initial version to run with megatron

fmoe/megatron.py

@@ -5,8 +5,9 @@ from .moe import FFFN
 def create_moe_mlp(args):
     assert args.num_experts % args.model_parallel_size == 0, 'Num experts should be multiple of mp size'
     num_experts = args.num_experts // args.model_parallel_size 
-    fmoe = FFFN(num_experts, in_feat=args.hidden_size, 
-            hidden_feat=args.hidden_size * 4, out_feat=args.hidden_size,
-            world_size = args.model_parallel_size)
+    fmoe = FFFN(num_experts, 
+            d_model=args.hidden_size, 
+            d_hidden=args.hidden_size * 4,
+            world_size=args.model_parallel_size)
     return fmoe
     

fmoe/moe.py

 import math
 from torch import nn
 import torch
+import torch.nn.functional as F
 
 from .moe_function import moe
 
@@ -27,20 +28,55 @@ class FMoE(nn.Module):
 
 
 class FFFN(nn.Module):
-    def __init__(self, num_expert=32, in_feat=1024, hidden_feat=4096, 
-            out_feat=1024, world_size=None, activation=torch.nn.functional.gelu):
+    def __init__(self, num_expert=32, d_model=1024, d_hidden=4096, 
+            world_size=None, activation=torch.nn.functional.gelu,
+            top_k=2, pre_lnorm=False):
         super(FFFN, self).__init__()
-        self.htoh4 = FMoE(num_expert, in_feat, hidden_feat,
-                world_size=world_size)
+        self.d_model = d_model
+        self.d_hidden = d_hidden
+        self.world_size = world_size
         self.activation = activation
-        self.h4toh = FMoE(num_expert, hidden_feat, out_feat, 
+        self.top_k = top_k
+        self.pre_lnorm = pre_lnorm
+
+        self.htoh4 = FMoE(num_expert, d_model, d_hidden,
                 world_size=world_size)
+        self.h4toh = FMoE(num_expert, d_hidden, d_model, 
+                world_size=world_size)
+        self.gate = nn.Linear(d_model, num_expert)
+        self.layer_norm = nn.LayerNorm(d_model)
+        self.bias = torch.nn.parameter.Parameter(torch.zeros(d_model,
+                dtype=torch.float32)) 
 
-    def forward(self, inp, gate):
-        x = self.htoh4(inp)
+    def forward(self, inp):
+        # import pdb; pdb.set_trace()
+        residual = inp
+        if self.pre_lnorm:
+            inp = self.layer_norm(inp)
+
+        gate = self.gate(inp)
+        gate_top_k_val, gate_top_k_idx = torch.topk(gate, k=self.top_k, dim=-1,
+                largest=True, sorted=False) # [.. x top_k]
+        gate_top_k_val = gate_top_k_val.view(-1, self.top_k)
+
+        # (BxL) x 1 x top_k 
+        gate_score = F.softmax(gate_top_k_val, dim=-1).unsqueeze(1) 
+        gate_top_k_idx = gate_top_k_idx.view(-1) # (BxLxtop_k)
+
+        inp = inp.view(-1, self.d_model).repeat_interleave(repeats=self.top_k, 
+                dim=0) # (BxLxtop_k) x d_model
+        x = self.htoh4(inp, gate_top_k_idx)
         x = self.activation(x)
-        x = self.h4toh(x)
-        return x
+        x = self.h4toh(x, gate_top_k_idx)
+
+        core_out = x.view(-1, self.top_k, self.d_model) # (BxL) x top_k x d_model 
+        core_out = torch.bmm(gate_score, core_out) # (BxL) x 1 x d_model
+        core_out = core_out.view(residual.size(0), residual.size(1), self.d_model)
+        output = core_out + residual
+
+        if not self.pre_lnorm:
+            output = self.layer_norm(output)
+        return output, self.bias
 
 
 class BruteForceMoE(nn.Module):