zhangyiss/mlx
1
0
Fork 0
mirror of https://github.com/ml-explore/mlx.git synced 2025-11-02 09:18:11 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

Use async cuda malloc managed with cuda 13

Browse Source
This commit is contained in:
Awni Hannun
2025-10-26 16:17:27 -07:00
parent 8f8af61a37
commit 58ccbaaf12
19 changed files with 110 additions and 54 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
mlx/backend/cuda/indexing.cpp
View File

@@ -59,7 +59,9 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
assert(inputs.size() > 0);
const auto& src = inputs[0];
out.set_data(allocator::malloc(out.nbytes()));
auto& s = stream();
auto& encoder = cu::get_command_encoder(s);
out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
if (out.size() == 0) {
return;
}
@@ -80,7 +82,6 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
dtype_to_string(idx_dtype),
nidx);
auto& s = stream();
cu::JitModule& mod = cu::get_jit_module(s.device, module_name, [&]() {
std::vector<std::string> kernel_names;
for (int ndim = 0; ndim <= MAX_NDIM; ++ndim) {
@@ -121,7 +122,6 @@ void Gather::eval_gpu(const std::vector<array>& inputs, array& out) {
idx_ndim,
large ? "int64_t" : "int32_t");
auto& encoder = cu::get_command_encoder(s);
for (const auto& in : inputs) {
encoder.set_input_array(in);
}
@@ -239,7 +239,9 @@ void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
const auto& src = inputs[0];
const auto& idx = inputs[1];
out.set_data(allocator::malloc(out.nbytes()));
auto& s = stream();
auto& encoder = cu::get_command_encoder(s);
out.set_data(cu::malloc_async(out.nbytes(), encoder.stream()));
if (out.size() == 0) {
return;
}
@@ -251,7 +253,6 @@ void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
dtype_to_string(out.dtype()),
dtype_to_string(idx.dtype()));
auto& s = stream();
cu::JitModule& mod = cu::get_jit_module(s.device, module_name, [&]() {
std::vector<std::string> kernel_names;
for (int ndim = 0; ndim <= MAX_NDIM; ++ndim) {
@@ -312,7 +313,6 @@ void GatherAxis::eval_gpu(const std::vector<array>& inputs, array& out) {
idx.flags().row_contiguous,
large ? "int64_t" : "int32_t");
auto& encoder = cu::get_command_encoder(s);
for (const auto& in : inputs) {
encoder.set_input_array(in);
}