Add 4 verified GPUs and GPU compatibility test script

[slacki-ai]

Summary

• Adds 4 new GPUs to VERIFIED_GPUs after testing both finetuning and inference with our CUDA 12.8 docker image: RTX 6000 Ada, RTX 3090, A40, RTX A4500
• Adds tests/gpu_compatibility_check.py — a manual test script that spawns RunPod pods directly (bypassing the cluster manager) to validate GPU/docker image compatibility

Test results

Each GPU was tested by spawning a RunPod pod with our production docker image and running a minimal 3-step SFT finetuning job and a 3-prompt vLLM inference job on unsloth/Llama-3.2-1B-Instruct.

GPU	Finetuning	Inference	Verdict
RTX 6000 Ada (6000Ada)	✅	✅	Added
A40	✅	✅	Added
RTX 3090	✅	✅	Added
RTX A4500	✅	✅	Added
RTX 5090	❌	❌	Broken — CUDA 12.8 incompatible (workers crash with vram=0)
RTX 4090	❌	❌	Broken — same CUDA detection failure
RTX 2000 Ada	❌	❌	Broken — same CUDA detection failure
RTX A5000	❌	❌	"Does not have the resources"
RTX 4000 Ada	❓	❓	No RunPod stock available
RTX A4000	❓	❓	No RunPod stock available
L4	❓	❓	No RunPod stock available
L40S	❓	❓	No RunPod stock available

About the test script

tests/gpu_compatibility_check.py is a standalone manual script (not picked up by pytest). Usage:

python tests/gpu_compatibility_check.py --gpus L40 A100 A40
python tests/gpu_compatibility_check.py --gpus A40 --job-types inference

It creates RunPod pods directly, creates targeted OpenWeights jobs, polls for completion, and reports a results table + JSON output.

Test plan

• Verified finetuning completes on all 4 GPUs
• Verified inference completes on all 4 GPUs
• Test script syntax validated
• Reviewer: confirm new GPUs are desirable for the verified list

🤖 Generated with Claude Code