vLLM 0.24.0 breaking change: CUDA_VISIBLE_DEVICES no longer set

vLLM v0.24.0 (2026-06-29) stops setting CUDA_VISIBLE_DEVICES internally. GPU selection moves to a new device_ids argument, with the old ROCm behavior on a deprecation window rather than removed outright. Existing multi-GPU launch configs that relied on vLLM managing device visibility can silently break - the process may grab the wrong GPUs, oversubscribe a card that another service is using, or fail to start with no obvious error pointing at the cause. The release also adds MiniMax-M3 and DiffusionGemma support and makes Model Runner V2 handle quantized models by default, which changes the execution path for most Q4/Q8 deployments even if you never touch the new flag.