⚡Under $0.50/hr🧠VRAM Estimator⚖Compare GPUs🎁Free LLM APIs🎯Model Index
Learn

Local LLM Serving Engines

Architecture comparison of PagedAttention, GGUF CPU offloading, and RadixAttention. Throughput benchmarks and hardware sizing.

Engine Architecture Comparison

EngineArchitectureThroughputMin HardwareBest For
vLLMPagedAttention~3,000 tok/s (batch 8)2x H100Production serving
Ollamallama.cpp backend~120 tok/s (single)RTX 4090Local development
SGLangRadixAttention~2,500 tok/s (batch 4)2x H100High-throughput serving
llama.cppGGUF offloading~60 tok/s (single CPU)Any x86 CPUEdge / laptop

Architecture Deep Dive

PagedAttention (vLLM)

PagedAttention allocates KV cache in fixed-size blocks managed by a block table, similar to virtual memory paging. This eliminates external fragmentation and enables continuous batching, achieving 2-4x throughput over naive implementations. Prefix caching further optimizes repeated prompt processing.

GGUF CPU Offloading (llama.cpp)

GGUF stores quantized model weights with metadata allowing CPU-to-GPU offloading. Each layer can reside on different devices. Supports INT4/INT8/FP16 quantization with minimal overhead. Enables inference on any platform from Raspberry Pi to data-center GPU clusters.

RadixAttention (SGLang)

RadixAttention uses a tree-structured KV cache for multi-step decoding. It enables efficient speculative decoding where a small draft model generates tokens that the larger target model verifies in parallel. Combined with Radix tree lookup, it achieves near-optimal prefix reuse across concurrent requests.

Hardware Sizing Table

ModelVRAM (INT4)GPUCPU RAMDisk
7B (INT4)6 GBRTX 40908 GB RAM5 GB
14B (INT4)10 GBRTX 409016 GB RAM10 GB
32B (INT4)20 GBRTX 4090 ×232 GB RAM20 GB
70B (INT4)40 GB2x H10064 GB RAM40 GB
Benchmark data: Verified via community benchmarks and official documentationMethodology →