update legion version (#1307)

* update legion version * legion version update * update legion version
flexflow · Aug 12, 2024 · 9784b5c · 9784b5c
1 parent 6a1a188
commit 9784b5c
Show file tree

Hide file tree

Showing 41 changed files with 59 additions and 50 deletions.
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -425,7 +425,7 @@ if(NOT BUILD_LEGION_ONLY)
       # generate the Legion Python bindings library. When building from pip, we need to do this post-install to prevent Legion from overwriting the path to the Legion shared library
       add_custom_command(TARGET flexflow
         POST_BUILD	
-        COMMAND ${Python_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/deps/legion/bindings/python/setup.py build --cmake-build-dir ${Legion_BINARY_DIR}/runtime --prefix ${Legion_BINARY_DIR} --build-lib=${Legion_BINARY_DIR}/bindings/python ${Legion_PYTHON_EXTRA_INSTALL_ARGS}
+        COMMAND CMAKE_BUILD_DIR=${Legion_BINARY_DIR}/runtime CMAKE_INSTALL_PREFIX=${Legion_BINARY_DIR} ${Python_EXECUTABLE} ${CMAKE_CURRENT_SOURCE_DIR}/deps/legion/bindings/python/setup.py build --build-lib=${Legion_BINARY_DIR}/bindings/python ${Legion_PYTHON_EXTRA_INSTALL_ARGS}
         WORKING_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/deps/legion/bindings/python
       )
       # create flexflow_python interpreter. When building from pip, we install the FF_HOME/python/flexflow_python script instead.

diff --git a/deps/legion b/deps/legion
diff --git a/examples/cpp/AlexNet/alexnet.cc b/examples/cpp/AlexNet/alexnet.cc
@@ -26,7 +26,7 @@ using FlexFlow::ParallelTensor;
 using FlexFlow::SGDOptimizer;
 using FlexFlow::Tensor;
 
-LegionRuntime::Logger::Category log_app("AlexNet");
+Legion::Logger log_app("AlexNet");
 
 void parse_input_args(char **argv, int argc, AlexNetConfig &config) {
   for (int i = 1; i < argc; i++) {

diff --git a/examples/cpp/DLRM/dlrm.cc b/examples/cpp/DLRM/dlrm.cc
@@ -19,7 +19,7 @@
 
 using namespace Legion;
 
-LegionRuntime::Logger::Category log_app("DLRM");
+Legion::Logger log_app("DLRM");
 
 void parse_input_args(char **argv, int argc, DLRMConfig &apConfig);
 

diff --git a/examples/cpp/InceptionV3/inception.cc b/examples/cpp/InceptionV3/inception.cc
@@ -21,7 +21,7 @@
 using namespace Legion;
 using namespace FlexFlow;
 
-LegionRuntime::Logger::Category log_app("Inceptionv3");
+Legion::Logger log_app("Inceptionv3");
 
 Tensor InceptionA(FFModel &ff, Tensor input, int pool_features) {
   Tensor t1 = input;

diff --git a/examples/cpp/ResNet/resnet.cc b/examples/cpp/ResNet/resnet.cc
@@ -24,7 +24,7 @@ using FlexFlow::Optimizer;
 using FlexFlow::SGDOptimizer;
 using FlexFlow::Tensor;
 
-LegionRuntime::Logger::Category log_app("ResNet");
+Legion::Logger log_app("ResNet");
 
 void parse_input_args(char **argv, int argc, ResNetConfig &config) {
   for (int i = 1; i < argc; i++) {

diff --git a/examples/cpp/Transformer/transformer.cc b/examples/cpp/Transformer/transformer.cc
@@ -17,7 +17,7 @@
 
 using namespace Legion;
 
-LegionRuntime::Logger::Category log_app("Transformer");
+Legion::Logger log_app("Transformer");
 
 Tensor create_emb(FFModel *model,
                   Tensor const &input,

diff --git a/examples/cpp/XDL/xdl.cc b/examples/cpp/XDL/xdl.cc
@@ -18,7 +18,7 @@
 
 using namespace Legion;
 
-LegionRuntime::Logger::Category log_app("XDL");
+Legion::Logger log_app("XDL");
 
 void parse_input_args(char **argv, int argc, XDLConfig &apConfig);
 

diff --git a/examples/cpp/candle_uno/candle_uno.cc b/examples/cpp/candle_uno/candle_uno.cc
@@ -21,7 +21,7 @@
 using namespace Legion;
 using namespace std;
 
-LegionRuntime::Logger::Category log_app("Candle_Uno");
+Legion::Logger log_app("Candle_Uno");
 
 void parse_input_args(char **argv, int argc, CandleConfig &apConfig);
 

diff --git a/examples/cpp/mixture_of_experts/moe.cc b/examples/cpp/mixture_of_experts/moe.cc
@@ -20,7 +20,7 @@
 
 using namespace Legion;
 
-LegionRuntime::Logger::Category log_app("MoE");
+Legion::Logger log_app("MoE");
 
 void parse_input_args(char **argv, int argc, MoeConfig &config) {
   for (int i = 1; i < argc; i++) {

diff --git a/examples/cpp/resnext50/resnext.cc b/examples/cpp/resnext50/resnext.cc
@@ -7,7 +7,7 @@ using FlexFlow::Optimizer;
 using FlexFlow::SGDOptimizer;
 using FlexFlow::Tensor;
 
-LegionRuntime::Logger::Category log_app("resnext");
+Legion::Logger log_app("resnext");
 
 Tensor resnext_block(FFModel &ff,
                      Tensor input,

diff --git a/examples/cpp/split_test/split_test.cc b/examples/cpp/split_test/split_test.cc
@@ -3,7 +3,7 @@
 using namespace Legion;
 using namespace FlexFlow;
 
-LegionRuntime::Logger::Category log_app("split_test");
+Legion::Logger log_app("split_test");
 
 void FlexFlow::top_level_task(Task const *task,
                               std::vector<PhysicalRegion> const &regions,

diff --git a/examples/cpp/split_test_2/split_test_2.cc b/examples/cpp/split_test_2/split_test_2.cc
@@ -9,7 +9,7 @@ using FlexFlow::PCG::Graph;
 using FlexFlow::PCG::GraphSearchHelper;
 using FlexFlow::PCG::Node;
 
-LegionRuntime::Logger::Category log_app("split_test_2");
+Legion::Logger log_app("split_test_2");
 
 void top_level_task(Task const *task,
                     std::vector<PhysicalRegion> const &regions,

diff --git a/include/flexflow/graph.h b/include/flexflow/graph.h
@@ -24,7 +24,7 @@
 #include "legion/legion_utilities.h"
 #include <unordered_set>
 
-extern LegionRuntime::Logger::Category log_dp;
+extern Legion::Logger log_dp;
 
 namespace FlexFlow::PCG {
 

diff --git a/include/flexflow/operator.h b/include/flexflow/operator.h
@@ -19,7 +19,7 @@
 
 namespace FlexFlow {
 
-extern LegionRuntime::Logger::Category log_measure;
+extern Legion::Logger log_measure;
 
 class OpMeta;
 class Simulator;
@@ -233,6 +233,8 @@ class Op {
                                       std::vector<ParallelTensor> const &,
                                       MachineView const *mv = nullptr) {
     assert(false);
+    Legion::FutureMap empty_map;
+    return empty_map;
   };
   virtual void print_layer(FFModel const &model) = 0;
   template <typename OpMetaType>

diff --git a/include/flexflow/utils/recursive_logger.h b/include/flexflow/utils/recursive_logger.h
@@ -26,7 +26,7 @@ class DepthTag {
 
 class RecursiveLogger {
 public:
-  /* RecursiveLogger(LegionRuntime::Logger::Category const &); */
+  /* RecursiveLogger(Legion::Logger const &); */
   RecursiveLogger(std::string const &category_name);
 
   Realm::LoggerMessage info();
@@ -42,7 +42,7 @@ class RecursiveLogger {
 
   void print_prefix(Realm::LoggerMessage &) const;
 
-  LegionRuntime::Logger::Category logger;
+  Legion::Logger logger;
 };
 
 };     // namespace FlexFlow

diff --git a/inference/incr_decoding/incr_decoding.cc b/inference/incr_decoding/incr_decoding.cc
@@ -28,7 +28,7 @@ using namespace FlexFlow;
 using namespace Legion;
 using json = nlohmann::json;
 
-LegionRuntime::Logger::Category log_app("llama");
+Legion::Logger log_app("llama");
 
 struct FilePaths {
   std::string cache_folder_path;

diff --git a/inference/spec_infer/spec_infer.cc b/inference/spec_infer/spec_infer.cc
@@ -26,7 +26,7 @@ using namespace FlexFlow;
 using namespace Legion;
 using json = nlohmann::json;
 
-LegionRuntime::Logger::Category log_app("llama");
+Legion::Logger log_app("llama");
 
 struct FilePaths {
   std::string cache_folder_path;

diff --git a/src/mapper/mapper.cc b/src/mapper/mapper.cc
@@ -20,7 +20,7 @@ namespace FlexFlow {
 using namespace Legion;
 using namespace Mapping;
 
-LegionRuntime::Logger::Category log_ff_mapper("Mapper");
+Legion::Logger log_ff_mapper("Mapper");
 
 FFShardingFunctor::FFShardingFunctor(int _gpus_per_node,
                                      int _cpus_per_node,
@@ -296,6 +296,7 @@ void FFMapper::select_task_options(const MapperContext ctx,
     // control replicate top level task
     if (enable_control_replication) {
       output.replicate = true;
+      output.map_locally = false;
     }
     return;
   }
@@ -560,6 +561,10 @@ void FFMapper::map_task(const MapperContext ctx,
       assert(output.target_procs[i].address_space() == node_id);
     }
   }
+  if (input.shard_processor.exists()) {
+    output.target_procs = std::vector<Processor>{input.shard_processor};
+  }
+
   // Find instances that still need to be mapped
   std::vector<std::set<FieldID>> missing_fields(task.regions.size());
   runtime->filter_instances(ctx,

diff --git a/src/ops/beam_topk.cpp b/src/ops/beam_topk.cpp
@@ -25,7 +25,7 @@ using Legion::coord_t;
 enum class HeapType { kMinHeap, kMaxHeap };
 enum class PreferIndices { kLower, kHigher };
 
-LegionRuntime::Logger::Category log_beam_topk("BeamTopK");
+Legion::Logger log_beam_topk("BeamTopK");
 
 template <typename T>
 struct Entry {

diff --git a/src/ops/beam_topk.cu b/src/ops/beam_topk.cu
@@ -25,7 +25,7 @@ using Legion::coord_t;
 enum class HeapType { kMinHeap, kMaxHeap };
 enum class PreferIndices { kLower, kHigher };
 
-LegionRuntime::Logger::Category log_beam_topk("BeamTopK");
+Legion::Logger log_beam_topk("BeamTopK");
 
 template <typename T>
 struct Entry {

diff --git a/src/ops/inc_multihead_self_attention.cc b/src/ops/inc_multihead_self_attention.cc
@@ -46,7 +46,7 @@ using Legion::TaskArgument;
 using Legion::TaskLauncher;
 using PCG::Node;
 
-LegionRuntime::Logger::Category log_inc_mha("IncrementalMHA");
+Legion::Logger log_inc_mha("IncrementalMHA");
 
 bool IncMultiHeadSelfAttentionParams::is_valid(
     ParallelTensorShape const &input) const {

diff --git a/src/ops/tree_inc_multihead_self_attention.cc b/src/ops/tree_inc_multihead_self_attention.cc
@@ -46,7 +46,7 @@ using Legion::TaskArgument;
 using Legion::TaskLauncher;
 using PCG::Node;
 
-LegionRuntime::Logger::Category log_tree_verify("TreeVerifyIncMHA");
+Legion::Logger log_tree_verify("TreeVerifyIncMHA");
 
 bool TreeIncMultiHeadSelfAttentionParams::is_valid(
     ParallelTensorShape const &input) const {

diff --git a/src/runtime/batch_config.cc b/src/runtime/batch_config.cc
@@ -21,7 +21,7 @@
 
 namespace FlexFlow {
 
-LegionRuntime::Logger::Category log_bc("BatchConfig");
+Legion::Logger log_bc("BatchConfig");
 using Legion::Future;
 using Legion::Memory;
 

diff --git a/src/runtime/beam_search_batch_config.cc b/src/runtime/beam_search_batch_config.cc
@@ -24,7 +24,7 @@
 
 namespace FlexFlow {
 
-LegionRuntime::Logger::Category log_beam_bc("BeamSearchBatchConfig");
+Legion::Logger log_beam_bc("BeamSearchBatchConfig");
 
 BeamSearchBatchConfig::BeamSearchBatchConfig() : BatchConfig() {
   this->beam_width = DEFAULT_BEAM_WIDTH;

diff --git a/src/runtime/graph.cc b/src/runtime/graph.cc
@@ -66,8 +66,8 @@ namespace FlexFlow::PCG {
 using namespace Legion;
 using FlexFlow::MachineView;
 
-LegionRuntime::Logger::Category log_graph("graph");
-LegionRuntime::Logger::Category log_simplify("graph_simplify");
+Legion::Logger log_graph("graph");
+Legion::Logger log_simplify("graph_simplify");
 
 const Node Node::INVALID_NODE = Node();
 

diff --git a/src/runtime/inference_manager.cc b/src/runtime/inference_manager.cc
@@ -25,8 +25,8 @@ namespace FlexFlow {
 
 using namespace Legion;
 
-LegionRuntime::Logger::Category log_inf_mgr("InferenceManager");
-LegionRuntime::Logger::Category log_offload("Offloading");
+Legion::Logger log_inf_mgr("InferenceManager");
+Legion::Logger log_offload("Offloading");
 
 InferenceManager::InferenceManager() {}
 

diff --git a/src/runtime/model.cc b/src/runtime/model.cc
@@ -82,8 +82,8 @@ namespace FlexFlow {
 
 using namespace Legion;
 
-LegionRuntime::Logger::Category log_model("Model");
-LegionRuntime::Logger::Category log_measure("measure");
+Legion::Logger log_model("Model");
+Legion::Logger log_measure("measure");
 
 Op::Op(FFModel &model,
        OperatorType otype,
@@ -6748,6 +6748,7 @@ void register_flexflow_internal_tasks(Runtime *runtime,
     TaskVariantRegistrar registrar(SGD_UPD_NCCL_TASK_ID, "SGD NCCL Update");
     registrar.add_constraint(ProcessorConstraint(Processor::TOC_PROC));
     registrar.set_leaf();
+    registrar.set_concurrent();
     if (pre_register) {
       Runtime::preregister_task_variant<SGDOptimizer::nccl_update_task>(
           registrar, "SGD NCCL Update Task");
@@ -6898,6 +6899,7 @@ void register_flexflow_internal_tasks(Runtime *runtime,
                                    "NCCL Init Communicators");
     registrar.add_constraint(ProcessorConstraint(Processor::TOC_PROC));
     registrar.set_leaf();
+    registrar.set_concurrent();
     if (pre_register) {
       Runtime::preregister_task_variant<ncclComm_t, Op::init_nccl_comms_task>(
           registrar, "NCCL Init Communicators Task");

diff --git a/src/runtime/optimizer_kernel.cpp b/src/runtime/optimizer_kernel.cpp
@@ -21,7 +21,7 @@
 
 namespace FlexFlow {
 
-LegionRuntime::Logger::Category log_optimizer("optimizer");
+Legion::Logger log_optimizer("optimizer");
 
 __global__ void sgd_update(size_t count,
                            float lr,
@@ -247,4 +247,4 @@ __host__ void AdamOptimizer::nccl_update_task_gpu(AdamOptimizer const *op,
 }
 #endif
 
-}; // namespace FlexFlow
+}; // namespace FlexFlow
diff --git a/src/runtime/optimizer_kernel.cu b/src/runtime/optimizer_kernel.cu
@@ -20,7 +20,7 @@
 
 namespace FlexFlow {
 
-LegionRuntime::Logger::Category log_optimizer("optimizer");
+Legion::Logger log_optimizer("optimizer");
 
 __global__ void sgd_update(size_t count,
                            float lr,

diff --git a/src/runtime/request_manager.cc b/src/runtime/request_manager.cc
@@ -29,7 +29,7 @@ namespace FlexFlow {
 using namespace Legion;
 using tokenizers::Tokenizer;
 
-LegionRuntime::Logger::Category log_req_mgr("RequestManager");
+Legion::Logger log_req_mgr("RequestManager");
 
 std::string LoadBytesFromFile(std::string const &path) {
   std::ifstream fs(path, std::ios::in | std::ios::binary);

diff --git a/src/runtime/simulator.cc b/src/runtime/simulator.cc
@@ -31,10 +31,10 @@ namespace FlexFlow {
 
 using namespace Legion;
 
-LegionRuntime::Logger::Category log_sim("sim");
-LegionRuntime::Logger::Category log_ps_sim("ps_sim");
-LegionRuntime::Logger::Category log_xfer_sim("xfer_sim");
-LegionRuntime::Logger::Category log_xfer_est("xfer_est");
+Legion::Logger log_sim("sim");
+Legion::Logger log_ps_sim("ps_sim");
+Legion::Logger log_xfer_sim("xfer_sim");
+Legion::Logger log_xfer_est("xfer_est");
 
 // template class std::map<const Op*, ParallelConfig>; // for debugging in gdb
 // template class std::map<const Op*, MachineView>; // for debugging in gdb

diff --git a/src/runtime/substitution.cc b/src/runtime/substitution.cc
@@ -54,8 +54,8 @@ namespace FlexFlow::PCG {
 
 using namespace Legion;
 
-LegionRuntime::Logger::Category log_xfers("xfers");
-LegionRuntime::Logger::Category log_xfer_matches("xfer_matches");
+Legion::Logger log_xfers("xfers");
+Legion::Logger log_xfer_matches("xfer_matches");
 
 const TensorX TensorX::NO_TX = TensorX();
 

diff --git a/src/runtime/tree_verify_batch_config.cc b/src/runtime/tree_verify_batch_config.cc
@@ -21,7 +21,7 @@
 
 namespace FlexFlow {
 
-LegionRuntime::Logger::Category log_tree_bc("TreeVerifyBatchConfig");
+Legion::Logger log_tree_bc("TreeVerifyBatchConfig");
 
 TreeVerifyBatchConfig::TreeVerifyBatchConfig() : BatchConfig() {}
 

diff --git a/tests/ops/batch_matmul_test.cc b/tests/ops/batch_matmul_test.cc
@@ -5,7 +5,7 @@
 #include <iostream>
 #include <sstream>
 using namespace Legion;
-LegionRuntime::Logger::Category log_app("bmm_test");
+Legion::Logger log_app("bmm_test");
 
 struct BMMTestMeta {
   int m, k, n, d;

diff --git a/tests/ops/concat_test.cc b/tests/ops/concat_test.cc
@@ -5,7 +5,7 @@
 #include <iostream>
 #include <sstream>
 using namespace Legion;
-LegionRuntime::Logger::Category log_app("concat_test");
+Legion::Logger log_app("concat_test");
 
 struct ConcatTestMeta {
   int batch_size, i_dim, num_channels, projected_num_channels,