NetX-lab / Frontier

What it does

• Co-located & Disaggregated Serving. This branch supports monolithic co-location, PDD serving, and AFD serving with separate prefill, decode-attention, and decode-FFN roles.
• Modern Runtime Optimizations. Frontier captures production techniques such as CUDA Graph, speculative decoding / MTP, prefix caching, quantization, chunked prefill, and hierarchical caching as part of the scheduler-batch-engine loop. These
• Fidelity. Frontier combines calibrated operator, communication, transfer, and KV-cache memory models to make simulation results useful for deployment decisions. This helps users compare configurations under SLA constraints, explore large GP