Experiment Sequence Starting Point

Experiment Sequence Starting Point | Generated by AI

Home 2026.07

Which experiment to run first

The experiments are designed sequentially — but per the project notes, the star is Experiment 4. Here’s the recommended order:

🥇 Start with Experiment 4 — Scheduler Watch

cd /mnt/data/nano-vllm
python3.12 experiments/04_scheduler_watch.py

This is described as “the ‘loss curve moment’ for inference systems” — the single most important experiment. You watch continuous batching happen in real time: 10 requests start together, then progressively finish and leave the batch. It’s the core vLLM insight made visible.

After that, follow this order for the full picture

#	Experiment	Why
4	`04_scheduler_watch.py` ⭐	See continuous batching with your own eyes
1	`01_input_length.py`	Understand prefill (compute-bound) vs decode (memory-bound)
2	`02_generation_length.py`	See KV cache grow and slow down decode
3	`03_batch_scaling.py`	See why batching helps but also creates problems
5	`05_kv_cache_pressure.py`	See what happens when KV cache runs out (preemption)
6	`06_config_sweep.py`	Tune the trade-off knobs
7	`07_mixed_length_batch.py` ⭐	See why continuous batching beats static batching 10x

Commands for any experiment

cd /mnt/data/nano-vllm
python3.12 experiments/04_scheduler_watch.py   # Start here
python3.12 experiments/07_mixed_length_batch.py # Also starred
python3.12 experiments/01_input_length.py       # Then fundamentals
# ... etc

Important: Always use python3.12 (not python3). The system python3 is 3.14.3 which is outside nano-vllm’s supported range (>=3.10,<3.13).

Back Donate