_matrix)/usec;
         printf("%9i;%8i;%6i;%6i;%6i;%15lli;%18lli;%19.2f\n",
             i,

examples/common.cpp

@@ -1,13 +1,18 @@
 #include "common.h"
 
 #include <cassert>
+#include <iostream>
 #include <cstring>
 #include <fstream>
 #include <string>
 #include <iterator>
 #include <algorithm>
 #include <sstream>
-#include <iostream>
+
+#if defined(__APPLE__) && defined(__MACH__)
+#include <sys/types.h>
+#include <sys/sysctl.h>
+#endif
 
 #if defined (_WIN32)
 #include <fcntl.h>
@@ -25,19 +30,43 @@ extern "C" __declspec(dllimport) int __stdcall WideCharToMultiByte(unsigned int
 #define CP_UTF8 65001
 #endif
 
-bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
-    // determine sensible default number of threads.
-    // std::thread::hardware_concurrency may not be equal to the number of cores, or may return 0.
+int32_t get_num_physical_cores() {
 #ifdef __linux__
     std::ifstream cpuinfo("/proc/cpuinfo");
-    params.n_threads = std::count(std::istream_iterator<std::string>(cpuinfo),
-                                  std::istream_iterator<std::string>(),
-                                  std::string("processor"));
-#endif
-    if (params.n_threads == 0) {
-        params.n_threads = std::max(1, (int32_t) std::thread::hardware_concurrency());
+    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
+                }
+            }
+        }
     }
+#elif defined(__APPLE__) && defined(__MACH__)
+    int32_t num_physical_cores;
+    size_t len = sizeof(num_physical_cores);
+    int result = sysctlbyname("hw.perflevel0.physicalcpu", &num_physical_cores, &len, NULL, 0);
+    if (result == 0) {
+        return num_physical_cores;
+    }
+    result = sysctlbyname("hw.physicalcpu", &num_physical_cores, &len, NULL, 0);
+    if (result == 0) {
+        return num_physical_cores;
+    }
+#elif defined(_WIN32)
+    //TODO: Implement
+#endif
+    unsigned int n_threads = std::thread::hardware_concurrency();
+    return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
+}
 
+bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
     bool invalid_param = false;
     std::string arg;
     gpt_params default_params;

examples/common.h

@@ -13,11 +13,12 @@
 //
 // CLI argument parsing
 //
+int32_t get_num_physical_cores();
 
 struct gpt_params {
     int32_t seed          = -1;   // RNG seed
-    int32_t n_threads     = std::min(4, (int32_t) std::thread::hardware_concurrency());
-    int32_t n_predict     = -1;   // new tokens to predict
+    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)

examples/main/main.cpp

@@ -161,23 +161,22 @@ int main(int argc, char ** argv) {
     std::vector<llama_token> session_tokens;
 
     if (!path_session.empty()) {
-        fprintf(stderr, "%s: attempting to load saved session from %s..\n", __func__, path_session.c_str());
+        fprintf(stderr, "%s: attempting to load saved session from '%s'\n", __func__, path_session.c_str());
 
-        // REVIEW - fopen to check for existing session
+        // fopen to check for existing session
         FILE * fp = std::fopen(path_session.c_str(), "rb");
         if (fp != NULL) {
             std::fclose(fp);
 
             session_tokens.resize(params.n_ctx);
             size_t n_token_count_out = 0;
-            const size_t n_session_bytes = llama_load_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out);
+            if (!llama_load_session_file(ctx, path_session.c_str(), session_tokens.data(), session_tokens.capacity(), &n_token_count_out)) {
+                fprintf(stderr, "%s: error: failed to load session file '%s'\n", __func__, path_session.c_str());
+                return 1;
+            }
             session_tokens.resize(n_token_count_out);
 
-            if (n_session_bytes > 0) {
-                fprintf(stderr, "%s: loaded %zu bytes of session data!\n", __func__, n_session_bytes);
-            } else {
-                fprintf(stderr, "%s: could not load session file, will recreate\n", __func__);
-            }
+            fprintf(stderr, "%s: loaded a session with prompt size of %d tokens\n", __func__, (int) session_tokens.size());
         } else {
             fprintf(stderr, "%s: session file does not exist, will create\n", __func__);
         }
@@ -214,7 +213,7 @@ int main(int argc, char ** argv) {
     }
 
     // number of tokens to keep when resetting context
-    if (params.n_keep < 0 || params.n_keep > (int)embd_inp.size() || params.instruct) {
+    if (params.n_keep < 0 || params.n_keep > (int) embd_inp.size() || params.instruct) {
         params.n_keep = (int)embd_inp.size();
     }
 
@@ -329,7 +328,7 @@ int main(int argc, char ** argv) {
                 // insert n_left/2 tokens at the start of embd from last_n_tokens
                 embd.insert(embd.begin(), last_n_tokens.begin() + n_ctx - n_left/2 - embd.size(), last_n_tokens.end() - embd.size());
 
-                // REVIEW - stop saving session if we run out of context
+                // stop saving session if we run out of context
                 path_session = "";
 
                 //printf("\n---\n");
@@ -355,6 +354,7 @@ int main(int argc, char ** argv) {
                     n_session_consumed++;
 
                     if (n_session_consumed >= (int) session_tokens.size()) {
+                        ++i;
                         break;
                     }
                 }

ggml-cuda.cu

@@ -355,8 +355,18 @@ cudaError_t ggml_cuda_h2d_tensor_2d(void * dst, const struct ggml_tensor * src,
 }
 
 void * ggml_cuda_host_malloc(size_t size) {
-    void * ptr;
-    CUDA_CHECK(cudaMallocHost((void **) &ptr, size));
+    if (getenv("GGML_CUDA_NO_PINNED") != nullptr) {
+        return nullptr;
+    }
+
+    void * ptr = nullptr;
+    cudaError_t err = cudaMallocHost((void **) &ptr, size);
+    if (err != cudaSuccess) {
+        fprintf(stderr, "WARNING: failed to allocate %.2f MB of pinned memory: %s\n",
+            size/1024.0/1024.0, cudaGetErrorString(err));
+        return nullptr;
+    }
+
     return ptr;
 }
 

ggml-opencl-dequant.cl

@@ -1,63 +0,0 @@
-#define MULTILINE_QUOTE(...) #__VA_ARGS__
-const char * clblast_dequant = MULTILINE_QUOTE(
-
-struct block_q4_0
-{
-    float d;
-    uchar qs[16];
-};
-
-__kernel void dequantize_row_q4_0(__global struct block_q4_0* blocks, __global float* result) {
-    const uint i = get_global_id(0) / 32;
-    const uint l = get_local_id(0);
-
-    const float d = blocks[i].d;
-
-    const uchar vi = blocks[i].qs[l];
-
-    const uint index = i*32 + l*2;
-    result[index + 0] = ((vi & 0xf) - 8)*d;
-    result[index + 1] = ((vi >> 4) - 8)*d;
-}
-
-struct block_q4_1
-{
-    float d;
-    float m;
-    uchar qs[16];
-};
-
-__kernel void dequantize_row_q4_1(__global struct block_q4_1* blocks, __global float* result) {
-    const uint i = get_global_id(0) / 32;
-    const uint l = get_local_id(0);
-
-    const float d = blocks[i].d;
-    const float m = blocks[i].m;
-
-    const uchar vi = blocks[i].qs[l];
-
-    const uint index = i*32 + l*2;
-    result[index + 0] = (vi & 0xf) * d + m;
-    result[index + 1] = (vi >> 4) * d + m;
-}
-
-struct block_q4_2
-{
-    ushort d;
-    uchar qs[8];
-};
-
-__kernel void dequantize_row_q4_2(__global struct block_q4_2* blocks, __global float* result) {
-    const uint i = get_global_id(0) / 16;
-    const uint l = get_local_id(0);
-
-    const float d = vload_half(0, (__global half*) &blocks[i].d);;
-
-    const uchar vi = blocks[i].qs[l];
-
-    const uint index = i*16 + l*2;
-    result[index + 0] = ((vi & 0xf) - 8)*d;
-    result[index + 1] = ((vi >> 4) - 8)*d;
-}
-
-);

ggml-opencl.c

@@ -3,12 +3,141 @@
 #define CL_TARGET_OPENCL_VERSION 110
 #include <clblast_c.h>
 
+#include <stdlib.h>
 #include <stdio.h>
 #include <string.h>
 
 #include "ggml.h"
 
-#include "ggml-opencl-dequant.cl"
+#define MULTILINE_QUOTE(...) #__VA_ARGS__
+const char * clblast_dequant = MULTILINE_QUOTE(
+
+struct block_q4_0
+{
+    float d;
+    uchar qs[16];
+};
+
+__kernel void dequantize_row_q4_0(__global struct block_q4_0* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 32;
+    const uint l = get_local_id(0);
+
+    const float d = blocks[i].d;
+
+    const uchar vi = blocks[i].qs[l];
+
+    const uint index = i*32 + l*2;
+    result[index + 0] = ((vi & 0xf) - 8)*d;
+    result[index + 1] = ((vi >> 4) - 8)*d;
+}
+
+struct block_q4_1
+{
+    float d;
+    float m;
+    uchar qs[16];
+};
+
+__kernel void dequantize_row_q4_1(__global struct block_q4_1* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 32;
+    const uint l = get_local_id(0);
+
+    const float d = blocks[i].d;
+    const float m = blocks[i].m;
+
+    const uchar vi = blocks[i].qs[l];
+
+    const uint index = i*32 + l*2;
+    result[index + 0] = (vi & 0xf) * d + m;
+    result[index + 1] = (vi >> 4) * d + m;
+}
+
+struct block_q4_2
+{
+    ushort d;
+    uchar qs[8];
+};
+
+__kernel void dequantize_row_q4_2(__global struct block_q4_2* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 16;
+    const uint l = get_local_id(0);
+
+    const float d = vload_half(0, (__global half*) &blocks[i].d);
+
+    const uchar vi = blocks[i].qs[l];
+
+    const uint index = i*16 + l*2;
+    result[index + 0] = ((vi & 0xf) - 8)*d;
+    result[index + 1] = ((vi >> 4) - 8)*d;
+}
+
+
+struct block_q5_0
+{
+    float d;
+    uint qh;
+    uchar qs[16];
+};
+
+__kernel void dequantize_row_q5_0(__global struct block_q5_0* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 32;
+    const uint l = get_local_id(0);
+
+    const float d = blocks[i].d;
+
+    const uchar vi = blocks[i].qs[l];
+
+    const uint l2 = l * 2;
+
+    const uchar vh0 = ((blocks[i].qh & (1 << (l2 + 0))) >> (l2 + 0)) << 4;
+    const uchar vh1 = ((blocks[i].qh & (1 << (l2 + 1))) >> (l2 + 1)) << 4;
+
+    const uint index = i*32 + l2;
+    result[index + 0] = (((vi & 0xf) | vh0) - 16)*d;
+    result[index + 1] = (((vi >>  4) | vh1) - 16)*d;
+}
+
+struct block_q5_1
+{
+    ushort d;
+    ushort m;
+    uint qh;
+    uchar qs[16];
+};
+
+__kernel void dequantize_row_q5_1(__global struct block_q5_1* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 32;
+    const uint l = get_local_id(0);
+
+    const float d = vload_half(0, (__global half*) &blocks[i].d);
+    const float m = vload_half(0, (__global half*) &blocks[i].m);
+
+    const uchar vi = blocks[i].qs[l];
+
+    const uint l2 = l * 2;
+
+    const uchar vh0 = ((blocks[i].qh & (1 << (l2 + 0))) >> (l2 + 0)) << 4;
+    const uchar vh1 = ((blocks[i].qh & (1 << (l2 + 1))) >> (l2 + 1)) << 4;
+
+    const uint index = i*32 + l2;
+    result[index + 0] = ((vi & 0xf) | vh0)*d + m;
+    result[index + 1] = ((vi >>  4) | vh1)*d + m;
+}
+
+struct block_q8_0
+{
+    float d;
+    char qs[32];
+};
+
+__kernel void dequantize_row_q8_0(__global struct block_q8_0* blocks, __global float* result) {
+    const uint i = get_global_id(0) / 32;
+    const uint l = get_local_id(0);
+
+    result[i*32 + l] = blocks[i].qs[l] * blocks[i].d;
+}
+
+);
 
 #define CL_CHECK(err, name)                                                                     \
     do {                                                                                        \
@@ -19,12 +148,26 @@
         }                                                                                       \
     } while (0)
 
+#define QK5_0 32
+typedef struct {
+    ggml_fp16_t d;         // delta
+    uint8_t qh[4];         // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2]; // nibbles / quants
+} block_q5_0;
+
+
+typedef struct {
+    float d;                // delta
+    uint32_t qh;          // 5-th bit of quants
+    uint8_t qs[QK5_0 / 2];  // nibbles / quants
+} cl_block_q5_0;
+
 static cl_platform_id platform;
 static cl_device_id device;
 static cl_context context;
 static cl_command_queue queue;
 static cl_program program;
-static cl_kernel kernel_q4_0, kernel_q4_1, kernel_q4_2;
+static cl_kernel kernel_q4_0, kernel_q4_1, kernel_q4_2, kernel_q5_0, kernel_q5_1, kernel_q8_0;
 static cl_mem cl_buffer_a, cl_buffer_qb, cl_buffer_b, cl_buffer_c;
 static size_t cl_size_a = 0, cl_size_qb = 0, cl_size_b = 0, cl_size_c = 0;
 
@@ -97,6 +240,12 @@ void ggml_cl_init(void) {
     CL_CHECK(err, "clCreateKernel");
     kernel_q4_2 = clCreateKernel(program, "dequantize_row_q4_2", &err);
     CL_CHECK(err, "clCreateKernel");
+    kernel_q5_0 = clCreateKernel(program, "dequantize_row_q5_0", &err);
+    CL_CHECK(err, "clCreateKernel");
+    kernel_q5_1 = clCreateKernel(program, "dequantize_row_q5_1", &err);
+    CL_CHECK(err, "clCreateKernel");
+    kernel_q8_0 = clCreateKernel(program, "dequantize_row_q8_0", &err);
+    CL_CHECK(err, "clCreateKernel");
 }
 
 static void ggml_cl_malloc(size_t req_size, size_t* cur_size, cl_mem_flags flags, cl_mem* buf) {
@@ -125,6 +274,7 @@ void ggml_cl_sgemm_wrapper(
     cl_kernel kernel;
     size_t global = n * k, local, size_qb;
     bool dequant;
+    cl_block_q5_0* cl_host_b;
 
     switch (btype) {
     case GGML_TYPE_F32:
@@ -146,7 +296,36 @@ void ggml_cl_sgemm_wrapper(
         dequant = true;
         kernel = kernel_q4_2;
         local = 8;
-        size_qb = global * (sizeof(short) + local) / 16;
+        size_qb = global * (sizeof(ggml_fp16_t) + local) / 16;
+        break;
+    case GGML_TYPE_Q5_0:
+        dequant = true;
+        kernel = kernel_q5_0;
+        local = 16;
+        // For some reason OpenCL seems to be incapable of working with structs of size 22.
+        // 20 and 24 bytes are fine. Workaround to do the fp16 to fp32 step on CPU...
+        // TODO Find the reason, fix and remove workaround.
+        const block_q5_0* b = (const block_q5_0*) host_b;
+        cl_host_b = (cl_block_q5_0*) malloc(sizeof(cl_block_q5_0) * global / 32);
+        for (size_t i = 0; i < global / 32; i++) {
+            cl_host_b[i].d = ggml_fp16_to_fp32(b[i].d);
+            memcpy(&cl_host_b[i].qh, b[i].qh, sizeof(uint32_t));
+            memcpy(&cl_host_b[i].qs, b[i].qs, QK5_0 / 2);
+        }
+        host_b = (const float*) cl_host_b;
+        size_qb = global * (sizeof(float) + sizeof(uint32_t) + local) / 32;
+        break;
+    case GGML_TYPE_Q5_1:
+        dequant = true;
+        kernel = kernel_q5_1;
+        local = 16;
+        size_qb = global * (sizeof(ggml_fp16_t) * 2 + sizeof(uint32_t) + local) / 32;
+        break;
+    case GGML_TYPE_Q8_0:
+        dequant = true;
+        kernel = kernel_q8_0;
+        local = 32;
+        size_qb = global * (sizeof(float) + local) / 32;
         break;
     default:
         fprintf(stderr, "Error: Unsupported OpenCL btype %d\n", btype);
@@ -171,12 +350,15 @@ void ggml_cl_sgemm_wrapper(
         err = clSetKernelArg(kernel, 0, sizeof(cl_mem), &cl_buffer_qb);
         err |= clSetKernelArg(kernel, 1, sizeof(cl_mem), &cl_buffer_b);
         CL_CHECK(err, "clSetKernelArg");
-        clEnqueueWriteBuffer(queue, cl_buffer_qb, CL_FALSE, 0, size_qb, host_b, 0, NULL, &ev_qb);
+        err = clEnqueueWriteBuffer(queue, cl_buffer_qb, CL_FALSE, 0, size_qb, host_b, 0, NULL, &ev_qb);
+        CL_CHECK(err, "clEnqueueWriteBuffer qb");
     } else {
-        clEnqueueWriteBuffer(queue, cl_buffer_b, CL_FALSE, 0, size_b, host_b, 0, NULL, &ev_b);
+        err = clEnqueueWriteBuffer(queue, cl_buffer_b, CL_FALSE, 0, size_b, host_b, 0, NULL, &ev_b);
+        CL_CHECK(err, "clEnqueueWriteBuffer b");
     }
 
-    clEnqueueWriteBuffer(queue, cl_buffer_a, CL_FALSE, 0, size_a, host_a, 0, NULL, &ev_a);
+    err = clEnqueueWriteBuffer(queue, cl_buffer_a, CL_FALSE, 0, size_a, host_a, 0, NULL, &ev_a);
+    CL_CHECK(err, "clEnqueueWriteBuffer a");
     if (dequant) {
         err = clEnqueueNDRangeKernel(queue, kernel, 1, NULL, &global, &local, 1, &ev_qb, &ev_b);
         CL_CHECK(err, "clEnqueueNDRangeKernel");
@@ -188,15 +370,20 @@ void ggml_cl_sgemm_wrapper(
     clReleaseEvent(ev_b);
 
     cl_event ev_sgemm;
-    CLBlastSgemm((CLBlastLayout)order,
-                 (CLBlastTranspose)trans_a, (CLBlastTranspose)trans_b,
-                 m, n, k,
-                 alpha,
-                 cl_buffer_a, 0, lda,
-                 cl_buffer_b, 0, ldb,
-                 beta,
-                 cl_buffer_c, 0, ldc,
-                 &queue, &ev_sgemm);
+    CLBlastStatusCode status = CLBlastSgemm((CLBlastLayout)order,
+                                            (CLBlastTranspose)trans_a, (CLBlastTranspose)trans_b,
+                                            m, n, k,
+                                            alpha,
+                                            cl_buffer_a, 0, lda,
+                                            cl_buffer_b, 0, ldb,
+                                            beta,
+                                            cl_buffer_c, 0, ldc,
+                                            &queue, &ev_sgemm);
+
+    if (status != CLBlastSuccess) {
+        fprintf(stderr, "Error: CLBlast SGEMM %d\n", status);
+        abort();
+    }
 
     cl_event ev_c;
     clEnqueueReadBuffer(queue, cl_buffer_c, CL_TRUE, 0, size_c, host_c, 1, &ev_sgemm, &ev_c);
@@ -205,4 +392,7 @@ void ggml_cl_sgemm_wrapper(
     clWaitForEvents(1, &ev_c);
     clReleaseEvent(ev_sgemm);
     clReleaseEvent(ev_c);
+    if (btype == GGML_TYPE_Q5_0) {
+        free((void*) cl_host_b);
+    }
 }

ggml.c

@@ -330,7 +330,7 @@ static ggml_fp16_t table_exp_f16[1 << 16];
 // precomputed f32 table for f16 (256 KB)
 static float table_f32_f16[1 << 16];
 
-#if defined(__ARM_NEON)
+#if defined(__ARM_NEON) || defined(__wasm_simd128__)
 #define B1(c,s,n)  0x ## n ## c ,  0x ## n ## s
 #define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
 #define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
@@ -668,6 +668,33 @@ uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
     return vget_high_u8(vcombine_u8(a, b));
 }
 
+int8x16_t vzip1q_s8(int8x16_t a, int8x16_t b) {
+    return vcombine_s8(vget_low_s8(a), vget_low_s8(b));
+}
+
+int8x16_t vzip2q_s8(int8x16_t a, int8x16_t b) {
+    return vcombine_s8(vget_high_s8(a), vget_high_s8(b));
+}
+
+uint8x16_t vzip1q_u8(uint8x16_t a, uint8x16_t b) {
+    return vcombine_u8(vget_low_u8(a), vget_low_u8(b));
+}
+
+uint8x16_t vzip2q_u8(uint8x16_t a, uint8x16_t b) {
+    return vcombine_u8(vget_high_u8(a), vget_high_u8(b));
+}
+
+int32x4_t vcvtnq_s32_f32(float32x4_t v) {
+    int32x4_t res;
+
+    res[0] = roundf(vgetq_lane_f32(v, 0));
+    res[1] = roundf(vgetq_lane_f32(v, 1));
+    res[2] = roundf(vgetq_lane_f32(v, 2));
+    res[3] = roundf(vgetq_lane_f32(v, 3));
+
+    return res;
+}
+
 #endif
 #endif
 
@@ -1060,7 +1087,7 @@ static void quantize_row_q4_0(const float * restrict x, void * restrict vy, int
             const v128_t v  = wasm_f32x4_mul(srcv[l], wasm_f32x4_splat(id));
             const v128_t vf = wasm_f32x4_add(v, wasm_f32x4_splat(8.5f));
             const v128_t vi = wasm_i32x4_trunc_sat_f32x4(vf);
-            const v128_t vc = wasm_i32x4_min_u(vi, wasm_i32x4_splat(15));
+            const v128_t vc = wasm_i32x4_min(vi, wasm_i32x4_splat(15));
 
             y[i].qs[2*l + 0] = wasm_i32x4_extract_lane(vc, 0) | (wasm_i32x4_extract_lane(vc, 1) << 4);
             y[i].qs[2*l + 1] = wasm_i32x4_extract_lane(vc, 2) | (wasm_i32x4_extract_lane(vc, 3) << 4);
@@ -1884,8 +1911,8 @@ static void dequantize_row_q5_0(const void * restrict vx, float * restrict y, in
             const uint8_t vi = pp[l/2];
 
             // extract the 5-th bit from qh
-            const uint8_t vh0 = ((qh & (1 << (l + 0))) >> (l + 0)) << 4;
-            const uint8_t vh1 = ((qh & (1 << (l + 1))) >> (l + 1)) << 4;
+            const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4;
+            const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4;
 
             const int8_t vi0 = (vi & 0x0F) | vh0;
             const int8_t vi1 = (vi >>   4) | vh1;
@@ -1921,8 +1948,8 @@ static void dequantize_row_q5_1(const void * restrict vx, float * restrict y, in
             const uint8_t vi = pp[l/2];
 
             // extract the 5-th bit from qh
-            const uint8_t vh0 = ((qh & (1 << (l + 0))) >> (l + 0)) << 4;
-            const uint8_t vh1 = ((qh & (1 << (l + 1))) >> (l + 1)) << 4;
+            const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4;
+            const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4;
 
             const uint8_t vi0 = (vi & 0x0F) | vh0;
             const uint8_t vi1 = (vi >>   4) | vh1;
@@ -2658,35 +2685,35 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
         const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b);
         const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b);
 
+        // interleave
+        const int8x16_t v0_0lz = vzip1q_s8(v0_0ls, v0_0hs);
+        const int8x16_t v0_0hz = vzip2q_s8(v0_0ls, v0_0hs);
+        const int8x16_t v0_1lz = vzip1q_s8(v0_1ls, v0_1hs);
+        const int8x16_t v0_1hz = vzip2q_s8(v0_1ls, v0_1hs);
+
         // load y
         const int8x16_t v1_0l = vld1q_s8(y0->qs);
         const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
         const int8x16_t v1_1l = vld1q_s8(y1->qs);
         const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
 
-        // interleave
-        const int8x16_t v1_0ls = vuzp1q_s8(v1_0l, v1_0h);
-        const int8x16_t v1_0hs = vuzp2q_s8(v1_0l, v1_0h);
-        const int8x16_t v1_1ls = vuzp1q_s8(v1_1l, v1_1h);
-        const int8x16_t v1_1hs = vuzp2q_s8(v1_1l, v1_1h);
-
 #if defined(__ARM_FEATURE_DOTPROD)
         // dot product into int32x4_t
-        const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0ls), v0_0hs, v1_0hs);
-        const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1ls), v0_1hs, v1_1hs);
+        const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l), v0_0hz, v1_0h);
+        const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l), v0_1hz, v1_1h);
 
         sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d);
         sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d);
 #else
-        const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0ls));
-        const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0ls));
-        const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0hs));
-        const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0hs));
+        const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l));
+        const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l));
+        const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h));
+        const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h));
 
-        const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1ls));
-        const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1ls));
-        const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1hs));
-        const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1hs));
+        const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l));
+        const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l));
+        const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h));
+        const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h));
 
         const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
         const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
@@ -3153,6 +3180,72 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
     }
 
     *s = vaddvq_f32(sumv);
+#elif defined(__wasm_simd128__)
+    v128_t sumv = wasm_f32x4_splat(0.0f);
+
+    uint64_t tmp[4];
+
+    for (int i = 0; i < nb; ++i) {
+        const block_q5_0 * restrict x0 = &x[i];
+        const block_q8_0 * restrict y0 = &y[i];
+
+        const v128_t m4b  = wasm_i8x16_splat(0x0F);
+        const v128_t s16b = wasm_i8x16_splat(0x10);
+
+        // extract the 5th bit
+        uint32_t qh;
+        memcpy(&qh, x0->qh, sizeof(qh));
+
+        tmp[0] = table_b2b_u[(qh >>  0) & 0xFF];
+        tmp[1] = table_b2b_u[(qh >>  8) & 0xFF];
+        tmp[2] = table_b2b_u[(qh >> 16) & 0xFF];
+        tmp[3] = table_b2b_u[(qh >> 24)       ];
+
+        const v128_t qhl = wasm_v128_load(tmp + 0);
+        const v128_t qhh = wasm_v128_load(tmp + 2);
+
+        const v128_t v0 = wasm_v128_load(x0->qs);
+
+        // 4-bit -> 8-bit
+        const v128_t v0l = wasm_v128_and (v0, m4b);
+        const v128_t v0h = wasm_u8x16_shr(v0, 4);
+
+        // interleave
+        const v128_t v0lz = wasm_v8x16_shuffle(v0l, v0h,  0, 16,  1, 17,  2, 18,  3, 19,  4, 20,  5, 21,  6, 22,  7, 23);
+        const v128_t v0hz = wasm_v8x16_shuffle(v0l, v0h,  8, 24,  9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31);
+
+        // add high bit and sub 16
+        const v128_t v0lf = wasm_i8x16_sub(wasm_v128_or(v0lz, qhl), s16b);
+        const v128_t v0hf = wasm_i8x16_sub(wasm_v128_or(v0hz, qhh), s16b);
+
+        // load y
+        const v128_t v1l = wasm_v128_load(y0->qs);
+        const v128_t v1h = wasm_v128_load(y0->qs + 16);
+
+        // int8x16 -> int16x8
+        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
+        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
+        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
+        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
+
+        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
+        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
+        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
+        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
+
+        const float x0d = GGML_FP16_TO_FP32(x0->d);
+
+        // dot product
+        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
+                        wasm_i32x4_add(
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
+                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
+                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_f32x4_splat(x0d*y0->d)));
+    }
+
+    *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
+         wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
 #elif defined(__AVX2__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -3193,8 +3286,8 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
         for (int j = 0; j < QK8_0/2; j++) {
             const uint8_t v0 = x0[j];
 
-            const int x0_0h = ((qh & (1 << (2*j + 0))) >> (2*j + 0)) << 4;
-            const int x1_0h = ((qh & (1 << (2*j + 1))) >> (2*j + 1)) << 4;
+            const int x0_0h = ((qh & (1u << (2*j + 0))) >> (2*j + 0)) << 4;
+            const int x1_0h = ((qh & (1u << (2*j + 1))) >> (2*j + 1)) << 4;
 
             const int x0_0 = ((v0 & 0x0F) | x0_0h) - 16;
             const int x1_0 = ((v0 >>   4) | x1_0h) - 16;
@@ -3284,6 +3377,77 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
     }
 
     *s = vaddvq_f32(sumv) + summs;
+#elif defined(__wasm_simd128__)
+    v128_t sumv = wasm_f32x4_splat(0.0f);
+
+    float summs = 0.0f;
+
+    uint64_t tmp[4];
+
+    for (int i = 0; i < nb; ++i) {
+        const block_q5_1 * restrict x0 = &x[i];
+        const block_q8_1 * restrict y0 = &y[i];
+
+        summs += GGML_FP16_TO_FP32(x0->m) * (y0->s0 + y0->s1);
+
+        const v128_t m4b = wasm_i8x16_splat(0x0F);
+
+        // extract the 5th bit
+        uint32_t qh;
+        memcpy(&qh, x0->qh, sizeof(qh));
+
+        tmp[0] = table_b2b_u[(qh >>  0) & 0xFF];
+        tmp[1] = table_b2b_u[(qh >>  8) & 0xFF];
+        tmp[2] = table_b2b_u[(qh >> 16) & 0xFF];
+        tmp[3] = table_b2b_u[(qh >> 24)       ];
+
+        const v128_t qhl = wasm_v128_load(tmp + 0);
+        const v128_t qhh = wasm_v128_load(tmp + 2);
+
+        const v128_t v0 = wasm_v128_load(x0->qs);
+
+        // 4-bit -> 8-bit
+        const v128_t v0l = wasm_v128_and (v0, m4b);
+        const v128_t v0h = wasm_u8x16_shr(v0, 4);
+
+        static bool x = true;
+
+        // interleave
+        const v128_t v0lz = wasm_v8x16_shuffle(v0l, v0h,  0, 16,  1, 17,  2, 18,  3, 19,  4, 20,  5, 21,  6, 22,  7, 23);
+        const v128_t v0hz = wasm_v8x16_shuffle(v0l, v0h,  8, 24,  9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31);
+
+        // add high bit
+        const v128_t v0lf = wasm_v128_or(v0lz, qhl);
+        const v128_t v0hf = wasm_v128_or(v0hz, qhh);
+
+        // load y
+        const v128_t v1l = wasm_v128_load(y0->qs);
+        const v128_t v1h = wasm_v128_load(y0->qs + 16);
+
+        // int8x16 -> int16x8
+        const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
+        const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
+        const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
+        const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
+
+        const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
+        const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
+        const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
+        const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
+
+        const float x0d = GGML_FP16_TO_FP32(x0->d);
+
+        // dot product
+        sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
+                        wasm_i32x4_add(
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
+                                           wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
+                            wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
+                                           wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_f32x4_splat(x0d*y0->d)));
+    }
+
+    *s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
+         wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs;
 #elif defined(__AVX2__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -3327,8 +3491,8 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
         for (int j = 0; j < QK8_1/2; j++) {
             const uint8_t v0 = x0[j];
 
-            const int x0_0h = ((qh & (1 << (2*j + 0))) >> (2*j + 0)) << 4;
-            const int x1_0h = ((qh & (1 << (2*j + 1))) >> (2*j + 1)) << 4;
+            const int x0_0h = ((qh & (1u << (2*j + 0))) >> (2*j + 0)) << 4;
+            const int x1_0h = ((qh & (1u << (2*j + 1))) >> (2*j + 1)) << 4;
 
             const int x0_0 = (v0 & 0x0F) | x0_0h;
             const int x1_0 = (v0 >>   4) | x1_0h;
@@ -3800,6 +3964,7 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = {
     "DIAG_MASK_INF",
     "SOFT_MAX",
     "ROPE",
+    "ALIBI",
     "CONV_1D_1S",
     "CONV_1D_2S",
 
@@ -3848,6 +4013,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "diag_mask_inf(x)",
     "soft_max(x)",
     "rope(x)",
+    "alibi(x)",
     "conv_1d_1s(x)",
     "conv_1d_2s(x)",
 
@@ -4028,6 +4194,27 @@ bool ggml_is_quantized(enum ggml_type type) {
     return GGML_IS_QUANTIZED[type];
 }
 
+enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
+    enum ggml_type wtype = GGML_TYPE_COUNT;
+
+    switch (ftype) {
+        case GGML_FTYPE_ALL_F32:              wtype = GGML_TYPE_F32;   break;
+        case GGML_FTYPE_MOSTLY_F16:           wtype = GGML_TYPE_F16;   break;
+        case GGML_FTYPE_MOSTLY_Q4_0:          wtype = GGML_TYPE_Q4_0;  break;
+        case GGML_FTYPE_MOSTLY_Q4_1:          wtype = GGML_TYPE_Q4_1;  break;
+        case GGML_FTYPE_MOSTLY_Q4_2:          wtype = GGML_TYPE_Q4_2;  break;
+        case GGML_FTYPE_MOSTLY_Q5_0:          wtype = GGML_TYPE_Q5_0;  break;
+        case GGML_FTYPE_MOSTLY_Q5_1:          wtype = GGML_TYPE_Q5_1;  break;
+        case GGML_FTYPE_MOSTLY_Q8_0:          wtype = GGML_TYPE_Q8_0;  break;
+        case GGML_FTYPE_UNKNOWN:              wtype = GGML_TYPE_COUNT; break;
+        case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
+    }
+
+    GGML_ASSERT(wtype != GGML_TYPE_COUNT);
+
+    return wtype;
+}
+
 static inline bool ggml_is_transposed(const struct ggml_tensor * tensor) {
     return tensor->nb[0] > tensor->nb[1];
 }
@@ -4224,7 +4411,7 @@ void ggml_free(struct ggml_context * ctx) {
 }
 
 size_t ggml_used_mem(const struct ggml_context * ctx) {
-    return ctx->objects_end->offs + ctx->objects_end->size;
+    return ctx->objects_end == NULL ? 0 : ctx->objects_end->offs + ctx->objects_end->size;
 }
 
 size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch) {
@@ -8245,8 +8432,6 @@ static void ggml_compute_forward_mul_mat_f16_f32(
         ggml_fp16_t * d_X = ggml_cuda_pool_malloc(sizeof(float) * x_ne, &x_size);
         ggml_fp16_t * d_Y = ggml_cuda_pool_malloc(sizeof(float) * y_ne, &y_size);
         float       * d_D = ggml_cuda_pool_malloc(sizeof(float) * d_ne, &d_size);
-#else
-        float * const wdata = params->wdata;
 #endif
         for (int64_t i03 = 0; i03 < ne03; i03++) {
             for (int64_t i02 = 0; i02 < ne02; i02++) {
@@ -8263,8 +8448,11 @@ static void ggml_compute_forward_mul_mat_f16_f32(
                             wdata[id++] = GGML_FP32_TO_FP16(*(float *) ((char *) src1->data + i03*nb13 + i02*nb12 + i01*nb11 + i00*nb10));
                         }
                     }
+
+                    assert(id*sizeof(ggml_fp16_t) <= params->wsize);
                 }
 #else
+                float * const wdata = params->wdata;
                 {
                     size_t id = 0;
                     for (int64_t i01 = 0; i01 < ne01; ++i01) {
@@ -8272,6 +8460,8 @@ static void ggml_compute_forward_mul_mat_f16_f32(
                             wdata[id++] = GGML_FP16_TO_FP32(*(ggml_fp16_t *) ((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01 + i00*nb00));
                         }
                     }
+
+                    assert(id*sizeof(float) <= params->wsize);
                 }
 #endif
 
@@ -8537,7 +8727,10 @@ static void ggml_compute_forward_mul_mat_q_f32(
                         dequantize_row_q((char *) src0->data + i03*nb03 + i02*nb02 + i01*nb01, wdata + id, ne00);
                         id += ne00;
                     }
+
+                    assert(id*sizeof(float) <= params->wsize);
                 }
+
                 const float * x = wdata;
 #endif
 
@@ -9118,7 +9311,7 @@ static void ggml_compute_forward_alibi_f32(
     //const int nb3 = src0->nb[3];
 
     assert(nb0 == sizeof(float));
-    assert(ne1+n_past == ne0);
+    assert(ne1 + n_past == ne0); (void) n_past;
 
     // add alibi to src0 (KQ_scaled)
     const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
@@ -9179,7 +9372,7 @@ static void ggml_compute_forward_alibi_f16(
     //const int nb3 = src0->nb[3];
 
     assert(nb0 == sizeof(ggml_fp16_t));
-    assert(ne1+n_past == ne0);
+    assert(ne1 + n_past == ne0); (void) n_past;
 
     // add alibi to src0 (KQ_scaled)
     const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
@@ -11571,10 +11764,13 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
                             if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
                                 node->n_tasks = 1; // TODO: this actually is doing nothing
                                                    //       the threads are still spinning
-                                cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*MAX(ggml_nelements(node->src1), ggml_nelements(node->src0));
-                                //printf("src0: ne0 = %d, ne1 = %d, ne = %d\n", node->src0->ne[0], node->src0->ne[1], node->src0->ne[0]*node->src0->ne[1]);
-                                //printf("src1: ne0 = %d, ne1 = %d, ne = %d\n", node->src1->ne[0], node->src1->ne[1], node->src1->ne[0]*node->src1->ne[1]);
-                                //printf("cur = %zu\n", cur);
+#if defined(GGML_USE_CUBLAS)
+                                // with cuBLAS, we need memory for the full 3D / 4D data of src1
+                                cur = GGML_TYPE_SIZE[GGML_TYPE_F16]*ggml_nelements(node->src1);
+#else
+                                // here we need memory just for single 2D matrix from src0
+                                cur = GGML_TYPE_SIZE[GGML_TYPE_F32]*(node->src0->ne[0]*node->src0->ne[1]);
+#endif
                             } else {
                                 cur = GGML_TYPE_SIZE[GGML_TYPE_F16]*ggml_nelements(node->src1);
                             }
@@ -11583,7 +11779,7 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
 #endif
                         } else if (node->src0->type == GGML_TYPE_F32 && node->src1->type == GGML_TYPE_F32) {
                             cur = 0;
-#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS)
+#if defined(GGML_USE_ACCELERATE) || defined(GGML_USE_OPENBLAS) || defined(GGML_USE_CUBLAS) || defined(GGML_USE_CLBLAST)
                             if (ggml_compute_forward_mul_mat_use_blas(node->src0, node->src1, node)) {
                                 node->n_tasks = 1;
                             }
@@ -12861,8 +13057,8 @@ size_t ggml_quantize_q5_0(const float * src, void * dst, int n, int k, int64_t *
             memcpy(&qh, &y[i].qh, sizeof(qh));
 
             for (int l = 0; l < QK5_0; l += 2) {
-                const uint8_t vh0 = ((qh & (1 << (l + 0))) >> (l + 0)) << 4;
-                const uint8_t vh1 = ((qh & (1 << (l + 1))) >> (l + 1)) << 4;
+                const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4;
+                const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4;
 
                 // cast to 16 bins
                 const uint8_t vi0 = ((y[i].qs[l/2] & 0x0F) | vh0) / 2;
@@ -12891,8 +13087,8 @@ size_t ggml_quantize_q5_1(const float * src, void * dst, int n, int k, int64_t *
             memcpy(&qh, &y[i].qh, sizeof(qh));
 
             for (int l = 0; l < QK5_1; l += 2) {
-                const uint8_t vh0 = ((qh & (1 << (l + 0))) >> (l + 0)) << 4;
-                const uint8_t vh1 = ((qh & (1 << (l + 1))) >> (l + 1)) << 4;
+                const uint8_t vh0 = ((qh & (1u << (l + 0))) >> (l + 0)) << 4;
+                const uint8_t vh1 = ((qh & (1u << (l + 1))) >> (l + 1)) << 4;
 
                 // cast to 16 bins
                 const uint8_t vi0 = ((y[i].qs[l/2] & 0x0F) | vh0) / 2;

ggml.h

@@ -232,6 +232,20 @@ extern "C" {
         GGML_TYPE_COUNT,
     };
 
+    // model file types
+    enum ggml_ftype {
+        GGML_FTYPE_UNKNOWN     = -1,
+        GGML_FTYPE_ALL_F32     = 0,
+        GGML_FTYPE_MOSTLY_F16  = 1,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_0 = 2,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_1 = 3,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
+        GGML_FTYPE_MOSTLY_Q4_2 = 5,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q8_0 = 7,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_0 = 8,  // except 1d tensors
+        GGML_FTYPE_MOSTLY_Q5_1 = 9,  // except 1d tensors
+    };
+
     // available tensor operations:
     enum ggml_op {
         GGML_OP_NONE = 0,
@@ -385,6 +399,9 @@ extern "C" {
 
     GGML_API bool    ggml_is_quantized(enum ggml_type type);
 
+    // TODO: temporary until model loading of ggml examples is refactored
+    GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype);
+
     // main
 
     GGML_API struct ggml_context * ggml_init(struct ggml_init_params params);
@@ -701,8 +718,8 @@ extern "C" {
             struct ggml_tensor  * c1);
 
     // Mapping operations
-    GGML_API typedef void (*ggml_unary_op_f32_t)(const int, float *, const float *);
-    GGML_API typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
+    typedef void (*ggml_unary_op_f32_t)(const int, float *, const float *);
+    typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
 
     GGML_API struct ggml_tensor * ggml_map_unary_f32(
             struct ggml_context        * ctx,

llama-util.h

@@ -243,7 +243,8 @@ struct llama_mmap {
 #else
     static constexpr bool SUPPORTED = false;
 
-    llama_mmap(struct llama_file *) {
+    llama_mmap(struct llama_file *, bool prefetch = true) {
+        (void)prefetch;
         throw std::string("mmap not supported");
     }
 #endif
@@ -382,8 +383,13 @@ struct llama_mlock {
 #else
     static constexpr bool SUPPORTED = false;
 
-    void raw_lock(const void * addr, size_t size) {
+    size_t lock_granularity() {
+        return (size_t) 65536;
+    }
+
+    bool raw_lock(const void * addr, size_t size) {
         fprintf(stderr, "warning: mlock not supported on this system\n");
+        return false;
     }
 
     void raw_unlock(const void * addr, size_t size) {}
@@ -395,6 +401,8 @@ struct llama_buffer {
     uint8_t * addr = NULL;
     size_t size = 0;
 
+    llama_buffer() = default;
+
     void resize(size_t size) {
         delete[] addr;
         addr = new uint8_t[size];
@@ -404,27 +412,59 @@ struct llama_buffer {
     ~llama_buffer() {
         delete[] addr;
     }
+
+    // disable copy and move
+    llama_buffer(const llama_buffer&) = delete;
+    llama_buffer(llama_buffer&&) = delete;
+    llama_buffer& operator=(const llama_buffer&) = delete;
+    llama_buffer& operator=(llama_buffer&&) = delete;
 };
 
 #ifdef GGML_USE_CUBLAS
 #include "ggml-cuda.h"
 struct llama_ctx_buffer {
     uint8_t * addr = NULL;
+    bool is_cuda;
     size_t size = 0;
 
+    llama_ctx_buffer() = default;
+
     void resize(size_t size) {
-        if (addr) {
-            ggml_cuda_host_free(addr);
-        }
+        free();
+
         addr = (uint8_t *) ggml_cuda_host_malloc(size);
+        if (addr) {
+            is_cuda = true;
+        }
+        else {
+            // fall back to pageable memory
+            addr = new uint8_t[size];
+            is_cuda = false;
+        }
         this->size = size;
     }
 
-    ~llama_ctx_buffer() {
+    void free() {
         if (addr) {
-            ggml_cuda_host_free(addr);
+            if (is_cuda) {
+                ggml_cuda_host_free(addr);
+            }
+            else {
+                delete[] addr;
+            }
         }
+        addr = NULL;
     }
+
+    ~llama_ctx_buffer() {
+        free();
+    }
+
+    // disable copy and move
+    llama_ctx_buffer(const llama_ctx_buffer&) = delete;
+    llama_ctx_buffer(llama_ctx_buffer&&) = delete;
+    llama_ctx_buffer& operator=(const llama_ctx_buffer&) = delete;
+    llama_ctx_buffer& operator=(llama_ctx_buffer&&) = delete;
 };
 #else
 typedef llama_buffer llama_ctx_buffer;

llama.cpp

@@ -727,8 +727,7 @@ struct llama_model_loader {
             LLAMA_ASSERT(offset == lt.size);
         } else if (lt.split_type == SPLIT_BY_COLUMNS) {
             // Let's load the data into temporary buffers to ensure the OS performs large loads.
-            std::vector<llama_buffer> tmp_bufs;
-            tmp_bufs.resize(lt.shards.size());
+            std::vector<llama_buffer> tmp_bufs(lt.shards.size());
             for (size_t i = 0; i < lt.shards.size(); i++) {
                 llama_load_tensor_shard & shard = lt.shards.at(i);
                 llama_file & file = file_loaders.at(shard.file_idx)->file;
@@ -780,7 +779,7 @@ static bool kv_cache_init(
     const int n_embd  = hparams.n_embd;
     const int n_layer = hparams.n_layer;
 
-    const int64_t n_mem      = (int64_t)n_layer*n_ctx;
+    const int64_t n_mem      = n_layer*n_ctx;
     const int64_t n_elements = n_embd*n_mem;
 
     cache.buf.resize(2u*n_elements*ggml_type_size(wtype) + 2u*MB);
@@ -2373,7 +2372,7 @@ int llama_apply_lora_from_file(struct llama_context * ctx, const char * path_lor
     }
 }
 
-int llama_get_kv_cache_token_count(struct llama_context * ctx) {
+int llama_get_kv_cache_token_count(const struct llama_context * ctx) {
     return ctx->model.kv_self.n;
 }
 
@@ -2387,7 +2386,7 @@ void llama_set_rng_seed(struct llama_context * ctx, int seed) {
 }
 
 // Returns the size of the state
-size_t llama_get_state_size(struct llama_context * ctx) {
+size_t llama_get_state_size(const struct llama_context * ctx) {
     // we don't know size of rng until we actually serialize it. so reserve more than enough memory for its serialized state.
     // for reference, std::mt19937(1337) serializes to 6701 bytes.
     const size_t s_rng_size        = sizeof(size_t);
@@ -2567,6 +2566,85 @@ size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src) {
     return nread;
 }
 
+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) {
+    llama_file file(path_session, "rb");
+
+    // sanity checks
+    {
+        const uint32_t magic   = file.read_u32();
+        const uint32_t version = file.read_u32();
+
+        if (!(magic == LLAMA_SESSION_MAGIC && version == LLAMA_SESSION_VERSION)) {
+            fprintf(stderr, "%s : unknown (magic, version) for session file: %08x, %08x\n", __func__, magic, version);
+            return false;
+        }
+
+        llama_hparams session_hparams;
+        file.read_raw(&session_hparams, sizeof(llama_hparams));
+
+        if (session_hparams != ctx->model.hparams) {
+            fprintf(stderr, "%s : model hparams didn't match from session file!\n", __func__);
+            return false;
+        }
+    }
+
+    // load the prompt
+    {
+        const uint32_t n_token_count = file.read_u32();
+
+        if (n_token_count > n_token_capacity) {
+            fprintf(stderr, "%s : token count in session file exceeded capacity! %u > %zu\n", __func__, n_token_count, n_token_capacity);
+            return false;
+        }
+
+        file.read_raw(tokens_out, sizeof(llama_token) * n_token_count);
+        *n_token_count_out = n_token_count;
+    }
+
+    // restore the context state
+    {
+        const size_t n_state_size_cur = file.size - file.tell();
+        const size_t n_state_size_exp = llama_get_state_size(ctx);
+
+        if (n_state_size_cur != n_state_size_exp) {
+            fprintf(stderr, "%s : the state size in session file didn't match! expected %zu, got %zu\n", __func__, n_state_size_exp, n_state_size_cur);
+            return false;
+        }
+
+        std::vector<uint8_t> state_data(n_state_size_cur);
+        file.read_raw(state_data.data(), n_state_size_cur);
+
+        llama_set_state_data(ctx, state_data.data());
+    }
+
+    return true;
+}
+
+bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) {
+    llama_file file(path_session, "wb");
+
+    file.write_u32(LLAMA_SESSION_MAGIC);
+    file.write_u32(LLAMA_SESSION_VERSION);
+
+    file.write_raw(&ctx->model.hparams, sizeof(llama_hparams));
+
+    // save the prompt
+    file.write_u32((uint32_t) n_token_count);
+    file.write_raw(tokens, sizeof(llama_token) * n_token_count);
+
+    // save the context state
+    {
+        const size_t n_state_size = llama_get_state_size(ctx);
+
+        std::vector<uint8_t> state_data(n_state_size);
+        llama_copy_state_data(ctx, state_data.data());
+
+        file.write_raw(state_data.data(), n_state_size);
+    }
+
+    return true;
+}
+
 int llama_eval(
         struct llama_context * ctx,
            const llama_token * tokens,
@@ -2605,15 +2683,15 @@ int llama_tokenize(
     return res.size();
 }
 
-int llama_n_vocab(struct llama_context * ctx) {
+int llama_n_vocab(const struct llama_context * ctx) {
     return ctx->vocab.id_to_token.size();
 }
 
-int llama_n_ctx(struct llama_context * ctx) {
+int llama_n_ctx(const struct llama_context * ctx) {
     return ctx->model.hparams.n_ctx;
 }
 
-int llama_n_embd(struct llama_context * ctx) {
+int llama_n_embd(const struct llama_context * ctx) {
     return ctx->model.hparams.n_embd;
 }
 
@@ -2625,7 +2703,7 @@ float * llama_get_embeddings(struct llama_context * ctx) {
     return ctx->embedding.data();
 }
 
-const char * llama_token_to_str(struct llama_context * ctx, llama_token token) {
+const char * llama_token_to_str(const struct llama_context * ctx, llama_token token) {
     if (token >= llama_n_vocab(ctx)) {
         return nullptr;
     }
@@ -2694,57 +2772,3 @@ const char * llama_print_system_info(void) {
 std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx) {
     return ctx->model.tensors_by_name;
 }
-
-size_t 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) {
-    // TODO leverage mmap
-    llama_file file(path_session, "rb");
-    const uint32_t magic = file.read_u32();
-    const uint32_t version = file.read_u32();
-
-    if (!(magic == 'ggsn' && version == 0)) {
-        fprintf(stderr, "%s : unknown (magic, version) for session file: %08x, %08x\n", __func__, magic, version);
-        return 0;
-    }
-
-    llama_hparams session_hparams;
-    file.read_raw(&session_hparams, sizeof(llama_hparams));
-
-    // REVIEW
-    if (session_hparams != ctx->model.hparams) {
-        fprintf(stderr, "%s : model hparams didn't match from session file!\n", __func__);
-        return 0;
-    }
-
-    const uint32_t n_token_count = file.read_u32();
-    LLAMA_ASSERT(n_token_capacity >= n_token_count);
-    file.read_raw(tokens_out, sizeof(llama_token) * n_token_count);
-    *n_token_count_out = n_token_count;
-
-    const size_t n_state_size = file.size - file.tell();
-    const size_t n_orig_state_size = llama_get_state_size(ctx);
-    if (n_state_size != n_orig_state_size) {
-        fprintf(stderr, "%s : failed to validate state size\n", __func__);
-    }
-    std::unique_ptr<uint8_t[]> state_data(new uint8_t[n_state_size]);
-    file.read_raw(state_data.get(), n_state_size);
-    return llama_set_state_data(ctx, state_data.get());
-}
-
-size_t llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count) {
-    // TODO save temp & swap
-    llama_file file(path_session, "wb");
-
-    const size_t n_state_size = llama_get_state_size(ctx);
-    std::unique_ptr<uint8_t[]> state_data(new uint8_t[n_state_size]);
-    llama_copy_state_data(ctx, state_data.get());
-
-    file.write_u32('ggsn'); // magic
-    file.write_u32(0); // version
-    file.write_raw(&ctx->model.hparams, sizeof(llama_hparams));
-
-    file.write_u32((uint32_t) n_token_count); // REVIEW
-    file.write_raw(tokens, sizeof(llama_token) * n_token_count);
-
-    file.write_raw(state_data.get(), n_state_size);
-    return n_state_size; // REVIEW
-}

llama.h

@@ -19,9 +19,11 @@
 #    define LLAMA_API
 #endif
 
-#define LLAMA_FILE_VERSION 1
-#define LLAMA_FILE_MAGIC 0x67676a74 // 'ggjt' in hex
-#define LLAMA_FILE_MAGIC_UNVERSIONED 0x67676d6c // pre-versioned files
+#define LLAMA_FILE_VERSION           1
+#define LLAMA_FILE_MAGIC             'ggjt'
+#define LLAMA_FILE_MAGIC_UNVERSIONED 'ggml'
+#define LLAMA_SESSION_MAGIC          'ggsn'
+#define LLAMA_SESSION_VERSION        0
 
 #ifdef __cplusplus
 extern "C" {
@@ -120,13 +122,13 @@ extern "C" {
                              int   n_threads);
 
     // Returns the number of tokens in the KV cache
-    LLAMA_API int llama_get_kv_cache_token_count(struct llama_context * ctx);
+    LLAMA_API int llama_get_kv_cache_token_count(const struct llama_context * ctx);
 
     // Sets the current rng seed.
     LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, int seed);
 
     // Returns the size in bytes of the state (rng, logits, embedding and kv_cache)
-    LLAMA_API size_t llama_get_state_size(struct llama_context * ctx);
+    LLAMA_API size_t llama_get_state_size(const struct llama_context * ctx);
 
     // Copies the state to the specified destination address.
     // Destination needs to have allocated enough memory.
@@ -138,8 +140,8 @@ extern "C" {
     LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src);
 
     // Save/load session file
-    LLAMA_API size_t 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);
-    LLAMA_API size_t llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
+    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);
+    LLAMA_API bool llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
 
     // Run the llama inference to obtain the logits and probabilities for the next token.
     // tokens + n_tokens is the provided batch of new tokens to process
@@ -164,9 +166,9 @@ extern "C" {
                              int   n_max_tokens,
                             bool   add_bos);
 
-    LLAMA_API int llama_n_vocab(struct llama_context * ctx);
-    LLAMA_API int llama_n_ctx  (struct llama_context * ctx);
-    LLAMA_API int llama_n_embd (struct llama_context * ctx);
+    LLAMA_API int llama_n_vocab(const struct llama_context * ctx);
+    LLAMA_API int llama_n_ctx  (const struct llama_context * ctx);
+    LLAMA_API int llama_n_embd (const struct llama_context * ctx);
 
     // Token logits obtained from the last call to llama_eval()
     // The logits for the last token are stored in the last row
@@ -180,7 +182,7 @@ extern "C" {
     LLAMA_API float * llama_get_embeddings(struct llama_context * ctx);
 
     // Token Id -> String. Uses the vocabulary in the provided context
-    LLAMA_API const char * llama_token_to_str(struct llama_context * ctx, llama_token token);
+    LLAMA_API const char * llama_token_to_str(const struct llama_context * ctx, llama_token token);
 
     // Special tokens
     LLAMA_API llama_token llama_token_bos();