llama_cpp 0.14.0 → 0.14.1

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

@@ -221,29 +221,29 @@ void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_ten
 GGML_CALL void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(buf != NULL && "tensor buffer not set");
+    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
 
     if (!size) {
         return;
     }
 
-    tensor->buffer->iface.set_tensor(buf, tensor, data, offset, size);
+    buf->iface.set_tensor(buf, tensor, data, offset, size);
 }
 
 GGML_CALL void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
 
+    GGML_ASSERT(buf != NULL && "tensor buffer not set");
     GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
-    GGML_ASSERT(tensor->buffer != NULL && "tensor buffer not set");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
 
     if (!size) {
         return;
     }
 
-    tensor->buffer->iface.get_tensor(buf, tensor, data, offset, size);
+    buf->iface.get_tensor(buf, tensor, data, offset, size);
 }
 
 void ggml_backend_synchronize(ggml_backend_t backend) {
@@ -255,18 +255,30 @@ void ggml_backend_synchronize(ggml_backend_t backend) {
 }
 
 ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    GGML_ASSERT(backend->iface.graph_plan_create != NULL);
+
     return backend->iface.graph_plan_create(backend, cgraph);
 }
 
 void ggml_backend_graph_plan_free(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    GGML_ASSERT(backend->iface.graph_plan_free != NULL);
+
     backend->iface.graph_plan_free(backend, plan);
 }
 
 enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan) {
+    GGML_ASSERT(backend->iface.graph_plan_compute != NULL);
+
     return backend->iface.graph_plan_compute(backend, plan);
 }
 
 enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    enum ggml_status err = ggml_backend_graph_compute_async(backend, cgraph);
+    ggml_backend_synchronize(backend);
+    return err;
+}
+
+bool ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     return backend->iface.graph_compute(backend, cgraph);
 }
 
@@ -314,34 +326,68 @@ void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst
     }
 }
 
-void ggml_backend_tensor_copy_async(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst) {
+void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, struct ggml_tensor * src, struct ggml_tensor * dst) {
     GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
 
     if (src == dst) {
         return;
     }
 
-    if (ggml_backend_buft_supports_backend(src->buffer->buft, backend) && ggml_backend_buft_supports_backend(dst->buffer->buft, backend)) {
-        if (backend->iface.cpy_tensor_async != NULL) {
-            if (backend->iface.cpy_tensor_async(backend, src, dst)) {
-                return;
-            }
+    if (backend_dst->iface.cpy_tensor_async != NULL) {
+        if (backend_dst->iface.cpy_tensor_async(backend_src, backend_dst, src, dst)) {
+            return;
         }
     }
 
-    size_t nbytes = ggml_nbytes(src);
+    // an async copy would normally happen after all the queued operations on both backends are completed
+    // sync src, set_async dst
     if (ggml_backend_buffer_is_host(src->buffer)) {
-        ggml_backend_tensor_set_async(backend, dst, src->data, 0, nbytes);
-    }
-    else {
+        ggml_backend_synchronize(backend_src);
+        ggml_backend_tensor_set_async(backend_dst, dst, src->data, 0, ggml_nbytes(src));
+    } else {
+        ggml_backend_synchronize(backend_src);
         ggml_backend_tensor_copy(src, dst);
+        ggml_backend_synchronize(backend_dst);
+    }
+}
+
+// events
+
+ggml_backend_event_t ggml_backend_event_new(ggml_backend_t backend) {
+    if (backend->iface.event_new == NULL) {
+        return NULL;
+    }
+    return backend->iface.event_new(backend);
+}
+
+void ggml_backend_event_free(ggml_backend_event_t event) {
+    if (event == NULL) {
+        return;
     }
+    event->backend->iface.event_free(event);
+}
+
+void ggml_backend_event_record(ggml_backend_event_t event) {
+    GGML_ASSERT(event->backend->iface.event_record != NULL);
+
+    event->backend->iface.event_record(event);
+}
+
+void ggml_backend_event_synchronize(ggml_backend_event_t event) {
+    GGML_ASSERT(event->backend->iface.event_synchronize != NULL);
+
+    event->backend->iface.event_synchronize(event);
 }
 
+void ggml_backend_event_wait(ggml_backend_t backend, ggml_backend_event_t event) {
+    GGML_ASSERT(backend->iface.event_wait != NULL);
+
+    backend->iface.event_wait(backend, event);
+}
 
 // backend registry
 
-#define GGML_MAX_BACKENDS_REG 16
+#define GGML_REG_MAX_BACKENDS 16
 
 struct ggml_backend_reg {
     char name[128];
@@ -350,7 +396,7 @@ struct ggml_backend_reg {
     void * user_data;
 };
 
-static struct ggml_backend_reg ggml_backend_registry[GGML_MAX_BACKENDS_REG];
+static struct ggml_backend_reg ggml_backend_registry[GGML_REG_MAX_BACKENDS];
 static size_t ggml_backend_registry_count = 0;
 
 GGML_CALL static ggml_backend_t ggml_backend_reg_cpu_init(const char * params, void * user_data);
@@ -395,7 +441,7 @@ GGML_CALL static void ggml_backend_registry_init(void) {
 }
 
 GGML_CALL void ggml_backend_register(const char * name, ggml_backend_init_fn init_fn, ggml_backend_buffer_type_t default_buffer_type, void * user_data) {
-    GGML_ASSERT(ggml_backend_registry_count < GGML_MAX_BACKENDS_REG);
+    GGML_ASSERT(ggml_backend_registry_count < GGML_REG_MAX_BACKENDS);
 
     size_t id = ggml_backend_registry_count;
 
@@ -746,8 +792,12 @@ GGML_CALL static enum ggml_status ggml_backend_cpu_graph_compute(ggml_backend_t
     struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
 
     if (cpu_ctx->work_size < cplan.work_size) {
-        // TODO: may be faster to free and use malloc to avoid the copy
-        cpu_ctx->work_data = realloc(cpu_ctx->work_data, cplan.work_size);
+        free(cpu_ctx->work_data);
+        cpu_ctx->work_data = malloc(cplan.work_size);
+        if (cpu_ctx->work_data == NULL) {
+            cpu_ctx->work_size = 0;
+            return GGML_STATUS_ALLOC_FAILED;
+        }
         cpu_ctx->work_size = cplan.work_size;
     }
     cplan.work_data = cpu_ctx->work_data;
@@ -784,6 +834,11 @@ static struct ggml_backend_i cpu_backend_i = {
     /* .graph_plan_compute      = */ ggml_backend_cpu_graph_plan_compute,
     /* .graph_compute           = */ ggml_backend_cpu_graph_compute,
     /* .supports_op             = */ ggml_backend_cpu_supports_op,
+    /* .event_new               = */ NULL,
+    /* .event_free              = */ NULL,
+    /* .event_record            = */ NULL,
+    /* .event_wait              = */ NULL,
+    /* .event_synchronize       = */ NULL,
 };
 
 static ggml_guid_t ggml_backend_cpu_guid(void) {
@@ -939,15 +994,27 @@ static bool ggml_is_view_op(enum ggml_op op) {
 
 // scheduler
 
-#define GGML_MAX_BACKENDS 16
-#define GGML_MAX_SPLITS 256
-#define GGML_MAX_SPLIT_INPUTS 16
+#ifndef GGML_SCHED_MAX_BACKENDS
+#define GGML_SCHED_MAX_BACKENDS 16
+#endif
+
+#ifndef GGML_SCHED_MAX_SPLITS
+#define GGML_SCHED_MAX_SPLITS 256
+#endif
+
+#ifndef GGML_SCHED_MAX_SPLIT_INPUTS
+#define GGML_SCHED_MAX_SPLIT_INPUTS 16
+#endif
+
+#ifndef GGML_SCHED_MAX_COPIES
+#define GGML_SCHED_MAX_COPIES 4
+#endif
 
 struct ggml_backend_sched_split {
     int backend_id;
     int i_start;
     int i_end;
-    struct ggml_tensor * inputs[GGML_MAX_SPLIT_INPUTS];
+    struct ggml_tensor * inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
     int n_inputs;
     // graph view of this split
     struct ggml_cgraph graph;
@@ -955,45 +1022,53 @@ struct ggml_backend_sched_split {
 
 struct ggml_backend_sched {
     bool is_reset; // true if the scheduler has been reset since the last graph split
+    bool is_alloc;
 
     int n_backends;
-    ggml_backend_t backends[GGML_MAX_BACKENDS];
-    ggml_backend_buffer_type_t bufts[GGML_MAX_BACKENDS];
 
+    ggml_backend_t backends[GGML_SCHED_MAX_BACKENDS];
+    ggml_backend_buffer_type_t bufts[GGML_SCHED_MAX_BACKENDS];
     ggml_gallocr_t galloc;
 
     // hash keys of the nodes in the graph
     struct ggml_hash_set    hash_set;
     // hash values
     int * tensor_backend_id;
-    struct ggml_tensor * (* tensor_copies)[GGML_MAX_BACKENDS];
+    struct ggml_tensor * (* tensor_copies)[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
 
-    int * node_backend_ids; // [n_nodes]
-    int n_nodes;
+    int * node_backend_ids; // [graph_size]
+    int * leaf_backend_ids; // [graph_size]
 
     // copy of the graph with modified inputs
     struct ggml_cgraph * graph;
 
-    struct ggml_backend_sched_split splits[GGML_MAX_SPLITS];
+    // graph splits
+    struct ggml_backend_sched_split splits[GGML_SCHED_MAX_SPLITS];
     int n_splits;
 
+    // pipeline parallelism support
+    int n_copies;
+    int cur_copy;
+    ggml_backend_event_t events[GGML_SCHED_MAX_BACKENDS][GGML_SCHED_MAX_COPIES];
+    struct ggml_tensor * graph_inputs[GGML_SCHED_MAX_SPLIT_INPUTS];
+    int n_graph_inputs;
+
     struct ggml_context * ctx;
 
     ggml_backend_sched_eval_callback callback_eval;
     void * callback_eval_user_data;
 
     // align context_buffer to GGML_MEM_ALIGN
-    #ifdef _MSC_VER
+#ifdef _MSC_VER
     __declspec(align(GGML_MEM_ALIGN))
-    #else
+#else
     __attribute__((aligned(GGML_MEM_ALIGN)))
-    #endif
-    char context_buffer[GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + sizeof(struct ggml_cgraph)];
+#endif
+    char context_buffer[GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + sizeof(struct ggml_cgraph)];
 };
 
-#define hash_id(node) ggml_hash_find_or_insert(sched->hash_set, node)
-#define tensor_backend_id(node) sched->tensor_backend_id[hash_id(node)]
-#define tensor_backend(node) (tensor_backend_id(node) == -1 ? NULL : sched->backends[tensor_backend_id(node)])
+#define hash_id(tensor) ggml_hash_find_or_insert(sched->hash_set, tensor)
+#define tensor_backend_id(tensor) sched->tensor_backend_id[hash_id(tensor)]
 
 // returns the priority of the backend, lower id is higher priority
 static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backend_t backend) {
@@ -1005,7 +1080,8 @@ static int ggml_backend_sched_backend_id(ggml_backend_sched_t sched, ggml_backen
     return -1;
 }
 
-static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, ggml_backend_buffer_t buffer) {
+static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, const struct ggml_tensor * tensor) {
+    ggml_backend_buffer_t buffer = tensor->buffer;
     if (buffer == NULL) {
         return -1;
     }
@@ -1016,12 +1092,16 @@ static int ggml_backend_sched_backend_from_buffer(ggml_backend_sched_t sched, gg
             return i;
         }
     }
-    GGML_ASSERT(false && "tensor buffer type not supported by any backend");
-    return -1; // silence warning
+
+    fprintf(stderr, "%s: error: no backend supports buffer type %s used in tensor %s\n",
+        __func__, ggml_backend_buffer_name(buffer), tensor->name);
+    GGML_ASSERT(false);
+
+    return -1;
 }
 
 #if 0
-static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS][128]; // debug only
+static char causes[GGML_DEFAULT_GRAPH_SIZE*16 + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS][128]; // debug only
 #define SET_CAUSE(node, ...) sprintf(causes[hash_id(node)], __VA_ARGS__)
 #define GET_CAUSE(node) causes[hash_id(node)]
 #else
@@ -1035,19 +1115,28 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
 
     // assign pre-allocated nodes to their backend
     // dst
-    int cur_backend = ggml_backend_sched_backend_from_buffer(sched, tensor->buffer);
+    int cur_backend = ggml_backend_sched_backend_from_buffer(sched, tensor);
     if (cur_backend != -1) {
-        SET_CAUSE(node, "1.dst");
+        SET_CAUSE(tensor, "1.dst");
         return cur_backend;
     }
+
     // view_src
     if (tensor->view_src != NULL) {
-        cur_backend = ggml_backend_sched_backend_from_buffer(sched, tensor->view_src->buffer);
+        cur_backend = ggml_backend_sched_backend_from_buffer(sched, tensor->view_src);
         if (cur_backend != -1) {
-            SET_CAUSE(node, "1.vsrc");
+            SET_CAUSE(tensor, "1.vsrc");
             return cur_backend;
         }
     }
+
+    // input
+    if (tensor->flags & GGML_TENSOR_FLAG_INPUT) {
+        cur_backend = sched->n_backends - 1; // last backend (assumed CPU)
+        SET_CAUSE(tensor, "1.inp");
+        return cur_backend;
+    }
+
     // assign nodes that use weights to the backend of the weights
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         const struct ggml_tensor * src = tensor->src[i];
@@ -1055,9 +1144,9 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
             continue;
         }
         if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
-            int src_backend = ggml_backend_sched_backend_from_buffer(sched, src->buffer);
+            int src_backend = ggml_backend_sched_backend_from_buffer(sched, src);
             // operations with weights are always run on the same backend as the weights
-            SET_CAUSE(node, "1.wgt%d", i);
+            SET_CAUSE(tensor, "1.wgt%d", i);
             return src_backend;
         }
     }
@@ -1093,7 +1182,7 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
         if (ggml_is_view_op(node->op)) {
             continue;
         }
-        ggml_backend_t tensor_backend = tensor_backend(node);
+        ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
         fprintf(stderr, "node #%3d (%10.10s): %20.20s (%5.5s) [%5.5s %8.8s]:", i, ggml_op_name(node->op), node->name,
             fmt_size(ggml_nbytes(node)), tensor_backend ? ggml_backend_name(tensor_backend) : "NULL", GET_CAUSE(node));
         for (int j = 0; j < GGML_MAX_SRC; j++) {
@@ -1101,7 +1190,7 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
             if (src == NULL) {
                 continue;
             }
-            ggml_backend_t src_backend = tensor_backend(src);
+            ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src);
             fprintf(stderr, " %20.20s (%5.5s) [%5.5s %8.8s]", src->name,
                 fmt_size(ggml_nbytes(src)), src_backend ? ggml_backend_name(src_backend) : "NULL", GET_CAUSE(src));
         }
@@ -1118,6 +1207,7 @@ static void ggml_backend_sched_print_assignments(ggml_backend_sched_t sched, str
 static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
     // reset splits
     sched->n_splits = 0;
+    sched->n_graph_inputs = 0;
     sched->is_reset = false;
 
     struct ggml_init_params params = {
@@ -1163,7 +1253,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         }
     }
 #ifdef DEBUG_PASS1
-    fprintf(stderr, "PASS 1 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+    fprintf(stderr, "PASS 1 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
 #endif
 
     // pass 2: expand current backend assignments
@@ -1171,10 +1261,11 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
     // expand gpu backends (i.e. non last prio) up and down, ignoring cpu (the lowest priority backend)
     // thus, cpu will never be used unless weights are on cpu, or there are no gpu ops between cpu ops
 
-    // pass 2.1 expand gpu up
+
+    // pass 2.2 expand gpu down
     {
         int cur_backend_id = -1;
-        for (int i = graph->n_nodes - 1; i >= 0; i--) {
+        for (int i = 0; i < graph->n_nodes; i++) {
             struct ggml_tensor * node = graph->nodes[i];
             if (ggml_is_view_op(node->op)) {
                 continue;
@@ -1189,15 +1280,15 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                 }
             } else {
                 tensor_backend_id(node) = cur_backend_id;
-                SET_CAUSE(node, "2.1");
+                SET_CAUSE(node, "2.2");
             }
         }
     }
 
-    // pass 2.2 expand gpu down
+    // pass 2.1 expand gpu up
     {
         int cur_backend_id = -1;
-        for (int i = 0; i < graph->n_nodes; i++) {
+        for (int i = graph->n_nodes - 1; i >= 0; i--) {
             struct ggml_tensor * node = graph->nodes[i];
             if (ggml_is_view_op(node->op)) {
                 continue;
@@ -1212,15 +1303,16 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                 }
             } else {
                 tensor_backend_id(node) = cur_backend_id;
-                SET_CAUSE(node, "2.2");
+                SET_CAUSE(node, "2.1");
             }
         }
     }
 
-    // pass 2.3 expand rest up
+
+    // pass 2.4 expand rest down
     {
         int cur_backend_id = -1;
-        for (int i = graph->n_nodes - 1; i >= 0; i--) {
+        for (int i = 0; i < graph->n_nodes; i++) {
             struct ggml_tensor * node = graph->nodes[i];
             if (ggml_is_view_op(node->op)) {
                 continue;
@@ -1230,15 +1322,14 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                 cur_backend_id = tensor_backend_id;
             } else {
                 tensor_backend_id(node) = cur_backend_id;
-                SET_CAUSE(node, "2.3");
+                SET_CAUSE(node, "2.4");
             }
         }
     }
-
-    // pass 2.4 expand rest down
+        // pass 2.3 expand rest up
     {
         int cur_backend_id = -1;
-        for (int i = 0; i < graph->n_nodes; i++) {
+        for (int i = graph->n_nodes - 1; i >= 0; i--) {
             struct ggml_tensor * node = graph->nodes[i];
             if (ggml_is_view_op(node->op)) {
                 continue;
@@ -1248,12 +1339,13 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                 cur_backend_id = tensor_backend_id;
             } else {
                 tensor_backend_id(node) = cur_backend_id;
-                SET_CAUSE(node, "2.4");
+                SET_CAUSE(node, "2.3");
             }
         }
     }
+
 #ifdef DEBUG_PASS2
-    fprintf(stderr, "PASS 2 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+    fprintf(stderr, "PASS 2 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
 #endif
 
     // pass 3: assign backends to remaining src from dst and view_src
@@ -1283,7 +1375,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         }
     }
 #ifdef DEBUG_PASS3
-    fprintf(stderr, "PASS 3 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+    fprintf(stderr, "PASS 3 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
 #endif
 
     // pass 4: split graph, find tensors that need to be copied
@@ -1315,7 +1407,7 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
             if (tensor_backend_id != cur_backend_id) {
                 sched->splits[cur_split].i_end = i;
                 cur_split++;
-                GGML_ASSERT(cur_split < GGML_MAX_SPLITS);
+                GGML_ASSERT(cur_split < GGML_SCHED_MAX_SPLITS);
                 sched->splits[cur_split].backend_id = tensor_backend_id;
                 sched->splits[cur_split].i_start = i;
                 sched->splits[cur_split].n_inputs = 0;
@@ -1328,25 +1420,57 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
                 if (src == NULL) {
                     continue;
                 }
+
                 int src_backend_id = tensor_backend_id(src);
                 assert(src_backend_id != -1); // all inputs should be assigned by now
+
+                if (src->flags & GGML_TENSOR_FLAG_INPUT)  {
+                    size_t id = hash_id(src);
+                    if (sched->tensor_copies[id][src_backend_id][0] == NULL) {
+                        ggml_backend_t backend = sched->backends[src_backend_id];
+                        for (int c = 0; c < sched->n_copies; c++) {
+                            struct ggml_tensor * tensor_copy;
+                            if (c == sched->cur_copy) {
+                                tensor_copy = src; // use the original tensor as the current copy
+                            } else {
+                                tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
+                                ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c);
+                            }
+                            if (sched->n_copies > 1) {
+                                ggml_set_input(tensor_copy);
+                                ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
+                            }
+                            sched->tensor_copies[id][src_backend_id][c] = tensor_copy;
+                            tensor_backend_id(tensor_copy) = src_backend_id;
+                            SET_CAUSE(tensor_copy, "4.cpy");
+                        }
+                        int n_graph_inputs = sched->n_graph_inputs++;
+                        GGML_ASSERT(n_graph_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
+                        sched->graph_inputs[n_graph_inputs] = src;
+                    }
+                }
+
                 if (src_backend_id != tensor_backend_id) {
                     // create a copy of the input in the split's backend
                     size_t id = hash_id(src);
-                    if (sched->tensor_copies[id][cur_backend_id] == NULL) {
+                    if (sched->tensor_copies[id][cur_backend_id][0] == NULL) {
                         ggml_backend_t backend = sched->backends[cur_backend_id];
-                        struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
-                        ggml_format_name(tensor_copy, "%s#%s", ggml_backend_name(backend), src->name);
-
-                        sched->tensor_copies[id][cur_backend_id] = tensor_copy;
-                        tensor_backend_id(tensor_copy) = cur_backend_id;
-                        SET_CAUSE(tensor_copy, "4.cpy");
-
+                        for (int c = 0; c < sched->n_copies; c++) {
+                            struct ggml_tensor * tensor_copy = ggml_dup_tensor_layout(sched->ctx, src);
+                            ggml_format_name(tensor_copy, "%s#%s#%d", ggml_backend_name(backend), src->name, c);
+                            if (sched->n_copies > 1) {
+                                ggml_set_input(tensor_copy);
+                                ggml_set_output(tensor_copy); // prevent ggml-alloc from overwriting the tensor
+                            }
+                            sched->tensor_copies[id][cur_backend_id][c] = tensor_copy;
+                            tensor_backend_id(tensor_copy) = cur_backend_id;
+                            SET_CAUSE(tensor_copy, "4.cpy");
+                        }
                         int n_inputs = sched->splits[cur_split].n_inputs++;
-                        GGML_ASSERT(n_inputs < GGML_MAX_SPLIT_INPUTS);
+                        GGML_ASSERT(n_inputs < GGML_SCHED_MAX_SPLIT_INPUTS);
                         sched->splits[cur_split].inputs[n_inputs] = src;
                     }
-                    node->src[j] = sched->tensor_copies[id][cur_backend_id];
+                    node->src[j] = sched->tensor_copies[id][cur_backend_id][sched->cur_copy];
                 }
             }
         }
@@ -1354,37 +1478,39 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
         sched->n_splits = cur_split + 1;
     }
 #ifdef DEBUG_PASS4
-    fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); sched_print_assignments(sched, graph);
+    fprintf(stderr, "PASS 4 ASSIGNMENTS\n"); ggml_backend_sched_print_assignments(sched, graph);
 #endif
 
 #ifndef NDEBUG
     // sanity check: all sources should have the same backend as the node
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
-        ggml_backend_t tensor_backend = tensor_backend(node);
+        ggml_backend_t tensor_backend = ggml_backend_sched_get_tensor_backend(sched, node);
         if (tensor_backend == NULL) {
             fprintf(stderr, "!!!!!!! %s has no backend\n", node->name);
         }
-        if (node->view_src != NULL && tensor_backend != tensor_backend(node->view_src)) {
+        if (node->view_src != NULL && tensor_backend != ggml_backend_sched_get_tensor_backend(sched, node->view_src)) {
             fprintf(stderr, "!!!!!!! %s has backend %s, view_src %s has backend %s\n",
                 node->name, tensor_backend ? ggml_backend_name(tensor_backend) : "NULL",
-                node->view_src->name, tensor_backend(node->view_src) ? ggml_backend_name(tensor_backend(node->view_src)) : "NULL");
+                node->view_src->name, ggml_backend_sched_get_tensor_backend(sched, node->view_src) ?
+                    ggml_backend_name(ggml_backend_sched_get_tensor_backend(sched, node->view_src)) : "NULL");
         }
         for (int j = 0; j < GGML_MAX_SRC; j++) {
             struct ggml_tensor * src = node->src[j];
             if (src == NULL) {
                 continue;
             }
-            ggml_backend_t src_backend = tensor_backend(src);
+            ggml_backend_t src_backend = ggml_backend_sched_get_tensor_backend(sched, src);
             if (src_backend != tensor_backend /* && src_backend != NULL */) {
                 fprintf(stderr, "!!!! %s has backend %s, src %d (%s) has backend %s\n",
                     node->name, tensor_backend ? ggml_backend_name(tensor_backend) : "NULL",
                     j, src->name, src_backend ? ggml_backend_name(src_backend) : "NULL");
             }
-            if (src->view_src != NULL && src_backend != tensor_backend(src->view_src)) {
+            if (src->view_src != NULL && src_backend != ggml_backend_sched_get_tensor_backend(sched, src->view_src)) {
                 fprintf(stderr, "!!!!!!! [src] %s has backend %s, view_src %s has backend %s\n",
                     src->name, src_backend ? ggml_backend_name(src_backend) : "NULL",
-                    src->view_src->name, tensor_backend(src->view_src) ? ggml_backend_name(tensor_backend(src->view_src)) : "NULL");
+                    src->view_src->name, ggml_backend_sched_get_tensor_backend(sched, src->view_src) ?
+                        ggml_backend_name(ggml_backend_sched_get_tensor_backend(sched, src->view_src)) : "NULL");
             }
         }
     }
@@ -1392,18 +1518,20 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 #endif
 
     // create copies of the graph for each split
-    // FIXME: avoid this copy, pass split inputs to ggml_gallocr_alloc_graph_n in some other way
-    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_MAX_SPLIT_INPUTS, false);
+    // TODO: avoid this copy
+    struct ggml_cgraph * graph_copy = ggml_new_graph_custom(sched->ctx, graph->n_nodes + sched->n_splits*GGML_SCHED_MAX_SPLIT_INPUTS, false);
     for (int i = 0; i < sched->n_splits; i++) {
         struct ggml_backend_sched_split * split = &sched->splits[i];
         split->graph = ggml_graph_view(graph, split->i_start, split->i_end);
 
+        // add inputs to the graph copy so that they are allocated by ggml-alloc at the start of the split
         for (int j = 0; j < split->n_inputs; j++) {
             struct ggml_tensor * input = split->inputs[j];
-            struct ggml_tensor * input_cpy = sched->tensor_copies[hash_id(input)][split->backend_id];
+            struct ggml_tensor * input_cpy = sched->tensor_copies[hash_id(input)][split->backend_id][sched->cur_copy];
 
             // add a dependency to the input source so that it is not freed before the copy is done
             struct ggml_tensor * input_dep = ggml_view_tensor(sched->ctx, input);
+            input_dep->src[0] = input;
             sched->node_backend_ids[graph_copy->n_nodes] = tensor_backend_id(input);
             graph_copy->nodes[graph_copy->n_nodes++] = input_dep;
 
@@ -1417,18 +1545,56 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
             graph_copy->nodes[graph_copy->n_nodes++] = graph->nodes[j];
         }
     }
+
+    if (sched->n_copies > 1) {
+        // add input copies as leafs so that they are allocated first
+        for (int i = 0; i < sched->n_graph_inputs; i++) {
+            struct ggml_tensor * input = sched->graph_inputs[i];
+            size_t id = hash_id(input);
+            int backend_id = tensor_backend_id(input);
+            for (int c = 0; c < sched->n_copies; c++) {
+                struct ggml_tensor * input_cpy = sched->tensor_copies[id][backend_id][c];
+                sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
+                graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
+            }
+        }
+
+        for (int i = 0; i < sched->n_splits; i++) {
+            struct ggml_backend_sched_split * split = &sched->splits[i];
+            int backend_id = split->backend_id;
+            for (int j = 0; j < split->n_inputs; j++) {
+                struct ggml_tensor * input = split->inputs[j];
+                size_t id = hash_id(input);
+                for (int c = 0; c < sched->n_copies; c++) {
+                    struct ggml_tensor * input_cpy = sched->tensor_copies[id][backend_id][c];
+                    sched->leaf_backend_ids[graph_copy->n_leafs] = backend_id;
+                    graph_copy->leafs[graph_copy->n_leafs++] = input_cpy;
+                }
+            }
+        }
+    }
+
+    // add leafs from the original graph
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        sched->leaf_backend_ids[graph_copy->n_leafs] = tensor_backend_id(leaf);
+        graph_copy->leafs[graph_copy->n_leafs++] = leaf;
+    }
+
     sched->graph = graph_copy;
 }
 
 static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
-    // ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids);
+    // allocate graph
     if (!ggml_gallocr_alloc_graph(sched->galloc, sched->graph)) {
+        // the re-allocation may cause the split inputs to be moved to a different address
+        ggml_backend_sched_synchronize(sched);
 #ifndef NDEBUG
-        fprintf(stderr, "ggml_backend_sched: failed to allocate graph, reserving\n");
+        fprintf(stderr, "%s: failed to allocate graph, reserving\n", __func__);
 #endif
-        ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids);
+        ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids, sched->leaf_backend_ids);
         if (!ggml_gallocr_alloc_graph(sched->galloc, sched->graph)) {
-            fprintf(stderr, "ggml_backend_sched: failed to allocate graph\n");
+            fprintf(stderr, "%s: failed to allocate graph\n", __func__);
             return false;
         }
     }
@@ -1437,9 +1603,6 @@ static bool ggml_backend_sched_alloc_splits(ggml_backend_sched_t sched) {
 }
 
 static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t sched) {
-    uint64_t copy_us[GGML_MAX_BACKENDS] = {0};
-    uint64_t compute_us[GGML_MAX_BACKENDS] = {0};
-
     struct ggml_backend_sched_split * splits = sched->splits;
 
     for (int i = 0; i < sched->n_splits; i++) {
@@ -1448,34 +1611,36 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
         ggml_backend_t split_backend = sched->backends[split_backend_id];
 
         // copy the input tensors to the split backend
-        uint64_t copy_start_us = ggml_time_us();
         for (int j = 0; j < split->n_inputs; j++) {
+            ggml_backend_t input_backend = ggml_backend_sched_get_tensor_backend(sched, split->inputs[j]);
             struct ggml_tensor * input = split->inputs[j];
-            struct ggml_tensor * input_cpy = sched->tensor_copies[hash_id(input)][split_backend_id];
+            struct ggml_tensor * input_cpy = sched->tensor_copies[hash_id(input)][split_backend_id][sched->cur_copy];
 
-            GGML_ASSERT(input->buffer != NULL);
-            GGML_ASSERT(input_cpy->buffer != NULL);
+            if (input->flags & GGML_TENSOR_FLAG_INPUT) {
+                // inputs from the user must be copied immediately to prevent the user overwriting the data before the copy is done
+                if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                    ggml_backend_event_synchronize(sched->events[split_backend_id][sched->cur_copy]);
+                } else {
+                    ggml_backend_synchronize(split_backend);
+                }
+                ggml_backend_tensor_copy(input, input_cpy);
+            } else {
+                if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                    ggml_backend_event_wait(split_backend, sched->events[split_backend_id][sched->cur_copy]);
+                } else {
+                    ggml_backend_synchronize(split_backend);
+                    ggml_backend_synchronize(input_backend);
+                }
 
-            ggml_backend_tensor_copy_async(split_backend, input, input_cpy);
+                ggml_backend_tensor_copy_async(input_backend, split_backend, input, input_cpy);
+            }
         }
-        //ggml_backend_synchronize(split_backend); // necessary to measure copy time
-        int64_t copy_end_us = ggml_time_us();
-        copy_us[split_backend_id] += copy_end_us - copy_start_us;
 
-#if 0
-        char split_filename[GGML_MAX_NAME];
-        snprintf(split_filename, GGML_MAX_NAME, "split_%i_%s.dot", i, ggml_backend_name(split_backend));
-        ggml_graph_dump_dot(split->graph, NULL, split_filename);
-#endif
-
-
-        uint64_t compute_start_us = ggml_time_us();
         if (!sched->callback_eval) {
-            enum ggml_status ec = ggml_backend_graph_compute(split_backend, &split->graph);
+            enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &split->graph);
             if (ec != GGML_STATUS_SUCCESS) {
                 return ec;
             }
-            //ggml_backend_synchronize(split_backend); // necessary to measure compute time
         } else {
             // similar to ggml_backend_compare_graph_backend
             for (int j0 = 0; j0 < split->graph.n_nodes; j0++) {
@@ -1494,11 +1659,14 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
 
                 struct ggml_cgraph gv = ggml_graph_view(&split->graph, j0, j1 + 1);
 
-                enum ggml_status ec = ggml_backend_graph_compute(split_backend, &gv);
+                enum ggml_status ec = ggml_backend_graph_compute_async(split_backend, &gv);
                 if (ec != GGML_STATUS_SUCCESS) {
                     return ec;
                 }
 
+                // TODO: pass backend to the callback, then the user can decide if they want to synchronize
+                ggml_backend_synchronize(split_backend);
+
                 if (need && !sched->callback_eval(t, false, sched->callback_eval_user_data)) {
                     break;
                 }
@@ -1506,39 +1674,54 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
                 j0 = j1;
             }
         }
-        uint64_t compute_end_us = ggml_time_us();
-        compute_us[split_backend_id] += compute_end_us - compute_start_us;
-    }
 
-#if 0
-    // per-backend timings
-    fprintf(stderr, "sched_compute_splits times (%d splits):\n", sched->n_splits);
-    for (int i = 0; i < sched->n_backends; i++) {
-        if (copy_us[i] > 0 || compute_us[i] > 0) {
-            fprintf(stderr, "\t%5.5s: %lu us copy, %lu us compute\n", ggml_backend_name(sched->backends[i]), copy_us[i], compute_us[i]);
+        // record the event of this copy
+        if (split->n_inputs > 0) {
+            if (sched->events[split_backend_id][sched->cur_copy] != NULL) {
+                ggml_backend_event_record(sched->events[split_backend_id][sched->cur_copy]);
+            }
         }
     }
-#endif
+
+    sched->cur_copy = (sched->cur_copy + 1) % sched->n_copies;
 
     return GGML_STATUS_SUCCESS;
 }
 
-ggml_backend_sched_t ggml_backend_sched_new(ggml_backend_t * backends, ggml_backend_buffer_type_t * bufts, int n_backends, size_t graph_size) {
+ggml_backend_sched_t ggml_backend_sched_new(
+        ggml_backend_t * backends,
+        ggml_backend_buffer_type_t * bufts,
+        int n_backends,
+        size_t graph_size,
+        bool parallel) {
     GGML_ASSERT(n_backends > 0);
-    GGML_ASSERT(n_backends <= GGML_MAX_BACKENDS);
+    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);
 
     // initialize hash table
-    sched->hash_set          = ggml_hash_set_new(graph_size + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS);
+    sched->hash_set          = ggml_hash_set_new(graph_size + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS);
     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->node_backend_ids  = calloc(sizeof(sched->node_backend_ids[0]), graph_size);
+    sched->leaf_backend_ids  = calloc(sizeof(sched->leaf_backend_ids[0]), graph_size);
 
     sched->n_backends = n_backends;
-    for (int i = 0; i < n_backends; i++) {
-        sched->backends[i] = backends[i];
-        sched->bufts[i] = bufts ? bufts[i] : ggml_backend_get_default_buffer_type(backends[i]);
+
+    sched->n_copies = parallel ? GGML_SCHED_MAX_COPIES : 1;
+
+    GGML_ASSERT(sched->n_copies <= GGML_SCHED_MAX_COPIES);
+
+    for (int b = 0; b < n_backends; b++) {
+        sched->backends[b] = backends[b];
+        sched->bufts[b] = bufts ? bufts[b] : ggml_backend_get_default_buffer_type(backends[b]);
+        GGML_ASSERT(ggml_backend_buft_supports_backend(sched->bufts[b], backends[b]));
+        if (sched->n_copies > 1) {
+            for (int c = 0; c < sched->n_copies; c++) {
+                sched->events[b][c] = ggml_backend_event_new(backends[b]);
+            }
+        }
     }
 
     sched->galloc = ggml_gallocr_new_n(sched->bufts, n_backends);
@@ -1552,12 +1735,18 @@ void ggml_backend_sched_free(ggml_backend_sched_t sched) {
     if (sched == NULL) {
         return;
     }
+    for (int b = 0; b < sched->n_backends; b++) {
+        for (int c = 0; c < sched->n_copies; c++) {
+            ggml_backend_event_free(sched->events[b][c]);
+        }
+    }
     ggml_gallocr_free(sched->galloc);
     ggml_free(sched->ctx);
     free(sched->hash_set.keys);
     free(sched->tensor_backend_id);
     free(sched->tensor_copies);
     free(sched->node_backend_ids);
+    free(sched->leaf_backend_ids);
     free(sched);
 }
 
@@ -1569,34 +1758,63 @@ void ggml_backend_sched_reset(ggml_backend_sched_t sched) {
     memset(sched->tensor_copies,      0, sizeof(sched->tensor_copies[0])     * hash_size);
 
     sched->is_reset = true;
+    sched->is_alloc = false;
 }
 
 bool ggml_backend_sched_reserve(ggml_backend_sched_t sched, struct ggml_cgraph * measure_graph) {
     ggml_backend_sched_split_graph(sched, measure_graph);
 
-    if (!ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids)) {
+    // TODO: extract this to a separate function
+    if (!ggml_gallocr_reserve_n(sched->galloc, sched->graph, sched->node_backend_ids, sched->leaf_backend_ids)) {
         return false;
     }
 
     ggml_backend_sched_reset(sched);
+    ggml_backend_sched_synchronize(sched);
+
+    return true;
+}
+
+bool ggml_backend_sched_alloc_graph(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS);
+
+    ggml_backend_sched_split_graph(sched, graph);
+
+    if (!ggml_backend_sched_alloc_splits(sched)) {
+        return false;
+    }
+
+    sched->is_alloc = true;
+
     return true;
 }
 
 enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
-    GGML_ASSERT((int)sched->hash_set.size >= graph->n_nodes + GGML_MAX_SPLITS*GGML_MAX_SPLIT_INPUTS);
+    enum ggml_status err = ggml_backend_sched_graph_compute_async(sched, graph);
+    ggml_backend_sched_synchronize(sched);
+    return err;
+}
 
-    if (!sched->is_reset) {
+enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+    if (!sched->is_reset && !sched->is_alloc) {
         ggml_backend_sched_reset(sched);
     }
 
-    ggml_backend_sched_split_graph(sched, graph);
-    if (!ggml_backend_sched_alloc_splits(sched)) {
-        return GGML_STATUS_ALLOC_FAILED;
+    if (!sched->is_alloc) {
+        if (!ggml_backend_sched_alloc_graph(sched, graph)) {
+            return GGML_STATUS_ALLOC_FAILED;
+        }
     }
 
     return ggml_backend_sched_compute_splits(sched);
 }
 
+void ggml_backend_sched_synchronize(ggml_backend_sched_t sched) {
+    for (int i = 0; i < sched->n_backends; i++) {
+        ggml_backend_synchronize(sched->backends[i]);
+    }
+}
+
 void ggml_backend_sched_set_eval_callback(ggml_backend_sched_t sched, ggml_backend_sched_eval_callback callback, void * user_data) {
     sched->callback_eval = callback;
     sched->callback_eval_user_data = user_data;
@@ -1606,19 +1824,24 @@ int ggml_backend_sched_get_n_splits(ggml_backend_sched_t sched) {
     return sched->n_splits;
 }
 
+int ggml_backend_sched_get_n_copies(ggml_backend_sched_t sched) {
+    return sched->n_copies;
+}
+
 size_t ggml_backend_sched_get_buffer_size(ggml_backend_sched_t sched, ggml_backend_t backend) {
     int backend_index = ggml_backend_sched_backend_id(sched, backend);
     GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
+
     return ggml_gallocr_get_buffer_size(sched->galloc, backend_index);
 }
 
-void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
+void ggml_backend_sched_set_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node, ggml_backend_t backend) {
     int backend_index = ggml_backend_sched_backend_id(sched, backend);
     GGML_ASSERT(backend_index >= 0 && backend_index < sched->n_backends);
     tensor_backend_id(node) = backend_index;
 }
 
-ggml_backend_t ggml_backend_sched_get_node_backend(ggml_backend_sched_t sched, struct ggml_tensor * node) {
+ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched, struct ggml_tensor * node) {
     int backend_index = tensor_backend_id(node);
     if (backend_index == -1) {
         return NULL;