Merge branch 'ggerganov:master' into master

README.md

@@ -5,12 +5,15 @@
 [![Actions Status](https://github.com/ggerganov/llama.cpp/workflows/CI/badge.svg)](https://github.com/ggerganov/llama.cpp/actions)
 [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](https://opensource.org/licenses/MIT)
 
+[Roadmap](https://github.com/users/ggerganov/projects/7) / [Manifesto](https://github.com/ggerganov/llama.cpp/discussions/205) / [ggml](https://github.com/ggerganov/ggml)
+
 Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 **Hot topics:**
 
+- New roadmap: https://github.com/users/ggerganov/projects/7
+- Azure CI brainstorming: https://github.com/ggerganov/llama.cpp/discussions/1985
 - p1 : LLM-based code completion engine at the edge : https://github.com/ggml-org/p1/discussions/1
-- Roadmap June 2023: https://github.com/ggerganov/llama.cpp/discussions/1729
 
 <details>
   <summary>Table of Contents</summary>

build.zig

@@ -1,61 +1,58 @@
 const std = @import("std");
 
+// Zig Version: 0.11.0-dev.3379+629f0d23b
 pub fn build(b: *std.build.Builder) void {
     const target = b.standardTargetOptions(.{});
-    const optimize = b.standardReleaseOptions();
+    const optimize = b.standardOptimizeOption(.{});
-    const want_lto = b.option(bool, "lto", "Want -fLTO");
+    const lib = b.addStaticLibrary(.{
-
+        .name = "llama",
-    const lib = b.addStaticLibrary("llama", null);
+        .target = target,
-    lib.want_lto = want_lto;
+        .optimize = optimize,
-    lib.setTarget(target);
+    });
-    lib.setBuildMode(optimize);
+    lib.linkLibC();
     lib.linkLibCpp();
     lib.addIncludePath(".");
-    lib.addIncludePath("examples");
+    lib.addIncludePath("./examples");
     lib.addCSourceFiles(&.{
         "ggml.c",
     }, &.{"-std=c11"});
     lib.addCSourceFiles(&.{
         "llama.cpp",
     }, &.{"-std=c++11"});
-    lib.install();
+    b.installArtifact(lib);
 
-    const build_args = .{ .b = b, .lib = lib, .target = target, .optimize = optimize, .want_lto = want_lto };
+    const examples = .{
+        "main",
+        "baby-llama",
+        "embedding",
+        // "metal",
+        "perplexity",
+        "quantize",
+        "quantize-stats",
+        "save-load-state",
+        // "server",
+        "simple",
+        "train-text-from-scratch",
+    };
 
-    const exe = build_example("main", build_args);
+    inline for (examples) |example_name| {
-    _ = build_example("quantize", build_args);
+        const exe = b.addExecutable(.{
-    _ = build_example("perplexity", build_args);
+            .name = example_name,
-    _ = build_example("embedding", build_args);
+            .target = target,
-
+            .optimize = optimize,
-    // create "zig build run" command for ./main
+        });
-
+        exe.addIncludePath(".");
-    const run_cmd = exe.run();
+        exe.addIncludePath("./examples");
-    run_cmd.step.dependOn(b.getInstallStep());
+        exe.addCSourceFiles(&.{
-    if (b.args) |args| {
+            std.fmt.comptimePrint("examples/{s}/{s}.cpp", .{example_name, example_name}),
-        run_cmd.addArgs(args);
+            "examples/common.cpp",
+        }, &.{"-std=c++11"});
+        exe.linkLibrary(lib);
+        b.installArtifact(exe);
+        const run_cmd = b.addRunArtifact(exe);
+        run_cmd.step.dependOn(b.getInstallStep());
+        if (b.args) |args| run_cmd.addArgs(args);
+        const run_step = b.step("run_" ++ example_name, "Run the app");
+        run_step.dependOn(&run_cmd.step);
     }
-
-    const run_step = b.step("run", "Run the app");
-    run_step.dependOn(&run_cmd.step);
-}
-
-fn build_example(comptime name: []const u8, args: anytype) *std.build.LibExeObjStep {
-    const b = args.b;
-    const lib = args.lib;
-    const want_lto = args.want_lto;
-
-    const exe = b.addExecutable(name, null);
-    exe.want_lto = want_lto;
-    lib.setTarget(args.target);
-    lib.setBuildMode(args.optimize);
-    exe.addIncludePath(".");
-    exe.addIncludePath("examples");
-    exe.addCSourceFiles(&.{
-        std.fmt.comptimePrint("examples/{s}/{s}.cpp", .{name, name}),
-        "examples/common.cpp",
-    }, &.{"-std=c++11"});
-    exe.linkLibrary(lib);
-    exe.install();
-
-    return exe;
 }

examples/server/server.cpp

@@ -374,10 +374,10 @@ struct llama_server_context {
                     result.tok = llama_sample_token_mirostat_v2(ctx, &candidates_p, mirostat_tau, mirostat_eta, &mirostat_mu);
                 } else {
                     // Temperature sampling
+                    llama_sample_top_k(ctx, &candidates_p, top_k, 1);
                     llama_sample_tail_free(ctx, &candidates_p, tfs_z, 1);
                     llama_sample_typical(ctx, &candidates_p, typical_p, 1);
                     llama_sample_top_p(ctx, &candidates_p, top_p, 1);
-                    llama_sample_top_k(ctx, &candidates_p, top_k, 1);
                     llama_sample_temperature(ctx, &candidates_p, temp);
                     result.tok = llama_sample_token(ctx, &candidates_p);
                 }

ggml-cuda.cu

@@ -2635,7 +2635,7 @@ void ggml_cuda_free_scratch() {
 bool ggml_cuda_compute_forward(struct ggml_compute_params * params, struct ggml_tensor * tensor){
     ggml_cuda_func_t func;
     const bool any_on_device = tensor->backend == GGML_BACKEND_GPU
-        || tensor->src0->backend == GGML_BACKEND_GPU || tensor->src0->backend == GGML_BACKEND_GPU_SPLIT
+        || (tensor->src0 != nullptr && (tensor->src0->backend == GGML_BACKEND_GPU || tensor->src0->backend == GGML_BACKEND_GPU_SPLIT))
         || (tensor->src1 != nullptr && tensor->src1->backend == GGML_BACKEND_GPU);
 
     switch (tensor->op) {

ggml.c

@@ -1,5 +1,5 @@
-// Defines CLOCK_MONOTONIC on Linux
+#define _GNU_SOURCE // Defines CLOCK_MONOTONIC on Linux
-#define _GNU_SOURCE
+#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnigns on Windows
 
 #include "ggml.h"
 
@@ -131,6 +131,34 @@ typedef void* thread_ret_t;
     #define GGML_MEM_ALIGN 16
 #endif
 
+//
+// logging
+//
+
+#if (GGML_DEBUG >= 1)
+#define GGML_PRINT_DEBUG(...) printf(__VA_ARGS__)
+#else
+#define GGML_PRINT_DEBUG(...)
+#endif
+
+#if (GGML_DEBUG >= 5)
+#define GGML_PRINT_DEBUG_5(...) printf(__VA_ARGS__)
+#else
+#define GGML_PRINT_DEBUG_5(...)
+#endif
+
+#if (GGML_DEBUG >= 10)
+#define GGML_PRINT_DEBUG_10(...) printf(__VA_ARGS__)
+#else
+#define GGML_PRINT_DEBUG_10(...)
+#endif
+
+#define GGML_PRINT(...) printf(__VA_ARGS__)
+
+//
+// end of logging block
+//
+
 #if defined(_MSC_VER) || defined(__MINGW32__)
 #define GGML_ALIGNED_MALLOC(size)  _aligned_malloc(size, GGML_MEM_ALIGN)
 #define GGML_ALIGNED_FREE(ptr)     _aligned_free(ptr)
@@ -144,6 +172,17 @@ inline static void* ggml_aligned_malloc(size_t size) {
 #endif
     if (result != 0) {
         // Handle allocation failure
+        const char *error_desc = "unknown allocation error";
+        switch (result) {
+            case EINVAL:
+                error_desc = "invalid alignment value";
+                break;
+            case ENOMEM:
+                error_desc = "insufficient memory";
+                break;
+        }
+        GGML_PRINT("%s: %s (attempted to allocate %6.2f MB)\n",
+            __func__, error_desc, size/(1024.0*1024.0));
         return NULL;
     }
     return aligned_memory;
@@ -3530,30 +3569,6 @@ inline static void ggml_vec_norm_inv_f32(const int n, float * s, const float * x
     *s = 1.f/(*s);
 }
 
-//
-// logging
-//
-
-#if (GGML_DEBUG >= 1)
-#define GGML_PRINT_DEBUG(...) printf(__VA_ARGS__)
-#else
-#define GGML_PRINT_DEBUG(...)
-#endif
-
-#if (GGML_DEBUG >= 5)
-#define GGML_PRINT_DEBUG_5(...) printf(__VA_ARGS__)
-#else
-#define GGML_PRINT_DEBUG_5(...)
-#endif
-
-#if (GGML_DEBUG >= 10)
-#define GGML_PRINT_DEBUG_10(...) printf(__VA_ARGS__)
-#else
-#define GGML_PRINT_DEBUG_10(...)
-#endif
-
-#define GGML_PRINT(...) printf(__VA_ARGS__)
-
 //
 // data types
 //
@@ -3713,11 +3728,15 @@ static const char * GGML_OP_NAME[GGML_OP_COUNT] = {
     "MAP_UNARY",
     "MAP_BINARY",
 
+    "MAP_CUSTOM1",
+    "MAP_CUSTOM2",
+    "MAP_CUSTOM3",
+
     "CROSS_ENTROPY_LOSS",
     "CROSS_ENTROPY_LOSS_BACK",
 };
 
-static_assert(GGML_OP_COUNT == 61, "GGML_OP_COUNT != 61");
+static_assert(GGML_OP_COUNT == 64, "GGML_OP_COUNT != 64");
 
 static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "none",
@@ -3785,11 +3804,15 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
     "f(x)",
     "f(x,y)",
 
+    "custom(x)",
+    "custom(x,y)",
+    "custom(x,y,z)",
+
     "cross_entropy_loss(x,y)",
     "cross_entropy_loss_back(x,y)",
 };
 
-static_assert(GGML_OP_COUNT == 61, "GGML_OP_COUNT != 61");
+static_assert(GGML_OP_COUNT == 64, "GGML_OP_COUNT != 64");
 
 static_assert(sizeof(struct ggml_object)%GGML_MEM_ALIGN == 0, "ggml_object size must be a multiple of GGML_MEM_ALIGN");
 static_assert(sizeof(struct ggml_tensor)%GGML_MEM_ALIGN == 0, "ggml_tensor size must be a multiple of GGML_MEM_ALIGN");
@@ -7094,9 +7117,14 @@ struct ggml_tensor * ggml_map_unary_impl_f32(
         is_node = true;
     }
 
+    struct ggml_tensor *result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    ggml_scratch_save(ctx);
+
     struct ggml_tensor * addr_tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sizeof(void *) / sizeof(int32_t));
     *((void (**)(void))addr_tensor->data) = (void (*)(void))fun;
-    struct ggml_tensor *result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    ggml_scratch_load(ctx);
 
     result->op = GGML_OP_MAP_UNARY;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -7136,9 +7164,14 @@ struct ggml_tensor * ggml_map_binary_impl_f32(
         is_node = true;
     }
 
+    struct ggml_tensor *result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    ggml_scratch_save(ctx);
+
     struct ggml_tensor * addr_tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sizeof(void *) / sizeof(int32_t));
     *((void (**)(void))addr_tensor->data) = (void (*)(void))fun;
-    struct ggml_tensor *result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    ggml_scratch_load(ctx);
 
     result->op = GGML_OP_MAP_BINARY;
     result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
@@ -7165,6 +7198,150 @@ struct ggml_tensor * ggml_map_binary_inplace_f32(
     return ggml_map_binary_impl_f32(ctx, a, b, fun, true);
 }
 
+// ggml_map_custom1
+
+struct ggml_tensor * ggml_map_custom1_impl_f32(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        const  ggml_custom1_op_f32_t   fun,
+        bool   inplace) {
+    bool is_node = false;
+
+    if (!inplace && a->grad) {
+        is_node = true;
+    }
+
+    struct ggml_tensor *result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    ggml_scratch_save(ctx);
+
+    struct ggml_tensor * addr_tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sizeof(void *) / sizeof(int32_t));
+    *((void (**)(void))addr_tensor->data) = (void (*)(void))fun;
+
+    ggml_scratch_load(ctx);
+
+    result->op = GGML_OP_MAP_CUSTOM1;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src0 = a;
+    result->opt[0] = addr_tensor;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_map_custom1_f32(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        const  ggml_custom1_op_f32_t   fun) {
+    return ggml_map_custom1_impl_f32(ctx, a, fun, false);
+}
+
+struct ggml_tensor * ggml_map_custom1_inplace_f32(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        const  ggml_custom1_op_f32_t   fun) {
+    return ggml_map_custom1_impl_f32(ctx, a, fun, true);
+}
+
+// ggml_map_custom2
+
+struct ggml_tensor * ggml_map_custom2_impl_f32(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        const  ggml_custom2_op_f32_t   fun,
+        bool   inplace) {
+    bool is_node = false;
+
+    if (!inplace && (a->grad || b->grad)) {
+        is_node = true;
+    }
+
+    struct ggml_tensor *result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    ggml_scratch_save(ctx);
+
+    struct ggml_tensor * addr_tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sizeof(void *) / sizeof(int32_t));
+    *((void (**)(void))addr_tensor->data) = (void (*)(void))fun;
+
+    ggml_scratch_load(ctx);
+
+    result->op = GGML_OP_MAP_CUSTOM2;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src0 = a;
+    result->src1 = b;
+    result->opt[0] = addr_tensor;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_map_custom2_f32(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        const  ggml_custom2_op_f32_t   fun) {
+    return ggml_map_custom2_impl_f32(ctx, a, b, fun, false);
+}
+
+struct ggml_tensor * ggml_map_custom2_inplace_f32(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        const  ggml_custom2_op_f32_t   fun) {
+    return ggml_map_custom2_impl_f32(ctx, a, b, fun, true);
+}
+
+// ggml_map_custom3
+
+struct ggml_tensor * ggml_map_custom3_impl_f32(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        struct ggml_tensor           * c,
+        const  ggml_custom3_op_f32_t   fun,
+        bool   inplace) {
+    bool is_node = false;
+
+    if (!inplace && (a->grad || b->grad || c->grad)) {
+        is_node = true;
+    }
+
+    struct ggml_tensor *result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+
+    ggml_scratch_save(ctx);
+
+    struct ggml_tensor * addr_tensor = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, sizeof(void *) / sizeof(int32_t));
+    *((void (**)(void))addr_tensor->data) = (void (*)(void))fun;
+
+    ggml_scratch_load(ctx);
+
+    result->op = GGML_OP_MAP_CUSTOM3;
+    result->grad = is_node ? ggml_dup_tensor(ctx, result) : NULL;
+    result->src0 = a;
+    result->src1 = b;
+    result->opt[0] = addr_tensor;
+    result->opt[1] = c;
+
+    return result;
+}
+
+struct ggml_tensor * ggml_map_custom3_f32(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        struct ggml_tensor           * c,
+        const  ggml_custom3_op_f32_t   fun) {
+    return ggml_map_custom3_impl_f32(ctx, a, b, c, fun, false);
+}
+
+struct ggml_tensor * ggml_map_custom3_inplace_f32(
+        struct ggml_context          * ctx,
+        struct ggml_tensor           * a,
+        struct ggml_tensor           * b,
+        struct ggml_tensor           * c,
+        const  ggml_custom3_op_f32_t   fun) {
+    return ggml_map_custom3_impl_f32(ctx, a, b, c, fun, true);
+}
+
 // ggml_cross_entropy_loss
 
 struct ggml_tensor * ggml_cross_entropy_loss(
@@ -14621,6 +14798,114 @@ static void ggml_compute_forward_map_binary(
     }
 }
 
+// ggml_compute_forward_map_custom1
+
+static void ggml_compute_forward_map_custom1_f32(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * a,
+        struct ggml_tensor * dst,
+        const ggml_custom1_op_f32_t fun) {
+    assert(params->ith == 0);
+
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    fun(dst, a);
+}
+
+
+static void ggml_compute_forward_map_custom1(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * a,
+        struct ggml_tensor * dst,
+        const ggml_custom1_op_f32_t fun) {
+    switch (a->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_map_custom1_f32(params, a, dst, fun);
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
+// ggml_compute_forward_map_custom2
+
+static void ggml_compute_forward_map_custom2_f32(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * a,
+        const struct ggml_tensor * b,
+        struct ggml_tensor * dst,
+        const ggml_custom2_op_f32_t fun) {
+    assert(params->ith == 0);
+
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    fun(dst, a, b);
+}
+
+
+static void ggml_compute_forward_map_custom2(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * a,
+        const struct ggml_tensor * b,
+        struct ggml_tensor * dst,
+        const ggml_custom2_op_f32_t fun) {
+    switch (a->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_map_custom2_f32(params, a, b, dst, fun);
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
+// ggml_compute_forward_map_custom3
+
+static void ggml_compute_forward_map_custom3_f32(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * a,
+        const struct ggml_tensor * b,
+        const struct ggml_tensor * c,
+        struct ggml_tensor * dst,
+        const ggml_custom3_op_f32_t fun) {
+    assert(params->ith == 0);
+
+    if (params->type == GGML_TASK_INIT || params->type == GGML_TASK_FINALIZE) {
+        return;
+    }
+
+    fun(dst, a, b, c);
+}
+
+
+static void ggml_compute_forward_map_custom3(
+        const struct ggml_compute_params * params,
+        const struct ggml_tensor * a,
+        const struct ggml_tensor * b,
+        const struct ggml_tensor * c,
+        struct ggml_tensor * dst,
+        const ggml_custom3_op_f32_t fun) {
+    switch (a->type) {
+        case GGML_TYPE_F32:
+            {
+                ggml_compute_forward_map_custom3_f32(params, a, b, c, dst, fun);
+            } break;
+        default:
+            {
+                GGML_ASSERT(false);
+            } break;
+    }
+}
+
 // ggml_compute_forward_cross_entropy_loss
 
 static void ggml_compute_forward_cross_entropy_loss_f32(
@@ -14911,7 +15196,7 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
     if (skip_cpu) {
         return;
     }
-    GGML_ASSERT(tensor->src0->backend == GGML_BACKEND_CPU);
+    GGML_ASSERT(tensor->src0 == NULL || tensor->src0->backend == GGML_BACKEND_CPU);
     GGML_ASSERT(tensor->src1 == NULL || tensor->src1->backend == GGML_BACKEND_CPU);
 #endif // GGML_USE_CUBLAS
 
@@ -15158,6 +15443,24 @@ static void ggml_compute_forward(struct ggml_compute_params * params, struct ggm
                 ggml_compute_forward_map_binary(params, tensor->src0, tensor->src1, tensor, fun);
             }
             break;
+        case GGML_OP_MAP_CUSTOM1:
+            {
+                const ggml_custom1_op_f32_t fun = *((ggml_custom1_op_f32_t *)tensor->opt[0]->data);
+                ggml_compute_forward_map_custom1(params, tensor->src0, tensor, fun);
+            }
+            break;
+        case GGML_OP_MAP_CUSTOM2:
+            {
+                const ggml_custom2_op_f32_t fun = *((ggml_custom2_op_f32_t *)tensor->opt[0]->data);
+                ggml_compute_forward_map_custom2(params, tensor->src0, tensor->src1, tensor, fun);
+            }
+            break;
+        case GGML_OP_MAP_CUSTOM3:
+            {
+                const ggml_custom3_op_f32_t fun = *((ggml_custom3_op_f32_t *)tensor->opt[0]->data);
+                ggml_compute_forward_map_custom3(params, tensor->src0, tensor->src1, tensor->opt[1], tensor, fun);
+            }
+            break;
         case GGML_OP_CROSS_ENTROPY_LOSS:
             {
                 ggml_compute_forward_cross_entropy_loss(params, tensor->src0, tensor->src1, tensor);
@@ -15964,6 +16267,9 @@ static void ggml_compute_backward(struct ggml_context * ctx, struct ggml_tensor
         case GGML_OP_WIN_UNPART:
         case GGML_OP_MAP_UNARY:
         case GGML_OP_MAP_BINARY:
+        case GGML_OP_MAP_CUSTOM1:
+        case GGML_OP_MAP_CUSTOM2:
+        case GGML_OP_MAP_CUSTOM3:
             {
                 GGML_ASSERT(false); // not supported
             } break;
@@ -16605,6 +16911,9 @@ void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph)
                 case GGML_OP_WIN_UNPART:
                 case GGML_OP_MAP_UNARY:
                 case GGML_OP_MAP_BINARY:
+                case GGML_OP_MAP_CUSTOM1:
+                case GGML_OP_MAP_CUSTOM2:
+                case GGML_OP_MAP_CUSTOM3:
                     {
                         node->n_tasks = 1;
                     } break;

ggml.h

@@ -345,6 +345,10 @@ extern "C" {
         GGML_OP_MAP_UNARY,
         GGML_OP_MAP_BINARY,
 
+        GGML_OP_MAP_CUSTOM1,
+        GGML_OP_MAP_CUSTOM2,
+        GGML_OP_MAP_CUSTOM3,
+
         GGML_OP_CROSS_ENTROPY_LOSS,
         GGML_OP_CROSS_ENTROPY_LOSS_BACK,
 
@@ -1167,21 +1171,73 @@ extern "C" {
             int                   h0,
             int                   w);
 
-    // Mapping operations
+    // custom operators
-    typedef void (*ggml_unary_op_f32_t)(const int, float *, const float *);
+
+    typedef void (*ggml_unary_op_f32_t) (const int, float *, const float *);
     typedef void (*ggml_binary_op_f32_t)(const int, float *, const float *, const float *);
 
+    typedef void (*ggml_custom1_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *);
+    typedef void (*ggml_custom2_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
+    typedef void (*ggml_custom3_op_f32_t)(struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *, const struct ggml_tensor *);
+
     GGML_API struct ggml_tensor * ggml_map_unary_f32(
             struct ggml_context        * ctx,
             struct ggml_tensor         * a,
                    ggml_unary_op_f32_t   fun);
 
+    GGML_API struct ggml_tensor * ggml_map_unary_inplace_f32(
+            struct ggml_context        * ctx,
+            struct ggml_tensor         * a,
+                   ggml_unary_op_f32_t   fun);
+
     GGML_API struct ggml_tensor * ggml_map_binary_f32(
             struct ggml_context         * ctx,
             struct ggml_tensor          * a,
             struct ggml_tensor          * b,
                    ggml_binary_op_f32_t   fun);
 
+    GGML_API struct ggml_tensor * ggml_map_binary_inplace_f32(
+            struct ggml_context         * ctx,
+            struct ggml_tensor          * a,
+            struct ggml_tensor          * b,
+                   ggml_binary_op_f32_t   fun);
+
+    GGML_API struct ggml_tensor * ggml_map_custom1_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+                   ggml_custom1_op_f32_t   fun);
+
+    GGML_API struct ggml_tensor * ggml_map_custom1_inplace_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+                   ggml_custom1_op_f32_t   fun);
+
+    GGML_API struct ggml_tensor * ggml_map_custom2_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+            struct ggml_tensor           * b,
+                   ggml_custom2_op_f32_t   fun);
+
+    GGML_API struct ggml_tensor * ggml_map_custom2_inplace_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+            struct ggml_tensor           * b,
+                   ggml_custom2_op_f32_t   fun);
+
+    GGML_API struct ggml_tensor * ggml_map_custom3_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+            struct ggml_tensor           * b,
+            struct ggml_tensor           * c,
+                   ggml_custom3_op_f32_t   fun);
+
+    GGML_API struct ggml_tensor * ggml_map_custom3_inplace_f32(
+            struct ggml_context          * ctx,
+            struct ggml_tensor           * a,
+            struct ggml_tensor           * b,
+            struct ggml_tensor           * c,
+                   ggml_custom3_op_f32_t   fun);
+
     // loss function
 
     GGML_API struct ggml_tensor * ggml_cross_entropy_loss(

tests/test-grad0.c

@@ -1,3 +1,4 @@
+#define _CRT_SECURE_NO_DEPRECATE // Disables ridiculous "unsafe" warnigns on Windows
 #include "ggml.h"
 
 #include <math.h>
@@ -5,6 +6,10 @@
 #include <stdlib.h>
 #include <assert.h>
 
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
 #define MAX_NARGS 3
 
 #undef MIN
@@ -197,8 +202,23 @@ bool check_gradient(
         float max_error_abs,
         float max_error_rel) {
 
+    static int n_threads = -1;
+    if (n_threads < 0) {
+        n_threads = GGML_DEFAULT_N_THREADS;
+
+        const char *env = getenv("GGML_N_THREADS");
+        if (env) {
+            n_threads = atoi(env);
+        }
+
+        printf("GGML_N_THREADS = %d\n", n_threads);
+    }
+
     struct ggml_cgraph gf = ggml_build_forward (f);
+    gf.n_threads = n_threads;
+
     struct ggml_cgraph gb = ggml_build_backward(ctx0, &gf, false);
+    gb.n_threads = n_threads;
 
     ggml_graph_compute(ctx0, &gf);
     ggml_graph_reset  (&gf);