What makes this different from the base model?
#1
by
nmkd
- opened
What exactly does Fast-Preprocessor mean?
By default, the preprocessor here is set to "slow". This was required when Qwen 2.5 VL was released, as there were large deviations in results between the fast and slow preprocessors. This issue has been somewhat fixed now (PR).
In vLLM, I was experiencing strange behavior and couldn't fully saturate the GPU, so I wanted to try the fast preprocessor. Since you can't override this value at runtime in vLLM, I uploaded a copy with the preprocessor set to "fast".