npm - @fugood/llama.node - Versions diffs - 0.3.3 → 0.3.4 - Mend

@fugood/llama.node 0.3.3 → 0.3.4

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (225) hide show

package/CMakeLists.txt +5 -0
package/bin/darwin/arm64/llama-node.node +0 -0
package/bin/darwin/x64/llama-node.node +0 -0
package/bin/linux/arm64/llama-node.node +0 -0
package/bin/linux/x64/llama-node.node +0 -0
package/bin/linux-vulkan/arm64/llama-node.node +0 -0
package/bin/linux-vulkan/x64/llama-node.node +0 -0
package/bin/win32/arm64/llama-node.node +0 -0
package/bin/win32/arm64/node.lib +0 -0
package/bin/win32/x64/llama-node.node +0 -0
package/bin/win32/x64/node.lib +0 -0
package/bin/win32-vulkan/arm64/llama-node.node +0 -0
package/bin/win32-vulkan/arm64/node.lib +0 -0
package/bin/win32-vulkan/x64/llama-node.node +0 -0
package/bin/win32-vulkan/x64/node.lib +0 -0
package/lib/binding.ts +18 -1
package/package.json +1 -1
package/src/EmbeddingWorker.cpp +15 -5
package/src/EmbeddingWorker.h +2 -1
package/src/LlamaCompletionWorker.cpp +1 -1
package/src/LlamaContext.cpp +81 -18
package/src/LlamaContext.h +2 -0
package/src/llama.cpp/.github/workflows/build.yml +197 -159
package/src/llama.cpp/.github/workflows/docker.yml +5 -8
package/src/llama.cpp/.github/workflows/python-lint.yml +8 -1
package/src/llama.cpp/.github/workflows/server.yml +21 -14
package/src/llama.cpp/CMakeLists.txt +11 -6
package/src/llama.cpp/Sources/llama/llama.h +4 -0
package/src/llama.cpp/cmake/common.cmake +33 -0
package/src/llama.cpp/cmake/x64-windows-llvm.cmake +11 -0
package/src/llama.cpp/common/CMakeLists.txt +6 -2
package/src/llama.cpp/common/arg.cpp +426 -245
package/src/llama.cpp/common/common.cpp +143 -80
package/src/llama.cpp/common/common.h +81 -24
package/src/llama.cpp/common/sampling.cpp +53 -19
package/src/llama.cpp/common/sampling.h +22 -1
package/src/llama.cpp/common/speculative.cpp +274 -0
package/src/llama.cpp/common/speculative.h +28 -0
package/src/llama.cpp/docs/build.md +101 -148
package/src/llama.cpp/examples/CMakeLists.txt +32 -13
package/src/llama.cpp/examples/batched/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/batched/batched.cpp +5 -4
package/src/llama.cpp/examples/batched-bench/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/convert-llama2c-to-ggml/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/cvector-generator/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/deprecation-warning/deprecation-warning.cpp +1 -1
package/src/llama.cpp/examples/embedding/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/eval-callback/CMakeLists.txt +3 -2
package/src/llama.cpp/examples/export-lora/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/gbnf-validator/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/gbnf-validator/gbnf-validator.cpp +4 -7
package/src/llama.cpp/examples/gen-docs/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/gguf/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/gguf-hash/CMakeLists.txt +8 -1
package/src/llama.cpp/examples/gguf-split/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/gguf-split/gguf-split.cpp +2 -2
package/src/llama.cpp/examples/gritlm/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/gritlm/gritlm.cpp +1 -1
package/src/llama.cpp/examples/imatrix/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/imatrix/imatrix.cpp +11 -2
package/src/llama.cpp/examples/infill/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/infill/infill.cpp +1 -1
package/src/llama.cpp/examples/llama-bench/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/llama-bench/llama-bench.cpp +405 -316
package/src/llama.cpp/examples/llama.android/llama/build.gradle.kts +1 -0
package/src/llama.cpp/examples/llava/CMakeLists.txt +10 -3
package/src/llama.cpp/examples/llava/clip.cpp +262 -66
package/src/llama.cpp/examples/llava/clip.h +8 -2
package/src/llama.cpp/examples/llava/llava-cli.cpp +1 -1
package/src/llama.cpp/examples/llava/llava.cpp +46 -19
package/src/llama.cpp/examples/llava/minicpmv-cli.cpp +1 -1
package/src/llama.cpp/examples/llava/qwen2vl-cli.cpp +581 -0
package/src/llama.cpp/examples/lookahead/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/lookahead/lookahead.cpp +1 -1
package/src/llama.cpp/examples/lookup/CMakeLists.txt +4 -4
package/src/llama.cpp/examples/lookup/lookup-stats.cpp +2 -1
package/src/llama.cpp/examples/lookup/lookup.cpp +2 -2
package/src/llama.cpp/examples/main/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/main/main.cpp +9 -5
package/src/llama.cpp/examples/main-cmake-pkg/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/parallel/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/parallel/parallel.cpp +1 -1
package/src/llama.cpp/examples/passkey/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/perplexity/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/quantize/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/quantize/quantize.cpp +0 -3
package/src/llama.cpp/examples/quantize-stats/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/retrieval/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/retrieval/retrieval.cpp +4 -4
package/src/llama.cpp/examples/run/CMakeLists.txt +5 -0
package/src/llama.cpp/examples/run/run.cpp +911 -0
package/src/llama.cpp/examples/save-load-state/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/save-load-state/save-load-state.cpp +4 -4
package/src/llama.cpp/examples/server/CMakeLists.txt +3 -7
package/src/llama.cpp/examples/server/server.cpp +1758 -886
package/src/llama.cpp/examples/server/tests/requirements.txt +2 -2
package/src/llama.cpp/examples/server/utils.hpp +94 -304
package/src/llama.cpp/examples/simple/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/simple/simple.cpp +4 -0
package/src/llama.cpp/examples/simple-chat/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/simple-chat/simple-chat.cpp +3 -0
package/src/llama.cpp/examples/speculative/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/speculative/speculative.cpp +16 -15
package/src/llama.cpp/examples/speculative-simple/CMakeLists.txt +5 -0
package/src/llama.cpp/examples/speculative-simple/speculative-simple.cpp +265 -0
package/src/llama.cpp/examples/tokenize/CMakeLists.txt +1 -1
package/src/llama.cpp/examples/tokenize/tokenize.cpp +1 -1
package/src/llama.cpp/examples/tts/CMakeLists.txt +5 -0
package/src/llama.cpp/examples/tts/tts.cpp +932 -0
package/src/llama.cpp/ggml/CMakeLists.txt +46 -34
package/src/llama.cpp/ggml/include/ggml-backend.h +16 -0
package/src/llama.cpp/ggml/include/ggml-cpu.h +7 -49
package/src/llama.cpp/ggml/include/ggml-opencl.h +26 -0
package/src/llama.cpp/ggml/include/ggml.h +106 -24
package/src/llama.cpp/ggml/src/CMakeLists.txt +73 -24
package/src/llama.cpp/ggml/src/ggml-alloc.c +0 -1
package/src/llama.cpp/ggml/src/ggml-backend-impl.h +51 -11
package/src/llama.cpp/ggml/src/ggml-backend-reg.cpp +379 -22
package/src/llama.cpp/ggml/src/ggml-backend.cpp +4 -4
package/src/llama.cpp/ggml/src/ggml-blas/CMakeLists.txt +3 -7
package/src/llama.cpp/ggml/src/ggml-blas/ggml-blas.cpp +5 -2
package/src/llama.cpp/ggml/src/ggml-cann/CMakeLists.txt +33 -3
package/src/llama.cpp/ggml/src/ggml-cann/aclnn_ops.cpp +456 -111
package/src/llama.cpp/ggml/src/ggml-cann/common.h +6 -3
package/src/llama.cpp/ggml/src/ggml-cann/ggml-cann.cpp +95 -35
package/src/llama.cpp/ggml/src/ggml-cann/kernels/CMakeLists.txt +2 -5
package/src/llama.cpp/ggml/src/ggml-cann/kernels/dup.cpp +22 -9
package/src/llama.cpp/ggml/src/ggml-cann/kernels/get_row_f16.cpp +24 -13
package/src/llama.cpp/ggml/src/ggml-cann/kernels/get_row_f32.cpp +23 -13
package/src/llama.cpp/ggml/src/ggml-cann/kernels/get_row_q4_0.cpp +11 -0
package/src/llama.cpp/ggml/src/ggml-cann/kernels/quantize_f16_q8_0.cpp +10 -0
package/src/llama.cpp/ggml/src/ggml-cann/kernels/quantize_f32_q8_0.cpp +10 -0
package/src/llama.cpp/ggml/src/ggml-cann/kernels/quantize_float_to_q4_0.cpp +17 -0
package/src/llama.cpp/ggml/src/ggml-common.h +42 -42
package/src/llama.cpp/ggml/src/ggml-cpu/CMakeLists.txt +288 -213
package/src/llama.cpp/ggml/src/ggml-cpu/amx/amx.cpp +220 -0
package/src/llama.cpp/ggml/src/ggml-cpu/amx/amx.h +8 -0
package/src/llama.cpp/ggml/src/{ggml-amx → ggml-cpu/amx}/common.h +19 -22
package/src/llama.cpp/ggml/src/{ggml-amx → ggml-cpu/amx}/mmq.cpp +93 -92
package/src/llama.cpp/ggml/src/{ggml-amx → ggml-cpu/amx}/mmq.h +2 -9
package/src/llama.cpp/ggml/src/ggml-cpu/cpu-feats-x86.cpp +323 -0
package/src/llama.cpp/ggml/src/ggml-cpu/{ggml-cpu-aarch64.c → ggml-cpu-aarch64.cpp} +892 -190
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-aarch64.h +2 -24
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-hbm.cpp +55 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-hbm.h +8 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-impl.h +15 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-quants.c +38 -25
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-traits.cpp +36 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu-traits.h +38 -0
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c +552 -399
package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.cpp +101 -136
package/src/llama.cpp/ggml/src/ggml-cpu/llamafile/sgemm.cpp +2 -2
package/src/llama.cpp/ggml/src/ggml-cuda/CMakeLists.txt +7 -10
package/src/llama.cpp/ggml/src/ggml-cuda/vendors/hip.h +8 -0
package/src/llama.cpp/ggml/src/ggml-hip/CMakeLists.txt +4 -6
package/src/llama.cpp/ggml/src/ggml-impl.h +32 -11
package/src/llama.cpp/ggml/src/ggml-kompute/CMakeLists.txt +13 -9
package/src/llama.cpp/ggml/src/ggml-kompute/ggml-kompute.cpp +131 -64
package/src/llama.cpp/ggml/src/ggml-metal/CMakeLists.txt +3 -6
package/src/llama.cpp/ggml/src/ggml-metal/ggml-metal-impl.h +39 -0
package/src/llama.cpp/ggml/src/ggml-musa/CMakeLists.txt +14 -7
package/src/llama.cpp/ggml/src/ggml-opencl/CMakeLists.txt +147 -0
package/src/llama.cpp/ggml/src/ggml-opencl/ggml-opencl.cpp +4004 -0
package/src/llama.cpp/ggml/src/ggml-opt.cpp +67 -80
package/src/llama.cpp/ggml/src/ggml-quants.c +0 -9
package/src/llama.cpp/ggml/src/ggml-rpc/CMakeLists.txt +3 -5
package/src/llama.cpp/ggml/src/ggml-rpc/ggml-rpc.cpp +5 -2
package/src/llama.cpp/ggml/src/ggml-sycl/CMakeLists.txt +13 -10
package/src/llama.cpp/ggml/src/ggml-sycl/common.cpp +2 -11
package/src/llama.cpp/ggml/src/ggml-sycl/common.hpp +1 -0
package/src/llama.cpp/ggml/src/ggml-sycl/concat.cpp +2 -2
package/src/llama.cpp/ggml/src/ggml-sycl/convert.cpp +1 -1
package/src/llama.cpp/ggml/src/ggml-sycl/dmmv.cpp +5 -5
package/src/llama.cpp/ggml/src/ggml-sycl/dpct/helper.hpp +32 -13
package/src/llama.cpp/ggml/src/ggml-sycl/element_wise.cpp +80 -61
package/src/llama.cpp/ggml/src/ggml-sycl/gemm.hpp +4 -4
package/src/llama.cpp/ggml/src/ggml-sycl/ggml-sycl.cpp +159 -114
package/src/llama.cpp/ggml/src/ggml-sycl/im2col.cpp +3 -2
package/src/llama.cpp/ggml/src/ggml-sycl/mmq.cpp +6 -6
package/src/llama.cpp/ggml/src/ggml-sycl/mmvq.cpp +6 -20
package/src/llama.cpp/ggml/src/ggml-sycl/norm.cpp +4 -3
package/src/llama.cpp/ggml/src/ggml-sycl/outprod.cpp +8 -8
package/src/llama.cpp/ggml/src/ggml-sycl/rope.cpp +4 -3
package/src/llama.cpp/ggml/src/ggml-sycl/softmax.cpp +7 -7
package/src/llama.cpp/ggml/src/ggml-sycl/tsembd.cpp +1 -0
package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.cpp +4 -1
package/src/llama.cpp/ggml/src/ggml-threading.h +4 -2
package/src/llama.cpp/ggml/src/ggml-vulkan/CMakeLists.txt +21 -7
package/src/llama.cpp/ggml/src/ggml-vulkan/ggml-vulkan.cpp +1718 -399
package/src/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/CMakeLists.txt +3 -1
package/src/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp +105 -31
package/src/llama.cpp/ggml/src/ggml.c +367 -207
package/src/llama.cpp/include/llama-cpp.h +25 -0
package/src/llama.cpp/include/llama.h +26 -19
package/src/llama.cpp/models/ggml-vocab-roberta-bpe.gguf.inp +112 -0
package/src/llama.cpp/models/ggml-vocab-roberta-bpe.gguf.out +46 -0
package/src/llama.cpp/pocs/CMakeLists.txt +3 -1
package/src/llama.cpp/pocs/vdot/CMakeLists.txt +2 -2
package/src/llama.cpp/src/CMakeLists.txt +2 -7
package/src/llama.cpp/src/llama-grammar.cpp +15 -15
package/src/llama.cpp/src/llama-grammar.h +2 -5
package/src/llama.cpp/src/llama-sampling.cpp +35 -90
package/src/llama.cpp/src/llama-vocab.cpp +6 -1
package/src/llama.cpp/src/llama.cpp +1748 -640
package/src/llama.cpp/src/unicode.cpp +62 -51
package/src/llama.cpp/src/unicode.h +9 -10
package/src/llama.cpp/tests/CMakeLists.txt +48 -37
package/src/llama.cpp/tests/test-arg-parser.cpp +2 -2
package/src/llama.cpp/tests/test-backend-ops.cpp +140 -21
package/src/llama.cpp/tests/test-chat-template.cpp +50 -4
package/src/llama.cpp/tests/test-gguf.cpp +1303 -0
package/src/llama.cpp/tests/test-grammar-integration.cpp +3 -6
package/src/llama.cpp/tests/test-llama-grammar.cpp +2 -4
package/src/llama.cpp/tests/test-quantize-fns.cpp +3 -3
package/src/llama.cpp/tests/test-rope.cpp +61 -20
package/src/llama.cpp/tests/test-sampling.cpp +2 -2
package/src/llama.cpp/.github/workflows/nix-ci-aarch64.yml +0 -72
package/src/llama.cpp/.github/workflows/nix-ci.yml +0 -79
package/src/llama.cpp/.github/workflows/nix-flake-update.yml +0 -22
package/src/llama.cpp/.github/workflows/nix-publish-flake.yml +0 -36
package/src/llama.cpp/ggml/include/ggml-amx.h +0 -25
package/src/llama.cpp/ggml/src/ggml-aarch64.c +0 -129
package/src/llama.cpp/ggml/src/ggml-aarch64.h +0 -19
package/src/llama.cpp/ggml/src/ggml-amx/CMakeLists.txt +0 -107
package/src/llama.cpp/ggml/src/ggml-amx/ggml-amx.cpp +0 -446

package/src/llama.cpp/ggml/src/ggml-cpu/ggml-cpu.c CHANGED Viewed

@@ -3,13 +3,14 @@
 #include "ggml-backend-impl.h"
 #include "ggml-backend.h"
-#include "ggml-cpu-aarch64.h"
+#include "ggml-cpu-traits.h"
 #include "ggml-cpu-impl.h"
 #include "ggml-cpu.h"
 #include "ggml-impl.h"
 #include "ggml-quants.h"
 #include "ggml-cpu-quants.h"
 #include "ggml-threading.h"
+#include "amx/amx.h"
 #include "ggml.h"
 #if defined(_MSC_VER) || defined(__MINGW32__)
@@ -109,10 +110,11 @@ static ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16];
 #if defined(__ARM_ARCH)
 struct ggml_arm_arch_features_type {
     int has_neon;
+    int has_dotprod;
     int has_i8mm;
     int has_sve;
     int sve_cnt;
-} ggml_arm_arch_features = {-1, -1, -1, 0};
+} ggml_arm_arch_features = {-1, -1, -1, -1, 0};
 #endif
@@ -124,8 +126,7 @@ struct ggml_arm_arch_features_type {
 #endif
 #include <windows.h>
-#if !defined(__clang__)
+#if defined(_MSC_VER) && !defined(__clang__)
 #define GGML_CACHE_ALIGN __declspec(align(GGML_CACHE_LINE))
 typedef volatile LONG atomic_int;
@@ -222,10 +223,6 @@ typedef void * thread_ret_t;
 typedef pthread_t ggml_thread_t;
-#ifdef GGML_USE_CPU_HBM
-#include <hbwmalloc.h>
-#endif
 #if defined(__APPLE__)
 #include <unistd.h>
 #include <mach/mach.h>
@@ -299,7 +296,6 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
     },
     [GGML_TYPE_Q8_0] = {
         .from_float               = quantize_row_q8_0,
-        .from_float_to_mat        = quantize_mat_q8_0,
         .vec_dot                  = ggml_vec_dot_q8_0_q8_0,
         .vec_dot_type             = GGML_TYPE_Q8_0,
 #if defined (__ARM_FEATURE_MATMUL_INT8)
@@ -407,33 +403,6 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .vec_dot_type             = GGML_TYPE_BF16,
         .nrows                    = 1,
     },
-    [GGML_TYPE_Q4_0_4_4] = {
-        .from_float               = NULL,
-        .vec_dot                  = NULL,
-        .vec_dot_type             = GGML_TYPE_Q8_0,
-        .nrows                    = 1,
-        .ncols                    = 4,
-        .gemv                     = ggml_gemv_q4_0_4x4_q8_0,
-        .gemm                     = ggml_gemm_q4_0_4x4_q8_0,
-    },
-    [GGML_TYPE_Q4_0_4_8] = {
-        .from_float               = NULL,
-        .vec_dot                  = NULL,
-        .vec_dot_type             = GGML_TYPE_Q8_0,
-        .nrows                    = 1,
-        .ncols                    = 4,
-        .gemv                     = ggml_gemv_q4_0_4x8_q8_0,
-        .gemm                     = ggml_gemm_q4_0_4x8_q8_0,
-    },
-    [GGML_TYPE_Q4_0_8_8] = {
-        .from_float               = NULL,
-        .vec_dot                  = NULL,
-        .vec_dot_type             = GGML_TYPE_Q8_0,
-        .nrows                    = 1,
-        .ncols                    = 8,
-        .gemv                     = ggml_gemv_q4_0_8x8_q8_0,
-        .gemm                     = ggml_gemm_q4_0_8x8_q8_0,
-    },
     [GGML_TYPE_TQ1_0] = {
         .from_float               = quantize_row_tq1_0,
         .vec_dot                  = ggml_vec_dot_tq1_0_q8_K,
@@ -485,21 +454,21 @@ const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type
 #define GGML_F32x4_ADD          vaddq_f32
 #define GGML_F32x4_MUL          vmulq_f32
 #define GGML_F32x4_REDUCE_ONE(x) vaddvq_f32(x)
-#define GGML_F32x4_REDUCE(res, x)                  \
-{                                                  \
-    int offset = GGML_F32_ARR >> 1;                \
-    for (int i = 0; i < offset; ++i) {             \
-        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
-    }                                              \
-    offset >>= 1;                                  \
-    for (int i = 0; i < offset; ++i) {             \
-        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
-    }                                              \
-    offset >>= 1;                                  \
-    for (int i = 0; i < offset; ++i) {             \
-        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
-    }                                              \
-    (res) = GGML_F32x4_REDUCE_ONE((x)[0]);         \
+#define GGML_F32x4_REDUCE(res, x)                       \
+{                                                       \
+    int offset = GGML_F32_ARR >> 1;                     \
+    for (int i = 0; i < offset; ++i) {                  \
+        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]);      \
+    }                                                   \
+    offset >>= 1;                                       \
+    for (int i = 0; i < offset; ++i) {                  \
+        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]);      \
+    }                                                   \
+    offset >>= 1;                                       \
+    for (int i = 0; i < offset; ++i) {                  \
+        (x)[i] = vaddq_f32((x)[i], (x)[offset+i]);      \
+    }                                                   \
+    (res) = (ggml_float) GGML_F32x4_REDUCE_ONE((x)[0]); \
 }
 #define GGML_F32_VEC        GGML_F32x4
@@ -614,7 +583,7 @@ do {                                                                  \
     for (int i = 0; i < offset; ++i) {                                \
         x[i] = _mm512_add_ps(x[i], x[offset+i]);                      \
     }                                                                 \
-    res = _mm512_reduce_add_ps(x[0]);                                 \
+    res = (ggml_float) _mm512_reduce_add_ps(x[0]);                    \
 } while (0)
 // TODO: is this optimal ?
@@ -664,7 +633,7 @@ do {                                                              \
     for (int i = 0; i < offset; ++i) {                            \
         x[i] = _mm512_add_ps(x[i], x[offset+i]);                  \
     }                                                             \
-    res = _mm512_reduce_add_ps(x[0]);                             \
+    res = (ggml_float) _mm512_reduce_add_ps(x[0]);                \
 } while (0)
 #define GGML_F16_VEC                GGML_F32Cx16
@@ -675,8 +644,8 @@ do {                                                              \
 #define GGML_F16_VEC_FMA            GGML_F32Cx16_FMA
 #define GGML_F16_VEC_ADD            GGML_F32Cx16_ADD
 #define GGML_F16_VEC_MUL            GGML_F32Cx16_MUL
-#define GGML_F16_VEC_REDUCE         GGML_F32Cx16_REDUCE
+#define GGML_F16_VEC_REDUCE         GGML_F32Cx16_REDUCE
 #elif defined(__AVX__)
 #define GGML_SIMD
@@ -745,7 +714,7 @@ do {                                                              \
 #define GGML_F32Cx8_LOAD(x)     _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(x)))
 #define GGML_F32Cx8_STORE(x, y) _mm_storeu_si128((__m128i *)(x), _mm256_cvtps_ph(y, 0))
 #else
-static inline __m256 __avx_f32cx8_load(ggml_fp16_t *x) {
+static inline __m256 __avx_f32cx8_load(const ggml_fp16_t * x) {
     float tmp[8];
     for (int i = 0; i < 8; i++) {
@@ -1168,28 +1137,28 @@ static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
 #define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a)
 #define GGML_F32x4_ADD     __lsx_vfadd_s
 #define GGML_F32x4_MUL     __lsx_vfmul_s
-#define GGML_F32x4_REDUCE(res, x)                                 \
-{                                                                 \
-    int offset = GGML_F32_ARR >> 1;                               \
-    for (int i = 0; i < offset; ++i) {                            \
-        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                     \
-    }                                                             \
-    offset >>= 1;                                                 \
-    for (int i = 0; i < offset; ++i) {                            \
-        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                     \
-    }                                                             \
-    offset >>= 1;                                                 \
-    for (int i = 0; i < offset; ++i) {                            \
-        x[i] = __lsx_vfadd_s(x[i], x[offset+i]);                     \
-    }                                                             \
-    __m128i tmp = __lsx_vsrli_d((__m128i)x[0], 32); \
-    tmp = (__m128i)__lsx_vfadd_s((__m128)tmp, x[0]); \
-    tmp = __lsx_vpickev_w(__lsx_vldi(0), tmp); \
-    const __m128 t0 = __lsx_vshuf4i_w(tmp, 0x88); \
-    tmp = __lsx_vsrli_d((__m128i)t0, 32); \
-    tmp = (__m128i)__lsx_vfadd_s((__m128)tmp, t0); \
-    tmp = __lsx_vpickev_w(__lsx_vldi(0), tmp); \
-    res = (ggml_float) __lsx_vpickve2gr_w(__lsx_vshuf4i_w(tmp, 0x88), 0);        \
+#define GGML_F32x4_REDUCE(res, x)                                                     \
+{                                                                                     \
+    int offset = GGML_F32_ARR >> 1;                                                   \
+    for (int i = 0; i < offset; ++i) {                                                \
+        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
+    }                                                                                 \
+    offset >>= 1;                                                                     \
+    for (int i = 0; i < offset; ++i) {                                                \
+        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
+    }                                                                                 \
+    offset >>= 1;                                                                     \
+    for (int i = 0; i < offset; ++i) {                                                \
+        x[i] = __lsx_vfadd_s(x[i], x[offset + i]);                                    \
+    }                                                                                 \
+    __m128i tmp     = __lsx_vsrli_d((__m128i) x[0], 32);                              \
+    tmp             = (__m128i) __lsx_vfadd_s((__m128) tmp, x[0]);                    \
+    tmp             = __lsx_vpickev_w(__lsx_vldi(0), tmp);                            \
+    const __m128 t0 = __lsx_vshuf4i_w(tmp, 0x88);                                     \
+    tmp             = __lsx_vsrli_d((__m128i) t0, 32);                                \
+    tmp             = (__m128i) __lsx_vfadd_s((__m128) tmp, t0);                      \
+    tmp             = __lsx_vpickev_w(__lsx_vldi(0), tmp);                            \
+    res             = (ggml_float) __lsx_vpickve2gr_w(__lsx_vshuf4i_w(tmp, 0x88), 0); \
 }
 #define GGML_F32_VEC        GGML_F32x4
@@ -1357,31 +1326,18 @@ struct ggml_compute_state {
     int ith;
 };
-struct ggml_compute_params {
-    // ith = thread index, nth = number of threads
-    int ith, nth;
-    // work buffer for all threads
-    size_t wsize;
-    void * wdata;
-    struct ggml_threadpool * threadpool;
-};
 //
 // fundamental operations
 //
 inline static void ggml_vec_set_i8(const int n, int8_t * x, const int8_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
 inline static void ggml_vec_set_i16(const int n, int16_t * x, const int16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
-inline static void ggml_vec_set_i32(const int n, int32_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
+inline static void ggml_vec_set_i32(const int n, int32_t * x, const int32_t   v) { for (int i = 0; i < n; ++i) x[i] = v;    }
+inline static void ggml_vec_cpy_i32(const int n, int32_t * y, const int32_t * x) { for (int i = 0; i < n; ++i) y[i] = x[i]; }
 inline static void ggml_vec_set_f16(const int n, ggml_fp16_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
 inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
 inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i]  = x[i] + y[i]; }
 inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float   v) { for (int i = 0; i < n; ++i) z[i]  = x[i] + v;    }
 inline static void ggml_vec_acc_f32 (const int n, float * y, const float * x)                  { for (int i = 0; i < n; ++i) y[i] += x[i];        }
@@ -2276,7 +2232,7 @@ struct ggml_state {
 static struct ggml_state g_state = {0};
-static void ggml_barrier(struct ggml_threadpool * tp) {
+void ggml_barrier(struct ggml_threadpool * tp) {
     int n_threads = atomic_load_explicit(&tp->n_threads_cur, memory_order_relaxed);
     if (n_threads == 1) {
         return;
@@ -2430,7 +2386,7 @@ bool ggml_is_numa(void) {
 #endif
 #if !defined(HWCAP2_I8MM)
-#define HWCAP2_I8MM 0
+#define HWCAP2_I8MM (1 << 13)
 #endif
 static void ggml_init_arm_arch_features(void) {
@@ -2439,6 +2395,7 @@ static void ggml_init_arm_arch_features(void) {
     uint32_t hwcap2 = getauxval(AT_HWCAP2);
     ggml_arm_arch_features.has_neon = !!(hwcap & HWCAP_ASIMD);
+    ggml_arm_arch_features.has_dotprod = !!(hwcap & HWCAP_ASIMDDP);
     ggml_arm_arch_features.has_i8mm = !!(hwcap2 & HWCAP2_I8MM);
     ggml_arm_arch_features.has_sve  = !!(hwcap & HWCAP_SVE);
@@ -2453,6 +2410,11 @@ static void ggml_init_arm_arch_features(void) {
     }
     ggml_arm_arch_features.has_neon = oldp;
+    if (sysctlbyname("hw.optional.arm.FEAT_DotProd", &oldp, &size, NULL, 0) != 0) {
+        oldp = 0;
+    }
+    ggml_arm_arch_features.has_dotprod = oldp;
     if (sysctlbyname("hw.optional.arm.FEAT_I8MM", &oldp, &size, NULL, 0) != 0) {
         oldp = 0;
     }
@@ -4505,9 +4467,6 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_Q4_0_4_4:
-        case GGML_TYPE_Q4_0_4_8:
-        case GGML_TYPE_Q4_0_8_8:
             {
                 ggml_compute_forward_add_q_f32(params, dst);
             } break;
@@ -4885,9 +4844,6 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_Q4_0_4_4:
-        case GGML_TYPE_Q4_0_4_8:
-        case GGML_TYPE_Q4_0_8_8:
             {
                 ggml_compute_forward_add1_q_f32(params, dst);
             } break;
@@ -5015,9 +4971,6 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_Q4_0_4_4:
-        case GGML_TYPE_Q4_0_4_8:
-        case GGML_TYPE_Q4_0_8_8:
         default:
             {
                 GGML_ABORT("fatal error");
@@ -7433,20 +7386,9 @@ static void ggml_compute_forward_mul_mat(
     const int ith = params->ith;
     const int nth = params->nth;
-    enum ggml_type type = src0->type;
-    if (src0->buffer && ggml_backend_cpu_buft_is_aarch64(src0->buffer->buft)) {
-        type = (enum ggml_type)(intptr_t)src0->extra;
-    }
-    enum ggml_type           const vec_dot_type         = type_traits_cpu[type].vec_dot_type;
+    enum ggml_type           const vec_dot_type         = type_traits_cpu[src0->type].vec_dot_type;
     ggml_from_float_t        const from_float           = type_traits_cpu[vec_dot_type].from_float;
-    ggml_from_float_to_mat_t const from_float_to_mat    = type_traits_cpu[vec_dot_type].from_float_to_mat;
-    int64_t                  const vec_dot_num_rows     = type_traits_cpu[type].nrows;
-    int64_t                  const matmul_num_cols      = type_traits_cpu[type].ncols;
-    int64_t                  const blck_size_interleave = ggml_get_type_traits(type)->blck_size_interleave;
-    ggml_gemv_t              const gemv                 = type_traits_cpu[type].gemv;
-    ggml_gemm_t              const gemm                 = type_traits_cpu[type].gemm;
+    int64_t                  const vec_dot_num_rows     = type_traits_cpu[src0->type].nrows;
     GGML_ASSERT(ne0 == ne01);
     GGML_ASSERT(ne1 == ne11);
@@ -7454,7 +7396,7 @@ static void ggml_compute_forward_mul_mat(
     GGML_ASSERT(ne3 == ne13);
     // we don't support permuted src0 or src1
-    GGML_ASSERT(nb00 == ggml_type_size(type));
+    GGML_ASSERT(nb00 == ggml_type_size(src0->type));
     GGML_ASSERT(nb10 == ggml_type_size(src1->type));
     // dst cannot be transposed or permuted
@@ -7466,6 +7408,7 @@ static void ggml_compute_forward_mul_mat(
     // nb01 >= nb00 - src0 is not transposed
     //   compute by src0 rows
+    // TODO: extract to "extra_op"
 #if GGML_USE_LLAMAFILE
     // broadcast factors
     const int64_t r2 = ne12 / ne02;
@@ -7476,15 +7419,15 @@ static void ggml_compute_forward_mul_mat(
     if (src1_cont) {
         for (int64_t i13 = 0; i13 < ne13; i13++)
             for (int64_t i12 = 0; i12 < ne12; i12++)
-                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(type),
+                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
                                      (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
-                                     nb01/ggml_type_size(type),
+                                     nb01/ggml_type_size(src0->type),
                                      (const char *)src1->data + i12*nb12 + i13*nb13,
                                      nb11/ggml_type_size(src1->type),
                                      (char *)dst->data + i12*nb2 + i13*nb3,
                                      nb1/ggml_type_size(dst->type),
                                      ith, nth,
-                                     type,
+                                     src0->type,
                                      src1->type,
                                      dst->type))
                     goto UseGgmlGemm1;
@@ -7505,19 +7448,10 @@ UseGgmlGemm1:;
         for (int64_t i13 = 0; i13 < ne13; ++i13) {
             for (int64_t i12 = 0; i12 < ne12; ++i12) {
-                int64_t i11_processed = 0;
-                if ((ggml_n_dims(src1) == 2) && from_float_to_mat && gemm) {
-                    for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
-                        from_float_to_mat((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11),
-                                          (void *)               (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
-                                          4, ne10, blck_size_interleave);
-                    }
-                    i11_processed = ne11 - ne11 % 4;
-                }
-                for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) {
+                for (int64_t i11 = ith; i11 < ne11; i11 += nth) {
                     from_float((float *)((char *) src1->data + i13*nb13 + i12*nb12 + i11*nb11),
-                           (void *)               (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
-                           ne10);
+                               (void *)               (wdata + i13*nbw3 + i12*nbw2 + i11*nbw1),
+                                ne10);
                 }
             }
         }
@@ -7537,15 +7471,15 @@ UseGgmlGemm1:;
         for (int64_t i13 = 0; i13 < ne13; i13++)
             for (int64_t i12 = 0; i12 < ne12; i12++)
-                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(type),
+                if (!llamafile_sgemm(ne01, ne11, ne00/ggml_blck_size(src0->type),
                                      (const char *)src0->data + i12/r2*nb02 + i13/r3*nb03,
-                                     nb01/ggml_type_size(type),
+                                     nb01/ggml_type_size(src0->type),
                                      (const char *)wdata + (i12*ne11 + i13*ne12*ne11)*row_size,
                                      row_size/ggml_type_size(vec_dot_type),
                                      (char *)dst->data + i12*nb2 + i13*nb3,
                                      nb1/ggml_type_size(dst->type),
                                      ith, nth,
-                                     type,
+                                     src0->type,
                                      vec_dot_type,
                                      dst->type))
                     goto UseGgmlGemm2;
@@ -7560,14 +7494,6 @@ UseGgmlGemm2:;
     // This is the size of the rest of the dimensions of the result
     const int64_t nr1 = ne1 * ne2 * ne3;
-    // dot kernels can handle 1 row and col at a time, but mmla kernels can process 2 rows and cols
-    int64_t num_rows_per_vec_dot = vec_dot_num_rows;
-    // TODO: currently the mmla kernels support only even numbered rows/cols.
-    // this check can be removed once they are extended to support odd numbered rows/cols too
-    if ((nr0 % 2 != 0) || (ne11 % 2 != 0)) {
-        num_rows_per_vec_dot = 1;
-    }
     // Now select a reasonable chunk size.
     int chunk_size = 16;
@@ -7595,28 +7521,6 @@ UseGgmlGemm2:;
     const int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0;
     const int64_t dr1 = (nr1 + nchunk1 - 1) / nchunk1;
-    if ((ggml_n_dims(src0) == 2) && gemv) {
-        const void * src1_wdata      = (src1->type == vec_dot_type) ? src1->data : params->wdata;
-        const size_t src1_col_stride = ggml_is_contiguous(src1) || src1->type != vec_dot_type ? ggml_row_size(vec_dot_type, ne10) : nb11;
-        int64_t src0_start = (ith * ne01) / nth;
-        int64_t src0_end   = ((ith + 1) * ne01) / nth;
-        src0_start = (src0_start % matmul_num_cols) ? src0_start + matmul_num_cols - (src0_start % matmul_num_cols): src0_start;
-        src0_end   = (src0_end   % matmul_num_cols) ? src0_end   + matmul_num_cols - (src0_end   % matmul_num_cols): src0_end;
-        if (src0_start >= src0_end) return;
-        // If there are more than three rows in src1, use gemm; otherwise, use gemv.
-        if (gemm && (ne11 > 3)) {
-            gemm(ne00, (float *)((char *) dst->data) + src0_start, ne01, (const char *) src0->data + src0_start * nb01,
-                 (const char *) src1_wdata, ne11 - ne11 % 4, src0_end - src0_start);
-        }
-        for (int iter = gemm ? ne11 - ne11 % 4 : 0; iter < ne11; iter++) {
-            gemv(ne00, (float *)((char *) dst->data + (iter * nb1)) + src0_start, ne01,
-                 (const char *) src0->data + src0_start * nb01, (const char *) src1_wdata + (src1_col_stride * iter), 1,
-                 src0_end - src0_start);
-        }
-        return;
-    }
     // The first chunk comes from our thread_id, the rest will get auto-assigned.
     int current_chunk = ith;
@@ -7630,7 +7534,16 @@ UseGgmlGemm2:;
         const int64_t ir1_start = dr1 * ith1;
         const int64_t ir1_end = MIN(ir1_start + dr1, nr1);
-        ggml_compute_forward_mul_mat_one_chunk(params, dst, type, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
+        // dot kernels can handle 1 row and col at a time, but mmla kernels can process 2 rows and cols
+        int64_t num_rows_per_vec_dot = vec_dot_num_rows;
+        // these checks are needed to avoid crossing dim1 boundaries
+        // can be optimized, but the logic would become more complicated, so keeping it like this for simplicity
+        if ((nr0 % 2 != 0) || (ne11 % 2 != 0) || ((ir0_end - ir0_start) % 2 != 0) || ((ir1_end - ir1_start) % 2 != 0)) {
+            num_rows_per_vec_dot = 1;
+        }
+        ggml_compute_forward_mul_mat_one_chunk(params, dst, src0->type, num_rows_per_vec_dot, ir0_start, ir0_end, ir1_start, ir1_end);
         if (nth >= nchunk0 * nchunk1) {
             break;
@@ -7662,8 +7575,6 @@ static void ggml_compute_forward_mul_mat_id(
     ggml_vec_dot_t    const vec_dot         = type_traits_cpu[type].vec_dot;
     enum ggml_type    const vec_dot_type    = type_traits_cpu[type].vec_dot_type;
     ggml_from_float_t const from_float      = type_traits_cpu[vec_dot_type].from_float;
-    int64_t           const matmul_num_cols = type_traits_cpu[type].ncols;
-    ggml_gemv_t       const gemv            = type_traits_cpu[type].gemv;
     // we don't support permuted src0 or src1
     GGML_ASSERT(nb00 == ggml_type_size(type));
@@ -7749,34 +7660,6 @@ static void ggml_compute_forward_mul_mat_id(
         const int64_t nr0 = ne01; // src0 rows
         const int64_t nr1 = cne1; // src1 rows
-        if (((ggml_n_dims(src0) - 1) == 2) && gemv) {
-            int64_t src0_cur_start = (ith * ne01) / nth;
-            int64_t src0_cur_end   = ((ith + 1) * ne01) / nth;
-            src0_cur_start = (src0_cur_start % matmul_num_cols) ? src0_cur_start + matmul_num_cols - (src0_cur_start % matmul_num_cols): src0_cur_start;
-            src0_cur_end   = (src0_cur_end % matmul_num_cols) ? src0_cur_end + matmul_num_cols - (src0_cur_end % matmul_num_cols): src0_cur_end;
-            if (src0_cur_start >= src0_cur_end) return;
-            for (int ir1 = 0; ir1 < nr1; ir1++) {
-                struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1);
-                const int id       = row_mapping.i1; // selected expert index
-                const int64_t  i11 = id % ne11;
-                const int64_t  i12 = row_mapping.i2; // row index in src1
-                const int64_t  i1 = id;  // selected expert index
-                const int64_t  i2 = i12; // row
-                const char * src1_col = (const char *) wdata +
-                    (src1_cont || src1->type != vec_dot_type
-                    ? (i11        + i12 * ne11) * row_size
-                    : (i11 * nb11 + i12 * nb12));
-                gemv(ne00, (float *)((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01,
-                     (const char *) src0_cur + src0_cur_start * nb01, src1_col, 1, src0_cur_end - src0_cur_start);
-            }
-            continue;
-        }
         // distribute the thread work across the inner or outer loop based on which one is larger
         const int64_t nth0 = nr0 > nr1 ? nth : 1; // parallelize by src0 rows
@@ -8084,9 +7967,6 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_Q4_0_4_4:
-        case GGML_TYPE_Q4_0_4_8:
-        case GGML_TYPE_Q4_0_8_8:
             {
                 ggml_compute_forward_out_prod_q_f32(params, dst);
             } break;
@@ -8239,6 +8119,77 @@ static void ggml_compute_forward_set_f32(
     }
 }
+static void ggml_compute_forward_set_i32(
+        const struct ggml_compute_params * params,
+        struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    const struct ggml_tensor * src1 = dst->src[1];
+    GGML_ASSERT(ggml_are_same_shape(src0, dst));
+    GGML_ASSERT(ggml_is_contiguous(dst) && ggml_is_contiguous(src0));
+    // view src0 and dst with these strides and data offset inbytes during set
+    // nb0 is implicitly element_size because src0 and dst are contiguous
+    size_t nb1     = ((int32_t *) dst->op_params)[0];
+    size_t nb2     = ((int32_t *) dst->op_params)[1];
+    size_t nb3     = ((int32_t *) dst->op_params)[2];
+    size_t offset  = ((int32_t *) dst->op_params)[3];
+    bool   inplace = (bool) ((int32_t *) dst->op_params)[4];
+    if (!inplace) {
+        if (params->ith == 0) {
+            // memcpy needs to be synchronized across threads to avoid race conditions.
+            // => do it in INIT phase
+            memcpy(
+                ((char *)  dst->data),
+                ((char *) src0->data),
+                ggml_nbytes(dst));
+        }
+        ggml_barrier(params->threadpool);
+    }
+    const int ith = params->ith;
+    const int nth = params->nth;
+    const int nr = ggml_nrows(src1);
+    const int nc = src1->ne[0];
+    GGML_TENSOR_LOCALS(int64_t, ne1, src1, ne)
+    GGML_TENSOR_LOCALS(size_t,  nb1, src1, nb)
+    // src0 and dst as viewed during set
+    const size_t nb0 = ggml_element_size(src0);
+    const int im0 = (ne10 == 0 ? 0 : ne10-1);
+    const int im1 = (ne11 == 0 ? 0 : ne11-1);
+    const int im2 = (ne12 == 0 ? 0 : ne12-1);
+    const int im3 = (ne13 == 0 ? 0 : ne13-1);
+    GGML_ASSERT(offset + im0*nb0  + im1*nb1  + im2*nb2  + im3*nb3  <= ggml_nbytes(dst));
+    GGML_ASSERT(nb10 == sizeof(int32_t));
+    // rows per thread
+    const int dr = (nr + nth - 1)/nth;
+    // row range for this thread
+    const int ir0 = dr*ith;
+    const int ir1 = MIN(ir0 + dr, nr);
+    for (int ir = ir0; ir < ir1; ++ir) {
+        // src0 and dst are viewed with shape of src1 and offset
+        // => same indices
+        const int i3 = ir/(ne12*ne11);
+        const int i2 = (ir - i3*ne12*ne11)/ne11;
+        const int i1 = (ir - i3*ne12*ne11 - i2*ne11);
+        ggml_vec_cpy_i32(nc,
+                (int32_t *) ((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + offset),
+                (int32_t *) ((char *) src1->data + i3*nb13 + i2*nb12 + i1*nb11));
+    }
+}
 static void ggml_compute_forward_set(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst) {
@@ -8250,6 +8201,10 @@ static void ggml_compute_forward_set(
             {
                 ggml_compute_forward_set_f32(params, dst);
             } break;
+        case GGML_TYPE_I32:
+            {
+                ggml_compute_forward_set_i32(params, dst);
+            } break;
         case GGML_TYPE_F16:
         case GGML_TYPE_BF16:
         case GGML_TYPE_Q4_0:
@@ -8274,9 +8229,6 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_Q4_0_4_4:
-        case GGML_TYPE_Q4_0_4_8:
-        case GGML_TYPE_Q4_0_8_8:
         default:
             {
                 GGML_ABORT("fatal error");
@@ -8538,9 +8490,6 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_IQ4_XS:
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
-        case GGML_TYPE_Q4_0_4_4:
-        case GGML_TYPE_Q4_0_4_8:
-        case GGML_TYPE_Q4_0_8_8:
             {
                 ggml_compute_forward_get_rows_q(params, dst);
             } break;
@@ -9130,9 +9079,6 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_IQ3_S:
         case GGML_TYPE_IQ2_S:
         case GGML_TYPE_Q8_K:
-        case GGML_TYPE_Q4_0_4_4:
-        case GGML_TYPE_Q4_0_4_8:
-        case GGML_TYPE_Q4_0_8_8:
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
         case GGML_TYPE_I32:
@@ -9187,6 +9133,64 @@ static void ggml_rope_cache_init(
     }
 }
+static void ggml_mrope_cache_init(
+     float theta_base_t, float theta_base_h, float theta_base_w, float theta_base_e, int sections[4], bool indep_sects,
+     float freq_scale, const float * freq_factors, float corr_dims[2], int64_t ne0, float ext_factor, float mscale,
+     float * cache, float sin_sign, float theta_scale) {
+    // ref: https://github.com/jquesnelle/yarn/blob/master/scaled_rope/LlamaYaRNScaledRotaryEmbedding.py
+    float theta_t = theta_base_t;
+    float theta_h = theta_base_h;
+    float theta_w = theta_base_w;
+    float theta_e = theta_base_e;  // extra position id for vision encoder
+    int sect_dims = sections[0] + sections[1] + sections[2] + sections[3];
+    int sec_w = sections[1] + sections[0];
+    int sec_e = sections[2] + sec_w;
+    GGML_ASSERT(sect_dims <= ne0);
+    for (int64_t i0 = 0; i0 < ne0; i0 += 2) {
+        const float ff = freq_factors ? freq_factors[i0/2] : 1.0f;
+        int sector = (i0 / 2) % sect_dims;
+        if (indep_sects) {
+            // compute theta independently for each dim sections
+            // (i.e. reset corresponding theta when `i0` go from one section to another)
+            if (sector == 0) {
+                theta_t = theta_base_t;
+            }
+            else if (sector == sections[0]) {
+                theta_h = theta_base_h;;
+            }
+            else if (sector == sec_w) {
+                theta_w = theta_base_w;
+            }
+            else if (sector == sec_e) {
+                theta_e = theta_base_e;
+            }
+        }
+        float theta = theta_t;
+        if (sector >= sections[0] && sector < sec_w) {
+            theta = theta_h;
+        }
+        else if (sector >= sec_w && sector < sec_w + sections[2]) {
+            theta = theta_w;
+        }
+        else if (sector >= sec_w + sections[2]) {
+            theta = theta_e;
+        }
+        rope_yarn(
+            theta/ff, freq_scale, corr_dims, i0, ext_factor, mscale, &cache[i0 + 0], &cache[i0 + 1]
+        );
+        cache[i0 + 1] *= sin_sign;
+        theta_t *= theta_scale;
+        theta_w *= theta_scale;
+        theta_h *= theta_scale;
+        theta_e *= theta_scale;
+    }
+}
 static void ggml_compute_forward_rope_f32(
         const struct ggml_compute_params * params,
         struct ggml_tensor * dst,
@@ -9197,6 +9201,7 @@ static void ggml_compute_forward_rope_f32(
     const struct ggml_tensor * src2 = dst->src[2];
     float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+    int sections[4];
     //const int n_past     = ((int32_t *) dst->op_params)[0];
     const int n_dims     = ((int32_t *) dst->op_params)[1];
@@ -9210,6 +9215,7 @@ static void ggml_compute_forward_rope_f32(
     memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int)*4);
     GGML_TENSOR_UNARY_OP_LOCALS
@@ -9242,6 +9248,16 @@ static void ggml_compute_forward_rope_f32(
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;  // ggml_rope_multi, multimodal rotary position embedding
+    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+    if (is_mrope) {
+        GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
+    }
+    if (is_vision) {
+        GGML_ASSERT(n_dims == ne0/2);
+    }
     const float * freq_factors = NULL;
     if (src2 != NULL) {
@@ -9257,18 +9273,63 @@ static void ggml_compute_forward_rope_f32(
     const int32_t * pos = (const int32_t *) src1->data;
-    for (int64_t i3 = 0; i3 < ne3; i3++) {
-        for (int64_t i2 = 0; i2 < ne2; i2++) {
-            const int64_t p = pos[i2];
+    for (int64_t i3 = 0; i3 < ne3; i3++) { // batch
+        for (int64_t i2 = 0; i2 < ne2; i2++) { // seq-len
             float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
-            ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            if (!is_mrope) {
+                const int64_t p = pos[i2];
+                ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            }
+            else {
+                const int64_t p_t = pos[i2];
+                const int64_t p_h = pos[i2 + ne2];
+                const int64_t p_w = pos[i2 + ne2 * 2];
+                const int64_t p_e = pos[i2 + ne2 * 3];
+                ggml_mrope_cache_init(
+                    p_t, p_h, p_w, p_e, sections, is_vision,
+                    freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            }
-            for (int64_t i1 = 0; i1 < ne1; i1++) {
+            for (int64_t i1 = 0; i1 < ne1; i1++) { // attn-heads
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
-                if (!is_neox) {
+                if (is_neox || is_mrope) {
+                    if (is_vision){
+                        for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                            const int64_t ic = i0/2;
+                            const float cos_theta = cache[i0 + 0];
+                            const float sin_theta = cache[i0 + 1];
+                            const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+                            float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+                            const float x0 = src[0];
+                            const float x1 = src[n_dims];
+                            dst_data[0]      = x0*cos_theta - x1*sin_theta;
+                            dst_data[n_dims] = x0*sin_theta + x1*cos_theta;
+                        }
+                    } else {
+                        for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                            const int64_t ic = i0/2;
+                            const float cos_theta = cache[i0 + 0];
+                            const float sin_theta = cache[i0 + 1];
+                            const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+                            float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+                            const float x0 = src[0];
+                            const float x1 = src[n_dims/2];
+                            dst_data[0]        = x0*cos_theta - x1*sin_theta;
+                            dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
+                        }
+                    }
+                } else {
                     for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
                         const float cos_theta = cache[i0 + 0];
                         const float sin_theta = cache[i0 + 1];
@@ -9282,8 +9343,10 @@ static void ggml_compute_forward_rope_f32(
                         dst_data[0] = x0*cos_theta - x1*sin_theta;
                         dst_data[1] = x0*sin_theta + x1*cos_theta;
                     }
-                } else {
-                    for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                }
+                if (is_vision) {
+                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
                         const int64_t ic = i0/2;
                         const float cos_theta = cache[i0 + 0];
@@ -9293,19 +9356,20 @@ static void ggml_compute_forward_rope_f32(
                         float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
                         const float x0 = src[0];
-                        const float x1 = src[n_dims/2];
+                        const float x1 = src[n_dims];
-                        dst_data[0]        = x0*cos_theta - x1*sin_theta;
-                        dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
+                        dst_data[0]      = x0*cos_theta - x1*sin_theta;
+                        dst_data[n_dims] = x0*sin_theta + x1*cos_theta;
                     }
-                }
-                for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
-                    const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                    float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+                } else {
+                    // fill the remain channels with data from src tensor
+                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
+                        const float * const src = (float *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                        float * dst_data  = (float *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-                    dst_data[0] = src[0];
-                    dst_data[1] = src[1];
+                        dst_data[0] = src[0];
+                        dst_data[1] = src[1];
+                    }
                 }
             }
         }
@@ -9323,6 +9387,7 @@ static void ggml_compute_forward_rope_f16(
     const struct ggml_tensor * src2 = dst->src[2];
     float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
+    int sections[4];
     //const int n_past     = ((int32_t *) dst->op_params)[0];
     const int n_dims     = ((int32_t *) dst->op_params)[1];
@@ -9335,6 +9400,8 @@ static void ggml_compute_forward_rope_f16(
     memcpy(&attn_factor, (int32_t *) dst->op_params +  8, sizeof(float));
     memcpy(&beta_fast,   (int32_t *) dst->op_params +  9, sizeof(float));
     memcpy(&beta_slow,   (int32_t *) dst->op_params + 10, sizeof(float));
+    memcpy(&sections,    (int32_t *) dst->op_params + 11, sizeof(int)*4);
     GGML_TENSOR_UNARY_OP_LOCALS
@@ -9367,6 +9434,16 @@ static void ggml_compute_forward_rope_f16(
     ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
     const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
+    const bool is_mrope = mode & GGML_ROPE_TYPE_MROPE;
+    const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
+    if (is_mrope) {
+        GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
+    }
+    if (is_vision) {
+        GGML_ASSERT(n_dims == ne0/2);
+    }
     const float * freq_factors = NULL;
     if (src2 != NULL) {
@@ -9384,16 +9461,61 @@ static void ggml_compute_forward_rope_f16(
     for (int64_t i3 = 0; i3 < ne3; i3++) {
         for (int64_t i2 = 0; i2 < ne2; i2++) {
-            const int64_t p = pos[i2];
             float * cache = (float *) params->wdata + (ne0 + CACHE_LINE_SIZE_F32)*ith;
-            ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            if (!is_mrope) {
+                const int64_t p = pos[i2];
+                ggml_rope_cache_init(p, freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            }
+            else {
+                const int64_t p_t = pos[i2];
+                const int64_t p_h = pos[i2 + ne2];
+                const int64_t p_w = pos[i2 + ne2 * 2];
+                const int64_t p_e = pos[i2 + ne2 * 3];
+                ggml_mrope_cache_init(
+                    p_t, p_h, p_w, p_e, sections, is_vision,
+                    freq_scale, freq_factors, corr_dims, ne0, ext_factor, attn_factor, cache, sin_sign, theta_scale);
+            }
             for (int64_t i1 = 0; i1 < ne1; i1++) {
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
-                if (!is_neox) {
+                if (is_neox || is_mrope) {
+                    if (is_vision) {
+                        for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                            const int64_t ic = i0/2;
+                            const float cos_theta = cache[i0 + 0];
+                            const float sin_theta = cache[i0 + 1];
+                            const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+                            ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+                            const float x0 = GGML_FP16_TO_FP32(src[0]);
+                            const float x1 = GGML_FP16_TO_FP32(src[n_dims]);
+                            dst_data[0]      = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                            dst_data[n_dims] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                        }
+                    } else {
+                        for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                            const int64_t ic = i0/2;
+                            const float cos_theta = cache[i0 + 0];
+                            const float sin_theta = cache[i0 + 1];
+                            const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + ic*nb00);
+                            ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
+                            const float x0 = GGML_FP16_TO_FP32(src[0]);
+                            const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]);
+                            dst_data[0]        = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                            dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                        }
+                    }
+                } else {
                     for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
                         const float cos_theta = cache[i0 + 0];
                         const float sin_theta = cache[i0 + 1];
@@ -9407,8 +9529,10 @@ static void ggml_compute_forward_rope_f16(
                         dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
                         dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                     }
-                } else {
-                    for (int64_t i0 = 0; i0 < n_dims; i0 += 2) {
+                }
+                if (is_vision) {
+                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
                         const int64_t ic = i0/2;
                         const float cos_theta = cache[i0 + 0];
@@ -9418,19 +9542,19 @@ static void ggml_compute_forward_rope_f16(
                         ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + ic*nb0);
                         const float x0 = GGML_FP16_TO_FP32(src[0]);
-                        const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]);
+                        const float x1 = GGML_FP16_TO_FP32(src[n_dims]);
-                        dst_data[0]        = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                        dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                        dst_data[0]      = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                        dst_data[n_dims] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
                     }
-                }
-                for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
-                    const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
-                    ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
+                } else {
+                    for (int64_t i0 = n_dims; i0 < ne0; i0 += 2) {
+                        const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01 + i0*nb00);
+                        ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3  + i2*nb2  + i1*nb1  + i0*nb0);
-                    dst_data[0] = src[0];
-                    dst_data[1] = src[1];
+                        dst_data[0] = src[0];
+                        dst_data[1] = src[1];
+                    }
                 }
             }
         }
@@ -10429,6 +10553,40 @@ static void ggml_compute_forward_pad(
     }
 }
+// ggml_compute_forward_pad_reflect_1d
+static void ggml_compute_forward_pad_reflect_1d(
+        const struct ggml_compute_params * params,
+              struct ggml_tensor * dst) {
+    const struct ggml_tensor * src0 = dst->src[0];
+    GGML_ASSERT(src0->type == GGML_TYPE_F32);
+    GGML_ASSERT( dst->type == GGML_TYPE_F32);
+    const int ith = params->ith;
+    const int nth = params->nth;
+    const int32_t * opts = (const int32_t *) dst->op_params;
+    const int p0 = opts[0];
+    const int p1 = opts[1];
+    GGML_TENSOR_UNARY_OP_LOCALS
+    for (int64_t i3 = 0; i3 < ne3; i3++) {
+        for (int64_t i2 = 0; i2 < ne2; i2++) {
+            for (int64_t i1 = ith; i1 < ne1; i1 += nth) {
+                float * left  = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 +         p0*nb0);
+                float * right = (float *) ((char *) dst->data + i3*nb3 + i2*nb2 + i1*nb1 + (ne0-p1-1)*nb0);
+                ggml_vec_cpy_f32(ne00, left, (float *) ((char *) src0->data + i3*nb03 + i2*nb02 + i1*nb01));
+                for (int i0 = 1; i0 <= p0; i0++) { left[-i0] = left[i0];   }
+                for (int i0 = 1; i0 <= p1; i0++) { right[i0] = right[-i0]; }
+            }
+        }
+    }
+}
 // ggml_compute_forward_arange
@@ -12304,6 +12462,9 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
         return;
     }
+    // extra_buffer op?
+    if (ggml_cpu_extra_compute_forward(params, tensor)) return;
     switch (tensor->op) {
         case GGML_OP_DUP:
             {
@@ -12525,6 +12686,10 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
             {
                 ggml_compute_forward_pad(params, tensor);
             } break;
+        case GGML_OP_PAD_REFLECT_1D:
+            {
+                ggml_compute_forward_pad_reflect_1d(params, tensor);
+            } break;
         case GGML_OP_ARANGE:
             {
                 ggml_compute_forward_arange(params, tensor);
@@ -12867,6 +13032,7 @@ static int ggml_get_n_tasks(struct ggml_tensor * node, int n_threads) {
             } break;
         case GGML_OP_UPSCALE:
         case GGML_OP_PAD:
+        case GGML_OP_PAD_REFLECT_1D:
         case GGML_OP_ARANGE:
         case GGML_OP_TIMESTEP_EMBEDDING:
         case GGML_OP_ARGSORT:
@@ -12956,7 +13122,7 @@ static thread_ret_t ggml_graph_compute_secondary_thread(void* data);
 #include "windows.h"
 // TODO: support > 64 CPUs
-bool ggml_thread_apply_affinity(bool * mask) {
+static bool ggml_thread_apply_affinity(bool * mask) {
     HANDLE    h = GetCurrentThread();
     uint64_t  bitmask = 0ULL;
@@ -13246,140 +13412,142 @@ struct ggml_cplan ggml_graph_plan(
         size_t cur = 0;
-        switch (node->op) {
-            case GGML_OP_CPY:
-            case GGML_OP_DUP:
-                {
-                    if (ggml_is_quantized(node->type) ||
-                        // F16 -> BF16 and BF16 -> F16 copies go through intermediate F32
-                        (node->src[0]->type == GGML_TYPE_F16  && node->src[1] && node->src[1]->type == GGML_TYPE_BF16) ||
-                        (node->src[0]->type == GGML_TYPE_BF16 && node->src[1] && node->src[1]->type == GGML_TYPE_F16)) {
+        if (!ggml_cpu_extra_work_size(n_threads, node, &cur)) {
+            switch (node->op) {
+                case GGML_OP_CPY:
+                case GGML_OP_DUP:
+                    {
+                        if (ggml_is_quantized(node->type) ||
+                            // F16 -> BF16 and BF16 -> F16 copies go through intermediate F32
+                            (node->src[0]->type == GGML_TYPE_F16  && node->src[1] && node->src[1]->type == GGML_TYPE_BF16) ||
+                            (node->src[0]->type == GGML_TYPE_BF16 && node->src[1] && node->src[1]->type == GGML_TYPE_F16)) {
+                            cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
+                        }
+                    } break;
+                case GGML_OP_ADD:
+                case GGML_OP_ADD1:
+                    {
+                        if (ggml_is_quantized(node->src[0]->type)) {
+                            cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
+                        }
+                    } break;
+                case GGML_OP_ACC:
+                    {
+                        if (ggml_is_quantized(node->src[0]->type)) {
+                            cur = ggml_type_size(GGML_TYPE_F32) * node->src[1]->ne[0] * n_tasks;
+                        }
+                    } break;
+                case GGML_OP_COUNT_EQUAL:
+                    {
+                        cur = ggml_type_size(node->type)*n_tasks;
+                    } break;
+                case GGML_OP_MUL_MAT:
+                    {
+                        const enum ggml_type vec_dot_type = type_traits_cpu[node->src[0]->type].vec_dot_type;
+                        if (node->src[1]->type != vec_dot_type) {
+                            cur = ggml_row_size(vec_dot_type, ggml_nelements(node->src[1]));
+                        }
+                    } break;
+                case GGML_OP_MUL_MAT_ID:
+                    {
+                        cur = 0;
+                        const struct ggml_tensor * src0 = node->src[0];
+                        const struct ggml_tensor * src1 = node->src[1];
+                        const enum ggml_type vec_dot_type = type_traits_cpu[src0->type].vec_dot_type;
+                        if (src1->type != vec_dot_type) {
+                            cur += ggml_row_size(vec_dot_type, ggml_nelements(src1));
+                        }
+                        const int n_as = src0->ne[2];
+                        cur += GGML_PAD(cur, sizeof(int64_t));       // align
+                        cur += n_as * sizeof(int64_t);               // matrix_row_counts
+                        cur += n_as * src1->ne[2] * sizeof(int64_t); // matrix_rows
+                    } break;
+                case GGML_OP_OUT_PROD:
+                    {
+                        if (ggml_is_quantized(node->src[0]->type)) {
+                            cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
+                        }
+                    } break;
+                case GGML_OP_SOFT_MAX:
+                case GGML_OP_ROPE:
+                    {
                         cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
-                    }
-                } break;
-            case GGML_OP_ADD:
-            case GGML_OP_ADD1:
-                {
-                    if (ggml_is_quantized(node->src[0]->type)) {
-                        cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
-                    }
-                } break;
-            case GGML_OP_ACC:
-                {
-                    if (ggml_is_quantized(node->src[0]->type)) {
-                        cur = ggml_type_size(GGML_TYPE_F32) * node->src[1]->ne[0] * n_tasks;
-                    }
-                } break;
-            case GGML_OP_COUNT_EQUAL:
-                {
-                    cur = ggml_type_size(node->type)*n_tasks;
-                } break;
-            case GGML_OP_MUL_MAT:
-                {
-                    const enum ggml_type vec_dot_type = type_traits_cpu[node->src[0]->type].vec_dot_type;
+                    } break;
+                case GGML_OP_CONV_TRANSPOSE_1D:
+                    {
+                        GGML_ASSERT(node->src[0]->ne[3] == 1);
+                        GGML_ASSERT(node->src[1]->ne[2] == 1);
+                        GGML_ASSERT(node->src[1]->ne[3] == 1);
+                        const int64_t ne00 = node->src[0]->ne[0];  // K
+                        const int64_t ne01 = node->src[0]->ne[1];  // Cout
+                        const int64_t ne02 = node->src[0]->ne[2];  // Cin
+                        const int64_t ne10 = node->src[1]->ne[0];  // L
+                        const int64_t ne11 = node->src[1]->ne[1];  // Cin
+                        if ((node->src[0]->type == GGML_TYPE_F16 ||
+                             node->src[0]->type == GGML_TYPE_BF16) &&
+                            node->src[1]->type == GGML_TYPE_F32) {
+                            cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02;
+                            cur += sizeof(ggml_fp16_t)*ne10*ne11;
+                        } else if (node->src[0]->type == GGML_TYPE_F32 &&
+                                   node->src[1]->type == GGML_TYPE_F32) {
+                            cur += sizeof(float)*ne00*ne01*ne02;
+                            cur += sizeof(float)*ne10*ne11;
+                        } else {
+                            GGML_ABORT("fatal error");
+                        }
+                    } break;
+                case GGML_OP_CONV_TRANSPOSE_2D:
+                    {
+                        const int64_t ne00 = node->src[0]->ne[0]; // W
+                        const int64_t ne01 = node->src[0]->ne[1]; // H
+                        const int64_t ne02 = node->src[0]->ne[2]; // Channels Out
+                        const int64_t ne03 = node->src[0]->ne[3]; // Channels In
-                    if (node->src[1]->type != vec_dot_type) {
-                        cur = ggml_row_size(vec_dot_type, ggml_nelements(node->src[1]));
-                    }
-                } break;
-            case GGML_OP_MUL_MAT_ID:
-                {
-                    cur = 0;
-                    const struct ggml_tensor * src0 = node->src[0];
-                    const struct ggml_tensor * src1 = node->src[1];
-                    const enum ggml_type vec_dot_type = type_traits_cpu[src0->type].vec_dot_type;
-                    if (src1->type != vec_dot_type) {
-                        cur += ggml_row_size(vec_dot_type, ggml_nelements(src1));
-                    }
-                    const int n_as = src0->ne[2];
-                    cur += GGML_PAD(cur, sizeof(int64_t));       // align
-                    cur += n_as * sizeof(int64_t);               // matrix_row_counts
-                    cur += n_as * src1->ne[2] * sizeof(int64_t); // matrix_rows
-                } break;
-            case GGML_OP_OUT_PROD:
-                {
-                    if (ggml_is_quantized(node->src[0]->type)) {
-                        cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
-                    }
-                } break;
-            case GGML_OP_SOFT_MAX:
-            case GGML_OP_ROPE:
-                {
-                    cur = ggml_type_size(GGML_TYPE_F32) * node->ne[0] * n_tasks;
-                } break;
-            case GGML_OP_CONV_TRANSPOSE_1D:
-                {
-                    GGML_ASSERT(node->src[0]->ne[3] == 1);
-                    GGML_ASSERT(node->src[1]->ne[2] == 1);
-                    GGML_ASSERT(node->src[1]->ne[3] == 1);
-                    const int64_t ne00 = node->src[0]->ne[0];  // K
-                    const int64_t ne01 = node->src[0]->ne[1];  // Cout
-                    const int64_t ne02 = node->src[0]->ne[2];  // Cin
-                    const int64_t ne10 = node->src[1]->ne[0];  // L
-                    const int64_t ne11 = node->src[1]->ne[1];  // Cin
-                    if ((node->src[0]->type == GGML_TYPE_F16 ||
-                         node->src[0]->type == GGML_TYPE_BF16) &&
-                        node->src[1]->type == GGML_TYPE_F32) {
-                        cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02;
-                        cur += sizeof(ggml_fp16_t)*ne10*ne11;
-                    } else if (node->src[0]->type == GGML_TYPE_F32 &&
-                               node->src[1]->type == GGML_TYPE_F32) {
-                        cur += sizeof(float)*ne00*ne01*ne02;
-                        cur += sizeof(float)*ne10*ne11;
-                    } else {
-                        GGML_ABORT("fatal error");
-                    }
-                } break;
-            case GGML_OP_CONV_TRANSPOSE_2D:
-                {
-                    const int64_t ne00 = node->src[0]->ne[0]; // W
-                    const int64_t ne01 = node->src[0]->ne[1]; // H
-                    const int64_t ne02 = node->src[0]->ne[2]; // Channels Out
-                    const int64_t ne03 = node->src[0]->ne[3]; // Channels In
-                    const int64_t ne10 = node->src[1]->ne[0]; // W
-                    const int64_t ne11 = node->src[1]->ne[1]; // H
-                    const int64_t ne12 = node->src[1]->ne[2]; // Channels In
-                    cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03;
-                    cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12;
-                } break;
-            case GGML_OP_FLASH_ATTN_EXT:
-                {
-                    const int64_t ne00 = node->src[0]->ne[0]; // D
+                        const int64_t ne10 = node->src[1]->ne[0]; // W
+                        const int64_t ne11 = node->src[1]->ne[1]; // H
+                        const int64_t ne12 = node->src[1]->ne[2]; // Channels In
-                    cur = 3*sizeof(float)*ne00*n_tasks; // 3x head size/thread
-                } break;
-            case GGML_OP_FLASH_ATTN_BACK:
-                {
-                    const int64_t    D = node->src[0]->ne[0];
-                    const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL);
-                    const int64_t mxDn = MAX(D, ne11) * 2; // *2 because of S and SM in ggml_compute_forward_flash_attn_back
-                    if (node->src[1]->type == GGML_TYPE_F32) {
-                        cur  = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
-                        cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
-                    } else if (node->src[1]->type == GGML_TYPE_F16) {
-                        cur  = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
-                        cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
-                    } else if (node->src[1]->type == GGML_TYPE_BF16) {
-                        cur  = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
-                        cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
-                    }
-                } break;
+                        cur += sizeof(ggml_fp16_t)*ne00*ne01*ne02*ne03;
+                        cur += sizeof(ggml_fp16_t)*ne10*ne11*ne12;
+                    } break;
+                case GGML_OP_FLASH_ATTN_EXT:
+                    {
+                        const int64_t ne00 = node->src[0]->ne[0]; // D
-            case GGML_OP_CROSS_ENTROPY_LOSS:
-                {
-                    cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks);
-                } break;
-            case GGML_OP_COUNT:
-                {
-                    GGML_ABORT("fatal error");
-                }
-            default:
-                break;
+                        cur = 3*sizeof(float)*ne00*n_tasks; // 3x head size/thread
+                    } break;
+                case GGML_OP_FLASH_ATTN_BACK:
+                    {
+                        const int64_t    D = node->src[0]->ne[0];
+                        const int64_t ne11 = ggml_up(node->src[1]->ne[1], GGML_SOFT_MAX_UNROLL);
+                        const int64_t mxDn = MAX(D, ne11) * 2; // *2 because of S and SM in ggml_compute_forward_flash_attn_back
+                        if (node->src[1]->type == GGML_TYPE_F32) {
+                            cur  = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
+                            cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
+                        } else if (node->src[1]->type == GGML_TYPE_F16) {
+                            cur  = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
+                            cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
+                        } else if (node->src[1]->type == GGML_TYPE_BF16) {
+                            cur  = sizeof(float)*mxDn*n_tasks; // TODO: this can become (n_tasks-1)
+                            cur += sizeof(float)*mxDn*n_tasks; // this is overestimated by x2
+                        }
+                    } break;
+                case GGML_OP_CROSS_ENTROPY_LOSS:
+                    {
+                        cur = ggml_type_size(node->type)*(n_tasks + node->src[0]->ne[0]*n_tasks);
+                    } break;
+                case GGML_OP_COUNT:
+                    {
+                        GGML_ABORT("fatal error");
+                    }
+                default:
+                    break;
+            }
         }
         work_size = MAX(work_size, cur);
@@ -13578,29 +13746,6 @@ static void ggml_graph_compute_kickoff(struct ggml_threadpool * threadpool, int
 #endif // GGML_USE_OPENMP
-void ggml_threadpool_params_init(struct ggml_threadpool_params * p, int n_threads) {
-    p->n_threads  = n_threads;
-    p->prio       = 0;     // default priority (usually means normal or inherited)
-    p->poll       = 50;    // hybrid-polling enabled
-    p->strict_cpu = false; // no strict placement (all threads share same cpumask)
-    p->paused     = false; // threads are ready to go
-    memset(p->cpumask, 0, GGML_MAX_N_THREADS); // all-zero means use the default affinity (usually inherited)
-}
-struct ggml_threadpool_params ggml_threadpool_params_default(int n_threads) {
-    struct ggml_threadpool_params p;
-    ggml_threadpool_params_init(&p, n_threads);
-    return p;
-}
-bool ggml_threadpool_params_match(const struct ggml_threadpool_params * p0, const struct ggml_threadpool_params * p1) {
-    if (p0->n_threads      != p1->n_threads  )    return false;
-    if (p0->prio           != p1->prio       )    return false;
-    if (p0->poll           != p1->poll       )    return false;
-    if (p0->strict_cpu     != p1->strict_cpu )    return false;
-    return memcmp(p0->cpumask, p1->cpumask, GGML_MAX_N_THREADS) == 0;
-}
 static struct ggml_threadpool * ggml_threadpool_new_impl(
     struct ggml_threadpool_params * tpp,
                struct ggml_cgraph * cgraph,
@@ -13896,15 +14041,23 @@ int ggml_cpu_has_vsx(void) {
 }
 int ggml_cpu_has_neon(void) {
-#if defined(__ARM_ARCH)
+#if defined(__ARM_ARCH) && defined(__ARM_NEON)
     return ggml_arm_arch_features.has_neon;
 #else
     return 0;
 #endif
 }
+int ggml_cpu_has_dotprod(void) {
+#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_DOTPROD)
+    return ggml_arm_arch_features.has_dotprod;
+#else
+    return 0;
+#endif
+}
 int ggml_cpu_has_sve(void) {
-#if defined(__ARM_ARCH)
+#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE)
     return ggml_arm_arch_features.has_sve;
 #else
     return 0;
@@ -13912,7 +14065,7 @@ int ggml_cpu_has_sve(void) {
 }
 int ggml_cpu_has_matmul_int8(void) {
-#if defined(__ARM_ARCH)
+#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_MATMUL_INT8)
     return ggml_arm_arch_features.has_i8mm;
 #else
     return 0;
@@ -13920,7 +14073,7 @@ int ggml_cpu_has_matmul_int8(void) {
 }
 int ggml_cpu_get_sve_cnt(void) {
-#if defined(__ARM_ARCH)
+#if defined(__ARM_ARCH) && defined(__ARM_FEATURE_SVE)
     return ggml_arm_arch_features.sve_cnt;
 #else
     return 0;