gpt_neox_client 0.1.0

data/ext/gpt_neox_client/src/ggml/ggml-metal.m

@@ -0,0 +1,1195 @@
+#import "ggml-metal.h"
+
+#import "ggml.h"
+
+#import <Foundation/Foundation.h>
+
+#import <Metal/Metal.h>
+
+#undef MIN
+#undef MAX
+#define MIN(a, b) ((a) < (b) ? (a) : (b))
+#define MAX(a, b) ((a) > (b) ? (a) : (b))
+
+#ifdef GGML_METAL_NDEBUG
+#define metal_printf(...)
+#else
+#define metal_printf(...) fprintf(stderr, __VA_ARGS__)
+#endif
+
+#define UNUSED(x) (void)(x)
+
+#define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
+
+struct ggml_metal_buffer {
+    const char * name;
+
+    void   * data;
+    size_t   size;
+
+    id<MTLBuffer> metal;
+};
+
+struct ggml_metal_context {
+    int n_cb;
+
+    id<MTLDevice>       device;
+    id<MTLCommandQueue> queue;
+    id<MTLLibrary>      library;
+
+    id<MTLCommandBuffer>         command_buffers [GGML_METAL_MAX_COMMAND_BUFFERS];
+    id<MTLComputeCommandEncoder> command_encoders[GGML_METAL_MAX_COMMAND_BUFFERS];
+
+    dispatch_queue_t d_queue;
+
+    int n_buffers;
+    struct ggml_metal_buffer buffers[GGML_METAL_MAX_BUFFERS];
+
+    int concur_list[GGML_MAX_CONCUR];
+    int concur_list_len;
+
+    // custom kernels
+#define GGML_METAL_DECL_KERNEL(name) \
+    id<MTLFunction>             function_##name; \
+    id<MTLComputePipelineState> pipeline_##name
+
+    GGML_METAL_DECL_KERNEL(add);
+    GGML_METAL_DECL_KERNEL(add_row); // TODO: avoid this extra kernel, instead extend the "add" kernel to support broadcast
+    GGML_METAL_DECL_KERNEL(mul);
+    GGML_METAL_DECL_KERNEL(mul_row); // TODO: avoid this extra kernel, instead extend the "mul" kernel to support broadcast
+    GGML_METAL_DECL_KERNEL(scale);
+    GGML_METAL_DECL_KERNEL(silu);
+    GGML_METAL_DECL_KERNEL(relu);
+    GGML_METAL_DECL_KERNEL(gelu);
+    GGML_METAL_DECL_KERNEL(soft_max);
+    GGML_METAL_DECL_KERNEL(diag_mask_inf);
+    GGML_METAL_DECL_KERNEL(get_rows_f16);
+    GGML_METAL_DECL_KERNEL(get_rows_q4_0);
+    GGML_METAL_DECL_KERNEL(get_rows_q4_1);
+    GGML_METAL_DECL_KERNEL(get_rows_q8_0);
+    GGML_METAL_DECL_KERNEL(get_rows_q2_K);
+    GGML_METAL_DECL_KERNEL(get_rows_q3_K);
+    GGML_METAL_DECL_KERNEL(get_rows_q4_K);
+    GGML_METAL_DECL_KERNEL(get_rows_q5_K);
+    GGML_METAL_DECL_KERNEL(get_rows_q6_K);
+    GGML_METAL_DECL_KERNEL(rms_norm);
+    GGML_METAL_DECL_KERNEL(norm);
+    GGML_METAL_DECL_KERNEL(mul_mat_f16_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q4_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q4_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q8_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q2_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q3_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q4_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q5_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mat_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q4_1_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q8_0_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q2_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q3_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q4_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q5_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(rope);
+    GGML_METAL_DECL_KERNEL(alibi_f32);
+    GGML_METAL_DECL_KERNEL(cpy_f32_f16);
+    GGML_METAL_DECL_KERNEL(cpy_f32_f32);
+    GGML_METAL_DECL_KERNEL(cpy_f16_f16);
+
+#undef GGML_METAL_DECL_KERNEL
+};
+
+// MSL code
+// TODO: move the contents here when ready
+//       for now it is easier to work in a separate file
+static NSString * const msl_library_source = @"see metal.metal";
+
+// Here to assist with NSBundle Path Hack
+@interface GGMLMetalClass : NSObject
+@end
+@implementation GGMLMetalClass
+@end
+
+struct ggml_metal_context * ggml_metal_init(int n_cb) {
+    fprintf(stderr, "%s: allocating\n", __func__);
+
+    struct ggml_metal_context * ctx = malloc(sizeof(struct ggml_metal_context));
+
+    ctx->n_cb   = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
+    ctx->device = MTLCreateSystemDefaultDevice();
+    ctx->queue  = [ctx->device newCommandQueue];
+    ctx->n_buffers = 0;
+    ctx->concur_list_len = 0;
+
+    ctx->d_queue = dispatch_queue_create("llama.cpp", DISPATCH_QUEUE_CONCURRENT);
+
+#if 0
+    // compile from source string and show compile log
+    {
+        NSError * error = nil;
+
+        ctx->library = [ctx->device newLibraryWithSource:msl_library_source options:nil error:&error];
+        if (error) {
+            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            return NULL;
+        }
+    }
+#else
+    UNUSED(msl_library_source);
+
+    // read the source from "ggml-metal.metal" into a string and use newLibraryWithSource
+    {
+        NSError * error = nil;
+
+        //NSString * path = [[NSBundle mainBundle] pathForResource:@"../../examples/metal/metal" ofType:@"metal"];
+        NSBundle * bundle = [NSBundle bundleForClass:[GGMLMetalClass class]];
+        NSString * path = [bundle pathForResource:@"ggml-metal" ofType:@"metal"];
+        fprintf(stderr, "%s: loading '%s'\n", __func__, [path UTF8String]);
+
+        NSString * src  = [NSString stringWithContentsOfFile:path encoding:NSUTF8StringEncoding error:&error];
+        if (error) {
+            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            return NULL;
+        }
+
+#ifdef GGML_QKK_64
+        MTLCompileOptions* options = [MTLCompileOptions new];
+        options.preprocessorMacros = @{ @"QK_K" : @(64) };
+        ctx->library = [ctx->device newLibraryWithSource:src options:options error:&error];
+#else
+        ctx->library = [ctx->device newLibraryWithSource:src options:nil error:&error];
+#endif
+        if (error) {
+            fprintf(stderr, "%s: error: %s\n", __func__, [[error description] UTF8String]);
+            return NULL;
+        }
+    }
+#endif
+
+    // load kernels
+    {
+        NSError * error = nil;
+#define GGML_METAL_ADD_KERNEL(name) \
+        ctx->function_##name = [ctx->library newFunctionWithName:@"kernel_"#name]; \
+        ctx->pipeline_##name = [ctx->device newComputePipelineStateWithFunction:ctx->function_##name error:&error]; \
+        fprintf(stderr, "%s: loaded %-32s %16p | th_max = %4d | th_width = %4d\n", __func__, "kernel_"#name, (void *) ctx->pipeline_##name, \
+                (int) ctx->pipeline_##name.maxTotalThreadsPerThreadgroup, \
+                (int) ctx->pipeline_##name.threadExecutionWidth); \
+        if (error) { \
+            fprintf(stderr, "%s: load pipeline error: %s\n", __func__, [[error description] UTF8String]); \
+            return NULL; \
+        }
+
+        GGML_METAL_ADD_KERNEL(add);
+        GGML_METAL_ADD_KERNEL(add_row);
+        GGML_METAL_ADD_KERNEL(mul);
+        GGML_METAL_ADD_KERNEL(mul_row);
+        GGML_METAL_ADD_KERNEL(scale);
+        GGML_METAL_ADD_KERNEL(silu);
+        GGML_METAL_ADD_KERNEL(relu);
+        GGML_METAL_ADD_KERNEL(gelu);
+        GGML_METAL_ADD_KERNEL(soft_max);
+        GGML_METAL_ADD_KERNEL(diag_mask_inf);
+        GGML_METAL_ADD_KERNEL(get_rows_f16);
+        GGML_METAL_ADD_KERNEL(get_rows_q4_0);
+        GGML_METAL_ADD_KERNEL(get_rows_q4_1);
+        GGML_METAL_ADD_KERNEL(get_rows_q8_0);
+        GGML_METAL_ADD_KERNEL(get_rows_q2_K);
+        GGML_METAL_ADD_KERNEL(get_rows_q3_K);
+        GGML_METAL_ADD_KERNEL(get_rows_q4_K);
+        GGML_METAL_ADD_KERNEL(get_rows_q5_K);
+        GGML_METAL_ADD_KERNEL(get_rows_q6_K);
+        GGML_METAL_ADD_KERNEL(rms_norm);
+        GGML_METAL_ADD_KERNEL(norm);
+        GGML_METAL_ADD_KERNEL(mul_mat_f16_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q4_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q4_1_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q8_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q2_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q3_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q4_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q5_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mat_q6_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q4_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q8_0_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q4_1_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q2_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q3_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q4_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q5_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mm_q6_K_f32);
+        GGML_METAL_ADD_KERNEL(rope);
+        GGML_METAL_ADD_KERNEL(alibi_f32);
+        GGML_METAL_ADD_KERNEL(cpy_f32_f16);
+        GGML_METAL_ADD_KERNEL(cpy_f32_f32);
+        GGML_METAL_ADD_KERNEL(cpy_f16_f16);
+
+#undef GGML_METAL_ADD_KERNEL
+    }
+
+    fprintf(stderr, "%s: recommendedMaxWorkingSetSize  = %8.2f MB\n", __func__, ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+    fprintf(stderr, "%s: hasUnifiedMemory              = %s\n",       __func__, ctx->device.hasUnifiedMemory ? "true" : "false");
+    if (ctx->device.maxTransferRate != 0) {
+        fprintf(stderr, "%s: maxTransferRate               = %8.2f MB/s\n", __func__, ctx->device.maxTransferRate / 1024.0 / 1024.0);
+    } else {
+        fprintf(stderr, "%s: maxTransferRate               = built-in GPU\n", __func__);
+    }
+
+    return ctx;
+}
+
+void ggml_metal_free(struct ggml_metal_context * ctx) {
+    fprintf(stderr, "%s: deallocating\n", __func__);
+#define GGML_METAL_DEL_KERNEL(name) \
+    [ctx->function_##name release]; \
+    [ctx->pipeline_##name release];
+
+    GGML_METAL_DEL_KERNEL(add);
+    GGML_METAL_DEL_KERNEL(add_row);
+    GGML_METAL_DEL_KERNEL(mul);
+    GGML_METAL_DEL_KERNEL(mul_row);
+    GGML_METAL_DEL_KERNEL(scale);
+    GGML_METAL_DEL_KERNEL(silu);
+    GGML_METAL_DEL_KERNEL(relu);
+    GGML_METAL_DEL_KERNEL(gelu);
+    GGML_METAL_DEL_KERNEL(soft_max);
+    GGML_METAL_DEL_KERNEL(diag_mask_inf);
+    GGML_METAL_DEL_KERNEL(get_rows_f16);
+    GGML_METAL_DEL_KERNEL(get_rows_q4_0);
+    GGML_METAL_DEL_KERNEL(get_rows_q4_1);
+    GGML_METAL_DEL_KERNEL(get_rows_q8_0);
+    GGML_METAL_DEL_KERNEL(get_rows_q2_K);
+    GGML_METAL_DEL_KERNEL(get_rows_q3_K);
+    GGML_METAL_DEL_KERNEL(get_rows_q4_K);
+    GGML_METAL_DEL_KERNEL(get_rows_q5_K);
+    GGML_METAL_DEL_KERNEL(get_rows_q6_K);
+    GGML_METAL_DEL_KERNEL(rms_norm);
+    GGML_METAL_DEL_KERNEL(norm);
+    GGML_METAL_DEL_KERNEL(mul_mat_f16_f32);
+    GGML_METAL_DEL_KERNEL(mul_mat_q4_0_f32);
+    GGML_METAL_DEL_KERNEL(mul_mat_q4_1_f32);
+    GGML_METAL_DEL_KERNEL(mul_mat_q8_0_f32);
+    GGML_METAL_DEL_KERNEL(mul_mat_q2_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mat_q3_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mat_q4_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mat_q5_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mat_q6_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mm_f16_f32);
+    GGML_METAL_DEL_KERNEL(mul_mm_q4_0_f32);
+    GGML_METAL_DEL_KERNEL(mul_mm_q8_0_f32);
+    GGML_METAL_DEL_KERNEL(mul_mm_q4_1_f32);
+    GGML_METAL_DEL_KERNEL(mul_mm_q2_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mm_q3_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mm_q4_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mm_q5_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mm_q6_K_f32);
+    GGML_METAL_DEL_KERNEL(rope);
+    GGML_METAL_DEL_KERNEL(alibi_f32);
+    GGML_METAL_DEL_KERNEL(cpy_f32_f16);
+    GGML_METAL_DEL_KERNEL(cpy_f32_f32);
+    GGML_METAL_DEL_KERNEL(cpy_f16_f16);
+
+#undef GGML_METAL_DEL_KERNEL
+
+    for (int i = 0; i < ctx->n_buffers; ++i) {
+        [ctx->buffers[i].metal release];
+    }
+
+    [ctx->library release];
+    [ctx->queue release];
+    [ctx->device release];
+
+    dispatch_release(ctx->d_queue);
+
+    free(ctx);
+}
+
+void * ggml_metal_host_malloc(size_t n) {
+    void * data = NULL;
+    const int result = posix_memalign((void **) &data, getpagesize(), n);
+    if (result != 0) {
+        fprintf(stderr, "%s: error: posix_memalign failed\n", __func__);
+        return NULL;
+    }
+
+    return data;
+}
+
+void ggml_metal_host_free(void * data) {
+    free(data);
+}
+
+void ggml_metal_set_n_cb(struct ggml_metal_context * ctx, int n_cb) {
+    ctx->n_cb = MIN(n_cb, GGML_METAL_MAX_BUFFERS);
+}
+
+int ggml_metal_if_optimized(struct ggml_metal_context * ctx) {
+    return ctx->concur_list_len;
+}
+
+int * ggml_metal_get_concur_list(struct ggml_metal_context * ctx) {
+    return ctx->concur_list;
+}
+
+// finds the Metal buffer that contains the tensor data on the GPU device
+// the assumption is that there is 1-to-1 mapping between the host and device memory buffers, so we can find the
+// Metal buffer based on the host memory pointer
+//
+static id<MTLBuffer> ggml_metal_get_buffer(struct ggml_metal_context * ctx, struct ggml_tensor * t, size_t * offs) {
+    //fprintf(stderr, "%s: data tensor '%16s', offs_data = %8ld, offs_eval = %8ld, offs_cach = %8ld\n", __func__, t->name, offs_data, offs_eval, offs_cach);
+
+    const int64_t tsize = ggml_nbytes(t);
+
+    // find the view that contains the tensor fully
+    for (int i = 0; i < ctx->n_buffers; ++i) {
+        const int64_t ioffs = (int64_t) t->data - (int64_t) ctx->buffers[i].data;
+
+        if (ioffs >= 0 && ioffs + tsize <= (int64_t) ctx->buffers[i].size) {
+            *offs = (size_t) ioffs;
+
+            //fprintf(stderr, "%s: '%s' tensor '%16s', offs = %8ld\n", __func__, ctx->buffers[i].name, t->name, *offs);
+
+            return ctx->buffers[i].metal;
+        }
+    }
+
+    fprintf(stderr, "%s: error: buffer is nil\n", __func__);
+
+    return nil;
+}
+
+bool ggml_metal_add_buffer(
+        struct ggml_metal_context * ctx,
+                     const char * name,
+                           void * data,
+                         size_t   size,
+                         size_t   max_size) {
+    if (ctx->n_buffers >= GGML_METAL_MAX_BUFFERS) {
+        fprintf(stderr, "%s: too many buffers\n", __func__);
+        return false;
+    }
+
+    if (data) {
+        // verify that the buffer does not overlap with any of the existing buffers
+        for (int i = 0; i < ctx->n_buffers; ++i) {
+            const int64_t ioffs = (int64_t) data - (int64_t) ctx->buffers[i].data;
+
+            if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
+                fprintf(stderr, "%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
+                return false;
+            }
+        }
+
+        const size_t size_page = getpagesize();
+
+        size_t size_aligned = size;
+        if ((size_aligned % size_page) != 0) {
+            size_aligned += (size_page - (size_aligned % size_page));
+        }
+
+        // the buffer fits into the max buffer size allowed by the device
+        if (size_aligned <= ctx->device.maxBufferLength) {
+            ctx->buffers[ctx->n_buffers].name = name;
+            ctx->buffers[ctx->n_buffers].data = data;
+            ctx->buffers[ctx->n_buffers].size = size;
+
+            ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:data length:size_aligned options:MTLResourceStorageModeShared deallocator:nil];
+
+            if (ctx->buffers[ctx->n_buffers].metal == nil) {
+                fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_aligned / 1024.0 / 1024.0);
+                return false;
+            }
+
+            fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB", __func__, name, size_aligned / 1024.0 / 1024.0);
+
+            ++ctx->n_buffers;
+        } else {
+            // this overlap between the views will guarantee that the tensor with the maximum size will fully fit into
+            // one of the views
+            const size_t size_ovlp = ((max_size + size_page - 1) / size_page + 1) * size_page; // round-up 2 pages just in case
+            const size_t size_step = ctx->device.maxBufferLength - size_ovlp;
+            const size_t size_view = ctx->device.maxBufferLength;
+
+            for (size_t i = 0; i < size; i += size_step) {
+                const size_t size_step_aligned = (i + size_view <= size) ? size_view : (size_aligned - i);
+
+                ctx->buffers[ctx->n_buffers].name = name;
+                ctx->buffers[ctx->n_buffers].data = (void *) ((uint8_t *) data + i);
+                ctx->buffers[ctx->n_buffers].size = size_step_aligned;
+
+                ctx->buffers[ctx->n_buffers].metal = [ctx->device newBufferWithBytesNoCopy:(void *) ((uint8_t *) data + i) length:size_step_aligned options:MTLResourceStorageModeShared deallocator:nil];
+
+                if (ctx->buffers[ctx->n_buffers].metal == nil) {
+                    fprintf(stderr, "%s: failed to allocate '%-16s' buffer, size = %8.2f MB\n", __func__, name, size_step_aligned / 1024.0 / 1024.0);
+                    return false;
+                }
+
+                fprintf(stderr, "%s: allocated '%-16s' buffer, size = %8.2f MB, offs = %12ld", __func__, name, size_step_aligned / 1024.0 / 1024.0, i);
+                if (i + size_step < size) {
+                    fprintf(stderr, "\n");
+                }
+
+                ++ctx->n_buffers;
+            }
+        }
+
+        fprintf(stderr, ", (%8.2f / %8.2f)",
+                ctx->device.currentAllocatedSize / 1024.0 / 1024.0,
+                ctx->device.recommendedMaxWorkingSetSize / 1024.0 / 1024.0);
+
+        if (ctx->device.currentAllocatedSize > ctx->device.recommendedMaxWorkingSetSize) {
+            fprintf(stderr, ", warning: current allocated size is greater than the recommended max working set size\n");
+        } else {
+            fprintf(stderr, "\n");
+        }
+    }
+
+    return true;
+}
+
+void ggml_metal_set_tensor(
+        struct ggml_metal_context * ctx,
+        struct ggml_tensor * t) {
+    metal_printf("%s: set input for tensor '%s'\n", __func__, t->name);
+
+    size_t offs;
+    id<MTLBuffer> id_dst = ggml_metal_get_buffer(ctx, t, &offs);
+
+    memcpy((void *) ((uint8_t *) id_dst.contents + offs), t->data, ggml_nbytes(t));
+}
+
+void ggml_metal_get_tensor(
+        struct ggml_metal_context * ctx,
+        struct ggml_tensor * t) {
+    metal_printf("%s: extract results for tensor '%s'\n", __func__, t->name);
+
+    size_t offs;
+    id<MTLBuffer> id_src = ggml_metal_get_buffer(ctx, t, &offs);
+
+    memcpy(t->data, (void *) ((uint8_t *) id_src.contents + offs), ggml_nbytes(t));
+}
+
+void ggml_metal_graph_find_concurrency(
+        struct ggml_metal_context * ctx,
+        struct ggml_cgraph * gf, bool check_mem) {
+    int search_depth = gf->n_nodes; //we only find concurrency in this range to avoid wasting too much time
+    int nodes_unused[GGML_MAX_CONCUR];
+
+    for (int i = 0; i < GGML_MAX_CONCUR; i++) { ctx->concur_list[i] = 0; }
+    for (int i = 0; i < gf->n_nodes;     i++) { nodes_unused[i]     = 1; }
+    ctx->concur_list_len = 0;
+
+    int n_left    = gf->n_nodes;
+    int n_start   = 0; // all nodes before n_start at nodes_unused array have been sorted and store back to ctx->concur_list
+    int level_pos = 0; // at ctx->concur_list, the last layer (level) ends at level_pos
+
+    while (n_left > 0) {
+        // number of nodes at a layer (that can be issued concurrently)
+        int concurrency = 0;
+        for (int i = n_start; i < ((n_start + search_depth > gf->n_nodes) ? gf->n_nodes : n_start + search_depth); i++) {
+            if (nodes_unused[i]) {
+                // if the requirements for gf->nodes[i] are satisfied
+                int exe_flag = 1;
+
+                // scan all srcs
+                for (int src_ind = 0; src_ind < GGML_MAX_SRC; src_ind++) {
+                    struct ggml_tensor * src_cur = gf->nodes[i]->src[src_ind];
+                    if (src_cur) {
+                        // if is leaf nodes it's satisfied.
+                        // TODO: ggml_is_leaf()
+                        if (src_cur->op == GGML_OP_NONE && src_cur->grad == NULL) {
+                            continue;
+                        }
+
+                        // otherwise this src should be the output from previous nodes.
+                        int is_found = 0;
+
+                        // scan 2*search_depth back because we inserted barrier.
+                        //for (int j = ((level_pos - 2*search_depth) < 0 ? 0 : (level_pos - 2*search_depth)); j < level_pos; j++) {
+                        for (int j = MAX(0, level_pos - 2*search_depth); j < level_pos; j++) {
+                            if (ctx->concur_list[j] >= 0 && gf->nodes[ctx->concur_list[j]] == src_cur) {
+                                is_found = 1;
+                                break;
+                            }
+                        }
+                        if (is_found == 0) {
+                            exe_flag = 0;
+                            break;
+                        }
+                    }
+                }
+                if (exe_flag && check_mem) {
+                    // check if nodes[i]'s data will be overwritten by a node before nodes[i].
+                    // if node[5] and node[3] write to the same memory region, then we can't issue node[5] before node[3]
+                    int64_t data_start = (int64_t) gf->nodes[i]->data;
+                    int64_t length     = (int64_t) ggml_nbytes(gf->nodes[i]);
+                    for (int j = n_start; j < i; j++) {
+                        if (nodes_unused[j] && gf->nodes[j]->op != GGML_OP_RESHAPE \
+                                            && gf->nodes[j]->op != GGML_OP_VIEW \
+                                            && gf->nodes[j]->op != GGML_OP_TRANSPOSE \
+                                            && gf->nodes[j]->op != GGML_OP_PERMUTE) {
+                            if (((int64_t)gf->nodes[j]->data) >= data_start + length || \
+                                ((int64_t)gf->nodes[j]->data) + (int64_t) ggml_nbytes(gf->nodes[j]) <= data_start) {
+                                continue;
+                            }
+
+                            exe_flag = 0;
+                        }
+                    }
+                }
+                if (exe_flag) {
+                    ctx->concur_list[level_pos + concurrency] = i;
+                    nodes_unused[i] = 0;
+                    concurrency++;
+                    ctx->concur_list_len++;
+                }
+            }
+        }
+        n_left -= concurrency;
+        // adding a barrier different layer
+        ctx->concur_list[level_pos + concurrency] = -1;
+        ctx->concur_list_len++;
+        // jump all sorted nodes at nodes_bak
+        while (!nodes_unused[n_start]) {
+            n_start++;
+        }
+        level_pos += concurrency + 1;
+    }
+
+    if (ctx->concur_list_len > GGML_MAX_CONCUR) {
+        fprintf(stderr, "%s: too many elements for metal ctx->concur_list!\n", __func__);
+    }
+}
+
+void ggml_metal_graph_compute(
+        struct ggml_metal_context * ctx,
+               struct ggml_cgraph * gf) {
+    metal_printf("%s: evaluating graph\n", __func__);
+
+    @autoreleasepool {
+
+    // if there is ctx->concur_list, dispatch concurrently
+    // else fallback to serial dispatch
+    MTLComputePassDescriptor * edesc = MTLComputePassDescriptor.computePassDescriptor;
+
+    const bool has_concur = ctx->concur_list_len && ctx->concur_list_len <= GGML_MAX_CONCUR;
+
+    const int n_nodes  = has_concur ? ctx->concur_list_len      : gf->n_nodes;
+    edesc.dispatchType = has_concur ? MTLDispatchTypeConcurrent : MTLDispatchTypeSerial;
+
+    // create multiple command buffers and enqueue them
+    // then, we encode the graph into the command buffers in parallel
+
+    const int n_cb = ctx->n_cb;
+
+    for (int i = 0; i < n_cb; ++i) {
+        ctx->command_buffers[i] = [ctx->queue commandBuffer];
+
+        // enqueue the command buffers in order to specify their execution order
+        [ctx->command_buffers[i] enqueue];
+
+        ctx->command_encoders[i] = [ctx->command_buffers[i] computeCommandEncoderWithDescriptor: edesc];
+    }
+
+    for (int cb_idx = 0; cb_idx < n_cb; ++cb_idx) {
+        const int n_nodes_per_cb = (n_nodes + n_cb - 1) / n_cb;
+
+        dispatch_async(ctx->d_queue, ^{
+            size_t offs_src0 = 0;
+            size_t offs_src1 = 0;
+            size_t offs_dst  = 0;
+
+            id<MTLCommandBuffer> command_buffer  = ctx->command_buffers[cb_idx];
+            id<MTLComputeCommandEncoder> encoder = ctx->command_encoders[cb_idx];
+
+            const int node_start =                                      (cb_idx + 0) * n_nodes_per_cb;
+            const int node_end   = MIN((cb_idx == n_cb - 1) ? n_nodes : (cb_idx + 1) * n_nodes_per_cb, n_nodes);
+
+            for (int ind = node_start; ind < node_end; ++ind) {
+                const int i = has_concur ? ctx->concur_list[ind] : ind;
+
+                if (i == -1) {
+                    [encoder memoryBarrierWithScope:MTLBarrierScopeBuffers];
+                    continue;
+                }
+
+                metal_printf("%s: encoding node %3d, op = %8s\n", __func__, i, ggml_op_name(gf->nodes[i]->op));
+
+                struct ggml_tensor * src0 = gf->nodes[i]->src[0];
+                struct ggml_tensor * src1 = gf->nodes[i]->src[1];
+                struct ggml_tensor * dst  = gf->nodes[i];
+
+                const int64_t  ne00 = src0 ? src0->ne[0] : 0;
+                const int64_t  ne01 = src0 ? src0->ne[1] : 0;
+                const int64_t  ne02 = src0 ? src0->ne[2] : 0;
+                const int64_t  ne03 = src0 ? src0->ne[3] : 0;
+
+                const uint64_t nb00 = src0 ? src0->nb[0] : 0;
+                const uint64_t nb01 = src0 ? src0->nb[1] : 0;
+                const uint64_t nb02 = src0 ? src0->nb[2] : 0;
+                const uint64_t nb03 = src0 ? src0->nb[3] : 0;
+
+                const int64_t  ne10 = src1 ? src1->ne[0] : 0;
+                const int64_t  ne11 = src1 ? src1->ne[1] : 0;
+                const int64_t  ne12 = src1 ? src1->ne[2] : 0;
+                const int64_t  ne13 = src1 ? src1->ne[3] : 0; UNUSED(ne13);
+
+                const uint64_t nb10 = src1 ? src1->nb[0] : 0;
+                const uint64_t nb11 = src1 ? src1->nb[1] : 0;
+                const uint64_t nb12 = src1 ? src1->nb[2] : 0;
+                const uint64_t nb13 = src1 ? src1->nb[3] : 0; UNUSED(nb13);
+
+                const int64_t  ne0  = dst ? dst->ne[0] : 0;
+                const int64_t  ne1  = dst ? dst->ne[1] : 0;
+                const int64_t  ne2  = dst ? dst->ne[2] : 0;
+                const int64_t  ne3  = dst ? dst->ne[3] : 0;
+
+                const uint64_t nb0  = dst ? dst->nb[0] : 0;
+                const uint64_t nb1  = dst ? dst->nb[1] : 0;
+                const uint64_t nb2  = dst ? dst->nb[2] : 0;
+                const uint64_t nb3  = dst ? dst->nb[3] : 0;
+
+                const enum ggml_type src0t = src0 ? src0->type : GGML_TYPE_COUNT;
+                const enum ggml_type src1t = src1 ? src1->type : GGML_TYPE_COUNT;
+                const enum ggml_type dstt  = dst  ? dst->type  : GGML_TYPE_COUNT;
+
+                id<MTLBuffer> id_src0 = src0 ? ggml_metal_get_buffer(ctx, src0, &offs_src0) : nil;
+                id<MTLBuffer> id_src1 = src1 ? ggml_metal_get_buffer(ctx, src1, &offs_src1) : nil;
+                id<MTLBuffer> id_dst  = dst  ? ggml_metal_get_buffer(ctx, dst,  &offs_dst)  : nil;
+
+                //metal_printf("%s: op - %s\n", __func__, ggml_op_name(dst->op));
+                //if (src0) {
+                //    metal_printf("%s: src0 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src0t), ne00, ne01, ne02,
+                //            ggml_is_contiguous(src0), src0->name);
+                //}
+                //if (src1) {
+                //    metal_printf("%s: src1 - %4s [%5lld, %5lld, %5lld], %d, %s\n", __func__, ggml_type_name(src1t), ne10, ne11, ne12,
+                //            ggml_is_contiguous(src1), src1->name);
+                //}
+                //if (dst) {
+                //    metal_printf("%s: dst  - %4s [%5lld, %5lld, %5lld], 1, %s\n",  __func__, ggml_type_name(dstt),  ne0,  ne1,  ne2,
+                //            dst->name);
+                //}
+
+                switch (dst->op) {
+                    case GGML_OP_NONE:
+                    case GGML_OP_RESHAPE:
+                    case GGML_OP_VIEW:
+                    case GGML_OP_TRANSPOSE:
+                    case GGML_OP_PERMUTE:
+                        {
+                            // noop
+                        } break;
+                    case GGML_OP_ADD:
+                        {
+                            if (ggml_nelements(src1) == ne10) {
+                                // src1 is a row
+                                [encoder setComputePipelineState:ctx->pipeline_add_row];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_add];
+                            }
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+
+                            const int64_t n = ggml_nelements(dst);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_MUL:
+                        {
+                            if (ggml_nelements(src1) == ne10) {
+                                // src1 is a row
+                                [encoder setComputePipelineState:ctx->pipeline_mul_row];
+                            } else {
+                                [encoder setComputePipelineState:ctx->pipeline_mul];
+                            }
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+
+                            const int64_t n = ggml_nelements(dst);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_SCALE:
+                        {
+                            const float scale = *(const float *) src1->data;
+
+                            [encoder setComputePipelineState:ctx->pipeline_scale];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&scale length:sizeof(scale) atIndex:2];
+
+                            const int64_t n = ggml_nelements(dst);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_UNARY:
+                        switch (ggml_get_unary_op(gf->nodes[i])) {
+                            case GGML_UNARY_OP_SILU:
+                                {
+                                    [encoder setComputePipelineState:ctx->pipeline_silu];
+                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                                    const int64_t n = ggml_nelements(dst);
+
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                } break;
+                            case GGML_UNARY_OP_RELU:
+                                {
+                                    [encoder setComputePipelineState:ctx->pipeline_relu];
+                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                                    const int64_t n = ggml_nelements(dst);
+
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                } break;
+                            case GGML_UNARY_OP_GELU:
+                                {
+                                    [encoder setComputePipelineState:ctx->pipeline_gelu];
+                                    [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                    [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+
+                                    const int64_t n = ggml_nelements(dst);
+
+                                    [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                                } break;
+                            default:
+                                {
+                                    fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                                    GGML_ASSERT(false);
+                                }
+                        } break;
+                    case GGML_OP_SOFT_MAX:
+                        {
+                            const int nth = 32;
+
+                            [encoder setComputePipelineState:ctx->pipeline_soft_max];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:2];
+                            [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:3];
+                            [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:4];
+                            [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
+                    case GGML_OP_DIAG_MASK_INF:
+                        {
+                            const int n_past = ((int32_t *)(dst->op_params))[0];
+
+                            [encoder setComputePipelineState:ctx->pipeline_diag_mask_inf];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00   length:sizeof(ne00) atIndex:2];
+                            [encoder setBytes:&ne01   length:sizeof(ne01) atIndex:3];
+                            [encoder setBytes:&n_past length:sizeof(int)  atIndex:4];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_MUL_MAT:
+                        {
+                            // TODO: needs to be updated after PR: https://github.com/ggerganov/ggml/pull/224
+
+                            GGML_ASSERT(ne00 == ne10);
+                            // GGML_ASSERT(ne02 == ne12); // Should be checked on individual data types until broadcast is implemented everywhere
+                            uint gqa = ne12/ne02;
+                            GGML_ASSERT(ne03 == ne13);
+
+                            // for now the matrix-matrix multiplication kernel only works on A14+/M1+ SoCs
+                            // AMD GPU and older A-chips will reuse matrix-vector multiplication kernel
+                            if (ggml_is_contiguous(src0) &&
+                                ggml_is_contiguous(src1) &&
+                                src1t == GGML_TYPE_F32 &&
+                                [ctx->device supportsFamily:MTLGPUFamilyApple7] &&
+                                ne00%32 == 0 &&
+                                ne11 > 1) {
+                                switch (src0->type) {
+                                    case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_mul_mm_f16_f32];  break;
+                                    case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_0_f32]; break;
+                                    case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_1_f32]; break;
+                                    case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q8_0_f32]; break;
+                                    case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q2_K_f32]; break;
+                                    case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q3_K_f32]; break;
+                                    case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break;
+                                    case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break;
+                                    case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break;
+                                    default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
+                                }
+                                [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
+                                [encoder setBuffer:id_src1 offset:offs_src1    atIndex:1];
+                                [encoder setBuffer:id_dst  offset:offs_dst     atIndex:2];
+                                [encoder setBytes:&ne00    length:sizeof(ne00) atIndex:3];
+                                [encoder setBytes:&ne02    length:sizeof(ne02) atIndex:4];
+                                [encoder setBytes:&nb01    length:sizeof(nb01) atIndex:5];
+                                [encoder setBytes:&nb02    length:sizeof(nb02) atIndex:6];
+                                [encoder setBytes:&ne12    length:sizeof(ne12) atIndex:7];
+                                [encoder setBytes:&ne0     length:sizeof(ne0)  atIndex:8];
+                                [encoder setBytes:&ne1     length:sizeof(ne1)  atIndex:9];
+                                [encoder setBytes:&gqa     length:sizeof(gqa)  atIndex:10];
+                                [encoder setThreadgroupMemoryLength:8192 atIndex:0];
+                                [encoder dispatchThreadgroups:MTLSizeMake( (ne11+31)/32, (ne01+63) / 64, ne12) threadsPerThreadgroup:MTLSizeMake(128, 1, 1)];
+                            } else {
+                                int nth0 = 32;
+                                int nth1 = 1;
+
+                                // use custom matrix x vector kernel
+                                switch (src0t) {
+                                    case GGML_TYPE_F16:
+                                        {
+                                            nth0 = 64;
+                                            nth1 = 1;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_f16_f32];
+                                        } break;
+                                    case GGML_TYPE_Q4_0:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_0_f32];
+                                        } break;
+                                    case GGML_TYPE_Q4_1:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_1_f32];
+                                        } break;
+                                    case GGML_TYPE_Q8_0:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 8;
+                                            nth1 = 8;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q8_0_f32];
+                                        } break;
+                                    case GGML_TYPE_Q2_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 2;
+                                            nth1 = 32;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q2_K_f32];
+                                        } break;
+                                    case GGML_TYPE_Q3_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 2;
+                                            nth1 = 32;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q3_K_f32];
+                                        } break;
+                                    case GGML_TYPE_Q4_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 2;
+                                            nth1 = 32;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q4_K_f32];
+                                        } break;
+                                    case GGML_TYPE_Q5_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 2;
+                                            nth1 = 32;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q5_K_f32];
+                                        } break;
+                                    case GGML_TYPE_Q6_K:
+                                        {
+                                            GGML_ASSERT(ne02 == 1);
+                                            GGML_ASSERT(ne12 == 1);
+
+                                            nth0 = 2;
+                                            nth1 = 32;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mat_q6_K_f32];
+                                        } break;
+                                    default:
+                                        {
+                                            fprintf(stderr, "Asserting on type %d\n",(int)src0t);
+                                            GGML_ASSERT(false && "not implemented");
+                                        }
+                                };
+
+                                [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                                [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                                [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                                [encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
+                                [encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
+                                [encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
+                                [encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
+                                [encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
+                                [encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
+                                [encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9];
+                                [encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10];
+                                [encoder setBytes:&ne12 length:sizeof(ne12) atIndex:11];
+                                [encoder setBytes:&nb10 length:sizeof(nb10) atIndex:12];
+                                [encoder setBytes:&nb11 length:sizeof(nb11) atIndex:13];
+                                [encoder setBytes:&nb12 length:sizeof(nb12) atIndex:14];
+                                [encoder setBytes:&ne0  length:sizeof(ne0)  atIndex:15];
+                                [encoder setBytes:&ne1  length:sizeof(ne1)  atIndex:16];
+                                [encoder setBytes:&gqa  length:sizeof(gqa)  atIndex:17];
+
+                                if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 || src0t == GGML_TYPE_Q8_0 ||
+                                    src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_Q4_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                                else if (src0t == GGML_TYPE_Q3_K) {
+#ifdef GGML_QKK_64
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+#else
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+#endif
+                                }
+                                else if (src0t == GGML_TYPE_Q5_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                                else if (src0t == GGML_TYPE_Q6_K) {
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 1)/2, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                } else {
+                                    [encoder setThreadgroupMemoryLength:nth0*sizeof(float) atIndex:0];
+                                    [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne11, ne12) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
+                            }
+                        } break;
+                    case GGML_OP_GET_ROWS:
+                        {
+                            switch (src0->type) {
+                                case GGML_TYPE_F16:  [encoder setComputePipelineState:ctx->pipeline_get_rows_f16];  break;
+                                case GGML_TYPE_Q4_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_0]; break;
+                                case GGML_TYPE_Q4_1: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_1]; break;
+                                case GGML_TYPE_Q8_0: [encoder setComputePipelineState:ctx->pipeline_get_rows_q8_0]; break;
+                                case GGML_TYPE_Q2_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q2_K]; break;
+                                case GGML_TYPE_Q3_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q3_K]; break;
+                                case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q4_K]; break;
+                                case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_K]; break;
+                                case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_K]; break;
+                                default: GGML_ASSERT(false && "not implemented");
+                            }
+
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:2];
+                            [encoder setBytes:&(src0->ne[0]) length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&(src0->nb[1]) length:sizeof(uint64_t) atIndex:4];
+                            [encoder setBytes:&(dst->nb[1])  length:sizeof(uint64_t) atIndex:5];
+
+                            const int64_t n = ggml_nelements(src1);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(n, 1, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_RMS_NORM:
+                        {
+                            float eps;
+                            memcpy(&eps, dst->op_params, sizeof(float));
+
+                            const int nth = 512;
+
+                            [encoder setComputePipelineState:ctx->pipeline_rms_norm];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:3];
+                            [encoder setBytes:&eps  length:sizeof(   float) atIndex:4];
+                            [encoder setThreadgroupMemoryLength:nth/32*sizeof(float) atIndex:0];
+
+                            const int64_t nrows = ggml_nrows(src0);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
+                    case GGML_OP_NORM:
+                        {
+                            float eps;
+                            memcpy(&eps, dst->op_params, sizeof(float));
+
+                            const int nth = 256;
+
+                            [encoder setComputePipelineState:ctx->pipeline_norm];
+                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
+                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:3];
+                            [encoder setBytes:&eps     length:sizeof(   float) atIndex:4];
+                            [encoder setThreadgroupMemoryLength:nth*sizeof(float) atIndex:0];
+
+                            const int64_t nrows = ggml_nrows(src0);
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(nrows, 1, 1) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
+                    case GGML_OP_ALIBI:
+                        {
+                            GGML_ASSERT((src0t == GGML_TYPE_F32));
+
+                            const int n_past = ((int32_t *) dst->op_params)[0]; UNUSED(n_past);
+                            const int n_head = ((int32_t *) dst->op_params)[1];
+                            float max_bias;
+                            memcpy(&max_bias, (int32_t *) dst->op_params + 2, sizeof(float));
+
+                            if (__builtin_popcount(n_head) != 1) {
+                                GGML_ASSERT(false && "only power-of-two n_head implemented");
+                            }
+
+                            const int n_heads_log2_floor = 1 << (int) floor(log2(n_head));
+                            const float m0 = powf(2.0f, -(max_bias) / n_heads_log2_floor);
+
+                            [encoder setComputePipelineState:ctx->pipeline_alibi_f32];
+                            [encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst  atIndex:1];
+                            [encoder setBytes:&ne00 length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&ne01 length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne02 length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne03 length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&nb00 length:sizeof(uint64_t) atIndex:6];
+                            [encoder setBytes:&nb01 length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb02 length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb03 length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&ne0  length:sizeof( int64_t) atIndex:10];
+                            [encoder setBytes:&ne1  length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne2  length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne3  length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&nb0  length:sizeof(uint64_t) atIndex:14];
+                            [encoder setBytes:&nb1  length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb2  length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb3  length:sizeof(uint64_t) atIndex:17];
+                            [encoder setBytes:&m0  length:sizeof(    float) atIndex:18];
+
+                            const int nth = 32;
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
+                    case GGML_OP_ROPE:
+                        {
+                            const int n_past = ((int32_t *) dst->op_params)[0];
+                            const int n_dims = ((int32_t *) dst->op_params)[1];
+                            const int mode   = ((int32_t *) dst->op_params)[2];
+
+                            float freq_base;
+                            float freq_scale;
+                            memcpy(&freq_base,  (int32_t *) dst->op_params + 4, sizeof(float));
+                            memcpy(&freq_scale, (int32_t *) dst->op_params + 5, sizeof(float));
+
+                            [encoder setComputePipelineState:ctx->pipeline_rope];
+                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
+                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
+                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
+                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
+                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
+                            [encoder setBytes:&n_past  length:sizeof(     int) atIndex:18];
+                            [encoder setBytes:&n_dims  length:sizeof(     int) atIndex:19];
+                            [encoder setBytes:&mode    length:sizeof(     int) atIndex:20];
+                            [encoder setBytes:&freq_base  length:sizeof(float) atIndex:21];
+                            [encoder setBytes:&freq_scale length:sizeof(float) atIndex:22];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
+                        } break;
+                    case GGML_OP_DUP:
+                    case GGML_OP_CPY:
+                    case GGML_OP_CONT:
+                        {
+                            const int nth = 32;
+
+                            switch (src0t) {
+                                case GGML_TYPE_F32:
+                                    {
+                                        switch (dstt) {
+                                            case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f16]; break;
+                                            case GGML_TYPE_F32: [encoder setComputePipelineState:ctx->pipeline_cpy_f32_f32]; break;
+                                            default: GGML_ASSERT(false && "not implemented");
+                                        };
+                                    } break;
+                                case GGML_TYPE_F16:
+                                    {
+                                        switch (dstt) {
+                                            case GGML_TYPE_F16: [encoder setComputePipelineState:ctx->pipeline_cpy_f16_f16]; break;
+                                            case GGML_TYPE_F32: GGML_ASSERT(false && "cpy_f16_f32 not implemented"); break;
+                                            default: GGML_ASSERT(false && "not implemented");
+                                        };
+                                    } break;
+                                default: GGML_ASSERT(false && "not implemented");
+                            }
+
+                            [encoder setBuffer:id_src0 offset:offs_src0        atIndex:0];
+                            [encoder setBuffer:id_dst  offset:offs_dst         atIndex:1];
+                            [encoder setBytes:&ne00    length:sizeof( int64_t) atIndex:2];
+                            [encoder setBytes:&ne01    length:sizeof( int64_t) atIndex:3];
+                            [encoder setBytes:&ne02    length:sizeof( int64_t) atIndex:4];
+                            [encoder setBytes:&ne03    length:sizeof( int64_t) atIndex:5];
+                            [encoder setBytes:&nb00    length:sizeof(uint64_t) atIndex:6];
+                            [encoder setBytes:&nb01    length:sizeof(uint64_t) atIndex:7];
+                            [encoder setBytes:&nb02    length:sizeof(uint64_t) atIndex:8];
+                            [encoder setBytes:&nb03    length:sizeof(uint64_t) atIndex:9];
+                            [encoder setBytes:&ne0     length:sizeof( int64_t) atIndex:10];
+                            [encoder setBytes:&ne1     length:sizeof( int64_t) atIndex:11];
+                            [encoder setBytes:&ne2     length:sizeof( int64_t) atIndex:12];
+                            [encoder setBytes:&ne3     length:sizeof( int64_t) atIndex:13];
+                            [encoder setBytes:&nb0     length:sizeof(uint64_t) atIndex:14];
+                            [encoder setBytes:&nb1     length:sizeof(uint64_t) atIndex:15];
+                            [encoder setBytes:&nb2     length:sizeof(uint64_t) atIndex:16];
+                            [encoder setBytes:&nb3     length:sizeof(uint64_t) atIndex:17];
+
+                            [encoder dispatchThreadgroups:MTLSizeMake(ne01, ne02, ne03) threadsPerThreadgroup:MTLSizeMake(nth, 1, 1)];
+                        } break;
+                    default:
+                        {
+                            fprintf(stderr, "%s: node %3d, op = %8s not implemented\n", __func__, i, ggml_op_name(dst->op));
+                            GGML_ASSERT(false);
+                        }
+                }
+            }
+
+            if (encoder != nil) {
+                [encoder endEncoding];
+                encoder = nil;
+            }
+
+            [command_buffer commit];
+        });
+    }
+
+    // wait for all threads to finish
+    dispatch_barrier_sync(ctx->d_queue, ^{});
+
+    // check status of command buffers
+    // needed to detect if the device ran out-of-memory for example (#1881)
+    for (int i = 0; i < n_cb; i++) {
+        [ctx->command_buffers[i] waitUntilCompleted];
+
+        MTLCommandBufferStatus status = (MTLCommandBufferStatus) [ctx->command_buffers[i] status];
+        if (status != MTLCommandBufferStatusCompleted) {
+            fprintf(stderr, "%s: command buffer %d failed with status %lu\n", __func__, i, status);
+            GGML_ASSERT(false);
+        }
+    }
+
+    }
+}