{"model":"deepseek-4.1-flash","title":"Benchmark #3 - arrangements + DSpark real-text correction","status":"positive","summary":"Random-token acceptance (4-8%) was misleading. Real prose (TP4xPP2, k=5): c1 43.1 tok/s (accept 41.2%), c4 111.0, c8 153.4 - c1 matches Schaka's 43.8. Arrangements: TP8xPP1 wins c1 (20.2), TP2xPP3 wins aggregate + TTFT. Long-context prefill 1,704 @8k to 2,873 tok/s @256k.","occurred_at":"2026-09-11T05:22:00Z","config":{"serving":"TP4xPP2, DSpark k=5, fp8_ds_mla, 1M ctx, CUDA graphs","engine":"localhost/vllm-backport-v41:sm80 (vLLM 0.12.0-sm80 + PR#56201 + Ampere shims + PP relay)"},"benchmarks":[{"metric":"c1_decode_tok_s","value":43.1,"unit":"tok/s","context":"real prose, TP4xPP2, DSpark k=5"},{"metric":"agg_decode_c4_tok_s","value":111,"unit":"tok/s","context":"real prose, TP4xPP2"},{"metric":"agg_decode_c8_tok_s","value":153.4,"unit":"tok/s","context":"real prose, TP4xPP2"},{"metric":"prefill_tok_s","value":2280,"unit":"tok/s","context":"32k ctx (1,704 @8k; 2,247 @128k; 2,873 @256k)"}]}