sync : llama.cpp

CMakeLists.txt

@@ -128,14 +128,9 @@ option(GGML_LLAMAFILE                       "ggml: use LLAMAFILE"
 
 option(GGML_CUDA                            "ggml: use CUDA"                                  OFF)
 option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
-option(GGML_CUDA_FORCE_DMMV                 "ggml: use dmmv instead of mmvq CUDA kernels"     OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
 option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
-set   (GGML_CUDA_DMMV_X   "32" CACHE STRING "ggml: x stride for dmmv CUDA kernels")
-set   (GGML_CUDA_MMV_Y     "1" CACHE STRING "ggml: y block size for mmv CUDA kernels")
 option(GGML_CUDA_F16                        "ggml: use 16 bit floats for some calculations"   OFF)
-set   (GGML_CUDA_KQUANTS_ITER "2" CACHE STRING
-                                            "ggml: iters./thread per block for Q2_K/Q6_K")
 set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
                                             "ggml: max. batch size for using peer access")
 option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)

scripts/sync-llama.last

@@ -1 +1 @@
-883d206fbd2c5b2b9b589a9328503b9005e146c9
+ce2e59ba107cf71ed566040ff20a15d1c58e09c2

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;
             }
         }

src/ggml-amx/ggml-amx.cpp

@@ -317,16 +317,13 @@ static bool ggml_backend_amx_device_supports_op(ggml_backend_dev_t dev, const st
             const enum ggml_type type = src0->type;
             const int64_t ne0 = op->ne[0];
 
-            bool is_training = src0->grad || src1->grad;
-
             // amx kernels enables for Q4_0, Q4_1, Q8_0, F16
             // Q4_K, Q5_K, Q6_K, IQ4_XS enabled for QK_K = 256
             bool has_amx_kernels = qtype_has_amx_kernels(type) || (type == GGML_TYPE_F16);
 
             bool can_use_amx =
                 is_contiguous_2d(src0) &&       // src0 must be contiguous
                 is_contiguous_2d(src1) &&       // src1 must be contiguous
-                !is_training &&                 // inference only
                 src1->type == GGML_TYPE_F32 &&  // src1 must be float32
                 has_amx_kernels &&              // with amx kernel impls
                 ne0 % (TILE_N * 2) == 0;        // out_features is 32x

src/ggml-backend.cpp

@@ -689,7 +689,7 @@ static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backen
 }
 
 static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor, const struct ggml_tensor * op) {
-    ggml_backend_buffer_t buffer = tensor->buffer;
+    ggml_backend_buffer_t buffer = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
     if (buffer == NULL) {
         return -1;
     }
@@ -722,8 +722,6 @@ static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS_DEBUG*GGML
 
 // returns the backend that should be used for the node based on the current locations
 static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, struct ggml_tensor * tensor) {
-    // TODO: use supports_op to check if the backend supports the op
-
     // assign pre-allocated nodes to their backend
     int cur_backend_id = ggml_backend_sched_backend_from_buffer(sched, tensor, tensor);
     if (cur_backend_id != -1) {
@@ -742,7 +740,7 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
 
     if (tensor->buffer || (tensor->view_src && tensor->view_src->buffer)) {
         // since the tensor is pre-allocated, it cannot be moved to another backend
-        GGML_ABORT("pre-allocated tensor in a backend that cannot run the operation");
+        GGML_ABORT("pre-allocated tensor (%s) in a backend that cannot run the operation", tensor->name);
     }
 
     // graph input
@@ -886,6 +884,9 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
         int * node_backend_id = &tensor_backend_id(node);
+        if (ggml_is_view_op(node->op)) {
+            continue;
+        }
         // do not overwrite user assignments
         if (*node_backend_id == -1) {
             *node_backend_id = ggml_backend_sched_backend_id_from_cur(sched, node);
@@ -1538,12 +1539,13 @@ bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph *
 
     ggml_backend_sched_split_graph(sched, measure_graph);
 
+    ggml_backend_sched_synchronize(sched);
+
     if (!ggml_gallocr_reserve_n(sched->galloc, &sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
         return false;
     }
 
     ggml_backend_sched_reset(sched);
-    ggml_backend_sched_synchronize(sched);
 
     return true;
 }

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__>)

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;
     }

src/ggml-cpu/ggml-cpu.c

@@ -2369,7 +2369,7 @@ void ggml_numa_init(enum ggml_numa_strategy numa_flag) {
     // figure out which node we're on
     uint current_cpu;
     int getcpu_ret = 0;
-#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 28) || defined(__COSMOPOLITAN__)
+#if __GLIBC__ > 2 || (__GLIBC__ == 2 && __GLIBC_MINOR__ > 33) || defined(__COSMOPOLITAN__)
     getcpu_ret = getcpu(&current_cpu, &g_state.numa.current_node);
 #else
     // old glibc doesn't have a wrapper for this call. Fall back on direct syscall

src/ggml-cpu/llamafile/sgemm.cpp

@@ -50,7 +50,8 @@
 
 #include "sgemm.h"
 #include "ggml-impl.h"
-#include "ggml-cpu-impl.h"
+// hack until moved into the CPU backend
+#include "../ggml-cpu-impl.h"
 #include "ggml-quants.h"
 
 #ifdef _MSC_VER

src/ggml-cuda/CMakeLists.txt

@@ -6,15 +6,18 @@ if (CUDAToolkit_FOUND)
     message(STATUS "CUDA Toolkit found")
 
     if (NOT DEFINED CMAKE_CUDA_ARCHITECTURES)
-        # 52 == lowest CUDA 12 standard
-        # 60 == FP16 CUDA intrinsics
-        # 61 == integer CUDA intrinsics
-        # 70 == compute capability at which unrolling a loop in mul_mat_q kernels is faster
-        if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
+        # native == GPUs available at build time
+        # 52     == Maxwell, lowest CUDA 12 standard
+        # 60     == P100, FP16 CUDA intrinsics
+        # 61     == Pascal, __dp4a instruction (per-byte integer dot product)
+        # 70     == V100, FP16 tensor cores
+        # 75     == Turing, int8 tensor cores
+        if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6")
+            set(CMAKE_CUDA_ARCHITECTURES "native")
+        elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
             set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")
         else()
             set(CMAKE_CUDA_ARCHITECTURES "52;61;70;75")
-            #set(CMAKE_CUDA_ARCHITECTURES "OFF") # use this to compile much faster, but only F16 models work
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
@@ -51,21 +54,12 @@ if (CUDAToolkit_FOUND)
     target_link_libraries(ggml-cuda PRIVATE ggml-base)
     target_include_directories(ggml-cuda PRIVATE . ..)
 
-    # TODO: change the definitions to this target only
-
-    add_compile_definitions(GGML_CUDA_DMMV_X=${GGML_CUDA_DMMV_X})
-    add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_MMV_Y})
-    add_compile_definitions(K_QUANTS_PER_ITERATION=${GGML_CUDA_KQUANTS_ITER})
     add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
 
     if (GGML_CUDA_GRAPHS)
         add_compile_definitions(GGML_CUDA_USE_GRAPHS)
     endif()
 
-    if (GGML_CUDA_FORCE_DMMV)
-        add_compile_definitions(GGML_CUDA_FORCE_DMMV)
-    endif()
-
     if (GGML_CUDA_FORCE_MMQ)
         add_compile_definitions(GGML_CUDA_FORCE_MMQ)
     endif()
@@ -78,10 +72,6 @@ if (CUDAToolkit_FOUND)
         add_compile_definitions(GGML_CUDA_NO_VMM)
     endif()
 
-    if (DEFINED GGML_CUDA_DMMV_Y)
-        add_compile_definitions(GGML_CUDA_MMV_Y=${GGML_CUDA_DMMV_Y}) # for backwards compatibility
-    endif()
-
     if (GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
         add_compile_definitions(GGML_CUDA_F16)
     endif()