rxnn 0.1.31__tar.gz → 0.1.32__tar.gz

{rxnn-0.1.31 → rxnn-0.1.32}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: rxnn
-Version: 0.1.31
+Version: 0.1.32
 Summary: RxNN: Reactive Neural Networks Platform
 License: Apache-2.0
 Keywords: deep-learning,ai,machine-learning

{rxnn-0.1.31 → rxnn-0.1.32}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "poetry.core.masonry.api"
 
 [tool.poetry]
 name = "rxnn"
-version = "0.1.31"
+version = "0.1.32"
 description = "RxNN: Reactive Neural Networks Platform"
 
 license = "Apache-2.0"

{rxnn-0.1.31 → rxnn-0.1.32}/src/rxnn/experimental/attention.py

@@ -323,7 +323,7 @@ class GroupedMoeAttentionVectorized(GroupedQueryAttention):
 
         # Key/Value MoE routing
         B, S, D = key.shape
-        key_flat = key.reshape(B * S, D)
+        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)
@@ -331,35 +331,28 @@ class GroupedMoeAttentionVectorized(GroupedQueryAttention):
         # Compute all experts' projections
         # Shape: (B*S, num_experts, head_dim)
         k_all = torch.einsum('bd,edh->beh', key_flat, self.wk)  # [B*S, num_experts, head_dim]
-        v_all = torch.einsum('bd,edh->beh', value.view(B*S, D), self.wv)
+        v_all = torch.einsum('bd,edh->beh', value.view(-1, D), self.wv)
 
-        # Reshape to [B, S, num_experts, head_dim]
-        k_all = k_all.view(B, S, self.num_experts, -1)
-        v_all = v_all.view(B, S, self.num_experts, -1)
-
-        print('k_all', k_all.size())
-        print('v_all', v_all.size())
-
-
-        # Gather top-k experts and weights
-        # Expand indices to [B, S, num_groups, head_dim]
-        expanded_indices = indices.unsqueeze(-1).expand(-1, -1, -1, k_all.size(-1))
-        selected_k = torch.gather(k_all, 2, expanded_indices)  # [B, S, num_groups, head_dim]
-        selected_v = torch.gather(v_all, 2, expanded_indices)
+        if self.use_bias:
+            k_all += self.bk
+            v_all += self.bv
 
-        print('selected_k', selected_k.size())
-        print('selected_v', selected_v.size())
+        # Get results for all heads
+        k_all = k_all.view(B, S, self.num_experts, -1) # [B, S, num_experts, head_dim]
+        v_all = v_all.view(B, S, self.num_experts, -1) # [B, S, num_experts, head_dim]
 
-        # Weighted sum
-        weighted_k = (selected_k * weights).sum(dim=2)  # [B, S, head_dim]
-        weighted_v = (selected_v * weights).sum(dim=2)
+        # Gather top-k experts using expanded indices
+        expanded_indices = indices.unsqueeze(-1).expand(-1, -1, -1, k_all.size(-1)) # [B, S, num_groups, head_dim]
+        selected_k = torch.gather(k_all, 2, expanded_indices) # [B, S, num_groups, head_dim]
+        selected_v = torch.gather(v_all, 2, expanded_indices) # [B, S, num_groups, head_dim]
 
-        print('weighted_k', weighted_k.size())
-        print('weighted_v', weighted_v.size())
+        # Weighted
+        weighted_k = selected_k * weights # [B, S, num_groups, head_dim]
+        weighted_v = selected_v * weights # [B, S, num_groups, head_dim]
 
         # Reshape to GQA format
-        k = weighted_k.view(B, S, self.num_groups, -1).permute(0, 2, 1, 3)  # [B, G, S, head_dim]
-        v = weighted_v.view(B, S, self.num_groups, -1).permute(0, 2, 1, 3)
+        k = weighted_k.view(B, S, self.num_groups, -1).permute(0, 2, 1, 3) # [B, num_groups, S, head_dim]
+        v = weighted_v.view(B, S, self.num_groups, -1).permute(0, 2, 1, 3) # [B, num_groups, S, head_dim]
 
         if not self.use_flash_attention:
             group_heads = self.num_heads // self.num_groups
@@ -370,10 +363,6 @@ class GroupedMoeAttentionVectorized(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', q.size())
-        print('k', k.size())
-        print('v', v.size())
-
         return q, k, v
 
 
@@ -458,13 +447,16 @@ class DeepMoeAttentionVectorized(GroupedMoeAttentionVectorized):
 
     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(B * T, D)
+        query_flat = query.reshape(-1, D)
         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)
 
         # Compute all query experts
         q_all = torch.einsum('bd,edh->beh', query_flat, self.wq)  # [B*T, num_query_experts, head_dim]
+        if self.use_bias:
+            q_all += self.bq
+
         q_all = q_all.view(B, T, self.num_query_experts, -1)
 
         # Gather top-k experts
@@ -472,12 +464,11 @@ class DeepMoeAttentionVectorized(GroupedMoeAttentionVectorized):
         selected_q = torch.gather(q_all, 2, expanded_indices)  # [B, T, num_query_groups, head_dim]
 
         # Weighted sum
-        q = (selected_q * weights_q).sum(dim=2)  # [B, T, head_dim]
-        q = q.view(B, T, self.num_query_groups, -1).permute(0, 2, 1, 3)  # [B, H_q, T, head_dim]
+        q = selected_q * weights_q  # [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)
 
-
 # Others
 
 class FlexAttention(MultiHeadAttention):