CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-04 09:32:00 +00:00
Code Issues 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
2 changed files with 18 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

51
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) {
GGML_ASSERT(!params.load || params.malloc || params.ctx != NULL);
FILE * file = fopen(fname, "rb");
if (!file) {
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);
// TODO: pad size_cur to alignment
ctx->size_data += size_cur;
ctx->size_data += GGML_PAD(size_cur, ctx->alignment);
}
// TODO: simplify
@@ -18528,28 +18529,18 @@ struct gguf_context * gguf_init(const char * fname, struct gguf_init_params para
ctx->data = GGML_ALIGNED_MALLOC(ctx->size_data);
fseek(file, ctx->offset, SEEK_SET);
ok = ok && gguf_fread_el(ctx->data, ctx->size_data, file, &offset);
} else if (params.ctx != NULL) {
bool ctx_new = false;
bool ctx_no_alloc = false;
} else {
const size_t mem_size =
ctx->header.n_tensors*ggml_tensor_overhead() + 1 +
ctx->size_data;
if (*params.ctx == NULL) {
const size_t mem_size =
ctx->header.n_tensors*ggml_tensor_overhead() + 1 +
ctx->size_data;
struct ggml_init_params pdata = {
.mem_size = mem_size,
.mem_buffer = NULL,
.no_alloc = false,
};
struct ggml_init_params pdata = {
.mem_size = mem_size,
.mem_buffer = NULL,
.no_alloc = false,
};
*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);
}
*params.ctx = ggml_init(pdata);
struct ggml_context * ctx_data = *params.ctx;
@@ -18561,11 +18552,7 @@ struct gguf_context * gguf_init(const char * fname, struct gguf_init_params para
if (!ok) {
fprintf(stderr, "%s: failed to read tensor data\n", __func__);
fclose(file);
if (ctx_new) {
ggml_free(ctx_data);
} else {
ggml_set_no_alloc(ctx_data, ctx_no_alloc);
}
ggml_free(ctx_data);
gguf_free(ctx);
return NULL;
}
@@ -18597,18 +18584,12 @@ struct gguf_context * gguf_init(const char * fname, struct gguf_init_params para
if (!ok) {
fprintf(stderr, "%s: failed to create tensors\n", __func__);
fclose(file);
if (ctx_new) {
ggml_free(ctx_data);
} else {
ggml_set_no_alloc(ctx_data, ctx_no_alloc);
}
ggml_free(ctx_data);
gguf_free(ctx);
return NULL;
}
ggml_set_no_alloc(ctx_data, ctx_no_alloc);
} else {
GGML_ASSERT("gguf: invalid params - load requires malloc or ctx");
ggml_set_no_alloc(ctx_data, false);
}
}

5
ggml.h
View File

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