跑通单轮推理

Author: onenewcode

models/minicpm3/common/src/compute.rs

@@ -191,7 +191,10 @@ where
             // norm
             let w = self.weights.attn_norm(iblk, queue);
             self.rms_norm(&mut x1, &x, &w, workspace, queue_alloc)?;
-
+            // if iblk==1{
+            //     Ops::debug(&x1, queue);
+            //     todo!();
+            // }
             drop(w);
             {
                 let w = self.weights.attn_qa(iblk, queue).transpose(&[1, 0]);