rxnn 0.1.48__py3-none-any.whl → 0.1.49__py3-none-any.whl

rxnn/experimental/attention.py

@@ -94,16 +94,17 @@ class GroupedMoeAttention(GroupedQueryAttention):
         B, S, D = key.shape
         key_flat = key.reshape(-1, D)
         weights, indices = self.router(key_flat)  # (B*S, num_groups), (B*S, num_groups)
-        weights = weights.view(B, self.num_groups, S, 1)
-        indices = indices.view(B, self.num_groups, S).unsqueeze(-1).expand(-1, -1, S, -1)
+        weights = weights.view(B, S, self.num_groups, 1)
+        indices = indices.view(B, S, self.num_groups)
 
         # 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]
+        k_all = self.k_proj(key_flat).view(B, S, self.num_experts, -1) # [B, S, num_experts, head_dim]
+        v_all = self.v_proj(value).view(B, S, self.num_experts, -1) # [B, S, num_experts, head_dim]
 
         # Gather top-k experts using expanded indices
-        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]
+        expanded_indices = indices.unsqueeze(-1).expand(-1, -1, -1, k_all.size(-1)) # [B, num_groups, S, head_dim]
+        selected_k = torch.gather(k_all, 2, expanded_indices) # [B, num_groups, S, head_dim]
+        selected_v = torch.gather(v_all, 2, expanded_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]
@@ -122,8 +123,6 @@ class GroupedMoeAttention(GroupedQueryAttention):
             k = k.flatten(start_dim=1, end_dim=2)  # (B, H, S, head_dim)
             v = v.flatten(start_dim=1, end_dim=2)  # (B, H, S, head_dim)
 
-        print(q.size(), k.size(), v.size())
-
         return q, k, v
 
 
@@ -208,17 +207,18 @@ class DeepMoeAttention(GroupedMoeAttention):
     def _forward_qkv(self, query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, b: int, t: int, d: int):
         B, T, D = query.shape
         query_flat = query.reshape(-1, D)
-        weights, indices = self.query_router(query_flat)
-        weights = weights.view(B, self.num_query_groups, T, 1)
-        indices = indices.view(B, self.num_query_groups, T).unsqueeze(-1).expand(-1, -1, T, -1) # [B, num_query_groups, T, head_dim]
+        weights_q, indices_q = self.query_router(query_flat)
+        weights_q = weights_q.view(B, T, self.num_query_groups, 1)
+        indices_q = indices_q.view(B, T, self.num_query_groups)
 
-        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]
+        q_all = self.q_proj(query_flat).view(B, T, self.num_query_experts, -1) # [B, num_groups, S, head_dim]
 
-        # Gather top-k experts using expanded indices
-        selected_q = torch.gather(q_all, 1, indices) # [B, num_query_groups, T, head_dim]
+        # Gather top-k experts
+        expanded_indices = indices_q.unsqueeze(-1).expand(-1, -1, -1, q_all.size(-1)) # [B, T, num_query_groups, head_dim]
+        selected_q = torch.gather(q_all, 2, expanded_indices)  # [B, T, num_query_groups, head_dim]
 
         # Weighted sum
-        q = (selected_q * weights).to(selected_q.device, dtype=selected_q.dtype)  # [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-0.1.48.dist-info → rxnn-0.1.49.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: rxnn
-Version: 0.1.48
+Version: 0.1.49
 Summary: RxNN: Reactive Neural Networks Platform
 License: Apache-2.0
 Keywords: deep-learning,ai,machine-learning

{rxnn-0.1.48.dist-info → rxnn-0.1.49.dist-info}/RECORD

@@ -1,6 +1,6 @@
 rxnn/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 rxnn/experimental/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-rxnn/experimental/attention.py,sha256=XaN7IAHhp1DxFtZxeTEP-EZN7PWnjocC42ndSDB9RvY,29432
+rxnn/experimental/attention.py,sha256=22Qb4jYN6QaqibTU8bwD8x2FaOKCxvWglM2eK9EuOlo,29468
 rxnn/experimental/models.py,sha256=IzUVc5s-cA__8jsG2mVvzUDmzPRcfBcI5btaOjnPYhA,4598
 rxnn/experimental/moe.py,sha256=jHZ1QhpWiVQOswVpFmuH7b2IUOPf0Uuf-I2Ddwsd7Us,6140
 rxnn/memory/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -25,7 +25,7 @@ rxnn/transformers/moe.py,sha256=6Cffyo0QjmEWc4rK1ncOmLRCQbY0OpQJ4D7xH_4nTN4,4738
 rxnn/transformers/positional.py,sha256=2l38RS0Dini3f6Z3LUHr3XwWzg1UK7fO2C6wazWDAYU,4292
 rxnn/transformers/sampler.py,sha256=poWBpxg1iuK5gEJtxHkk5VVfS9V48hs2Olqdhy_Gw8c,6548
 rxnn/utils.py,sha256=d5U8i5ukovgDyqiycc2AoxObTz_eF_bgo2MKvdtJ98s,467
-rxnn-0.1.48.dist-info/LICENSE,sha256=C8coDFIUYuOcke4JLPwTqahQUCyXyGq6WOaigOkx8tY,11275
-rxnn-0.1.48.dist-info/METADATA,sha256=YelXCeEWnK9llWGzSHOiThlUIz8ttWH-KsL_68pJ9-Y,16627
-rxnn-0.1.48.dist-info/WHEEL,sha256=fGIA9gx4Qxk2KDKeNJCbOEwSrmLtjWCwzBz351GyrPQ,88
-rxnn-0.1.48.dist-info/RECORD,,
+rxnn-0.1.49.dist-info/LICENSE,sha256=C8coDFIUYuOcke4JLPwTqahQUCyXyGq6WOaigOkx8tY,11275
+rxnn-0.1.49.dist-info/METADATA,sha256=PijR2z5P5nuTlOaWn-ylU_Loluy-e2HRgpMEc4TCohk,16627
+rxnn-0.1.49.dist-info/WHEEL,sha256=fGIA9gx4Qxk2KDKeNJCbOEwSrmLtjWCwzBz351GyrPQ,88
+rxnn-0.1.49.dist-info/RECORD,,