ggml : change ggml_graph_compute() API to not require context (#1999)

* ggml_graph_compute: deprecate using ggml_context, try resolve issue #287

* rewrite: no longer consider backward compitability; plan and make_plan

* minor: rename ctx as plan; const

* remove ggml_graph_compute from tests/test-grad0.c, but current change breaks backward

* add static ggml_graph_compute_sugar()

* minor: update comments

* reusable buffers

* ggml : more consistent naming + metal fixes

* ggml : fix docs

* tests : disable grad / opt + minor naming changes

* ggml : add ggml_graph_compute_with_ctx()

- backwards compatible API
- deduplicates a lot of copy-paste

* ci : enable test-grad0

* examples : factor out plan allocation into a helper function

* llama : factor out plan stuff into a helper function

* ci : fix env

* llama : fix duplicate symbols + refactor example benchmark

* ggml : remove obsolete assert + refactor n_tasks section

* ggml : fix indentation in switch

* llama : avoid unnecessary bool

* ggml : remove comments from source file and match order in header

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

1 files changed, 39 insertions, 15 deletions

diff --git a/llama.cpp b/llama.cpp
index 02afdeb..ee6ec09 100644
--- a/llama.cpp
+++ b/llama.cpp
@@ -79,6 +79,25 @@ void llama_nop(struct ggml_tensor * tensor) { // don't offload by default
     (void) tensor;
 }
 
+//
+// ggml helpers
+//
+
+static void ggml_graph_compute_helper(std::vector<uint8_t> & buf, ggml_cgraph * graph, int n_threads) {
+    struct ggml_cplan plan = ggml_graph_plan(graph, n_threads);
+
+    if (plan.work_size > 0) {
+        buf.resize(plan.work_size);
+        plan.work_data = buf.data();
+    }
+
+    ggml_graph_compute(graph, &plan);
+}
+
+//
+// memory sizes
+//
+
 static const std::map<e_model, size_t> & MEM_REQ_SCRATCH0()
 {
     static std::map<e_model, size_t> k_sizes = {
@@ -321,6 +340,9 @@ struct llama_context {
     // input embedding (1-dimensional array: [n_embd])
     std::vector<float> embedding;
 
+    // reusable buffer for `struct ggml_graph_plan.work_data`
+    std::vector<uint8_t> work_buffer;
+
     // memory buffers used to evaluate the model
     // TODO: move in llama_state
     llama_ctx_buffer buf_compute;
@@ -758,7 +780,6 @@ struct llama_model_loader {
 
 };
 
-
 //
 // kv cache
 //
@@ -1265,7 +1286,7 @@ static bool llama_eval_internal(
            const float * embd,
              const int   n_tokens,
              const int   n_past,
-             const int   n_threads,
+                   int   n_threads,
             const char * cgraph_fname) {
 
     LLAMA_ASSERT((!tokens && embd) || (tokens && !embd));
@@ -1306,10 +1327,11 @@ static bool llama_eval_internal(
 
     struct ggml_context * ctx0 = ggml_init(params);
 
+    ggml_cgraph gf = {};
+
     // for big prompts, if BLAS is enabled, it is better to use only one thread
     // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance
-    ggml_cgraph gf = {};
-    gf.n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas() ? 1 : n_threads;
+    n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_gpublas() ? 1 : n_threads;
 
     struct ggml_tensor * cur;
     struct ggml_tensor * inpL;
@@ -1593,6 +1615,7 @@ static bool llama_eval_internal(
 
 #ifdef GGML_USE_METAL
     if (lctx.ctx_metal && N == 1) {
+        ggml_metal_set_n_cb     (lctx.ctx_metal, n_threads);
         ggml_metal_graph_compute(lctx.ctx_metal, &gf);
         ggml_metal_get_tensor   (lctx.ctx_metal, cur);
     } else {
@@ -1612,10 +1635,10 @@ static bool llama_eval_internal(
             ggml_metal_get_tensor(lctx.ctx_metal, kv_self.v);
         }
 
-        ggml_graph_compute(ctx0, &gf);
+        ggml_graph_compute_helper(lctx.work_buffer, &gf, n_threads);
     }
 #else
-    ggml_graph_compute(ctx0, &gf);
+    ggml_graph_compute_helper(lctx.work_buffer, &gf, n_threads);
 #endif
 
     if (cgraph_fname) {
@@ -2575,8 +2598,8 @@ void llama_free_model(struct llama_model * model) {
 }
 
 struct llama_context * llama_new_context_with_model(
-                             struct llama_model * model,
-            struct llama_context_params   params) {
+                 struct llama_model * model,
+        struct llama_context_params   params) {
 
     if (!model) {
         return nullptr;
@@ -2645,7 +2668,7 @@ struct llama_context * llama_new_context_with_model(
 #ifdef GGML_USE_METAL
     if (params.n_gpu_layers > 0) {
         // this allocates all Metal resources and memory buffers
-        ctx->ctx_metal = ggml_metal_init();
+        ctx->ctx_metal = ggml_metal_init(1);
 
         void * data_ptr  = NULL;
         size_t data_size = 0;
@@ -2802,6 +2825,9 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
     // read tensors and apply
     bool warned = false;
     int n_tensors = 0;
+
+    std::vector<uint8_t> work_buffer;
+
     while (true) {
         int32_t n_dims;
         int32_t length;
@@ -2966,8 +2992,8 @@ int llama_apply_lora_from_file_internal(const struct llama_model & model, const
             }
 
             struct ggml_cgraph gf = ggml_build_forward(r);
-            gf.n_threads = n_threads;
-            ggml_graph_compute(lora_ctx, &gf);
+
+            ggml_graph_compute_helper(work_buffer, &gf, n_threads);
 
             // we won't need these tensors again, reset the context to save memory
             ggml_free(lora_ctx);
@@ -3120,7 +3146,6 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
 
             ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true });
             ggml_cgraph gf{};
-            gf.n_threads = 1;
 
             ggml_tensor * kout3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_ntok, n_layer);
             kout3d->data = out;
@@ -3140,7 +3165,7 @@ size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dst) {
 
             ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, k3d, kout3d));
             ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, v3d, vout3d));
-            ggml_graph_compute(cpy_ctx, &gf);
+            ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1);
 
             ggml_free(cpy_ctx);
         }
@@ -3226,7 +3251,6 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
 
             ggml_context * cpy_ctx = ggml_init({ 4096, NULL, /* no_alloc */ true });
             ggml_cgraph gf{};
-            gf.n_threads = 1;
 
             ggml_tensor * kin3d = ggml_new_tensor_3d(cpy_ctx, kv_self.k->type, n_embd, kv_ntok, n_layer);
             kin3d->data = (void *) inp;
@@ -3246,7 +3270,7 @@ size_t llama_set_state_data(struct llama_context * ctx, uint8_t * src) {
 
             ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, kin3d, k3d));
             ggml_build_forward_expand(&gf, ggml_cpy(cpy_ctx, vin3d, v3d));
-            ggml_graph_compute(cpy_ctx, &gf);
+            ggml_graph_compute_helper(ctx->work_buffer, &gf, /*n_threads*/ 1);
 
             ggml_free(cpy_ctx);
         }