diff --git a/.github/ci_commit_pins/xla.txt b/.github/ci_commit_pins/xla.txt
index c84777ca1e690..509cbb61824cc 100644
--- a/.github/ci_commit_pins/xla.txt
+++ b/.github/ci_commit_pins/xla.txt
@@ -1 +1 @@
-709c75a04d9b35d83a9509e1534a8aa2046b8912
+c3c16ccac41cb2db6ba88fb31342f4af62c7e15a
diff --git a/aten/src/ATen/DLConvertor.cpp b/aten/src/ATen/DLConvertor.cpp
index 57b9ce050eecf..f940170639abb 100644
--- a/aten/src/ATen/DLConvertor.cpp
+++ b/aten/src/ATen/DLConvertor.cpp
@@ -1,9 +1,6 @@
 #include <ATen/DLConvertor.h>
 #include <ATen/Functions.h>
 
-#include <iostream>
-#include <sstream>
-
 using namespace std;
 namespace at {
 
diff --git a/aten/src/ATen/autocast_mode.cpp b/aten/src/ATen/autocast_mode.cpp
index da09e1f6c0c65..bb4bad9f57879 100644
--- a/aten/src/ATen/autocast_mode.cpp
+++ b/aten/src/ATen/autocast_mode.cpp
@@ -1,6 +1,5 @@
 #include <ATen/autocast_mode.h>
 
-#include <iostream>
 #include <exception>
 #include <mutex>
 #include <ATen/CachedTensorUtils.h>
diff --git a/aten/src/ATen/core/Range.cpp b/aten/src/ATen/core/Range.cpp
index bed964f6809fe..06a79a9c7d063 100644
--- a/aten/src/ATen/core/Range.cpp
+++ b/aten/src/ATen/core/Range.cpp
@@ -1,6 +1,6 @@
 #include <ATen/core/Range.h>
 
-#include <iostream>
+#include <ostream>
 
 namespace at {
 
diff --git a/aten/src/ATen/core/dispatch/Dispatcher.h b/aten/src/ATen/core/dispatch/Dispatcher.h
index b794f37404b0f..4ca8f08ceed7d 100644
--- a/aten/src/ATen/core/dispatch/Dispatcher.h
+++ b/aten/src/ATen/core/dispatch/Dispatcher.h
@@ -18,6 +18,10 @@
 #include <ATen/core/grad_mode.h>
 #include <ATen/core/enum_tag.h>
 
+#ifndef NDEBUG
+#include <iostream>
+#endif
+
 namespace c10 {
 
 TORCH_API bool show_dispatch_trace();
diff --git a/aten/src/ATen/core/function_schema_inl.h b/aten/src/ATen/core/function_schema_inl.h
index 3daefc1de2e5a..a6959c661af15 100644
--- a/aten/src/ATen/core/function_schema_inl.h
+++ b/aten/src/ATen/core/function_schema_inl.h
@@ -1,5 +1,6 @@
 #pragma once
-#include <iostream>
+#include <ostream>
+#include <sstream>
 
 // note: windows build doesn't find symbols in operator files unless
 // this is a header file
diff --git a/aten/src/ATen/core/interned_strings.cpp b/aten/src/ATen/core/interned_strings.cpp
index ff9361f462a1a..2457b45c1a371 100644
--- a/aten/src/ATen/core/interned_strings.cpp
+++ b/aten/src/ATen/core/interned_strings.cpp
@@ -4,7 +4,6 @@
 #include <ATen/core/interned_strings.h>
 #include <cstdint>
 #include <cstring>
-#include <iostream>
 #include <mutex>
 #include <sstream>
 #include <string>
diff --git a/aten/src/ATen/core/union_type.cpp b/aten/src/ATen/core/union_type.cpp
index fee0d8b4bcc38..1bffd7b076374 100644
--- a/aten/src/ATen/core/union_type.cpp
+++ b/aten/src/ATen/core/union_type.cpp
@@ -7,7 +7,8 @@
 #include <ATen/core/type_factory.h>
 #include <c10/macros/Macros.h>
 #include <c10/util/irange.h>
-#include <iostream>
+#include <ostream>
+#include <sstream>
 #include <utility>
 
 namespace c10 {
diff --git a/aten/src/ATen/cpu/vec/vec256/vec256_int.h b/aten/src/ATen/cpu/vec/vec256/vec256_int.h
index de543651a0540..ade140ecc93f6 100644
--- a/aten/src/ATen/cpu/vec/vec256/vec256_int.h
+++ b/aten/src/ATen/cpu/vec/vec256/vec256_int.h
@@ -7,7 +7,6 @@
 #include <ATen/cpu/vec/vec_base.h>
 #include <c10/macros/Macros.h>
 #include <c10/util/irange.h>
-#include <iostream>
 
 namespace at {
 namespace vec {
diff --git a/aten/src/ATen/cpu/vec/vec256/vec256_qint.h b/aten/src/ATen/cpu/vec/vec256/vec256_qint.h
index 58068953c81ac..c153da26072b0 100644
--- a/aten/src/ATen/cpu/vec/vec256/vec256_qint.h
+++ b/aten/src/ATen/cpu/vec/vec256/vec256_qint.h
@@ -14,7 +14,6 @@
 
 #include <array>
 #include <cmath>
-#include <iostream>
 
 // This file defines Vectorized<> for the quantized types.
 //
diff --git a/aten/src/ATen/cpu/vec/vec512/vec512.h b/aten/src/ATen/cpu/vec/vec512/vec512.h
index 6ac1d84187368..88deea313a200 100644
--- a/aten/src/ATen/cpu/vec/vec512/vec512.h
+++ b/aten/src/ATen/cpu/vec/vec512/vec512.h
@@ -18,7 +18,7 @@
 #include <cstddef>
 #include <cstdint>
 #include <cstring>
-#include <iostream>
+#include <ostream>
 
 namespace at {
 namespace vec {
diff --git a/aten/src/ATen/functorch/TensorWrapper.cpp b/aten/src/ATen/functorch/TensorWrapper.cpp
index 6d0f7f8aef93a..a8411570801db 100644
--- a/aten/src/ATen/functorch/TensorWrapper.cpp
+++ b/aten/src/ATen/functorch/TensorWrapper.cpp
@@ -11,6 +11,8 @@
 #include <torch/library.h>
 #include <ATen/core/dispatch/Dispatcher.h>
 
+#include <iostream>
+
 namespace at {
 namespace functorch {
 
diff --git a/c10/util/logging_is_not_google_glog.h b/c10/util/logging_is_not_google_glog.h
index d92f163453e93..3b1b869756eb5 100644
--- a/c10/util/logging_is_not_google_glog.h
+++ b/c10/util/logging_is_not_google_glog.h
@@ -4,9 +4,9 @@
 #include <chrono>
 #include <climits>
 #include <ctime>
-#include <fstream>
 #include <iomanip>
 #include <map>
+#include <ostream>
 #include <set>
 #include <sstream>
 #include <string>
diff --git a/functorch/csrc/dim/minpybind.h b/functorch/csrc/dim/minpybind.h
index f4a0c65595db7..de82b5af95a43 100644
--- a/functorch/csrc/dim/minpybind.h
+++ b/functorch/csrc/dim/minpybind.h
@@ -8,7 +8,7 @@
 #define PY_SSIZE_T_CLEAN
 #include <Python.h>
 #include <utility>
-#include <iostream>
+#include <ostream>
 #include <memory>
 
 #define PY_BEGIN try {
diff --git a/torch/csrc/Module.cpp b/torch/csrc/Module.cpp
index 0511f599ddb14..e628cc3867c5d 100644
--- a/torch/csrc/Module.cpp
+++ b/torch/csrc/Module.cpp
@@ -28,6 +28,7 @@
 #include <torch/csrc/THConcat.h>
 #include <torch/csrc/utils/pybind.h>
 #include <cstdlib>
+#include <iostream>
 #include <unordered_map>
 
 #include <ATen/ThreadLocalPythonObjects.h>
diff --git a/torch/csrc/dynamo/compiled_autograd.h b/torch/csrc/dynamo/compiled_autograd.h
index e9d5117711c54..24d49260c7dc4 100644
--- a/torch/csrc/dynamo/compiled_autograd.h
+++ b/torch/csrc/dynamo/compiled_autograd.h
@@ -4,7 +4,6 @@
 #include <torch/csrc/autograd/engine.h>
 #include <torch/csrc/utils/python_stub.h>
 #include <torch/csrc/utils/torch_dispatch_mode.h>
-#include <iostream>
 #include <typeindex>
 #include <vector>
 
diff --git a/torch/csrc/functorch/init.cpp b/torch/csrc/functorch/init.cpp
index 6acd309771d66..a94801cf7ac4f 100644
--- a/torch/csrc/functorch/init.cpp
+++ b/torch/csrc/functorch/init.cpp
@@ -18,6 +18,8 @@
 #include <ATen/functorch/TensorWrapper.h>
 #include <c10/core/AutogradState.h>
 
+#include <iostream>
+
 // This file contains functorch's Python bindings.
 
 namespace torch {
diff --git a/torch/csrc/jit/api/module.cpp b/torch/csrc/jit/api/module.cpp
index 8a396602f74e0..19f67e2f54d11 100644
--- a/torch/csrc/jit/api/module.cpp
+++ b/torch/csrc/jit/api/module.cpp
@@ -19,6 +19,8 @@
 #include <torch/csrc/jit/passes/inliner.h>
 #include <torch/csrc/jit/runtime/operator.h>
 
+#include <iostream>
+
 namespace torch::jit {
 
 namespace {
diff --git a/torch/csrc/jit/codegen/fuser/codegen.h b/torch/csrc/jit/codegen/fuser/codegen.h
index 2b96264549ff6..fc0b34e55fe7e 100644
--- a/torch/csrc/jit/codegen/fuser/codegen.h
+++ b/torch/csrc/jit/codegen/fuser/codegen.h
@@ -6,9 +6,7 @@
 #include <torch/csrc/jit/codegen/fuser/tensor_desc.h>
 #include <torch/csrc/jit/ir/ir.h>
 
-#include <iostream>
 #include <string>
-#include <tuple>
 #include <vector>
 
 namespace torch {
diff --git a/torch/csrc/jit/codegen/fuser/tensor_desc.h b/torch/csrc/jit/codegen/fuser/tensor_desc.h
index 65f456e27ad54..ffc405244a71e 100644
--- a/torch/csrc/jit/codegen/fuser/tensor_desc.h
+++ b/torch/csrc/jit/codegen/fuser/tensor_desc.h
@@ -7,7 +7,7 @@
 #include <torch/csrc/Export.h>
 
 #include <algorithm>
-#include <iostream>
+#include <ostream>
 #include <vector>
 
 namespace torch {
diff --git a/torch/csrc/jit/frontend/concrete_module_type.cpp b/torch/csrc/jit/frontend/concrete_module_type.cpp
index 5e6aeb306a4d6..b18917d0dc01f 100644
--- a/torch/csrc/jit/frontend/concrete_module_type.cpp
+++ b/torch/csrc/jit/frontend/concrete_module_type.cpp
@@ -3,6 +3,8 @@
 #include <c10/util/irange.h>
 #include <torch/csrc/jit/python/pybind_utils.h>
 
+#include <iostream>
+
 namespace torch::jit {
 
 ClassTypePtr ConcreteModuleTypeBuilder::createTypeFromThis() const {
diff --git a/torch/csrc/jit/frontend/source_range.h b/torch/csrc/jit/frontend/source_range.h
index d278d1ab2cb35..2e544e189dd73 100644
--- a/torch/csrc/jit/frontend/source_range.h
+++ b/torch/csrc/jit/frontend/source_range.h
@@ -3,10 +3,11 @@
 #include <c10/util/Optional.h>
 
 #include <algorithm>
-#include <iostream>
 #include <iterator>
 #include <memory>
 #include <numeric>
+#include <ostream>
+#include <sstream>
 #include <unordered_map>
 
 namespace torch {
diff --git a/torch/csrc/jit/frontend/tracer.h b/torch/csrc/jit/frontend/tracer.h
index 7c355857e5b1c..97cdbb237c641 100644
--- a/torch/csrc/jit/frontend/tracer.h
+++ b/torch/csrc/jit/frontend/tracer.h
@@ -12,7 +12,6 @@
 #include <torch/csrc/utils/variadic.h>
 
 #include <cstdint>
-#include <iostream>
 #include <memory>
 #include <mutex>
 #include <unordered_map>
diff --git a/torch/csrc/jit/ir/alias_analysis.cpp b/torch/csrc/jit/ir/alias_analysis.cpp
index 87031ec5867fb..f49dc2c40d09c 100644
--- a/torch/csrc/jit/ir/alias_analysis.cpp
+++ b/torch/csrc/jit/ir/alias_analysis.cpp
@@ -10,6 +10,7 @@
 #include <torch/csrc/jit/runtime/operator.h>
 #include <torch/csrc/utils/memory.h>
 #include <fstream>
+#include <iostream>
 
 namespace torch::jit {
 
diff --git a/torch/csrc/jit/ir/ir.h b/torch/csrc/jit/ir/ir.h
index 91da53c0daeeb..ad8f081c23ba5 100644
--- a/torch/csrc/jit/ir/ir.h
+++ b/torch/csrc/jit/ir/ir.h
@@ -23,7 +23,7 @@
 #include <c10/util/Optional.h>
 
 #include <functional>
-#include <iostream>
+#include <iosfwd>
 #include <unordered_set>
 #include <vector>
 
diff --git a/torch/csrc/jit/jit_log.cpp b/torch/csrc/jit/jit_log.cpp
index d520ee2fa7ec4..716efb7b0c28e 100644
--- a/torch/csrc/jit/jit_log.cpp
+++ b/torch/csrc/jit/jit_log.cpp
@@ -1,5 +1,6 @@
 #include <cstdlib>
 #include <iomanip>
+#include <iostream>
 #include <sstream>
 #include <string>
 #include <unordered_map>
diff --git a/torch/csrc/jit/mobile/model_tracer/MobileModelRunner.h b/torch/csrc/jit/mobile/model_tracer/MobileModelRunner.h
index 403d336534f1b..b6abe86c0fdca 100644
--- a/torch/csrc/jit/mobile/model_tracer/MobileModelRunner.h
+++ b/torch/csrc/jit/mobile/model_tracer/MobileModelRunner.h
@@ -1,6 +1,5 @@
 #pragma once
 
-#include <iostream>
 #include <mutex>
 #include <sstream>
 
diff --git a/torch/csrc/jit/mobile/register_ops_common_utils.h b/torch/csrc/jit/mobile/register_ops_common_utils.h
index c9e40ff7268c2..b0ecaf055f5ee 100644
--- a/torch/csrc/jit/mobile/register_ops_common_utils.h
+++ b/torch/csrc/jit/mobile/register_ops_common_utils.h
@@ -7,8 +7,6 @@
 #include <torch/csrc/jit/runtime/jit_exception.h>
 #include <torch/csrc/jit/runtime/vararg_functions.h>
 
-#include <iostream>
-
 namespace torch {
 namespace jit {
 
diff --git a/torch/csrc/jit/operator_upgraders/upgraders.h b/torch/csrc/jit/operator_upgraders/upgraders.h
index a6c4b81b15f55..e84602160f587 100644
--- a/torch/csrc/jit/operator_upgraders/upgraders.h
+++ b/torch/csrc/jit/operator_upgraders/upgraders.h
@@ -1,7 +1,6 @@
 #pragma once
 #include <c10/macros/Export.h>
 #include <torch/csrc/jit/ir/ir.h>
-#include <iostream>
 #include <mutex>
 #include <string>
 #include <unordered_map>
diff --git a/torch/csrc/jit/operator_upgraders/upgraders_entry.h b/torch/csrc/jit/operator_upgraders/upgraders_entry.h
index e29f854a5028b..e732ca21d0462 100644
--- a/torch/csrc/jit/operator_upgraders/upgraders_entry.h
+++ b/torch/csrc/jit/operator_upgraders/upgraders_entry.h
@@ -1,11 +1,8 @@
 #pragma once
 #include <c10/macros/Export.h>
 #include <torch/csrc/jit/ir/ir.h>
-#include <iostream>
-#include <mutex>
 #include <string>
 #include <unordered_map>
-#include <vector>
 
 namespace torch {
 namespace jit {
diff --git a/torch/csrc/jit/operator_upgraders/utils.h b/torch/csrc/jit/operator_upgraders/utils.h
index 98550330d0861..ddf6e817b5d09 100644
--- a/torch/csrc/jit/operator_upgraders/utils.h
+++ b/torch/csrc/jit/operator_upgraders/utils.h
@@ -2,8 +2,6 @@
 #include <c10/macros/Export.h>
 #include <c10/util/Optional.h>
 #include <torch/csrc/jit/operator_upgraders/version_map.h>
-#include <iostream>
-#include <regex>
 #include <string>
 #include <vector>
 
diff --git a/torch/csrc/jit/passes/liveness.cpp b/torch/csrc/jit/passes/liveness.cpp
index 2aed7cbe3aab0..7aa6ac045749a 100644
--- a/torch/csrc/jit/passes/liveness.cpp
+++ b/torch/csrc/jit/passes/liveness.cpp
@@ -3,6 +3,7 @@
 #include <torch/csrc/jit/ir/alias_analysis.h>
 #include <torch/csrc/jit/ir/ir_views.h>
 #include <torch/csrc/jit/passes/constant_pooling.h>
+#include <iostream>
 #include <memory>
 
 namespace torch {
diff --git a/torch/csrc/jit/passes/onnx/pattern_conversion/pattern_conversion.cpp b/torch/csrc/jit/passes/onnx/pattern_conversion/pattern_conversion.cpp
index 25b97ef908ec1..a1a3d8356d1d0 100644
--- a/torch/csrc/jit/passes/onnx/pattern_conversion/pattern_conversion.cpp
+++ b/torch/csrc/jit/passes/onnx/pattern_conversion/pattern_conversion.cpp
@@ -7,6 +7,8 @@
 
 #include <ATen/ScalarOps.h>
 
+#include <iostream>
+
 // EDITING THIS FILE? READ THIS FIRST!
 // see Note [Edit Pattern Conversion] in pattern_conversion.h
 
diff --git a/torch/csrc/jit/passes/remove_inplace_ops.cpp b/torch/csrc/jit/passes/remove_inplace_ops.cpp
index 2c58c5089686a..cf93fd8a24d0a 100644
--- a/torch/csrc/jit/passes/remove_inplace_ops.cpp
+++ b/torch/csrc/jit/passes/remove_inplace_ops.cpp
@@ -1,4 +1,5 @@
 #include <torch/csrc/jit/passes/remove_inplace_ops.h>
+#include <iostream>
 
 namespace torch {
 namespace jit {
diff --git a/torch/csrc/jit/runtime/argument_spec.cpp b/torch/csrc/jit/runtime/argument_spec.cpp
index 48db6ac397875..2b04ec3186481 100644
--- a/torch/csrc/jit/runtime/argument_spec.cpp
+++ b/torch/csrc/jit/runtime/argument_spec.cpp
@@ -1,6 +1,8 @@
 #include <c10/util/irange.h>
 #include <torch/csrc/jit/runtime/argument_spec.h>
 
+#include <iostream>
+
 namespace torch {
 namespace jit {
 
diff --git a/torch/csrc/jit/runtime/argument_spec.h b/torch/csrc/jit/runtime/argument_spec.h
index b73b136bca5cc..02b443eb4e451 100644
--- a/torch/csrc/jit/runtime/argument_spec.h
+++ b/torch/csrc/jit/runtime/argument_spec.h
@@ -7,7 +7,7 @@
 #include <torch/csrc/Export.h>
 #include <torch/csrc/autograd/variable.h>
 #include <torch/csrc/jit/ir/ir.h>
-#include <iostream>
+#include <ostream>
 #include <vector>
 
 C10_CLANG_DIAGNOSTIC_PUSH()
diff --git a/torch/csrc/jit/runtime/print_handler.cpp b/torch/csrc/jit/runtime/print_handler.cpp
index 9452589f9e390..8288b93a6bb45 100644
--- a/torch/csrc/jit/runtime/print_handler.cpp
+++ b/torch/csrc/jit/runtime/print_handler.cpp
@@ -1,5 +1,6 @@
 #include <torch/csrc/jit/runtime/print_handler.h>
 
+#include <atomic>
 #include <iostream>
 #include <string>
 
diff --git a/torch/csrc/jit/runtime/print_handler.h b/torch/csrc/jit/runtime/print_handler.h
index 2f1f3ee92e069..ba1da4f22ed26 100644
--- a/torch/csrc/jit/runtime/print_handler.h
+++ b/torch/csrc/jit/runtime/print_handler.h
@@ -2,9 +2,7 @@
 
 #include <torch/csrc/Export.h>
 
-#include <atomic>
-#include <functional>
-#include <iostream>
+#include <string>
 
 namespace torch {
 namespace jit {
diff --git a/torch/csrc/jit/runtime/static/impl.cpp b/torch/csrc/jit/runtime/static/impl.cpp
index e770b971992fb..08e864386aa9d 100644
--- a/torch/csrc/jit/runtime/static/impl.cpp
+++ b/torch/csrc/jit/runtime/static/impl.cpp
@@ -28,6 +28,7 @@
 #include <torch/csrc/jit/runtime/vararg_functions.h>
 #include <algorithm>
 #include <cstdint>
+#include <iostream>
 
 #ifndef AT_PER_OPERATOR_HEADERS
 #include <ATen/NativeFunctions.h>
diff --git a/torch/csrc/jit/serialization/python_print.h b/torch/csrc/jit/serialization/python_print.h
index 94fb83d7aab77..dd9f3d4e06507 100644
--- a/torch/csrc/jit/serialization/python_print.h
+++ b/torch/csrc/jit/serialization/python_print.h
@@ -2,7 +2,6 @@
 #include <torch/csrc/Export.h>
 #include <torch/csrc/jit/api/module.h>
 #include <torch/csrc/jit/ir/ir.h>
-#include <iostream>
 #include <vector>
 
 namespace torch {
diff --git a/torch/csrc/jit/tensorexpr/bounds_inference.cpp b/torch/csrc/jit/tensorexpr/bounds_inference.cpp
index 360818bbae742..fe4a26662db67 100644
--- a/torch/csrc/jit/tensorexpr/bounds_inference.cpp
+++ b/torch/csrc/jit/tensorexpr/bounds_inference.cpp
@@ -10,6 +10,8 @@
 
 #include <c10/util/irange.h>
 
+#include <iostream>
+
 namespace torch::jit::tensorexpr {
 
 using namespace analysis;
diff --git a/torch/csrc/jit/tensorexpr/bounds_overlap.cpp b/torch/csrc/jit/tensorexpr/bounds_overlap.cpp
index ef24566f8a42e..dbcdfd10ee5fd 100644
--- a/torch/csrc/jit/tensorexpr/bounds_overlap.cpp
+++ b/torch/csrc/jit/tensorexpr/bounds_overlap.cpp
@@ -3,6 +3,8 @@
 #include <torch/csrc/jit/tensorexpr/ir_visitor.h>
 #include <torch/csrc/jit/tensorexpr/stmt.h>
 
+#include <iostream>
+
 namespace torch::jit::tensorexpr::analysis {
 
 // Returns true if the given expression is guaranteed to be positive.
diff --git a/torch/csrc/jit/tensorexpr/ir_printer.cpp b/torch/csrc/jit/tensorexpr/ir_printer.cpp
index 5ed247b6881a2..fc9e8116072c3 100644
--- a/torch/csrc/jit/tensorexpr/ir_printer.cpp
+++ b/torch/csrc/jit/tensorexpr/ir_printer.cpp
@@ -6,6 +6,8 @@
 
 #include <c10/util/irange.h>
 
+#include <iostream>
+
 namespace torch::jit::tensorexpr {
 
 std::string IRPrinter::dtypeToCppString(const Dtype& dtype) {
diff --git a/torch/csrc/jit/tensorexpr/ir_printer.h b/torch/csrc/jit/tensorexpr/ir_printer.h
index 661d3a463d137..8eff1abd1b262 100644
--- a/torch/csrc/jit/tensorexpr/ir_printer.h
+++ b/torch/csrc/jit/tensorexpr/ir_printer.h
@@ -1,6 +1,6 @@
 #pragma once
 
-#include <iostream>
+#include <ostream>
 
 #include <torch/csrc/jit/tensorexpr/fwd_decls.h>
 #include <torch/csrc/jit/tensorexpr/ir.h>
diff --git a/torch/csrc/jit/tensorexpr/ir_verifier.h b/torch/csrc/jit/tensorexpr/ir_verifier.h
index 90bca0eedeafb..03b6d9a1f0cce 100644
--- a/torch/csrc/jit/tensorexpr/ir_verifier.h
+++ b/torch/csrc/jit/tensorexpr/ir_verifier.h
@@ -1,7 +1,5 @@
 #pragma once
 
-#include <iostream>
-
 #include <torch/csrc/jit/tensorexpr/fwd_decls.h>
 #include <torch/csrc/jit/tensorexpr/ir_visitor.h>
 
diff --git a/torch/csrc/jit/tensorexpr/loopnest.cpp b/torch/csrc/jit/tensorexpr/loopnest.cpp
index e23c0285634dc..1b08286fbd9f4 100644
--- a/torch/csrc/jit/tensorexpr/loopnest.cpp
+++ b/torch/csrc/jit/tensorexpr/loopnest.cpp
@@ -1,6 +1,7 @@
 #include <torch/csrc/jit/tensorexpr/loopnest.h>
 
 #include <algorithm>
+#include <iostream>
 #include <stdexcept>
 #include <typeinfo>
 #include <unordered_map>
diff --git a/torch/csrc/jit/tensorexpr/loopnest_randomization.cpp b/torch/csrc/jit/tensorexpr/loopnest_randomization.cpp
index 80a9cb7203fcd..110d4151a364c 100644
--- a/torch/csrc/jit/tensorexpr/loopnest_randomization.cpp
+++ b/torch/csrc/jit/tensorexpr/loopnest_randomization.cpp
@@ -1,4 +1,5 @@
 #include <algorithm>
+#include <iostream>
 #include <random>
 #include <stdexcept>
 #include <typeinfo>
diff --git a/torch/csrc/jit/tensorexpr/mem_dependency_checker.cpp b/torch/csrc/jit/tensorexpr/mem_dependency_checker.cpp
index 45025db8ca3f8..1d687a6c2b30e 100644
--- a/torch/csrc/jit/tensorexpr/mem_dependency_checker.cpp
+++ b/torch/csrc/jit/tensorexpr/mem_dependency_checker.cpp
@@ -3,6 +3,7 @@
 #include <c10/util/irange.h>
 
 #include <fstream>
+#include <iostream>
 
 namespace torch::jit::tensorexpr::analysis {
 
diff --git a/torch/csrc/jit/tensorexpr/registerizer.cpp b/torch/csrc/jit/tensorexpr/registerizer.cpp
index caf325d732d78..939f82c616dc9 100644
--- a/torch/csrc/jit/tensorexpr/registerizer.cpp
+++ b/torch/csrc/jit/tensorexpr/registerizer.cpp
@@ -1,4 +1,5 @@
 #include <torch/csrc/jit/tensorexpr/registerizer.h>
+#include <iostream>
 
 namespace torch::jit::tensorexpr {
 namespace registerizer {
diff --git a/torch/csrc/jit/tensorexpr/types.h b/torch/csrc/jit/tensorexpr/types.h
index e4c95424f67e2..ee1b8e213adb4 100644
--- a/torch/csrc/jit/tensorexpr/types.h
+++ b/torch/csrc/jit/tensorexpr/types.h
@@ -1,7 +1,7 @@
 #pragma once
 
 #include <cstdint>
-#include <iostream>
+#include <iosfwd>
 
 #include <c10/core/ScalarType.h>
 #include <c10/util/Logging.h>
diff --git a/torch/csrc/lazy/core/debug_util.h b/torch/csrc/lazy/core/debug_util.h
index 50ba1dae8c9ed..ef4b81e1ca9c5 100644
--- a/torch/csrc/lazy/core/debug_util.h
+++ b/torch/csrc/lazy/core/debug_util.h
@@ -1,6 +1,5 @@
 #pragma once
 
-#include <iostream>
 #include <string>
 #include <vector>
 
diff --git a/torch/csrc/profiler/unwind/action.h b/torch/csrc/profiler/unwind/action.h
index 6672625a3da33..e1ed407384fc9 100644
--- a/torch/csrc/profiler/unwind/action.h
+++ b/torch/csrc/profiler/unwind/action.h
@@ -1,6 +1,6 @@
 #pragma once
 #include <stdint.h>
-#include <iostream>
+#include <ostream>
 
 enum {
   A_UNDEFINED = 0x0,
diff --git a/torch/csrc/profiler/unwind/eh_frame_hdr.h b/torch/csrc/profiler/unwind/eh_frame_hdr.h
index 89a08480c3250..9800166675093 100644
--- a/torch/csrc/profiler/unwind/eh_frame_hdr.h
+++ b/torch/csrc/profiler/unwind/eh_frame_hdr.h
@@ -1,6 +1,6 @@
 #pragma once
 #include <stdint.h>
-#include <iostream>
+#include <ostream>
 
 #include <torch/csrc/profiler/unwind/lexer.h>
 #include <torch/csrc/profiler/unwind/unwind_error.h>
diff --git a/torch/csrc/profiler/unwind/unwinder.h b/torch/csrc/profiler/unwind/unwinder.h
index 7d31de56e5af7..87cb7e9053282 100644
--- a/torch/csrc/profiler/unwind/unwinder.h
+++ b/torch/csrc/profiler/unwind/unwinder.h
@@ -2,7 +2,6 @@
 #include <stdint.h>
 #include <torch/csrc/profiler/unwind/action.h>
 #include <torch/csrc/profiler/unwind/unwind_error.h>
-#include <iostream>
 #include <limits>
 
 struct UnwindState {
diff --git a/torch/csrc/utils/throughput_benchmark.h b/torch/csrc/utils/throughput_benchmark.h
index e540022e0f50f..d7c716a2df02f 100644
--- a/torch/csrc/utils/throughput_benchmark.h
+++ b/torch/csrc/utils/throughput_benchmark.h
@@ -7,7 +7,7 @@
 
 #include <torch/csrc/jit/python/pybind_utils.h>
 
-#include <iostream>
+#include <iosfwd>
 #include <memory>
 #include <string>
 #include <vector>
diff --git a/torch/custom_class.h b/torch/custom_class.h
index 1277ca61a94cc..a7f5dd50dd605 100644
--- a/torch/custom_class.h
+++ b/torch/custom_class.h
@@ -12,7 +12,6 @@
 #include <c10/util/TypeTraits.h>
 #include <torch/custom_class_detail.h>
 #include <torch/library.h>
-#include <iostream>
 #include <sstream>
 
 namespace torch {
diff --git a/torch/lib/libshm/socket.h b/torch/lib/libshm/socket.h
index ee36f460b6ed9..aeb0a8bb9d4e7 100644
--- a/torch/lib/libshm/socket.h
+++ b/torch/lib/libshm/socket.h
@@ -9,8 +9,6 @@
 #include <cstddef>
 #include <cstdio>
 #include <cstring>
-#include <iostream>
-#include <sstream>
 #include <string>
 
 #include <libshm/alloc_info.h>