using the same eps in layernorm as default torch (#221)

CHANGELOG.md

@@ -7,6 +7,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 ## [0.0.x] - TBD
 ### Fixed
 - Expose bias flag for feedforwards, same default as Timm [#220]
+- Update eps value for layernormm, same default as torch [#221]
 
 ## [0.0.9] - 2022-02-09
 ### Added

tests/test_triton_layernorm.py

@@ -53,8 +53,8 @@ def test_layernorm_parity(shape, amp):
     eps = 1e-5
 
     # Initialize the two layers, weights are 1 and 0 by default, no randomness
-    torch_layernorm = torch.nn.LayerNorm(X.shape[-1], eps).to("cuda")
-    triton_layernorm = FusedLayerNorm(X.shape[-1], eps).to("cuda")
+    torch_layernorm = torch.nn.LayerNorm(X.shape[-1], eps=eps).to("cuda")
+    triton_layernorm = FusedLayerNorm(X.shape[-1], affine=True, eps=eps).to("cuda")
 
     with autocast(enabled=amp):
         assert torch.allclose(X, X_)  # sanity checking, else all hell breaks loose

xformers/triton/layer_norm.py

@@ -32,7 +32,7 @@ class FusedLayerNorm(nn.Module):
 
     """
 
-    def __init__(self, normalized_shape, affine=True, eps=1e-05):
+    def __init__(self, normalized_shape, affine=True, eps=1e-06):
         super().__init__()
         if affine:
             self.weight = nn.Parameter(torch.ones(normalized_shape))
@@ -49,7 +49,7 @@ def layer_norm(
     x: torch.Tensor,
     weight: Optional[torch.Tensor] = None,
     bias: Optional[torch.Tensor] = None,
-    eps: float = 1e-05,
+    eps: float = 1e-06,
 ) -> torch.Tensor:
 
     global _triton_registered_warnings