[dierksen]

What

This PR adds a gpu-g7e-ondemand pool to the multi-runner configs, for a FlashInfer SM120 unit-test lane on GitHub
Actions.

Labels	[self-hosted, linux, x64, gpu, sm120, on-demand] (alias rtx-pro-6000 in place of sm120)
Instances	g7e.4xlarge (16 vCPU, 128 GiB, $3.998/hr), with g7e.8xlarge (32 vCPU, 256 GiB, $5.268/hr) as fallback
AMI	Deep Learning Base OSS Nvidia Driver GPU AMI (Ubuntu 24.04) * (owner 898082745236)
User data	the existing templates/user-data-gpu.sh, unchanged
Cap	runners_maximum_count: 4 (pilot)
Tagging	RunnerPool: gpu-g7e-ondemand on instances, for Cost Explorer

This change only adds the pool. No workflow targets these labels yet, so it launches nothing until a flashinfer PR adds
the lane.

Why these choices

• No g7e.2xlarge. Its 64 GiB is below the peak host RSS of the host-memory-heavy FlashInfer tests. It would also
  cap JIT builds at MAX_JOBS=4.
• Base OSS AMI instead of PyTorch 22.04. The Base OSS Ubuntu 24.04 DLAMI lists G7e as supported and ships the open
  kernel modules that Blackwell needs. Nothing on the host needs PyTorch, because tests run in the flashinfer-ci-*
  images.
• On-demand only. Full-suite jobs take about 5 hours, and the PR workflow dropped spot in flashinfer#4454.
• Pilot cap of 4. The cap rises to about 16 after the full-suite pilot runs and the G/VT vCPU quota increase.

Before merging

Checked in us-west-2 on 2026-10-22:

• Quota. L-DB2E81BA (Running On-Demand G and VT instances) is 1,000 vCPUs. G-instance usage peaked at 528 vCPUs
  over the last 30 days (2026-10-13), leaving about 470. The pilot's worst case is 4 × g7e.8xlarge = 128 vCPUs, so
  it needs no increase.
• AMI. The newest match is Deep Learning Base OSS Nvidia Driver GPU AMI (Ubuntu 24.04) 20261016
  (ami-0c524a5ec025460c1). Its root device is /dev/sda1 and its description lists G7e as supported.
• Driver. Driver ≥ 580 is needed for the cu129/cu130 images. The newest documented build (20260915) ships
  595.91.07 (CUDA 13.2), and the 595 line has shipped since April. The cu134 image needs R615 and is out of scope here.
• Placement. g7e.4xlarge and g7e.8xlarge are offered in all four runner AZs (us-west-2a–d).
• Plan. terraform plan shows only the new runner's resources (launch template, lambdas, SSM, IAM) and the
  matcher config update, with no API Gateway or webhook changes.

Raising the cap to about 16 later would mean up to 256 vCPUs on 4xlarge, or 512 if every runner falls back to
8xlarge. 512 exceeds the current headroom at peak, so request a quota increase to about 1,500 vCPUs at that point.

Test plan

After apply, run a workflow_dispatch smoke job on the new labels:

• nvidia-smi;
• ci/bash.sh with the cu129 and cu130 CI images;
• a small targeted pytest.

Then run the full-suite pilot on cu130 to measure wall time, peak host RSS and cost per run.

🤖 Generated with Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

• Source mode: Auto
• Repository: Unavailable
• Branch: Unavailable
• File: No configuration file from this repository applied
• Commit: Unavailable
• Layers: defaults
• Review profile: CHILL
• Plan: Advanced
• Run ID: a727ee68-eca8-4346-ad9d-4e1f77492987

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}