diff --git a/README.md b/README.md
index df5fb968..f90c6abd 100644
--- a/README.md
+++ b/README.md
@@ -23,6 +23,17 @@ I'm fortunate to be able to dedicate significant time and money of my own suppor
 
 ## What's New
 
+### May 13, 2022
+* Official Swin-V2 models and weights added from (https://github.com/microsoft/Swin-Transformer). Cleaned up to support torchscript.
+* Some refactoring for existing `timm` Swin-V2-CR impl, will likely do a bit more to bring parts closer to official and decide whether to merge some aspects.
+* More Vision Transformer relative position / residual post-norm experiments w/ 512 dim
+  * `vit_relpos_small_patch16_224` - 81.5 @ 224, 82.5 @ 320 -- rel pos, layer scale, no class token, avg pool
+  * `vit_relpos_medium_patch16_rpn_224` - 82.3 @ 224, 83.1 @ 320 -- rel pos + res-post-norm, no class token, avg pool
+  * `vit_relpos_medium_patch16_224` - 82.5 @ 224, 83.3 @ 320 -- rel pos, layer scale, no class token, avg pool
+  * `vit_relpos_base_patch16_gapcls_224` - 82.8 @ 224, 83.9 @ 320 -- rel pos, layer scale, class token, avg pool (by mistake)
+* Bring 512 dim, 8-head 'medium' ViT model variant back to life (after using in a pre DeiT 'small' model for first ViT impl back in 2020)
+* Add ViT relative position support for switching btw existing impl and some additions in official Swin-V2 impl for future trials
+* Sequencer2D impl (https://arxiv.org/abs/2205.01972), added via PR from author (https://github.com/okojoalg)
 
 ### May 2, 2022
 * Vision Transformer experiments adding Relative Position (Swin-V2 log-coord) (`vision_transformer_relpos.py`) and Residual Post-Norm branches (from Swin-V2) (`vision_transformer*.py`)
@@ -390,6 +401,7 @@ A full version of the list below with source links can be found in the [document
 * ReXNet - https://arxiv.org/abs/2007.00992
 * SelecSLS - https://arxiv.org/abs/1907.00837
 * Selective Kernel Networks - https://arxiv.org/abs/1903.06586
+* Sequencer2D - https://arxiv.org/abs/2205.01972
 * Swin S3 (AutoFormerV2) - https://arxiv.org/abs/2111.14725
 * Swin Transformer - https://arxiv.org/abs/2103.14030
 * Swin Transformer V2 - https://arxiv.org/abs/2111.09883
diff --git a/tests/test_models.py b/tests/test_models.py
index 6489892c..7ea9af6e 100644
--- a/tests/test_models.py
+++ b/tests/test_models.py
@@ -25,7 +25,7 @@ if hasattr(torch._C, '_jit_set_profiling_executor'):
 NON_STD_FILTERS = [
     'vit_*', 'tnt_*', 'pit_*', 'swin_*', 'coat_*', 'cait_*', '*mixer_*', 'gmlp_*', 'resmlp_*', 'twins_*',
     'convit_*', 'levit*', 'visformer*', 'deit*', 'jx_nest_*', 'nest_*', 'xcit_*', 'crossvit_*', 'beit_*',
-    'poolformer_*', 'volo_*', 'sequencer2d_*']
+    'poolformer_*', 'volo_*', 'sequencer2d_*', 'swinv2_*']
 NUM_NON_STD = len(NON_STD_FILTERS)
 
 # exclude models that cause specific test failures
diff --git a/timm/models/__init__.py b/timm/models/__init__.py
index 32212fca..8cb6c70a 100644
--- a/timm/models/__init__.py
+++ b/timm/models/__init__.py
@@ -42,6 +42,7 @@ from .senet import *
 from .sequencer import *
 from .sknet import *
 from .swin_transformer import *
+from .swin_transformer_v2 import *
 from .swin_transformer_v2_cr import *
 from .tnt import *
 from .tresnet import *
diff --git a/timm/models/helpers.py b/timm/models/helpers.py
index c4f48d6a..1276b68e 100644
--- a/timm/models/helpers.py
+++ b/timm/models/helpers.py
@@ -477,7 +477,7 @@ def build_model_with_cfg(
         pretrained_cfg: Optional[Dict] = None,
         model_cfg: Optional[Any] = None,
         feature_cfg: Optional[Dict] = None,
-        pretrained_strict: bool = False,
+        pretrained_strict: bool = True,
         pretrained_filter_fn: Optional[Callable] = None,
         pretrained_custom_load: bool = False,
         kwargs_filter: Optional[Tuple[str]] = None,
diff --git a/timm/models/layers/mlp.py b/timm/models/layers/mlp.py
index a85e28d0..91e80a84 100644
--- a/timm/models/layers/mlp.py
+++ b/timm/models/layers/mlp.py
@@ -10,16 +10,17 @@ from .helpers import to_2tuple
 class Mlp(nn.Module):
     """ MLP as used in Vision Transformer, MLP-Mixer and related networks
     """
-    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, bias=True, drop=0.):
         super().__init__()
         out_features = out_features or in_features
         hidden_features = hidden_features or in_features
+        bias = to_2tuple(bias)
         drop_probs = to_2tuple(drop)
 
-        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias[0])
         self.act = act_layer()
         self.drop1 = nn.Dropout(drop_probs[0])
-        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias[1])
         self.drop2 = nn.Dropout(drop_probs[1])
 
     def forward(self, x):
@@ -35,17 +36,18 @@ class GluMlp(nn.Module):
     """ MLP w/ GLU style gating
     See: https://arxiv.org/abs/1612.08083, https://arxiv.org/abs/2002.05202
     """
-    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.Sigmoid, drop=0.):
+    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.Sigmoid, bias=True, drop=0.):
         super().__init__()
         out_features = out_features or in_features
         hidden_features = hidden_features or in_features
         assert hidden_features % 2 == 0
+        bias = to_2tuple(bias)
         drop_probs = to_2tuple(drop)
 
-        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias[0])
         self.act = act_layer()
         self.drop1 = nn.Dropout(drop_probs[0])
-        self.fc2 = nn.Linear(hidden_features // 2, out_features)
+        self.fc2 = nn.Linear(hidden_features // 2, out_features, bias=bias[1])
         self.drop2 = nn.Dropout(drop_probs[1])
 
     def init_weights(self):
@@ -67,14 +69,16 @@ class GluMlp(nn.Module):
 class GatedMlp(nn.Module):
     """ MLP as used in gMLP
     """
-    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU,
-                 gate_layer=None, drop=0.):
+    def __init__(
+            self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU,
+            gate_layer=None, bias=True, drop=0.):
         super().__init__()
         out_features = out_features or in_features
         hidden_features = hidden_features or in_features
+        bias = to_2tuple(bias)
         drop_probs = to_2tuple(drop)
 
-        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias[0])
         self.act = act_layer()
         self.drop1 = nn.Dropout(drop_probs[0])
         if gate_layer is not None:
@@ -83,7 +87,7 @@ class GatedMlp(nn.Module):
             hidden_features = hidden_features // 2  # FIXME base reduction on gate property?
         else:
             self.gate = nn.Identity()
-        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.fc2 = nn.Linear(hidden_features, out_features, bias=bias[1])
         self.drop2 = nn.Dropout(drop_probs[1])
 
     def forward(self, x):
@@ -100,15 +104,18 @@ class ConvMlp(nn.Module):
     """ MLP using 1x1 convs that keeps spatial dims
     """
     def __init__(
-            self, in_features, hidden_features=None, out_features=None, act_layer=nn.ReLU, norm_layer=None, drop=0.):
+            self, in_features, hidden_features=None, out_features=None, act_layer=nn.ReLU,
+            norm_layer=None, bias=True, drop=0.):
         super().__init__()
         out_features = out_features or in_features
         hidden_features = hidden_features or in_features
-        self.fc1 = nn.Conv2d(in_features, hidden_features, kernel_size=1, bias=True)
+        bias = to_2tuple(bias)
+
+        self.fc1 = nn.Conv2d(in_features, hidden_features, kernel_size=1, bias=bias[0])
         self.norm = norm_layer(hidden_features) if norm_layer else nn.Identity()
         self.act = act_layer()
-        self.fc2 = nn.Conv2d(hidden_features, out_features, kernel_size=1, bias=True)
         self.drop = nn.Dropout(drop)
+        self.fc2 = nn.Conv2d(hidden_features, out_features, kernel_size=1, bias=bias[1])
 
     def forward(self, x):
         x = self.fc1(x)
diff --git a/timm/models/swin_transformer_v2.py b/timm/models/swin_transformer_v2.py
new file mode 100644
index 00000000..0c9db3dd
--- /dev/null
+++ b/timm/models/swin_transformer_v2.py
@@ -0,0 +1,753 @@
+""" Swin Transformer V2
+A PyTorch impl of : `Swin Transformer V2: Scaling Up Capacity and Resolution`
+    - https://arxiv.org/abs/2111.09883
+
+Code/weights from https://github.com/microsoft/Swin-Transformer, original copyright/license info below
+
+Modifications and additions for timm hacked together by / Copyright 2022, Ross Wightman
+"""
+# --------------------------------------------------------
+# Swin Transformer V2
+# Copyright (c) 2022 Microsoft
+# Licensed under The MIT License [see LICENSE for details]
+# Written by Ze Liu
+# --------------------------------------------------------
+import math
+from typing import Tuple, Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint as checkpoint
+
+from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
+from .fx_features import register_notrace_function
+from .helpers import build_model_with_cfg, named_apply
+from .layers import PatchEmbed, Mlp, DropPath, to_2tuple, to_ntuple, trunc_normal_, _assert
+from .registry import register_model
+
+
+def _cfg(url='', **kwargs):
+    return {
+        'url': url,
+        'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,
+        'crop_pct': .9, 'interpolation': 'bicubic', 'fixed_input_size': True,
+        'mean': IMAGENET_DEFAULT_MEAN, 'std': IMAGENET_DEFAULT_STD,
+        'first_conv': 'patch_embed.proj', 'classifier': 'head',
+        **kwargs
+    }
+
+
+default_cfgs = {
+    'swinv2_tiny_window8_256': _cfg(
+        url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_tiny_patch4_window8_256.pth',
+        input_size=(3, 256, 256)
+    ),
+    'swinv2_tiny_window16_256': _cfg(
+        url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_tiny_patch4_window16_256.pth',
+        input_size=(3, 256, 256)
+    ),
+    'swinv2_small_window8_256': _cfg(
+        url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_small_patch4_window8_256.pth',
+        input_size=(3, 256, 256)
+    ),
+    'swinv2_small_window16_256': _cfg(
+        url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_small_patch4_window16_256.pth',
+        input_size=(3, 256, 256)
+    ),
+    'swinv2_base_window8_256': _cfg(
+        url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_base_patch4_window8_256.pth',
+        input_size=(3, 256, 256)
+    ),
+    'swinv2_base_window16_256': _cfg(
+        url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_base_patch4_window16_256.pth',
+        input_size=(3, 256, 256)
+    ),
+
+    'swinv2_base_window12_192_22k': _cfg(
+        url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_base_patch4_window12_192_22k.pth',
+        num_classes=21841, input_size=(3, 192, 192)
+    ),
+    'swinv2_base_window12to16_192to256_22kft1k': _cfg(
+        url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_base_patch4_window12to16_192to256_22kto1k_ft.pth',
+        input_size=(3, 256, 256)
+    ),
+    'swinv2_base_window12to24_192to384_22kft1k': _cfg(
+        url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_base_patch4_window12to24_192to384_22kto1k_ft.pth',
+        input_size=(3, 384, 384), crop_pct=1.0,
+    ),
+    'swinv2_large_window12_192_22k': _cfg(
+        url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_large_patch4_window12_192_22k.pth',
+        num_classes=21841, input_size=(3, 192, 192)
+    ),
+    'swinv2_large_window12to16_192to256_22kft1k': _cfg(
+        url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_large_patch4_window12to16_192to256_22kto1k_ft.pth',
+        input_size=(3, 256, 256)
+    ),
+    'swinv2_large_window12to24_192to384_22kft1k': _cfg(
+        url='https://github.com/SwinTransformer/storage/releases/download/v2.0.0/swinv2_large_patch4_window12to24_192to384_22kto1k_ft.pth',
+        input_size=(3, 384, 384), crop_pct=1.0,
+    ),
+}
+
+
+def window_partition(x, window_size: Tuple[int, int]):
+    """
+    Args:
+        x: (B, H, W, C)
+        window_size (int): window size
+
+    Returns:
+        windows: (num_windows*B, window_size, window_size, C)
+    """
+    B, H, W, C = x.shape
+    x = x.view(B, H // window_size[0], window_size[0], W // window_size[1], window_size[1], C)
+    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size[0], window_size[1], C)
+    return windows
+
+
+@register_notrace_function  # reason: int argument is a Proxy
+def window_reverse(windows, window_size: Tuple[int, int], img_size: Tuple[int, int]):
+    """
+    Args:
+        windows: (num_windows * B, window_size[0], window_size[1], C)
+        window_size (Tuple[int, int]): Window size
+        img_size (Tuple[int, int]): Image size
+
+    Returns:
+        x: (B, H, W, C)
+    """
+    H, W = img_size
+    B = int(windows.shape[0] / (H * W / window_size[0] / window_size[1]))
+    x = windows.view(B, H // window_size[0], W // window_size[1], window_size[0], window_size[1], -1)
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
+    return x
+
+
+class WindowAttention(nn.Module):
+    r""" Window based multi-head self attention (W-MSA) module with relative position bias.
+    It supports both of shifted and non-shifted window.
+
+    Args:
+        dim (int): Number of input channels.
+        window_size (tuple[int]): The height and width of the window.
+        num_heads (int): Number of attention heads.
+        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
+        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
+        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
+        pretrained_window_size (tuple[int]): The height and width of the window in pre-training.
+    """
+
+    def __init__(
+            self, dim, window_size, num_heads, qkv_bias=True, attn_drop=0., proj_drop=0.,
+            pretrained_window_size=[0, 0]):
+
+        super().__init__()
+        self.dim = dim
+        self.window_size = window_size  # Wh, Ww
+        self.pretrained_window_size = pretrained_window_size
+        self.num_heads = num_heads
+
+        self.logit_scale = nn.Parameter(torch.log(10 * torch.ones((num_heads, 1, 1))))
+
+        # mlp to generate continuous relative position bias
+        self.cpb_mlp = nn.Sequential(
+            nn.Linear(2, 512, bias=True),
+            nn.ReLU(inplace=True),
+            nn.Linear(512, num_heads, bias=False)
+        )
+
+        # get relative_coords_table
+        relative_coords_h = torch.arange(-(self.window_size[0] - 1), self.window_size[0], dtype=torch.float32)
+        relative_coords_w = torch.arange(-(self.window_size[1] - 1), self.window_size[1], dtype=torch.float32)
+        relative_coords_table = torch.stack(torch.meshgrid([
+            relative_coords_h,
+            relative_coords_w])).permute(1, 2, 0).contiguous().unsqueeze(0)  # 1, 2*Wh-1, 2*Ww-1, 2
+        if pretrained_window_size[0] > 0:
+            relative_coords_table[:, :, :, 0] /= (pretrained_window_size[0] - 1)
+            relative_coords_table[:, :, :, 1] /= (pretrained_window_size[1] - 1)
+        else:
+            relative_coords_table[:, :, :, 0] /= (self.window_size[0] - 1)
+            relative_coords_table[:, :, :, 1] /= (self.window_size[1] - 1)
+        relative_coords_table *= 8  # normalize to -8, 8
+        relative_coords_table = torch.sign(relative_coords_table) * torch.log2(
+            torch.abs(relative_coords_table) + 1.0) / math.log2(8)
+
+        self.register_buffer("relative_coords_table", relative_coords_table, persistent=False)
+
+        # get pair-wise relative position index for each token inside the window
+        coords_h = torch.arange(self.window_size[0])
+        coords_w = torch.arange(self.window_size[1])
+        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
+        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
+        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
+        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
+        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
+        relative_coords[:, :, 1] += self.window_size[1] - 1
+        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
+        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
+        self.register_buffer("relative_position_index", relative_position_index, persistent=False)
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=False)
+        if qkv_bias:
+            self.q_bias = nn.Parameter(torch.zeros(dim))
+            self.register_buffer('k_bias', torch.zeros(dim), persistent=False)
+            self.v_bias = nn.Parameter(torch.zeros(dim))
+        else:
+            self.q_bias = None
+            self.k_bias = None
+            self.v_bias = None
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+        self.softmax = nn.Softmax(dim=-1)
+
+    def forward(self, x, mask: Optional[torch.Tensor] = None):
+        """
+        Args:
+            x: input features with shape of (num_windows*B, N, C)
+            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
+        """
+        B_, N, C = x.shape
+        qkv_bias = None
+        if self.q_bias is not None:
+            qkv_bias = torch.cat((self.q_bias, self.k_bias, self.v_bias))
+        qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)
+        qkv = qkv.reshape(B_, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)
+        q, k, v = qkv.unbind(0)
+
+        # cosine attention
+        attn = (F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(-2, -1))
+        logit_scale = torch.clamp(self.logit_scale, max=math.log(1. / 0.01)).exp()
+        attn = attn * logit_scale
+
+        relative_position_bias_table = self.cpb_mlp(self.relative_coords_table).view(-1, self.num_heads)
+        relative_position_bias = relative_position_bias_table[self.relative_position_index.view(-1)].view(
+            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
+        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
+        relative_position_bias = 16 * torch.sigmoid(relative_position_bias)
+        attn = attn + relative_position_bias.unsqueeze(0)
+
+        if mask is not None:
+            nW = mask.shape[0]
+            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
+            attn = attn.view(-1, self.num_heads, N, N)
+            attn = self.softmax(attn)
+        else:
+            attn = self.softmax(attn)
+
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class SwinTransformerBlock(nn.Module):
+    r""" Swin Transformer Block.
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        num_heads (int): Number of attention heads.
+        window_size (int): Window size.
+        shift_size (int): Shift size for SW-MSA.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float, optional): Stochastic depth rate. Default: 0.0
+        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+        pretrained_window_size (int): Window size in pretraining.
+    """
+
+    def __init__(
+            self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
+            mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0., drop_path=0.,
+            act_layer=nn.GELU, norm_layer=nn.LayerNorm, pretrained_window_size=0):
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = to_2tuple(input_resolution)
+        self.num_heads = num_heads
+        ws, ss = self._calc_window_shift(window_size, shift_size)
+        self.window_size: Tuple[int, int] = ws
+        self.shift_size: Tuple[int, int] = ss
+        self.window_area = self.window_size[0] * self.window_size[1]
+        self.mlp_ratio = mlp_ratio
+
+        self.attn = WindowAttention(
+            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
+            qkv_bias=qkv_bias, attn_drop=attn_drop, proj_drop=drop,
+            pretrained_window_size=to_2tuple(pretrained_window_size))
+        self.norm1 = norm_layer(dim)
+        self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+
+        self.mlp = Mlp(in_features=dim, hidden_features=int(dim * mlp_ratio), act_layer=act_layer, drop=drop)
+        self.norm2 = norm_layer(dim)
+        self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
+
+        if any(self.shift_size):
+            # calculate attention mask for SW-MSA
+            H, W = self.input_resolution
+            img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
+            cnt = 0
+            for h in (
+                    slice(0, -self.window_size[0]),
+                    slice(-self.window_size[0], -self.shift_size[0]),
+                    slice(-self.shift_size[0], None)):
+                for w in (
+                        slice(0, -self.window_size[1]),
+                        slice(-self.window_size[1], -self.shift_size[1]),
+                        slice(-self.shift_size[1], None)):
+                    img_mask[:, h, w, :] = cnt
+                    cnt += 1
+            mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
+            mask_windows = mask_windows.view(-1, self.window_area)
+            attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
+            attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
+        else:
+            attn_mask = None
+
+        self.register_buffer("attn_mask", attn_mask)
+
+    def _calc_window_shift(self, target_window_size, target_shift_size) -> Tuple[Tuple[int, int], Tuple[int, int]]:
+        target_window_size = to_2tuple(target_window_size)
+        target_shift_size = to_2tuple(target_shift_size)
+        window_size = [r if r <= w else w for r, w in zip(self.input_resolution, target_window_size)]
+        shift_size = [0 if r <= w else s for r, w, s in zip(self.input_resolution, window_size, target_shift_size)]
+        return tuple(window_size), tuple(shift_size)
+
+    def _attn(self, x):
+        H, W = self.input_resolution
+        B, L, C = x.shape
+        _assert(L == H * W, "input feature has wrong size")
+        x = x.view(B, H, W, C)
+
+        # cyclic shift
+        has_shift = any(self.shift_size)
+        if has_shift:
+            shifted_x = torch.roll(x, shifts=(-self.shift_size[0], -self.shift_size[1]), dims=(1, 2))
+        else:
+            shifted_x = x
+
+        # partition windows
+        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
+        x_windows = x_windows.view(-1, self.window_area, C)  # nW*B, window_size*window_size, C
+
+        # W-MSA/SW-MSA
+        attn_windows = self.attn(x_windows, mask=self.attn_mask)  # nW*B, window_size*window_size, C
+
+        # merge windows
+        attn_windows = attn_windows.view(-1, self.window_size[0], self.window_size[1], C)
+        shifted_x = window_reverse(attn_windows, self.window_size, self.input_resolution)  # B H' W' C
+
+        # reverse cyclic shift
+        if has_shift:
+            x = torch.roll(shifted_x, shifts=self.shift_size, dims=(1, 2))
+        else:
+            x = shifted_x
+        x = x.view(B, H * W, C)
+        return x
+
+    def forward(self, x):
+        x = x + self.drop_path1(self.norm1(self._attn(x)))
+        x = x + self.drop_path2(self.norm2(self.mlp(x)))
+        return x
+
+
+class PatchMerging(nn.Module):
+    r""" Patch Merging Layer.
+
+    Args:
+        input_resolution (tuple[int]): Resolution of input feature.
+        dim (int): Number of input channels.
+        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
+    """
+
+    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
+        super().__init__()
+        self.input_resolution = input_resolution
+        self.dim = dim
+        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
+        self.norm = norm_layer(2 * dim)
+
+    def forward(self, x):
+        """
+        x: B, H*W, C
+        """
+        H, W = self.input_resolution
+        B, L, C = x.shape
+        _assert(L == H * W, "input feature has wrong size")
+        _assert(H % 2 == 0, f"x size ({H}*{W}) are not even.")
+        _assert(W % 2 == 0, f"x size ({H}*{W}) are not even.")
+
+        x = x.view(B, H, W, C)
+
+        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
+        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
+        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
+        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
+        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
+        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C
+
+        x = self.reduction(x)
+        x = self.norm(x)
+
+        return x
+
+
+class BasicLayer(nn.Module):
+    """ A basic Swin Transformer layer for one stage.
+
+    Args:
+        dim (int): Number of input channels.
+        input_resolution (tuple[int]): Input resolution.
+        depth (int): Number of blocks.
+        num_heads (int): Number of attention heads.
+        window_size (int): Local window size.
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
+        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
+        drop (float, optional): Dropout rate. Default: 0.0
+        attn_drop (float, optional): Attention dropout rate. Default: 0.0
+        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
+        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
+        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
+        pretrained_window_size (int): Local window size in pre-training.
+    """
+
+    def __init__(
+            self, dim, input_resolution, depth, num_heads, window_size,
+            mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0., drop_path=0.,
+            norm_layer=nn.LayerNorm, downsample=None, pretrained_window_size=0):
+
+        super().__init__()
+        self.dim = dim
+        self.input_resolution = input_resolution
+        self.depth = depth
+        self.grad_checkpointing = False
+
+        # build blocks
+        self.blocks = nn.ModuleList([
+            SwinTransformerBlock(
+                dim=dim, input_resolution=input_resolution,
+                num_heads=num_heads, window_size=window_size,
+                shift_size=0 if (i % 2 == 0) else window_size // 2,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                drop=drop, attn_drop=attn_drop,
+                drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
+                norm_layer=norm_layer,
+                pretrained_window_size=pretrained_window_size)
+            for i in range(depth)])
+
+        # patch merging layer
+        if downsample is not None:
+            self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
+        else:
+            self.downsample = nn.Identity()
+
+    def forward(self, x):
+        for blk in self.blocks:
+            if self.grad_checkpointing and not torch.jit.is_scripting():
+                x = checkpoint.checkpoint(blk, x)
+            else:
+                x = blk(x)
+        x = self.downsample(x)
+        return x
+
+    def _init_respostnorm(self):
+        for blk in self.blocks:
+            nn.init.constant_(blk.norm1.bias, 0)
+            nn.init.constant_(blk.norm1.weight, 0)
+            nn.init.constant_(blk.norm2.bias, 0)
+            nn.init.constant_(blk.norm2.weight, 0)
+
+
+class SwinTransformerV2(nn.Module):
+    r""" Swin Transformer V2
+        A PyTorch impl of : `Swin Transformer V2: Scaling Up Capacity and Resolution`
+            - https://arxiv.org/abs/2111.09883
+    Args:
+        img_size (int | tuple(int)): Input image size. Default 224
+        patch_size (int | tuple(int)): Patch size. Default: 4
+        in_chans (int): Number of input image channels. Default: 3
+        num_classes (int): Number of classes for classification head. Default: 1000
+        embed_dim (int): Patch embedding dimension. Default: 96
+        depths (tuple(int)): Depth of each Swin Transformer layer.
+        num_heads (tuple(int)): Number of attention heads in different layers.
+        window_size (int): Window size. Default: 7
+        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4
+        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
+        drop_rate (float): Dropout rate. Default: 0
+        attn_drop_rate (float): Attention dropout rate. Default: 0
+        drop_path_rate (float): Stochastic depth rate. Default: 0.1
+        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
+        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
+        patch_norm (bool): If True, add normalization after patch embedding. Default: True
+        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
+        pretrained_window_sizes (tuple(int)): Pretrained window sizes of each layer.
+    """
+
+    def __init__(
+            self, img_size=224, patch_size=4, in_chans=3, num_classes=1000, global_pool='avg',
+            embed_dim=96, depths=(2, 2, 6, 2), num_heads=(3, 6, 12, 24),
+            window_size=7, mlp_ratio=4., qkv_bias=True,
+            drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
+            norm_layer=nn.LayerNorm, ape=False, patch_norm=True,
+            pretrained_window_sizes=(0, 0, 0, 0), **kwargs):
+        super().__init__()
+
+        self.num_classes = num_classes
+        assert global_pool in ('', 'avg')
+        self.global_pool = global_pool
+        self.num_layers = len(depths)
+        self.embed_dim = embed_dim
+        self.patch_norm = patch_norm
+        self.num_features = int(embed_dim * 2 ** (self.num_layers - 1))
+
+        # split image into non-overlapping patches
+        self.patch_embed = PatchEmbed(
+            img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim,
+            norm_layer=norm_layer if self.patch_norm else None)
+        num_patches = self.patch_embed.num_patches
+
+        # absolute position embedding
+        if ape:
+            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
+            trunc_normal_(self.absolute_pos_embed, std=.02)
+        else:
+            self.absolute_pos_embed = None
+
+        self.pos_drop = nn.Dropout(p=drop_rate)
+
+        # stochastic depth
+        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule
+
+        # build layers
+        self.layers = nn.ModuleList()
+        for i_layer in range(self.num_layers):
+            layer = BasicLayer(
+                dim=int(embed_dim * 2 ** i_layer),
+                input_resolution=(
+                    self.patch_embed.grid_size[0] // (2 ** i_layer),
+                    self.patch_embed.grid_size[1] // (2 ** i_layer)),
+                depth=depths[i_layer],
+                num_heads=num_heads[i_layer],
+                window_size=window_size,
+                mlp_ratio=mlp_ratio,
+                qkv_bias=qkv_bias,
+                drop=drop_rate, attn_drop=attn_drop_rate,
+                drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
+                norm_layer=norm_layer,
+                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
+                pretrained_window_size=pretrained_window_sizes[i_layer]
+            )
+            self.layers.append(layer)
+
+        self.norm = norm_layer(self.num_features)
+        self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
+
+        self.apply(self._init_weights)
+        for bly in self.layers:
+            bly._init_respostnorm()
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+
+    @torch.jit.ignore
+    def no_weight_decay(self):
+        nod = {'absolute_pos_embed'}
+        for n, m in self.named_modules():
+            if any([kw in n for kw in ("cpb_mlp", "logit_scale", 'relative_position_bias_table')]):
+                nod.add(n)
+        return nod
+
+    @torch.jit.ignore
+    def group_matcher(self, coarse=False):
+        return dict(
+            stem=r'^absolute_pos_embed|patch_embed',  # stem and embed
+            blocks=r'^layers\.(\d+)' if coarse else [
+                (r'^layers\.(\d+).downsample', (0,)),
+                (r'^layers\.(\d+)\.\w+\.(\d+)', None),
+                (r'^norm', (99999,)),
+            ]
+        )
+
+    @torch.jit.ignore
+    def set_grad_checkpointing(self, enable=True):
+        for l in self.layers:
+            l.grad_checkpointing = enable
+
+    @torch.jit.ignore
+    def get_classifier(self):
+        return self.head
+
+    def reset_classifier(self, num_classes, global_pool=None):
+        self.num_classes = num_classes
+        if global_pool is not None:
+            assert global_pool in ('', 'avg')
+            self.global_pool = global_pool
+        self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()
+
+    def forward_features(self, x):
+        x = self.patch_embed(x)
+        if self.absolute_pos_embed is not None:
+            x = x + self.absolute_pos_embed
+        x = self.pos_drop(x)
+
+        for layer in self.layers:
+            x = layer(x)
+
+        x = self.norm(x)  # B L C
+        return x
+
+    def forward_head(self, x, pre_logits: bool = False):
+        if self.global_pool == 'avg':
+            x = x.mean(dim=1)
+        return x if pre_logits else self.head(x)
+
+    def forward(self, x):
+        x = self.forward_features(x)
+        x = self.forward_head(x)
+        return x
+
+
+def checkpoint_filter_fn(state_dict, model):
+    out_dict = {}
+    if 'model' in state_dict:
+        # For deit models
+        state_dict = state_dict['model']
+    for k, v in state_dict.items():
+        if any([n in k for n in ('relative_position_index', 'relative_coords_table')]):
+            continue  # skip buffers that should not be persistent
+        out_dict[k] = v
+    return out_dict
+
+
+def _create_swin_transformer_v2(variant, pretrained=False, **kwargs):
+    model = build_model_with_cfg(
+        SwinTransformerV2, variant, pretrained,
+        pretrained_filter_fn=checkpoint_filter_fn,
+        **kwargs)
+    return model
+
+
+@register_model
+def swinv2_tiny_window16_256(pretrained=False, **kwargs):
+    """
+    """
+    model_kwargs = dict(
+        window_size=16, embed_dim=96, depths=(2, 2, 6, 2), num_heads=(3, 6, 12, 24), **kwargs)
+    return _create_swin_transformer_v2('swinv2_tiny_window16_256', pretrained=pretrained, **model_kwargs)
+
+
+@register_model
+def swinv2_tiny_window8_256(pretrained=False, **kwargs):
+    """
+    """
+    model_kwargs = dict(
+        window_size=8, embed_dim=96, depths=(2, 2, 6, 2), num_heads=(3, 6, 12, 24), **kwargs)
+    return _create_swin_transformer_v2('swinv2_tiny_window8_256', pretrained=pretrained, **model_kwargs)
+
+
+@register_model
+def swinv2_small_window16_256(pretrained=False, **kwargs):
+    """
+    """
+    model_kwargs = dict(
+        window_size=16, embed_dim=96, depths=(2, 2, 18, 2), num_heads=(3, 6, 12, 24), **kwargs)
+    return _create_swin_transformer_v2('swinv2_small_window16_256', pretrained=pretrained, **model_kwargs)
+
+
+@register_model
+def swinv2_small_window8_256(pretrained=False, **kwargs):
+    """
+    """
+    model_kwargs = dict(
+        window_size=8, embed_dim=96, depths=(2, 2, 18, 2), num_heads=(3, 6, 12, 24), **kwargs)
+    return _create_swin_transformer_v2('swinv2_small_window8_256', pretrained=pretrained, **model_kwargs)
+
+
+@register_model
+def swinv2_base_window16_256(pretrained=False, **kwargs):
+    """
+    """
+    model_kwargs = dict(
+        window_size=16, embed_dim=128, depths=(2, 2, 18, 2), num_heads=(4, 8, 16, 32), **kwargs)
+    return _create_swin_transformer_v2('swinv2_base_window16_256', pretrained=pretrained, **model_kwargs)
+
+
+@register_model
+def swinv2_base_window8_256(pretrained=False, **kwargs):
+    """
+    """
+    model_kwargs = dict(
+        window_size=8, embed_dim=128, depths=(2, 2, 18, 2), num_heads=(4, 8, 16, 32), **kwargs)
+    return _create_swin_transformer_v2('swinv2_base_window8_256', pretrained=pretrained, **model_kwargs)
+
+
+@register_model
+def swinv2_base_window12_192_22k(pretrained=False, **kwargs):
+    """
+    """
+    model_kwargs = dict(
+        window_size=12, embed_dim=128, depths=(2, 2, 18, 2), num_heads=(4, 8, 16, 32), **kwargs)
+    return _create_swin_transformer_v2('swinv2_base_window12_192_22k', pretrained=pretrained, **model_kwargs)
+
+
+@register_model
+def swinv2_base_window12to16_192to256_22kft1k(pretrained=False, **kwargs):
+    """
+    """
+    model_kwargs = dict(
+        window_size=16, embed_dim=128, depths=(2, 2, 18, 2), num_heads=(4, 8, 16, 32),
+        pretrained_window_sizes=(12, 12, 12, 6), **kwargs)
+    return _create_swin_transformer_v2(
+        'swinv2_base_window12to16_192to256_22kft1k', pretrained=pretrained, **model_kwargs)
+
+
+@register_model
+def swinv2_base_window12to24_192to384_22kft1k(pretrained=False, **kwargs):
+    """
+    """
+    model_kwargs = dict(
+        window_size=24, embed_dim=128, depths=(2, 2, 18, 2), num_heads=(4, 8, 16, 32),
+        pretrained_window_sizes=(12, 12, 12, 6), **kwargs)
+    return _create_swin_transformer_v2(
+        'swinv2_base_window12to24_192to384_22kft1k', pretrained=pretrained, **model_kwargs)
+
+
+@register_model
+def swinv2_large_window12_192_22k(pretrained=False, **kwargs):
+    """
+    """
+    model_kwargs = dict(
+        window_size=12, embed_dim=192, depths=(2, 2, 18, 2), num_heads=(6, 12, 24, 48), **kwargs)
+    return _create_swin_transformer_v2('swinv2_large_window12_192_22k', pretrained=pretrained, **model_kwargs)
+
+
+@register_model
+def swinv2_large_window12to16_192to256_22kft1k(pretrained=False, **kwargs):
+    """
+    """
+    model_kwargs = dict(
+        window_size=16, embed_dim=192, depths=(2, 2, 18, 2), num_heads=(6, 12, 24, 48),
+        pretrained_window_sizes=(12, 12, 12, 6), **kwargs)
+    return _create_swin_transformer_v2(
+        'swinv2_large_window12to16_192to256_22kft1k', pretrained=pretrained, **model_kwargs)
+
+
+@register_model
+def swinv2_large_window12to24_192to384_22kft1k(pretrained=False, **kwargs):
+    """
+    """
+    model_kwargs = dict(
+        window_size=24, embed_dim=192, depths=(2, 2, 18, 2), num_heads=(6, 12, 24, 48),
+        pretrained_window_sizes=(12, 12, 12, 6), **kwargs)
+    return _create_swin_transformer_v2(
+        'swinv2_large_window12to24_192to384_22kft1k', pretrained=pretrained, **model_kwargs)
diff --git a/timm/models/swin_transformer_v2_cr.py b/timm/models/swin_transformer_v2_cr.py
index 472ae205..d143c14c 100644
--- a/timm/models/swin_transformer_v2_cr.py
+++ b/timm/models/swin_transformer_v2_cr.py
@@ -34,6 +34,7 @@ from typing import Tuple, Optional, List, Union, Any, Type
 
 import torch
 import torch.nn as nn
+import torch.nn.functional as F
 import torch.utils.checkpoint as checkpoint
 
 from timm.data import IMAGENET_DEFAULT_MEAN, IMAGENET_DEFAULT_STD
@@ -41,7 +42,7 @@ from .fx_features import register_notrace_function
 from .helpers import build_model_with_cfg, named_apply
 from .layers import DropPath, Mlp, to_2tuple, _assert
 from .registry import register_model
-from .vision_transformer import checkpoint_filter_fn
+
 
 _logger = logging.getLogger(__name__)
 
@@ -51,7 +52,7 @@ def _cfg(url='', **kwargs):
         'url': url,
         'num_classes': 1000,
         'input_size': (3, 224, 224),
-        'pool_size': None,
+        'pool_size': (7, 7),
         'crop_pct': 0.9,
         'interpolation': 'bicubic',
         'fixed_input_size': True,
@@ -64,38 +65,38 @@ def _cfg(url='', **kwargs):
 
 
 default_cfgs = {
-    'swin_v2_cr_tiny_384': _cfg(
-        url="", input_size=(3, 384, 384), crop_pct=1.0),
-    'swin_v2_cr_tiny_224': _cfg(
+    'swinv2_cr_tiny_384': _cfg(
+        url="", input_size=(3, 384, 384), crop_pct=1.0, pool_size=(12, 12)),
+    'swinv2_cr_tiny_224': _cfg(
         url="", input_size=(3, 224, 224), crop_pct=0.9),
-    'swin_v2_cr_tiny_ns_224': _cfg(
+    'swinv2_cr_tiny_ns_224': _cfg(
         url="https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights-swinv2/swin_v2_cr_tiny_ns_224-ba8166c6.pth",
         input_size=(3, 224, 224), crop_pct=0.9),
-    'swin_v2_cr_small_384': _cfg(
-        url="", input_size=(3, 384, 384), crop_pct=1.0),
-    'swin_v2_cr_small_224': _cfg(
+    'swinv2_cr_small_384': _cfg(
+        url="", input_size=(3, 384, 384), crop_pct=1.0, pool_size=(12, 12)),
+    'swinv2_cr_small_224': _cfg(
         url="https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights-swinv2/swin_v2_cr_small_224-0813c165.pth",
         input_size=(3, 224, 224), crop_pct=0.9),
-    'swin_v2_cr_small_ns_224': _cfg(
+    'swinv2_cr_small_ns_224': _cfg(
         url="https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-weights-swinv2/swin_v2_cr_small_ns_224_iv-2ce90f8e.pth",
         input_size=(3, 224, 224), crop_pct=0.9),
-    'swin_v2_cr_base_384': _cfg(
-        url="", input_size=(3, 384, 384), crop_pct=1.0),
-    'swin_v2_cr_base_224': _cfg(
+    'swinv2_cr_base_384': _cfg(
+        url="", input_size=(3, 384, 384), crop_pct=1.0, pool_size=(12, 12)),
+    'swinv2_cr_base_224': _cfg(
         url="", input_size=(3, 224, 224), crop_pct=0.9),
-    'swin_v2_cr_base_ns_224': _cfg(
+    'swinv2_cr_base_ns_224': _cfg(
         url="", input_size=(3, 224, 224), crop_pct=0.9),
-    'swin_v2_cr_large_384': _cfg(
-        url="", input_size=(3, 384, 384), crop_pct=1.0),
-    'swin_v2_cr_large_224': _cfg(
+    'swinv2_cr_large_384': _cfg(
+        url="", input_size=(3, 384, 384), crop_pct=1.0, pool_size=(12, 12)),
+    'swinv2_cr_large_224': _cfg(
         url="", input_size=(3, 224, 224), crop_pct=0.9),
-    'swin_v2_cr_huge_384': _cfg(
-        url="", input_size=(3, 384, 384), crop_pct=1.0),
-    'swin_v2_cr_huge_224': _cfg(
+    'swinv2_cr_huge_384': _cfg(
+        url="", input_size=(3, 384, 384), crop_pct=1.0, pool_size=(12, 12)),
+    'swinv2_cr_huge_224': _cfg(
         url="", input_size=(3, 224, 224), crop_pct=0.9),
-    'swin_v2_cr_giant_384': _cfg(
-        url="", input_size=(3, 384, 384), crop_pct=1.0),
-    'swin_v2_cr_giant_224': _cfg(
+    'swinv2_cr_giant_384': _cfg(
+        url="", input_size=(3, 384, 384), crop_pct=1.0, pool_size=(12, 12)),
+    'swinv2_cr_giant_224': _cfg(
         url="", input_size=(3, 224, 224), crop_pct=0.9),
 }
 
@@ -186,12 +187,13 @@ class WindowMultiHeadAttention(nn.Module):
             act_layer=nn.ReLU,
             drop=(0.125, 0.)  # FIXME should there be stochasticity, appears to 'overfit' without?
         )
-        self.register_parameter("tau", torch.nn.Parameter(torch.ones(num_heads)))
+        # NOTE old checkpoints used inverse of logit_scale ('tau') following the paper, see conversion fn
+        self.logit_scale = nn.Parameter(torch.log(10 * torch.ones(num_heads)))
         self._make_pair_wise_relative_positions()
 
     def _make_pair_wise_relative_positions(self) -> None:
         """Method initializes the pair-wise relative positions to compute the positional biases."""
-        device = self.tau.device
+        device = self.logit_scale.device
         coordinates = torch.stack(torch.meshgrid([
             torch.arange(self.window_size[0], device=device),
             torch.arange(self.window_size[1], device=device)]), dim=0).flatten(1)
@@ -250,10 +252,11 @@ class WindowMultiHeadAttention(nn.Module):
         query, key, value = qkv.unbind(0)
 
         # compute attention map with scaled cosine attention
-        denom = torch.norm(query, dim=-1, keepdim=True) @ torch.norm(key, dim=-1, keepdim=True).transpose(-2, -1)
-        attn = query @ key.transpose(-2, -1) / denom.clamp(min=1e-6)
-        attn = attn / self.tau.clamp(min=0.01).reshape(1, self.num_heads, 1, 1)
+        attn = (F.normalize(query, dim=-1) @ F.normalize(key, dim=-1).transpose(-2, -1))
+        logit_scale = torch.clamp(self.logit_scale.reshape(1, self.num_heads, 1, 1), max=math.log(1. / 0.01)).exp()
+        attn = attn * logit_scale
         attn = attn + self._relative_positional_encodings()
+
         if mask is not None:
             # Apply mask if utilized
             num_win: int = mask.shape[0]
@@ -309,7 +312,7 @@ class SwinTransformerBlock(nn.Module):
         window_size: Tuple[int, int],
         shift_size: Tuple[int, int] = (0, 0),
         mlp_ratio: float = 4.0,
-        init_values: float = 0,
+        init_values: Optional[float] = 0,
         drop: float = 0.0,
         drop_attn: float = 0.0,
         drop_path: float = 0.0,
@@ -323,7 +326,7 @@ class SwinTransformerBlock(nn.Module):
         self.target_shift_size: Tuple[int, int] = to_2tuple(shift_size)
         self.window_size, self.shift_size = self._calc_window_shift(to_2tuple(window_size))
         self.window_area = self.window_size[0] * self.window_size[1]
-        self.init_values: float = init_values
+        self.init_values: Optional[float] = init_values
 
         # attn branch
         self.attn = WindowMultiHeadAttention(
@@ -387,7 +390,7 @@ class SwinTransformerBlock(nn.Module):
 
     def init_weights(self):
         # extra, module specific weight init
-        if self.init_values:
+        if self.init_values is not None:
             nn.init.constant_(self.norm1.weight, self.init_values)
             nn.init.constant_(self.norm2.weight, self.init_values)
 
@@ -536,7 +539,7 @@ class SwinTransformerStage(nn.Module):
         feat_size: Tuple[int, int],
         window_size: Tuple[int, int],
         mlp_ratio: float = 4.0,
-        init_values: float = 0.0,
+        init_values: Optional[float] = 0.0,
         drop: float = 0.0,
         drop_attn: float = 0.0,
         drop_path: Union[List[float], float] = 0.0,
@@ -650,7 +653,7 @@ class SwinTransformerV2Cr(nn.Module):
         depths: Tuple[int, ...] = (2, 2, 6, 2),
         num_heads: Tuple[int, ...] = (3, 6, 12, 24),
         mlp_ratio: float = 4.0,
-        init_values: float = 0.0,
+        init_values: Optional[float] = 0.,
         drop_rate: float = 0.0,
         attn_drop_rate: float = 0.0,
         drop_path_rate: float = 0.0,
@@ -808,6 +811,21 @@ def init_weights(module: nn.Module, name: str = ''):
         module.init_weights()
 
 
+def checkpoint_filter_fn(state_dict, model):
+    """ convert patch embedding weight from manual patchify + linear proj to conv"""
+    out_dict = {}
+    if 'model' in state_dict:
+        # For deit models
+        state_dict = state_dict['model']
+    for k, v in state_dict.items():
+        if 'tau' in k:
+            # convert old tau based checkpoints -> logit_scale (inverse)
+            v = torch.log(1 / v)
+            k = k.replace('tau', 'logit_scale')
+        out_dict[k] = v
+    return out_dict
+
+
 def _create_swin_transformer_v2_cr(variant, pretrained=False, **kwargs):
     if kwargs.get('features_only', None):
         raise RuntimeError('features_only not implemented for Vision Transformer models.')
@@ -820,7 +838,7 @@ def _create_swin_transformer_v2_cr(variant, pretrained=False, **kwargs):
 
 
 @register_model
-def swin_v2_cr_tiny_384(pretrained=False, **kwargs):
+def swinv2_cr_tiny_384(pretrained=False, **kwargs):
     """Swin-T V2 CR @ 384x384, trained ImageNet-1k"""
     model_kwargs = dict(
         embed_dim=96,
@@ -828,11 +846,11 @@ def swin_v2_cr_tiny_384(pretrained=False, **kwargs):
         num_heads=(3, 6, 12, 24),
         **kwargs
     )
-    return _create_swin_transformer_v2_cr('swin_v2_cr_tiny_384', pretrained=pretrained, **model_kwargs)
+    return _create_swin_transformer_v2_cr('swinv2_cr_tiny_384', pretrained=pretrained, **model_kwargs)
 
 
 @register_model
-def swin_v2_cr_tiny_224(pretrained=False, **kwargs):
+def swinv2_cr_tiny_224(pretrained=False, **kwargs):
     """Swin-T V2 CR @ 224x224, trained ImageNet-1k"""
     model_kwargs = dict(
         embed_dim=96,
@@ -840,11 +858,11 @@ def swin_v2_cr_tiny_224(pretrained=False, **kwargs):
         num_heads=(3, 6, 12, 24),
         **kwargs
     )
-    return _create_swin_transformer_v2_cr('swin_v2_cr_tiny_224', pretrained=pretrained, **model_kwargs)
+    return _create_swin_transformer_v2_cr('swinv2_cr_tiny_224', pretrained=pretrained, **model_kwargs)
 
 
 @register_model
-def swin_v2_cr_tiny_ns_224(pretrained=False, **kwargs):
+def swinv2_cr_tiny_ns_224(pretrained=False, **kwargs):
     """Swin-T V2 CR @ 224x224, trained ImageNet-1k w/ extra stage norms.
     ** Experimental, may make default if results are improved. **
     """
@@ -855,11 +873,11 @@ def swin_v2_cr_tiny_ns_224(pretrained=False, **kwargs):
         extra_norm_stage=True,
         **kwargs
     )
-    return _create_swin_transformer_v2_cr('swin_v2_cr_tiny_ns_224', pretrained=pretrained, **model_kwargs)
+    return _create_swin_transformer_v2_cr('swinv2_cr_tiny_ns_224', pretrained=pretrained, **model_kwargs)
 
 
 @register_model
-def swin_v2_cr_small_384(pretrained=False, **kwargs):
+def swinv2_cr_small_384(pretrained=False, **kwargs):
     """Swin-S V2 CR @ 384x384, trained ImageNet-1k"""
     model_kwargs = dict(
         embed_dim=96,
@@ -867,12 +885,12 @@ def swin_v2_cr_small_384(pretrained=False, **kwargs):
         num_heads=(3, 6, 12, 24),
         **kwargs
     )
-    return _create_swin_transformer_v2_cr('swin_v2_cr_small_384', pretrained=pretrained, **model_kwargs
+    return _create_swin_transformer_v2_cr('swinv2_cr_small_384', pretrained=pretrained, **model_kwargs
     )
 
 
 @register_model
-def swin_v2_cr_small_224(pretrained=False, **kwargs):
+def swinv2_cr_small_224(pretrained=False, **kwargs):
     """Swin-S V2 CR @ 224x224, trained ImageNet-1k"""
     model_kwargs = dict(
         embed_dim=96,
@@ -880,25 +898,24 @@ def swin_v2_cr_small_224(pretrained=False, **kwargs):
         num_heads=(3, 6, 12, 24),
         **kwargs
     )
-    return _create_swin_transformer_v2_cr('swin_v2_cr_small_224', pretrained=pretrained, **model_kwargs)
+    return _create_swin_transformer_v2_cr('swinv2_cr_small_224', pretrained=pretrained, **model_kwargs)
 
 
 @register_model
-def swin_v2_cr_small_ns_224(pretrained=False, **kwargs):
+def swinv2_cr_small_ns_224(pretrained=False, **kwargs):
     """Swin-S V2 CR @ 224x224, trained ImageNet-1k"""
     model_kwargs = dict(
         embed_dim=96,
         depths=(2, 2, 18, 2),
         num_heads=(3, 6, 12, 24),
-        init_values=1e-5,
         extra_norm_stage=True,
         **kwargs
     )
-    return _create_swin_transformer_v2_cr('swin_v2_cr_small_ns_224', pretrained=pretrained, **model_kwargs)
+    return _create_swin_transformer_v2_cr('swinv2_cr_small_ns_224', pretrained=pretrained, **model_kwargs)
 
 
 @register_model
-def swin_v2_cr_base_384(pretrained=False, **kwargs):
+def swinv2_cr_base_384(pretrained=False, **kwargs):
     """Swin-B V2 CR @ 384x384, trained ImageNet-1k"""
     model_kwargs = dict(
         embed_dim=128,
@@ -906,11 +923,11 @@ def swin_v2_cr_base_384(pretrained=False, **kwargs):
         num_heads=(4, 8, 16, 32),
         **kwargs
     )
-    return _create_swin_transformer_v2_cr('swin_v2_cr_base_384', pretrained=pretrained, **model_kwargs)
+    return _create_swin_transformer_v2_cr('swinv2_cr_base_384', pretrained=pretrained, **model_kwargs)
 
 
 @register_model
-def swin_v2_cr_base_224(pretrained=False, **kwargs):
+def swinv2_cr_base_224(pretrained=False, **kwargs):
     """Swin-B V2 CR @ 224x224, trained ImageNet-1k"""
     model_kwargs = dict(
         embed_dim=128,
@@ -918,25 +935,24 @@ def swin_v2_cr_base_224(pretrained=False, **kwargs):
         num_heads=(4, 8, 16, 32),
         **kwargs
     )
-    return _create_swin_transformer_v2_cr('swin_v2_cr_base_224', pretrained=pretrained, **model_kwargs)
+    return _create_swin_transformer_v2_cr('swinv2_cr_base_224', pretrained=pretrained, **model_kwargs)
 
 
 @register_model
-def swin_v2_cr_base_ns_224(pretrained=False, **kwargs):
+def swinv2_cr_base_ns_224(pretrained=False, **kwargs):
     """Swin-B V2 CR @ 224x224, trained ImageNet-1k"""
     model_kwargs = dict(
         embed_dim=128,
         depths=(2, 2, 18, 2),
         num_heads=(4, 8, 16, 32),
-        init_values=1e-6,
         extra_norm_stage=True,
         **kwargs
     )
-    return _create_swin_transformer_v2_cr('swin_v2_cr_base_ns_224', pretrained=pretrained, **model_kwargs)
+    return _create_swin_transformer_v2_cr('swinv2_cr_base_ns_224', pretrained=pretrained, **model_kwargs)
 
 
 @register_model
-def swin_v2_cr_large_384(pretrained=False, **kwargs):
+def swinv2_cr_large_384(pretrained=False, **kwargs):
     """Swin-L V2 CR @ 384x384, trained ImageNet-1k"""
     model_kwargs = dict(
         embed_dim=192,
@@ -944,12 +960,12 @@ def swin_v2_cr_large_384(pretrained=False, **kwargs):
         num_heads=(6, 12, 24, 48),
         **kwargs
     )
-    return _create_swin_transformer_v2_cr('swin_v2_cr_large_384', pretrained=pretrained, **model_kwargs
+    return _create_swin_transformer_v2_cr('swinv2_cr_large_384', pretrained=pretrained, **model_kwargs
     )
 
 
 @register_model
-def swin_v2_cr_large_224(pretrained=False, **kwargs):
+def swinv2_cr_large_224(pretrained=False, **kwargs):
     """Swin-L V2 CR @ 224x224, trained ImageNet-1k"""
     model_kwargs = dict(
         embed_dim=192,
@@ -957,11 +973,11 @@ def swin_v2_cr_large_224(pretrained=False, **kwargs):
         num_heads=(6, 12, 24, 48),
         **kwargs
     )
-    return _create_swin_transformer_v2_cr('swin_v2_cr_large_224', pretrained=pretrained, **model_kwargs)
+    return _create_swin_transformer_v2_cr('swinv2_cr_large_224', pretrained=pretrained, **model_kwargs)
 
 
 @register_model
-def swin_v2_cr_huge_384(pretrained=False, **kwargs):
+def swinv2_cr_huge_384(pretrained=False, **kwargs):
     """Swin-H V2 CR @ 384x384, trained ImageNet-1k"""
     model_kwargs = dict(
         embed_dim=352,
@@ -970,11 +986,11 @@ def swin_v2_cr_huge_384(pretrained=False, **kwargs):
         extra_norm_period=6,
         **kwargs
     )
-    return _create_swin_transformer_v2_cr('swin_v2_cr_huge_384', pretrained=pretrained, **model_kwargs)
+    return _create_swin_transformer_v2_cr('swinv2_cr_huge_384', pretrained=pretrained, **model_kwargs)
 
 
 @register_model
-def swin_v2_cr_huge_224(pretrained=False, **kwargs):
+def swinv2_cr_huge_224(pretrained=False, **kwargs):
     """Swin-H V2 CR @ 224x224, trained ImageNet-1k"""
     model_kwargs = dict(
         embed_dim=352,
@@ -983,11 +999,11 @@ def swin_v2_cr_huge_224(pretrained=False, **kwargs):
         extra_norm_period=6,
         **kwargs
     )
-    return _create_swin_transformer_v2_cr('swin_v2_cr_huge_224', pretrained=pretrained, **model_kwargs)
+    return _create_swin_transformer_v2_cr('swinv2_cr_huge_224', pretrained=pretrained, **model_kwargs)
 
 
 @register_model
-def swin_v2_cr_giant_384(pretrained=False, **kwargs):
+def swinv2_cr_giant_384(pretrained=False, **kwargs):
     """Swin-G V2 CR @ 384x384, trained ImageNet-1k"""
     model_kwargs = dict(
         embed_dim=512,
@@ -996,12 +1012,12 @@ def swin_v2_cr_giant_384(pretrained=False, **kwargs):
         extra_norm_period=6,
         **kwargs
     )
-    return _create_swin_transformer_v2_cr('swin_v2_cr_giant_384', pretrained=pretrained, **model_kwargs
+    return _create_swin_transformer_v2_cr('swinv2_cr_giant_384', pretrained=pretrained, **model_kwargs
     )
 
 
 @register_model
-def swin_v2_cr_giant_224(pretrained=False, **kwargs):
+def swinv2_cr_giant_224(pretrained=False, **kwargs):
     """Swin-G V2 CR @ 224x224, trained ImageNet-1k"""
     model_kwargs = dict(
         embed_dim=512,
@@ -1010,4 +1026,4 @@ def swin_v2_cr_giant_224(pretrained=False, **kwargs):
         extra_norm_period=6,
         **kwargs
     )
-    return _create_swin_transformer_v2_cr('swin_v2_cr_giant_224', pretrained=pretrained, **model_kwargs)
+    return _create_swin_transformer_v2_cr('swinv2_cr_giant_224', pretrained=pretrained, **model_kwargs)
diff --git a/timm/models/vision_transformer_relpos.py b/timm/models/vision_transformer_relpos.py
index 9ecfd473..0c2ac376 100644
--- a/timm/models/vision_transformer_relpos.py
+++ b/timm/models/vision_transformer_relpos.py
@@ -1,5 +1,7 @@
 """ Relative Position Vision Transformer (ViT) in PyTorch
 
+NOTE: these models are experimental / WIP, expect changes
+
 Hacked together by / Copyright 2022, Ross Wightman
 """
 import math
@@ -37,9 +39,23 @@ default_cfgs = {
         url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/vit_replos_base_patch32_plus_rpn_256-sw-dd486f51.pth',
         input_size=(3, 256, 256)),
     'vit_relpos_base_patch16_plus_240': _cfg(url='', input_size=(3, 240, 240)),
-    'vit_relpos_base_patch16_rpn_224': _cfg(url=''),
+
+    'vit_relpos_small_patch16_224': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/vit_relpos_small_patch16_224-sw-ec2778b4.pth'),
+    'vit_relpos_medium_patch16_224': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/vit_relpos_medium_patch16_224-sw-11c174af.pth'),
     'vit_relpos_base_patch16_224': _cfg(
         url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/vit_relpos_base_patch16_224-sw-49049aed.pth'),
+
+    'vit_relpos_base_patch16_cls_224': _cfg(
+        url=''),
+    'vit_relpos_base_patch16_gapcls_224': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/vit_relpos_base_patch16_gapcls_224-sw-1a341d6c.pth'),
+
+    'vit_relpos_small_patch16_rpn_224': _cfg(url=''),
+    'vit_relpos_medium_patch16_rpn_224': _cfg(
+        url='https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-tpu-weights/vit_relpos_medium_patch16_rpn_224-sw-5d2befd8.pth'),
+    'vit_relpos_base_patch16_rpn_224': _cfg(url=''),
 }
 
 
@@ -66,43 +82,84 @@ def gen_relative_position_index(win_size: Tuple[int, int], class_token: int = 0)
     return relative_position_index
 
 
-def gen_relative_position_log(win_size: Tuple[int, int]) -> torch.Tensor:
-    """Method initializes the pair-wise relative positions to compute the positional biases."""
-    coordinates = torch.stack(torch.meshgrid([torch.arange(win_size[0]), torch.arange(win_size[1])])).flatten(1)
-    relative_coords = coordinates[:, :, None] - coordinates[:, None, :]
-    relative_coords = relative_coords.permute(1, 2, 0).float()
-    relative_coordinates_log = torch.sign(relative_coords) * torch.log(1.0 + relative_coords.abs())
-    return relative_coordinates_log
+def gen_relative_log_coords(
+        win_size: Tuple[int, int],
+        pretrained_win_size: Tuple[int, int] = (0, 0),
+        mode='swin'
+):
+    # as per official swin-v2 impl, supporting timm swin-v2-cr coords as well
+    relative_coords_h = torch.arange(-(win_size[0] - 1), win_size[0], dtype=torch.float32)
+    relative_coords_w = torch.arange(-(win_size[1] - 1), win_size[1], dtype=torch.float32)
+    relative_coords_table = torch.stack(torch.meshgrid([relative_coords_h, relative_coords_w]))
+    relative_coords_table = relative_coords_table.permute(1, 2, 0).contiguous()  # 2*Wh-1, 2*Ww-1, 2
+    if mode == 'swin':
+        if pretrained_win_size[0] > 0:
+            relative_coords_table[:, :, 0] /= (pretrained_win_size[0] - 1)
+            relative_coords_table[:, :, 1] /= (pretrained_win_size[1] - 1)
+        else:
+            relative_coords_table[:, :, 0] /= (win_size[0] - 1)
+            relative_coords_table[:, :, 1] /= (win_size[1] - 1)
+        relative_coords_table *= 8  # normalize to -8, 8
+        scale = math.log2(8)
+    else:
+        # FIXME we should support a form of normalization (to -1/1) for this mode?
+        scale = math.log2(math.e)
+    relative_coords_table = torch.sign(relative_coords_table) * torch.log2(
+        1.0 + relative_coords_table.abs()) / scale
+    return relative_coords_table
 
 
 class RelPosMlp(nn.Module):
-    # based on timm swin-v2 impl
-    def __init__(self, window_size, num_heads=8, hidden_dim=32, class_token=False):
+    def __init__(
+            self,
+            window_size,
+            num_heads=8,
+            hidden_dim=128,
+            class_token=False,
+            mode='cr',
+            pretrained_window_size=(0, 0)
+    ):
         super().__init__()
         self.window_size = window_size
         self.window_area = self.window_size[0] * self.window_size[1]
         self.class_token = 1 if class_token else 0
         self.num_heads = num_heads
+        self.bias_shape = (self.window_area,) * 2 + (num_heads,)
+        self.apply_sigmoid = mode == 'swin'
 
+        mlp_bias = (True, False) if mode == 'swin' else True
         self.mlp = Mlp(
             2,  # x, y
-            hidden_features=min(128, hidden_dim * num_heads),
+            hidden_features=hidden_dim,
             out_features=num_heads,
             act_layer=nn.ReLU,
+            bias=mlp_bias,
             drop=(0.125, 0.)
         )
 
         self.register_buffer(
-            'rel_coords_log',
-            gen_relative_position_log(window_size),
-            persistent=False
-        )
+            "relative_position_index",
+            gen_relative_position_index(window_size),
+            persistent=False)
+
+        # get relative_coords_table
+        self.register_buffer(
+            "rel_coords_log",
+            gen_relative_log_coords(window_size, pretrained_window_size, mode=mode),
+            persistent=False)
 
     def get_bias(self) -> torch.Tensor:
-        relative_position_bias = self.mlp(self.rel_coords_log).permute(2, 0, 1).unsqueeze(0)
+        relative_position_bias = self.mlp(self.rel_coords_log)
+        if self.relative_position_index is not None:
+            relative_position_bias = relative_position_bias.view(-1, self.num_heads)[
+                self.relative_position_index.view(-1)]  # Wh*Ww,Wh*Ww,nH
+            relative_position_bias = relative_position_bias.view(self.bias_shape)
+        relative_position_bias = relative_position_bias.permute(2, 0, 1)
+        if self.apply_sigmoid:
+            relative_position_bias = 16 * torch.sigmoid(relative_position_bias)
         if self.class_token:
             relative_position_bias = F.pad(relative_position_bias, [self.class_token, 0, self.class_token, 0])
-        return relative_position_bias
+        return relative_position_bias.unsqueeze(0).contiguous()
 
     def forward(self, attn, shared_rel_pos: Optional[torch.Tensor] = None):
         return attn + self.get_bias()
@@ -131,10 +188,10 @@ class RelPosBias(nn.Module):
         trunc_normal_(self.relative_position_bias_table, std=.02)
 
     def get_bias(self) -> torch.Tensor:
-        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
-            self.bias_shape)  # win_h * win_w, win_h * win_w, num_heads
-        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()
-        return relative_position_bias
+        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)]
+        # win_h * win_w, win_h * win_w, num_heads
+        relative_position_bias = relative_position_bias.view(self.bias_shape).permute(2, 0, 1)
+        return relative_position_bias.unsqueeze(0).contiguous()
 
     def forward(self, attn, shared_rel_pos: Optional[torch.Tensor] = None):
         return attn + self.get_bias()
@@ -250,8 +307,8 @@ class VisionTransformerRelPos(nn.Module):
 
     def __init__(
             self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, global_pool='avg',
-            embed_dim=768, depth=12, num_heads=12, mlp_ratio=4., qkv_bias=True, init_values=1e-5,
-            class_token=False, rel_pos_type='mlp', shared_rel_pos=False, fc_norm=False,
+            embed_dim=768, depth=12, num_heads=12, mlp_ratio=4., qkv_bias=True, init_values=1e-6,
+            class_token=False, fc_norm=False, rel_pos_type='mlp', shared_rel_pos=False, rel_pos_dim=None,
             drop_rate=0., attn_drop_rate=0., drop_path_rate=0., weight_init='skip',
             embed_layer=PatchEmbed, norm_layer=None, act_layer=None, block_fn=RelPosBlock):
         """
@@ -268,9 +325,9 @@ class VisionTransformerRelPos(nn.Module):
             qkv_bias (bool): enable bias for qkv if True
             init_values: (float): layer-scale init values
             class_token (bool): use class token (default: False)
+            fc_norm (bool): use pre classifier norm instead of pre-pool
             rel_pos_ty pe (str): type of relative position
             shared_rel_pos (bool): share relative pos across all blocks
-            fc_norm (bool): use pre classifier norm instead of pre-pool
             drop_rate (float): dropout rate
             attn_drop_rate (float): attention dropout rate
             drop_path_rate (float): stochastic depth rate
@@ -295,8 +352,15 @@ class VisionTransformerRelPos(nn.Module):
             img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)
         feat_size = self.patch_embed.grid_size
 
-        rel_pos_cls = RelPosMlp if rel_pos_type == 'mlp' else RelPosBias
-        rel_pos_cls = partial(rel_pos_cls, window_size=feat_size, class_token=class_token)
+        rel_pos_args = dict(window_size=feat_size, class_token=class_token)
+        if rel_pos_type.startswith('mlp'):
+            if rel_pos_dim:
+                rel_pos_args['hidden_dim'] = rel_pos_dim
+            if 'swin' in rel_pos_type:
+                rel_pos_args['mode'] = 'swin'
+            rel_pos_cls = partial(RelPosMlp, **rel_pos_args)
+        else:
+            rel_pos_cls = partial(RelPosBias, **rel_pos_args)
         self.shared_rel_pos = None
         if shared_rel_pos:
             self.shared_rel_pos = rel_pos_cls(num_heads=num_heads)
@@ -408,6 +472,26 @@ def vit_relpos_base_patch16_plus_240(pretrained=False, **kwargs):
     return model
 
 
+@register_model
+def vit_relpos_small_patch16_224(pretrained=False, **kwargs):
+    """ ViT-Base (ViT-B/16) w/ relative log-coord position, no class token
+    """
+    model_kwargs = dict(
+        patch_size=16, embed_dim=384, depth=12, num_heads=6, qkv_bias=False, fc_norm=True, **kwargs)
+    model = _create_vision_transformer_relpos('vit_relpos_small_patch16_224', pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_relpos_medium_patch16_224(pretrained=False, **kwargs):
+    """ ViT-Base (ViT-B/16) w/ relative log-coord position, no class token
+    """
+    model_kwargs = dict(
+        patch_size=16, embed_dim=512, depth=12, num_heads=8, qkv_bias=False, fc_norm=True, **kwargs)
+    model = _create_vision_transformer_relpos('vit_relpos_medium_patch16_224', pretrained=pretrained, **model_kwargs)
+    return model
+
+
 @register_model
 def vit_relpos_base_patch16_224(pretrained=False, **kwargs):
     """ ViT-Base (ViT-B/16) w/ relative log-coord position, no class token
@@ -418,11 +502,57 @@ def vit_relpos_base_patch16_224(pretrained=False, **kwargs):
     return model
 
 
+@register_model
+def vit_relpos_base_patch16_cls_224(pretrained=False, **kwargs):
+    """ ViT-Base (ViT-B/16) w/ relative log-coord position, class token present
+    """
+    model_kwargs = dict(
+        patch_size=16, embed_dim=768, depth=12, num_heads=12, qkv_bias=False,
+        class_token=True, global_pool='token', **kwargs)
+    model = _create_vision_transformer_relpos('vit_relpos_base_patch16_cls_224', pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_relpos_base_patch16_gapcls_224(pretrained=False, **kwargs):
+    """ ViT-Base (ViT-B/16) w/ relative log-coord position, class token present
+    NOTE this config is a bit of a mistake, class token was enabled but global avg-pool w/ fc-norm was not disabled
+    Leaving here for comparisons w/ a future re-train as it performs quite well.
+    """
+    model_kwargs = dict(
+        patch_size=16, embed_dim=768, depth=12, num_heads=12, qkv_bias=False, fc_norm=True, class_token=True, **kwargs)
+    model = _create_vision_transformer_relpos('vit_relpos_base_patch16_gapcls_224', pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_relpos_small_patch16_rpn_224(pretrained=False, **kwargs):
+    """ ViT-Base (ViT-B/16) w/ relative log-coord position and residual post-norm, no class token
+    """
+    model_kwargs = dict(
+        patch_size=16, embed_dim=384, depth=12, num_heads=6, qkv_bias=False, block_fn=ResPostRelPosBlock, **kwargs)
+    model = _create_vision_transformer_relpos(
+        'vit_relpos_small_patch16_rpn_224', pretrained=pretrained, **model_kwargs)
+    return model
+
+
+@register_model
+def vit_relpos_medium_patch16_rpn_224(pretrained=False, **kwargs):
+    """ ViT-Base (ViT-B/16) w/ relative log-coord position and residual post-norm, no class token
+    """
+    model_kwargs = dict(
+        patch_size=16, embed_dim=512, depth=12, num_heads=8, qkv_bias=False, block_fn=ResPostRelPosBlock, **kwargs)
+    model = _create_vision_transformer_relpos(
+        'vit_relpos_medium_patch16_rpn_224', pretrained=pretrained, **model_kwargs)
+    return model
+
+
 @register_model
 def vit_relpos_base_patch16_rpn_224(pretrained=False, **kwargs):
     """ ViT-Base (ViT-B/16) w/ relative log-coord position and residual post-norm, no class token
     """
     model_kwargs = dict(
         patch_size=16, embed_dim=768, depth=12, num_heads=12, qkv_bias=False, block_fn=ResPostRelPosBlock, **kwargs)
-    model = _create_vision_transformer_relpos('vit_relpos_base_patch16_rpn_224', pretrained=pretrained, **model_kwargs)
+    model = _create_vision_transformer_relpos(
+        'vit_relpos_base_patch16_rpn_224', pretrained=pretrained, **model_kwargs)
     return model
diff --git a/timm/utils/jit.py b/timm/utils/jit.py
index 8ebfdbff..a32cbd40 100644
--- a/timm/utils/jit.py
+++ b/timm/utils/jit.py
@@ -35,8 +35,8 @@ def set_jit_fuser(fuser):
         torch._C._jit_set_texpr_fuser_enabled(False)
     elif fuser == "nvfuser" or fuser == "nvf":
         os.environ['PYTORCH_NVFUSER_DISABLE_FALLBACK'] = '1'
-        os.environ['PYTORCH_NVFUSER_DISABLE_FMA'] = '1'
-        os.environ['PYTORCH_NVFUSER_JIT_OPT_LEVEL'] = '0'
+        #os.environ['PYTORCH_NVFUSER_DISABLE_FMA'] = '1'
+        #os.environ['PYTORCH_NVFUSER_JIT_OPT_LEVEL'] = '0'
         torch._C._jit_set_texpr_fuser_enabled(False)
         torch._C._jit_set_profiling_executor(True)
         torch._C._jit_set_profiling_mode(True)
diff --git a/timm/version.py b/timm/version.py
index 8411e551..aece342d 100644
--- a/timm/version.py
+++ b/timm/version.py
@@ -1 +1 @@
-__version__ = '0.6.1'
+__version__ = '0.6.2'