rxnn 0.1.40__tar.gz → 0.1.42__tar.gz

{rxnn-0.1.40 → rxnn-0.1.42}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: rxnn
-Version: 0.1.40
+Version: 0.1.42
 Summary: RxNN: Reactive Neural Networks Platform
 License: Apache-2.0
 Keywords: deep-learning,ai,machine-learning

{rxnn-0.1.40 → rxnn-0.1.42}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "poetry.core.masonry.api"
 
 [tool.poetry]
 name = "rxnn"
-version = "0.1.40"
+version = "0.1.42"
 description = "RxNN: Reactive Neural Networks Platform"
 
 license = "Apache-2.0"

{rxnn-0.1.40 → rxnn-0.1.42}/src/rxnn/experimental/attention.py

@@ -90,6 +90,7 @@ class GroupedMoeAttention(GroupedQueryAttention):
 
         # Key/Value MoE routing
         B, S, D = key.shape
+        print('key/value type', key.dtype, value.dtype)
         key_flat = key.reshape(-1, D)
         print('key flat type', key_flat.dtype)
         weights, indices = self.router(key_flat)  # (B*S, num_groups), (B*S, num_groups)
@@ -127,8 +128,8 @@ class GroupedMoeAttention(GroupedQueryAttention):
         print(selected_v.size(), selected_v.dtype)
 
         # Weighted
-        weighted_k = selected_k * weights # [B, S, num_groups, head_dim]
-        weighted_v = selected_v * weights # [B, S, num_groups, head_dim]
+        weighted_k = (selected_k * weights).to(selected_k.device, dtype=selected_k.dtype) # [B, S, num_groups, head_dim]
+        weighted_v = (selected_v * weights).to(selected_k.device, dtype=selected_k.dtype)  # [B, S, num_groups, head_dim]
 
         print('weighted')
         print(weighted_k.size(), weighted_k.dtype)
@@ -252,7 +253,7 @@ class DeepMoeAttention(GroupedMoeAttention):
         selected_q = torch.gather(q_all, 2, expanded_indices)  # [B, T, num_query_groups, head_dim]
 
         # Weighted sum
-        q = selected_q * weights_q  # [B, T, num_query_groups, head_dim]
+        q = (selected_q * weights_q).to(selected_q.device, dtype=selected_q.dtype)  # [B, T, num_query_groups, head_dim]
         q = q.view(B, T, self.num_query_groups, -1).permute(0, 2, 1, 3)  # [B, num_query_groups, T, head_dim]
 
         return super()._forward_qkv(q, key, value, b, t, d, skip_query_processing=True)

{rxnn-0.1.40 → rxnn-0.1.42}/src/rxnn/transformers/attention.py

@@ -137,6 +137,10 @@ class MultiHeadAttention(nn.Module):
         return self._calculate_output(attn_weights, v, b, t, d)
 
     def forward(self, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, mask: torch.Tensor = None):
+        print('MHA forward')
+        print(query.size(), query.dtype)
+        print(key.size(), key.dtype)
+        print(value.size(), value.dtype)
         b, t, d = query.size()
         q, k, v = self._forward_qkv(query, key, value, b, t, d)
         if not self.rel_embed: