Scalpel-VL-1.7B
A 1.7-billion-parameter vision-language model that applies a "scalpel"-style precision optimization approach. The method identifies and removes redundant visual tokens, attention heads, or intermediate-layer activations, then completes fine-tuning with a minimal number of recovery tokens, achieving inference acceleration while preserving model performance.
Timeline (last 90 days)
Scalpel-VL-1.7B claims to achieve a 20%–30% inference speedup while using only 0.1B (100 million) recovery tokens.
Scalpel-VL is a visual language model with 1.7B parameters.
The core of the Scalpel approach lies in identifying and removing redundant parts of the computation process, then fine-tuning the model with only a minimal number of recovery tokens to adapt it to the pruned structure.
If the speedup mainly comes from reducing the number of visual tokens, the improvement in prefill latency is more significant, while the impact on decoding speed for long text generation is limited.
Scalpel-VL is a community project with relatively limited public information available; benchmark details such as hardware, batch size, and tasks for its speedup data have not yet been specified.
All Facts (5)
Scalpel-VL-1.7B claims to achieve a 20%–30% inference speedup while using only 0.1B (100 million) recovery tokens.
50%UnverifiedScalpel-VL is a visual language model with 1.7B parameters.
50%UnverifiedThe core of the Scalpel approach lies in identifying and removing redundant parts of the computation process, then fine-tuning the model with only a minimal number of recovery tokens to adapt it to the pruned structure.
50%UnverifiedIf the speedup mainly comes from reducing the number of visual tokens, the improvement in prefill latency is more significant, while the impact on decoding speed for long text generation is limited.
50%UnverifiedScalpel-VL is a community project with relatively limited public information available; benchmark details such as hardware, batch size, and tasks for its speedup data have not yet been specified.
50%