From f87c1375cd69def940a7ac1f27d28a4caca66fec Mon Sep 17 00:00:00 2001
From: Nicholas Sarkauskas <nsarkauskas@PDX01-M03-I40-DGX-A100-DL-1.nvidia.com>
Date: Thu, 9 Jan 2025 10:33:49 -0800
Subject: [PATCH] linter

---
 benchmarks/cpp/transformer.cpp             | 17 ++++++++++-------
 tests/cpp/multidevice_transformer.cpp      |  2 +-
 tests/cpp/multidevice_transformer.h        |  2 +-
 tests/cpp/test_multidevice_transformer.cpp |  2 +-
 4 files changed, 13 insertions(+), 10 deletions(-)

diff --git a/benchmarks/cpp/transformer.cpp b/benchmarks/cpp/transformer.cpp
index 2e6ef6d9ff0..4728d6c114f 100644
--- a/benchmarks/cpp/transformer.cpp
+++ b/benchmarks/cpp/transformer.cpp
@@ -23,23 +23,23 @@
 
 #include <benchmarks/cpp/utils.h>
 #include <csrc/multidevice/utils.h>
-#include <tests/cpp/utils.h>
 #include <tests/cpp/multidevice_transformer.h>
+#include <tests/cpp/utils.h>
 
 using namespace nvfuser;
 
 namespace {
 // Note: We test on smaller model and input sizes to avoid high error
 // accumulation for validation.
-  constexpr int64_t B = 2, E = 768, H = 16, S = 128;
+constexpr int64_t B = 2, E = 768, H = 16, S = 128;
 // Note: Dropout probabilities are set to 0. Since the dropout mask is sharded
 // it throws off the seed offset between the sharded nvFuser program and the
 // unsharded reference.
-  constexpr double kDropoutProb = 0.0, kSdpaProb = 0.0, kSdpaScale = 1e-3;
+constexpr double kDropoutProb = 0.0, kSdpaProb = 0.0, kSdpaScale = 1e-3;
 // Note parameters scaled by kParamScale following weight initialization
 // recommendations:
 // https://huggingface.co/docs/transformers/en/model_doc/gpt2#transformers.GPT2Config.initializer_range
-  constexpr double kParamScale = 0.02;
+constexpr double kParamScale = 0.02;
 } // namespace
 
 // Return reduction tensor view and output of reduction
@@ -49,7 +49,7 @@ void setupTransformerForward(Fusion* fusion, DataType dtype) {
   const int64_t D = communicator_->size(); // number of devices
 
   auto model = std::make_unique<DistributedTransformer>(
-        D, B, E, H, S, kDropoutProb, kSdpaProb);
+      D, B, E, H, S, kDropoutProb, kSdpaProb);
 
   model->setupForward(fusion, dtype, /*sequence_parallel=*/false);
 }
@@ -96,8 +96,11 @@ void transformerFwd(
       x_,
       ln0_w_,
       ln0_b_,
-      transformerShardTensor_Mesh(mha_w0_.view({3, E, E}), 1, mesh, communicator_).view({1, 3 * E / D, E}),
-      transformerShardTensor_Mesh(mha_b0_.view({3, E}), 1, mesh, communicator_).view({1, 3 * E / D}),
+      transformerShardTensor_Mesh(
+          mha_w0_.view({3, E, E}), 1, mesh, communicator_)
+          .view({1, 3 * E / D, E}),
+      transformerShardTensor_Mesh(mha_b0_.view({3, E}), 1, mesh, communicator_)
+          .view({1, 3 * E / D}),
       transformerShardTensor_Mesh(mha_w1_, 1, mesh, communicator_).unsqueeze(0),
       mha_b1_,
       ln1_w_,
diff --git a/tests/cpp/multidevice_transformer.cpp b/tests/cpp/multidevice_transformer.cpp
index 67b3e48aedc..38917151004 100644
--- a/tests/cpp/multidevice_transformer.cpp
+++ b/tests/cpp/multidevice_transformer.cpp
@@ -395,7 +395,7 @@ std::vector<TensorView*> DistributedTransformer::mha_backwards(
 /* NVFuser benchmark manages the unique_ptr for Fusion and FusionExecutorCache,
    so update the raw pointer with this setupForward function */
 void DistributedTransformer::setupForward(
-    Fusion *fusion,
+    Fusion* fusion,
     DataType dtype,
     bool sequence_parallel) {
   FusionGuard fg(fusion);
diff --git a/tests/cpp/multidevice_transformer.h b/tests/cpp/multidevice_transformer.h
index 43c0b06bf80..caf2a250ced 100644
--- a/tests/cpp/multidevice_transformer.h
+++ b/tests/cpp/multidevice_transformer.h
@@ -46,7 +46,7 @@ class DistributedTransformer {
         kSdpaProb(sdpa_dropout_prob) {}
 
   void setupForward(
-      Fusion *fusion,
+      Fusion* fusion,
       DataType dtype,
       bool sequence_parallel = false);
   std::unique_ptr<FusionExecutorCache> forward(
diff --git a/tests/cpp/test_multidevice_transformer.cpp b/tests/cpp/test_multidevice_transformer.cpp
index 8db0db5915a..1b90f1d0b65 100644
--- a/tests/cpp/test_multidevice_transformer.cpp
+++ b/tests/cpp/test_multidevice_transformer.cpp
@@ -11,8 +11,8 @@
 #include <gtest/gtest.h>
 
 #include <fusion.h>
-#include <ops/all_ops.h>
 #include <multidevice/communicator.h>
+#include <ops/all_ops.h>
 #include <tests/cpp/multidevice.h>
 #include <tests/cpp/multidevice_transformer.h>
 #include <tests/cpp/validator.h>