root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-01-12 19:50:17 +00:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

llama : fix loading models with shared tok_embd and output (#5651)

Browse Source 
ggml-ci
This commit is contained in:
slaren 2024-02-22 00:42:09 +01:00 committed by GitHub
parent 7c8bcc11dc
commit 973053d8b0
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 4 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
llama.cpp
View File

@ -2791,13 +2791,7 @@ struct llama_model_loader {
std::vector<no_init<uint8_t>> read_buf; std::vector<no_init<uint8_t>> read_buf;
for (int i = 0; i < gguf_get_n_tensors(ctx_gguf); i++) { for (struct ggml_tensor * cur = ggml_get_first_tensor(ctx); cur != NULL; cur = ggml_get_next_tensor(ctx, cur)) {
struct ggml_tensor * cur = ggml_get_tensor(ctx, gguf_get_tensor_name(ctx_gguf, i));
if (!cur) {
// some tensors may be allocated in a different context
continue;
}
if (progress_callback) { if (progress_callback) {
if (!progress_callback((float) size_done / size_data, progress_callback_user_data)) { if (!progress_callback((float) size_done / size_data, progress_callback_user_data)) {
return false; return false;
@ -3722,7 +3716,7 @@ static bool llm_load_tensors(
} }
// create one context per buffer type // create one context per buffer type
size_t ctx_size = ggml_tensor_overhead()*ml.n_tensors; size_t ctx_size = ggml_tensor_overhead()*(ml.n_tensors + 1); // +1 for models where tok_embd is duplicated as output
std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map; std::map<ggml_backend_buffer_type_t, ggml_context *> ctx_map;
for (auto & it : buft_layer_count) { for (auto & it : buft_layer_count) {
struct ggml_init_params params = { struct ggml_init_params params = {
@ -3860,6 +3854,7 @@ static bool llm_load_tensors(
} else { } else {
model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // needs to be on GPU model.output = ml.create_tensor(ctx_output_split, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // needs to be on GPU
ml.n_created--; // artificial tensor ml.n_created--; // artificial tensor
ml.size_data += ggml_nbytes(model.output);
} }
} }
@ -4396,6 +4391,7 @@ static bool llm_load_tensors(
model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // same as tok_embd, duplicated to allow offloading model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); // same as tok_embd, duplicated to allow offloading
ml.n_created--; // artificial tensor ml.n_created--; // artificial tensor
ml.size_data += ggml_nbytes(model.output);
const int64_t n_ff = hparams.n_ff; const int64_t n_ff = hparams.n_ff;
const int64_t n_embd_head_k = hparams.n_embd_head_k; const int64_t n_embd_head_k = hparams.n_embd_head_k;