zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-09-29 01:58:12 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Support disable metal buffer cache to prevent performance degradation caused by large memory caching (#390)

Browse Source 
* support disable metal buffer cache, due to large unused memory buffered when llm generated long context tokens

* Run format and add "cache_enabled" feature tests
This commit is contained in:
Ethan
2024-01-19 00:33:34 +08:00
committed by GitHub
parent 49a52610b7
commit a749a91c75
4 changed files with 67 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

41
tests/metal_tests.cpp
View File

@@ -5,6 +5,7 @@
#include "mlx/backend/metal/device.h"
#include "mlx/backend/metal/metal.h"
#include "mlx/backend/metal/allocator.h"
#include "mlx/mlx.h"
using namespace mlx::core;
@@ -471,3 +472,43 @@ TEST_CASE("test metal validation") {
eval(scatter_max(array(1), {}, array(2), std::vector<int>{}));
}
TEST_CASE("test metal enable/disable cache") {
// Test enable metal cache
{
metal::set_cache_enabled(true);
CHECK(metal::cache_enabled());
auto &a = metal::allocator();
auto size = 100;
auto buf = a.malloc(size, false);
// Release a
a.free(buf);
// Check size should equals to size
CHECK_EQ(static_cast<MTL::Buffer*>(buf.ptr())->length(), size);
}
// Test disable metal cache
{
metal::set_cache_enabled(false);
CHECK(!metal::cache_enabled());
auto &a = metal::allocator();
auto size = 100;
auto buf = a.malloc(size, false);
auto buf_ptr = static_cast<MTL::Buffer*>(buf.ptr());
unsigned char first_byte = *reinterpret_cast<unsigned char*>(buf_ptr);
printf("first byte: %d\n", first_byte);
// Release a
a.free(buf);
// If release successfully, the first byte should be different from the first byte before release
unsigned char new_first_byte = *reinterpret_cast<unsigned char*>(buf_ptr);
printf("new first byte: %d\n", new_first_byte);
CHECK_NE(new_first_byte, first_byte);
}
}