Minions: A Dynamic and Distributed Virtual GPU Cloud Built on Idle Apple Silicon Devices

While ML model training and inference are both GPU-intensive, CPU-based data processing is often the bottleneck. Existing distributed data processing systems assume homogeneous resource requirements. They excel at CPU-based computation but either under-utilize heterogeneous resources or impose high overheads on failure and reconfiguration. We introduce Ray Data, a scalable system for heterogeneous workloads spanning batch inference and ML training. The key idea is to combine dynamic task creation for memory efficiency with dynamic task placement for operator elasticity. Making both decisions dynamically allows the system to adapt to heterogeneous pipelines where the memory and compute requirements per operator are unknown and variable over time. Ray Data addresses two challenges: (1) leveraging dynamic task creation for memory-efficient pipelining, and (2) designing a task placement policy that jointly allocates heterogeneous compute and memory resources to maximize throughput while limiting memory usage. Ray Data improves throughput on heterogeneous batch inference pipelines by 2.6-12$\times$ compared to traditional data processing systems. By leveraging heterogeneous clusters, Ray Data improves end-to-end training time for multimodal models such as Stable Diffusion by 31% compared to popular ML data loaders.

Apple Silicon is among the most widely used platforms for local LLM inference, yet nearly all inference optimizations are designed and validated on server-class NVIDIA GPUs under batched serving workloads. Which optimizations retain their benefits when moved to Apple Silicon, and what determines their applicability? We evaluate four optimization families: KV-cache pruning, speculative decoding, ported GPU kernels, and post-training quantization. Experiments span four Apple machines under a common single-request decode workload, with matched RTX 4090 comparisons where possible. KV-cache pruning can reach up to 1.75 times end-to-end throughput on Apple but can reduce throughput on desktop NVIDIA GPUs. Under the common benchmark, all evaluated speculative decoding methods reduce throughput on Apple, despite reported 2-6 times gains on server GPUs. CUDA kernel designs transfer only partially: public Metal lacks asynchronous copy, and batch-one decode makes several serving-oriented mechanisms inapplicable, although transferable components yield up to 1.20 times decode throughput. Quantization gains vary by machine. Across four families, a reported speedup is not a portable property of a technique alone but a joint function of technique, runtime, hardware contract, and workload regime.

Apple-Silicon SoCs share CPU, GPU, and Neural Engine over one unified memory system, raising the question of whether transformer inference can be accelerated by splitting single operators across units. Prior attempts, including our own, failed or produced precision-confounded wins. We identify the cause: MLX's lazy-graph scheduler serializes cross-stream work whenever a CPU-stream operation consumes an unmaterialized GPU result inside one evaluation graph, so a row-split matmul that runs \x{1.38} faster with materialized inputs runs \x{0.66} slower than GPU-only inside a lazy graph; an eager materialization boundary restores concurrency (\x{1.34}). \sys{} implements a per-layer, contention-aware CPU+GPU row split for transformer prefill built on this fix. Evaluated across five chips and three Apple-Silicon generations, community-replicated, the split accelerates Llama-shaped decoder-block prefill by \x{1.15}--\x{1.38}, unchanged at full 32-block depth, and reaches \x{1.18}--\x{1.25} faster time-to-first-token on a real Qwen2.5-7B checkpoint served through stock MLX-LM, with token-identical outputs and unchanged decode throughput. We characterize the boundaries equally carefully: decode cannot benefit, bound by shared bandwidth co-execution does not add; precision-matched training loses \x{0.86}--\x{0.97} on all five chips; ANE dispatch overhead excludes it at layer granularity; and a no-regression runtime gate becomes self-defeating under memory pressure, where probing an alternative mode evicts the active mode's working set. Code, raw results, and generation transcripts are released.