[CI] forward fix CRCR report step in the torch-nightly lane (#53732)
Signed-off-by: Andrey Talman <atalman@fb.com>
Andrey Talman committed
Aug 26, 2026 at 05:51 UTC
7a9993878cccc5b598d0b8934becf6fffb03e6fb
2 files changed
+7
-29
.buildkite/scripts/crcr-report.sh
+4
-1
@@ -63,7 +63,10 @@ fi
63
# One callback per job, matching HUD's per-job crcr_workflow_job schema.
64
# delivery_id is synthetic: the nightly path has no upstream dispatch to borrow
65
# one from, so build+job is used as the idempotency key.
66
-python3 .buildkite/scripts/crcr_report.py \
66
+# Resolved from this script's location: the pipeline runs it from
67
+# /vllm-workspace/tests, so a repo-relative path would not resolve.
68
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
69
+python3 "${SCRIPT_DIR}/crcr_report.py" \
70
--build-json "${BUILD_JSON}" \
71
--callback-url "${CALLBACK_URL}" \
72
--oidc-token "${OIDC_TOKEN}" \
.buildkite/test_areas/crcr_report.yaml
+3
-28
@@ -1,37 +1,12 @@
1
group: CRCR Report
2
-# Runs only in the torch-nightly lane; the script exits 0 immediately otherwise,
3
-# following the same in-script gating as image_build.sh.
4
-#
5
-# ORDERING (needs confirmation against the pipeline generator): this must be the
6
-# last step to run, because it reads final job states from the Buildkite REST
7
-# API and silently skips any job still in a non-terminal state. On build 83159
8
-# four jobs were still `scheduled` when the build was already 320 jobs deep, so
9
-# running early under-reports rather than erroring.
10
-#
11
-# depends_on across every group is brittle to maintain. If the generator emits a
12
-# `wait` between groups, placing this group last is enough. If not, the more
13
-# robust option is an agent pre-exit hook or a follow-up scheduled step, since
14
-# `CONTINUE_ON_FAILURE=1` means failures must not stop the report either way.
2
+depends_on:
3
+ - image-build
4
steps:
5
- label: ":satellite: Report nightly results to CRCR"
6
key: crcr-report
18
- # Belt and braces with the script's own TORCH_NIGHTLY check: this keeps the
19
- # step out of non-nightly builds entirely, while the script gate still holds
20
- # if the generator ever drops this key.
7
if: build.env("TORCH_NIGHTLY") == "1"
22
- # Reporting is observability, never a gate: a relay outage, an expired
23
- # pipeline mapping or a missing secret must not turn the nightly red.
8
soft_fail: true
9
allow_dependency_failure: true
10
timeout_in_minutes: 15
27
- # Reporting is a couple of REST calls and a curl: no GPU needed, so keep it
28
- # off the accelerator agents the test groups queue for.
29
- device: cpu-small
30
- # No source_file_dependencies: this step is lane-gated, not change-gated,
31
- # so it must not be skipped by the file-based step selection.
11
commands:
33
- - bash .buildkite/scripts/crcr-report.sh
34
- # CRCR_CALLBACK_URL and BUILDKITE_API_TOKEN are deliberately not declared
35
- # here. Buildkite interpolates a step-level env block at pipeline-upload
36
- # time, so declaring them would substitute empty strings and shadow the
37
- # values the agent already has from the pipeline/schedule environment.
12
+ - bash /vllm-workspace/.buildkite/scripts/crcr-report.sh