metal : minimize view overlap to try to utilize device memory better

Makefile

@@ -252,7 +252,7 @@ $(info )
 ggml.o: ggml.c ggml.h ggml-cuda.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-llama.o: llama.cpp ggml.h ggml-cuda.h llama.h llama-util.h
+llama.o: llama.cpp ggml.h ggml-cuda.h ggml-metal.h llama.h llama-util.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 common.o: examples/common.cpp examples/common.h

ggml-metal.h

@@ -41,12 +41,15 @@ void ggml_metal_free(struct ggml_metal_context * ctx);
 // - make sure to map all buffers used in the graph before calling ggml_metal_graph_compute
 // - the mapping is used during computation to determine the arguments of the compute kernels
 // - you don't need to keep the host memory buffer allocated as it is never accessed by Metal
+// - max_size specifies the maximum size of a tensor and is used to create shared views such
+//   that it is guaranteed that the tensor will fit in at least one of the views
 //
 bool ggml_metal_add_buffer(
         struct ggml_metal_context * ctx,
                        const char * name,
                              void * data,
-                           size_t   size);
+                           size_t   size,
+                           size_t   max_size);
 
 // set data from host memory into the device
 void ggml_metal_set_tensor(struct ggml_metal_context * ctx, struct ggml_tensor * t);

ggml-metal.m

@@ -240,14 +240,14 @@ bool ggml_metal_add_buffer(
 
     if (data) {
         // verify that the buffer does not overlap with any of the existing buffers
-        //for (int i = 0; i < ctx->n_buffers; ++i) {
-        //    const int64_t ioffs = (int64_t) data - (int64_t) ctx->buffers[i].data;
+        for (int i = 0; i < ctx->n_buffers; ++i) {
+            const int64_t ioffs = (int64_t) data - (int64_t) ctx->buffers[i].data;
 
-        //    if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
-        //        fprintf(stderr, "%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
-        //        return false;
-        //    }
-        //}
+            if (ioffs >= 0 && ioffs < (int64_t) ctx->buffers[i].size) {
+                fprintf(stderr, "%s: error: buffer '%s' overlaps with '%s'\n", __func__, name, ctx->buffers[i].name);
+                return false;
+            }
+        }
 
         const size_t size_page = getpagesize();
 
@@ -273,7 +273,9 @@ bool ggml_metal_add_buffer(
 
             ++ctx->n_buffers;
         } else {
-            const size_t size_ovlp = (max_size + size_page - 1) / size_page * size_page;
+            // this overlap between the views will guarantee that the tensor with the maximum size will fully fit into
+            // one of the views
+            const size_t size_ovlp = ((max_size + size_page - 1) / size_page + 1) * size_page; // round-up 2 pages just in case
             const size_t size_step = ctx->device.maxBufferLength - size_ovlp;
             const size_t size_view = ctx->device.maxBufferLength;
 

ggml.c

@@ -4154,14 +4154,34 @@ void ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc) {
     ctx->no_alloc = no_alloc;
 }
 
-void * ggml_get_mem_buffer(struct ggml_context * ctx) {
+void * ggml_get_mem_buffer(const struct ggml_context * ctx) {
     return ctx->mem_buffer;
 }
 
-size_t ggml_get_mem_size(struct ggml_context * ctx) {
+size_t ggml_get_mem_size(const struct ggml_context * ctx) {
     return ctx->mem_size;
 }
 
+size_t ggml_get_max_tensor_size(const struct ggml_context * ctx) {
+    size_t max_size = 0;
+
+    struct ggml_object * obj = ctx->objects_begin;
+
+    while (obj != NULL) {
+        struct ggml_tensor * tensor = (struct ggml_tensor *) ((char *) ctx->mem_buffer + obj->offs);
+
+        const size_t size = ggml_nbytes(tensor);
+
+        if (max_size < size) {
+            max_size = size;
+        }
+
+        obj = obj->next;
+    }
+
+    return max_size;
+}
+
 // IMPORTANT:
 // when creating "opt" tensors, always save and load the scratch buffer
 // this is an error prone process, but it is necessary to support inplace

ggml.h

@@ -500,8 +500,9 @@ extern "C" {
     GGML_API size_t  ggml_set_scratch (struct ggml_context * ctx, struct ggml_scratch scratch);
     GGML_API void    ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc);
 
-    GGML_API void *  ggml_get_mem_buffer(struct ggml_context * ctx);
-    GGML_API size_t  ggml_get_mem_size  (struct ggml_context * ctx);
+    GGML_API void *  ggml_get_mem_buffer     (const struct ggml_context * ctx);
+    GGML_API size_t  ggml_get_mem_size       (const struct ggml_context * ctx);
+    GGML_API size_t  ggml_get_max_tensor_size(const struct ggml_context * ctx);
 
     GGML_API struct ggml_tensor * ggml_new_tensor(
             struct ggml_context * ctx,

llama.cpp

@@ -2698,6 +2698,7 @@ struct llama_context * llama_init_from_file(
 
         void * data_ptr  = NULL;
         size_t data_size = 0;
+
         if (params.use_mmap) {
             data_ptr  = ctx->model.mapping->addr;
             data_size = ctx->model.mapping->size;
@@ -2706,6 +2707,10 @@ struct llama_context * llama_init_from_file(
             data_size = ggml_get_mem_size  (ctx->model.ctx);
         }
 
+        const size_t max_size = ggml_get_max_tensor_size(ctx->model.ctx);
+
+        printf("%s: max tensor size = %8.2f MB\n", __func__, max_size/1024.0/1024.0);
+
 #define LLAMA_METAL_CHECK_BUF(result)                                          \
     if (!(result)) {                                                           \
         fprintf(stderr, "%s: failed to add buffer\n", __func__);               \
@@ -2713,12 +2718,13 @@ struct llama_context * llama_init_from_file(
         return NULL;                                                           \
     }
 
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size));
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.addr, ctx->buf_compute.size));
+        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "data", data_ptr, data_size, max_size));
 
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv",   ctx->model.kv_self.buf.addr, ctx->model.kv_self.buf.size));
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr0", ctx->buf_scratch[0].addr,    ctx->buf_scratch[0].size));
-        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr1", ctx->buf_scratch[1].addr,    ctx->buf_scratch[1].size));
+        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "eval", ctx->buf_compute.addr,       ctx->buf_compute.size,       0));
+        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "kv",   ctx->model.kv_self.buf.addr, ctx->model.kv_self.buf.size, 0));
+
+        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr0", ctx->buf_scratch[0].addr, ctx->buf_scratch[0].size, 0));
+        LLAMA_METAL_CHECK_BUF(ggml_metal_add_buffer(ctx->ctx_metal, "scr1", ctx->buf_scratch[1].addr, ctx->buf_scratch[1].size, 0));
 #undef LLAMA_METAL_CHECK_BUF
     }
 #endif