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