From 497f0b0cdb3c85d503685bcb50a39f3958773175 Mon Sep 17 00:00:00 2001 From: ylsun <1654366371@qq.com> Date: Wed, 9 Apr 2025 19:15:49 +0800 Subject: [PATCH 1/2] update --- src/ggml-opt.cpp | 22 +++++++++++++++------- src/ggml.c | 17 +++++++++-------- 2 files changed, 24 insertions(+), 15 deletions(-) diff --git a/src/ggml-opt.cpp b/src/ggml-opt.cpp index 7c3e24103..12afbab7e 100644 --- a/src/ggml-opt.cpp +++ b/src/ggml-opt.cpp @@ -14,10 +14,10 @@ #include struct ggml_opt_dataset { - struct ggml_context * ctx = nullptr; - ggml_backend_buffer_t buf = nullptr; - struct ggml_tensor * data = nullptr; - struct ggml_tensor * labels = nullptr; + struct ggml_context * ctx = nullptr; //A memory context used to manage the tensors (data and labels) associated with the dataset. + ggml_backend_buffer_t buf = nullptr; // A backend buffer used to store the dataset's tensors in a specific memory backend (e.g., CPU or GPU) + struct ggml_tensor * data = nullptr; // A tensor that holds the input data for the dataset + struct ggml_tensor * labels = nullptr; // A tensor that holds the labels for the dataset int64_t ndata = -1; int64_t ndata_shard = -1; @@ -72,7 +72,15 @@ struct ggml_opt_result { }; // ====== Dataset ====== - +/** + * @brief a function to initialize a dataset + * + * @param ne_datapoint // The number of features (or dimensions) in each data point. + * @param ne_label //The number of labels associated with each data point. + * @param ndata // The total number of data points in the dataset + * @param ndata_shard The number of data points in each shard. Shards are used to divide the dataset into smaller chunks for distributed or batched processing. + * @return ggml_opt_dataset_t + */ ggml_opt_dataset_t ggml_opt_dataset_init(int64_t ne_datapoint, int64_t ne_label, int64_t ndata, int64_t ndata_shard) { GGML_ASSERT(ne_datapoint > 0); GGML_ASSERT(ne_label >= 0); @@ -85,11 +93,11 @@ ggml_opt_dataset_t ggml_opt_dataset_init(int64_t ne_datapoint, int64_t ne_label, { struct ggml_init_params params = { - /*.mem_size =*/ 2*ggml_tensor_overhead(), + /*.mem_size =*/ 2*ggml_tensor_overhead(), //data tensor and labels tensor /*.mem_buffer =*/ nullptr, /*.no_alloc =*/ true, }; - result->ctx = ggml_init(params); + result->ctx = ggml_init(params);//the brace hope params will be destructed immediately } result->data = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_datapoint, ndata); diff --git a/src/ggml.c b/src/ggml.c index 161dd3fa9..c6e90d572 100644 --- a/src/ggml.c +++ b/src/ggml.c @@ -880,15 +880,15 @@ static const size_t GGML_OBJECT_SIZE = sizeof(struct ggml_object); // struct ggml_context { - size_t mem_size; - void * mem_buffer; - bool mem_buffer_owned; - bool no_alloc; + size_t mem_size;// Total size of the memory buffer allocated for this context. + void * mem_buffer;// Total size of the memory buffer allocated for this context. + bool mem_buffer_owned;// Indicates if the memory buffer is owned by this context or not. + bool no_alloc;// Indicates if the context should allocate memory or not. - int n_objects; + int n_objects;//Number of objects (e.g., tensors) currently allocated in this context. - struct ggml_object * objects_begin; - struct ggml_object * objects_end; + struct ggml_object * objects_begin;//pointer to the beginning of the list of objects + struct ggml_object * objects_end;//pointer to the end of the list of objects }; struct ggml_context_container { @@ -1297,7 +1297,8 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) { } size_t ggml_tensor_overhead(void) { - return GGML_OBJECT_SIZE + GGML_TENSOR_SIZE; + //tensor overhead is the size of tensor struct + size of object struct， ggml_obj is used to describe the tensor/graph/buffer + return GGML_OBJECT_SIZE + GGML_TENSOR_SIZE; } bool ggml_is_transposed(const struct ggml_tensor * tensor) { From a056547aa1d26b7a77eddb0a08394c64577c4393 Mon Sep 17 00:00:00 2001 From: ylsun <1654366371@qq.com> Date: Sun, 13 Apr 2025 22:53:48 +0800 Subject: [PATCH 2/2] update --- .gitignore | 2 +- src/ggml-opt.cpp | 6 +++--- src/ggml.c | 20 ++++++++++---------- 3 files changed, 14 insertions(+), 14 deletions(-) diff --git a/.gitignore b/.gitignore index 4bf0fa095..6c5e3421b 100644 --- a/.gitignore +++ b/.gitignore @@ -10,7 +10,7 @@ CMakeSettings.json .vs/ .vscode/ .clangd - +.llama_env/ .venv/ ggml_env/ .exrc diff --git a/src/ggml-opt.cpp b/src/ggml-opt.cpp index 12afbab7e..96f06d3f6 100644 --- a/src/ggml-opt.cpp +++ b/src/ggml-opt.cpp @@ -93,14 +93,14 @@ ggml_opt_dataset_t ggml_opt_dataset_init(int64_t ne_datapoint, int64_t ne_label, { struct ggml_init_params params = { - /*.mem_size =*/ 2*ggml_tensor_overhead(), //data tensor and labels tensor + /*.mem_size =*/ 2*ggml_tensor_overhead(), //data tensor and labels tensor itself, not include the tensors' data /*.mem_buffer =*/ nullptr, /*.no_alloc =*/ true, }; result->ctx = ggml_init(params);//the brace hope params will be destructed immediately } - - result->data = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_datapoint, ndata); + // after init, the memory of ctx is allocated, the tensors memory is also allocated, but the tensors' data is not allocated yet + result->data = ggml_new_tensor_2d(result->ctx, GGML_TYPE_F32, ne_datapoint, ndata);//get data tensor memory from context, set tensor attributes result->nbs_data = ggml_nbytes(result->data) * ndata_shard/ndata; if (ne_label > 0) { diff --git a/src/ggml.c b/src/ggml.c index c6e90d572..e722bd4a1 100644 --- a/src/ggml.c +++ b/src/ggml.c @@ -1419,7 +1419,7 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { if (is_first_call) { // initialize time system (required on Windows) ggml_time_init(); - + //if is first called, calculate the mapping from fp16 to f32 and store it in the table for (int i = 0; i < (1 << 16); ++i) { union { uint16_t u16; @@ -1433,7 +1433,7 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { ggml_critical_section_end(); - struct ggml_context * ctx = GGML_MALLOC(sizeof(struct ggml_context)); + struct ggml_context * ctx = GGML_MALLOC(sizeof(struct ggml_context));//malloc for the context struct itself // allow to call ggml_init with 0 size if (params.mem_size == 0) { @@ -1441,7 +1441,7 @@ struct ggml_context * ggml_init(struct ggml_init_params params) { } const size_t mem_size = params.mem_buffer ? params.mem_size : GGML_PAD(params.mem_size, GGML_MEM_ALIGN); - + //allow initialize a context with a extern buffer, if no buffer is provided, malloc a buffer *ctx = (struct ggml_context) { /*.mem_size =*/ mem_size, /*.mem_buffer =*/ params.mem_buffer ? params.mem_buffer : ggml_aligned_malloc(mem_size), @@ -1520,7 +1520,7 @@ static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml // always insert objects at the end of the context's memory pool struct ggml_object * obj_cur = ctx->objects_end; - const size_t cur_offs = obj_cur == NULL ? 0 : obj_cur->offs; + const size_t cur_offs = obj_cur == NULL ? 0 : obj_cur->offs; // offs is the start, [cur_offs, cur_offs + cur_size) is the lastest object const size_t cur_size = obj_cur == NULL ? 0 : obj_cur->size; const size_t cur_end = cur_offs + cur_size; @@ -1529,7 +1529,7 @@ static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml char * const mem_buffer = ctx->mem_buffer; struct ggml_object * const obj_new = (struct ggml_object *)(mem_buffer + cur_end); - + //[cur_end, cur_end + GGML_OBJECT_SIZE) is the ggml_object( just a descriptor of the object ) if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) { GGML_LOG_WARN("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n", __func__, cur_end + size_needed + GGML_OBJECT_SIZE, ctx->mem_size); @@ -1538,14 +1538,14 @@ static struct ggml_object * ggml_new_object(struct ggml_context * ctx, enum ggml #endif return NULL; } - + // just assign value for the new object, the memory is prepared yet *obj_new = (struct ggml_object) { .offs = cur_end + GGML_OBJECT_SIZE, .size = size_needed, .next = NULL, .type = type, }; - + //[cur_end + GGML_OBJECT_SIZE, cur_end + GGML_OBJECT_SIZE + size_needed) is the actual memory of the object GGML_ASSERT_ALIGNED(mem_buffer + obj_new->offs); if (obj_cur != NULL) { @@ -1579,14 +1579,14 @@ static struct ggml_tensor * ggml_new_tensor_impl( view_src = view_src->view_src; } - size_t data_size = ggml_row_size(type, ne[0]); - for (int i = 1; i < n_dims; i++) { + size_t data_size = ggml_row_size(type, ne[0]);//ne[0] is the number of elements in the first dimension + for (int i = 1; i < n_dims; i++) { // first dimension multiple by each other dimension data_size *= ne[i]; } GGML_ASSERT(view_src == NULL || data_size == 0 || data_size + view_offs <= ggml_nbytes(view_src)); - void * data = view_src != NULL ? view_src->data : NULL; + void * data = view_src != NULL ? view_src->data : NULL; // handle virw_src, which is used to view another tensor if (data != NULL) { data = (char *) data + view_offs; }