attempt #1

Author: oppiliappan

crates/models/bert/src/lib.rs

@@ -241,7 +241,7 @@ impl KnownModel for Bert {
 
                 // self-attention
                 {
-                    current = ctx0.op_cont(&current);
+                    print_shape(&current, "current");
                     let q_current = ctx0.op_reshape_3d(
                         &ctx0.op_add(
                             &ctx0.op_mul_mat(&self.layers[il].q_w, &current),
@@ -263,6 +263,10 @@ impl KnownModel for Bert {
                         input_len,
                     );
                     let k = ctx0.op_permute(&k_current, (0, 2, 1, 3));
+                    let k = ctx0.op_cpy(
+                        &k,
+                        &ctx0.new_tensor_3d(ggml::Type::F32, d_head, input_len, n_head),
+                    );
 
                     let v_current = ctx0.op_reshape_3d(
                         &ctx0.op_add(
@@ -274,9 +278,13 @@ impl KnownModel for Bert {
                         input_len,
                     );
                     let mut v = ctx0.op_permute(&v_current, (0, 2, 1, 3));
+                    v = ctx0.op_cpy(
+                        &v,
+                        &ctx0.new_tensor_3d(ggml::Type::F32, d_head, input_len, n_head),
+                    );
 
-                    let k = ctx0.op_cont(&k);
-                    let q = ctx0.op_cont(&q);
+                    print_shape(&k, "k");
+                    print_shape(&q, "q");
                     let mut kq = ctx0.op_mul_mat(&k, &q);
 
                     // TODO: look into op_scale_inplace and op_soft_max_inplace
@@ -465,3 +473,7 @@ struct Layer {
     ff_o_w: ggml::Tensor,
     ff_o_b: ggml::Tensor,
 }
+
+fn print_shape(t: &ggml::Tensor, name: &str) {
+    println!("{name} {} {:?}", t.get_type(), t.get_ne());
+}