diff --git a/Source/1BitSGD/BlockMomentumDistributedLearner.h b/Source/1BitSGD/BlockMomentumDistributedLearner.h
index 1585bd4c2588..d0366cbac94e 100644
--- a/Source/1BitSGD/BlockMomentumDistributedLearner.h
+++ b/Source/1BitSGD/BlockMomentumDistributedLearner.h
@@ -524,38 +524,40 @@ namespace CNTK
                     ResetBuffer<double>(i, p);
                 else if (p->GetDataType() == DataType::Float)
                     ResetBuffer<float>(i, p);
+                else if (p->GetDataType() == DataType::Float16)
+                    ResetBuffer<half, float16>(i, p);
                 else
                     RuntimeError("Unsupported type.");
             }
         }
 
-        template<class ElemType>
+        template<class ElemTypeV1, class ElemTypeV2=ElemTypeV1>
         void ResetBuffer(size_t index, const NDArrayViewPtr& p)
         {
-            auto data = p->GetMatrix<ElemType>();
+            auto data = p->GetMatrix<ElemTypeV1>();
             if (!m_blockLevelSmoothedGradient[index])
             {
                 // has not been initialized yet
-                auto pSmoothedGrad = std::make_shared<NDArrayView>(AsDataType<ElemType>(), p->Shape(), AsDeviceDescriptor(data->GetDeviceId()));
-                pSmoothedGrad->SetValue(static_cast<ElemType>(0));
+                auto pSmoothedGrad = std::make_shared<NDArrayView>(AsDataType<ElemTypeV2>(), p->Shape(), AsDeviceDescriptor(data->GetDeviceId()));
+                pSmoothedGrad->SetValue(static_cast<ElemTypeV2>(0));
                 m_blockLevelSmoothedGradient[index] = pSmoothedGrad;
             }
 
             if (!m_prevParameters[index])
             {
-                NDArrayViewPtr newValue = std::make_shared<NDArrayView>(AsDataType<ElemType>(), p->Shape(), AsDeviceDescriptor(data->GetDeviceId()));
-                std::shared_ptr<Matrix<ElemType>> newData = newValue->GetWritableMatrix<ElemType>();
+                NDArrayViewPtr newValue = std::make_shared<NDArrayView>(AsDataType<ElemTypeV2>(), p->Shape(), AsDeviceDescriptor(data->GetDeviceId()));
+                std::shared_ptr<Matrix<ElemTypeV1>> newData = newValue->GetWritableMatrix<ElemTypeV1>();
                 newData->SetValue(*data);
                 m_prevParameters[index] = newValue;
             }
             else
             {
-                m_prevParameters[index]->GetWritableMatrix<ElemType>()->SetValue(*data);
+                m_prevParameters[index]->GetWritableMatrix<ElemTypeV1>()->SetValue(*data);
             }
 
             if (!m_tempBlockGradient[index])
             {
-                m_tempBlockGradient[index] = std::make_shared<NDArrayView>(AsDataType<ElemType>(), p->Shape(), AsDeviceDescriptor(data->GetDeviceId()));
+                m_tempBlockGradient[index] = std::make_shared<NDArrayView>(AsDataType<ElemTypeV2>(), p->Shape(), AsDeviceDescriptor(data->GetDeviceId()));
             }
         }
 
diff --git a/Source/CNTKv2LibraryDll/DistributedCommunicator.cpp b/Source/CNTKv2LibraryDll/DistributedCommunicator.cpp
index f37394aab64b..0cabaef81374 100644
--- a/Source/CNTKv2LibraryDll/DistributedCommunicator.cpp
+++ b/Source/CNTKv2LibraryDll/DistributedCommunicator.cpp
@@ -14,6 +14,7 @@
 #include "GPUDataTransferer.h"
 #include <numeric>
 #include "Utils.h"
+#include<iostream>
 
 using namespace Microsoft::MSR::CNTK;
 
@@ -732,6 +733,7 @@ namespace CNTK
     {
         if (m_nccl->IsSupported() && !dataOnCPU)
         {
+            std::cerr << " NCCL fp16 allreduce" << endl;
             m_nccl->AllReduce(inputData, outputData, numElements, op);
 
             return;
diff --git a/Source/Common/Include/Sequences.h b/Source/Common/Include/Sequences.h
index 2571fd6f5499..d6e330e6f63f 100644
--- a/Source/Common/Include/Sequences.h
+++ b/Source/Common/Include/Sequences.h
@@ -546,7 +546,8 @@ struct MBLayout
             LogicError("GetColumnIndex: t out of sequence bounds.");
         if (seq.s > GetNumParallelSequences())
             LogicError("GetColumnIndex: seq.s out of sequence bounds."); // can only happen if 'seq' does not come out of our own m_sequences array, which is verboten
-        ptrdiff_t tIn = (ptrdiff_t)t + seq.tBegin;       // shifted time index
+        //ptrdiff_t tIn = (ptrdiff_t)t + seq.tBegin;       // shifted time index
+        ptrdiff_t tIn = (ptrdiff_t)t + (seq.tBegin > 0 ?  seq.tBegin : 0 );       // shifted time index
         if (tIn < 0 || (size_t)tIn >= GetNumTimeSteps())
             LogicError("GetColumnIndex: Attempted to access a time step that is accessing a portion of a sequence that is not included in current minibatch."); // we may encounter this for truncated BPTT
         size_t col = (size_t)tIn * GetNumParallelSequences() + seq.s;
diff --git a/bindings/python/cntk/logging/progress_print.py b/bindings/python/cntk/logging/progress_print.py
index 4d2649fc501c..199e2b9c8f44 100644
--- a/bindings/python/cntk/logging/progress_print.py
+++ b/bindings/python/cntk/logging/progress_print.py
@@ -220,6 +220,8 @@ def write(self, key, value):
     def ___logprint(self, logline):
         if self.log_to_file == None:
             # to stdout.  if distributed, all ranks merge output into stdout
+            t = time.localtime()
+            print(str(t.tm_year) + '-' + str(t.tm_mon) + '-' + str(t.tm_mday) + ' ' + str(t.tm_hour) + ':' + str(t.tm_min) + ':' + str(t.tm_sec), end = ' ')
             print(logline)
         else:
             # to named file.  if distributed, one file per rank