Frontier
A discrete-event simulator for modern LLM serving systems, used to study parallelism, runtime optimizations and disaggregated serving without deploying to real GPU clusters.
Frontier is a discrete-event simulator for modern LLM serving systems that combine complex parallelism, runtime optimizations, sparse mixture-of-experts architectures, and stateful workloads such as reasoning agents and RL rollouts. It currently models vLLM-style serving behavior, with support for AMD accelerators and other engines such as SGLang planned.
It supports co-located and disaggregated serving, including prefill-decode disaggregation and attention-FFN disaggregation with separate prefill, decode-attention and decode-FFN roles. It captures production runtime techniques such as CUDA graphs, speculative decoding, prefix caching, quantization, chunked prefill and hierarchical caching as part of the scheduler-batch-engine loop, since these change batch shape and memory state over time.
Frontier combines calibrated operator, communication, transfer and KV-cache memory models so simulation results can inform deployment decisions and SLA-constrained comparisons without repeatedly provisioning GPU clusters. Simulation runs entirely on CPU using pre-compiled profiling databases; only the profiling step that builds those databases needs a GPU. It is written in Python and licensed under MIT.