[Feature Request] Llama 4

[yzh119]

Checklist

• Support temperature tuning
• chunk attention mask, we don't really need new kernels, just separate different chunk into different instances in the batch (considering they are totally independency).
  • We an example about this.

[yzh119]

For chunk attention layers, we don't need to create page tables for these layers, using a fixed-length (chunk size) buffer is enough.