Lint

Author: nsarka

csrc/host_ir/evaluator.cpp

@@ -376,7 +376,8 @@ void HostIrEvaluator::handle(Communication* communication) {
          communication->type() == CommunicationType::Reduce)
         ? root_val
         : -1;
-    // For Broadcast and Reduce, non-roots may have no output; use input for cache key
+    // For Broadcast and Reduce, non-roots may have no output; use input for
+    // cache key
     at::Tensor cache_buffer =
         output_tensor.defined() ? output_tensor : input_tensor;
     // For Reduce specifically, use the input tensor for the cache key

csrc/multidevice/cuda_p2p.cpp

@@ -914,7 +914,8 @@ void postReduceWithCudaBackend(
         cuStreamBatchMemOp(stream, world_size - 1, ops.data(), 0));
 
     // Root launches the ld_reduce kernel
-    NVF_ERROR(output.defined(), "Output must be defined for reduce on the root");
+    NVF_ERROR(
+        output.defined(), "Output must be defined for reduce on the root");
     launchMulticastReduceKernel(
         reduce_handle->multicastPtr(), output.data_ptr(), size, stream);
 

tests/cpp/test_multidevice_stream_parallel_type.cpp

@@ -764,7 +764,8 @@ TEST_P(RSMatmulTest, ReduceScatterReduceBased) {
   C->axis(1)->parallelize(ParallelType::DIDx);
   C->axis(0)->parallelize(ParallelType::Stream);
 
-  // Set the matmul's output to be symmetric in order to use NVLS multimem reduce.
+  // Set the matmul's output to be symmetric in order to use NVLS multimem
+  // reduce.
   if (communicator_backend == CommunicatorBackend::kCuda) {
     C_unreduced->setMemoryType(MemoryType::Symmetric);
   }