AIP-5176: Schedule GPT-OSS 20B on a 2-GPU L40S worker

[vavarshn]

Summary

This change schedules GPT-OSS 20B on an L40S worker from the 2-GPU worker group while keeping its GPU allocation at one GPU.

It also adjusts the default L40S worker topology so that separate 2-GPU workers are available for GPT-OSS and Gemma.

Changes

• Added explicit Ray resource requirements to the GPT-OSS L40S configuration:
  • gpu_count:2: 0.001
  • accelerator_type:L40S: 0.001
• Kept GPT-OSS at num_gpus: 1.
• Updated the default L40S worker topology:
  • l40s-0-gpu: 1
  • l40s-1-gpu: changed from 2 to 1
  • l40s-2-gpu: changed from 1 to 2
  • l40s-4-gpu: 0
• Updated the Ray builder scale-factor parity test to reflect the new defaults.

Reason

GPT-OSS requires one GPU, but placing it on the 1-GPU worker group allows fractional-GPU models to consume the same scheduling tier. Depending on deployment order, this can leave insufficient contiguous GPU capacity for GPT-OSS.

Targeting the 2-GPU worker group allows:

• Gemma to use both GPUs on one 2-GPU worker.
• GPT-OSS to use one GPU on the second 2-GPU worker.
• Fractional-GPU models to remain consolidated on the 1-GPU worker.

Validation

Automated test

GOCACHE=/private/tmp/codex-go-cache go test ./pkg/ai/raybuilder

Result: passed.

AI Tier deployment

The updated operator image was deployed in an air-gapped AI Tier environment:

172.31.13.147:5000/splunk/splunk-ai-operator:gpt-oss-2gpu-20260917-44b41884

Verified:

• Installer completed with exit code 0.
• All four Kubernetes nodes reported Ready.
• RayCluster reported 4/4 workers and 5 GPUs.
• RayService reported Running with five Serve endpoints.
• All 13 Ray Serve applications reported RUNNING.
• All Ray and AI Tier pods were healthy with zero restarts.
• No Ray nodes were pending or reporting recent failures.

Placement

• GPT-OSS 20B was scheduled on:
  ai-cluster-ai-platform-raycluster-jdqkn-l40s-2-gpu-worker-2rtw6
• GPT-OSS allocated one GPU on that worker.
• Gemma was scheduled on the other 2-GPU worker.
• Fractional-GPU models were placed on the 1-GPU worker.

Inference

GPT-OSS request:

POST /gpt_oss_20b/v1/chat/completions

Result:

HTTP 200
GPT_OSS_20B_OK

Gemma request:

POST /gemma4_31b_it/v1/chat/completions

Result:

HTTP 200
GEMMA_4_31B_OK

Notes

• This change does not increase GPT-OSS GPU allocation; it still uses one GPU.
• The second GPU on the GPT-OSS worker is intentionally left available at the physical worker level.
• Gemma remains configured to use two L40S GPUs in this change.

Deployment evidence

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-18T05:55:54.068887Z	e87f88c	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.