add ability to use differentiable topk

setup.py

@@ -3,7 +3,7 @@
 setup(
   name = 'st-moe-pytorch',
   packages = find_packages(exclude=[]),
-  version = '0.0.21',
+  version = '0.0.22',
   license='MIT',
   description = 'ST - Mixture of Experts - Pytorch',
   author = 'Phil Wang',

st_moe_pytorch/st_moe_pytorch.py

@@ -11,7 +11,7 @@
 
 from einops import rearrange, repeat, reduce, pack, unpack
 
-from colt5_attention import topk as differentiable_topk
+from colt5_attention import topk as maybe_differentiable_topk
 
 # constants
 
@@ -156,13 +156,23 @@ def __init__(
         threshold_eval: Union[float, Tuple[float, ...]] = 0.2,
         capacity_factor_train = 1.25,
         capacity_factor_eval = 2.,
-        straight_through_dispatch_tensor = True
+        straight_through_dispatch_tensor = True,
+        differentiable_topk = False,
+        differentiable_topk_fused = True
     ):
         super().__init__()
         self.eps = eps
         self.num_gates = num_gates
         self.to_gates = nn.Linear(dim, num_gates, bias = False)
 
+        self.differentiable_topk = differentiable_topk
+
+        self.topk = partial(
+            maybe_differentiable_topk,
+            non_differentiable = not differentiable_topk,
+            fused = differentiable_topk_fused # use triton fused coordinate descent if possible by default
+        )
+
         assert top_n >= 2, 'must be 2 or more experts'
         self.top_n = top_n
         top_n_minus_1 = top_n - 1
@@ -214,7 +224,17 @@ def forward(self, x):
 
         # find top N experts per position
 
-        gates, gate_indices = raw_gates.topk(k = top_n, dim = -1)
+        topk_return = self.topk(raw_gates, k = top_n)
+
+        gate_indices = topk_return.indices
+
+        if self.differentiable_topk:
+            # allow for differentiable topk using coordinate descent
+            # used successfully for routing from CoLT5 paper https://github.com/lucidrains/CoLT5-attention
+
+            gates = topk_return.coor_descent_values
+        else:
+            gates = topk_return.values
 
         # move the top-n dimension to be first
 
@@ -342,25 +362,24 @@ def __init__(self,
         loss_coef = 1e-2,
         router_z_loss_coef = 1e-3,
         experts: Optional[Module] = None,
-        straight_through_dispatch_tensor = True
+        straight_through_dispatch_tensor = True,
+        differentiable_topk = False,
+        differentiable_topk_fused = True
     ):
         super().__init__()
         self.dim = dim
         self.num_experts = num_experts
 
-        gating_kwargs = dict(
-            threshold_train = threshold_train,
-            threshold_eval = threshold_eval,
-            capacity_factor_train = capacity_factor_train,
-            capacity_factor_eval = capacity_factor_eval
-        )
-
         self.gate = TopNGating(
             dim,
             top_n = gating_top_n,
             num_gates = num_experts,
             straight_through_dispatch_tensor = straight_through_dispatch_tensor,
-            **gating_kwargs
+            differentiable_topk = differentiable_topk,
+            threshold_train = threshold_train,
+            threshold_eval = threshold_eval,
+            capacity_factor_train = capacity_factor_train,
+            capacity_factor_eval = capacity_factor_eval
         )
 
         self.experts = default(experts, lambda: Experts(dim, num_experts = num_experts, hidden_mult = expert_hidden_mult))