EPS
← All docs·Reference

GCP quota requests — paste-ready justifications (2026-06-12)

updated: 2026-06-12

File from the console (IAM & Admin → Quotas, project eps-persona-gpu-jun2026) with your owner account — the eps-router service account cannot file quota requests. Filter the quota list by the names below; exact metric IDs for H100/H200 only appear in the console / Cloud Quotas API, not in the classic region-quota list.

Background (deep-research, 2026-06-12): the 1-4 GPU H100 shapes (a3-highgpu-1g/2g/4g) are Spot/DWS-flex-start-only — no on-demand path exists — and flex-start consumes PREEMPTIBLE quota, so the preemptible metrics below are the ones that unlock H100s. Code-side support is tracked in task #631.

Requests to file (priority order)

#Quota (console filter)RegionAskWhy
1Preemptible NVIDIA H100 GPUsus-central18DWS flex-start / Spot on a3-highgpu-1g/2g — primary H100 route
2Preemptible NVIDIA H100 GPUsus-east4 (backup)8Same; capacity fallback region (a3 shapes present)
3Preemptible NVIDIA A100 80GB GPUsus-central116 (from 8)Headroom for the Spot-A100 lane (#631 item 3)
4NVIDIA A100 80GB GPUs (on-demand)us-central116 (from 8)Proven bottleneck: #597/#606 queued at 8/8
5GPUs (all regions)global≥24Global cap binds on top of every per-model grant
6Preemptible NVIDIA H100 MEGA / H200 metrics (if exposed)us-central18Future a3-mega/a3-ultra flex-start use
7Vertex AI: custom_model_training_preemptible_nvidia_h100_gpusus-central18Independent parallel pool (Vertex FLEX_START jobs)

Justification text (paste into the request form)

Academic AI-safety research project (university PhD fellowship, Google-credits-backed billing) running single-user ML fine-tuning experiments: LoRA and full fine-tunes of 7B-parameter open-weights LLMs (Qwen-2.5-7B), 1-4 GPUs per job. Jobs run on ephemeral instances created per experiment with --max-run-duration (typically 2-24 h) and --instance-termination-action=DELETE; all workloads checkpoint to external storage and are resumable, so they fit Spot / DWS flex-start semantics. Current usage: sustained 5-8 of 8 NVIDIA A100 80GB GPUs in us-central1 over the past two weeks, with experiments regularly queued behind the 8-GPU cap. Requesting preemptible-class H100 quota to run 1-2 GPU a3-highgpu shapes via DWS flex-start (shorter wall-clock per experiment at similar cost), plus modest A100 headroom. Expected steady-state: 2-4 concurrent jobs, 8-16 GPUs peak.

For the A100 on-demand bump (row 4), trim to the usage + queueing sentences.

After grants land

  • H100 rows → unblock task #631 items 1-2 (flex-start + a3-highgpu intent mappings).
  • Row 3 → #631 item 3 (Spot A100) gets headroom; the existing 8 are usable as soon as the SPOT idempotency proofs land.
  • Caveats from the research: quota ≠ capacity (flex-start is queued, validity 90 s-2 h); Google documents the preemptible-quota switch for predefined-run-time standard VMs as one-way; whether fellowship credits cover DWS charges is undocumented — check the first bill.