enable generic dimentionality for input

Author: Vitaliy Chiley

megablocks/layers/dmoe.py

@@ -132,10 +132,9 @@ def sparse_forward_once(self, x, expert_weights, top_experts):
         with torch.no_grad():
             indices, bin_ids, bins, padded_bins, tokens_per_expert = (
                 self.indices_and_padded_bins(top_experts))
-        sl, bs, hs = x.size()
 
         # Route the tokens for MoE computation.
-        x = x.view(sl * bs, hs)
+        x = x.view(-1, x.shape[-1])
         x = ops.padded_gather(
             x,
             indices,

megablocks/layers/moe.py

@@ -419,13 +419,13 @@ def parallel_forward_once(self, x, expert_weights, top_experts):
         return x, tokens_per_expert.flatten()
 
     def forward(self, x, scores, expert_weights, top_experts):
-        sl, bs, hs = x.size()
+        in_shape = x.size()
 
         # Compute the experts.
         x, tokens_per_expert = self.forward_fn(
             x, expert_weights, top_experts)
         save_load_balancing_loss((tokens_per_expert, scores))
-        x = x.view(sl, bs, hs)
+        x = x.view(in_shape)
         if self.bias is not None:
             if self.args.return_bias:
                 return x, self.bias
@@ -448,7 +448,6 @@ def forward(self, x):
         # NOTE: If we're going to cast the activations to lower precision
         # do it before we permute the tokens to save bandwidth.
         x = common.cast_if_autocast_enabled(x)
-        sl, bs, hs = x.size()
 
         # Compute the expert scores and assignments.
         scores, expert_weights, top_experts = self.router(x)

megablocks/layers/router.py

@@ -53,8 +53,7 @@ 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)
+        scores = self.layer(x.view(-1, x.shape[-1])).softmax(dim=-1)
         expert_weights, expert_indices = self._top_k(scores)
 
         expert_indices = (