jdelony/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-06-28 04:30:15 -05:00
Code Issues Packages Projects Releases Wiki Activity

allow user to use THP for host allocations with GGML_CUDA_HOST_MALLOC_THP (#2010)

Browse Source 
* allow user to use THP for host allocations with GGML_CUDA_HOST_MALLOC_THP

* Remove useless symbol check
This commit is contained in:
Farmadupe 2026-06-23 14:13:41 +01:00 committed by GitHub
parent 2d3ecd5e19
commit 7ccf1d2095
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 11 additions and 13 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

24
ggml/src/ggml-cuda.cu
View File

@ -1424,17 +1424,16 @@ static void * ggml_cuda_host_malloc(size_t size) {
return nullptr; return nullptr;
} }
// Whether to request the kernel to attempt to defragment memory to back the region with 2M hugepages. // Whether to request the kernel to attempt to defragment memory to back the region with 2M hugepages.
// Otherwise dependent on kernel settings: // Otherwise dependent on kernel settings:
// * enabled="always": Hand over whatever 2M pages it has on hand and the rest will be 4k // * enabled="always": Hand over whatever 2M pages it has on hand and the rest will be 4k
// * enabled="madvise": 4k pages // * enabled="madvise": 4k pages
// * enabled="never": 4k pages // * enabled="never": 4k pages
// Potluck on performance. If there's not much defragmentation to do, then you win. Otherwise come back in an hour. // Potluck on performance. If there's not much defragmentation to do, then you win. Otherwise come back in an hour.
#if 0 // Defaults to disabled unless GGML_CUDA_HOST_MALLOC_THP is set.
#ifdef MADV_HUGEPAGE if (getenv("GGML_CUDA_HOST_MALLOC_THP") != nullptr) {
madvise(ptr, size, MADV_HUGEPAGE); madvise(ptr, size, MADV_HUGEPAGE);
#endif }
#endif
// prefault the whole region. If the kernel knows how to do this then let it do so. // prefault the whole region. If the kernel knows how to do this then let it do so.
// Might be worth spawning threads to speed up this process on huge allocations. // Might be worth spawning threads to speed up this process on huge allocations.
@ -1442,8 +1441,7 @@ static void * ggml_cuda_host_malloc(size_t size) {
#ifdef MADV_POPULATE_WRITE #ifdef MADV_POPULATE_WRITE
needs_manual_prefault = madvise(ptr, size, MADV_POPULATE_WRITE); needs_manual_prefault = madvise(ptr, size, MADV_POPULATE_WRITE);
#endif #endif
if (needs_manual_prefault) if (needs_manual_prefault) {
{
char * p = (char *) ptr; char * p = (char *) ptr;
for (size_t off = 0; off < size; off += 4096) { for (size_t off = 0; off < size; off += 4096) {
p[off] = 0; p[off] = 0;