mirrors/ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2025-05-11 18:36:41 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

ollamarunner: Re-enable worst case graph preallocation.

Browse source 
Worst case graph preallocation was disabled by a27462b
"ollamarunner: Temporarily disable worst case graph preallocation"
since it caused crashes with large batches when not using the GPU.

This backports upstream llama.cpp commit f057808
"ggml: Don't assert fail when tensor data changes (#13222)", which
fixes the underlying bug and allows reverting the previous workaround.
This commit is contained in:
Jesse Gross 2025-05-02 11:24:19 -07:00 • committed by Jesse Gross
parent 57fb759f3c
commit c2f5d6662b
3 changed files with 46 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

5
ml/backend/ggml/ggml/src/ggml-alloc.c vendored
View file

@ -816,7 +816,10 @@ static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor *
static bool ggml_gallocr_node_needs_realloc(ggml_gallocr_t galloc, struct ggml_tensor * node, struct tensor_alloc * talloc) {
size_t node_size = 0;
if (!node->data && !node->view_src) {
GGML_ASSERT(talloc->buffer_id >= 0); // prevent segfault when misusing the API
// If we previously had data but don't now then reallocate
if (talloc->buffer_id < 0) {
return false;
}
node_size = ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
}
return talloc->size_max >= node_size;