Update modeling_gemmoe.py

modeling_gemmoe.py

@@ -629,23 +629,32 @@ GEMMOE_ATTENTION_CLASSES = {
     "sdpa": GemmoeSdpaAttention,
 }
 
-class GemmoeBlockSparseTop2MLP(nn.Module):
-    def __init__(self, config: GemmoeConfig):
-        super().__init__()
-        self.ffn_dim = config.intermediate_size
-        self.hidden_dim = config.hidden_size
-        
-        self.w1 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
-        self.w2 = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False)
-        self.w3 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False)
-
-        self.act_fn = approx_gelu
-
-    def forward(self, hidden_states):
-        current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(hidden_states)
-        current_hidden_states = self.w2(current_hidden_states)
-        return current_hidden_states.to(hidden_states.dtype)
+scatter2scatter=False
 
+class ParallelLinear(nn.Module):
+    def __init__(self, in_features, out_features, num_experts):
+        super().__init__()
+        self.in_features = in_features
+        self.out_features = out_features
+        self.num_experts = num_experts
+        self.weight = nn.Parameter(torch.Tensor(num_experts, in_features, out_features))
+        nn.init.xavier_uniform_(self.weight)
+
+    def forward(self, input, ordering, top_k, routing_weights=None, grouped_in=False, grouped_out=False, use_scatter2scatter=False):
+        if use_scatter2scatter:
+            output = scatter2scatter(input, self.weight, ordering, top_k, grouped_in=grouped_in, grouped_out=grouped_out)
+        else:
+            if not grouped_in:
+                input = input[ordering]
+            output = torch.bmm(input.view(-1, top_k, self.in_features), self.weight[ordering].view(-1, self.in_features, self.out_features))
+            if not grouped_out:
+                output = output.view(-1, top_k * self.out_features)
+                output = output[torch.argsort(ordering)]
+
+        if routing_weights is not None:
+            output = output.view(-1, top_k, self.out_features)
+            output = torch.bmm(routing_weights.unsqueeze(1), output).squeeze(1)
+        return output
 
 class GemmoeSparseMoeBlock(nn.Module):
     def __init__(self, config):
@@ -654,11 +663,14 @@ class GemmoeSparseMoeBlock(nn.Module):
         self.ffn_dim = config.intermediate_size
         self.num_experts = config.num_local_experts
         self.top_k = 2
+        self.use_scatter2scatter = config.use_scatter2scatter if hasattr(config, 'use_scatter2scatter') else False
 
         # gating
         self.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False)
 
-        self.experts = nn.ModuleList([GemmoeBlockSparseTop2MLP(config) for _ in range(self.num_experts)])
+        self.expert_mlp1 = ParallelLinear(self.hidden_dim, self.ffn_dim, self.num_experts)
+        self.expert_mlp2 = ParallelLinear(self.ffn_dim, self.hidden_dim, self.num_experts)
+        self.activation = nn.GELU()
 
     def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
         batch_size, sequence_length, hidden_dim = hidden_states.shape
@@ -667,22 +679,18 @@ class GemmoeSparseMoeBlock(nn.Module):
         # router_logits: (batch * sequence_length, n_experts)
         router_logits = self.gate(hidden_states)
         routing_weights = F.softmax(router_logits, dim=1)
-        topk_weight, topk_idx = torch.topk(routing_weights, self.top_k, dim=-1, sorted=False)
+        _, topk_idx = torch.topk(routing_weights, self.top_k, dim=-1, sorted=False)
+        topk_weight = routing_weights.gather(1, topk_idx)
         topk_weight /= topk_weight.sum(dim=-1, keepdim=True)
 
-        hidden_states = hidden_states.repeat_interleave(self.top_k, dim=0)
-
-        y = torch.empty_like(hidden_states)
-
-        flat_topk_idx = topk_idx.view(-1)
-        for i in range(self.num_experts):
-            expert = self.experts[i]
-            expert_output = expert(hidden_states[flat_topk_idx == i])
-            y[flat_topk_idx == i] = expert_output
+        ordering = topk_idx.view(-1)
 
-        y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)
+        # ParallelLinear for expert MLP
+        hidden_states = self.expert_mlp1(hidden_states, ordering, self.top_k, grouped_in=False, grouped_out=True, use_scatter2scatter=self.use_scatter2scatter)
+        hidden_states = self.activation(hidden_states)
+        hidden_states = self.expert_mlp2(hidden_states, ordering, self.top_k, topk_weight, grouped_in=True, grouped_out=False, use_scatter2scatter=self.use_scatter2scatter)
 
-        final_hidden_states = y.reshape(batch_size, sequence_length, hidden_dim)
+        final_hidden_states = hidden_states.reshape(batch_size, sequence_length, hidden_dim)
         return final_hidden_states.to(hidden_states.dtype), router_logits.to(hidden_states.dtype)