vLLM mixed 3090, 3080 rigs questioned

My read: this is a plausible hack, not a cleanly supported setup. vLLM’s docs clearly support multi-GPU tensor parallelism, but the project’s public guidance does not promise smooth behavior with mixed cards, and a GitHub feature request explicitly calls out heterogeneous TP as a gap.

• –vLLM is optimized for sharding model weights across GPUs, so the software path exists; the question is whether the hardware mix behaves well under NCCL and TP scheduling.
• –A 3090 plus modded 3080 20G are close in architecture, but the weaker card will still cap effective throughput and may waste some of the 3090’s headroom.
• –Mixed-GPU setups tend to be more fragile on PCIe-only consumer rigs, where communication overhead can erase the benefit of adding cheaper VRAM.
• –If it works, it is more likely to be acceptable for capacity expansion than for maximizing tokens/sec.
• –The discussion is useful because it reflects a common local-LLM tradeoff: buy matched expensive cards, or accept compatibility risk to stretch budget and memory.