REDDIT · REDDIT// 35d agoINFRASTRUCTURE

AMD Instinct MI50 passthrough issue hits LocalLLaMA setup

A LocalLLaMA user says an AMD Instinct MI50 that previously worked for local LLM experiments stopped loading after they swapped in an NVIDIA Tesla V100 on a QEMU GPU passthrough VM. The card still appears in the system, but `modprobe amdgpu` now fails with error `-12`, turning the thread into a mixed-vendor GPU troubleshooting case rather than a product announcement.

// ANALYSIS

This is the unglamorous side of local LLM infrastructure: cheap datacenter GPUs are attractive, but mixed AMD/NVIDIA passthrough setups can break in ways that are hard to unwind.

–The post is about hardware and driver stability, not a new launch or release
–Error `-12` points more toward resource allocation, BAR mapping, IOMMU, or passthrough configuration trouble than a clear “dead GPU” verdict
–Swapping between AMD Instinct and Tesla cards in the same host is exactly the sort of edge-case many homelab LLM builders run into
–The thread is useful as a signal that second-hand accelerator setups still carry real operational friction for self-hosted inference

// TAGS

amd-instinct-mi50gpuinferenceself-hostedllm

DISCOVERED

35d ago

2026-03-07

PUBLISHED

35d ago

2026-03-07

RELEVANCE

6/ 10

AUTHOR

WhatererBlah555