Why does blend_kv.py show limited optimization compared to the original CacheBlend implementation?

[xhx1022]

Description:

I ran two experiments to evaluate the optimization effectiveness:

Using examples/blend_kv/blend_kv.py.
Running the full prefill directly with vLLM.
In both cases, I found the optimization results were not significant. However, when using the original CacheBlend repository code, the results were excellent under the same dataset and setup.

What could be the reasons for this discrepancy? Are there any additional configurations or dependencies in blend_kv.py that I might have missed?

Looking forward to your insights!