Merge branch 'master' into upsampling-expose-recompute_scale_factor

.github/templates/linux_ci_workflow.yml.j2

@@ -223,6 +223,8 @@ jobs:
         # Time out the test phase after !{{ timeout_after }} minutes
         timeout-minutes: !{{ timeout_after }}
         run: |
+          set -x
+
           if [[ $TEST_CONFIG == 'multigpu' ]]; then
             TEST_COMMAND=.jenkins/pytorch/multigpu-test.sh
           elif [[ $BUILD_ENVIRONMENT == *onnx* ]]; then

.jenkins/pytorch/common_utils.sh

@@ -70,7 +70,7 @@ function file_diff_from_base() {
 
 function get_bazel() {
   # download bazel version
-  wget https://github.com/bazelbuild/bazel/releases/download/4.2.1/bazel-4.2.1-linux-x86_64 -O tools/bazel
+  wget https://ossci-linux.s3.amazonaws.com/bazel-4.2.1-linux-x86_64 -O tools/bazel
   # verify content
   echo '1a4f3a3ce292307bceeb44f459883859c793436d564b95319aacb8af1f20557c tools/bazel' | sha256sum --quiet -c
 

.jenkins/pytorch/test.sh

@@ -4,12 +4,11 @@
 # (This is set by default in the Docker images we build, so you don't
 # need to set it yourself.
 
+set -ex
+
 # shellcheck disable=SC2034
 COMPACT_JOB_NAME="${BUILD_ENVIRONMENT}"
 
-# Get fully qualified path using realpath
-CUSTOM_TEST_ARTIFACT_BUILD_DIR=$(realpath "${CUSTOM_TEST_ARTIFACT_BUILD_DIR:-${PWD}/../}")
-
 TORCH_INSTALL_DIR=$(python -c "import site; print(site.getsitepackages()[0])")/torch
 TORCH_BIN_DIR="$TORCH_INSTALL_DIR"/bin
 TORCH_LIB_DIR="$TORCH_INSTALL_DIR"/lib
@@ -24,6 +23,12 @@ if [[ -n "${TEST_CONFIG}" ]]; then
     BUILD_ENVIRONMENT="${BUILD_ENVIRONMENT}-${TEST_CONFIG}"
 fi
 
+# Get fully qualified path using realpath
+if [[ "$BUILD_ENVIRONMENT" != *bazel* ]]; then
+  CUSTOM_TEST_ARTIFACT_BUILD_DIR=$(realpath "${CUSTOM_TEST_ARTIFACT_BUILD_DIR:-${PWD}/../}")
+fi
+
+
 # shellcheck source=./common.sh
 source "$(dirname "${BASH_SOURCE[0]}")/common.sh"
 
@@ -69,6 +74,9 @@ fi
 if [[ "$BUILD_ENVIRONMENT" == *rocm* ]]; then
   # Print GPU info
   rocminfo | grep -E 'Name:.*\sgfx|Marketing'
+
+  # Manually set NUM_TEST_SHARDS since Jenkins doesn't do it
+  export NUM_TEST_SHARDS=2
 fi
 
 # --user breaks ppc64le builds and these packages are already in ppc64le docker
@@ -154,13 +162,12 @@ test_python_legacy_jit() {
   assert_git_not_dirty
 }
 
-test_python_shard1() {
-  time python test/run_test.py --exclude-jit-executor --exclude-distributed-tests --shard 1 2 --verbose --determine-from="$DETERMINE_FROM"
-  assert_git_not_dirty
-}
-
-test_python_shard2() {
-  time python test/run_test.py --exclude-jit-executor --exclude-distributed-tests --shard 2 2 --verbose --determine-from="$DETERMINE_FROM"
+test_python_shard() {
+  if [[ -z "$NUM_TEST_SHARDS" ]]; then
+    echo "NUM_TEST_SHARDS must be defined to run a Python test shard"
+    exit 1
+  fi
+  time python test/run_test.py --exclude-jit-executor --exclude-distributed-tests --shard "$1" "$NUM_TEST_SHARDS" --verbose --determine-from="$DETERMINE_FROM"
   assert_git_not_dirty
 }
 
@@ -415,7 +422,7 @@ test_xla() {
   assert_git_not_dirty
 }
 
-# Do NOT run this test before any other tests, like test_python_shard1, etc.
+# Do NOT run this test before any other tests, like test_python_shard, etc.
 # Because this function uninstalls the torch built from branch, and install
 # nightly version.
 test_backward_compatibility() {
@@ -521,11 +528,11 @@ elif [[ "${BUILD_ENVIRONMENT}" == *-test1 || "${JOB_BASE_NAME}" == *-test1 || ("
   fi
   test_without_numpy
   install_torchvision
-  test_python_shard1
+  test_python_shard 1
   test_aten
 elif [[ "${BUILD_ENVIRONMENT}" == *-test2 || "${JOB_BASE_NAME}" == *-test2 || ("${SHARD_NUMBER}" == 2 && $NUM_TEST_SHARDS -gt 1) ]]; then
   install_torchvision
-  test_python_shard2
+  test_python_shard 2
   test_libtorch
   test_aot_compilation
   test_custom_script_ops
@@ -535,7 +542,7 @@ elif [[ "${BUILD_ENVIRONMENT}" == *vulkan* ]]; then
   test_vulkan
 elif [[ "${BUILD_ENVIRONMENT}" == *-bazel-* ]]; then
   test_bazel
-elif [[ "${BUILD_ENVIRONMENT}" == *distributed* ]]; then
+elif [[ "${BUILD_ENVIRONMENT}" == *distributed* || "${JOB_BASE_NAME}" == *distributed* ]]; then
   test_distributed
   test_rpc
 elif [[ "${TEST_CONFIG}" = docs_test ]]; then

android/pytorch_android/src/main/cpp/pytorch_jni_common.cpp

@@ -4,6 +4,7 @@
 #include <string>
 
 #include <c10/core/MemoryFormat.h>
+#include <c10/util/irange.h>
 
 #include <fbjni/ByteBuffer.h>
 #include <fbjni/fbjni.h>
@@ -97,7 +98,7 @@ static at::Tensor newAtTensor(
   std::vector<int64_t> shapeVec{};
   shapeVec.reserve(rank);
   auto numel = 1;
-  for (auto i = 0; i < rank; ++i) {
+  for (const auto i : c10::irange(rank)) {
     shapeVec.push_back(shapeArr[i]);
     numel *= shapeArr[i];
   }
@@ -521,7 +522,7 @@ at::IValue JIValue::JIValueToAtIValue(
 
     std::vector<at::IValue> elements;
     elements.reserve(n);
-    for (auto i = 0; i < n; ++i) {
+    for (const auto i : c10::irange(n)) {
       auto jivalue_element = jarray->getElement(i);
       auto element = JIValue::JIValueToAtIValue(jivalue_element);
       elements.push_back(std::move(element));
@@ -535,7 +536,7 @@ at::IValue JIValue::JIValueToAtIValue(
     size_t n = jArrayPinned.size();
     c10::List<bool> list{};
     list.reserve(n);
-    for (size_t i = 0; i < n; ++i) {
+    for (const auto i : c10::irange(n)) {
       list.push_back(jArrayPinned[i]);
     }
     return at::IValue{std::move(list)};
@@ -547,7 +548,7 @@ at::IValue JIValue::JIValueToAtIValue(
     size_t n = jArrayPinned.size();
     c10::List<int64_t> list{};
     list.reserve(n);
-    for (size_t i = 0; i < n; ++i) {
+    for (const auto i : c10::irange(n)) {
       list.push_back(jArrayPinned[i]);
     }
     return at::IValue{std::move(list)};
@@ -559,7 +560,7 @@ at::IValue JIValue::JIValueToAtIValue(
     size_t n = jArrayPinned.size();
     c10::List<double> list{};
     list.reserve(n);
-    for (size_t i = 0; i < n; ++i) {
+    for (const auto i : c10::irange(n)) {
       list.push_back(jArrayPinned[i]);
     }
     return at::IValue{std::move(list)};
@@ -572,7 +573,7 @@ at::IValue JIValue::JIValueToAtIValue(
     size_t n = jArray->size();
     c10::List<at::Tensor> list{};
     list.reserve(n);
-    for (size_t i = 0; i < n; ++i) {
+    for (const auto i : c10::irange(n)) {
       list.push_back(
           TensorHybrid::newAtTensorFromJTensor(jArray->getElement(i)));
     }
@@ -594,7 +595,7 @@ at::IValue JIValue::JIValueToAtIValue(
     c10::impl::GenericList list{c10::unshapedType(first_element.type())};
     list.reserve(n);
     list.push_back(first_element);
-    for (auto i = 1; i < n; ++i) {
+    for (const auto i : c10::irange(1, n)) {
       auto jivalue_element = jarray->getElement(i);
       auto element = JIValue::JIValueToAtIValue(jivalue_element);
       list.push_back(element);

android/pytorch_android/src/main/cpp/pytorch_jni_lite.cpp

@@ -6,6 +6,7 @@
 #include <fbjni/ByteBuffer.h>
 #include <fbjni/fbjni.h>
 
+#include <c10/util/irange.h>
 #include <torch/csrc/jit/mobile/import.h>
 #include <torch/csrc/jit/mobile/module.h>
 #include <torch/script.h>
@@ -157,7 +158,7 @@ class PytorchJni : public facebook::jni::HybridClass<PytorchJni> {
     std::vector<at::IValue> inputs{};
     size_t n = jinputs->size();
     inputs.reserve(n);
-    for (size_t i = 0; i < n; i++) {
+    for (const auto i : c10::irange(n)) {
       at::IValue atIValue = JIValue::JIValueToAtIValue(jinputs->getElement(i));
       if (at::kVulkan == deviceType_) {
         inputs.push_back(
@@ -186,7 +187,7 @@ class PytorchJni : public facebook::jni::HybridClass<PytorchJni> {
     std::vector<at::IValue> inputs{};
     size_t n = jinputs->size();
     inputs.reserve(n);
-    for (size_t i = 0; i < n; i++) {
+    for (const auto i : c10::irange(n)) {
       at::IValue atIValue = JIValue::JIValueToAtIValue(jinputs->getElement(i));
       if (at::kVulkan == deviceType_) {
         inputs.push_back(

aten/src/ATen/BatchingRegistrations.cpp

@@ -3,6 +3,7 @@
 #include <ATen/BatchedFallback.h>
 #include <ATen/native/ResizeCommon.h>
 #include <ATen/ATen.h>
+#include <c10/util/irange.h>
 
 namespace at {
 
@@ -329,7 +330,7 @@ Tensor permute_batching_rule(const Tensor& self, IntArrayRef dims) {
 
   VmapDimVector all_dims_physical;
   all_dims_physical.reserve(self_physical.tensor().dim());
-  for (int64_t bdim = 0; bdim < self_physical.numBatchDims(); bdim++) {
+  for (const auto bdim : c10::irange(self_physical.numBatchDims())) {
     all_dims_physical.push_back(bdim);
   }
   all_dims_physical.insert(

aten/src/ATen/CPUApplyUtils.h

@@ -2,6 +2,7 @@
 
 #include <ATen/Parallel.h>
 #include <ATen/TensorUtils.h>
+#include <c10/util/irange.h>
 #include <limits>
 #include <utility>
 #include <cstring>
@@ -130,7 +131,7 @@ inline Tensor sort_strides(Tensor& tensor_) {
   IntArrayRef strides = tensor_.strides();
   std::vector<int64_t> indices;
   indices.reserve(tensor_.ndimension());
-  for (int64_t i = 0; i < tensor_.ndimension(); i++) {
+  for (const auto i : c10::irange(tensor_.ndimension())) {
     indices.push_back(i);
   }
   std::sort(indices.begin(), indices.end(), [&strides](int64_t i1, int64_t i2) {
@@ -196,7 +197,7 @@ inline bool _all_equal_numel(at::ArrayRef<Tensor> tensors) {
   if (tensors.size() == 0)
     return true;
   int64_t all_numel = tensors[0].numel();
-  for (size_t i = 1; i < tensors.size(); i++) {
+  for (const auto i : c10::irange(1, tensors.size())) {
     if (tensors[i].numel() != all_numel)
       return false;
   }

aten/src/ATen/Context.cpp

@@ -62,18 +62,32 @@ bool Context::deterministicAlgorithms() const {
   return _deterministic_algorithms;
 }
 
-void Context::setDeterministicAlgorithms(bool b) {
+bool Context::deterministicAlgorithmsWarnOnly() const {
+  return _deterministic_algorithms_warn_only;
+}
+
+void Context::setDeterministicAlgorithms(bool b, bool warn_only=false) {
   _deterministic_algorithms = b;
+  _deterministic_algorithms_warn_only = warn_only;
 }
 
 void Context::alertNotDeterministic(c10::string_view const& caller) {
   if (globalContext().deterministicAlgorithms()) {
-    TORCH_CHECK(false,
-      caller, " does not have a deterministic implementation, but you set "
-      "'torch.use_deterministic_algorithms(True)'. You can turn off determinism ",
-      "just for this operation if that's acceptable for your application. You "
-      "can also file an issue at https://github.com/pytorch/pytorch/issues "
-      "to help us prioritize adding deterministic support for this operation.");
+    if (globalContext().deterministicAlgorithmsWarnOnly()) {
+      TORCH_WARN(
+        caller, " does not have a deterministic implementation, but you set "
+        "'torch.use_deterministic_algorithms(True, warn_only=True)'. "
+        "You can file an issue at https://github.com/pytorch/pytorch/issues "
+        "to help us prioritize adding deterministic support for this operation.");
+    } else {
+      TORCH_CHECK(false,
+        caller, " does not have a deterministic implementation, but you set "
+        "'torch.use_deterministic_algorithms(True)'. You can turn off "
+        "determinism just for this operation, or you can use the "
+        "'warn_only=True' option, if that's acceptable for your application. "
+        "You can also file an issue at https://github.com/pytorch/pytorch/issues "
+        "to help us prioritize adding deterministic support for this operation.");
+    }
   }
 }