fix

flexflow · Nov 27, 2024 · 98e025c · 98e025c
1 parent 450c98f
commit 98e025c
Show file tree

Hide file tree

Showing 5 changed files with 9 additions and 8 deletions.
diff --git a/benchmarking/debug.sh b/benchmarking/debug.sh
@@ -26,14 +26,15 @@ export FF_DEBG_NO_WEIGHTS=1
 gdb -ex run --args ./inference/incr_decoding/incr_decoding \
     -ll:cpu $NCPUS -ll:gpu $NGPUS -ll:util $NCPUS \
     -ll:fsize 20000 -ll:zsize 10000 \
-    --verbose -lg:prof 1 -lg:prof_logfile prof_%.gz \
-    -llm-model $MODEL_NAME \
+    -llm-model $MODEL_NAME --verbose \
     -prompt $PROMPT \
     -tensor-parallelism-degree $NGPUS \
     -log-file ../inference/output/test.out \
     -output-file ../inference/output/test.json \
     --max-requests-per-batch 1 --max-tokens-per-batch 3000 --max-sequence-length 3000
 
+#--verbose -lg:prof 1 -lg:prof_logfile prof_%.gz \
+
 # ./inference/peft/peft \
 #     -ll:cpu 4 -ll:gpu $NGPUS -ll:util 2 \
 #     -ll:fsize 10000 -ll:zsize 10000 \

diff --git a/deps/legion b/deps/legion
diff --git a/include/flexflow/request_manager.h b/include/flexflow/request_manager.h
@@ -341,7 +341,7 @@ class RequestManager {
   //     std::vector<Legion::PhysicalRegion> const &regions,
   //     Legion::Context ctx,
   //     Legion::Runtime *runtime);
-  static void process_work_from_old_batches_task(
+  static bool process_work_from_old_batches_task(
       Legion::Task const *task,
       std::vector<Legion::PhysicalRegion> const &regions,
       Legion::Context ctx,

diff --git a/src/runtime/model.cc b/src/runtime/model.cc
@@ -4690,14 +4690,14 @@ void register_flexflow_internal_tasks(Runtime *runtime,
     registrar.add_constraint(ProcessorConstraint(Processor::LOC_PROC));
     registrar.set_leaf();
     if (pre_register) {
-      Runtime::preregister_task_variant<
+      Runtime::preregister_task_variant<bool,
           RequestManager::process_work_from_old_batches_task>(
           registrar, "RequestManager Process Work from Old Batches Task");
     } else {
       if (enable_control_replication) {
         registrar.global_registration = false;
       }
-      runtime->register_task_variant<
+      runtime->register_task_variant<bool,
           RequestManager::process_work_from_old_batches_task>(registrar);
     }
   }

diff --git a/src/runtime/request_manager.cc b/src/runtime/request_manager.cc
@@ -672,15 +672,14 @@ std::pair<BatchConfigFuture, BatchConfigFuture>
   launcher3.add_future(std::get<3>(batch_pipeline_entry));
   launcher3.add_future(pwfobf);
   BatchConfigFuture bcbf = runtime->execute_task(ctx, launcher3);
-  // return pair of batch futures
   return std::make_pair(bcff, bcbf);
 }
 
 // future[0]: old_fwd_bc
 // future[1]: old_bwd_bc
 // future[2]: inference result
 // future[3]: wait for bwd to finish
-void RequestManager::process_work_from_old_batches_task(
+bool RequestManager::process_work_from_old_batches_task(
     Task const *task,
     std::vector<PhysicalRegion> const &regions,
     Context ctx,
@@ -693,6 +692,7 @@ void RequestManager::process_work_from_old_batches_task(
       Future(task->futures[2]).get_result<InferenceResult>();
   Future(task->futures[3]).get_void_result(); // wait until bwd is done
   rm->process_work_from_old_batches(*old_fwd_bc, *old_bwd_bc, result);
+  return true;
 }
 
 // future[0]: old_fwd_bc