PyTorch migrates vLLM benchmark CI from EC2 to OSDC runners
Two key jobs move to OSDC, ditching EC2 for cheaper, faster builds.
PyTorch migrated the last two EC2 jobs (set-parameters and build) of its vLLM benchmark workflow onto OSDC ARC runners (the benchmarks job was already on OSDC). The set‑parameters job replaced linux.c7i.2xlarge with mt‑l‑x86iamx‑8‑16, added a workspace‑permission fix, and swapped calculate‑docker‑image for inline ECR‑path construction. The build job replaced linux.24xlarge.memory with mt‑l‑x86iavx512‑94‑768, added id‑token permission and role/arc OIDC, the workspace‑permission fix, SCCACHE_S3_NO_CREDENTIALS from the creds‑step outcome, and switched artifact upload to upload‑build‑artifacts@main (S3 with GHA fallback). Artifacts continue to land in gha‑artifacts under the same build_environment name that the benchmarks job downloads.
- Migrated two vLLM benchmark CI jobs from EC2 to OSDC ARC runners.
- Replaced linux.c7i.2xlarge and linux.24xlarge with smaller OSDC VMs.
- Introduced OIDC permissions, S3 artifact upload with GHA fallback.
Why It Matters
Reduces CI costs and complexity for PyTorch, a critical infrastructure for thousands of ML projects.