Merge branch 'main' into pmannan/hcp_fix

Author: parthmannan

megatron/core/transformer/moe/moe_layer.py

@@ -451,10 +451,25 @@ def custom_forward(hidden_states, intermediate_tensors=None, padding_mask=None):
 
     def backward_dw(self, routed_experts: bool = True, shared_experts: bool = False):
         """Compute weight gradients for experts and shared experts."""
+        # TODO(Wohox): replace the "routed_experts" and "shared_experts" arguments with better
+        # naming to better explain that they are actually from different fine-grained callables,
+        # or use scanning to decide which backward_dw should be called.
         if routed_experts:
             self.experts.backward_dw()
-        if shared_experts and self.use_shared_expert and not self.shared_expert_overlap:
-            self.shared_experts.backward_dw()
+            if self.config.moe_latent_size:
+                # TODO(Wohox): fc2_latent_proj forward and backward are executed in comm stream,
+                # so we execute its backward_dw in the comm stream too. But this may harm the
+                # EP overlap performance. Better to check if there is a better way to handle this.
+                from megatron.core.pipeline_parallel.utils import get_comm_stream
+
+                comm_stream = get_comm_stream()
+                with torch.cuda.stream(comm_stream):
+                    self.fc2_latent_proj.backward_dw()
+        if shared_experts:
+            if self.use_shared_expert and not self.shared_expert_overlap:
+                self.shared_experts.backward_dw()
+            if self.config.moe_latent_size:
+                self.fc1_latent_proj.backward_dw()
 
     def set_for_recompute_pre_mlp_layernorm(self):
         """Set the MoE layer for recompute pre_mlp_layernorm. Only needed for fp8/fp4."""

megatron/training/training.py

@@ -364,6 +364,7 @@ def transformer_flops():
             if args.moe_ffn_hidden_size is not None
             else args.ffn_hidden_size
         )
+        moe_latent_size = args.moe_latent_size
         shared_expert_ffn_hidden_size = (
             0
             if args.moe_shared_expert_intermediate_size is None
@@ -545,7 +546,20 @@ def transformer_flops():
                     (args.ffn_hidden_size * ffn_expansion_factor)
                     * num_dense_layers
                     # routed experts
-                    + (moe_ffn_hidden_size * num_experts_routed_to * ffn_expansion_factor)
+                    + (
+                        (moe_ffn_hidden_size * num_experts_routed_to * ffn_expansion_factor)
+                        if moe_latent_size is None
+                        else (
+                            (
+                                moe_ffn_hidden_size
+                                * num_experts_routed_to
+                                * ffn_expansion_factor
+                                * moe_latent_size
+                                / args.hidden_size
+                            )  # Routed experts run on moe_latent_size.
+                            + 2 * moe_latent_size  # Up proj and down proj.
+                        )
+                    )
                     * num_moe_layers
                     # Shared Experts.
                     + (shared_expert_ffn_hidden_size * ffn_expansion_factor)