2024-10-02 nightly release (18525e1)

Author: pytorchbot

.ci/docker/build.sh

@@ -379,6 +379,14 @@ case "$image" in
     GCC_VERSION=11
     CONDA_CMAKE=yes
     HALIDE=yes
+    TRITON=yes
+    ;;
+  pytorch-linux-jammy-py3.12-triton-cpu)
+    CUDA_VERSION=12.4
+    ANACONDA_PYTHON_VERSION=3.12
+    GCC_VERSION=11
+    CONDA_CMAKE=yes
+    TRITON_CPU=yes
     ;;
   pytorch-linux-focal-linter)
     # TODO: Use 3.9 here because of this issue https://github.com/python/mypy/issues/13627.
@@ -509,6 +517,7 @@ docker build \
        --build-arg "UCC_COMMIT=${UCC_COMMIT}" \
        --build-arg "CONDA_CMAKE=${CONDA_CMAKE}" \
        --build-arg "TRITON=${TRITON}" \
+       --build-arg "TRITON_CPU=${TRITON_CPU}" \
        --build-arg "ONNX=${ONNX}" \
        --build-arg "DOCS=${DOCS}" \
        --build-arg "INDUCTOR_BENCHMARKS=${INDUCTOR_BENCHMARKS}" \

.ci/docker/ci_commit_pins/triton-cpu.txt

@@ -0,0 +1 @@
+6a333f1b05671f6fada4ba7bbfae4a02a9d96f4f

.ci/docker/common/install_triton.sh

@@ -15,8 +15,11 @@ conda_reinstall() {
 if [ -n "${XPU_VERSION}" ]; then
   TRITON_REPO="https://github.com/intel/intel-xpu-backend-for-triton"
   TRITON_TEXT_FILE="triton-xpu"
+elif [ -n "${TRITON_CPU}" ]; then
+  TRITON_REPO="https://github.com/triton-lang/triton-cpu"
+  TRITON_TEXT_FILE="triton-cpu"
 else
-  TRITON_REPO="https://github.com/openai/triton"
+  TRITON_REPO="https://github.com/triton-lang/triton"
   TRITON_TEXT_FILE="triton"
 fi
 
@@ -44,9 +47,10 @@ chown -R jenkins /var/lib/jenkins/triton
 chgrp -R jenkins /var/lib/jenkins/triton
 pushd /var/lib/jenkins/
 
-as_jenkins git clone ${TRITON_REPO} triton
+as_jenkins git clone --recursive ${TRITON_REPO} triton
 cd triton
 as_jenkins git checkout ${TRITON_PINNED_COMMIT}
+as_jenkins git submodule update --init --recursive
 cd python
 
 # TODO: remove patch setup.py once we have a proper fix for https://github.com/triton-lang/triton/issues/4527

.ci/docker/ubuntu/Dockerfile

@@ -147,6 +147,13 @@ COPY ci_commit_pins/triton.txt triton.txt
 RUN if [ -n "${TRITON}" ]; then bash ./install_triton.sh; fi
 RUN rm install_triton.sh common_utils.sh triton.txt
 
+ARG TRITON_CPU
+COPY ./common/install_triton.sh install_triton.sh
+COPY ./common/common_utils.sh common_utils.sh
+COPY ci_commit_pins/triton-cpu.txt triton-cpu.txt
+RUN if [ -n "${TRITON_CPU}" ]; then bash ./install_triton.sh; fi
+RUN rm install_triton.sh common_utils.sh triton-cpu.txt
+
 ARG EXECUTORCH
 # Build and install executorch
 COPY ./common/install_executorch.sh install_executorch.sh

.ci/pytorch/test.sh

@@ -376,7 +376,7 @@ test_inductor_cpp_wrapper_abi_compatible() {
 
   echo "Testing Inductor cpp wrapper mode with TORCHINDUCTOR_ABI_COMPATIBLE=1"
   PYTORCH_TESTING_DEVICE_ONLY_FOR="" python test/run_test.py --include inductor/test_cpu_cpp_wrapper
-  python test/run_test.py --include inductor/test_cuda_cpp_wrapper inductor/test_cpu_repro
+  python test/run_test.py --include inductor/test_cuda_cpp_wrapper inductor/test_cpu_repro inductor/test_extension_backend
 
   TORCHINDUCTOR_CPP_WRAPPER=1 python benchmarks/dynamo/timm_models.py --device cuda --accuracy --amp \
     --training --inductor --disable-cudagraphs --only vit_base_patch16_224 \
@@ -403,7 +403,7 @@ pr_time_benchmarks() {
   PYTHONPATH=$(pwd)/benchmarks/dynamo/pr_time_benchmarks source benchmarks/dynamo/pr_time_benchmarks/benchmark_runner.sh "$TEST_REPORTS_DIR/pr_time_benchmarks_results.csv" "benchmarks/dynamo/pr_time_benchmarks/benchmarks"
   echo "benchmark results on current PR: "
   cat  "$TEST_REPORTS_DIR/pr_time_benchmarks_results.csv"
-
+  PYTHONPATH=$(pwd)/benchmarks/dynamo/pr_time_benchmarks python benchmarks/dynamo/pr_time_benchmarks/check_results.py "benchmarks/dynamo/pr_time_benchmarks/expected_results.csv" "$TEST_REPORTS_DIR/pr_time_benchmarks_results.csv"
 }
 
 if [[ "${TEST_CONFIG}" == *pr_time_benchmarks* ]]; then
@@ -606,6 +606,11 @@ test_inductor_halide() {
   assert_git_not_dirty
 }
 
+test_inductor_triton_cpu() {
+  python test/run_test.py --include inductor/test_triton_cpu_backend.py --verbose
+  assert_git_not_dirty
+}
+
 test_dynamo_benchmark() {
   # Usage: test_dynamo_benchmark huggingface 0
   TEST_REPORTS_DIR=$(pwd)/test/test-reports
@@ -1439,6 +1444,8 @@ elif [[ "${TEST_CONFIG}" == *inductor_distributed* ]]; then
   test_inductor_distributed
 elif [[ "${TEST_CONFIG}" == *inductor-halide* ]]; then
   test_inductor_halide
+elif [[ "${TEST_CONFIG}" == *inductor-triton-cpu* ]]; then
+  test_inductor_triton_cpu
 elif [[ "${TEST_CONFIG}" == *inductor-micro-benchmark* ]]; then
   test_inductor_micro_benchmark
 elif [[ "${TEST_CONFIG}" == *huggingface* ]]; then

.ci/pytorch/win-build.sh

@@ -26,7 +26,7 @@ fi
 export SCRIPT_HELPERS_DIR=$SCRIPT_PARENT_DIR/win-test-helpers
 
 set +ex
-grep -E -R 'PyLong_(From|As)(Unsigned|)Long\(' --exclude=python_numbers.h --exclude=eval_frame.c torch/
+grep -E -R 'PyLong_(From|As)(Unsigned|)Long\(' --exclude=python_numbers.h  --exclude=pythoncapi_compat.h --exclude=eval_frame.c torch/
 PYLONG_API_CHECK=$?
 if [[ $PYLONG_API_CHECK == 0 ]]; then
   echo "Usage of PyLong_{From,As}{Unsigned}Long API may lead to overflow errors on Windows"

.github/ISSUE_TEMPLATE.md

@@ -5,7 +5,8 @@ about: Tracking incidents for PyTorch's CI infra.
 
 > NOTE: Remember to label this issue with "`ci: sev`"
 
-**MERGE BLOCKING** <!-- remove this line if you don't want this SEV to block merges -->
+ <!-- uncomment the below line if you don't want this SEV to block merges -->
+ <!--  **MERGE BLOCKING** -->
 
 ## Current Status
 *Status could be: preemptive, ongoing, mitigated, closed. Also tell people if they need to take action to fix it (i.e. rebase)*.

.github/ISSUE_TEMPLATE/ci-sev.md

@@ -1,5 +1,9 @@
 # flake8: noqa: G004
 
+# Note: Copies of this script in runner_determinator.py and _runner-determinator.yml
+#       must be kept in sync. You can do it easily by running the following command:
+#           python .github/scripts/update_runner_determinator.py
+
 """
 This runner determinator is used to determine which set of runners to run a
 GitHub job on. It uses the first comment of a GitHub issue (by default
@@ -79,6 +83,9 @@ class Experiment(NamedTuple):
     rollout_perc: float = (
         0  # Percentage of workflows to experiment on when user is not opted-in.
     )
+    all_branches: bool = (
+        False  # If True, the experiment is also enabled on the exception branches
+    )
 
     # Add more fields as needed
 
@@ -212,7 +219,7 @@ def get_potential_pr_author(
 
 def is_exception_branch(branch: str) -> bool:
     """
-    Branches that get opted out of all experiments and should always use Meta runners
+    Branches that get opted out of experiments by default, until they're explicitly enabled.
     """
     return branch.split("/")[0] in {"main", "nightly", "release", "landchecks"}
 
@@ -338,7 +345,10 @@ def is_user_opted_in(user: str, user_optins: UserOptins, experiment_name: str) -
 
 
 def get_runner_prefix(
-    rollout_state: str, workflow_requestors: Iterable[str], is_canary: bool = False
+    rollout_state: str,
+    workflow_requestors: Iterable[str],
+    branch: str,
+    is_canary: bool = False,
 ) -> str:
     settings = parse_settings(rollout_state)
     user_optins = parse_users(rollout_state)
@@ -348,6 +358,12 @@ def get_runner_prefix(
     for experiment_name, experiment_settings in settings.experiments.items():
         enabled = False
 
+        if not experiment_settings.all_branches and is_exception_branch(branch):
+            log.info(
+                f"Branch {branch} is an exception branch. Not enabling experiment {experiment_name}."
+            )
+            continue
+
         # Is any workflow_requestor opted in to this experiment?
         opted_in_users = [
             requestor
@@ -407,35 +423,34 @@ def get_rollout_state_from_issue(github_token: str, repo: str, issue_num: int) -
 def main() -> None:
     args = parse_args()
 
-    if args.github_ref_type == "branch" and is_exception_branch(args.github_branch):
-        log.info(
-            f"Exception branch: '{args.github_branch}', using Meta runners and no experiments."
+    runner_label_prefix = DEFAULT_LABEL_PREFIX
+
+    try:
+        rollout_state = get_rollout_state_from_issue(
+            args.github_token, args.github_issue_repo, args.github_issue
         )
-        runner_label_prefix = DEFAULT_LABEL_PREFIX
-    else:
-        try:
-            rollout_state = get_rollout_state_from_issue(
-                args.github_token, args.github_issue_repo, args.github_issue
-            )
 
-            username = get_potential_pr_author(
-                args.github_token,
-                args.github_repo,
-                args.github_actor,
-                args.github_ref_type,
-                args.github_branch,
-            )
+        username = get_potential_pr_author(
+            args.github_token,
+            args.github_repo,
+            args.github_actor,
+            args.github_ref_type,
+            args.github_branch,
+        )
 
-            is_canary = args.github_repo == "pytorch/pytorch-canary"
+        is_canary = args.github_repo == "pytorch/pytorch-canary"
 
-            runner_label_prefix = get_runner_prefix(
-                rollout_state, (args.github_issue_owner, username), is_canary
-            )
+        runner_label_prefix = get_runner_prefix(
+            rollout_state,
+            (args.github_issue_owner, username),
+            args.github_branch,
+            is_canary,
+        )
 
-        except Exception as e:
-            log.error(
-                f"Failed to get issue. Defaulting to Meta runners and no experiments. Exception: {e}"
-            )
+    except Exception as e:
+        log.error(
+            f"Failed to get issue. Defaulting to Meta runners and no experiments. Exception: {e}"
+        )
 
     set_github_output(GH_OUTPUT_KEY_LABEL_TYPE, runner_label_prefix)