Merge branch 'ggerganov:master' into server-chat-templates

Author: MaggotHATE

CMakePresets.json

@@ -24,11 +24,12 @@
             "CMAKE_INSTALL_RPATH": "$ORIGIN;$ORIGIN/.."
         }
     },
-    { "name": "debug",   "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
-    { "name": "release", "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
-    { "name": "reldbg",  "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
-    { "name": "static",  "hidden": true, "cacheVariables": { "GGML_STATIC": "ON" } },
-    { "name": "sycl_f16",  "hidden": true, "cacheVariables": { "GGML_SYCL_F16": "ON" } },
+    { "name": "debug",    "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Debug" } },
+    { "name": "release",  "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "Release" } },
+    { "name": "reldbg",   "hidden": true, "cacheVariables": { "CMAKE_BUILD_TYPE": "RelWithDebInfo" } },
+    { "name": "static",   "hidden": true, "cacheVariables": { "GGML_STATIC":      "ON" } },
+    { "name": "sycl_f16", "hidden": true, "cacheVariables": { "GGML_SYCL_F16":    "ON" } },
+    { "name": "vulkan",   "hidden": true, "cacheVariables": { "GGML_VULKAN":      "ON" } },
 
     {
         "name": "arm64-windows-msvc", "hidden": true,
@@ -57,25 +58,28 @@
         }
     },
 
-    { "name": "arm64-windows-llvm-debug"  , "inherits": [ "base", "arm64-windows-llvm",  "debug"   ] },
-    { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm",  "reldbg" ] },
-    { "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm",  "reldbg", "static" ] },
+    { "name": "arm64-windows-llvm-debug", "inherits": [ "base", "arm64-windows-llvm", "debug" ] },
+    { "name": "arm64-windows-llvm-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg" ] },
+    { "name": "arm64-windows-llvm+static-release", "inherits": [ "base", "arm64-windows-llvm", "reldbg", "static" ] },
 
-    { "name": "arm64-apple-clang-debug"  , "inherits": [ "base", "arm64-apple-clang",  "debug"   ] },
-    { "name": "arm64-apple-clang-release"  , "inherits": [ "base", "arm64-apple-clang",  "reldbg"   ] },
-    { "name": "arm64-apple-clang+static-release"  , "inherits": [ "base", "arm64-apple-clang",  "reldbg", "static" ] },
+    { "name": "arm64-apple-clang-debug", "inherits": [ "base", "arm64-apple-clang", "debug" ] },
+    { "name": "arm64-apple-clang-release", "inherits": [ "base", "arm64-apple-clang", "reldbg" ] },
+    { "name": "arm64-apple-clang+static-release", "inherits": [ "base", "arm64-apple-clang",  "reldbg", "static" ] },
 
-    { "name": "arm64-windows-msvc-debug"  , "inherits": [ "base", "arm64-windows-msvc",  "debug"   ] },
+    { "name": "arm64-windows-msvc-debug", "inherits": [ "base", "arm64-windows-msvc",  "debug"   ] },
     { "name": "arm64-windows-msvc-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg" ] },
     { "name": "arm64-windows-msvc+static-release", "inherits": [ "base", "arm64-windows-msvc",  "reldbg", "static" ] },
 
-    { "name": "x64-windows-msvc-debug"  , "inherits": [ "base", "debug"   ] },
+    { "name": "x64-windows-msvc-debug", "inherits": [ "base", "debug" ] },
     { "name": "x64-windows-msvc-release", "inherits": [ "base", "reldbg" ] },
     { "name": "x64-windows-msvc+static-release", "inherits": [ "base", "reldbg", "static" ] },
 
-    { "name": "x64-windows-sycl-debug"  , "inherits": [ "sycl-base", "debug"   ] },
+    { "name": "x64-windows-sycl-debug", "inherits": [ "sycl-base", "debug" ] },
     { "name": "x64-windows-sycl-debug-f16", "inherits": [ "sycl-base", "debug", "sycl_f16" ] },
     { "name": "x64-windows-sycl-release", "inherits": [ "sycl-base", "release" ] },
-    { "name": "x64-windows-sycl-release-f16", "inherits": [ "sycl-base", "release", "sycl_f16" ] }
+    { "name": "x64-windows-sycl-release-f16", "inherits": [ "sycl-base", "release", "sycl_f16" ] },
+
+    { "name": "x64-windows-vulkan-debug", "inherits": [ "base", "vulkan", "debug" ] },
+    { "name": "x64-windows-vulkan-release", "inherits": [ "base", "vulkan", "release" ] }
   ]
 }

Makefile

@@ -359,6 +359,10 @@ ifdef LLAMA_SERVER_SSL
 	MK_LDFLAGS += -lssl -lcrypto
 endif
 
+ifndef GGML_NO_CPU_AARCH64
+	MK_CPPFLAGS += -DGGML_USE_CPU_AARCH64
+endif
+
 # warnings
 WARN_FLAGS = \
 	-Wall \
@@ -940,10 +944,6 @@ ggml/src/ggml-cuda/%.o: \
 	$(MCC) $(CXXFLAGS) $(MUSAFLAGS) -x musa -mtgpu -c -o $@ $<
 endif # GGML_MUSA
 
-ifndef GGML_NO_CPU_AARCH64
-	MK_CPPFLAGS += -DGGML_USE_CPU_AARCH64
-endif
-
 ifdef GGML_METAL
 	MK_CPPFLAGS += -DGGML_USE_METAL
 	MK_LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit

ggml/src/ggml-aarch64.c

@@ -8,19 +8,42 @@
 
 #define UNUSED GGML_UNUSED
 
-static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
+static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
     block_q4_0x4 out;
 
     for (int i = 0; i < 4; i++) {
         out.d[i] = in[i].d;
     }
 
-    for (int i = 0; i < QK4_0 * 2; i++) {
-        int src_offset = (i / (4 * blck_size_interleave)) * blck_size_interleave;
-        int src_id = (i % (4 * blck_size_interleave)) / blck_size_interleave;
-        src_offset += (i % blck_size_interleave);
+    const int end = QK4_0 * 2 / blck_size_interleave;
 
-        out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
+    if (blck_size_interleave == 8) {
+        const uint64_t xor_mask = 0x8888888888888888ULL;
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            uint64_t elems;
+            // Using memcpy to avoid unaligned memory accesses
+            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+            elems ^= xor_mask;
+            memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+        }
+    } else if (blck_size_interleave == 4) {
+        const uint32_t xor_mask = 0x88888888;
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            uint32_t elems;
+            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t));
+            elems ^= xor_mask;
+            memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t));
+        }
+    } else {
+        GGML_ASSERT(false);
     }
 
     return out;
@@ -30,19 +53,25 @@ static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_in
 // returns an interleaved block_q4_0x8
 // in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
 // first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
-static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
+static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave) {
     block_q4_0x8 out;
 
     for (int i = 0; i < 8; i++) {
         out.d[i] = in[i].d;
     }
 
-    for (int i = 0; i < QK4_0 * 4; i++) {
-        int src_offset = (i / (8 * blck_size_interleave)) * blck_size_interleave;
-        int src_id = (i % (8 * blck_size_interleave)) / blck_size_interleave;
-        src_offset += (i % blck_size_interleave);
+    const int end = QK4_0 * 4 / blck_size_interleave;
+    const uint64_t xor_mask = 0x8888888888888888ULL;
+
+    for (int i = 0; i < end; ++i) {
+        int src_id = i % 8;
+        int src_offset = (i / 8) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
 
-        out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
+        uint64_t elems;
+        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+        elems ^= xor_mask;
+        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
     }
 
     return out;
@@ -71,11 +100,11 @@ static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict ds
             }
 
             if (nrows_interleaved == 8) {
-                *(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave, 0x88);
+                *(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave);
                 out_ptr = (block_q4_0x8 *) out_ptr + 1;
             }
             else if (nrows_interleaved == 4) {
-                *(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave, 0x88);
+                *(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave);
                 out_ptr = (block_q4_0x4 *) out_ptr + 1;
             }
         }

ggml/src/ggml-cpu/CMakeLists.txt

@@ -143,14 +143,23 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
             if (GGML_AVX512_VBMI)
                 add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VBMI__>)
                 add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VBMI__>)
+                if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
+                    list(APPEND ARCH_FLAGS -mavx512vbmi)
+                endif()
             endif()
             if (GGML_AVX512_VNNI)
                 add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512VNNI__>)
                 add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512VNNI__>)
+                if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
+                    list(APPEND ARCH_FLAGS -mavx512vnni)
+                endif()
             endif()
             if (GGML_AVX512_BF16)
                 add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AVX512BF16__>)
                 add_compile_definitions($<$<COMPILE_LANGUAGE:CXX>:__AVX512BF16__>)
+                if (CMAKE_C_COMPILER_ID STREQUAL "Clang")
+                    list(APPEND ARCH_FLAGS -mavx512bf16)
+                endif()
             endif()
             if (GGML_AMX_TILE)
                 add_compile_definitions($<$<COMPILE_LANGUAGE:C>:__AMX_TILE__>)
@@ -211,10 +220,13 @@ elseif (CMAKE_OSX_ARCHITECTURES STREQUAL "x86_64" OR CMAKE_GENERATOR_PLATFORM_LW
     endif()
 elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64")
     message(STATUS "PowerPC detected")
-    execute_process(COMMAND bash -c "grep POWER10 /proc/cpuinfo | head -n 1"
-                   OUTPUT_VARIABLE POWER10_M)
-    string(FIND ${POWER10_M} "POWER10" substring_index)
-    if(${substring_index} GREATER_EQUAL 0)
+    execute_process(COMMAND bash -c "grep POWER10 /proc/cpuinfo | head -n 1" OUTPUT_VARIABLE POWER10_M)
+    string(FIND "${POWER10_M}" "POWER10" substring_index)
+    if (NOT DEFINED substring_index OR "${substring_index}" STREQUAL "")
+        set(substring_index -1)
+    endif()
+
+    if (${substring_index} GREATER_EQUAL 0)
        list(APPEND ARCH_FLAGS -mcpu=power10)
     elseif (${CMAKE_SYSTEM_PROCESSOR} MATCHES "ppc64le")
        list(APPEND ARCH_FLAGS -mcpu=powerpc64le)

ggml/src/ggml-cpu/ggml-cpu-aarch64.c

@@ -3387,19 +3387,42 @@ void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void *
 }
 
 // FIXME: this code is duplicated from ggml-aarch64.c
-static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
+static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) {
     block_q4_0x4 out;
 
     for (int i = 0; i < 4; i++) {
         out.d[i] = in[i].d;
     }
 
-    for (int i = 0; i < QK4_0 * 2; i++) {
-        int src_offset = (i / (4 * blck_size_interleave)) * blck_size_interleave;
-        int src_id = (i % (4 * blck_size_interleave)) / blck_size_interleave;
-        src_offset += (i % blck_size_interleave);
+    const int end = QK4_0 * 2 / blck_size_interleave;
 
-        out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
+    if (blck_size_interleave == 8) {
+        const uint64_t xor_mask = 0x8888888888888888ULL;
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            uint64_t elems;
+            // Using memcpy to avoid unaligned memory accesses
+            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+            elems ^= xor_mask;
+            memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
+        }
+    } else if (blck_size_interleave == 4) {
+        const uint32_t xor_mask = 0x88888888;
+        for (int i = 0; i < end; ++i) {
+            int src_id = i % 4;
+            int src_offset = (i / 4) * blck_size_interleave;
+            int dst_offset = i * blck_size_interleave;
+
+            uint32_t elems;
+            memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t));
+            elems ^= xor_mask;
+            memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t));
+        }
+    } else {
+        GGML_ASSERT(false);
     }
 
     return out;
@@ -3409,19 +3432,25 @@ static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_in
 // returns an interleaved block_q4_0x8
 // in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
 // first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
-static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
+static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave) {
     block_q4_0x8 out;
 
     for (int i = 0; i < 8; i++) {
         out.d[i] = in[i].d;
     }
 
-    for (int i = 0; i < QK4_0 * 4; i++) {
-        int src_offset = (i / (8 * blck_size_interleave)) * blck_size_interleave;
-        int src_id = (i % (8 * blck_size_interleave)) / blck_size_interleave;
-        src_offset += (i % blck_size_interleave);
+    const int end = QK4_0 * 4 / blck_size_interleave;
+    const uint64_t xor_mask = 0x8888888888888888ULL;
+
+    for (int i = 0; i < end; ++i) {
+        int src_id = i % 8;
+        int src_offset = (i / 8) * blck_size_interleave;
+        int dst_offset = i * blck_size_interleave;
 
-        out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
+        uint64_t elems;
+        memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t));
+        elems ^= xor_mask;
+        memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t));
     }
 
     return out;
@@ -3449,7 +3478,7 @@ static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block
             for (int i = 0; i < nrows_interleaved; i++) {
                 dst_tmp[i] = src[x + i * nblocks];
             }
-            *dst++ = make_block_q4_0x4(dst_tmp, interleave_block, 0x88);
+            *dst++ = make_block_q4_0x4(dst_tmp, interleave_block);
         }
         src += nrows_interleaved * nblocks;
     }
@@ -3480,7 +3509,7 @@ static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor *t, int interleave_block,
             for (int i  = 0; i < nrows_interleaved; i++ ) {
                 dst_tmp[i] = src[x + i * nblocks];
             }
-            *dst++ = make_block_q4_0x8(dst_tmp, interleave_block, 0x88);
+            *dst++ = make_block_q4_0x8(dst_tmp, interleave_block);
         }
         src += nrows_interleaved * nblocks;
     }

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -1763,7 +1763,8 @@ static void ggml_vk_print_gpu_info(size_t idx) {
     fp16 = fp16 && vk12_features.shaderFloat16;
 
     std::string device_name = props2.properties.deviceName.data();
-    std::cerr << GGML_VK_NAME << idx << ": " << device_name << " (" << driver_props.driverName << ") | uma: " << uma << " | fp16: " << fp16 << " | warp size: " << subgroup_size << std::endl;
+    GGML_LOG_DEBUG("ggml_vulkan: %d = %s (%s) | uma: %d | fp16: %d | warp size: %d\n",
+              idx, device_name.c_str(), driver_props.driverName, uma, fp16, subgroup_size);
 
     if (props2.properties.deviceType == vk::PhysicalDeviceType::eCpu) {
         std::cerr << "ggml_vulkan: Warning: Device type is CPU. This is probably not the device you want." << std::endl;
@@ -1821,8 +1822,7 @@ void ggml_vk_instance_init() {
         };
         validation_features.setPNext(nullptr);
         instance_create_info.setPNext(&validation_features);
-
-        std::cerr << "ggml_vulkan: Validation layers enabled" << std::endl;
+        GGML_LOG_DEBUG("ggml_vulkan: Validation layers enabled\n");
     }
     vk_instance.instance = vk::createInstance(instance_create_info);
 
@@ -1936,8 +1936,8 @@ void ggml_vk_instance_init() {
             vk_instance.device_indices.push_back(0);
         }
     }
+    GGML_LOG_DEBUG("ggml_vulkan: Found %d Vulkan devices:\n", vk_instance.device_indices.size());
 
-    std::cerr << "ggml_vulkan: Found " << vk_instance.device_indices.size() << " Vulkan devices:" << std::endl;
 
     for (size_t i = 0; i < vk_instance.device_indices.size(); i++) {
         ggml_vk_print_gpu_info(i);

ggml/src/ggml.c

@@ -49,6 +49,14 @@
 
 #define UNUSED GGML_UNUSED
 
+#if defined(_MSC_VER)
+#define m512bh(p) p
+#define m512i(p) p
+#else
+#define m512bh(p) (__m512bh)(p)
+#define m512i(p) (__m512i)(p)
+#endif
+
 // precomputed f32 table for f16 (256 KB) (ggml-impl.h)
 float ggml_table_f32_f16[1 << 16];
 

scripts/compare-llama-bench.py

@@ -19,22 +19,22 @@
 
 # Properties by which to differentiate results per commit:
 KEY_PROPERTIES = [
-    "cpu_info", "gpu_info", "n_gpu_layers", "cuda", "vulkan", "kompute", "metal", "sycl", "rpc", "gpu_blas",
-    "blas", "model_filename", "model_type", "n_batch", "n_ubatch", "embeddings", "n_threads",
-    "type_k", "type_v", "use_mmap", "no_kv_offload", "split_mode", "main_gpu", "tensor_split", "flash_attn", "n_prompt", "n_gen"
+    "cpu_info", "gpu_info", "backends", "n_gpu_layers", "model_filename", "model_type", "n_batch", "n_ubatch",
+    "embeddings", "cpu_mask", "cpu_strict", "poll", "n_threads", "type_k", "type_v", "use_mmap", "no_kv_offload",
+    "split_mode", "main_gpu", "tensor_split", "flash_attn", "n_prompt", "n_gen"
 ]
 
 # Properties that are boolean and are converted to Yes/No for the table:
-BOOL_PROPERTIES = ["cuda", "vulkan", "kompute", "metal", "sycl", "gpu_blas", "blas", "embeddings", "use_mmap", "no_kv_offload", "flash_attn"]
+BOOL_PROPERTIES = ["embeddings", "cpu_strict", "use_mmap", "no_kv_offload", "flash_attn"]
 
 # Header names for the table:
 PRETTY_NAMES = {
-    "cuda": "CUDA", "vulkan": "Vulkan", "kompute": "Kompute", "metal": "Metal", "sycl": "SYCL", "rpc": "RPC",
-    "gpu_blas": "GPU BLAS", "blas": "BLAS", "cpu_info": "CPU", "gpu_info": "GPU", "model_filename": "File", "model_type": "Model",
-    "model_size": "Model Size [GiB]", "model_n_params": "Num. of Par.", "n_batch": "Batch size", "n_ubatch": "Microbatch size",
-    "n_threads": "Threads", "type_k": "K type", "type_v": "V type", "n_gpu_layers": "GPU layers", "split_mode": "Split mode",
-    "main_gpu": "Main GPU", "no_kv_offload": "NKVO", "flash_attn": "FlashAttention", "tensor_split": "Tensor split",
-    "use_mmap": "Use mmap", "embeddings": "Embeddings",
+    "cpu_info": "CPU", "gpu_info": "GPU", "backends": "Backends", "n_gpu_layers": "GPU layers",
+    "model_filename": "File", "model_type": "Model", "model_size": "Model size [GiB]",
+    "model_n_params": "Num. of par.", "n_batch": "Batch size", "n_ubatch": "Microbatch size",
+    "embeddings": "Embeddings", "cpu_mask": "CPU mask", "cpu_strict": "CPU strict", "poll": "Poll",
+    "n_threads": "Threads", "type_k": "K type", "type_v": "V type", "split_mode": "Split mode", "main_gpu": "Main GPU",
+    "no_kv_offload": "NKVO", "flash_attn": "FlashAttention", "tensor_split": "Tensor split", "use_mmap": "Use mmap",
 }
 
 DEFAULT_SHOW = ["model_type"]  # Always show these properties by default.

scripts/sync-ggml.last

@@ -1 +1 @@
-8a3d799484d861748f86eb87c8314fa2dbccc254
+9d0708e863f3aa2fc1eb0b75d433303c30bd0dbc