[Benchmark] DFlash2 vs MTP comparison. 5090RTX, Qwen 3.8 27B, Dynamic v3 GGUF, llama.cpp. Token generation, latency and available context.

TLDR: Caution advised. DFlash2 gives +20% generation speed at the expense of 38% of context size. Long running tasks and large context work (>100k) makes DFlash2 lose it's advantage due to compaction and diminishing returns. TPS comparison Input Prompt parsing DFlash2 Prompt parsing MTP Parsing winner Generation DFlash2 Generation MTP Generation winner 4.2K 1,916.65 t/s 2,139.07 t/s MTP +11.6% 104.43 t/s 86.50 t/s DFlash2 +20.7% 16.4K 2,226.55 t/s 2,299.56 t/s MTP +3.3% 101.97 t/s 87.04 t/s DFlash2 +17.1% 65.6K 1,950.37 t/s 1,956.50 t/s Effectively tied 88.33 t/s 72.18 t/s DFlash2 +22.4% Weighted 1,995.84 t/s 2,022.41 t/s MTP +1.3% 97.71 t/s 81.29 t/s DFlash2 +20.2% Total round-trip latency Input DFlash2 MTP DFlash2 advantage 4.2K 7.09 s 7.88 s 11.1% 16.4K 12.40 s 13.03 s 5.1% 65.6K 39.42 s 40.62 s 3.0% Context size DFlash2: 90,112 MTP: 124,416 MTP provides 34,304 more tokens, or 38.1% more usable context. DFlash2 config (MTP is similar, with spec-draft-n-max 2) llama-server \ --model /home/human/.cache/llama.cpp/unsloth-qwen38-27b-dynamic3/Qwen3.8-27B-UD-Q8_K_L.gguf \ --mmproj /home/human/.cache/llama.cpp/unsloth-qwen38-27b/mmproj-F16.gguf \ --no-mmproj-offload \ --image-min-tokens 1024 \ --alias qwen3.8-27b-q8-0 \ --ctx-size 90112 \ --gpu-layers all \ --fit off \ --parallel 1 \ --batch-size 512 \ --ubatch-size 128 \ --cache-type-k q8_0 \ --cache-type-v q8_0 \ --kv-offload \ --kv-unified \ --cache-ram 8192 \ --cache-idle-slots \ --flash-attn on \ --jinja \ --reasoning on \ --reasoning-preserve \ --spec-type draft-dflash \ --spec-draft-model /home/human/.cache/llama.cpp/unsloth-qwen38-27b-dynamic3/Qwen3.8-27B-DFlash2-Q8_0.gguf \ --spec-draft-n-max 4 \ --spec-draft-type-k q8_0 \ --spec-draft-type-v q8_0 \ --spec-draft-ngl all \