RTX6k Qwen3.5 Benchmark Updated

After upgrading llama.cpp from b8198 -> d05fe1d, Qwen3.5-122B-A10B (MXFP4_MOE) was re-measured on RTX6k (Server, 450W) with the same model, same hardware, and same flags. CUDA toolkit remained at 13.0, and the new build includes hybrid SSM/MoE speculative decode, native NVFP4 MMQ for SM120, server prompt cache, and MoE/MXFP4 kernel improvements from the last 2 months.

Measurements were performed by sending synthetic prompts to the v1/completions endpoint of a running server. cache_prompt:false, n_predict=128-256, temperature=0, and chat_template_kwargs.enable_thinking=false were used, and speed was parsed from the server response. The production setting was changed from --parallel 4 to --parallel 2.

The key results are as follows.

• pp512: 2,188 -> 3,176 t/s, +45%
• tg single-stream: 80.0 -> 106.6 t/s, +33%
• tg per-req @ c=2: 55.7 -> 89.3 t/s, +60%
• Total tg @ c=2: 111.4 -> 178.6 t/s, +60%
• pp 8K / 32K / 65K: +69% / +65% / +59%
• tg 8K / 32K / 65K: +35% / +35% / +28%
• TTFT 8K / 32K / 65K: 2,780 / 10,780 / 23,161 ms -> 1,877 / 7,955 / 17,737 ms

The improvement in prompt processing was larger than in token generation, and TG dropped by about -13% from 0->65K, maintaining a slope similar to the previous -10%.