Merge branch 'main' into btk-setup-deps

megablocks/layers/arguments.py

@@ -35,6 +35,9 @@ class Arguments:
     init_method : InitFn =  partial(torch.nn.init.normal_, mean=0.0, std=0.02)
     output_layer_init_method : InitFn = init_method
 
+    # Benchmarking arguments.
+    uniform_expert_assignment : bool = False
+
 
 def from_megatron(megatron_args):
     args = Arguments()

megablocks/layers/router.py

@@ -2,6 +2,22 @@
 from megablocks.layers.arguments import Arguments
 import torch
 
+
+# NOTE: To enable end-to-end benchmarking without convergence we
+# support a flag to force the router to assign tokens uniformly
+# across the experts. We do this with a custom autograd operation
+# so that PyTorch still executes the full set of router operation.
+class _UniformExpertAssignment(torch.autograd.Function):
+
+
+    @staticmethod
+    def forward(ctx, x, num_experts):
+        out = torch.arange(x.numel(), dtype=x.dtype, device=x.device)
+        out = torch.remainder(out, num_experts)
+        return out.view(x.shape)
+_uniform_expert_assignment = _UniformExpertAssignment.apply
+
+
 class LearnedRouter(torch.nn.Module):
 
     def __init__(self, args : Arguments):
@@ -27,12 +43,22 @@ def jitter(self, x):
         noise = torch.rand(x.size(), dtype=x.dtype, device=x.device)
         return low + noise * (high - low)
 
+    def _top_k(self, scores):
+        if self.args.moe_top_k == 1:
+            return scores.max(dim=-1)
+        return torch.topk(scores, self.args.moe_top_k, dim=-1)
+
+
     def forward(self, x):
         if self.training and self.args.moe_jitter_eps is not None:
             x = x * self.jitter(x)
 
         sl, bs, hs = x.size()
         scores = self.layer(x.view(-1, hs)).softmax(dim=-1)
-        if self.args.moe_top_k == 1:
-            return scores, *scores.max(dim=-1)
-        return scores, *torch.topk(scores, self.args.moe_top_k, dim=-1)
+        expert_weights, expert_indices = self._top_k(scores)
+
+        expert_indices = (
+            _uniform_expert_assignment(expert_indices, self.args.moe_num_experts)
+            if self.args.uniform_expert_assignment else expert_indices
+        )
+        return scores, expert_weights, expert_indices