Torchscript fixes/hacks for rms_norm, refactor ParallelScalingBlock with manual combination of input projections, closer paper match

timm/layers/fast_norm.py

@@ -90,7 +90,15 @@ def rms_norm(
     weight: Optional[torch.Tensor] = None,
     eps: float = 1e-5,
 ):
-    dims = tuple(i for i in range(-1, -len(normalized_shape) - 1, -1))
+    norm_ndim = len(normalized_shape)
+    if torch.jit.is_scripting():
+        # ndim = len(x.shape)
+        # dims = list(range(ndim - norm_ndim, ndim))  # this doesn't work on pytorch <= 1.13.x
+        # NOTE -ve dims cause torchscript to crash in some cases, out of options to work around
+        assert norm_ndim == 1
+        v = torch.var(x, dim=-1).unsqueeze(-1)  # ts crashes with -ve dim + keepdim=True
+    else:
+        dims = tuple(range(-1, -norm_ndim - 1, -1))
         v = torch.var(x, dim=dims, keepdim=True)
     x = x * torch.rsqrt(v + eps)
     if weight is not None:
@@ -104,10 +112,15 @@ def fast_rms_norm(
     weight: Optional[torch.Tensor] = None,
     eps: float = 1e-5,
 ) -> torch.Tensor:
-    if torch.jit.is_scripting() or not has_apex_rmsnorm:
+    if torch.jit.is_scripting():
+        # this must be by itself, cannot merge with has_apex_rmsnorm
         return rms_norm(x, normalized_shape, weight, eps)
 
+    if has_apex_rmsnorm:
         if weight is None:
             return fused_rms_norm(x, normalized_shape, eps)
         else:
             return fused_rms_norm_affine(x, weight, normalized_shape, eps)
+
+    # fallback
+    return rms_norm(x, normalized_shape, weight, eps)

timm/layers/norm.py

@@ -122,6 +122,7 @@ class LayerNormExp2d(nn.LayerNorm):
 class RmsNorm(nn.Module):
     """ RmsNorm w/ fast (apex) norm if available
     """
+    __constants__ = ['normalized_shape', 'eps', 'elementwise_affine']
     normalized_shape: Tuple[int, ...]
     eps: float
     elementwise_affine: bool

timm/models/vision_transformer.py

@@ -217,6 +217,8 @@ class ParallelScalingBlock(nn.Module):
     Based on:
       'Scaling Vision Transformers to 22 Billion Parameters` - https://arxiv.org/abs/2302.05442
     """
+    fast_attn: Final[bool]
+
     def __init__(
             self,
             dim,
@@ -232,33 +234,76 @@ class ParallelScalingBlock(nn.Module):
             norm_layer=nn.LayerNorm
     ):
         super().__init__()
-        self.norm1 = norm_layer(dim)
+        assert dim % num_heads == 0, 'dim should be divisible by num_heads'
-        self.attn = Attention(
+        self.num_heads = num_heads
-            dim,
+        self.head_dim = dim // num_heads
-            num_heads=num_heads,
+        self.scale = self.head_dim ** -0.5
-            qkv_bias=qkv_bias,
+        self.fast_attn = hasattr(torch.nn.functional, 'scaled_dot_product_attention')  # FIXME
-            qk_norm=qk_norm,
+        mlp_hidden_dim = int(mlp_ratio * dim)
-            attn_drop=attn_drop,
+        in_proj_out_dim = mlp_hidden_dim + 3 * dim
-            proj_drop=drop,
+        out_proj_in_dim = mlp_hidden_dim + dim
-            norm_layer=norm_layer,
-        )
-        self.ls1 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
-        self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
 
-        self.norm2 = norm_layer(dim)
+        self.in_norm = norm_layer(dim)
-        self.mlp = Mlp(
+        self.in_proj = nn.Linear(dim, in_proj_out_dim, bias=qkv_bias)
-            in_features=dim,
+        self.in_split = [mlp_hidden_dim] + [dim] * 3
-            hidden_features=int(dim * mlp_ratio),
+        if qkv_bias:
-            act_layer=act_layer,
+            self.register_buffer('qkv_bias', None)
-            drop=drop,
+            self.register_parameter('mlp_bias', None)
-        )
+        else:
-        self.ls2 = LayerScale(dim, init_values=init_values) if init_values else nn.Identity()
+            self.register_buffer('qkv_bias', torch.zeros(3 * dim), persistent=False)
-        self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+            self.mlp_bias = nn.Parameter(torch.zeros(mlp_hidden_dim))
+
+        self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.attn_out_proj = nn.Linear(dim, dim)
+
+        self.mlp_drop = nn.Dropout(drop)
+        self.mlp_act = act_layer()
+        self.mlp_out_proj = nn.Linear(mlp_hidden_dim, dim)
+
+        self.ls = LayerScale(dim, init_values=init_values) if init_values is not None else nn.Identity()
+        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
 
     def forward(self, x):
-        y1 = self.drop_path1(self.ls1(self.attn(self.norm1(x))))
+        B, N, C = x.shape
-        y2 = self.drop_path2(self.ls2(self.mlp(self.norm2(x))))
+
-        x = x + y1 + y2
+        # Combined MLP fc1 & qkv projections
+        y = self.in_norm(x)
+        if self.mlp_bias is not None:
+            # Concat constant zero-bias for qkv w/ trainable mlp_bias.
+            # Appears faster than adding to x_mlp separately
+            y = F.linear(y, self.in_proj.weight, torch.cat((self.qkv_bias, self.mlp_bias)))
+        else:
+            y = self.in_proj(y)
+        x_mlp, q, k, v = torch.split(y, self.in_split, dim=-1)
+
+        # Dot product attention w/ qk norm
+        q = self.q_norm(q.view(B, N, self.num_heads, self.head_dim)).transpose(1, 2)
+        k = self.k_norm(k.view(B, N, self.num_heads, self.head_dim)).transpose(1, 2)
+        v = v.view(B, N, self.num_heads, self.head_dim).transpose(1, 2)
+        if self.fast_attn:
+            x_attn = F.scaled_dot_product_attention(
+                q, k, v,
+                dropout_p=self.attn_drop.p,
+            )
+        else:
+            q = q * self.scale
+            attn = q @ k.transpose(-2, -1)
+            attn = attn.softmax(dim=-1)
+            attn = self.attn_drop(attn)
+            x_attn = attn @ v
+        x_attn = x_attn.transpose(1, 2).reshape(B, N, C)
+        x_attn = self.attn_out_proj(x_attn)
+
+        # MLP activation, dropout, fc2
+        x_mlp = self.mlp_act(x_mlp)
+        x_mlp = self.mlp_drop(x_mlp)
+        x_mlp = self.mlp_out_proj(x_mlp)
+
+        # Add residual w/ drop path & layer scale applied
+        y = self.drop_path(self.ls(x_attn + x_mlp))
+        x = x + y
         return x
 
 
@@ -1249,6 +1294,7 @@ default_cfgs = generate_default_cfgs({
         hf_hub_id='timm/',
         input_size=(3, 240, 240), crop_pct=0.95, num_classes=21843),
 
+    'vit_base_patch16_xp_224.untrained': _cfg(url=''),
     'vit_large_patch14_xp_224.untrained': _cfg(url=''),
     'vit_huge_patch14_xp_224.untrained': _cfg(url=''),
 })
@@ -1750,6 +1796,19 @@ def flexivit_large(pretrained=False, **kwargs):
     return model
 
 
+@register_model
+def vit_base_patch16_xp_224(pretrained=False, **kwargs):
+    """ ViT-Large model (ViT-L/14) w/ parallel blocks and qk norm enabled.
+    """
+    model_kwargs = dict(
+        patch_size=16, embed_dim=768, depth=12, num_heads=12, pre_norm=True, no_embed_class=True,
+        norm_layer=RmsNorm, block_fn=ParallelScalingBlock, qkv_bias=False, qk_norm=True,
+    )
+    model = _create_vision_transformer(
+        'vit_base_patch16_xp_224', pretrained=pretrained, **dict(model_kwargs, **kwargs))
+    return model
+
+
 @register_model
 def vit_large_patch14_xp_224(pretrained=False, **kwargs):
     """ ViT-Large model (ViT-L/14) w/ parallel blocks and qk norm enabled.