Skip to content

Commit c40ec5c

Browse files
ggerganovggerganov
committed
llama : add hparams.ctx_train + no longer print ftype
1 parent 8be49fd commit c40ec5c

File tree

1 file changed

+30
-27
lines changed

1 file changed

+30
-27
lines changed

‎llama.cpp

Lines changed: 30 additions & 27 deletions
Original file line numberDiff line numberDiff line change
@@ -676,22 +676,21 @@ static const std::map<e_model, size_t> & VRAM_REQ_SCRATCH_PER_CONTEXT()
676676

677677
// default hparams (LLaMA 7B)
678678
struct llama_hparams {
679-
uint32_t n_vocab = 32000;
680-
uint32_t n_ctx = 512;
681-
uint32_t n_embd = 4096;
682-
uint32_t n_head = 32;
683-
uint32_t n_head_kv = 32;
684-
uint32_t n_layer = 32;
685-
uint32_t n_rot = 64;
686-
uint32_t n_ff = 11008;
679+
uint32_t n_vocab = 32000;
680+
uint32_t n_ctx_train = 2048; // the context size used during training
681+
uint32_t n_ctx = 512; // the context size used during inference
682+
uint32_t n_embd = 4096;
683+
uint32_t n_head = 32;
684+
uint32_t n_head_kv = 32;
685+
uint32_t n_layer = 32;
686+
uint32_t n_rot = 64;
687+
uint32_t n_ff = 11008;
687688

688689
float f_norm_rms_eps = 1e-5;
689690

690691
float rope_freq_base = 10000.0f;
691692
float rope_freq_scale = 1.0f;
692693

693-
enum llama_ftype ftype = LLAMA_FTYPE_MOSTLY_F16;
694-
695694
bool operator!=(const llama_hparams & other) const {
696695
return static_cast<bool>(memcmp(this, &other, sizeof(llama_hparams))); // NOLINT
697696
}
@@ -1325,7 +1324,7 @@ static void llama_model_load_internal(
13251324
}
13261325

13271326
GGUF_GET(hparams.n_vocab, gguf_get_arr_n, GGUF_TYPE_ARRAY, true, "tokenizer.ggml.tokens");
1328-
GGUF_GET(hparams.n_ctx, gguf_get_val_u32, GGUF_TYPE_UINT32, true, "llama.context_length");
1327+
GGUF_GET(hparams.n_ctx_train, gguf_get_val_u32, GGUF_TYPE_UINT32, true, "llama.context_length");
13291328
GGUF_GET(hparams.n_embd, gguf_get_val_u32, GGUF_TYPE_UINT32, true, "llama.embedding_length");
13301329
GGUF_GET(hparams.n_ff, gguf_get_val_u32, GGUF_TYPE_UINT32, true, "llama.feed_forward_length");
13311330
GGUF_GET(hparams.n_head, gguf_get_val_u32, GGUF_TYPE_UINT32, true, "llama.attention.head_count");
@@ -1399,21 +1398,23 @@ static void llama_model_load_internal(
13991398
}
14001399

14011400
{
1402-
LLAMA_LOG_INFO("%s: format = %s\n", __func__, llama_file_version_name(ml->file_version));
1403-
LLAMA_LOG_INFO("%s: n_vocab = %u\n", __func__, hparams.n_vocab);
1404-
LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, hparams.n_ctx);
1405-
LLAMA_LOG_INFO("%s: n_embd = %u\n", __func__, hparams.n_embd);
1406-
LLAMA_LOG_INFO("%s: n_head = %u\n", __func__, hparams.n_head);
1407-
LLAMA_LOG_INFO("%s: n_head_kv = %u\n", __func__, hparams.n_head_kv);
1408-
LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer);
1409-
LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot); // a.k.a. n_embd_head, n_head_dim
1410-
LLAMA_LOG_INFO("%s: n_gqa = %u\n", __func__, hparams.n_gqa());
1411-
LLAMA_LOG_INFO("%s: f_norm_eps = %.1e\n", __func__, hparams.f_norm_rms_eps);
1412-
LLAMA_LOG_INFO("%s: n_ff = %u\n", __func__, hparams.n_ff);
1413-
LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, hparams.rope_freq_base);
1414-
LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, hparams.rope_freq_scale);
1415-
LLAMA_LOG_INFO("%s: ftype = %u (%s)\n", __func__, hparams.ftype, llama_ftype_name(hparams.ftype));
1416-
LLAMA_LOG_INFO("%s: model size = %s\n", __func__, llama_model_type_name(model.type));
1401+
LLAMA_LOG_INFO("%s: format = %s\n", __func__, llama_file_version_name(ml->file_version));
1402+
LLAMA_LOG_INFO("%s: n_vocab = %u\n", __func__, hparams.n_vocab);
1403+
LLAMA_LOG_INFO("%s: n_ctx_train = %u\n", __func__, hparams.n_ctx_train);
1404+
LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, hparams.n_ctx);
1405+
LLAMA_LOG_INFO("%s: n_embd = %u\n", __func__, hparams.n_embd);
1406+
LLAMA_LOG_INFO("%s: n_head = %u\n", __func__, hparams.n_head);
1407+
LLAMA_LOG_INFO("%s: n_head_kv = %u\n", __func__, hparams.n_head_kv);
1408+
LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer);
1409+
LLAMA_LOG_INFO("%s: n_rot = %u\n", __func__, hparams.n_rot); // a.k.a. n_embd_head, n_head_dim
1410+
LLAMA_LOG_INFO("%s: n_gqa = %u\n", __func__, hparams.n_gqa());
1411+
LLAMA_LOG_INFO("%s: f_norm_eps = %.1e\n", __func__, hparams.f_norm_rms_eps);
1412+
LLAMA_LOG_INFO("%s: n_ff = %u\n", __func__, hparams.n_ff);
1413+
LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, hparams.rope_freq_base);
1414+
LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, hparams.rope_freq_scale);
1415+
LLAMA_LOG_INFO("%s: model size = %s\n", __func__, llama_model_type_name(model.type));
1416+
1417+
// TODO: print number of tensors for each quantization
14171418
}
14181419

14191420
if (vocab_only) {
@@ -3365,7 +3366,6 @@ static void llama_convert_tensor_internal(struct ggml_tensor * tensor, std::vect
33653366
static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, const llama_model_quantize_params * params) {
33663367
ggml_type quantized_type;
33673368
llama_ftype ftype = params->ftype;
3368-
int nthread = params->nthread;
33693369

33703370
switch (params->ftype) {
33713371
case LLAMA_FTYPE_MOSTLY_Q4_0: quantized_type = GGML_TYPE_Q4_0; break;
@@ -3391,6 +3391,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
33913391
default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
33923392
}
33933393

3394+
int nthread = params->nthread;
3395+
33943396
if (nthread <= 0) {
33953397
nthread = std::thread::hardware_concurrency();
33963398
}
@@ -3661,6 +3663,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
36613663
}
36623664
}
36633665

3666+
// TODO: after the GGUF PR, this likely won't work and needs to be updated
36643667
int llama_apply_lora_from_file_internal(const struct llama_model & model, const char * path_lora, const char * path_base_model, int n_threads) {
36653668
LLAMA_LOG_INFO("%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora);
36663669

0 commit comments

Comments
 (0)