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

train-text-from-scratch can train (full finetune) gguf models

Browse source 
just pass the gguf model via `--checkpoint-in FN`.
after this, to continue training, pass the generated checkpoint instead of the original gguf model.

tested with smaller models, bigger models may exceed available memory.
use (LORA) finetune for those.
This commit is contained in:
xaedes 2023-08-30 15:57:17 +02:00
parent e6b7158123
commit fc456edda6
No known key found for this signature in database
GPG key ID: 30030EDD817EA2B1
1 changed files with 14 additions and 10 deletions
Download patch file Download diff file Expand all files Collapse all files

6
examples/train-text-from-scratch/train-text-from-scratch.cpp
View file

@ -1188,7 +1188,8 @@ void save_llama_model_file(const char * filename, const char * fn_vocab_model, s
void load_checkpoint_gguf(struct gguf_context * fctx, struct ggml_context * f_ggml_ctx, struct my_llama_model * model, struct ggml_opt_context * opt) {
load_llama_model_gguf(fctx, f_ggml_ctx, model);
uint32_t file_version;
if (gguf_find_key(fctx, LLM_KV_TRAINING_FILE_VERSION) >= 0) {
uint32_t file_version = 0xFFFFFFFFu;
GGUF_GET_KEY(fctx, file_version, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_FILE_VERSION);
GGML_ASSERT(file_version == 0);
@ -1201,6 +1202,9 @@ void load_checkpoint_gguf(struct gguf_context * fctx, struct ggml_context * f_gg
GGUF_GET_KEY(fctx, model->train_tokens, gguf_get_val_u32, GGUF_TYPE_UINT32, true, LLM_KV_TRAINING_TOKEN_COUNT);
load_opt_context_gguf(fctx, f_ggml_ctx, opt);
} else {
printf("%s: loaded llama model as checkpoint\n", __func__);
}
}
void save_checkpoint_gguf(struct gguf_context * fctx, const char * fn_vocab_model, struct my_llama_model * model, struct ggml_opt_context * opt) {