Update modeling_gemmoe.py

modeling_gemmoe.py

@@ -25,8 +25,6 @@ import torch.utils.checkpoint
 from torch import nn
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
-from .mlp import MLP
-
 from transformers.activations import ACT2FN
 from transformers.cache_utils import Cache, DynamicCache, StaticCache
 from transformers.modeling_attn_mask_utils import (
@@ -631,39 +629,61 @@ 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)
+
+
 class GemmoeSparseMoeBlock(nn.Module):
     def __init__(self, config):
         super().__init__()
         self.hidden_dim = config.hidden_size
         self.ffn_dim = config.intermediate_size
         self.num_experts = config.num_local_experts
-        self.top_k = config.num_experts_per_tok
+        self.top_k = 2
 
         # gating
         self.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False)
 
-        self.mlp = MLP(
-            input_size=self.hidden_dim,
-            hidden_size=self.ffn_dim,
-            activation=nn.GELU(),
-            num_experts=self.num_experts,
-            top_k=self.top_k
-        )
+        self.experts = nn.ModuleList([GemmoeBlockSparseTop2MLP(config) for _ in range(self.num_experts)])
 
-    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+    def forward(self, hidden_states: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
         batch_size, sequence_length, hidden_dim = hidden_states.shape
         hidden_states = hidden_states.view(-1, hidden_dim)
-        
+
+        # router_logits: (batch * sequence_length, n_experts)
         router_logits = self.gate(hidden_states)
-        routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
-        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
-        routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
-        routing_weights = routing_weights.to(hidden_states.dtype)
+        routing_weights = F.softmax(router_logits, dim=1)
+        topk_weight, topk_idx = torch.topk(routing_weights, self.top_k, dim=-1, sorted=False)
+        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
 
-        hidden_states = self.mlp(hidden_states, routing_weights, selected_experts)
-        hidden_states = hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+        y = (y.view(*topk_weight.shape, -1) * topk_weight.unsqueeze(-1)).sum(dim=1)
 
-        return hidden_states, router_logits
+        final_hidden_states = y.reshape(batch_size, sequence_length, hidden_dim)
+        return final_hidden_states.to(hidden_states.dtype), router_logits.to(hidden_states.dtype)
 
     
 # Copied from transformers.models.llama.modeling_llama.LlamaDecoderLayer with LLAMA->GEMMOE,Llama->Gemmoe