Reproducible inference across rented NVIDIA GPUs: a methods note on the November 9, 2026 GPU Swap Test
Year: 2026
Type: article
Abstract: This methods note reports an internal measurement of reproducible large language model inference on rented NVIDIA GPUs. On November 9, 2026, a frozen Luxi engine v27 job (engine identifier 795c71736e22e2f6) ran DeepSeek-R1-Distill-Qwen-7B w8a8 on nine GPU types in one cloud, RunPod. Each GPU reproduced 136/136 sampled answers and 546,114 generated tokens against a signed H100 NVL reference, with matching recorded per-step logit fingerprints, and each passed a separate 92-case Qwen2-7B gate. These were 17 problems with eight sampled answers each, not 136 distinct questions. Seven Intel and AMD host CPU models independently matched 2/2 small Qwen checks and a short signed R1 replay capped at 32 tokens. The CPU checks used a smaller workload than the full GPU test. A public closed-binary replay demo, a new build identity of the same v27 source, was tested on November 10, 2026 on three of those GPU types and does not inherit the nine-GPU result. The measurements are internal, not independent validation. Answer accuracy is not the claim. vLLM 0.32 normal mode was faster and cheaper per output token on every GPU where the tested configuration ran. The engine source is private.
Author: Waller, Eric
Language: English
Subfield: Hardware and Architecture
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7237182575
