add ability to set flash = True on RIN

Author: lucidrains

README.md

@@ -149,3 +149,12 @@ sampled_images.shape # (4, 3, 128, 128)
     year    = {2023}
 }
 ```
+
+```bibtex
+@inproceedings{dao2022flashattention,
+    title   = {Flash{A}ttention: Fast and Memory-Efficient Exact Attention with {IO}-Awareness},
+    author  = {Dao, Tri and Fu, Daniel Y. and Ermon, Stefano and Rudra, Atri and R{\'e}, Christopher},
+    booktitle = {Advances in Neural Information Processing Systems},
+    year    = {2022}
+}
+```

rin_pytorch/attend.py

@@ -0,0 +1,126 @@
+from functools import wraps
+from packaging import version
+from collections import namedtuple
+
+import torch
+from torch import nn, einsum
+import torch.nn.functional as F
+
+from einops import rearrange, reduce
+
+# constants
+
+FlashAttentionConfig = namedtuple('FlashAttentionConfig', ['enable_flash', 'enable_math', 'enable_mem_efficient'])
+
+# helpers
+
+def exists(val):
+    return val is not None
+
+def once(fn):
+    called = False
+    @wraps(fn)
+    def inner(x):
+        nonlocal called
+        if called:
+            return
+        called = True
+        return fn(x)
+    return inner
+
+print_once = once(print)
+
+# main class
+
+class Attend(nn.Module):
+    def __init__(
+        self,
+        dropout = 0.,
+        flash = False
+    ):
+        super().__init__()
+        self.dropout = dropout
+        self.attn_dropout = nn.Dropout(dropout)
+
+        self.flash = flash
+        assert not (flash and version.parse(torch.__version__) < version.parse('2.0.0')), 'in order to use flash attention, you must be using pytorch 2.0 or above'
+
+        # determine efficient attention configs for cuda and cpu
+
+        self.cpu_config = FlashAttentionConfig(True, True, True)
+        self.cuda_config = None
+
+        if not torch.cuda.is_available() or not flash:
+            return
+
+        device_properties = torch.cuda.get_device_properties(torch.device('cuda'))
+
+        if device_properties.major == 8 and device_properties.minor == 0:
+            print_once('A100 GPU detected, using flash attention if input tensor is on cuda')
+            self.cuda_config = FlashAttentionConfig(True, False, False)
+        else:
+            print_once('Non-A100 GPU detected, using math or mem efficient attention if input tensor is on cuda')
+            self.cuda_config = FlashAttentionConfig(False, True, True)
+
+    def flash_attn(self, q, k, v, mask = None):
+        _, heads, q_len, _, k_len, is_cuda, device = *q.shape, k.shape[-2], q.is_cuda, q.device
+
+        # Check if mask exists and expand to compatible shape
+        # The mask is B L, so it would have to be expanded to B H N L
+
+        if exists(mask):
+            mask = mask.expand(-1, heads, q_len, -1)
+
+        # Check if there is a compatible device for flash attention
+
+        config = self.cuda_config if is_cuda else self.cpu_config
+
+        # pytorch 2.0 flash attn: q, k, v, mask, dropout, softmax_scale
+
+        with torch.backends.cuda.sdp_kernel(**config._asdict()):
+            out = F.scaled_dot_product_attention(
+                q, k, v,
+                attn_mask = mask,
+                dropout_p = self.dropout if self.training else 0.
+            )
+
+        return out
+
+    def forward(self, q, k, v, mask = None):
+        """
+        einstein notation
+        b - batch
+        h - heads
+        n, i, j - sequence length (base sequence length, source, target)
+        d - feature dimension
+        """
+
+        q_len, k_len, device = q.shape[-2], k.shape[-2], q.device
+
+        scale = q.shape[-1] ** -0.5
+
+        if exists(mask) and mask.ndim != 4:
+            mask = rearrange(mask, 'b j -> b 1 1 j')
+
+        if self.flash:
+            return self.flash_attn(q, k, v, mask = mask)
+
+        # similarity
+
+        sim = einsum(f"b h i d, b h j d -> b h i j", q, k) * scale
+
+        # key padding mask
+
+        if exists(mask):
+            sim = sim.masked_fill(~mask, -torch.finfo(sim.dtype).max)
+
+        # attention
+
+        attn = sim.softmax(dim=-1)
+        attn = self.attn_dropout(attn)
+
+        # aggregate values
+
+        out = einsum(f"b h i j, b h j d -> b h i d", attn, v)
+
+        return out

rin_pytorch/rin_pytorch.py

@@ -18,6 +18,8 @@
 from einops import rearrange, reduce, repeat
 from einops.layers.torch import Rearrange
 
+from rin_pytorch.attend import Attend
+
 from PIL import Image
 from tqdm.auto import tqdm
 from ema_pytorch import EMA
@@ -166,7 +168,8 @@ def __init__(
         dim_head = 32,
         norm = False,
         norm_context = False,
-        time_cond_dim = None
+        time_cond_dim = None,
+        flash = False
     ):
         super().__init__()
         hidden_dim = dim_head * heads
@@ -194,6 +197,8 @@ def __init__(
         self.to_kv = nn.Linear(dim_context, hidden_dim * 2, bias = False)
         self.to_out = nn.Linear(hidden_dim, dim, bias = False)
 
+        self.attend = Attend(flash = flash)
+
     def forward(
         self,
         x,
@@ -217,12 +222,8 @@ def forward(
         qkv = (self.to_q(x), *self.to_kv(context).chunk(2, dim = -1))
         q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b h n d', h = h), qkv)
 
-        q = q * self.scale
-
-        sim = einsum('b h i d, b h j d -> b h i j', q, k)
-        attn = sim.softmax(dim = -1)
+        out = self.attend(q, k, v)
 
-        out = einsum('b h i j, b h j d -> b h i d', attn, v)
         out = rearrange(out, 'b h n d -> b n (h d)')
         return self.to_out(out)
 

setup.py

@@ -3,7 +3,7 @@
 setup(
   name = 'RIN-pytorch',
   packages = find_packages(exclude=[]),
-  version = '0.7.4',
+  version = '0.7.5',
   license='MIT',
   description = 'RIN - Recurrent Interface Network - Pytorch',
   author = 'Phil Wang',