# Probes and harnesses
Everything here lived only on the benchmark box and would have been lost when it was terminated.
| `fp8_vs_bf16.cu` | **The FP8 spike.** Self-contained C + cuBLASLt on our exact gate|up shape. Measured **FP8 1064 TFLOP/s vs BF16 627 = 1.70x** on an H100 PCIe, which is what makes the prefill plan viable. Also the working reference for the Lt call sequence, since `cublasGemmEx` cannot do FP8 and cudarc's safe `Matmul` trait has no FP8 impl. Build: `nvcc -O3 -arch=sm_90 -lcublasLt -o fp8probe fp8_vs_bf16.cu`. |
| `pf.py` | Prefill throughput / TTFT from streamed TTFT, exact token ids, leading id perturbed per repeat so no prefix cache can serve it. |
| `conc.py` | Concurrency: M simultaneous streams, aggregate tokens over the burst's wall clock. **Note: its wall clock INCLUDES the prefill**, which matters when reading long-context numbers. |
| `strict.py` | Single-stream head-to-head that also diffs generated ids, doubling as a correctness gate. |
| `lctx.sh` | Long-context single-stream across all three engines (ours, vLLM, SGLang), start to finish. |
| `ttft.sh` | TTFT across ours and vLLM, short and long prompt. |
| `resume.sh` | Prints every command line for the box (build, gates, ncu, all three servers). |
All were run against each engine's OpenAI-compatible endpoint with **identical prompt token ids**
(ids, not text) and prefix/radix caching disabled everywhere.