llama_cpp 0.14.6 → 0.14.7

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 5c4bd6bcb93b98a00f94dcdf93d04f853174f73e281d96fce8f837a6ba7f250e
-  data.tar.gz: 6d184e9ce927c06ba794bea63a09007a175a72e477366ffb1c5763ceb2c7c71e
+  metadata.gz: 243241c78383cb68d4fb5027ffc54ea7f6789bd74bfe85fae8e62d45e7c3145d
+  data.tar.gz: b7c792c6fb2287b71a72ff823a31706dc0830aa704d86e6f8a92d1d0630649d9
 SHA512:
-  metadata.gz: 953fe2777a759e5467694b8afb9d3f929a42603e81b2c3e38ba0fda4bb6dca78b2d147345023f99c2c9fb899cc746bf6729ad2726c2cb473d7094e93c13caf73
-  data.tar.gz: 71eb3cd5a5c619e9cc8a3418be745a8b76dc5e8cabe5b26a766230a8533df9a11c3981601b0be4ec0adb34a49f86ad741503ffc9f3b0d7ba021a7e9ddc3246a7
+  metadata.gz: 59565cd5e6bd79d98d31dcf1ce505c8388a97296f607c2a114cf92a614a2cd39291a8a18a3f58993606ea3f0970d1eadbfe670280c5261c5826a54d77a2eb85d
+  data.tar.gz: 228bc19181b0163ef922e847f67e7b6a52dc1311c4e8173586dfca82eb402c5a08c104b5bac5ba0eee4772f615f8fd17f2d06cbc6db5323d133a46d3de85eeb4

data/CHANGELOG.md

@@ -1,3 +1,12 @@
+## [[0.14.7](https://github.com/yoshoku/llama_cpp.rb/compare/v0.14.6...v0.14.7)] - 2024-04-27
+
+- Bump llama.cpp from b2698 to b2740.
+  - Add `keep_split` accessor to `ModelQuantizeParams`.
+  - Add `pooling_type` method to `Context`.
+  - Add `token_is_eog?` method to `Model`.
+
+Implementation binding for llama_sample_token_with_rng has been skipped.
+
 ## [[0.14.6](https://github.com/yoshoku/llama_cpp.rb/compare/v0.14.5...v0.14.6)] - 2024-04-20
 
 - Bump llama.cpp from b2658 to b2698.

data/ext/llama_cpp/llama_cpp.cpp

@@ -1321,6 +1321,8 @@ public:
     rb_define_method(rb_cLLaMAModelQuantizeParams, "only_copy", RUBY_METHOD_FUNC(_llama_model_quantize_params_get_only_copy), 0);
     rb_define_method(rb_cLLaMAModelQuantizeParams, "pure=", RUBY_METHOD_FUNC(_llama_model_quantize_params_set_pure), 1);
     rb_define_method(rb_cLLaMAModelQuantizeParams, "pure", RUBY_METHOD_FUNC(_llama_model_quantize_params_get_pure), 0);
+    rb_define_method(rb_cLLaMAModelQuantizeParams, "keep_split=", RUBY_METHOD_FUNC(_llama_model_quantize_params_set_keep_split), 1);
+    rb_define_method(rb_cLLaMAModelQuantizeParams, "keep_split", RUBY_METHOD_FUNC(_llama_model_quantize_params_get_keep_split), 0);
   }
 
 private:
@@ -1405,6 +1407,18 @@ private:
     LLaMAModelQuantizeParamsWrapper* ptr = get_llama_model_quantize_params(self);
     return ptr->params.pure ? Qtrue : Qfalse;
   }
+
+  // keep_split
+  static VALUE _llama_model_quantize_params_set_keep_split(VALUE self, VALUE keep_split) {
+    LLaMAModelQuantizeParamsWrapper* ptr = get_llama_model_quantize_params(self);
+    ptr->params.keep_split = RTEST(keep_split) ? true : false;
+    return ptr->params.keep_split ? Qtrue : Qfalse;
+  }
+
+  static VALUE _llama_model_quantize_params_get_keep_split(VALUE self) {
+    LLaMAModelQuantizeParamsWrapper* ptr = get_llama_model_quantize_params(self);
+    return ptr->params.keep_split ? Qtrue : Qfalse;
+  }
 };
 
 const rb_data_type_t RbLLaMAModelQuantizeParams::llama_model_quantize_params_type = {
@@ -1487,6 +1501,7 @@ public:
     rb_define_method(rb_cLLaMAModel, "token_middle", RUBY_METHOD_FUNC(_llama_model_token_middle), 0);
     rb_define_method(rb_cLLaMAModel, "token_suffix", RUBY_METHOD_FUNC(_llama_model_token_suffix), 0);
     rb_define_method(rb_cLLaMAModel, "token_eot", RUBY_METHOD_FUNC(_llama_model_token_eot), 0);
+    rb_define_method(rb_cLLaMAModel, "token_is_eog?", RUBY_METHOD_FUNC(_llama_model_token_is_eog), 1);
   }
 
 private:
@@ -1634,10 +1649,10 @@ private:
     const llama_token token = NUM2INT(token_);
     LLaMAModelWrapper* ptr = get_llama_model(self);
     std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_piece(ptr->model, token, result.data(), result.size());
+    const int n_tokens = llama_token_to_piece(ptr->model, token, result.data(), result.size(), false);
     if (n_tokens < 0) {
       result.resize(-n_tokens);
-      const int check = llama_token_to_piece(ptr->model, token, result.data(), result.size());
+      const int check = llama_token_to_piece(ptr->model, token, result.data(), result.size(), false);
       if (check != -n_tokens) {
         rb_raise(rb_eRuntimeError, "failed to convert");
         return Qnil;
@@ -1789,6 +1804,16 @@ private:
     LLaMAModelWrapper* ptr = get_llama_model(self);
     return INT2NUM(llama_token_eot(ptr->model));
   }
+
+  static VALUE _llama_model_token_is_eog(VALUE self, VALUE token_) {
+    if (!RB_INTEGER_TYPE_P(token_)) {
+      rb_raise(rb_eArgError, "token must be an integer");
+      return Qnil;
+    }
+    const llama_token token = NUM2INT(token_);
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return llama_token_is_eog(ptr->model, token) ? Qtrue : Qfalse;
+  }
 };
 
 const rb_data_type_t RbLLaMAModel::llama_model_type = {
@@ -2102,6 +2127,7 @@ public:
     rb_define_method(rb_cLLaMAContext, "sample_grammar", RUBY_METHOD_FUNC(_llama_context_sample_grammar), -1);
     rb_define_method(rb_cLLaMAContext, "grammar_accept_token", RUBY_METHOD_FUNC(_llama_context_grammar_accept_token), -1);
     rb_define_method(rb_cLLaMAContext, "apply_control_vector", RUBY_METHOD_FUNC(_llama_context_apply_control_vector), -1);
+    rb_define_method(rb_cLLaMAContext, "pooling_type", RUBY_METHOD_FUNC(_llama_context_pooling_type), 0);
   }
 
 private:
@@ -3225,6 +3251,15 @@ private:
 
     return Qnil;
   }
+
+  static VALUE _llama_context_pooling_type(VALUE self) {
+    LLaMAContextWrapper* ptr = get_llama_context(self);
+    if (ptr->ctx == NULL) {
+      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
+      return Qnil;
+    }
+    return INT2NUM(static_cast<int>(llama_pooling_type(ptr->ctx)));
+  }
 };
 
 const rb_data_type_t RbLLaMAContext::llama_context_type = {

data/lib/llama_cpp/version.rb

@@ -3,8 +3,8 @@
 # llama_cpp.rb provides Ruby bindings for the llama.cpp.
 module LLaMACpp
   # The version of llama_cpp.rb you install.
-  VERSION = '0.14.6'
+  VERSION = '0.14.7'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'b2698'
+  LLAMA_CPP_VERSION = 'b2740'
 end

data/sig/llama_cpp.rbs

@@ -141,6 +141,7 @@ module LLaMACpp
     def token_middle: () -> Integer
     def token_suffix: () -> Integer
     def token_eot: () -> Integer
+    def token_is_eog?: (Integer) -> bool
   end
 
   class Timings
@@ -260,6 +261,7 @@ module LLaMACpp
     def sample_grammar: (::LLaMACpp::TokenDataArray, grammar: ::LLaMACpp::Grammar) -> void
     def grammar_accept_token: (grammar: ::LLaMACpp::Grammar, token: Integer) -> void
     def apply_control_vector: (data: Array[Float], n_embd: Integer, il_start: Integer, il_end: Integer) -> void
+    def pooling_type: () -> Integer
   end
 
   class ContextParams
@@ -328,6 +330,8 @@ module LLaMACpp
     def only_copy=: (bool) -> bool
     def pure: () -> bool
     def pure=: (bool) -> bool
+    def keep_split: () -> bool
+    def keep_split=: (bool) -> bool
   end
 
   class Params = ContextParams

data/vendor/tmp/llama.cpp/Makefile

@@ -386,10 +386,6 @@ ifdef LLAMA_OPENBLAS
 	MK_LDFLAGS  += $(shell pkg-config --libs openblas)
 endif # LLAMA_OPENBLAS
 
-# TODO: temporary disable until MoE is fixed
-#       https://github.com/ggerganov/llama.cpp/pull/6716
-LLAMA_NO_LLAMAFILE := 1
-
 ifndef LLAMA_NO_LLAMAFILE
 	MK_CPPFLAGS += -DGGML_USE_LLAMAFILE
 	OBJS        += sgemm.o
@@ -701,7 +697,7 @@ OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o unicode-data.o
 llama.o: llama.cpp unicode.h ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-COMMON_H_DEPS = common/common.h common/sampling.h common/log.h
+COMMON_H_DEPS = common/common.h common/sampling.h common/log.h llama.h
 COMMON_DEPS   = common.o sampling.o grammar-parser.o build-info.o json-schema-to-grammar.o
 
 common.o: common/common.cpp $(COMMON_H_DEPS)
@@ -805,10 +801,19 @@ save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(C
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h common/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/utils.hpp examples/server/httplib.h common/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/server/json-schema-to-grammar.mjs.hpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h %.hpp $<,$^) -Iexamples/server $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) $(LWINSOCK2)
 
+# Portable equivalent of `cd examples/server/public && xxd -i $(notdir $<) ../$(notdir $<).hpp`:
+examples/server/%.hpp: examples/server/public/% Makefile
+	@( export NAME=$(subst .,_,$(subst -,_,$(notdir $<))) && \
+		echo "unsigned char $${NAME}[] = {" && \
+		cat $< | od -v -t x1 -An | sed -E 's/([0-9a-fA-F]+)/0x\1, /g' && \
+		echo "};" && \
+		echo "unsigned int $${NAME}_len = $(shell cat $< | wc -c );" \
+	) > $@
+
 gguf: examples/gguf/gguf.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)

data/vendor/tmp/llama.cpp/ggml-alloc.c

@@ -371,16 +371,16 @@ struct ggml_gallocr {
 };
 
 ggml_gallocr_t ggml_gallocr_new_n(ggml_backend_buffer_type_t * bufts, int n_bufs) {
-    ggml_gallocr_t galloc = (ggml_gallocr_t)calloc(sizeof(struct ggml_gallocr), 1);
+    ggml_gallocr_t galloc = (ggml_gallocr_t)calloc(1, sizeof(struct ggml_gallocr));
     GGML_ASSERT(galloc != NULL);
 
-    galloc->bufts = calloc(sizeof(ggml_backend_buffer_type_t) * n_bufs, 1);
+    galloc->bufts = calloc(n_bufs, sizeof(ggml_backend_buffer_type_t));
     GGML_ASSERT(galloc->bufts != NULL);
 
-    galloc->buffers = calloc(sizeof(ggml_backend_buffer_t) * n_bufs, 1);
+    galloc->buffers = calloc(n_bufs, sizeof(ggml_backend_buffer_t) * n_bufs);
     GGML_ASSERT(galloc->buffers != NULL);
 
-    galloc->buf_tallocs = calloc(sizeof(struct ggml_dyn_tallocr *) * n_bufs, 1);
+    galloc->buf_tallocs = calloc(n_bufs, sizeof(struct ggml_dyn_tallocr *));
     GGML_ASSERT(galloc->buf_tallocs != NULL);
 
     for (int i = 0; i < n_bufs; i++) {
@@ -646,8 +646,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
         free(galloc->hash_set.keys);
         free(galloc->hash_values);
         galloc->hash_set.size = hash_size;
-        galloc->hash_set.keys = calloc(sizeof(struct ggml_tensor *), hash_size);
-        galloc->hash_values   = calloc(sizeof(struct hash_node), hash_size);
+        galloc->hash_set.keys = calloc(hash_size, sizeof(struct ggml_tensor *));
+        galloc->hash_values   = calloc(hash_size, sizeof(struct hash_node));
         GGML_ASSERT(galloc->hash_set.keys != NULL);
         GGML_ASSERT(galloc->hash_values != NULL);
     } else {
@@ -667,7 +667,7 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
     // set the node_allocs from the hash table
     if (galloc->n_nodes < graph->n_nodes) {
         free(galloc->node_allocs);
-        galloc->node_allocs = calloc(sizeof(struct node_alloc), graph->n_nodes);
+        galloc->node_allocs = calloc(graph->n_nodes, sizeof(struct node_alloc));
         GGML_ASSERT(galloc->node_allocs != NULL);
     }
     galloc->n_nodes = graph->n_nodes;
@@ -697,7 +697,7 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
     }
     if (galloc->n_leafs < graph->n_leafs) {
         free(galloc->leaf_allocs);
-        galloc->leaf_allocs = calloc(sizeof(galloc->leaf_allocs[0]), graph->n_leafs);
+        galloc->leaf_allocs = calloc(graph->n_leafs, sizeof(galloc->leaf_allocs[0]));
         GGML_ASSERT(galloc->leaf_allocs != NULL);
     }
     galloc->n_leafs = graph->n_leafs;

data/vendor/tmp/llama.cpp/ggml-backend.c

@@ -822,7 +822,11 @@ GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t
 GGML_CALL static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
     switch (op->op) {
         case GGML_OP_CPY:
-            return op->type != GGML_TYPE_IQ2_XXS && op->type != GGML_TYPE_IQ2_XS && op->type != GGML_TYPE_IQ1_S; // missing type_traits.from_float
+            return
+                op->type != GGML_TYPE_IQ2_XXS &&
+                op->type != GGML_TYPE_IQ2_XS  &&
+                op->type != GGML_TYPE_IQ1_S   &&
+                op->type != GGML_TYPE_IQ1_M; // missing type_traits.from_float
         case GGML_OP_MUL_MAT:
             return op->src[1]->type == GGML_TYPE_F32 || op->src[1]->type == ggml_internal_get_type_traits(op->src[0]->type).vec_dot_type;
         default:
@@ -1721,23 +1725,23 @@ ggml_backend_sched_t ggml_backend_sched_new(
     GGML_ASSERT(n_backends <= GGML_SCHED_MAX_BACKENDS);
     GGML_ASSERT(ggml_backend_is_cpu(backends[n_backends - 1])); // last backend must be CPU
 
-    struct ggml_backend_sched * sched = calloc(sizeof(struct ggml_backend_sched), 1);
+    struct ggml_backend_sched * sched = calloc(1, sizeof(struct ggml_backend_sched));
 
     // initialize hash table
     sched->hash_set          = ggml_hash_set_new(graph_size);
-    sched->tensor_backend_id = calloc(sizeof(sched->tensor_backend_id[0]), sched->hash_set.size);
-    sched->tensor_copies     = calloc(sizeof(sched->tensor_copies[0]), sched->hash_set.size);
+    sched->tensor_backend_id = calloc(sched->hash_set.size, sizeof(sched->tensor_backend_id[0]));
+    sched->tensor_copies     = calloc(sched->hash_set.size, sizeof(sched->tensor_copies[0]));
 
     const size_t nodes_size = graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2;
-    sched->node_backend_ids  = calloc(sizeof(sched->node_backend_ids[0]), nodes_size);
-    sched->leaf_backend_ids  = calloc(sizeof(sched->leaf_backend_ids[0]), nodes_size);
+    sched->node_backend_ids  = calloc(nodes_size, sizeof(sched->node_backend_ids[0]));
+    sched->leaf_backend_ids  = calloc(nodes_size, sizeof(sched->leaf_backend_ids[0]));
 
     sched->n_backends = n_backends;
 
     sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
 
     const int initial_splits_capacity = 16;
-    sched->splits = calloc(sizeof(sched->splits[0]), initial_splits_capacity);
+    sched->splits = calloc(initial_splits_capacity, sizeof(sched->splits[0]));
     sched->splits_capacity = initial_splits_capacity;
 
     for (int b = 0; b < n_backends; b++) {
@@ -1968,10 +1972,10 @@ static void graph_copy_init_tensor(struct ggml_hash_set hash_set, struct ggml_te
 struct ggml_backend_graph_copy ggml_backend_graph_copy(ggml_backend_t backend, struct ggml_cgraph * graph) {
     struct ggml_hash_set hash_set = {
         /* .size = */ graph->visited_hash_table.size,
-        /* .keys = */ calloc(sizeof(hash_set.keys[0]), graph->visited_hash_table.size) // NOLINT
+        /* .keys = */ calloc(graph->visited_hash_table.size, sizeof(hash_set.keys[0])) // NOLINT
     };
-    struct ggml_tensor ** node_copies = calloc(sizeof(node_copies[0]), hash_set.size); // NOLINT
-    bool * node_init = calloc(sizeof(node_init[0]), hash_set.size);
+    struct ggml_tensor ** node_copies = calloc(hash_set.size, sizeof(node_copies[0])); // NOLINT
+    bool * node_init = calloc(hash_set.size, sizeof(node_init[0]));
 
     struct ggml_init_params params = {
         /* .mem_size   = */ ggml_tensor_overhead()*hash_set.size + ggml_graph_overhead_custom(graph->size, false),

data/vendor/tmp/llama.cpp/ggml-impl.h

@@ -11,6 +11,12 @@
 #include <string.h> // memcpy
 #include <math.h>   // fabsf
 
+#undef MIN
+#undef MAX
+
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
 #ifdef __cplusplus
 extern "C" {
 #endif
@@ -45,7 +51,7 @@ extern "C" {
 // 16-bit float
 // on Arm, we use __fp16
 // on x86, we use uint16_t
-#if defined(__ARM_NEON) && !defined(_MSC_VER)
+#if defined(__ARM_NEON)
 
 // if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
 //
@@ -53,8 +59,262 @@ extern "C" {
 //
 #include <arm_neon.h>
 
+#ifdef _MSC_VER
+
+typedef uint16_t ggml_fp16_internal_t;
+
+#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
+
+#else
+
 typedef __fp16 ggml_fp16_internal_t;
 
+#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
+
+#endif // _MSC_VER
+
+#if !defined(__aarch64__)
+
+// 32-bit ARM compatibility
+
+// vaddvq_s16
+// vpaddq_s16
+// vpaddq_s32
+// vaddvq_s32
+// vaddvq_f32
+// vmaxvq_f32
+// vcvtnq_s32_f32
+// vzip1_u8
+// vzip2_u8
+
+inline static int32_t vaddvq_s16(int16x8_t v) {
+    return
+        (int32_t)vgetq_lane_s16(v, 0) + (int32_t)vgetq_lane_s16(v, 1) +
+        (int32_t)vgetq_lane_s16(v, 2) + (int32_t)vgetq_lane_s16(v, 3) +
+        (int32_t)vgetq_lane_s16(v, 4) + (int32_t)vgetq_lane_s16(v, 5) +
+        (int32_t)vgetq_lane_s16(v, 6) + (int32_t)vgetq_lane_s16(v, 7);
+}
+
+inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
+    int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
+    int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
+    return vcombine_s16(a0, b0);
+}
+
+inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
+    int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
+    int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
+    return vcombine_s32(a0, b0);
+}
+
+inline static int32_t vaddvq_s32(int32x4_t v) {
+    return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
+}
+
+inline static float vaddvq_f32(float32x4_t v) {
+    return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
+}
+
+inline static float vmaxvq_f32(float32x4_t v) {
+    return
+        MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
+            MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
+}
+
+inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
+    int32x4_t res;
+
+    res[0] = roundf(vgetq_lane_f32(v, 0));
+    res[1] = roundf(vgetq_lane_f32(v, 1));
+    res[2] = roundf(vgetq_lane_f32(v, 2));
+    res[3] = roundf(vgetq_lane_f32(v, 3));
+
+    return res;
+}
+
+inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[0]; res[1] = b[0];
+    res[2] = a[1]; res[3] = b[1];
+    res[4] = a[2]; res[5] = b[2];
+    res[6] = a[3]; res[7] = b[3];
+
+    return res;
+}
+
+inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
+    uint8x8_t res;
+
+    res[0] = a[4]; res[1] = b[4];
+    res[2] = a[5]; res[3] = b[5];
+    res[4] = a[6]; res[5] = b[6];
+    res[6] = a[7]; res[7] = b[7];
+
+    return res;
+}
+
+// vld1q_s16_x2
+// vld1q_u8_x2
+// vld1q_u8_x4
+// vld1q_s8_x2
+// vld1q_s8_x4
+// TODO: double-check these work correctly
+
+typedef struct ggml_int16x8x2_t {
+    int16x8_t val[2];
+} ggml_int16x8x2_t;
+
+inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
+    ggml_int16x8x2_t res;
+
+    res.val[0] = vld1q_s16(ptr + 0);
+    res.val[1] = vld1q_s16(ptr + 8);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x2_t {
+    uint8x16_t val[2];
+} ggml_uint8x16x2_t;
+
+inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
+    ggml_uint8x16x2_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+
+    return res;
+}
+
+typedef struct ggml_uint8x16x4_t {
+    uint8x16_t val[4];
+} ggml_uint8x16x4_t;
+
+inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
+    ggml_uint8x16x4_t res;
+
+    res.val[0] = vld1q_u8(ptr + 0);
+    res.val[1] = vld1q_u8(ptr + 16);
+    res.val[2] = vld1q_u8(ptr + 32);
+    res.val[3] = vld1q_u8(ptr + 48);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x2_t {
+    int8x16_t val[2];
+} ggml_int8x16x2_t;
+
+inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
+    ggml_int8x16x2_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+
+    return res;
+}
+
+typedef struct ggml_int8x16x4_t {
+    int8x16_t val[4];
+} ggml_int8x16x4_t;
+
+inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
+    ggml_int8x16x4_t res;
+
+    res.val[0] = vld1q_s8(ptr + 0);
+    res.val[1] = vld1q_s8(ptr + 16);
+    res.val[2] = vld1q_s8(ptr + 32);
+    res.val[3] = vld1q_s8(ptr + 48);
+
+    return res;
+}
+
+// NOTE: not tested
+inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
+    int8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+// NOTE: not tested
+inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
+    uint8x16_t res;
+
+    res[ 0] = a[b[ 0]];
+    res[ 1] = a[b[ 1]];
+    res[ 2] = a[b[ 2]];
+    res[ 3] = a[b[ 3]];
+    res[ 4] = a[b[ 4]];
+    res[ 5] = a[b[ 5]];
+    res[ 6] = a[b[ 6]];
+    res[ 7] = a[b[ 7]];
+    res[ 8] = a[b[ 8]];
+    res[ 9] = a[b[ 9]];
+    res[10] = a[b[10]];
+    res[11] = a[b[11]];
+    res[12] = a[b[12]];
+    res[13] = a[b[13]];
+    res[14] = a[b[14]];
+    res[15] = a[b[15]];
+
+    return res;
+}
+
+#else
+
+#define ggml_int16x8x2_t  int16x8x2_t
+#define ggml_uint8x16x2_t uint8x16x2_t
+#define ggml_uint8x16x4_t uint8x16x4_t
+#define ggml_int8x16x2_t  int8x16x2_t
+#define ggml_int8x16x4_t  int8x16x4_t
+
+#define ggml_vld1q_s16_x2 vld1q_s16_x2
+#define ggml_vld1q_u8_x2  vld1q_u8_x2
+#define ggml_vld1q_u8_x4  vld1q_u8_x4
+#define ggml_vld1q_s8_x2  vld1q_s8_x2
+#define ggml_vld1q_s8_x4  vld1q_s8_x4
+#define ggml_vqtbl1q_s8   vqtbl1q_s8
+#define ggml_vqtbl1q_u8   vqtbl1q_u8
+
+#endif // !defined(__aarch64__)
+
+#if !defined(__ARM_FEATURE_DOTPROD)
+
+inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
+    const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
+    const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
+
+    return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
+}
+
+#else
+
+#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
+
+#endif // !defined(__ARM_FEATURE_DOTPROD)
+
+#endif // defined(__ARM_NEON)
+
+#if defined(__ARM_NEON) && !defined(__MSC_VER)
+
 #define GGML_COMPUTE_FP16_TO_FP32(x) ggml_compute_fp16_to_fp32(x)
 #define GGML_COMPUTE_FP32_TO_FP16(x) ggml_compute_fp32_to_fp16(x)
 
@@ -75,8 +335,6 @@ static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
 
 #else
 
-typedef uint16_t ggml_fp16_internal_t;
-
 #ifdef __wasm_simd128__
 #include <wasm_simd128.h>
 #else
@@ -221,7 +479,7 @@ static inline ggml_fp16_t ggml_compute_fp32_to_fp16(float f) {
 
 #endif // __F16C__
 
-#endif // __ARM_NEON
+#endif // defined(__ARM_NEON) && (!defined(__MSC_VER)
 
 // precomputed f32 table for f16 (256 KB)
 // defined in ggml.c, initialized in ggml_init()