ggml: avoid rebuild of GGML graph for each token (#7456)

Introduces caching of GGML graph to avoid unnecessary full rebuild
between each token. KV cache parameters, which change with each token,
are updated directly in cached GGML graph. Can be disabled with
GGML_DISABLE_GRAPH_CACHING environment variable.

Author: agray3

ggml/include/ggml-backend.h

@@ -230,6 +230,11 @@ extern "C" {
     GGML_API void ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
     GGML_API void ggml_backend_view_init(struct ggml_tensor * tensor);
 
+    // Utility to query whether cached GGML graph is in use
+    GGML_API bool ggml_use_cached_graph(ggml_backend_sched_t sched);
+
+    // Set whether or not to use GGML graph caching
+    GGML_API void ggml_set_cached_graph(ggml_backend_sched_t sched, bool set_value);
 
 #ifdef  __cplusplus
 }

ggml/src/ggml-backend.c

@@ -1036,6 +1036,13 @@ struct ggml_backend_sched_split {
     struct ggml_cgraph graph;
 };
 
+// Object to facilitate GML graph caching
+struct ggml_cached_graph {
+    bool is_active;
+    ggml_backend_t input_backend;
+    struct ggml_tensor * input_cpy[GGML_SCHED_MAX_SPLIT_INPUTS];
+};
+
 struct ggml_backend_sched {
     bool is_reset; // true if the scheduler has been reset since the last graph split
     bool is_alloc;
@@ -1087,6 +1094,8 @@ struct ggml_backend_sched {
     __attribute__((aligned(GGML_MEM_ALIGN)))
 #endif
     char context_buffer[GGML_SCHED_MAX_SPLITS*GGML_SCHED_MAX_SPLIT_INPUTS*2*sizeof(struct ggml_tensor) + sizeof(struct ggml_cgraph)];
+
+    struct ggml_cached_graph cached_graph;
 };
 
 #define hash_id(tensor) ggml_hash_find_or_insert(sched->hash_set, tensor)
@@ -1753,6 +1762,14 @@ static enum ggml_status ggml_backend_sched_compute_splits(ggml_backend_sched_t s
             struct ggml_tensor * input = split->inputs[j];
             struct ggml_tensor * input_cpy = sched->tensor_copies[hash_id(input)][split_backend_id][sched->cur_copy];
 
+            if (!sched->cached_graph.is_active) {
+                sched->cached_graph.input_backend = input_backend;
+                sched->cached_graph.input_cpy[j] = input_cpy;
+            }
+            else {
+                input_backend = sched->cached_graph.input_backend;
+                input_cpy = sched->cached_graph.input_cpy[j];
+            }
             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) {
@@ -1872,6 +1889,8 @@ ggml_backend_sched_t ggml_backend_sched_new(
 
     ggml_backend_sched_reset(sched);
 
+    sched->cached_graph.is_active = false;
+
     return sched;
 }
 
@@ -1947,6 +1966,9 @@ enum ggml_status ggml_backend_sched_graph_compute(ggml_backend_sched_t sched, st
 }
 
 enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sched, struct ggml_cgraph * graph) {
+
+    if(!sched->cached_graph.is_active)
+    {
     if (!sched->is_reset && !sched->is_alloc) {
         ggml_backend_sched_reset(sched);
     }
@@ -1956,7 +1978,7 @@ enum ggml_status ggml_backend_sched_graph_compute_async(ggml_backend_sched_t sch
             return GGML_STATUS_ALLOC_FAILED;
         }
     }
-
+    }
     return ggml_backend_sched_compute_splits(sched);
 }
 
@@ -2223,3 +2245,12 @@ bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t
 
     return true;
 }
+
+bool ggml_use_cached_graph(ggml_backend_sched_t sched) {
+    return sched->cached_graph.is_active;
+}
+
+void ggml_set_cached_graph(ggml_backend_sched_t sched, bool set_value) {
+    sched->cached_graph.is_active = set_value;
+}
+

src/llama.cpp

@@ -2712,6 +2712,16 @@ struct llama_model {
     }
 };
 
+// Object used to allow caching of GGML graph between tokens where possible.
+struct ggml_cached_graph {
+    ggml_cgraph * gf;
+    size_t n;
+    ggml_backend_t backend_res;
+    ggml_backend_t backend_embd;
+    struct ggml_tensor * res;
+    struct ggml_tensor * embd;
+};
+
 struct llama_context {
     llama_context(const llama_model & model) : model(model), t_start_us(model.t_start_us), t_load_us(model.t_load_us) {}
     ~llama_context() {
@@ -2813,6 +2823,8 @@ struct llama_context {
 
     // control vectors
     struct llama_control_vector cvec;
+
+    struct ggml_cached_graph cached_graph;
 };
 
 static size_t llama_get_device_count(const llama_model & model) {
@@ -14524,12 +14536,37 @@ static int llama_decode_internal(
         ggml_backend_sched_reset(lctx.sched);
         ggml_backend_sched_set_eval_callback(lctx.sched, lctx.cparams.cb_eval, lctx.cparams.cb_eval_user_data);
 
-        ggml_cgraph * gf = llama_build_graph(lctx, u_batch, false);
-
+        ggml_cgraph * gf;
         // the output is always the last tensor in the graph
-        struct ggml_tensor * res  = gf->nodes[gf->n_nodes - 1];
-        struct ggml_tensor * embd = gf->nodes[gf->n_nodes - 2];
+        struct ggml_tensor * res;
+        struct ggml_tensor * embd;
+
+        bool n_has_changed_since_last_token = false;
+        if(lctx.cached_graph.n != kv_self.n) n_has_changed_since_last_token = true;
+        lctx.cached_graph.n = kv_self.n;
+
+        // Re-build graph only if graph caching is not possible
+        if(!ggml_use_cached_graph(lctx.sched) || n_has_changed_since_last_token) {
+
+        gf = llama_build_graph(lctx, u_batch, false);
+
+        // disable future graph caching in presense of env var,
+        // if there are multiple devices, or if batch size is greater than 1
+        // TO DO enable graph caching for these cases
+        bool disable_cached_ggml_graph = (getenv("GGML_DISABLE_GRAPH_CACHING") != nullptr)
+            || (llama_get_device_count(model) > 1);
+        for (int i = 0 ; i < gf->n_nodes; i++) {
+            if (gf->nodes[i]->op == GGML_OP_ADD && gf->nodes[i]->src[1] && gf->nodes[i]->src[1]->ne[1] > 1) {
+                disable_cached_ggml_graph = true;
+                break;
+            }
+        }
+
+        if(!disable_cached_ggml_graph) ggml_set_cached_graph(lctx.sched,true);
 
+        // the output is always the last tensor in the graph
+        res  = gf->nodes[gf->n_nodes - 1];
+        embd = gf->nodes[gf->n_nodes - 2];
         if (lctx.n_outputs == 0) {
             // no output
             res  = nullptr;
@@ -14545,10 +14582,71 @@ static int llama_decode_internal(
             embd = nullptr; // do not extract embeddings when not needed
             GGML_ASSERT(strcmp(res->name, "result_output") == 0 && "missing result_output tensor");
         }
+        lctx.cached_graph.res = res;
+        lctx.cached_graph.embd = embd;
         // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
 
         ggml_backend_sched_alloc_graph(lctx.sched, gf);
 
+        }
+        else {
+            gf = lctx.cached_graph.gf;
+            res = lctx.cached_graph.res;
+            embd = lctx.cached_graph.embd;
+        }
+        lctx.cached_graph.gf = gf;
+
+        if(ggml_use_cached_graph(lctx.sched)) {
+
+            // If using flash attention, find mask node so it can be skipped when updating
+            // KV cache paramaters in cached graph nodes below
+            void * flash_attn_mask_node = nullptr;
+            if(cparams.flash_attn) {
+                for (int i = 0; i < gf->n_nodes; i++) {
+                    ggml_tensor * node = gf->nodes[i];
+                    if (node->op == GGML_OP_FLASH_ATTN_EXT) {
+                        flash_attn_mask_node = node->src[3];
+                        break;
+                    }
+                }
+            }
+
+            // Temporarily store KV cache parameters that will need updated in cached graph.
+            const struct llama_hparams & hparams = model.hparams;
+            const int64_t  n_layer = hparams.n_layer;
+            const int64_t kv_head = kv_self.head;
+            std::vector<void *> kv_cache_ptrs;
+            for (int il = 0; il < n_layer; ++il) {
+                const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa();
+                const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa();
+                ggml_tensor * tmp_tensor =  kv_self.k_l[il];
+                size_t tmp_offset = (ggml_row_size(kv_self.k_l[il]->type, n_embd_k_gqa))*kv_head;
+                kv_cache_ptrs.push_back(static_cast<char*>(tmp_tensor->data) + tmp_offset);
+                tmp_tensor = kv_self.v_l[il];
+                if (cparams.flash_attn) {
+                    tmp_offset = (kv_head)*ggml_row_size(kv_self.v_l[il]->type, n_embd_v_gqa);
+                } else {
+                    tmp_offset = (kv_head)*ggml_element_size(kv_self.v_l[il]);
+                }
+                kv_cache_ptrs.push_back(static_cast<char*>(tmp_tensor->data) + tmp_offset);
+            }
+
+            // Update KV cache parameters in cached graph.
+            int copy_op_count = 0;
+            if(gf != nullptr && gf->nodes != nullptr){
+                for (int i = 0; i < gf->n_nodes; i++) {
+                    ggml_tensor * node = gf->nodes[i];
+                    if (node->op == GGML_OP_CPY) {
+                        if (node != flash_attn_mask_node) {
+                            node->src[1]->data = kv_cache_ptrs[copy_op_count];
+                            copy_op_count++;
+                        }
+                    }
+                }
+            }
+
+        }
+
         llama_set_inputs(lctx, u_batch);
 
         llama_graph_compute(lctx, gf, n_threads);
@@ -14571,11 +14669,15 @@ static int llama_decode_internal(
         // extract logits
         if (res) {
             ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(lctx.sched, res);
-            GGML_ASSERT(backend_res != nullptr);
-            GGML_ASSERT(lctx.logits != nullptr);
-
             float * logits_out = lctx.logits + n_outputs_prev*n_vocab;
             const int32_t n_outputs_new = lctx.n_outputs;
+            if(!ggml_use_cached_graph(lctx.sched))
+                lctx.cached_graph.backend_res = backend_res;
+            else
+                backend_res = lctx.cached_graph.backend_res;
+
+            GGML_ASSERT(backend_res != nullptr);
+            GGML_ASSERT(lctx.logits != nullptr);
 
             if (n_outputs_new) {
                 GGML_ASSERT( n_outputs_prev + n_outputs_new <= n_outputs);
@@ -14587,6 +14689,12 @@ static int llama_decode_internal(
         // extract embeddings
         if (embd) {
             ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(lctx.sched, embd);
+
+
+            if(!ggml_use_cached_graph(lctx.sched))
+                lctx.cached_graph.backend_embd = backend_embd;
+            else
+                backend_embd = lctx.cached_graph.backend_embd;
             GGML_ASSERT(backend_embd != nullptr);
 
             switch (cparams.pooling_type) {