vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

no duplicated tensor while reading gguf

Browse source
This commit is contained in:
ngxson 2024-04-27 15:31:52 +02:00
parent 9cd09aa79b
commit 509c28f4a1
3 changed files with 24 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

10
ggml.c
View file

@ -20824,6 +20824,14 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
gguf_tensor_info_sanitize(info); gguf_tensor_info_sanitize(info);
// make sure there is no duplicated tensor names
for (uint64_t j = 0; j < i; ++j) {
if (strcmp(info->name.data, ctx->infos[j].name.data) == 0) {
fprintf(stderr, "%s: duplicated tensor name %s\n", __func__, info->name.data);
ok = false;
}
}
if (!ok) { if (!ok) {
fprintf(stderr, "%s: failed to read tensor info\n", __func__); fprintf(stderr, "%s: failed to read tensor info\n", __func__);
fclose(file); fclose(file);
@ -21363,7 +21371,7 @@ void gguf_add_tensor(
if (gguf_find_tensor(ctx, tensor->name) != -1) { if (gguf_find_tensor(ctx, tensor->name) != -1) {
GGML_ASSERT(false && "duplicated tensor name"); GGML_ASSERT(false && "duplicated tensor name");
} }
const int idx = ctx->header.n_tensors; const int idx = ctx->header.n_tensors;
ctx->infos = realloc(ctx->infos, (idx + 1)*sizeof(struct gguf_tensor_info)); ctx->infos = realloc(ctx->infos, (idx + 1)*sizeof(struct gguf_tensor_info));

8
gguf-py/gguf/gguf_reader.py
View file

@ -234,8 +234,14 @@ class GGUFReader:
def _build_tensors(self, start_offs: int, fields: list[ReaderField]) -> None: def _build_tensors(self, start_offs: int, fields: list[ReaderField]) -> None:
tensors = [] tensors = []
tensor_names = [] # keep track of name to prevent duplicated tensors
for field in fields: for field in fields:
_name_len, name_data, _n_dims, dims, raw_dtype, offset_tensor = field.parts _name_len, name_data, _n_dims, dims, raw_dtype, offset_tensor = field.parts
# check if there's any tensor having same name already in the list
tensor_name = str(bytes(name_data), encoding = 'utf-8')
if tensor_name in tensor_names:
raise ValueError(f'Found duplicated tensor with name {tensor_name}')
tensor_names.add(tensor_name)
ggml_type = GGMLQuantizationType(raw_dtype[0]) ggml_type = GGMLQuantizationType(raw_dtype[0])
n_elems = np.prod(dims) n_elems = np.prod(dims)
block_size, type_size = GGML_QUANT_SIZES[ggml_type] block_size, type_size = GGML_QUANT_SIZES[ggml_type]
@ -267,7 +273,7 @@ class GGUFReader:
item_count = n_bytes item_count = n_bytes
item_type = np.uint8 item_type = np.uint8
tensors.append(ReaderTensor( tensors.append(ReaderTensor(
name = str(bytes(name_data), encoding = 'utf-8'), name = tensor_name,
tensor_type = ggml_type, tensor_type = ggml_type,
shape = dims, shape = dims,
n_elements = n_elems, n_elements = n_elems,

8
llama.cpp
View file

@ -3110,9 +3110,17 @@ struct llama_model_loader {
fver = (enum llama_fver) gguf_get_version(meta); fver = (enum llama_fver) gguf_get_version(meta);
std::set<std::string> tensor_names;
for (auto & w : weights) { for (auto & w : weights) {
n_elements += ggml_nelements(w.tensor); n_elements += ggml_nelements(w.tensor);
n_bytes += ggml_nbytes(w.tensor); n_bytes += ggml_nbytes(w.tensor);
// make sure there is no duplicated tensor names
const std::string name(w.tensor->name);
auto found = tensor_names.find(name);
if (found != tensor_names.end()) {
LLAMA_LOG_ERROR("%s: Found duplicated tensor name %s", __func__, w.tensor->name);
}
tensor_names.insert(name);
} }
LLAMA_LOG_INFO("%s: loaded meta data with %d key-value pairs and %d tensors from %s (version %s)\n", LLAMA_LOG_INFO("%s: loaded meta data with %d key-value pairs and %d tensors from %s (version %s)\n",