From 093b29d3ed0e159bf1702de024370ee24b92a755 Mon Sep 17 00:00:00 2001
From: Zhihao Jia <zhihao@cmu.edu>
Date: Tue, 2 Jan 2024 18:50:49 -0500
Subject: [PATCH] bug fixes and update Legion version

---
 deps/legion                    | 2 +-
 src/ops/linear.cc              | 8 ++------
 src/runtime/model.cu           | 1 -
 src/runtime/request_manager.cc | 7 ++-----
 4 files changed, 5 insertions(+), 13 deletions(-)

diff --git a/deps/legion b/deps/legion
index 626b55689c..d065278678 160000
--- a/deps/legion
+++ b/deps/legion
@@ -1 +1 @@
-Subproject commit 626b55689c77848b246e1da19678c7ad58899f0c
+Subproject commit d0652786784249e933dd62f675591da99a5e960d
diff --git a/src/ops/linear.cc b/src/ops/linear.cc
index 63b26bfe7d..93e93a5953 100644
--- a/src/ops/linear.cc
+++ b/src/ops/linear.cc
@@ -467,12 +467,8 @@ OpMeta *Linear::init_task_with_dim(Task const *task,
                                        ctx,
                                        runtime,
                                        false /*readOutput*/);
-  TensorAccessorW<WT, NDIM> acc_kernel(regions[2],
-                                       task->regions[2],
-                                       FID_DATA,
-                                       ctx,
-                                       runtime,
-                                       false /*readOutput*/);
+  TensorAccessorR<WT, NDIM> acc_kernel(
+      regions[2], task->regions[2], FID_DATA, ctx, runtime);
 
   // TensorAccessorR<float, 1> acc_bias(
   //     regions[3], task->regions[3], FID_DATA, ctx, runtime);
diff --git a/src/runtime/model.cu b/src/runtime/model.cu
index c885b29db2..23b7f0efbe 100644
--- a/src/runtime/model.cu
+++ b/src/runtime/model.cu
@@ -175,7 +175,6 @@ FFHandler
   } else {
     handle.batch_config_metadata = nullptr;
   }
-   
 
   // checkCUDA(cudaMalloc(&handle.workSpace, handle.workSpaceSize));
 #ifdef FF_USE_NCCL
diff --git a/src/runtime/request_manager.cc b/src/runtime/request_manager.cc
index 88754f5a82..a285932b7f 100644
--- a/src/runtime/request_manager.cc
+++ b/src/runtime/request_manager.cc
@@ -1188,10 +1188,7 @@ BeamSearchBatchConfig
       int ssm_decoding_steps =
           profiling_requests[request.guid].ssm_decoding_steps;
 
-      new_bc.beamRequestsInfo[i].beam_size =
-          spec_infer_tree_width.size() > ssm_decoding_steps
-              ? spec_infer_tree_width[ssm_decoding_steps]
-              : 1;
+      new_bc.beamRequestsInfo[i].beam_size = 1;
       // printf("beam size: %d, %d\n",
       //        new_bc.beamRequestsInfo[i].beam_size,
       //        ssm_decoding_steps);
@@ -1820,7 +1817,7 @@ void RequestManager::updateBitMask(BatchConfig::BitMask &bitmask,
 void RequestManager::appendPendingRequest(BatchConfig::BitMask &bitmask,
                                           int initLength) {
   assert(initLength > 0);
-  std::cout << "append pending bit mask: " << initLength << "\n";
+  // std::cout << "append pending bit mask: " << initLength << "\n";
   // eg. 4 tokens: t1: 0000000..1111, t2: 0000000..1110, t3: 0000000..1100, t4:
   // 0000000..1000
   bitmask.non_tree_cache_size = 0;