MTPLX/Benchmarks/Receipt 2026-07-02

81.74 tok/s, the raw logs.

Seven mtplx bench tune --max-tokens 192 --depths 3 runs on 2 July 2026, Qwen 3.6 27B Optimized Speed, M5 Max. Each log records the fan ramp (target 7,826 RPM, actual 7,821 to 7,830), the plain-decode baseline, the depth-3 result, acceptance by depth and power telemetry. Machine paths are replaced with placeholders; nothing else is edited.

The runs are an A/B between MTPLX's own verify kernels, with and without the compiled verify graph, and a port of a third-party kernel used as the control. The record cell is MTPLX's own kernels plus the compiled graph, twin-confirmed. The plain-decode baseline is flat across all seven runs (30.23 to 30.43 tok/s), so the multiplier rests on a flat baseline.

LogConfigurationAR tok/sD3 tok/sMultiplier
oursW2-1.logMTPLX kernels + compiled verify (record)30.3781.742.69x
oursW2-2.logMTPLX kernels + compiled verify (twin)30.2381.732.70x
portW2-1.logported kernel + compiled verify30.3381.682.69x
portW2-2.logported kernel + compiled verify (twin)30.3381.642.69x
port1.logported kernel, eager30.3875.462.48x
port2.logported kernel, eager (twin)30.4375.482.48x
ours-plain.logMTPLX kernels, eager30.2574.772.47x

Acceptance by depth on every run: 50/53, 47/53, 42/53 (94.3, 88.7, 79.2 percent). Sampler: temperature 0.6, top-p 0.95, top-k 20, thinking off. The 192-token generation is the burst lane, the shape of a short chat reply. The same configuration on an uncapped 11,390-token Flappy Bird generation with reasoning on ran 62.95 tok/s the same afternoon (75.5 tok/s over the first 128 tokens, 39.9 over the last 128); the compiled verify graph shipped as the turbo default in MTPLX 2.0.0 four days later.