Skip to content
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

Jump to bottom

[PJRT:GPU] Treat GPU collective memory space as device memory space #16921

Closed
wants to merge 2 commits into from
Closed

[PJRT:GPU] Treat GPU collective memory space as device memory space #16921

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8113e6f
Treat collective memory space as device memory space when using as an…
zhenying-liu Sep 10, 2024
b5e43d6
fix the test
zhenying-liu Sep 12, 2024
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
62 changes: 62 additions & 0 deletions xla/pjrt/gpu/se_gpu_pjrt_client_test.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1131,6 +1131,24 @@ constexpr char const* kD2HProgramTupleOutput = R"(
}
)";

constexpr char const* kCollectiveMemorySpaceOutput = R"(

HloModule jit__psum, entry_computation_layout={(s32[1,4]{1,0})->s32[4]{0}}

region_0.3 {
Arg_0.0 = s32[] parameter(0)
Arg_1.0 = s32[] parameter(1)
ROOT add.0 = s32[] add(Arg_0.0, Arg_1.0)
}

ENTRY main.10_spmd {
param = s32[1,4]{1,0} parameter(0)
reshape = s32[4]{0} reshape(param)
ROOT all-reduce = s32[4]{0} all-reduce(reshape), channel_id=1, to_apply=region_0.3
}

)";

} // namespace

TEST(StreamExecutorGpuClientTest, ExecutePinnedHostOutputTest) {
Expand Down Expand Up @@ -1197,6 +1215,50 @@ TEST(StreamExecutorGpuClientTest, ExecutablePinnedHostOutputMemoryKindTest) {
EXPECT_EQ(memory_kinds[0][0], "pinned_host");
}

// Verify the output device memory kind with collective memory space shape when
// NCCL user buffer is enabled.
TEST(StreamExecutorGpuClientTest,
ExecutableCollectiveMemoryOutputMemoryKindTest) {
TF_ASSERT_OK_AND_ASSIGN(auto client,
GetStreamExecutorGpuClient(GpuClientOptions()));
xla::CompileOptions options;
options.executable_build_options.mutable_debug_options()
->set_xla_gpu_enable_nccl_user_buffers(true);

TF_ASSERT_OK_AND_ASSIGN(
auto executable,
CompileExecutable(kCollectiveMemorySpaceOutput, *client, options));
std::vector<int32_t> data{1, 2, 3, 4};
// Build the input shape with the correct memory space set.
Shape shape = ShapeUtil::MakeShapeWithDenseLayout(S32, {1, 4},
/*major_to_minor=*/{1, 0});
shape.mutable_layout()->set_memory_space(Layout::kDefaultMemorySpace);

auto device = client->addressable_devices()[0];
TF_EXPECT_OK(device->default_memory_space());
TF_ASSERT_OK_AND_ASSIGN(
auto input, client->BufferFromHostBuffer(
data.data(), shape.element_type(), shape.dimensions(),
/*byte_strides=*/std::nullopt,
PjRtClient::HostBufferSemantics::kImmutableOnlyDuringCall,
/*on_done_with_host_buffer=*/nullptr, device));
EXPECT_EQ(input->memory_space()->kind(), "device");

TF_ASSERT_OK_AND_ASSIGN(auto memory_kinds,
executable->GetOutputMemoryKinds());
EXPECT_EQ(memory_kinds.size(), 1);
EXPECT_EQ(memory_kinds[0].size(), 1);
EXPECT_EQ(memory_kinds[0][0], "device");

TF_ASSERT_OK_AND_ASSIGN(
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What happens if we execute it without the changes in this PR?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

It will have the assertion failure in pjrt_stream_execution_client.cc:MemoryKindFromSimpleShape():
INVALID_ARGUMENT: Unexpected memory space 1 in output layout

auto result, executable->Execute({{input.get()}}, ExecuteOptions()));
std::vector<std::unique_ptr<xla::PjRtBuffer>>& result_buffers = result[0];
EXPECT_EQ(result_buffers[0]->memory_space()->kind(), "device");
Shape result_shape = result_buffers[0]->on_device_shape();
auto memory_space = result_shape.layout().memory_space();
EXPECT_EQ(memory_space, 1);
}

TEST(StreamExecutorGpuClientTest,
ExecutablePinnedHostTupleOutputMemoryKindTest) {
TF_ASSERT_OK_AND_ASSIGN(auto client,
Expand Down
2 changes: 2 additions & 0 deletions xla/pjrt/pjrt_stream_executor_client.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2286,6 +2286,7 @@ absl::StatusOr<std::unique_ptr<PjRtBuffer>> OutputBufferHelper(
device->default_memory_space().value_or(nullptr);
if (shape.has_layout()) {
switch (shape.layout().memory_space()) {
case Layout::kGenericFastMemorySpace:
case Layout::kDefaultMemorySpace:
// Nothing to do, we have already set the default memory space.
break;
Expand Down Expand Up @@ -3322,6 +3323,7 @@ absl::StatusOr<absl::string_view> MemoryKindFromSimpleShape(
switch (shape.layout().memory_space()) {
case Layout::kHostMemorySpace:
return PinnedHostMemorySpace::kKind;
case Layout::kGenericFastMemorySpace:
case Layout::kDefaultMemorySpace:
return default_memory_kind;
default:
Expand Down
Loading