backend : offload large batches to GPU

ggerganov · Mar 15, 2024 · 19add48 · 19add48
1 parent 46acb36
commit 19add48
Show file tree

Hide file tree

Showing 8 changed files with 162 additions and 191 deletions.
diff --git a/ggml-alloc.c b/ggml-alloc.c
@@ -548,7 +548,11 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
 
-        if (ggml_is_view(node)) {
+        // TODO: better way to add external dependencies
+        // GGML_OP_NONE does not appear normally in the graph nodes, but is used by ggml-backend to add dependencies to
+        // control when some tensors are allocated and freed. in this case, the dependencies are in `src`, but the node
+        // itself is never used and should not be considered a dependency
+        if (ggml_is_view(node) && node->op != GGML_OP_NONE) {
             struct ggml_tensor * view_src = node->view_src;
             ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
         }
@@ -565,8 +569,8 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
 
             ggml_gallocr_hash_get(galloc, src)->n_children += 1;
 
-            // allocate explicit inputs and leafs
-            if (src->flags & GGML_TENSOR_FLAG_INPUT || src->op == GGML_OP_NONE) {
+            // allocate explicit inputs
+            if (src->flags & GGML_TENSOR_FLAG_INPUT) {
                 ggml_gallocr_allocate_node(galloc, src, get_node_buffer_id(node_buffer_ids, i));
             }
         }

diff --git a/ggml-backend-impl.h b/ggml-backend-impl.h
@@ -103,6 +103,11 @@ extern "C" {
         // check if the backend supports an operation
         bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
 
+        // check if the backend want to run an operation, even if the weights are allocated in a CPU buffer
+        // these should be expensive operations with large batch sizes that may benefit from running on this backend
+        // even if the weight has to be copied from the CPU temporarily
+        bool (*GGML_CALL offload_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+
         // (optional) event synchronization
         ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
         void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);

diff --git a/ggml-backend.c b/ggml-backend.c
@@ -278,14 +278,21 @@ enum ggml_status ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_
     return err;
 }
 
-bool ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     return backend->iface.graph_compute(backend, cgraph);
 }
 
 bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
     return backend->iface.supports_op(backend, op);
 }
 
+bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op) {
+    if (backend->iface.offload_op != NULL) {
+        return backend->iface.offload_op(backend, op);
+    }
+    return false;
+}
+
 // backend copy
 
 static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
@@ -834,6 +841,7 @@ 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,
+    /* .offload_op              = */ NULL,
     /* .event_new               = */ NULL,
     /* .event_free              = */ NULL,
     /* .event_record            = */ NULL,
@@ -999,7 +1007,7 @@ static bool ggml_is_view_op(enum ggml_op op) {
 #endif
 
 #ifndef GGML_SCHED_MAX_SPLITS
-#define GGML_SCHED_MAX_SPLITS 256
+#define GGML_SCHED_MAX_SPLITS 1024
 #endif
 
 #ifndef GGML_SCHED_MAX_SPLIT_INPUTS
@@ -1138,14 +1146,23 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
     }
 
     // assign nodes that use weights to the backend of the weights
+    // operations with weights are preferably run on the same backend as the weights
     for (int i = 0; i < GGML_MAX_SRC; i++) {
         const struct ggml_tensor * src = tensor->src[i];
         if (src == NULL) {
             continue;
         }
         if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
             int src_backend = ggml_backend_sched_backend_from_buffer(sched, src);
-            // operations with weights are always run on the same backend as the weights
+            // check if a backend with higher prio wants to run the op
+            if (src_backend == sched->n_backends - 1) {
+                for (int b = 0; b < src_backend; b++) {
+                    if (ggml_backend_offload_op(sched->backends[b], tensor)) {
+                        SET_CAUSE(tensor, "1.off");
+                        return b;
+                    }
+                }
+            }
             SET_CAUSE(tensor, "1.wgt%d", i);
             return src_backend;
         }
@@ -1404,7 +1421,25 @@ static void ggml_backend_sched_split_graph(ggml_backend_sched_t sched, struct gg
 
             GGML_ASSERT(tensor_backend_id != -1); // all nodes should be assigned by now
 
-            if (tensor_backend_id != cur_backend_id) {
+            // check if a weight is on a different backend and start a new split if so
+            bool offload = false;
+            if (tensor_backend_id == cur_backend_id && sched->splits[cur_split].n_inputs > 0) {
+                for (int j = 0; j < GGML_MAX_SRC; j++) {
+                    struct ggml_tensor * src = node->src[j];
+                    if (src == NULL) {
+                        continue;
+                    }
+                    if (src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
+                        int src_backend_id = tensor_backend_id(src);
+                        if (src_backend_id != -1 && src_backend_id != cur_backend_id) {
+                            offload = true;
+                            break;
+                        }
+                    }
+                }
+            }
+
+            if (tensor_backend_id != cur_backend_id || offload) {
                 sched->splits[cur_split].i_end = i;
                 cur_split++;
                 GGML_ASSERT(cur_split < GGML_SCHED_MAX_SPLITS);

diff --git a/ggml-backend.h b/ggml-backend.h
@@ -70,11 +70,11 @@ extern "C" {
     GGML_API ggml_backend_graph_plan_t ggml_backend_graph_plan_create(ggml_backend_t backend, struct ggml_cgraph * cgraph);
     GGML_API void                      ggml_backend_graph_plan_free  (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
 
-    GGML_API enum ggml_status ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API enum ggml_status ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
-
-    GGML_API bool ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API enum ggml_status ggml_backend_graph_plan_compute (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+    GGML_API enum ggml_status ggml_backend_graph_compute      (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API enum ggml_status ggml_backend_graph_compute_async(ggml_backend_t backend, struct ggml_cgraph * cgraph);
     GGML_API bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op);
+    GGML_API bool ggml_backend_offload_op(ggml_backend_t backend, const struct ggml_tensor * op);
 
     // tensor copy between different backends
     GGML_API void ggml_backend_tensor_copy(struct ggml_tensor * src, struct ggml_tensor * dst);