Frontier Lab Notes

Frontier Model Systems and Inference

Inference Serving Stack
Inference Serving Stack From request to token — vLLM-style paged KV cache Incoming requests Continuous batching scheduler packs many requests to fill the GPU Prefill whole prompt in one pass; fills KV cache compute-bound Decode one token per step, reads weights + cache memory-bandwidth-bound Paged KV cache fixed-size blocks, allocated on demand (PagedAttention) near-zero fragmentation → higher batch → more throughput Sampling temperature · top-p · (speculative decoding) Output tokens LATENCY dominated by decode: time-to-first-token = prefill, then per-token step time. Speculative decoding cuts it. THROUGHPUT ↔ bigger batches, but batch size fights latency.

The serving path from request batching through the prefill/decode split, a vLLM-style paged KV cache, and sampling, annotated with latency versus throughput tradeoffs.

Frontier architecture is constrained by hardware. A design that looks elegant on paper may lose if it is slow, memory-heavy, communication-heavy, or hard to serve.

Training Versus Inference

Training bottlenecks:

Inference bottlenecks:

Prefill And Decode

Prefill

The model processes the prompt and fills the KV cache. Long prompts make prefill expensive because attention sees the prompt sequence.

Decode

The model generates one token at a time using the KV cache. Decode is often memory-bandwidth-bound because it repeatedly reads model weights and KV cache for small increments of work.

Parallelism

Frontier models use combinations of:

The hard part is not naming these. The hard part is knowing which communication pattern dominates.

Quantization

Quantization reduces precision to lower memory and improve throughput.

Common targets:

Tradeoffs:

Speculative Decoding

Speculative decoding uses a cheaper draft model to propose tokens, then the larger model verifies them.

Goal:

Research question:

MoE Serving

MoE adds serving complications:

This links directly to Mixture of Experts Architectures.

Long Context Serving

Long context stresses:

Long context is not just "increase max sequence length." It is an infrastructure problem.

Research Questions

Related