TL;DR:

If you had to choose one for a professional dev who lives in HuggingFace weights, Unsloth scripts to fine-tune, and llama.cpp/vllm servers for local inference, which machine is the better long-term investment?

I’m currently at a crossroads and need some community wisdom. I’m looking to buy for a very specific AI development workflow, and I’m thinking between an NVIDIA RTX PRO 5000 48GB (Blackwell) workstation and a MacBook Pro M5 Max 128GB.

My job is just needing to fine-tune with small/quantized models (< 32B). I see the VGA is the clearly winner. But I want to get more opinions from the community.

My analysis so far:

1. The Model Size vs Speed Trade-off

The RTX has extremely good bandwidth 1,344 GB/s vs 614 GB/s (M5 Max) that denotes via inference speed.

The unified memory gives me more opportunities to run massive models (even with quantized/MoE models), then more headroom for larger context window.

2. The Unsloth Bottleneck

Unsloth is a CUDA masterpiece. Moving to a Mac means losing those specific kernels and potentially doubling my training time. Is the extra RAM on the Mac worth losing the "Unsloth edge"? Eventually, they will roll out to support MLX soon from their roadmap.

3. LLM Inference engine - llama.cpp and vllm

How should I optimize LLM inference for these two setups? I’m familiar with Windows (WSL2) and macOS.

Specifically, which engine provides the best performance for:

- MacBook M5 Max (128GB RAM): Should I use llama.cpp or vLLM?

- NVIDIA RTX Pro 5000 (48GB VRAM): Which engine best utilizes this hardware?

I would love to hear from anyone who has used both or moved from one to the other!