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

Increase GGML_MAX_NAME

Browse source 
Avoiding strncpy danger in train-text-from-scratch and reducing potential future name length issues
This commit is contained in:
David Yang 2023-06-26 11:26:48 +08:00
parent 0766ee3de4
commit 42bf841232
2 changed files with 4 additions and 15 deletions
Download patch file Download diff file Expand all files Collapse all files

17
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -294,20 +294,9 @@ void init_model(struct my_llama_model * model) {
ggml_set_name(layer.ffn_norm, (layers_i + ".ffn_norm.weight").c_str());
// 'layers.10.feed_forward.w1.weight' has length of 32.
// ggml_tensor->name only has 32 characters, but we need one more for the '\0' terminator.
// ggml_set_name will set the last character to '\0', so we can only store 'layers.10.feed_forward.w1.weigh'.
// when saving llama compatible model the tensors names will miss a character.
// ggml_set_name(layer.w1, (layers_i + ".feed_forward.w1.weight").c_str());
// ggml_set_name(layer.w2, (layers_i + ".feed_forward.w2.weight").c_str());
// ggml_set_name(layer.w3, (layers_i + ".feed_forward.w3.weight").c_str());
strncpy(layer.w1->name, (layers_i + ".feed_forward.w1.weight").c_str(), sizeof(layer.w1->name));
strncpy(layer.w2->name, (layers_i + ".feed_forward.w2.weight").c_str(), sizeof(layer.w2->name));
strncpy(layer.w3->name, (layers_i + ".feed_forward.w3.weight").c_str(), sizeof(layer.w3->name));
layer.w1->padding[0] = 0;
layer.w2->padding[0] = 0;
layer.w3->padding[0] = 0;
ggml_format_name(layer.w1, "%s.feed_forward.w1.weight", layers_i.c_str());
ggml_format_name(layer.w2, "%s.feed_forward.w2.weight", layers_i.c_str());
ggml_format_name(layer.w3, "%s.feed_forward.w3.weight", layers_i.c_str());
}
}

2
ggml.h
View file

@ -198,7 +198,7 @@
#define GGML_MAX_PARAMS 256
#define GGML_MAX_CONTEXTS 64
#define GGML_MAX_OPT 4
#define GGML_MAX_NAME 32
#define GGML_MAX_NAME 48
#define GGML_DEFAULT_N_THREADS 4
#define GGML_ASSERT(x) \