rxnn 0.1.44__tar.gz → 0.1.46__tar.gz

{rxnn-0.1.44 → rxnn-0.1.46}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: rxnn
-Version: 0.1.44
+Version: 0.1.46
 Summary: RxNN: Reactive Neural Networks Platform
 License: Apache-2.0
 Keywords: deep-learning,ai,machine-learning

{rxnn-0.1.44 → rxnn-0.1.46}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "poetry.core.masonry.api"
 
 [tool.poetry]
 name = "rxnn"
-version = "0.1.44"
+version = "0.1.46"
 description = "RxNN: Reactive Neural Networks Platform"
 
 license = "Apache-2.0"

{rxnn-0.1.44 → rxnn-0.1.46}/src/rxnn/experimental/attention.py

@@ -95,16 +95,15 @@ class GroupedMoeAttention(GroupedQueryAttention):
         key_flat = key.reshape(-1, D)
         weights, indices = self.router(key_flat)  # (B*S, num_groups), (B*S, num_groups)
         weights = weights.view(B, S, self.num_groups, 1)
-        indices = indices.view(B, S, self.num_groups)
+        indices = indices.view(B, S, self.num_groups).unsqueeze(-1).transpose(1, 2).expand(-1, -1, S, -1)
 
         # Compute all experts' projections
         k_all = self.k_proj(key_flat).view(B, S, self.num_experts, -1).permute(0, 2, 1, 3) # [B, num_experts, S, head_dim]
         v_all = self.v_proj(value).view(B, S, self.num_experts, -1).permute(0, 2, 1, 3) # [B, num_experts, S, head_dim]
 
         # Gather top-k experts using expanded indices
-        expanded_indices = indices.unsqueeze(-1).unsqueeze(-1).expand(-1, -1, S, -1) # [B, num_groups, S, head_dim]
-        selected_k = torch.gather(k_all, 1, expanded_indices) # [B, num_groups, S, head_dim]
-        selected_v = torch.gather(v_all, 1, expanded_indices) # [B, num_groups, S, head_dim]
+        selected_k = torch.gather(k_all, 1, indices) # [B, num_groups, S, head_dim]
+        selected_v = torch.gather(v_all, 1, indices) # [B, num_groups, S, head_dim]
 
         # Weighted
         weighted_k = (selected_k * weights).to(selected_k.device, dtype=selected_k.dtype) # [B, S, num_groups, head_dim]
@@ -209,13 +208,12 @@ class DeepMoeAttention(GroupedMoeAttention):
         query_flat = query.reshape(-1, D)
         weights, indices = self.query_router(query_flat)
         weights = weights.view(B, T, self.num_query_groups, 1)
-        indices = indices.view(B, T, self.num_query_groups)
+        indices = indices.view(B, T, self.num_query_groups).unsqueeze(-1).transpose(1, 2).expand(-1, -1, T, -1) # [B, num_query_groups, T, head_dim]
 
         q_all = self.q_proj(query_flat).view(B, T, self.num_query_experts, -1).permute(0, 2, 1, 3) # [B, num_query_experts, T, head_dim]
 
         # Gather top-k experts using expanded indices
-        expanded_indices = indices.unsqueeze(-1).unsqueeze(-1).expand(-1, -1, T, -1) # [B, num_query_groups, T, head_dim]
-        selected_q = torch.gather(q_all, 1, expanded_indices) # [B, num_query_groups, T, head_dim]
+        selected_q = torch.gather(q_all, 1, indices) # [B, num_query_groups, T, head_dim]
 
         # Weighted sum
         q = (selected_q * weights).to(selected_q.device, dtype=selected_q.dtype)  # [B, T, num_query_groups, head_dim]