From 8f397e04bfffde68c018a40b1121ef98c362783f Mon Sep 17 00:00:00 2001
From: Jingyue Wu <wujingyue@gmail.com>
Date: Thu, 1 Jan 2026 22:49:04 -0800
Subject: [PATCH 1/2] Reorder multidevice test includes

---
 tests/cpp/test_multidevice_communications.cpp | 28 ++++-----
 tests/cpp/test_multidevice_communicator.cpp   |  4 +-
 tests/cpp/test_multidevice_host_ir.cpp        | 16 ++---
 .../cpp/test_multidevice_host_ir_overlap.cpp  | 16 ++---
 tests/cpp/test_multidevice_ipc.cpp            | 17 ++---
 .../test_multidevice_lower_communication.cpp  | 21 +++----
 ...t_multidevice_lower_communication_cuda.cpp | 26 ++++----
 tests/cpp/test_multidevice_matmul.cpp         | 32 +++++-----
 tests/cpp/test_multidevice_pipeline.cpp       | 62 +++++++++----------
 tests/cpp/test_multidevice_sharding.cpp       | 16 ++---
 .../test_multidevice_stream_parallel_type.cpp | 22 +++----
 .../cpp/test_multidevice_symmetric_tensor.cpp |  4 +-
 tests/cpp/test_multidevice_transformer.cpp    | 12 ++--
 tests/cpp/test_multidevice_tutorial.cpp       | 16 ++---
 14 files changed, 145 insertions(+), 147 deletions(-)

diff --git a/tests/cpp/test_multidevice_communications.cpp b/tests/cpp/test_multidevice_communications.cpp
index c76539b806a..afd372500af 100644
--- a/tests/cpp/test_multidevice_communications.cpp
+++ b/tests/cpp/test_multidevice_communications.cpp
@@ -5,22 +5,22 @@
  * SPDX-License-Identifier: BSD-3-Clause
  */
 // clang-format on
+#include <cuda.h>
+
 #include <gtest/gtest.h>
 
-#include <cuda.h>
-#include <cuda_utils.h>
-#include <driver_api.h>
-#include <fusion.h>
-#include <ir/builder.h>
-#include <multidevice/communication.h>
-#include <multidevice/communicator.h>
-#include <multidevice/cuda_p2p.h>
-#include <tests/cpp/multidevice.h>
-#include <tests/cpp/validator.h>
-
-#include <ops/all_ops.h>
-#include <ops/arith.h>
-#include <ops/utils.h>
+#include "cuda_utils.h"
+#include "driver_api.h"
+#include "fusion.h"
+#include "ir/builder.h"
+#include "multidevice/communication.h"
+#include "multidevice/communicator.h"
+#include "multidevice/cuda_p2p.h"
+#include "ops/all_ops.h"
+#include "ops/arith.h"
+#include "ops/utils.h"
+#include "tests/cpp/multidevice.h"
+#include "tests/cpp/validator.h"
 
 namespace nvfuser {
 
diff --git a/tests/cpp/test_multidevice_communicator.cpp b/tests/cpp/test_multidevice_communicator.cpp
index a7222c760ea..d1749260b20 100644
--- a/tests/cpp/test_multidevice_communicator.cpp
+++ b/tests/cpp/test_multidevice_communicator.cpp
@@ -11,8 +11,8 @@
 #include <gmock/gmock-matchers.h>
 #include <gtest/gtest.h>
 
-#include <multidevice/communicator.h>
-#include <tests/cpp/multidevice.h>
+#include "multidevice/communicator.h"
+#include "tests/cpp/multidevice.h"
 
 namespace std {
 namespace chrono {
diff --git a/tests/cpp/test_multidevice_host_ir.cpp b/tests/cpp/test_multidevice_host_ir.cpp
index 5898a0365b1..d9ec18a1a5c 100644
--- a/tests/cpp/test_multidevice_host_ir.cpp
+++ b/tests/cpp/test_multidevice_host_ir.cpp
@@ -7,14 +7,14 @@
 // clang-format on
 #include <torch/torch.h>
 
-#include <fusion.h>
-#include <host_ir/container.h>
-#include <host_ir/evaluator.h>
-#include <host_ir/pass/stream_parallel_type.h>
-#include <ir/all_nodes.h>
-#include <ops/all_ops.h>
-#include <preseg_passes/reorder_sharded_axis.h>
-#include <tests/cpp/multidevice.h>
+#include "fusion.h"
+#include "host_ir/container.h"
+#include "host_ir/evaluator.h"
+#include "host_ir/pass/stream_parallel_type.h"
+#include "ir/all_nodes.h"
+#include "ops/all_ops.h"
+#include "preseg_passes/reorder_sharded_axis.h"
+#include "tests/cpp/multidevice.h"
 
 namespace nvfuser {
 
diff --git a/tests/cpp/test_multidevice_host_ir_overlap.cpp b/tests/cpp/test_multidevice_host_ir_overlap.cpp
index 8a96f9a0d04..bede474d821 100644
--- a/tests/cpp/test_multidevice_host_ir_overlap.cpp
+++ b/tests/cpp/test_multidevice_host_ir_overlap.cpp
@@ -9,14 +9,14 @@
 #include <c10/cuda/CUDAStream.h>
 #include <c10/util/ArrayRef.h>
 
-#include <fusion.h>
-#include <host_ir/container.h>
-#include <host_ir/evaluator.h>
-#include <host_ir/ir.h>
-#include <ir/utils.h>
-#include <multidevice/cuda_p2p.h>
-#include <ops/all_ops.h>
-#include <tests/cpp/multidevice.h>
+#include "fusion.h"
+#include "host_ir/container.h"
+#include "host_ir/evaluator.h"
+#include "host_ir/ir.h"
+#include "ir/utils.h"
+#include "multidevice/cuda_p2p.h"
+#include "ops/all_ops.h"
+#include "tests/cpp/multidevice.h"
 
 namespace nvfuser {
 
diff --git a/tests/cpp/test_multidevice_ipc.cpp b/tests/cpp/test_multidevice_ipc.cpp
index 50f438dd97d..90a615602d7 100644
--- a/tests/cpp/test_multidevice_ipc.cpp
+++ b/tests/cpp/test_multidevice_ipc.cpp
@@ -6,14 +6,15 @@
 */
 // clang-format on
 #include <cuda.h>
-#include <fusion.h>
-#include <host_ir/container.h>
-#include <host_ir/evaluator.h>
-#include <ir/all_nodes.h>
-#include <multidevice/ipc_utils.h>
-#include <multidevice/utils.h>
-#include <ops/all_ops.h>
-#include <tests/cpp/multidevice.h>
+
+#include "fusion.h"
+#include "host_ir/container.h"
+#include "host_ir/evaluator.h"
+#include "ir/all_nodes.h"
+#include "multidevice/ipc_utils.h"
+#include "multidevice/utils.h"
+#include "ops/all_ops.h"
+#include "tests/cpp/multidevice.h"
 
 namespace nvfuser {
 
diff --git a/tests/cpp/test_multidevice_lower_communication.cpp b/tests/cpp/test_multidevice_lower_communication.cpp
index 4a6e0443b51..ce7d3e5d25a 100644
--- a/tests/cpp/test_multidevice_lower_communication.cpp
+++ b/tests/cpp/test_multidevice_lower_communication.cpp
@@ -5,20 +5,19 @@
  * SPDX-License-Identifier: BSD-3-Clause
  */
 // clang-format on
-
-#include <cuda_utils.h>
-#include <driver_api.h>
 #include <gmock/gmock-matchers.h>
 #include <gtest/gtest.h>
 
-#include <multidevice/execution_utils.h>
-#include <ops/all_ops.h>
-#include <optimization_pass.h>
-#include <preseg_passes/mark_aliases_prepare.h>
-#include <runtime/communication_executor.h>
-#include <runtime/fusion_executor_cache.h>
-#include <tests/cpp/multidevice.h>
-#include <tests/cpp/validator.h>
+#include "cuda_utils.h"
+#include "driver_api.h"
+#include "multidevice/execution_utils.h"
+#include "ops/all_ops.h"
+#include "optimization_pass.h"
+#include "preseg_passes/mark_aliases_prepare.h"
+#include "runtime/communication_executor.h"
+#include "runtime/fusion_executor_cache.h"
+#include "tests/cpp/multidevice.h"
+#include "tests/cpp/validator.h"
 
 namespace nvfuser {
 
diff --git a/tests/cpp/test_multidevice_lower_communication_cuda.cpp b/tests/cpp/test_multidevice_lower_communication_cuda.cpp
index 3164f193132..000797aa0c3 100644
--- a/tests/cpp/test_multidevice_lower_communication_cuda.cpp
+++ b/tests/cpp/test_multidevice_lower_communication_cuda.cpp
@@ -6,24 +6,24 @@
  */
 // clang-format on
 
-#include <cuda_profiler_api.h>
-
 #include <chrono>
 
+#include <cuda_profiler_api.h>
+
 #include <gmock/gmock-matchers.h>
 #include <gtest/gtest.h>
 
-#include <cuda_utils.h>
-#include <driver_api.h>
-#include <multidevice/execution_utils.h>
-#include <multidevice/ipc_utils.h>
-#include <ops/all_ops.h>
-#include <optimization_pass.h>
-#include <preseg_passes/mark_aliases_prepare.h>
-#include <runtime/communication_executor.h>
-#include <runtime/fusion_executor_cache.h>
-#include <tests/cpp/multidevice.h>
-#include <tests/cpp/validator.h>
+#include "cuda_utils.h"
+#include "driver_api.h"
+#include "multidevice/execution_utils.h"
+#include "multidevice/ipc_utils.h"
+#include "ops/all_ops.h"
+#include "optimization_pass.h"
+#include "preseg_passes/mark_aliases_prepare.h"
+#include "runtime/communication_executor.h"
+#include "runtime/fusion_executor_cache.h"
+#include "tests/cpp/multidevice.h"
+#include "tests/cpp/validator.h"
 
 namespace nvfuser {
 
diff --git a/tests/cpp/test_multidevice_matmul.cpp b/tests/cpp/test_multidevice_matmul.cpp
index 20d5b0f7554..45affe5500d 100644
--- a/tests/cpp/test_multidevice_matmul.cpp
+++ b/tests/cpp/test_multidevice_matmul.cpp
@@ -8,22 +8,22 @@
 #include <gmock/gmock-matchers.h>
 #include <gtest/gtest.h>
 
-#include <expr_evaluator.h>
-#include <fusion.h>
-#include <fusion_segmenter.h>
-#include <ir/all_nodes.h>
-#include <ir/interface_nodes.h>
-#include <ir/iostream.h>
-#include <ir/printer.h>
-#include <ir/utils.h>
-#include <mma_type.h>
-#include <multidevice/execution_utils.h>
-#include <ops/all_ops.h>
-#include <runtime/executor.h>
-#include <scheduler/mma_utils.h>
-#include <scheduler/utils.h>
-#include <tests/cpp/multidevice.h>
-#include <tests/cpp/validator.h>
+#include "expr_evaluator.h"
+#include "fusion.h"
+#include "fusion_segmenter.h"
+#include "ir/all_nodes.h"
+#include "ir/interface_nodes.h"
+#include "ir/iostream.h"
+#include "ir/printer.h"
+#include "ir/utils.h"
+#include "mma_type.h"
+#include "multidevice/execution_utils.h"
+#include "ops/all_ops.h"
+#include "runtime/executor.h"
+#include "scheduler/mma_utils.h"
+#include "scheduler/utils.h"
+#include "tests/cpp/multidevice.h"
+#include "tests/cpp/validator.h"
 
 namespace nvfuser {
 
diff --git a/tests/cpp/test_multidevice_pipeline.cpp b/tests/cpp/test_multidevice_pipeline.cpp
index 1ec47e4986d..ccadcc61b85 100644
--- a/tests/cpp/test_multidevice_pipeline.cpp
+++ b/tests/cpp/test_multidevice_pipeline.cpp
@@ -5,41 +5,39 @@
  * SPDX-License-Identifier: BSD-3-Clause
  */
 // clang-format on
-#include <gtest/gtest.h>
-
 #include <algorithm>
 #include <iostream>
 
-#include <torch/csrc/jit/codegen/cuda/interface.h>
-
-#include <codegen.h>
-#include <device_lower/lower2device.h>
-#include <disjoint_set.h>
-#include <expr_evaluator.h>
-#include <fusion.h>
-#include <fusion_segmenter.h>
-#include <ir/all_nodes.h>
-#include <ir/graphviz.h>
-#include <ir/iostream.h>
-#include <ir/printer.h>
-#include <ir/utils.h>
-#include <iter_visitor.h>
-#include <kernel_ir.h>
-#include <logical_domain_map.h>
-#include <multidevice/execution_utils.h>
-#include <ops/all_ops.h>
-#include <optimization_pass.h>
-#include <preseg_passes/reorder_sharded_axis.h>
-#include <runtime/executor.h>
-#include <runtime/executor_params.h>
-#include <runtime/fusion_executor_cache.h>
-#include <scheduler/all_schedulers.h>
-#include <scheduler/reduction_utils.h>
-#include <scheduler/tools/inlining.h>
-#include <scheduler/utils.h>
-#include <tests/cpp/multidevice.h>
-#include <transform_replay.h>
-#include <transform_rfactor.h>
+#include <gtest/gtest.h>
+
+#include "codegen.h"
+#include "device_lower/lower2device.h"
+#include "disjoint_set.h"
+#include "expr_evaluator.h"
+#include "fusion.h"
+#include "fusion_segmenter.h"
+#include "ir/all_nodes.h"
+#include "ir/graphviz.h"
+#include "ir/iostream.h"
+#include "ir/printer.h"
+#include "ir/utils.h"
+#include "iter_visitor.h"
+#include "kernel_ir.h"
+#include "logical_domain_map.h"
+#include "multidevice/execution_utils.h"
+#include "ops/all_ops.h"
+#include "optimization_pass.h"
+#include "preseg_passes/reorder_sharded_axis.h"
+#include "runtime/executor.h"
+#include "runtime/executor_params.h"
+#include "runtime/fusion_executor_cache.h"
+#include "scheduler/all_schedulers.h"
+#include "scheduler/reduction_utils.h"
+#include "scheduler/tools/inlining.h"
+#include "scheduler/utils.h"
+#include "tests/cpp/multidevice.h"
+#include "transform_replay.h"
+#include "transform_rfactor.h"
 
 namespace nvfuser {
 
diff --git a/tests/cpp/test_multidevice_sharding.cpp b/tests/cpp/test_multidevice_sharding.cpp
index ff4c98936df..46731e11c6b 100644
--- a/tests/cpp/test_multidevice_sharding.cpp
+++ b/tests/cpp/test_multidevice_sharding.cpp
@@ -8,14 +8,14 @@
 #include <gmock/gmock-matchers.h>
 #include <gtest/gtest.h>
 
-#include <fusion.h>
-#include <multidevice/execution_utils.h>
-#include <ops/all_ops.h>
-#include <optimization_pass.h>
-#include <preseg_passes/finalize_multidevice_domains.h>
-#include <runtime/fusion_executor_cache.h>
-#include <tests/cpp/multidevice.h>
-#include <tests/cpp/validator.h>
+#include "fusion.h"
+#include "multidevice/execution_utils.h"
+#include "ops/all_ops.h"
+#include "optimization_pass.h"
+#include "preseg_passes/finalize_multidevice_domains.h"
+#include "runtime/fusion_executor_cache.h"
+#include "tests/cpp/multidevice.h"
+#include "tests/cpp/validator.h"
 
 namespace nvfuser {
 
diff --git a/tests/cpp/test_multidevice_stream_parallel_type.cpp b/tests/cpp/test_multidevice_stream_parallel_type.cpp
index a9d43bbd1a0..2e26cb34ee0 100644
--- a/tests/cpp/test_multidevice_stream_parallel_type.cpp
+++ b/tests/cpp/test_multidevice_stream_parallel_type.cpp
@@ -5,17 +5,17 @@
 * SPDX-License-Identifier: BSD-3-Clause
 */
 // clang-format on
-#include <fusion.h>
-#include <host_ir/container.h>
-#include <host_ir/evaluator.h>
-#include <host_ir/ir.h>
-#include <ir/all_nodes.h>
-#include <multidevice/communication.h>
-#include <multidevice/execution_utils.h>
-#include <ops/all_ops.h>
-#include <preseg_passes/reorder_sharded_axis.h>
-#include <tests/cpp/multidevice.h>
-#include <tests/cpp/validator.h>
+#include "fusion.h"
+#include "host_ir/container.h"
+#include "host_ir/evaluator.h"
+#include "host_ir/ir.h"
+#include "ir/all_nodes.h"
+#include "multidevice/communication.h"
+#include "multidevice/execution_utils.h"
+#include "ops/all_ops.h"
+#include "preseg_passes/reorder_sharded_axis.h"
+#include "tests/cpp/multidevice.h"
+#include "tests/cpp/validator.h"
 
 namespace nvfuser {
 
diff --git a/tests/cpp/test_multidevice_symmetric_tensor.cpp b/tests/cpp/test_multidevice_symmetric_tensor.cpp
index 9cd8409811b..6ffb275bb96 100644
--- a/tests/cpp/test_multidevice_symmetric_tensor.cpp
+++ b/tests/cpp/test_multidevice_symmetric_tensor.cpp
@@ -5,8 +5,8 @@
  * SPDX-License-Identifier: BSD-3-Clause
  */
 // clang-format on
-#include <multidevice/symmetric_tensor.h>
-#include <tests/cpp/multidevice.h>
+#include "multidevice/symmetric_tensor.h"
+#include "tests/cpp/multidevice.h"
 
 namespace nvfuser {
 
diff --git a/tests/cpp/test_multidevice_transformer.cpp b/tests/cpp/test_multidevice_transformer.cpp
index e13c806b89a..fff97ff9907 100644
--- a/tests/cpp/test_multidevice_transformer.cpp
+++ b/tests/cpp/test_multidevice_transformer.cpp
@@ -10,12 +10,12 @@
 #include <gmock/gmock-matchers.h>
 #include <gtest/gtest.h>
 
-#include <fusion.h>
-#include <multidevice/execution_utils.h>
-#include <ops/all_ops.h>
-#include <tests/cpp/multidevice.h>
-#include <tests/cpp/multidevice_transformer.h>
-#include <tests/cpp/validator.h>
+#include "fusion.h"
+#include "multidevice/execution_utils.h"
+#include "ops/all_ops.h"
+#include "tests/cpp/multidevice.h"
+#include "tests/cpp/multidevice_transformer.h"
+#include "tests/cpp/validator.h"
 
 namespace nvfuser {
 
diff --git a/tests/cpp/test_multidevice_tutorial.cpp b/tests/cpp/test_multidevice_tutorial.cpp
index 777d65c464a..2d9b60ccf59 100644
--- a/tests/cpp/test_multidevice_tutorial.cpp
+++ b/tests/cpp/test_multidevice_tutorial.cpp
@@ -8,14 +8,14 @@
 
 #include <torch/torch.h>
 
-#include <host_ir/container.h>
-#include <host_ir/evaluator.h>
-#include <host_ir/ir.h>
-#include <ir/iostream.h>
-#include <multidevice/communicator.h>
-#include <ops/all_ops.h>
-#include <ops/utils.h>
-#include <tests/cpp/multidevice.h>
+#include "host_ir/container.h"
+#include "host_ir/evaluator.h"
+#include "host_ir/ir.h"
+#include "ir/iostream.h"
+#include "multidevice/communicator.h"
+#include "ops/all_ops.h"
+#include "ops/utils.h"
+#include "tests/cpp/multidevice.h"
 
 namespace nvfuser {
 

From 49d1c0279a9fd77ae7e618a61a6eab5739a6254c Mon Sep 17 00:00:00 2001
From: Jingyue Wu <wujingyue@gmail.com>
Date: Sun, 4 Jan 2026 23:06:14 -0800
Subject: [PATCH 2/2] More

---
 tests/cpp/multidevice.cpp             | 23 ++++++++++++-----------
 tests/cpp/multidevice.h               | 12 ++++++------
 tests/cpp/multidevice_transformer.cpp | 15 ++++++++-------
 tests/cpp/multidevice_transformer.h   |  2 +-
 4 files changed, 27 insertions(+), 25 deletions(-)

diff --git a/tests/cpp/multidevice.cpp b/tests/cpp/multidevice.cpp
index 4d962ffa922..47374739557 100644
--- a/tests/cpp/multidevice.cpp
+++ b/tests/cpp/multidevice.cpp
@@ -5,6 +5,8 @@
  * SPDX-License-Identifier: BSD-3-Clause
  */
 // clang-format on
+#include "tests/cpp/multidevice.h"
+
 #include <sys/types.h>
 #include <unistd.h>
 #include <mutex>
@@ -12,20 +14,19 @@
 #ifdef NVFUSER_DISTRIBUTED
 #include <torch/csrc/distributed/c10d/debug.h>
 #else
-#include <multidevice/c10d_mock.h>
+#include "multidevice/c10d_mock.h"
 #endif
 #include <torch/cuda.h>
 
-#include <fusion_segmenter.h>
-#include <ir/all_nodes.h>
-#include <ir/iostream.h>
-#include <multidevice/execution_utils.h>
-#include <multidevice/utils.h>
-#include <ops/all_ops.h>
-#include <options.h>
-#include <runtime/allocations.h>
-#include <tests/cpp/multidevice.h>
-#include <tests/cpp/validator.h>
+#include "fusion_segmenter.h"
+#include "ir/all_nodes.h"
+#include "ir/iostream.h"
+#include "multidevice/execution_utils.h"
+#include "multidevice/utils.h"
+#include "ops/all_ops.h"
+#include "options.h"
+#include "runtime/allocations.h"
+#include "tests/cpp/validator.h"
 
 namespace nvfuser {
 
diff --git a/tests/cpp/multidevice.h b/tests/cpp/multidevice.h
index fa043ef3f6d..931076852d1 100644
--- a/tests/cpp/multidevice.h
+++ b/tests/cpp/multidevice.h
@@ -7,12 +7,12 @@
 // clang-format on
 #pragma once
 
-#include <multidevice/communication.h>
-#include <multidevice/communicator.h>
-#include <multidevice/execution_utils.h>
-#include <multidevice/executor.h>
-#include <multidevice/utils.h>
-#include <tests/cpp/utils.h>
+#include "multidevice/communication.h"
+#include "multidevice/communicator.h"
+#include "multidevice/execution_utils.h"
+#include "multidevice/executor.h"
+#include "multidevice/utils.h"
+#include "tests/cpp/utils.h"
 
 namespace nvfuser {
 
diff --git a/tests/cpp/multidevice_transformer.cpp b/tests/cpp/multidevice_transformer.cpp
index f7ddd3db9f7..e8a523f9526 100644
--- a/tests/cpp/multidevice_transformer.cpp
+++ b/tests/cpp/multidevice_transformer.cpp
@@ -5,16 +5,17 @@
  * SPDX-License-Identifier: BSD-3-Clause
  */
 // clang-format on
+#include "tests/cpp/multidevice_transformer.h"
+
 #include <unordered_set>
 #include <vector>
 
-#include <fusion.h>
-#include <ir/utils.h>
-#include <multidevice/utils.h>
-#include <ops/all_ops.h>
-#include <scheduler/utils.h>
-#include <tests/cpp/multidevice.h>
-#include <tests/cpp/multidevice_transformer.h>
+#include "fusion.h"
+#include "ir/utils.h"
+#include "multidevice/utils.h"
+#include "ops/all_ops.h"
+#include "scheduler/utils.h"
+#include "tests/cpp/multidevice.h"
 
 namespace nvfuser {
 namespace {
diff --git a/tests/cpp/multidevice_transformer.h b/tests/cpp/multidevice_transformer.h
index 47de2e2ac13..923c72958e0 100644
--- a/tests/cpp/multidevice_transformer.h
+++ b/tests/cpp/multidevice_transformer.h
@@ -8,7 +8,7 @@
 #include <cstdint>
 #include <vector>
 
-#include <runtime/fusion_executor_cache.h>
+#include "runtime/fusion_executor_cache.h"
 
 namespace nvfuser {