Examples
Worked end-to-end scenarios. Each one ships with a ready cluster config, a run command you can copy, and a short explanation of what's interesting about it. Use them as starting points, most real experiments are 80% one of these and 20% your own tweak.
Start here
Browse by topic
Parallelism
TP, PP, EP, and DP+EP, what each looks like in a config, when to reach for which, and how they compose for MoE.
Disaggregated serving
Multi-instance with router, prefill/decode disaggregation, and PIM attention offload. Model real production layouts and compute disaggregation.
Memory tiers
Prefix caching, CXL extended memory, FP8 KV cache. The extended-memory features.
Model families
Beyond dense transformers: hybrid linear attention (gated DeltaNet) and sparse attention (DSA, block-sparse MSA). What each changes about the trace, the memory model and the scheduler.
Advanced
Power modeling, sub-batch interleaving and speculative decoding, niche features that pay off in specific scenarios.
Looking for workloads? Building or generating the JSONL files that drive these examples lives in its own section: Workloads: flat ShareGPT traces, agentic sessions, and the generators that produce them.
How each example is laid out
Every example page follows the same template:
- What it demonstrates: one-sentence summary.
- Prerequisites: which install path you need (most just need the simulator container).
- Cluster config: the JSON file, annotated.
- Run command: copy-able, exact CLI invocation.
- Expected output: what the throughput logs and CSV look like.
- What's interesting: the takeaway.
- Related: pointers to similar or follow-on examples.
Quick onboarding flow if you're new: start with Cluster config explained → then Tensor parallel for the simplest non-trivial example → then whichever topic matches your research question.