root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-28 12:24:35 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

gguf : do not support passing existing ggml_context to gguf_init

Browse Source
This commit is contained in:
Georgi Gerganov 2023-07-26 18:48:52 +03:00
parent 860c9c63ce
commit cb871fa022
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735
2 changed files with 18 additions and 38 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

29
ggml.c
View File

@ -18388,6 +18388,8 @@ static bool gguf_fread_str(void * dst, FILE * file, size_t * offset) {
} }
struct gguf_context * gguf_init(const char * fname, struct gguf_init_params params) { struct gguf_context * gguf_init(const char * fname, struct gguf_init_params params) {
GGML_ASSERT(!params.load || params.malloc || params.ctx != NULL);
FILE * file = fopen(fname, "rb"); FILE * file = fopen(fname, "rb");
if (!file) { if (!file) {
return NULL; return NULL;
@ -18518,8 +18520,7 @@ struct gguf_context * gguf_init(const char * fname, struct gguf_init_params para
const size_t size_cur = (ne*ggml_type_size(info->type))/ggml_blck_size(info->type); const size_t size_cur = (ne*ggml_type_size(info->type))/ggml_blck_size(info->type);
// TODO: pad size_cur to alignment ctx->size_data += GGML_PAD(size_cur, ctx->alignment);
ctx->size_data += size_cur;
} }
// TODO: simplify // TODO: simplify
@ -18528,11 +18529,7 @@ struct gguf_context * gguf_init(const char * fname, struct gguf_init_params para
ctx->data = GGML_ALIGNED_MALLOC(ctx->size_data); ctx->data = GGML_ALIGNED_MALLOC(ctx->size_data);
fseek(file, ctx->offset, SEEK_SET); fseek(file, ctx->offset, SEEK_SET);
ok = ok && gguf_fread_el(ctx->data, ctx->size_data, file, &offset); ok = ok && gguf_fread_el(ctx->data, ctx->size_data, file, &offset);
} else if (params.ctx != NULL) { } else {
bool ctx_new = false;
bool ctx_no_alloc = false;
if (*params.ctx == NULL) {
const size_t mem_size = const size_t mem_size =
ctx->header.n_tensors*ggml_tensor_overhead() + 1 + ctx->header.n_tensors*ggml_tensor_overhead() + 1 +
ctx->size_data; ctx->size_data;
@ -18545,12 +18542,6 @@ struct gguf_context * gguf_init(const char * fname, struct gguf_init_params para
*params.ctx = ggml_init(pdata); *params.ctx = ggml_init(pdata);
ctx_new = true;
} else {
ctx_no_alloc = ggml_get_no_alloc(*params.ctx);
ggml_set_no_alloc(*params.ctx, false);
}
struct ggml_context * ctx_data = *params.ctx; struct ggml_context * ctx_data = *params.ctx;
struct ggml_tensor * data = ggml_new_tensor_1d(ctx_data, GGML_TYPE_I8, ctx->size_data); struct ggml_tensor * data = ggml_new_tensor_1d(ctx_data, GGML_TYPE_I8, ctx->size_data);
@ -18561,11 +18552,7 @@ struct gguf_context * gguf_init(const char * fname, struct gguf_init_params para
if (!ok) { if (!ok) {
fprintf(stderr, "%s: failed to read tensor data\n", __func__); fprintf(stderr, "%s: failed to read tensor data\n", __func__);
fclose(file); fclose(file);
if (ctx_new) {
ggml_free(ctx_data); ggml_free(ctx_data);
} else {
ggml_set_no_alloc(ctx_data, ctx_no_alloc);
}
gguf_free(ctx); gguf_free(ctx);
return NULL; return NULL;
} }
@ -18597,18 +18584,12 @@ struct gguf_context * gguf_init(const char * fname, struct gguf_init_params para
if (!ok) { if (!ok) {
fprintf(stderr, "%s: failed to create tensors\n", __func__); fprintf(stderr, "%s: failed to create tensors\n", __func__);
fclose(file); fclose(file);
if (ctx_new) {
ggml_free(ctx_data); ggml_free(ctx_data);
} else {
ggml_set_no_alloc(ctx_data, ctx_no_alloc);
}
gguf_free(ctx); gguf_free(ctx);
return NULL; return NULL;
} }
ggml_set_no_alloc(ctx_data, ctx_no_alloc); ggml_set_no_alloc(ctx_data, false);
} else {
GGML_ASSERT("gguf: invalid params - load requires malloc or ctx");
} }
} }

5
ggml.h
View File

@ -1636,9 +1636,8 @@ extern "C" {
struct gguf_init_params { struct gguf_init_params {
bool load; // load the tensor data bool load; // load the tensor data
bool malloc; // if false, use the provided ggml_context to allocate the tensor data bool malloc; // if false, create a ggml_context and allocate the tensor data in it
// it no ggml_context is provided, it will be created // if true, use malloc to allocate the tensor data instead
// if true, use malloc to allocate the tensor data
struct ggml_context ** ctx; struct ggml_context ** ctx;
}; };