vllm-pytorch-ci-triage

作成者： pytorch

Triage a failing vLLM Buildkite CI build for a PyTorch version-bump PR, isolate new regressions vs. pre-existing failures on main by comparing against recent…

npx skills add https://github.com/pytorch/test-infra --skill vllm-pytorch-ci-triage

ZIPダウンロード GitHub

110

vLLM × PyTorch version-bump CI triage

End-to-end workflow for triaging a vLLM CI run that tests a new torch/triton release and filing upstream issues for the real regressions. Derived from the multi-week triage of vLLM PR #40077 (torch 2.12.0 + triton 3.7.0) starting 2026-04-20 against Buildkite build #62138 → filed 16+ issues under umbrella pytorch/pytorch#180899 over a series of daily runs (62138 → 62232 → 62495 → 62583 → 62848 → 63095). The workflow handles both first-time triage and ongoing daily monitoring.

Prerequisites

Both tokens must exist on disk with 600 perms. Do NOT paste into chat.

# Buildkite API token — https://buildkite.com/user/api-access-tokens
# Scopes: read_builds, read_build_logs. Must be a member of the `vllm` org.
umask 077 && printf '%s\n' 'bkua_...' > ~/.buildkite_token && chmod 600 ~/.buildkite_token

# GitHub PAT — https://github.com/settings/tokens/new
# Classic token with `public_repo` scope is enough (pytorch/pytorch + vllm-project/vllm are public).
printf '%s\n' 'ghp_...' > ~/.github_vllm_token && chmod 600 ~/.github_vllm_token

Shell state does NOT persist between Bash tool calls — always read tokens per-invocation via $(cat ~/.foo_token).

gh CLI may not be installed (it wasn't in this env). Use raw curl against the REST API.

Proxy: if curl returns Recv failure: Connection reset by peer or Received HTTP code 0 from proxy after CONNECT for api.buildkite.com / api.github.com, the fwdproxy needs to be set explicitly per-invocation:

export https_proxy=http://fwdproxy:8080

Don't try to retry without it — sleeping/looping won't help. Set the env, then refetch.

Inputs the user usually provides

A failing Buildkite build URL, e.g. https://buildkite.com/vllm/ci/builds/62138
Optionally the PR (e.g. vllm-project/vllm#40077)
An umbrella issue (e.g. pytorch/pytorch#180899) — if present, append new issues to its checklist

If only the Buildkite URL is given, the build JSON contains the commit, branch, and PR link.

Workflow

Step 1 — Confirm build state, get totals

TOKEN=$(cat ~/.buildkite_token)
curl -sH "Authorization: Bearer $TOKEN" \
  "https://api.buildkite.com/v2/organizations/vllm/pipelines/ci/builds/<N>" > /tmp/bk_<N>.json

Pitfall: the anonymous JSON endpoint (https://buildkite.com/vllm/ci/builds/<N>.json) returns statistics only — jobs: [] is empty without auth. Always use api.buildkite.com.

Inspect jobs[*].{name,state,soft_failed,exit_status,id,web_url,log_url}.

Step 2 — Filter to BLOCKING failures only

A vLLM "failed" job can be non-blocking (soft_failed=True). User told us to ignore those. Also exclude waiting_failed (downstream cascades).

def hard_fails(path):
    d = json.load(open(path))
    return [(j['name'], j.get('exit_status'))
            for j in d['jobs']
            if j.get('state') == 'failed' and j.get('soft_failed') is False]

Step 3 — Compare against recent full builds on main

The full builds are commits on main whose message starts with Full CI run - nightly or Full CI run - daily. Pull the last ~week:

curl -sH "Authorization: Bearer $TOKEN" \
  "https://api.buildkite.com/v2/organizations/vllm/pipelines/ci/builds?branch=main&created_from=YYYY-MM-DDT00:00:00Z&per_page=100" \
  > /tmp/main_builds.json

Grep for "Full CI run - nightly" and "Full CI run - daily" in the message, then fetch each build's JSON with the same endpoint as step 1.

Compare SIGNATURES, not just job names. A job-name overlap with main is not sufficient to drop a failure — main may fail the same job for an entirely different reason. Examples seen in this engagement:

V1 Core + KV + Metrics failed on both PR (62495) and main (62254). PR signature: Expected: 0.54 \| Measured: 0.4806 (real accuracy regression). Main signature: Server exited unexpectedly (infra flake). Two different bugs sharing a job name.
Entrypoints Integration (API Server openai - Part 3) similarly: PR fails on test_multi_chunk_streaming[Voxtral], main fails on test_openapi_stateless schemathesis.

Workflow: if a job-name overlaps with main, fetch the main log and compare the FAILED test ID + exception. Only drop if signatures match.

Verify against ≥3 recent main builds before declaring a failure "new" — a single-build pass-or-fail isn't decisive (some jobs flake; some main builds get interrupted). Five-build crosscheck (5 × Full CI runs covering ~3 days) gives a confident verdict.

CRITICAL: an infra-killed main job is not a baseline. A main-build job that exits with exit_status=125 (nvidia-container-cli driver/container error), started_at=None + 0-byte log, or any other infra-kill signature never actually ran the test — treating it as "missing" or "passing" will produce false positives. If most/all recent main builds had a job killed by infra (especially the cluster of B200 nvidia-container-cli timeouts that hit many B200 jobs at once), the comparison is invalid:

Do not file a "new on test PR" issue based on this — the test could fail on torch 2.11 too, you just have no evidence either way.
Ask the user (or the release manager) to retry the main nightly for the specific job. A retried successful run on main is the only valid baseline.
Once retried main completes, compare signatures. Sometimes the failure on main matches the test PR exactly (i.e. it's a pre-existing vLLM bug, not a torch regression — retract).

Real example (2026-05-05): MoE Refactor Integration Test (B200 - TEMPORARY) was hit by exit 125 on every recent main build. Test PR 64468 ran it successfully and caught a GSM8K accuracy floor failure (0.2085 < 0.2100). I filed it as a torch 2.12 regression — wrongly. Once main 64432 was retried and actually ran the test, it produced the same signature with 0.2039 (worse than the test PR). Retracted as not_planned. The skill should reflect: when the main baseline is missing because of infra, demand a rerun before filing.

Step 4 — Pull logs for the survivors

Log endpoint (plain-text variant):

TOKEN=$(cat ~/.buildkite_token)
# IMPORTANT: read token INTO a variable first; inline `$(cat ...)` inside a
# backgrounded curl inside a `while read` loop silently fails (returns 0-byte
# files). Lesson from first attempt that produced all-empty logs.
for id in <JOB_IDS>; do
  curl -sL -H "Authorization: Bearer $TOKEN" \
    "https://api.buildkite.com/v2/organizations/vllm/pipelines/ci/builds/<N>/jobs/$id/log.txt" \
    -o "/tmp/logs_<N>/$id.log" &
done
wait

Step 5 — Strip ANSI + timestamp markers, extract root-cause lines

Buildkite logs contain \x1b_bk;t=...\x07 timestamp markers and ANSI color codes. Strip them first:

import re
ANSI = re.compile(r'\x1b\[[0-9;?]*[a-zA-Z]')
BKT  = re.compile(r'\x1b_bk;[^\x07]*\x07')
def clean(s): return BKT.sub('', ANSI.sub('', s))

Useful signal patterns to scan cleaned logs for:

pytest summary line: = \d+ failed
Explicit RuntimeError: ... / AssertionError: ... / ValueError: ... (skip the generic raise RuntimeError( frames — they hide the real message)
FAILED tests/.../...::test_name lines (pytest verbose output)
Infra: ECR, docker pull, Connection refused, no space left on device, exit status 137
GPU contention: Free memory on device cuda:0 (X/Y GiB) on startup is less than desired

Critical: the literal "Engine core initialization failed. See root cause above." is never the root cause — scan upward for the actual exception line logged by the EngineCore worker process.

Step 6 — Group by root cause

The goal is ONE issue per root cause, not per failing job. From the 2026-04-20 triage, 22 failing non-CPU jobs grouped into 10 distinct root causes. Several causes produced >4 failing jobs each (e.g. Inductor MetaProxy → 4 Fusion E2E variants).

Separate out:

Infra/resource contention → recommend restart, don't file.
Test-case assertions that look like real regressions (e.g. accuracy 0.48 < 0.54 threshold).
Torch/triton framework regressions → pytorch/pytorch.
vLLM-side application bugs (response APIs, multimodal) → vllm-project/vllm.

Step 7 — Draft and confirm before posting

Public issues are high-blast-radius. ALWAYS:

Draft the full title + body in chat.
Ask the user for explicit "post" / "edit: ..." / "skip".
Post one at a time, or in a single batch only after the user approves the whole set.

Title convention: always start with [vllm], then a sub-area tag, then a concise root-cause. Examples:

[vllm] [2.12 regression] torch.library.Library.impl("aten::bmm", ...) now fails ...
[vllm] [triton 3.7] PassManager::run failed in make_ttgir ...
[vllm] [2.12 regression][Inductor] prims.convert_element_type receives MetaProxy ...
[vllm] [2.12 regression][CPU] torch.compile fullgraph=True raises "found no compiled frames" under Intel SDE
[vllm] [2.12 regression][B200] test_batch_invariance: nondeterministic outputs 3/5 trials

Package name is triton, not pytorch-triton (common mistake — the PyPI name is triton).

Body sections to include:

Summary with the single-line exception message quoted.
Environment block: exact torch / triton / CUDA / Python / GPU.
Reproduction or the specific failing test IDs.
Traceback (trimmed — 10–20 relevant frames).
Question / diagnosis — invite the maintainer to clarify intentional behavior change vs. regression.
Links — vLLM PR, Buildkite build, the specific failed job (click-through URL uses the job id as fragment: …/builds/<N>#<job-uuid>), umbrella issue.

Step 8 — Post via GitHub REST API

curl -s -X POST \
  -H "Authorization: Bearer $(cat ~/.github_vllm_token)" \
  -H "Accept: application/vnd.github+json" \
  -H "X-GitHub-Api-Version: 2022-11-28" \
  https://api.github.com/repos/pytorch/pytorch/issues \
  -d @/tmp/issue_body.json

JSON body shape: {"title": "...", "body": "...markdown..."}. Labels and assignees intentionally omitted — let maintainers triage.

Step 9 — Link to umbrella

Fetch the umbrella body, find the last numbered checklist line (^\d+\. \[[ x]\] https://github.com/pytorch/pytorch/issues/\d+), insert the new link(s) with incremented numbers, PATCH:

curl -s -X PATCH \
  -H "Authorization: Bearer $(cat ~/.github_vllm_token)" \
  -H "Accept: application/vnd.github+json" \
  https://api.github.com/repos/pytorch/pytorch/issues/<UMBRELLA> \
  -d "{\"body\": <escaped new body>}"

Always re-fetch the umbrella body before patching — other people may have edited it in between.

Step 10 — Post-filing corrections

Titles and bodies can be bulk-PATCHed; simple string replacements work fine:

new_title = old_title.replace('pytorch-triton', 'triton')
new_body  = old_body.replace('pytorch-triton', 'triton')

Step 11 — Recurring runs (daily monitoring)

Once the umbrella exists, subsequent test-PR builds are not "open new issues per failing job" — they're delta analysis. For each new build:

Re-fetch the umbrella body and the JSON of every linked issue. Cache issue states (open/closed) keyed by number.
Match each hard-failed job in the new build against tracked-issue signatures (build a regex map from issue titles/bodies). Three buckets:
- Still reproducing: tracked issue still hits → no new issue. If the user wants, PATCH the existing issue body to append the new build link to a Reproducibility section.
- Newly silent: previously-failing job/test now passes. Don't immediately close — wait for ≥2 consecutive runs of "silent" before suggesting close.
- Unmatched: failing job whose signature isn't in any tracked issue. Cross-check against ≥3 main builds (per Step 3). If new on the torch-bump branch, draft + post a fresh issue and append to umbrella.
Maintain umbrella checklist hygiene: mark [x] on items that are closed upstream OR confirmed silent for ≥2 runs. Numbering continues — never reuse numbers.

Updating an existing issue's reproducibility list (PATCH pattern):

old = "## Links\n\n- vLLM PR: ...\n- Failing build: <single old build>\n..."
new = """## Reproducibility on torch 2.12 branch

Same `<exact signature>` on the same N tests across every test-PR run since YYYY-MM-DD:

- 2026-04-20: <buildkite URL with #job-uuid>
- 2026-04-22: <buildkite URL with #job-uuid>
- ...

Passes on same-day main builds (torch 2.11): <list of main build numbers>.

## Links

- vLLM PR: ..."""
new_body = body.replace(old, new)

Step 12 — "Closed upstream but still reproducing" check

When a tracked issue's state flips to closed but the same signature keeps appearing in subsequent builds, the fix is in pytorch main but not yet in the test-channel wheel that vLLM CI pulls. Verify by comparing timestamps:

# Closing commit timestamp from the issue's timeline
gh_close = events[event=='closed'].created_at  # commit that auto-closed
# Build start time on the test branch
build.created_at
# Test-channel wheel build timestamp (look at torch dist URL or PEP-503 index page)

If build.created_at > closing_commit.created_at but the failure persists, the wheel predates the fix. Recommendation: cherry-pick the fix to the release branch and rebuild the RC wheel. Don't reopen the issue — it really is fixed in main.

Step 12.5 — Reopen vs file new

Before drafting a "new" issue for an unmatched failure, search the umbrella's closed entries by exact failure-text fragment:

# Compare the failing-test signature ("Generated text X doesn't match...", op name, etc.)
# against every closed umbrella issue's title + body. If you get an exact match,
# REOPEN the closed issue + post a comment with new build links — do not file a duplicate.

A regression that re-appears (closed issue's signature reproduces in a later run) often means either (a) the upstream fix was reverted or (b) a new vLLM-side change re-exposed the same code path. Reopening preserves history and avoids fragmenting the discussion.

PATCH pattern:

# Reopen
curl -X PATCH .../issues/<N> -d '{"state":"open"}'
# Post comment with new build data
curl -X POST .../issues/<N>/comments -d @comment.json
# Update umbrella: change `[x]` back to `[ ]` and add a "reopened YYYY-MM-DD" note

Step 12.6 — Custom-op stride/shape mismatch is almost always a vLLM-side fake-kernel bug, NOT a torch regression

When you see this signature:

AssertionError: expected size N==N, stride A==B at dim=0
Error in op: torch.ops.vllm.<X>.default
This error most often comes from a incorrect fake (aka meta) kernel for a custom op.

Default assumption: vLLM's registered fake kernel for that custom op returns a different shape/stride than the runtime kernel. The check itself (assert_size_stride) is not new in any torch release — it's been there for years (see pytorch/pytorch#177719 discussion to disable it for vLLM perf).

Why torch X passes and torch Y fails on the same vLLM commit can mislead you here:

AOTAutograd cache hit may have bypassed recompile under torch X
torch.Tag.needs_fixed_stride_order only inserts the assert when Inductor sees a stride change
Torch Y might exercise a slightly different graph capture path

Look for the root cause in vLLM, not torch:

Find where the custom op is registered:

git grep -rn "direct_register_custom_op" vllm/ | grep -E "op_name=.<your_op>."
git grep -rn "register_fake|fake_impl|impl_abstract" vllm/ | grep <op>

Read the fake_impl= function — what shape does it return?
Read the actual implementation (op_func=) and any expert/dispatcher classes it calls — what shape do they actually allocate?
Bisect with git log --stat <last-passing>..<first-failing> -- <suspect dir> to find the vLLM commit that introduced the mismatch.

Real example from the gpt-oss MoE Blackwell triage:

Op: torch.ops.vllm.moe_forward.default
_moe_forward_fake returns torch.empty_like(hidden_states) → padded hidden_dim (3072)
TrtLlmMxfp4Experts{Monolithic,Modular} was changed by vllm-project/vllm#40960 to allocate at hidden_dim_unpadded (2880)
Inductor caught it with assert_size_stride(buf, (s72, 3072), (3072, 1), 'torch.ops.vllm.moe_forward.default')
Fix is in vLLM, not pytorch. Track at the vLLM repo, close any pytorch issue with state_reason: not_planned.

Before filing the upstream issue, also search the vLLM repo for an existing report — community filings can land independently:

curl -s "https://api.github.com/search/issues?q=repo:vllm-project/vllm+is:issue+<distinctive_signature>"

Step 13 — vLLM PR status comment

When meaningful events happen (a fix lands, a batch of issues filed, an umbrella checklist update), post a comment on the vLLM test PR (e.g. vllm-project/vllm#40077) summarizing:

Closed upstream: numbered issues no longer reproducing
Newly silent: candidates for close, awaiting verification
Still reproducing: open numbered issues
New: issues filed in the latest run
Dormant: filed but never re-reproduced

The comment is for human-readable status tracking by the release manager; keep it under ~30 bullet points and link to umbrella, not to every individual issue.

Gotchas observed

Don't trust jobs[*].state alone. failed + soft_failed=True is non-blocking. Always filter both.
waiting_failed jobs only mean "upstream I depended on failed" — they aren't independent signal.
started_at=None + exit_status=-1 + 0-byte log = job never ran (infra cancelled/aborted before it started). Don't treat as a real signal — it's an infra event masquerading as state=failed. Ignore.
"Engine core initialization failed. See root cause above." is a red herring — the actual exception is logged several lines earlier by the (EngineCore pid=...) worker.
GPU contention (~1 GiB free on an H100) can cascade dozens of unrelated tests into ValueError: Free memory ... less than desired. If you see this pattern widely, recommend a job rerun before filing; the real failures may be a subset.
CUDA OOM in tp=2 / B200 fusion tests is also commonly infra (the runner had 4–5 GiB free at start when 5 GiB was needed). Cross-check the same job on the same-day main build — if main fails the same way with the same OOM signature, it's contention, not torch. Skip filing.
Log timestamps differ per infrastructure: dgxB200 nodes prefix with _bk;t=… only; mithril/aws nodes prefix with [YYYY-MM-DDTHH:MM:SSZ]. The ANSI-strip + BKT-strip pair handles both.
PyPI vs test channel: ERROR: No matching distribution found for torch==2.12.0 isn't infra — the release isn't on PyPI yet. Tell the user; don't file a bug.
Python-only Installation job has multiple unrelated failure modes: (a) torch not on PyPI — expected, skip. (b) metadata is still not available after N attempts / precompiled wheel for commit X is available — vLLM's own precompiled-wheel infra hiccup, not torch. Both → ignore.
Parallel curl in while read needs the token in a shell variable first, not $(cat …) inline — otherwise backgrounded processes race the substitution and log files are 0 bytes.
Buildkite REST API rate-limit is 400/min. A burst of parallel-fetched logs will hit it; expect 161-byte error JSON instead of real logs. Switch to serial fetch (or sleep 15 between bursts) when rate-limited.
exit_status=125 on multiple B200 jobs simultaneously, all with nvidia-container-cli: initialization error: driver rpc error: timed out = B200 agent driver/container infra issue, not a regression. Recommend rerun, do not file. Often clusters across V1 attention, Fusion E2E, GPQA, LM Eval, MoE Refactor, Spec Decode B200 jobs at once.
Same B200 infra cluster wipes out main-build coverage too. When the main daily/nightly builds also have many B200 jobs killed by exit 125 / nvidia-container-cli (typical when the agent fleet has a bad day), do NOT use those main builds as a baseline. The pattern "test PR fails this job, main appears not to" is inconclusive — main never actually ran the test. ALWAYS ask the user (or release manager) to retry the corresponding main nightly job before drafting a new umbrella issue. Filing without that baseline produced a wrongful issue (#182549, retracted 2026-05-05): same Nemotron-Nano-30B-Fp8 GSM8K accuracy floor failed on both torch 2.11 (after main retry) and torch 2.12, but the torch 2.11 main builds had been killed by the B200 infra issue and looked "passing" by absence.
Compile-on vs --enforce-eager CI gap: fake-kernel / Inductor stride bugs only surface when compile is on. Many gpt-oss CI lanes (tests/evals/gpt_oss/test_gpqa_correctness.py, --enforce-eager parametrizations) bypass torch.compile entirely and never trace the fake kernel. If a custom-op stride mismatch only shows up on the torch-bump test PR, the bug almost certainly exists on main too — vLLM CI is just hiding it. When closing such an issue, mention this gap so vLLM can add coverage.
Closed upstream ≠ fixed in CI. The pytorch test-channel wheel is a snapshot; if the closing commit landed AFTER the wheel was built, the same signature keeps reproducing. Verify with timestamps before reopening — recommend a wheel rebuild instead.
Dockerfile.cpu seeds requirements/test/cpu.in from requirements/test/cuda.in (literal COPY ... cuda.in cpu.in), so the top-line --extra-index-url https://download.pytorch.org/whl/test/cu130 carries over to the CPU build. Combined with uv pip compile --torch-backend cpu (which forces stable cpu channel), torch 2.12 wheels go missing. Fix: sed-rewrite the index-url to whl/test/cpu AND drop --torch-backend cpu.
uv --torch-backend <name> overrides extra-index-url for torch. Only stable channels (cpu, cu128, etc.) are presets — there is no test-cpu preset. To pin torch to the test channel, use --extra-index-url explicitly (or UV_EXTRA_INDEX_URL env) and don't pass --torch-backend.

Token ownership

Tokens live in the user's home dir only. Never echo them to the conversation, never commit them, never include in issue bodies. If the user ever asks to rotate, remind them to delete the file + revoke at the provider UI.

Repo routing cheat-sheet

Error pattern	Repo
`torch.library.Library.impl ... already a kernel registered`	pytorch/pytorch
`MetaProxy` in `prims.*` / Inductor	pytorch/pytorch
`PassManager::run failed` inside `triton/` frames	pytorch/pytorch (triton)
`Pointer argument cannot be accessed from Triton`	pytorch/pytorch (triton)
`Cannot access data pointer of Tensor (FakeTensor…)`	pytorch/pytorch (AOTAutograd)
`_pickle.PicklingError` on triton `launcher`	pytorch/pytorch (triton + AOT cache)
`warm_artifacts_saved: got 0`, `KeyError: None` in standalone_compile	pytorch/pytorch (Inductor cache)
`assert 'no' == 'yes'` in `test_dynamic_shapes_compilation`	pytorch/pytorch (Dynamo), but rerun first if GPU was OOM
`torch.compile with fullgraph=True found no compiled frames` (when `TORCH_COMPILE_DISABLE=1` is in env)	vLLM-side fix usually correct (use `--enforce-eager` instead); upstream interest if behavior change is intentional
`RayChannelTimeoutError: System error: Timed out waiting for object available to read` on tp≥2 ray	pytorch/pytorch — likely torch.compile per-worker latency exceeds Ray channel timeout
`Failed: Nondeterministic outputs detected: N failed out of M trials` (B200-only)	pytorch/pytorch — Blackwell-specific kernel drift
`assert torch.allclose(golden_output, vllm_output, ...)` failure on reward / PRM models	pytorch/pytorch — numerical drift from triton update
`compare_two_settings(... cpu-offload-gb ...)` → "Results are not the same"	pytorch/pytorch (CPU↔GPU dequantize parity)
GSM8K accuracy collapses to 0.000 (not just degrades)	pytorch/pytorch — likely worker-side crash hidden behind unpickle error
`Generated text "X" doesn't match expected pattern "Y"` on Qwen2-VL / Qwen3-VL LoRA	pytorch/pytorch — multimodal LoRA path numerical drift (file separately if different LoRA target)
`AssertionError: expected size N==N, stride A==B at dim=0` + `Error in op: torch.ops.vllm.<X>.default` + "incorrect fake (aka meta) kernel" hint	vllm-project/vllm — fake kernel registered via `direct_register_custom_op(..., fake_impl=...)` returns wrong shape. Find the registration site and the recent vLLM commit that changed the runtime allocation. Close any pytorch issue with `state_reason: not_planned`.
Bulk `exit_status=125` + `nvidia-container-cli: initialization error: driver rpc error: timed out` across many B200 jobs	infra (B200 agent) — recommend rerun, do not file
Multi-modal per-model assertions (qwen2_vl, chameleon)	vllm-project/vllm first — may be torch-side once isolated
Responses API assertion (`'incomplete' == 'completed'`)	vllm-project/vllm
`test_lm_eval_accuracy_v1_engine` — measured below threshold	investigate both — often numerical drift from triton update
CUDA OOM in fusion-test parallel runs (~4–5 GiB free at start)	infra; check main same-day to confirm