Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

model-loader : fix quantized reshaped tensor strides (#26672)

Browse Source
This commit is contained in:
Georgi Gerganov
2026-08-06 15:21:44 +03:00
committed by GitHub
parent e700bfb37f
commit 3db4ff877d
1 changed files with 16 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama-model-loader.cpp
+16 -2
View File
@@ -1249,7 +1249,13 @@ struct ggml_tensor * llama_model_loader::create_tensor(
for (size_t dim = 0; dim < GGML_MAX_DIMS; dim++) { for (size_t dim = 0; dim < GGML_MAX_DIMS; dim++) {
t_meta.ne[dim] = dim < ne.size() ? ne.begin()[dim] : 1; t_meta.ne[dim] = dim < ne.size() ? ne.begin()[dim] : 1;
GGML_ASSERT(t_meta.ne[dim] >= 1); GGML_ASSERT(t_meta.ne[dim] >= 1);
t_meta.nb[dim] = dim == 0 ? ggml_type_size(type) : t_meta.ne[dim-1]*t_meta.nb[dim-1]; if (dim == 0) {
t_meta.nb[dim] = ggml_type_size(type);
} else if (dim == 1) {
t_meta.nb[dim] = ggml_row_size(type, t_meta.ne[dim-1]);
} else {
t_meta.nb[dim] = t_meta.nb[dim-1]*t_meta.ne[dim-1];
}
GGML_ASSERT(t_meta.nb[dim] >= 1); GGML_ASSERT(t_meta.nb[dim] >= 1);
} }
ggml_set_name(&t_meta, tn.str().c_str()); ggml_set_name(&t_meta, tn.str().c_str());
@@ -1272,10 +1278,18 @@ struct ggml_tensor * llama_model_loader::create_tensor(
if (flags & TENSOR_ALLOW_RESHAPE) { if (flags & TENSOR_ALLOW_RESHAPE) {
for (size_t dim = 0; dim < GGML_MAX_DIMS; dim++) { for (size_t dim = 0; dim < GGML_MAX_DIMS; dim++) {
t_meta.ne[dim] = dim < ne.size() ? ne.begin()[dim] : 1; t_meta.ne[dim] = dim < ne.size() ? ne.begin()[dim] : 1;
t_meta.nb[dim] = dim == 0 ? ggml_type_size(t_meta.type) : t_meta.ne[dim-1]*t_meta.nb[dim-1]; if (dim == 0) {
t_meta.nb[dim] = ggml_type_size(t_meta.type);
} else if (dim == 1) {
t_meta.nb[dim] = ggml_row_size(t_meta.type, t_meta.ne[dim-1]);
} else {
t_meta.nb[dim] = t_meta.ne[dim-1]*t_meta.nb[dim-1];
}
} }
} }
GGML_ASSERT(ggml_nbytes(&t_meta) == ggml_nbytes(cur));
ggml_backend_buffer_type_t buft = buft_for_tensor(&t_meta); ggml_backend_buffer_type_t buft = buft_for_tensor(&t_meta);
if (buft == nullptr) { if (buft == nullptr) {
return nullptr; return nullptr;