Skip to content

Fix custom metal kernel cache collision for same name, different source (#3832) #3833

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
katlun-lgtm wants to merge 1 commit into
base: main
Choose a base branch
from
+40 −28
Open

Fix custom metal kernel cache collision for same name, different source (#3832) #3833

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
36 changes: 8 additions & 28 deletions mlx/backend/metal/custom_kernel.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,16 +8,6 @@

namespace mlx::core::fast {

struct CustomKernelCache {
std::unordered_map<std::string, std::pair<std::string, CompileOptions::Data>>
libraries;
};

static CustomKernelCache& cache() {
static CustomKernelCache cache_;
return cache_;
};

void CustomKernel::eval_gpu(
const std::vector<array>& inputs,
std::vector<array>& outputs) {
Expand Down Expand Up @@ -55,25 +45,15 @@ void CustomKernel::eval_gpu(

auto& d = metal::device(s.device);

{
// Clear kernels from the device library cache if needed
auto& kernel_cache = cache();
if (auto it = kernel_cache.libraries.find(name_);
it != kernel_cache.libraries.end()) {
if (it->second.first != source_ ||
it->second.second != compile_options_) {
auto& d = metal::device(s.device);
d.clear_library(name_);
it->second = std::make_tuple(source_, compile_options_);
}
} else {
kernel_cache.libraries.emplace(
name_, std::make_tuple(source_, compile_options_));
}
}

// Key the library by (name, source, compile options) so kernels that differ
// in any of those coexist instead of thrashing a single cache slot within one
// eval batch (fixes #3832). The map key is never used as a Metal identifier,
// so the full source is a collision-free key (a content hash could alias
// distinct sources and silently reintroduce the bug).
std::string lib_name =
name_ + "\n" + source_ + "\n" + std::to_string(compile_options_);
auto lib = d.get_library(
name_, compile_options_, [this] { return metal::utils() + source_; });
lib_name, compile_options_, [this] { return metal::utils() + source_; });
auto kernel = d.get_kernel(name_, lib);
auto& compute_encoder = metal::get_command_encoder(s);
compute_encoder.set_compute_pipeline_state(kernel);
Expand Down
32 changes: 32 additions & 0 deletions python/tests/test_fast.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1026,6 +1026,38 @@ def call_kernel(a: mx.array, source):
out = call_kernel(a, source)
self.assertTrue(mx.array_equal(out, mx.ones_like(out)))

@unittest.skipIf(not mx.metal.is_available(), "Metal is not available")
def test_custom_kernel_same_name_different_source_one_eval(self):
# Regression test for #3832: two kernels sharing a name but with
# different sources, dispatched in a SINGLE eval batch, must each run
# their own compiled code instead of silently reusing the first's.
def call_kernel(a, source):
kernel = mx.fast.metal_kernel(
name="dup_name",
input_names=["inp"],
output_names=["out"],
source=source,
)
return kernel(
inputs=[a],
grid=(a.size, 1, 1),
threadgroup=(a.size, 1, 1),
output_shapes=[a.shape],
output_dtypes=[a.dtype],
stream=mx.gpu,
)[0]

a = mx.arange(32, dtype=mx.float32)
out_a = call_kernel(
a, "uint e = thread_position_in_grid.x; out[e] = inp[e] * 2.0f;"
)
out_b = call_kernel(
a, "uint e = thread_position_in_grid.x; out[e] = inp[e] + 100.0f;"
)
mx.eval(out_a, out_b) # one batch — the reported failure case
self.assertTrue(mx.array_equal(out_a, a * 2.0))
self.assertTrue(mx.array_equal(out_b, a + 100.0))

@unittest.skipIf(not mx.metal.is_available(), "Metal is not available")
def test_custom_metal_kernel_math_mode(self):
with self.assertRaises(ValueError):
Expand Down