Use transform to replace rms_norm

Author: Sheng Feng Wu

examples/models/llama2/TARGETS

@@ -71,6 +71,7 @@ runtime.python_library(
         "export_llama_lib.py",
         "model.py",
         "source_transformation/quantize.py",
+        "source_transformation/rms_norm.py",
         "source_transformation/rope.py",
         "source_transformation/sdpa.py",
     ],

examples/models/llama2/export_llama_lib.py

@@ -43,13 +43,13 @@
     get_qnn_quantizer,
 )
 from executorch.util.activation_memory_profiler import generate_memory_trace
-from torch._export import capture_pre_autograd_graph
 
 from ..model_factory import EagerModelFactory
 from .source_transformation.quantize import (
     get_quant_embedding_transform,
     get_quant_weight_transform,
 )
+from .source_transformation.rms_norm import replace_rms_norm_with_native_rms_norm
 from .source_transformation.rope import materialze_broadcast_of_rope_freq_cis
 from .source_transformation.sdpa import (
     replace_causal_mask,
@@ -409,6 +409,7 @@ def _prepare_for_llama_export(modelname: str, args) -> LLMEdgeManager:
             transforms.append(replace_kv_cache_with_simple_kv_cache)
             transforms.append(replace_sdpa_with_flex_sdpa)
             transforms.append(replace_causal_mask)
+            transforms.append(replace_rms_norm_with_native_rms_norm)
             transforms.append(convert_linear_to_conv2d)
 
         elif args.coreml or args.mps:

examples/models/llama2/llama_transformer.py

@@ -39,6 +39,7 @@ def __init__(self, dim: int, eps: float = 1e-6):
 
         """
         super().__init__()
+        self.dim = dim
         self.eps = eps
         self.weight = nn.Parameter(torch.ones(dim))
 
@@ -416,8 +417,8 @@ def __init__(self, layer_id: int, args: ModelArgs):
             self.block_sparse_moe = MOEFeedForward(args)
         else:
             self.feed_forward = FeedForward(args)
-        self.attention_norm = torch.nn.RMSNorm(args.dim, eps=args.norm_eps)
-        self.ffn_norm = torch.nn.RMSNorm(args.dim, eps=args.norm_eps)
+        self.attention_norm = RMSNorm(args.dim, eps=args.norm_eps)
+        self.ffn_norm = RMSNorm(args.dim, eps=args.norm_eps)
 
     def forward(self, x, freqs_cos, freqs_sin, input_pos=None):  # x: 1xN
         h = self.attention.forward(
@@ -443,7 +444,7 @@ def __init__(self, params: ModelArgs):
         self.layers = torch.nn.ModuleList()
         for layer_id in range(params.n_layers):
             self.layers.append(TransformerBlock(layer_id, params))
-        self.norm = torch.nn.RMSNorm(params.dim, eps=params.norm_eps)
+        self.norm = RMSNorm(params.dim, eps=params.norm_eps)
         self.output = nn.Linear(params.dim, params.vocab_size, bias=False)
         self.use_kv_cache = params.use_kv_cache
         self.generate_full_logits = params.generate_full_logits

examples/models/llama2/source_transformation/rms_norm.py

@@ -0,0 +1,23 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from executorch.examples.models.llama2.llama_transformer import RMSNorm
+
+
+def replace_rms_norm_with_native_rms_norm(module: torch.nn.Module):
+    for name, child in module.named_children():
+        if isinstance(child, RMSNorm):
+            rms_norm = torch.nn.RMSNorm(child.dim, eps=child.eps)
+            rms_norm.weight = child.weight
+            setattr(
+                module,
+                name,
+                rms_norm,
+            )
+        else:
+            replace_rms_norm_with_native_rms_norm(child)
+    return module

extension/llm/export/builder.py

@@ -71,7 +71,6 @@ def __init__(
         verbose: bool = False,
         metadata: Optional[dict] = None,
         dynamic_shapes: Optional[Any] = None,
-        export_fn=capture_pre_autograd_graph,
     ):
         self.model = model
         # graph module returned from capture_pre_autograd_graph