6983bacea1
Enabling this logic removes cross-worker send/recv dependencies required for TPUExecuteOp nodes to access a model's variables. This decreases overhead at the start of a training loop. The approach used is to replace remote variable reads with zero tensors on each worker, except for the primary worker. The zero tensors feed TPUExecute nodes that are local to that worker. For large distributed systems with large variables, this removes the need for the initial Send/Recv variable broadcast, which can be expensive. PiperOrigin-RevId: 351904109 Change-Id: I9f1ed63c2401f227646010a94a70c04f1c96cb7e
126 lines
5.0 KiB
C++
126 lines
5.0 KiB
C++
/* Copyright 2017 The TensorFlow Authors. All Rights Reserved.
|
|
|
|
Licensed under the Apache License, Version 2.0 (the "License");
|
|
you may not use this file except in compliance with the License.
|
|
You may obtain a copy of the License at
|
|
|
|
http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
Unless required by applicable law or agreed to in writing, software
|
|
distributed under the License is distributed on an "AS IS" BASIS,
|
|
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
See the License for the specific language governing permissions and
|
|
limitations under the License.
|
|
==============================================================================*/
|
|
|
|
#include "tensorflow/compiler/xla/service/compile_only_service.h"
|
|
|
|
#include <string>
|
|
#include <utility>
|
|
#include <vector>
|
|
|
|
#include "absl/strings/str_cat.h"
|
|
#include "tensorflow/compiler/xla/debug_options_flags.h"
|
|
#include "tensorflow/compiler/xla/service/backend.h"
|
|
#include "tensorflow/compiler/xla/service/computation_layout.h"
|
|
#include "tensorflow/compiler/xla/service/dump.h"
|
|
#include "tensorflow/compiler/xla/service/platform_util.h"
|
|
#include "tensorflow/compiler/xla/status_macros.h"
|
|
#include "tensorflow/compiler/xla/types.h"
|
|
#include "tensorflow/compiler/xla/util.h"
|
|
#include "tensorflow/core/lib/gtl/cleanup.h"
|
|
#include "tensorflow/core/lib/io/path.h"
|
|
#include "tensorflow/core/platform/host_info.h"
|
|
#include "tensorflow/core/platform/logging.h"
|
|
#include "tensorflow/core/platform/stream_executor_no_cuda.h"
|
|
|
|
namespace xla {
|
|
|
|
/* static */ StatusOr<std::unique_ptr<CompileOnlyService>>
|
|
CompileOnlyService::NewService(se::Platform* platform) {
|
|
ServiceOptions default_options;
|
|
default_options.set_platform(platform);
|
|
return NewService(default_options);
|
|
}
|
|
|
|
/* static */ StatusOr<std::unique_ptr<CompileOnlyService>>
|
|
CompileOnlyService::NewService(const ServiceOptions& options) {
|
|
se::Platform* platform = options.platform();
|
|
if (platform == nullptr) {
|
|
TF_ASSIGN_OR_RETURN(platform, PlatformUtil::GetDefaultPlatform());
|
|
}
|
|
|
|
TF_ASSIGN_OR_RETURN(auto compiler, Compiler::GetForPlatform(platform));
|
|
|
|
std::unique_ptr<CompileOnlyService> service(
|
|
new CompileOnlyService(options, compiler));
|
|
return std::move(service);
|
|
}
|
|
|
|
CompileOnlyService::CompileOnlyService(const ServiceOptions& options,
|
|
Compiler* compiler)
|
|
: Service(options, /*execute_backend=*/nullptr), compiler_(compiler) {}
|
|
|
|
StatusOr<std::vector<std::unique_ptr<AotCompilationResult>>>
|
|
CompileOnlyService::CompileAheadOfTime(
|
|
const absl::Span<const AotXlaComputationInstance> computations,
|
|
const AotCompilationOptions& options,
|
|
std::unique_ptr<AotCompilationMetadata>* metadata) {
|
|
std::vector<std::unique_ptr<HloModule>> hlo_modules;
|
|
|
|
const DebugOptions& debug_options = options.debug_options();
|
|
ExecutionOptions execution_options;
|
|
*execution_options.mutable_debug_options() = debug_options;
|
|
// Capture replica_count, num_cores, and device_assignment in ExecutionOptions
|
|
// to later save in a proto dump.
|
|
if (options.replica_count() > 0) {
|
|
execution_options.set_num_replicas(options.replica_count());
|
|
if (options.has_static_device_assignment()) {
|
|
CHECK_EQ(options.replica_count(),
|
|
options.static_device_assignment().replica_count());
|
|
}
|
|
}
|
|
if (options.num_cores() > 0) {
|
|
execution_options.set_num_partitions(options.num_cores());
|
|
if (options.has_static_device_assignment()) {
|
|
CHECK_EQ(options.num_cores(),
|
|
options.static_device_assignment().computation_count());
|
|
}
|
|
}
|
|
if (options.has_static_device_assignment()) {
|
|
TF_RETURN_IF_ERROR(options.static_device_assignment().Serialize(
|
|
execution_options.mutable_device_assignment()));
|
|
}
|
|
execution_options.set_use_spmd_partitioning(options.use_spmd_partitioning());
|
|
execution_options.set_deduplicate_hlo(options.deduplicate_hlo());
|
|
execution_options.set_broadcast_replicated_parameters_via_collectives(
|
|
options.broadcast_replicated_params());
|
|
for (const AotXlaComputationInstance& instance : computations) {
|
|
TF_RET_CHECK(instance.computation.has_host_program_shape());
|
|
*execution_options.mutable_shape_with_output_layout() =
|
|
instance.result_layout->ToProto();
|
|
|
|
TF_ASSIGN_OR_RETURN(
|
|
std::unique_ptr<HloModuleConfig> module_config,
|
|
CreateModuleConfig(
|
|
ProgramShape(instance.computation.host_program_shape()),
|
|
instance.argument_layouts, &execution_options, &options));
|
|
|
|
TF_ASSIGN_OR_RETURN(
|
|
std::unique_ptr<HloModule> hlo_module,
|
|
HloModule::CreateFromProto(instance.computation, *module_config));
|
|
DumpHloModuleIfEnabled(*hlo_module, "before_optimizations");
|
|
hlo_modules.push_back(std::move(hlo_module));
|
|
}
|
|
|
|
execution_options.clear_shape_with_output_layout();
|
|
DumpExecutionOptions(execution_options, debug_options);
|
|
|
|
return compiler_->CompileAheadOfTime(
|
|
absl::make_unique<HloModuleGroup>(hlo_modules[0]->name(),
|
|
absl::MakeSpan(hlo_modules)),
|
|
options, metadata);
|
|
}
|
|
|
|
} // namespace xla
|