Propagate status on OOM crashes and exception.

Author: ysiraichi

test/cpp/cpp_test_util.cpp

@@ -303,9 +303,8 @@ std::vector<torch_xla::runtime::ComputationClient::DataPtr> Execute(
 std::vector<at::Tensor> Fetch(
     absl::Span<const torch_xla::runtime::ComputationClient::DataPtr>
         device_data) {
-  std::vector<xla::Literal> literals =
-      torch_xla::runtime::GetComputationClientOrDie()->TransferFromDevice(
-          device_data);
+  std::vector<xla::Literal> literals = GetValueOrThrow(
+      runtime::GetComputationClientOrDie()->TransferFromDevice(device_data));
   std::vector<at::Tensor> tensors;
   for (auto& literal : literals) {
     tensors.push_back(MakeTensorFromXlaLiteral(

test/cpp/test_replication.cpp

@@ -79,9 +79,8 @@ void TestSingleReplication(
   counter.Wait();
 
   for (size_t i = 0; i < results.size(); ++i) {
-    std::vector<xla::Literal> literals =
-        torch_xla::runtime::GetComputationClientOrDie()->TransferFromDevice(
-            results[i]);
+    std::vector<xla::Literal> literals = GetValueOrThrow(
+        runtime::GetComputationClientOrDie()->TransferFromDevice(results[i]));
     ASSERT_EQ(literals.size(), 1);
 
     // The result must be the original tensor value, multiplied by the number of

test/test_operations.py

@@ -2458,6 +2458,14 @@ def test_add_broadcast_error(self):
       torch.add(a, b)
       torch_xla.sync()
 
+  def test_construct_large_tensor_raises_error(self):
+    a = torch.rand(1024, 1024, 1024, 1024, 1024, device=torch_xla.device())
+
+    # OOM is raised when we try to bring data from the device.
+    with self.assertRaisesRegex(RuntimeError, r"Out of memory allocating \d* bytes"):
+      b = a.sum()
+      b.cpu()
+
 
 class MNISTComparator(nn.Module):
 

torch_xla/csrc/init_python_bindings.cpp

@@ -1229,9 +1229,9 @@ class PyLoweringContext {
         lowering_ctx.GetParametersData();
 
     // Fetch this parameter data
-    std::vector<xla::Literal> literals =
+    std::vector<xla::Literal> literals = GetValueOrThrow(
         runtime::GetComputationClientOrDie()->TransferFromDevice(
-            UnwrapXlaData(device_data));
+            UnwrapXlaData(device_data)));
 
     // Create a mapping from paramater id to the tensor data
     std::unordered_map<int64_t, at::Tensor> results;

torch_xla/csrc/runtime/BUILD

@@ -123,6 +123,7 @@ cc_library(
         ":tf_logging",
         ":xla_coordinator",
         "//torch_xla/csrc:status",
+        "@com_google_absl//absl/log:absl_check",
         "@com_google_absl//absl/strings",
         "@com_google_absl//absl/synchronization",
         "@com_google_absl//absl/types:span",

torch_xla/csrc/runtime/computation_client.h

@@ -318,7 +318,7 @@ class ComputationClient {
   // Note: `TransferFromDevice` call will block until the `DataPtrs` are ready
   // if they were created by `TransferToDevice` or `Execute*`. Calling this from
   // python while holding the GIL can cause deadlocks!
-  virtual std::vector<xla::Literal> TransferFromDevice(
+  virtual absl::StatusOr<std::vector<xla::Literal>> TransferFromDevice(
       absl::Span<const DataPtr> handles) = 0;
 
   virtual std::uintptr_t UnsafeBufferPointer(const DataPtr handle) = 0;

torch_xla/csrc/runtime/ifrt_computation_client.cpp

@@ -436,8 +436,8 @@ std::shared_ptr<xla::PjRtBuffer> IfrtComputationClient::GetPjRtBuffer(
   XLA_ERROR() << __FUNCTION__ << " not implemented";
 }
 
-std::vector<xla::Literal> IfrtComputationClient::TransferFromDevice(
-    absl::Span<const DataPtr> handles) {
+absl::StatusOr<std::vector<xla::Literal>>
+IfrtComputationClient::TransferFromDevice(absl::Span<const DataPtr> handles) {
   metrics::TimedSection timed(TransferFromDeviceMetric());
   tsl::profiler::TraceMe activity("IfrtComputationClient::TransferFromDevice",
                                   tsl::profiler::TraceMeLevel::kInfo);
@@ -455,9 +455,9 @@ std::vector<xla::Literal> IfrtComputationClient::TransferFromDevice(
     auto& literal = literals.emplace_back(
         xla::ShapeUtil::DeviceShapeToHostShape(ifrt_data->shape()));
     std::vector<int64_t> byte_strides(literal.shape().dimensions_size());
-    XLA_CHECK_OK(xla::ShapeUtil::ByteStrides(literal.shape(),
-                                             absl::MakeSpan(byte_strides)));
-    XLA_CHECK_OK(
+    XLA_RETURN_IF_ERROR(xla::ShapeUtil::ByteStrides(
+        literal.shape(), absl::MakeSpan(byte_strides)));
+    XLA_RETURN_IF_ERROR(
         replicated_array
             ->CopyToHostBuffer(literal.untyped_data(), byte_strides,
                                xla::ifrt::ArrayCopySemantics::kAlwaysCopy)

torch_xla/csrc/runtime/ifrt_computation_client.h

@@ -62,7 +62,7 @@ class IfrtComputationClient : public ComputationClient {
     XLA_ERROR() << __FUNCTION__ << " not implemented";
   }
 
-  std::vector<xla::Literal> TransferFromDevice(
+  absl::StatusOr<std::vector<xla::Literal>> TransferFromDevice(
       absl::Span<const DataPtr> handles) override;
 
   std::uintptr_t UnsafeBufferPointer(const DataPtr handle) override;

torch_xla/csrc/runtime/ifrt_computation_client_test.cpp

@@ -70,7 +70,7 @@ TEST(PjRtComputationClientTest, Init) {
 
   // Copy the output from device back to host and assert correctness..
   ASSERT_EQ(results.size(), 1);
-  auto result_literals = client->TransferFromDevice(results);
+  auto result_literals = GetValueOrThrow(client->TransferFromDevice(results));
   ASSERT_THAT(result_literals, ::testing::SizeIs(1));
   EXPECT_TRUE(xla::LiteralTestUtil::Equal(
       xla::LiteralUtil::CreateR2<float>({{6.0f, 8.0f}, {10.0f, 12.0f}}),

torch_xla/csrc/runtime/pjrt_computation_client.cpp

@@ -4,6 +4,7 @@
 #include <stdexcept>
 #include <vector>
 
+#include "absl/log/absl_check.h"
 #include "absl/strings/ascii.h"
 #include "absl/synchronization/blocking_counter.h"
 #include "absl/types/span.h"
@@ -508,8 +509,8 @@ std::shared_ptr<xla::PjRtBuffer> PjRtComputationClient::GetPjRtBuffer(
   }
 }
 
-std::vector<xla::Literal> PjRtComputationClient::TransferFromDevice(
-    absl::Span<const DataPtr> handles) {
+absl::StatusOr<std::vector<xla::Literal>>
+PjRtComputationClient::TransferFromDevice(absl::Span<const DataPtr> handles) {
   metrics::TimedSection timed(TransferFromDeviceMetric());
   tsl::profiler::TraceMe activity("PjRtComputationClient::TransferFromDevice",
                                   tsl::profiler::TraceMeLevel::kInfo);
@@ -522,21 +523,21 @@ std::vector<xla::Literal> PjRtComputationClient::TransferFromDevice(
     // Use XLA replication to reassemble the sharded data. If input handle
     // is not sharded, then it is a no-op.
     std::shared_ptr<PjRtData> pjrt_data = ReplicateShardedData(handle);
-    XLA_CHECK(pjrt_data) << "PjRt_data is null in " << __FUNCTION__;
-    XLA_CHECK(pjrt_data->buffer != nullptr)
+    ABSL_CHECK(pjrt_data) << "PjRt_data is null in " << __FUNCTION__;
+    ABSL_CHECK(pjrt_data->buffer != nullptr)
         << "PjRt buffer is null in " << __FUNCTION__;
 
-    xla::Literal& literal =
-        literals.emplace_back(host_output_shape(pjrt_data->buffer.get()));
+    // Constructing a literal too large will make the whole program crash.
+    // Instead, we pass allocate_arrays=False, which makes this kind of
+    // error possible to be handled in the `Await()` call below.
+    xla::Literal& literal = literals.emplace_back(
+        xla::Literal(host_output_shape(pjrt_data->buffer.get()),
+                     /* allocate_arrays= */ false));
     futures.push_back(pjrt_data->buffer->ToLiteral(&literal));
 
     total_size += literal.size_bytes();
   }
-  for (auto& future : futures) {
-    absl::Status status = future.Await();
-    XLA_CHECK_OK(status) << "Failed to await future from buffer to literal in"
-                         << __FUNCTION__;
-  }
+  XLA_RETURN_IF_ERROR(xla::JoinFutures(futures).Await());
   InboundDataMetric()->AddSample(total_size);
 
   return literals;
@@ -773,10 +774,8 @@ PjRtComputationClient::ExecuteComputation(
 
   std::optional<xla::PjRtFuture<>> returned_future;
   std::vector<std::unique_ptr<xla::PjRtBuffer>> results =
-      pjrt_computation.executable
-          ->ExecuteSharded(buffers, pjrt_device, execute_options,
-                           returned_future)
-          .value();
+      GetValueOrThrow(pjrt_computation.executable->ExecuteSharded(
+          buffers, pjrt_device, execute_options, returned_future));
 
   returned_future->OnReady(std::move(
       [timed, op_tracker = std::move(op_tracker)](absl::Status unused) mutable {
@@ -878,10 +877,8 @@ PjRtComputationClient::ExecuteReplicated(
     tsl::profiler::TraceMe activity(
         "PjRtComputationClient::ExecuteReplicated_execute",
         tsl::profiler::TraceMeLevel::kInfo);
-    results = pjrt_computation.executable
-                  ->Execute(std::move(argument_handles), execute_options,
-                            returned_futures)
-                  .value();
+    results = GetValueOrThrow(pjrt_computation.executable->Execute(
+        std::move(argument_handles), execute_options, returned_futures));
 
     (*returned_futures)[0].OnReady(
         std::move([timed, op_tracker = std::move(op_tracker)](