KV Cache getting prefilled and emptied during training phase

Although the KV cache implementation for inference looks good, the block_kv_cache is getting prefilled in the training phase leading to increase in compute and memory consumption.

https://github.com/huggingface/nanoVLM/blob/098db57e18b867aee604d8e8b2923b6bff1e6657/models/vision_language_model.py#L51

Start in models/vision_language_model.py around line 51 and trace how block_kv_cache is populated during training. Confirm the training path avoids unnecessary cache prefill and that compute and memory consumption no longer increase; validate the behavior with a training run or the repository’s existing checks.

Written by the indexing model from the issue text.