Episode 02 · Planned
Equal-work runtime baseline
Compare HF/PyTorch, vLLM and SGLang under the same model, corpus, template and output contract.
Compare the same model, corpus, template, output contract and sampling settings across an HF/PyTorch control, vLLM and SGLang where compatible. Randomize or counterbalance execution order. Keep cold start separate from steady state, and compare output-length distributions before interpreting throughput.
Evidence: selected images and kernels, tokenizer/template hashes, per-repetition TTFT and end-to-end latency, output lengths, stop reasons, correctness checks, errors, device telemetry and attributable cost.