ggml : add context enumeration functions (#3605)

finetune : fix assert failure in ggml-alloc

examples/finetune/finetune.cpp

@@ -529,13 +529,14 @@ static void init_lora(const struct my_llama_model * model, struct my_llama_lora
     set_param_lora(lora);
 
     // measure data size
-    struct ggml_allocr * alloc = NULL;
-    alloc = ggml_allocr_new_measure(tensor_alignment);
-    alloc_lora(alloc, lora);
+    size_t size = 0;
+    for (struct ggml_tensor * t = ggml_get_first_tensor(ctx); t != NULL; t = ggml_get_next_tensor(ctx, t)) {
+        size += GGML_PAD(ggml_nbytes(t), tensor_alignment);
+    }
 
     // allocate data
-    lora->data.resize(ggml_allocr_max_size(alloc) + tensor_alignment);
-    ggml_allocr_free(alloc);
+    struct ggml_allocr * alloc = NULL;
+    lora->data.resize(size + tensor_alignment);
     alloc = ggml_allocr_new(lora->data.data(), lora->data.size(), tensor_alignment);
     alloc_lora(alloc, lora);
     ggml_allocr_free(alloc);
@@ -1714,11 +1715,9 @@ int main(int argc, char ** argv) {
     struct ggml_tensor * target_probs  = ggml_new_tensor_3d(ctx_input, GGML_TYPE_F32, n_vocab,  n_tokens, n_batch);
 
     // measure required memory for input tensors
-    alloc = ggml_allocr_new_measure(tensor_alignment);
-    ggml_allocr_alloc(alloc, tokens_input);
-    ggml_allocr_alloc(alloc, target_probs);
-    size_t max_input_size = ggml_allocr_max_size(alloc) + tensor_alignment;
-    ggml_allocr_free(alloc);
+    size_t max_input_size = GGML_PAD(ggml_nbytes(tokens_input), tensor_alignment) +
+                            GGML_PAD(ggml_nbytes(target_probs), tensor_alignment) +
+                            tensor_alignment;
     printf("%s: input_size = %zu bytes (%.1f MB)\n", __func__, max_input_size, (float) max_input_size / (1024.0f*1024.0f));
 
     // allocate input tensors

ggml.c

@@ -5494,6 +5494,39 @@ struct ggml_tensor * ggml_view_tensor(
     return result;
 }
 
+struct ggml_tensor * ggml_get_first_tensor(struct ggml_context * ctx) {
+    struct ggml_object * obj = ctx->objects_begin;
+
+    char * const mem_buffer = ctx->mem_buffer;
+
+    while (obj != NULL) {
+        if (obj->type == GGML_OBJECT_TENSOR) {
+            return (struct ggml_tensor *)(mem_buffer + obj->offs);
+        }
+
+        obj = obj->next;
+    }
+
+    return NULL;
+}
+
+struct ggml_tensor * ggml_get_next_tensor(struct ggml_context * ctx, struct ggml_tensor * tensor) {
+    struct ggml_object * obj = (struct ggml_object *) ((char *)tensor - GGML_OBJECT_SIZE);
+    obj = obj->next;
+
+    char * const mem_buffer = ctx->mem_buffer;
+
+    while (obj != NULL) {
+        if (obj->type == GGML_OBJECT_TENSOR) {
+            return (struct ggml_tensor *)(mem_buffer + obj->offs);
+        }
+
+        obj = obj->next;
+    }
+
+    return NULL;
+}
+
 struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name) {
     struct ggml_object * obj = ctx->objects_begin;
 
@@ -8647,6 +8680,7 @@ void ggml_set_param(
 
     GGML_ASSERT(tensor->grad == NULL);
     tensor->grad = ggml_dup_tensor(ctx, tensor);
+    ggml_format_name(tensor->grad, "%s (grad)", tensor->name);
 }
 
 // ggml_compute_forward_dup

ggml.h

@@ -704,6 +704,9 @@ extern "C" {
     GGML_API struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
     GGML_API struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, struct ggml_tensor * src);
 
+    // Context tensor enumeration and lookup
+    GGML_API struct ggml_tensor * ggml_get_first_tensor(struct ggml_context * ctx);
+    GGML_API struct ggml_tensor * ggml_get_next_tensor (struct ggml_context * ctx, struct ggml_tensor * tensor);
     GGML_API struct ggml_tensor * ggml_get_tensor(struct ggml_context * ctx, const char * name);
 
     GGML_API struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);