Create resampler.py

resampler.py的attn模块不适应zero3训练，在以往的finetuning脚本里只能强制性的聚拢参数，但当使用lora时这些变量名会修改，这里修改了attn模块的实现方式，主要将multihead attention的实现改成了使用直接召唤模型

resampler.py

@@ -1,27 +1,76 @@
+# Copyright (c) Alibaba Cloud.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from collections import OrderedDict
+import math
+import requests
+from io import BytesIO
 from functools import partial
+from PIL import Image
+from typing import Callable, Optional, Sequence, Tuple, List, Union
 import numpy as np
 
 import torch
 from torch import nn
+from torch.nn import functional as F
 from torch.nn.init import trunc_normal_
+from torchvision import transforms
+from torchvision.transforms import InterpolationMode
 
-def get_2d_sincos_pos_embed(embed_dim, image_size):
+from functools import partial
+import numpy as np
+import warnings
+from typing import Optional, Tuple
+import torch
+from torch import nn
+from torch import Tensor
+import deepspeed
+import torch.nn.functional as F 
+from torch.nn.functional import *
+from torch.nn.modules.activation import *
+from torch.nn.init import trunc_normal_
+from torch.nn.init import constant_, xavier_normal_, xavier_uniform_
+from transformers import PreTrainedModel
+from transformers.integrations import is_deepspeed_zero3_enabled
+def get_abs_pos(abs_pos, tgt_size):
+    # abs_pos: L, C
+    # tgt_size: (H, W)
+    # return: M, C
+    src_size = int(math.sqrt(abs_pos.size(0)))
+    # tgt_size = int(math.sqrt(tgt_size))
+    dtype = abs_pos.dtype
+
+    return F.interpolate(
+        abs_pos.float().reshape(1, src_size, src_size, -1).permute(0, 3, 1, 2),
+        size=(tgt_size[0], tgt_size[1]),
+        mode="bicubic",
+        align_corners=False,
+    ).permute(0, 2, 3, 1).flatten(0, 2).to(dtype=dtype)
+
+
+# https://github.com/facebookresearch/mae/blob/efb2a8062c206524e35e47d04501ed4f544c0ae8/util/pos_embed.py#L20
+def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
     """
-    image_size: image_size or (image_height, image_width)
+    grid_size: int of the grid height and width
     return:
-    pos_embed: [image_height, image_width, embed_dim]
+    pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
     """
-    if isinstance(image_size, int):
-        grid_h_size, grid_w_size = image_size, image_size
+    if isinstance(grid_size, int):
+        grid_h_size, grid_w_size = grid_size, grid_size
     else:
-        grid_h_size, grid_w_size = image_size[0], image_size[1]
+        grid_h_size, grid_w_size = grid_size[0], grid_size[1]
 
     grid_h = np.arange(grid_h_size, dtype=np.float32)
     grid_w = np.arange(grid_w_size, dtype=np.float32)
     grid = np.meshgrid(grid_w, grid_h)  # here w goes first
     grid = np.stack(grid, axis=0)
 
+    grid = grid.reshape([2, 1, grid_h_size, grid_w_size])
     pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+    if cls_token:
+        pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
     return pos_embed
 
 
@@ -29,57 +78,60 @@ def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
     assert embed_dim % 2 == 0
 
     # use half of dimensions to encode grid_h
-    emb_h = get_1d_sincos_pos_embed_from_grid_new(embed_dim // 2, grid[0])  # (H, W, D/2)
-    emb_w = get_1d_sincos_pos_embed_from_grid_new(embed_dim // 2, grid[1])  # (H, W, D/2)
+    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0])  # (H*W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1])  # (H*W, D/2)
 
-    emb = np.concatenate([emb_h, emb_w], axis=-1)  # (H, W, D)
+    emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)
     return emb
 
 
-def get_1d_sincos_pos_embed_from_grid_new(embed_dim, pos):
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
     """
     embed_dim: output dimension for each position
-    pos: a list of positions to be encoded: size (H, W)
-    out: (H, W, D)
+    pos: a list of positions to be encoded: size (M,)
+    out: (M, D)
     """
     assert embed_dim % 2 == 0
     omega = np.arange(embed_dim // 2, dtype=np.float32)
     omega /= embed_dim / 2.
     omega = 1. / 10000 ** omega  # (D/2,)
 
-    out = np.einsum('hw,d->hwd', pos, omega)  # (H, W, D/2), outer product
+    pos = pos.reshape(-1)  # (M,)
+    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product
 
-    emb_sin = np.sin(out)  # (H, W, D/2)
-    emb_cos = np.cos(out)  # (H, W, D/2)
+    emb_sin = np.sin(out)  # (M, D/2)
+    emb_cos = np.cos(out)  # (M, D/2)
 
-    emb = np.concatenate([emb_sin, emb_cos], axis=-1)  # (H, W, D)
+    emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
     return emb
 
 
 class Resampler(nn.Module):
     """
     A 2D perceiver-resampler network with one cross attention layers by
-       given learnable queries and 2d sincos pos_emb
+        (grid_size**2) learnable queries and 2d sincos pos_emb
     Outputs:
-        A tensor with the shape of (batch_size, num_queries, embed_dim)
+        A tensor with the shape of (grid_size**2, embed_dim)
     """
 
     def __init__(
             self,
-            num_queries,
+            grid_size,
             embed_dim,
             num_heads,
             kv_dim=None,
             norm_layer=partial(nn.LayerNorm, eps=1e-6),
-            adaptive=False,
-            max_size=(70, 70),
+            adaptive=False
     ):
         super().__init__()
-        self.num_queries = num_queries
+        self.num_queries = grid_size ** 2
         self.embed_dim = embed_dim
         self.num_heads = num_heads
         self.adaptive = adaptive
-        self.max_size = max_size
+
+        self.pos_embed = nn.Parameter(
+            torch.from_numpy(get_2d_sincos_pos_embed(embed_dim, grid_size)).float()
+        ).requires_grad_(False)
 
         self.query = nn.Parameter(torch.zeros(self.num_queries, embed_dim))
         trunc_normal_(self.query, std=.02)
@@ -89,27 +141,15 @@ class Resampler(nn.Module):
         else:
             self.kv_proj = nn.Identity()
 
-        self.attn = nn.MultiheadAttention(embed_dim, num_heads)
+        self.attn = MultiheadAttention(embed_dim, num_heads)
         self.ln_q = norm_layer(embed_dim)
         self.ln_kv = norm_layer(embed_dim)
 
         self.ln_post = norm_layer(embed_dim)
         self.proj = nn.Parameter((embed_dim ** -0.5) * torch.randn(embed_dim, embed_dim))
 
-        self._set_2d_pos_cache(self.max_size)
         self.apply(self._init_weights)
 
-    def _set_2d_pos_cache(self, max_size, device='cpu'):
-        pos_embed = torch.from_numpy(get_2d_sincos_pos_embed(self.embed_dim, max_size)).float().to(device)
-        self.register_buffer("pos_embed", pos_embed, persistent=False)
-
-    def _adjust_pos_cache(self, tgt_sizes, device):
-        max_h = torch.max(tgt_sizes[:, 0])
-        max_w = torch.max(tgt_sizes[:, 1])
-        if max_h > self.max_size[0] or max_w > self.max_size[1]:
-            self.max_size = [max(max_h, self.max_size[0]), max(max_w, self.max_size[1])]
-            self._set_2d_pos_cache(self.max_size, device)
-
     def _init_weights(self, m):
         if isinstance(m, nn.Linear):
             trunc_normal_(m.weight, std=.02)
@@ -119,45 +159,676 @@ class Resampler(nn.Module):
             nn.init.constant_(m.bias, 0)
             nn.init.constant_(m.weight, 1.0)
 
-    def forward(self, x, tgt_sizes=None):
-        assert x.shape[0] == tgt_sizes.shape[0]
-        bs = x.shape[0]
-
-        device = x.device
-        dtype = x.dtype
-
-        patch_len = tgt_sizes[:, 0] * tgt_sizes[:, 1]
-
-        self._adjust_pos_cache(tgt_sizes, device=device)
-
-        max_patch_len = torch.max(patch_len)
-        key_padding_mask = torch.zeros((bs, max_patch_len), dtype=torch.bool, device=device)
-
-        pos_embed = []
-        for i in range(bs):
-            tgt_h, tgt_w = tgt_sizes[i]
-            pos_embed.append(self.pos_embed[:tgt_h, :tgt_w, :].reshape((tgt_h * tgt_w, -1)).to(dtype))  # patches * D
-            key_padding_mask[i, patch_len[i]:] = True
-
-        pos_embed = torch.nn.utils.rnn.pad_sequence(
-            pos_embed, batch_first=True, padding_value=0.0).permute(1, 0, 2)  # BLD => L * B * D
-
-        x = self.kv_proj(x)  # B * L * D
-        x = self.ln_kv(x).permute(1, 0, 2)  # L * B * D
+    def forward(self, x, tgt_size=None, attn_mask=None):
+        if self.adaptive:
+            pos_embed = torch.Tensor(get_2d_sincos_pos_embed(self.embed_dim, tgt_size)).float().to(device=x.device, dtype=x.dtype)
+        else:
+            pos_embed = get_abs_pos(self.pos_embed, tgt_size)
 
-        q = self.ln_q(self.query)  # Q * D
+        x = self.kv_proj(x)
+        x = self.ln_kv(x).permute(1, 0, 2)
 
+        N = x.shape[1]
+        q = self.ln_q(self.query)
         out = self.attn(
-            self._repeat(q, bs),  # Q * B * D
-            x + pos_embed,  # L * B * D +  L * B * D
+            self._repeat(q, N) + self.pos_embed.unsqueeze(1),
+            x + pos_embed.unsqueeze(1),
             x,
-            key_padding_mask=key_padding_mask)[0]
-        #  out: Q * B * D
-        x = out.permute(1, 0, 2)  # B * Q * D
+            attn_mask=attn_mask)[0]
+        x = out.permute(1, 0, 2)
 
         x = self.ln_post(x)
         x = x @ self.proj
         return x
 
     def _repeat(self, query, N: int):
-        return query.unsqueeze(1).repeat(1, N, 1)
+        return query.unsqueeze(1).repeat(1, N, 1)
+
+
+
+class MultiheadAttention(nn.MultiheadAttention):
+    def __init__(self, embed_dim, num_heads, dropout=0., bias=True, add_bias_kv=False, 
+                 add_zero_attn=False, kdim=None, vdim=None, batch_first=False, device=None, dtype=None):
+        super().__init__(embed_dim, num_heads, dropout, bias, add_bias_kv, add_zero_attn, kdim, vdim, batch_first, device, dtype)
+
+        # rewrite out_proj layer，with nn.Linear
+        self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias, device=device, dtype=dtype)
+
+    def forward(
+                self,
+                query: Tensor,
+                key: Tensor,
+                value: Tensor,
+                key_padding_mask: Optional[Tensor] = None,
+                need_weights: bool = True,
+                attn_mask: Optional[Tensor] = None,
+                average_attn_weights: bool = True,
+                is_causal : bool = False) -> Tuple[Tensor, Optional[Tensor]]:
+        why_not_fast_path = ''
+        if ((attn_mask is not None and torch.is_floating_point(attn_mask))
+           or (key_padding_mask is not None) and torch.is_floating_point(key_padding_mask)):
+            why_not_fast_path = "floating-point masks are not supported for fast path."
+
+        is_batched = query.dim() == 3
+
+        key_padding_mask = F._canonical_mask(
+            mask=key_padding_mask,
+            mask_name="key_padding_mask",
+            other_type=F._none_or_dtype(attn_mask),
+            other_name="attn_mask",
+            target_type=query.dtype
+        )
+
+        attn_mask = F._canonical_mask(
+            mask=attn_mask,
+            mask_name="attn_mask",
+            other_type=None,
+            other_name="",
+            target_type=query.dtype,
+            check_other=False,
+        )
+
+
+        if not is_batched:
+            why_not_fast_path = f"input not batched; expected query.dim() of 3 but got {query.dim()}"
+        elif query is not key or key is not value:
+            # When lifting this restriction, don't forget to either
+            # enforce that the dtypes all match or test cases where
+            # they don't!
+            why_not_fast_path = "non-self attention was used (query, key, and value are not the same Tensor)"
+        elif self.in_proj_bias is not None and query.dtype != self.in_proj_bias.dtype:
+            why_not_fast_path = f"dtypes of query ({query.dtype}) and self.in_proj_bias ({self.in_proj_bias.dtype}) don't match"
+        elif self.in_proj_weight is None:
+            why_not_fast_path = "in_proj_weight was None"
+        elif query.dtype != self.in_proj_weight.dtype:
+            # this case will fail anyway, but at least they'll get a useful error message.
+            why_not_fast_path = f"dtypes of query ({query.dtype}) and self.in_proj_weight ({self.in_proj_weight.dtype}) don't match"
+        elif self.training:
+            why_not_fast_path = "training is enabled"
+        elif (self.num_heads % 2) != 0:
+            why_not_fast_path = "self.num_heads is not even"
+        elif not self.batch_first:
+            why_not_fast_path = "batch_first was not True"
+        elif self.bias_k is not None:
+            why_not_fast_path = "self.bias_k was not None"
+        elif self.bias_v is not None:
+            why_not_fast_path = "self.bias_v was not None"
+        elif self.add_zero_attn:
+            why_not_fast_path = "add_zero_attn was enabled"
+        elif not self._qkv_same_embed_dim:
+            why_not_fast_path = "_qkv_same_embed_dim was not True"
+        elif query.is_nested and (key_padding_mask is not None or attn_mask is not None):
+            why_not_fast_path = "supplying both src_key_padding_mask and src_mask at the same time \
+                                 is not supported with NestedTensor input"
+        elif torch.is_autocast_enabled():
+            why_not_fast_path = "autocast is enabled"
+
+        if not why_not_fast_path:
+            tensor_args = (
+                query,
+                key,
+                value,
+                self.in_proj_weight,
+                self.in_proj_bias,
+                self.out_proj.weight,
+                self.out_proj.bias,
+            )
+            # We have to use list comprehensions below because TorchScript does not support
+            # generator expressions.
+            if torch.overrides.has_torch_function(tensor_args):
+                why_not_fast_path = "some Tensor argument has_torch_function"
+            elif _is_make_fx_tracing():
+                why_not_fast_path = "we are running make_fx tracing"
+            elif not all(_check_arg_device(x) for x in tensor_args):
+                why_not_fast_path = ("some Tensor argument's device is neither one of "
+                                     f"cpu, cuda or {torch.utils.backend_registration._privateuse1_backend_name}")
+            elif torch.is_grad_enabled() and any(_arg_requires_grad(x) for x in tensor_args):
+                why_not_fast_path = ("grad is enabled and at least one of query or the "
+                                     "input/output projection weights or biases requires_grad")
+            if not why_not_fast_path:
+                merged_mask, mask_type = self.merge_masks(attn_mask, key_padding_mask, query)
+
+                if self.in_proj_bias is not None and self.in_proj_weight is not None:
+                    return torch._native_multi_head_attention(
+                        query,
+                        key,
+                        value,
+                        self.embed_dim,
+                        self.num_heads,
+                        self.in_proj_weight,
+                        self.in_proj_bias,
+                        self.out_proj.weight,
+                        self.out_proj.bias,
+                        merged_mask,
+                        need_weights,
+                        average_attn_weights,
+                        mask_type)
+
+        any_nested = query.is_nested or key.is_nested or value.is_nested
+        assert not any_nested, ("MultiheadAttention does not support NestedTensor outside of its fast path. " +
+                                f"The fast path was not hit because {why_not_fast_path}")
+
+        if self.batch_first and is_batched:
+            # make sure that the transpose op does not affect the "is" property
+            if key is value:
+                if query is key:
+                    query = key = value = query.transpose(1, 0)
+                else:
+                    query, key = (x.transpose(1, 0) for x in (query, key))
+                    value = key
+            else:
+                query, key, value = (x.transpose(1, 0) for x in (query, key, value))
+        
+        if not self._qkv_same_embed_dim:
+            attn_output, attn_output_weights = self.multi_head_attention_forward(
+                query, key, value, self.embed_dim, self.num_heads,
+                self.in_proj_weight, self.in_proj_bias,
+                self.bias_k, self.bias_v, self.add_zero_attn,
+                self.dropout, self.out_proj.weight, self.out_proj.bias,
+                training=self.training,
+                key_padding_mask=key_padding_mask, need_weights=need_weights,
+                attn_mask=attn_mask,
+                use_separate_proj_weight=True,
+                q_proj_weight=self.q_proj_weight, k_proj_weight=self.k_proj_weight,
+                v_proj_weight=self.v_proj_weight,
+                average_attn_weights=average_attn_weights,
+                is_causal=is_causal)
+        else:
+            attn_output, attn_output_weights = self.multi_head_attention_forward(
+                query, key, value, self.embed_dim, self.num_heads,
+                self.in_proj_weight, self.in_proj_bias,
+                self.bias_k, self.bias_v, self.add_zero_attn,
+                self.dropout, self.out_proj.weight, self.out_proj.bias,
+                training=self.training,
+                key_padding_mask=key_padding_mask,
+                need_weights=need_weights,
+                attn_mask=attn_mask,
+                average_attn_weights=average_attn_weights,
+                is_causal=is_causal)
+        if self.batch_first and is_batched:
+            return attn_output.transpose(1, 0), attn_output_weights
+        else:
+            return attn_output, attn_output_weights
+            
+    def multi_head_attention_forward(
+        self,
+        query: Tensor,
+        key: Tensor,
+        value: Tensor,
+        embed_dim_to_check: int,
+        num_heads: int,
+        in_proj_weight: Optional[Tensor],
+        in_proj_bias: Optional[Tensor],
+        bias_k: Optional[Tensor],
+        bias_v: Optional[Tensor],
+        add_zero_attn: bool,
+        dropout_p: float,
+        out_proj_weight: Tensor,
+        out_proj_bias: Optional[Tensor],
+        training: bool = True,
+        key_padding_mask: Optional[Tensor] = None,
+        need_weights: bool = True,
+        attn_mask: Optional[Tensor] = None,
+        use_separate_proj_weight: bool = False,
+        q_proj_weight: Optional[Tensor] = None,
+        k_proj_weight: Optional[Tensor] = None,
+        v_proj_weight: Optional[Tensor] = None,
+        static_k: Optional[Tensor] = None,
+        static_v: Optional[Tensor] = None,
+        average_attn_weights: bool = True,
+        is_causal: bool = False,
+    ) -> Tuple[Tensor, Optional[Tensor]]:
+        tens_ops = (query, key, value, in_proj_weight, in_proj_bias, bias_k, bias_v, out_proj_weight, out_proj_bias)
+        if has_torch_function(tens_ops):
+            return handle_torch_function(
+                multi_head_attention_forward,
+                tens_ops,
+                query,
+                key,
+                value,
+                embed_dim_to_check,
+                num_heads,
+                in_proj_weight,
+                in_proj_bias,
+                bias_k,
+                bias_v,
+                add_zero_attn,
+                dropout_p,
+                out_proj_weight,
+                out_proj_bias,
+                training=training,
+                key_padding_mask=key_padding_mask,
+                need_weights=need_weights,
+                attn_mask=attn_mask,
+                is_causal=is_causal,
+                use_separate_proj_weight=use_separate_proj_weight,
+                q_proj_weight=q_proj_weight,
+                k_proj_weight=k_proj_weight,
+                v_proj_weight=v_proj_weight,
+                static_k=static_k,
+                static_v=static_v,
+                average_attn_weights=average_attn_weights,
+            )
+    
+        is_batched = _mha_shape_check(query, key, value, key_padding_mask, attn_mask, num_heads)
+    
+        # For unbatched input, we unsqueeze at the expected batch-dim to pretend that the input
+        # is batched, run the computation and before returning squeeze the
+        # batch dimension so that the output doesn't carry this temporary batch dimension.
+        if not is_batched:
+            # unsqueeze if the input is unbatched
+            query = query.unsqueeze(1)
+            key = key.unsqueeze(1)
+            value = value.unsqueeze(1)
+            if key_padding_mask is not None:
+                key_padding_mask = key_padding_mask.unsqueeze(0)
+    
+        # set up shape vars
+        tgt_len, bsz, embed_dim = query.shape
+        src_len, _, _ = key.shape
+    
+        key_padding_mask = _canonical_mask(
+            mask=key_padding_mask,
+            mask_name="key_padding_mask",
+            other_type=_none_or_dtype(attn_mask),
+            other_name="attn_mask",
+            target_type=query.dtype
+        )
+    
+        if is_causal and attn_mask is None:
+            raise RuntimeError(
+                "Need attn_mask if specifying the is_causal hint. "
+                "You may use the Transformer module method "
+                "`generate_square_subsequent_mask` to create this mask."
+            )
+    
+        if is_causal and key_padding_mask is None and not need_weights:
+            # when we have a kpm or need weights, we need attn_mask
+            # Otherwise, we use the is_causal hint go as is_causal
+            # indicator to SDPA.
+            attn_mask = None
+        else:
+            attn_mask = _canonical_mask(
+                mask=attn_mask,
+                mask_name="attn_mask",
+                other_type=None,
+                other_name="",
+                target_type=query.dtype,
+                check_other=False,
+            )
+    
+            if key_padding_mask is not None:
+                # We have the attn_mask, and use that to merge kpm into it.
+                # Turn off use of is_causal hint, as the merged mask is no
+                # longer causal.
+                is_causal = False
+    
+        assert embed_dim == embed_dim_to_check, \
+            f"was expecting embedding dimension of {embed_dim_to_check}, but got {embed_dim}"
+        if isinstance(embed_dim, torch.Tensor):
+            # embed_dim can be a tensor when JIT tracing
+            head_dim = embed_dim.div(num_heads, rounding_mode='trunc')
+        else:
+            head_dim = embed_dim // num_heads
+        assert head_dim * num_heads == embed_dim, f"embed_dim {embed_dim} not divisible by num_heads {num_heads}"
+        if use_separate_proj_weight:
+            # allow MHA to have different embedding dimensions when separate projection weights are used
+            assert key.shape[:2] == value.shape[:2], \
+                f"key's sequence and batch dims {key.shape[:2]} do not match value's {value.shape[:2]}"
+        else:
+            assert key.shape == value.shape, f"key shape {key.shape} does not match value shape {value.shape}"
+    
+        #
+        # compute in-projection
+        #
+        if not use_separate_proj_weight:
+            assert in_proj_weight is not None, "use_separate_proj_weight is False but in_proj_weight is None"
+            q, k, v = _in_projection_packed(query, key, value, in_proj_weight, in_proj_bias)
+        else:
+            assert q_proj_weight is not None, "use_separate_proj_weight is True but q_proj_weight is None"
+            assert k_proj_weight is not None, "use_separate_proj_weight is True but k_proj_weight is None"
+            assert v_proj_weight is not None, "use_separate_proj_weight is True but v_proj_weight is None"
+            if in_proj_bias is None:
+                b_q = b_k = b_v = None
+            else:
+                b_q, b_k, b_v = in_proj_bias.chunk(3)
+            q, k, v = _in_projection(query, key, value, q_proj_weight, k_proj_weight, v_proj_weight, b_q, b_k, b_v)
+    
+        # prep attention mask
+    
+        if attn_mask is not None:
+            # ensure attn_mask's dim is 3
+            if attn_mask.dim() == 2:
+                correct_2d_size = (tgt_len, src_len)
+                if attn_mask.shape != correct_2d_size:
+                    raise RuntimeError(f"The shape of the 2D attn_mask is {attn_mask.shape}, but should be {correct_2d_size}.")
+                attn_mask = attn_mask.unsqueeze(0)
+            elif attn_mask.dim() == 3:
+                correct_3d_size = (bsz * num_heads, tgt_len, src_len)
+                if attn_mask.shape != correct_3d_size:
+                    raise RuntimeError(f"The shape of the 3D attn_mask is {attn_mask.shape}, but should be {correct_3d_size}.")
+            else:
+                raise RuntimeError(f"attn_mask's dimension {attn_mask.dim()} is not supported")
+    
+        # add bias along batch dimension (currently second)
+        if bias_k is not None and bias_v is not None:
+            assert static_k is None, "bias cannot be added to static key."
+            assert static_v is None, "bias cannot be added to static value."
+            k = torch.cat([k, bias_k.repeat(1, bsz, 1)])
+            v = torch.cat([v, bias_v.repeat(1, bsz, 1)])
+            if attn_mask is not None:
+                attn_mask = pad(attn_mask, (0, 1))
+            if key_padding_mask is not None:
+                key_padding_mask = pad(key_padding_mask, (0, 1))
+        else:
+            assert bias_k is None
+            assert bias_v is None
+    
+        #
+        # reshape q, k, v for multihead attention and make em batch first
+        #
+        q = q.view(tgt_len, bsz * num_heads, head_dim).transpose(0, 1)
+        if static_k is None:
+            k = k.view(k.shape[0], bsz * num_heads, head_dim).transpose(0, 1)
+        else:
+            # TODO finish disentangling control flow so we don't do in-projections when statics are passed
+            assert static_k.size(0) == bsz * num_heads, \
+                f"expecting static_k.size(0) of {bsz * num_heads}, but got {static_k.size(0)}"
+            assert static_k.size(2) == head_dim, \
+                f"expecting static_k.size(2) of {head_dim}, but got {static_k.size(2)}"
+            k = static_k
+        if static_v is None:
+            v = v.view(v.shape[0], bsz * num_heads, head_dim).transpose(0, 1)
+        else:
+            # TODO finish disentangling control flow so we don't do in-projections when statics are passed
+            assert static_v.size(0) == bsz * num_heads, \
+                f"expecting static_v.size(0) of {bsz * num_heads}, but got {static_v.size(0)}"
+            assert static_v.size(2) == head_dim, \
+                f"expecting static_v.size(2) of {head_dim}, but got {static_v.size(2)}"
+            v = static_v
+    
+        # add zero attention along batch dimension (now first)
+        if add_zero_attn:
+            zero_attn_shape = (bsz * num_heads, 1, head_dim)
+            k = torch.cat([k, torch.zeros(zero_attn_shape, dtype=k.dtype, device=k.device)], dim=1)
+            v = torch.cat([v, torch.zeros(zero_attn_shape, dtype=v.dtype, device=v.device)], dim=1)
+            if attn_mask is not None:
+                attn_mask = pad(attn_mask, (0, 1))
+            if key_padding_mask is not None:
+                key_padding_mask = pad(key_padding_mask, (0, 1))
+    
+        # update source sequence length after adjustments
+        src_len = k.size(1)
+    
+        # merge key padding and attention masks
+        if key_padding_mask is not None:
+            assert key_padding_mask.shape == (bsz, src_len), \
+                f"expecting key_padding_mask shape of {(bsz, src_len)}, but got {key_padding_mask.shape}"
+            key_padding_mask = key_padding_mask.view(bsz, 1, 1, src_len).   \
+                expand(-1, num_heads, -1, -1).reshape(bsz * num_heads, 1, src_len)
+            if attn_mask is None:
+                attn_mask = key_padding_mask
+            else:
+                attn_mask = attn_mask + key_padding_mask
+    
+        # adjust dropout probability
+        if not training:
+            dropout_p = 0.0
+    
+        #
+        # (deep breath) calculate attention and out projection
+        #
+    
+        if need_weights:
+            B, Nt, E = q.shape
+            q_scaled = q / math.sqrt(E)
+    
+            assert not (is_causal and attn_mask is None), "FIXME: is_causal not implemented for need_weights"
+    
+            if attn_mask is not None:
+                attn_output_weights = torch.baddbmm(attn_mask, q_scaled, k.transpose(-2, -1))
+            else:
+                attn_output_weights = torch.bmm(q_scaled, k.transpose(-2, -1))
+            attn_output_weights = softmax(attn_output_weights, dim=-1)
+            if dropout_p > 0.0:
+                attn_output_weights = dropout(attn_output_weights, p=dropout_p)
+    
+            attn_output = torch.bmm(attn_output_weights, v)
+    
+            attn_output = attn_output.transpose(0, 1).contiguous().view(tgt_len * bsz, embed_dim)
+            attn_output = self.out_proj(attn_output)
+            attn_output = attn_output.view(tgt_len, bsz, attn_output.size(1))
+    
+            # optionally average attention weights over heads
+            attn_output_weights = attn_output_weights.view(bsz, num_heads, tgt_len, src_len)
+            if average_attn_weights:
+                attn_output_weights = attn_output_weights.mean(dim=1)
+    
+            if not is_batched:
+                # squeeze the output if input was unbatched
+                attn_output = attn_output.squeeze(1)
+                attn_output_weights = attn_output_weights.squeeze(0)
+            return attn_output, attn_output_weights
+        else:
+            # attn_mask can be either (L,S) or (N*num_heads, L, S)
+            # if attn_mask's shape is (1, L, S) we need to unsqueeze to (1, 1, L, S)
+            # in order to match the input for SDPA of (N, num_heads, L, S)
+            if attn_mask is not None:
+                if attn_mask.size(0) == 1 and attn_mask.dim() == 3:
+                    attn_mask = attn_mask.unsqueeze(0)
+                else:
+                    attn_mask = attn_mask.view(bsz, num_heads, -1, src_len)
+    
+            q = q.view(bsz, num_heads, tgt_len, head_dim)
+            k = k.view(bsz, num_heads, src_len, head_dim)
+            v = v.view(bsz, num_heads, src_len, head_dim)
+    
+            attn_output = F.scaled_dot_product_attention(q, k, v, attn_mask, dropout_p, is_causal)
+            attn_output = attn_output.permute(2, 0, 1, 3).contiguous().view(bsz * tgt_len, embed_dim)
+    
+            attn_output = self.out_proj(attn_output)
+            attn_output = attn_output.view(tgt_len, bsz, attn_output.size(1))
+            if not is_batched:
+                # squeeze the output if input was unbatched
+                attn_output = attn_output.squeeze(1)
+            return attn_output, None
+
+
+def _mha_shape_check(query: Tensor, key: Tensor, value: Tensor,
+                     key_padding_mask: Optional[Tensor], attn_mask: Optional[Tensor], num_heads: int):
+    # Verifies the expected shape for `query, `key`, `value`, `key_padding_mask` and `attn_mask`
+    # and returns if the input is batched or not.
+    # Raises an error if `query` is not 2-D (unbatched) or 3-D (batched) tensor.
+
+    # Shape check.
+    if query.dim() == 3:
+        # Batched Inputs
+        is_batched = True
+        assert key.dim() == 3 and value.dim() == 3, \
+            ("For batched (3-D) `query`, expected `key` and `value` to be 3-D"
+             f" but found {key.dim()}-D and {value.dim()}-D tensors respectively")
+        if key_padding_mask is not None:
+            assert key_padding_mask.dim() == 2, \
+                ("For batched (3-D) `query`, expected `key_padding_mask` to be `None` or 2-D"
+                 f" but found {key_padding_mask.dim()}-D tensor instead")
+        if attn_mask is not None:
+            assert attn_mask.dim() in (2, 3), \
+                ("For batched (3-D) `query`, expected `attn_mask` to be `None`, 2-D or 3-D"
+                 f" but found {attn_mask.dim()}-D tensor instead")
+    elif query.dim() == 2:
+        # Unbatched Inputs
+        is_batched = False
+        assert key.dim() == 2 and value.dim() == 2, \
+            ("For unbatched (2-D) `query`, expected `key` and `value` to be 2-D"
+             f" but found {key.dim()}-D and {value.dim()}-D tensors respectively")
+
+        if key_padding_mask is not None:
+            assert key_padding_mask.dim() == 1, \
+                ("For unbatched (2-D) `query`, expected `key_padding_mask` to be `None` or 1-D"
+                 f" but found {key_padding_mask.dim()}-D tensor instead")
+
+        if attn_mask is not None:
+            assert attn_mask.dim() in (2, 3), \
+                ("For unbatched (2-D) `query`, expected `attn_mask` to be `None`, 2-D or 3-D"
+                 f" but found {attn_mask.dim()}-D tensor instead")
+            if attn_mask.dim() == 3:
+                expected_shape = (num_heads, query.shape[0], key.shape[0])
+                assert attn_mask.shape == expected_shape, \
+                    (f"Expected `attn_mask` shape to be {expected_shape} but got {attn_mask.shape}")
+    else:
+        raise AssertionError(
+            f"query should be unbatched 2D or batched 3D tensor but received {query.dim()}-D query tensor")
+
+    return is_batched
+
+
+def _canonical_mask(
+        mask: Optional[Tensor],
+        mask_name: str,
+        other_type: Optional[DType],
+        other_name: str,
+        target_type: DType,
+        check_other: bool = True,
+) -> Optional[Tensor]:
+
+    if mask is not None:
+        _mask_dtype = mask.dtype
+        _mask_is_float = torch.is_floating_point(mask)
+        if _mask_dtype != torch.bool and not _mask_is_float:
+            raise AssertionError(
+                f"only bool and floating types of {mask_name} are supported")
+        if check_other and other_type is not None:
+            if _mask_dtype != other_type:
+                warnings.warn(
+                    f"Support for mismatched {mask_name} and {other_name} "
+                    "is deprecated. Use same type for both instead."
+                )
+        if not _mask_is_float:
+            mask = (
+                torch.zeros_like(mask, dtype=target_type)
+                .masked_fill_(mask, float("-inf"))
+            )
+    return mask
+
+
+def _none_or_dtype(input: Optional[Tensor]) -> Optional[DType]:
+    if input is None:
+        return None
+    elif isinstance(input, torch.Tensor):
+        return input.dtype
+    raise RuntimeError("input to _none_or_dtype() must be None or torch.Tensor")
+
+def _in_projection_packed(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    w: Tensor,
+    b: Optional[Tensor] = None,
+) -> List[Tensor]:
+    r"""
+    Performs the in-projection step of the attention operation, using packed weights.
+    Output is a triple containing projection tensors for query, key and value.
+
+    Args:
+        q, k, v: query, key and value tensors to be projected. For self-attention,
+            these are typically the same tensor; for encoder-decoder attention,
+            k and v are typically the same tensor. (We take advantage of these
+            identities for performance if they are present.) Regardless, q, k and v
+            must share a common embedding dimension; otherwise their shapes may vary.
+        w: projection weights for q, k and v, packed into a single tensor. Weights
+            are packed along dimension 0, in q, k, v order.
+        b: optional projection biases for q, k and v, packed into a single tensor
+            in q, k, v order.
+
+    Shape:
+        Inputs:
+        - q: :math:`(..., E)` where E is the embedding dimension
+        - k: :math:`(..., E)` where E is the embedding dimension
+        - v: :math:`(..., E)` where E is the embedding dimension
+        - w: :math:`(E * 3, E)` where E is the embedding dimension
+        - b: :math:`E * 3` where E is the embedding dimension
+
+        Output:
+        - in output list :math:`[q', k', v']`, each output tensor will have the
+            same shape as the corresponding input tensor.
+    """
+    E = q.size(-1)
+    if k is v:
+        if q is k:
+            # self-attention
+            proj = linear(q, w, b)
+            # reshape to 3, E and not E, 3 is deliberate for better memory coalescing and keeping same order as chunk()
+            proj = proj.unflatten(-1, (3, E)).unsqueeze(0).transpose(0, -2).squeeze(-2).contiguous()
+            return proj[0], proj[1], proj[2]
+        else:
+            # encoder-decoder attention
+            w_q, w_kv = w.split([E, E * 2])
+            if b is None:
+                b_q = b_kv = None
+            else:
+                b_q, b_kv = b.split([E, E * 2])
+            q_proj = linear(q, w_q, b_q)
+            kv_proj = linear(k, w_kv, b_kv)
+            # reshape to 2, E and not E, 2 is deliberate for better memory coalescing and keeping same order as chunk()
+            kv_proj = kv_proj.unflatten(-1, (2, E)).unsqueeze(0).transpose(0, -2).squeeze(-2).contiguous()
+            return (q_proj, kv_proj[0], kv_proj[1])
+    else:
+        w_q, w_k, w_v = w.chunk(3)
+        if b is None:
+            b_q = b_k = b_v = None
+        else:
+            b_q, b_k, b_v = b.chunk(3)
+        return linear(q, w_q, b_q), linear(k, w_k, b_k), linear(v, w_v, b_v)
+
+
+def _in_projection(
+    q: Tensor,
+    k: Tensor,
+    v: Tensor,
+    w_q: Tensor,
+    w_k: Tensor,
+    w_v: Tensor,
+    b_q: Optional[Tensor] = None,
+    b_k: Optional[Tensor] = None,
+    b_v: Optional[Tensor] = None,
+) -> Tuple[Tensor, Tensor, Tensor]:
+    r"""
+    Performs the in-projection step of the attention operation. This is simply
+    a triple of linear projections, with shape constraints on the weights which
+    ensure embedding dimension uniformity in the projected outputs.
+    Output is a triple containing projection tensors for query, key and value.
+
+    Args:
+        q, k, v: query, key and value tensors to be projected.
+        w_q, w_k, w_v: weights for q, k and v, respectively.
+        b_q, b_k, b_v: optional biases for q, k and v, respectively.
+
+    Shape:
+        Inputs:
+        - q: :math:`(Qdims..., Eq)` where Eq is the query embedding dimension and Qdims are any
+            number of leading dimensions.
+        - k: :math:`(Kdims..., Ek)` where Ek is the key embedding dimension and Kdims are any
+            number of leading dimensions.
+        - v: :math:`(Vdims..., Ev)` where Ev is the value embedding dimension and Vdims are any
+            number of leading dimensions.
+        - w_q: :math:`(Eq, Eq)`
+        - w_k: :math:`(Eq, Ek)`
+        - w_v: :math:`(Eq, Ev)`
+        - b_q: :math:`(Eq)`
+        - b_k: :math:`(Eq)`
+        - b_v: :math:`(Eq)`
+
+        Output: in output triple :math:`(q', k', v')`,
+         - q': :math:`[Qdims..., Eq]`
+         - k': :math:`[Kdims..., Eq]`
+         - v': :math:`[Vdims..., Eq]`
+
+    """
+    Eq, Ek, Ev = q.size(-1), k.size(-1), v.size(-1)
+    assert w_q.shape == (Eq, Eq), f"expecting query weights shape of {(Eq, Eq)}, but got {w_q.shape}"
+    assert w_k.shape == (Eq, Ek), f"expecting key weights shape of {(Eq, Ek)}, but got {w_k.shape}"
+    assert w_v.shape == (Eq, Ev), f"expecting value weights shape of {(Eq, Ev)}, but got {w_v.shape}"
+    assert b_q is None or b_q.shape == (Eq,), f"expecting query bias shape of {(Eq,)}, but got {b_q.shape}"
+    assert b_k is None or b_k.shape == (Eq,), f"expecting key bias shape of {(Eq,)}, but got {b_k.shape}"
+    assert b_v is None or b_v.shape == (Eq,), f"expecting value bias shape of {(Eq,)}, but got {b_v.shape}"
+    return linear(q, w_q, b_q), linear(k, w_k, b_k), linear(v, w_v, b_v)