fix: refactor layer norm converter with INormalization Layer (#2755)

Author: zewenli98

py/torch_tensorrt/dynamo/conversion/impl/normalization/ops.py

@@ -9,6 +9,7 @@
 from torch_tensorrt.dynamo.conversion._ConversionContext import ConversionContext
 from torch_tensorrt.dynamo.conversion.converter_utils import (
     cast_trt_tensor,
+    get_axes_for_reduce_op,
     get_positive_dim,
     get_trt_tensor,
     to_numpy,
@@ -105,102 +106,30 @@ def layer_norm(
     cudnn_enable: bool,
     return_mean_rstd: bool,
 ) -> Union[TRTTensor, Tuple[TRTTensor, torch.Tensor, torch.Tensor]]:
-    if weight is None:
-        weight = to_numpy(1.0)
-
-    if bias is None:
-        bias = to_numpy(0.0)
-
-    shape = weight.shape
-    gamma = to_numpy(weight).reshape(shape)
-    beta = to_numpy(bias).reshape(shape)
-
-    dims = list(range(len(input.shape) - len(shape), len(input.shape)))
-
-    # E[x]
-    mean_expected_trt = impl.reduce.mean(
-        ctx, target, source_ir, f"{name}_mean_expected", input, dims, True
-    )
-
-    # X-E[x]
-    sub_trt = impl.elementwise.sub(
-        ctx,
-        target,
-        source_ir,
-        f"{name}_sub",
-        input,
-        mean_expected_trt,
-    )
-
-    # Variance = mean(pow(x_sub_mean, 2))
-    pow_trt = get_trt_tensor(ctx, 2, f"{name}_power", np.float32)
-    pow_var = impl.elementwise.pow(
-        ctx,
-        target,
-        source_ir,
-        f"{name}_pow_var",
-        sub_trt,
-        pow_trt,
-    )
-    mean_trt = impl.reduce.mean(
-        ctx, target, source_ir, f"{name}_mean", pow_var, dims, True
-    )
-
-    # sqrt((var + eps))
-    eps_trt = get_trt_tensor(ctx, eps, f"{name}_eps", np.float32)
-    add_trt = impl.elementwise.add(
-        ctx,
-        target,
-        source_ir,
-        f"{name}_add",
-        mean_trt,
-        eps_trt,
-    )
-    sqrt_trt = impl.unary.sqrt(
-        ctx,
-        target,
-        source_ir,
-        f"{name}_sqrt",
-        add_trt,
-    )
-
-    # (X - E[X]) / sqrt((var + eps))
-    div_trt = impl.elementwise.div(
-        ctx,
-        target,
-        source_ir,
-        f"{name}_div",
-        sub_trt,
-        sqrt_trt,
-    )
-
-    gamma_trt = get_trt_tensor(ctx, weight, f"{name}_gamma")
-    beta_trt = get_trt_tensor(ctx, bias, f"{name}_beta")
-
-    # y * gamma + beta
-    scaled_y = impl.elementwise.mul(
-        ctx,
-        target,
-        source_ir,
-        f"{name}_mul_gamma",
-        div_trt,
-        gamma_trt,
-    )
+    dims = list(range(len(input.shape) - len(normalized_shape), len(input.shape)))
+    axes = get_axes_for_reduce_op(dims)
+
+    weight = get_trt_tensor(ctx, weight, f"{name}_weight")
+    bias = get_trt_tensor(ctx, bias, f"{name}_bias")
+    if tuple(input.shape) != tuple(weight.shape):
+        weight = impl.slice.expand(
+            ctx, target, source_ir, f"{name}_expand_weight", weight, input.shape
+        )
+    if tuple(input.shape) != tuple(bias.shape):
+        bias = impl.slice.expand(
+            ctx, target, source_ir, f"{name}_expand_bias", bias, input.shape
+        )
 
-    output = impl.elementwise.add(
-        ctx,
-        target,
-        source_ir,
-        f"{name}_add_beta",
-        scaled_y,
-        beta_trt,
-    )
+    layer_norm = ctx.net.add_normalization(input, weight, bias, axes)
+    layer_norm.epsilon = eps
+    layer_norm.compute_precision = input.dtype
+    set_layer_name(layer_norm, target, f"{name}_layer_norm", source_ir)
 
     if return_mean_rstd:
         # return fake mean and rstd for now
-        return output, None, None
+        return layer_norm.get_output(0), None, None
 
-    return output
+    return layer_norm.get_output(0)
 
 
 def native_group_norm(

tests/py/dynamo/conversion/test_layer_norm_aten.py

@@ -1,32 +1,75 @@
 import torch
+from parameterized import parameterized
 from torch.testing._internal.common_utils import run_tests
 from torch_tensorrt import Input
 
 from .harness import DispatchTestCase
 
 
 class TestLayerNormConverter(DispatchTestCase):
-    def test_layer_norm(self):
+    @parameterized.expand(
+        [
+            (
+                (5, 3, 2, 4),
+                [
+                    4,
+                ],
+            ),
+            ((5, 3, 2, 4), [2, 4]),
+            ((5, 3, 2, 4), [3, 2, 4]),
+            ((5, 3, 2, 4), [5, 3, 2, 4]),
+        ]
+    )
+    def test_layer_norm(self, input_shape, normalized_shape, eps=1e-05):
         class LayerNorm(torch.nn.Module):
             def forward(self, x):
                 return torch.ops.aten.layer_norm.default(
                     x,
-                    torch.tensor([3, 224, 224]),
-                    torch.ones((3, 224, 224)),
-                    torch.zeros((3, 224, 224)),
-                    1e-05,
-                    True,
+                    normalized_shape,
+                    torch.randn(normalized_shape),
+                    torch.randn(normalized_shape),
+                    eps,
                 )
 
-        inputs = [torch.randn(1, 3, 224, 224)]
+        inputs = [torch.randn(input_shape)]
         self.run_test(
             LayerNorm(),
             inputs,
         )
 
 
 class TestNativeLayerNormConverter(DispatchTestCase):
-    def test_layer_norm(self):
+    @parameterized.expand(
+        [
+            (
+                (5, 3, 2, 4),
+                [
+                    4,
+                ],
+            ),
+            ((5, 3, 2, 4), [2, 4]),
+            ((5, 3, 2, 4), [3, 2, 4]),
+            ((5, 3, 2, 4), [5, 3, 2, 4]),
+        ]
+    )
+    def test_layer_norm(self, input_shape, normalized_shape, eps=1e-05):
+        class LayerNorm(torch.nn.Module):
+            def forward(self, x):
+                return torch.ops.aten.native_layer_norm.default(
+                    x,
+                    normalized_shape,
+                    torch.randn(normalized_shape),
+                    torch.randn(normalized_shape),
+                    eps,
+                )[0]
+
+        inputs = [torch.randn(input_shape)]
+        self.run_test(
+            LayerNorm(),
+            inputs,
+        )
+
+    def test_layernorm_with_dynamic_shape(self):
         class LayerNorm(torch.nn.Module):
             def forward(self, x):
                 return torch.ops.aten.native_layer_norm.default(
@@ -37,10 +80,17 @@ def forward(self, x):
                     1e-05,
                 )[0]
 
-        inputs = [torch.randn(1, 3, 224, 224)]
-        self.run_test(
+        input_specs = [
+            Input(
+                shape=(-1, 3, 224, 224),
+                dtype=torch.float32,
+                shape_ranges=[((1, 3, 224, 224), (5, 3, 224, 224), (10, 3, 224, 224))],
+            ),
+        ]
+
+        self.run_test_with_dynamic_shape(
             LayerNorm(),
-            inputs,
+            input_specs,
         )