Add per-layer MLP type support for executorch export (#18856)

examples/models/llama/llama_transformer.py

@@ -98,26 +98,33 @@ def forward(self, x: torch.Tensor) -> torch.Tensor:
 
 
 class TransformerBlock(nn.Module):
-    def __init__(self, args: ModelArgs, attention: Attention):
+    def __init__(
+        self, args: ModelArgs, attention: Attention, mlp_type: str = "default"
+    ):
         """
         Transformer block with support for pre-norm and post-norm.
         Args:
             args (ModelArgs): model configuration parameters.
             attention (Attention): attention object to use in the transformer
                 block. See `attention.py` for types of attention. Make sure
                 the attention type is registered in the ATTENTION_REGISTRY.
+            mlp_type (str): MLP type for this layer. "default" for standard
+                FFN, "skip" for no FFN block.
         """
         super().__init__()
         self.use_kv_cache = args.use_kv_cache
         self.n_heads = args.n_heads
         self.dim = args.dim
         self.head_dim = args.head_dim
         self.attention = attention
+        self.mlp_type = mlp_type.lower()
 
         assert (
             args.hidden_dim is not None
         ), "`hidden_dim` must be set in ModelArgs to construct a TransformerBlock."
-        if args.moe:
+        if self.mlp_type == "skip":
+            pass  # No FFN block for this layer
+        elif args.moe:
             self.block_sparse_moe = MOEFeedForward(args)
         elif args.target_modules is not None and (
             "down_proj" in args.target_modules
@@ -136,11 +143,12 @@ def __init__(self, args: ModelArgs, attention: Attention):
                 eps=args.norm_eps,
                 add_unit_offset=args.rms_norm_add_unit_offset,
             )
-        self.ffn_norm = RMSNorm(
-            args.dim,
-            eps=args.norm_eps,
-            add_unit_offset=args.rms_norm_add_unit_offset,
-        )
+        if self.mlp_type != "skip":
+            self.ffn_norm = RMSNorm(
+                args.dim,
+                eps=args.norm_eps,
+                add_unit_offset=args.rms_norm_add_unit_offset,
+            )
 
     @classmethod
     def from_type(cls, layer_id, args, rope) -> "TransformerBlock":
@@ -156,9 +164,12 @@ def from_type(cls, layer_id, args, rope) -> "TransformerBlock":
                 f"Unknown attention type: {args.attention_type}. "
                 f"Available: {list(ATTENTION_REGISTRY.keys())}"
             )
+        mlp_type = "default"
+        if args.mlp_type is not None and layer_id < len(args.mlp_type):
+            mlp_type = args.mlp_type[layer_id]
         cls = ATTENTION_REGISTRY[args.attention_type]
         attention = cls(args, layer_id, rope, **args.attention_kwargs)
-        return TransformerBlock(args, attention)
+        return TransformerBlock(args, attention, mlp_type=mlp_type)
 
     def forward(self, x, freqs_cos, freqs_sin, attn_options: ForwardOptions):  # x: 1xN
         h, attn_options_update = self.attention(
@@ -167,7 +178,9 @@ def forward(self, x, freqs_cos, freqs_sin, attn_options: ForwardOptions):  # x:
         if not isinstance(self.attention, AttentionSkip):
             h = x + h
 
-        if hasattr(self, "block_sparse_moe"):
+        if self.mlp_type == "skip":
+            out = h
+        elif hasattr(self, "block_sparse_moe"):
             out = h + self.block_sparse_moe(self.ffn_norm(h))
         else:
             out = h + self.feed_forward(self.ffn_norm(h))

examples/models/llama/model_args.py

@@ -145,6 +145,9 @@ class ModelArgs:
     attention_kwargs: Dict[str, Any] = dataclasses.field(default_factory=dict)
     # Hybrid models can have layer types different from attention
     layer_types: Optional[list] = None
+    # Per-layer MLP type: "default" for standard FFN, "skip" for no FFN block.
+    # Indexed by layer id (e.g. mlp_type[0] applies to layer 0).
+    mlp_type: Optional[list] = None
     model_architecture: Optional[str] = (
         None  # Architecture of model. For HF models, please refer to the HF model.config.architectures. This is used in QNN backend only for now.
     )