Public/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-19 17:24:57 +02:00
Code Issues Packages Projects Releases Wiki Activity

rename to TENSOR_READ_LAZY

Browse Source
This commit is contained in:
Xuan Son Nguyen
2026-08-27 13:24:29 +02:00
parent bb07e28f88
commit f3bb7618c5
5 changed files with 6 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-model-loader.cpp
+1 -1
View File
@@ -1282,7 +1282,7 @@ struct ggml_tensor * llama_model_loader::create_tensor(
return NULL;
}
if ((flags & TENSOR_GET_ROW_LAZY) && use_mmap && tensor_read_lazy != LLAMA_TENSOR_READ_LAZY_OFF) {
if ((flags & TENSOR_READ_LAZY) && use_mmap && tensor_read_lazy != LLAMA_TENSOR_READ_LAZY_OFF) {
// in auto mode, small tensors are cheap enough to keep resident
constexpr size_t auto_lazy_min_size = 4ull * 1024 * 1024 * 1024;
if (tensor_read_lazy == LLAMA_TENSOR_READ_LAZY_ON || ggml_nbytes(cur) > auto_lazy_min_size) {
src/llama-model-loader.h
+2 -2
View File
@@ -68,7 +68,7 @@ struct llama_model_loader {
static const int TENSOR_SKIP = 1 << 2;
static const int TENSOR_SKIP_IF_VIRTUAL = 1 << 3;
static const int TENSOR_ALLOW_RESHAPE = 1 << 4;
static const int TENSOR_GET_ROW_LAZY = 1 << 5; // read rows on demand instead of loading whole tensor; requires mmap for now
static const int TENSOR_READ_LAZY = 1 << 5; // read rows on demand instead of loading whole tensor; requires mmap for now
int n_kv = 0;
int n_tensors = 0;
@@ -92,7 +92,7 @@ struct llama_model_loader {
llama_mmaps mappings;
// byte ranges of TENSOR_GET_ROW_LAZY tensors, per file index
// byte ranges of TENSOR_READ_LAZY tensors, per file index
std::map<uint32_t, std::vector<std::pair<size_t, size_t>>> lazy_tensor_ranges;
std::map<std::string, llama_tensor_weight, weight_name_comparer> weights_map;
src/llama-model.cpp
+1 -1
View File
@@ -3069,7 +3069,7 @@ llama_model_base::llama_model_base(const struct llama_model_params & params) : l
TENSOR_SKIP (llama_model_loader::TENSOR_SKIP),
TENSOR_SKIP_IF_VIRTUAL(llama_model_loader::TENSOR_SKIP_IF_VIRTUAL),
TENSOR_ALLOW_RESHAPE (llama_model_loader::TENSOR_ALLOW_RESHAPE),
TENSOR_GET_ROW_LAZY (llama_model_loader::TENSOR_GET_ROW_LAZY) {}
TENSOR_READ_LAZY (llama_model_loader::TENSOR_READ_LAZY) {}
ggml_tensor * llama_model_base::create_tensor(const LLM_TN_IMPL & tn, const std::initializer_list<int64_t> & ne, int flags) {
GGML_ASSERT(ml != nullptr);
src/llama-model.h
+1 -1
View File
@@ -756,7 +756,7 @@ struct llama_model_base : public llama_model {
const int TENSOR_SKIP;
const int TENSOR_SKIP_IF_VIRTUAL;
const int TENSOR_ALLOW_RESHAPE;
const int TENSOR_GET_ROW_LAZY;
const int TENSOR_READ_LAZY;
explicit llama_model_base(const llama_model_params & params);
virtual ~llama_model_base() = default;
src/models/gemma4.cpp
+1 -1
View File
@@ -50,7 +50,7 @@ void llama_model_gemma4::load_arch_tensors(llama_model_loader &) {
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
if (n_embd_per_layer > 0) {
per_layer_tok_embd = create_tensor(tn(LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "weight"), {n_embd_per_layer * n_layer, n_vocab}, TENSOR_GET_ROW_LAZY);
per_layer_tok_embd = create_tensor(tn(LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "weight"), {n_embd_per_layer * n_layer, n_vocab}, TENSOR_READ_LAZY);
per_layer_model_proj = create_tensor(tn(LLM_TENSOR_PER_LAYER_MODEL_PROJ, "weight", 0), {n_embd, n_embd_per_layer * n_layer}, 0);
per_layer_proj_norm = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ_NORM, "weight", 0), {n_embd_per_layer}, 0);
}