Beware speed exaggeration

DFlash's 4-5x speed advantage appears to hold only in limited cases.

• With MoE, the gains aren't large, and the core value is closer to dense models.
• Applying quantization reduces the gains. Q8_0 still shows improvement, but Q4_0 shows almost none.
• As concurrent users increase, the speedup drops sharply. It's about half at 2 users, around 20% at 4 users, and close to 0% at 8 users.
• These results are based on very short context, so performance may degrade further as context grows longer.

From a practical standpoint, both general users with 8/16GB VRAM and small server environments may not feel much benefit, and the argument is that we should watch DDTree, which is expected to show more consistent results.