add tests, fix bugs

Author: lucidrains

memory_efficient_attention_pytorch/__init__.py

@@ -1,2 +1,3 @@
 from memory_efficient_attention_pytorch.memory_efficient_attention import Attention, memory_efficient_attention
 from memory_efficient_attention_pytorch.memory_efficient_cosine_sim_attention import CosineSimAttention, numerically_unstable_memory_efficient_attention
+from memory_efficient_attention_pytorch.flash_attention import FlashAttention

memory_efficient_attention_pytorch/cosine_sim_flash_attention.py

@@ -42,7 +42,8 @@ def forward(ctx, q, k, v, mask, scale, causal, q_bucket_size, k_bucket_size):
         if not exists(mask):
             mask = (None,) * math.ceil(q.shape[-2] / q_bucket_size)
         else:
-            mask = mask.split(q_bucket_size, dim = -2)
+            mask = rearrange(mask, 'b n -> b 1 1 n')
+            mask = mask.split(q_bucket_size, dim = -1)
 
         row_splits = zip(
             q.split(q_bucket_size, dim = -2),
@@ -184,7 +185,7 @@ def __init__(
 
         self.to_q = nn.Linear(dim, inner_dim, bias = False)
         self.to_kv = nn.Linear(dim, inner_dim * 2, bias = False)
-        self.to_out = nn.Linear(inner_dim, dim)
+        self.to_out = nn.Linear(inner_dim, dim, bias = False)
 
         # memory efficient attention related parameters
         # can be overriden on forward

memory_efficient_attention_pytorch/flash_attention.py

@@ -37,12 +37,12 @@ def forward(ctx, q, k, v, mask, causal, q_bucket_size, k_bucket_size):
         all_row_maxes = torch.full((*q.shape[:-1], 1), max_neg_value, device = device)
 
         scale = (q.shape[-1] ** -0.5)
-        q = q * scale
 
         if not exists(mask):
             mask = (None,) * math.ceil(q.shape[-2] / q_bucket_size)
         else:
-            mask = mask.split(q_bucket_size, dim = -2)
+            mask = rearrange(mask, 'b n -> b 1 1 n')
+            mask = mask.split(q_bucket_size, dim = -1)
 
         row_splits = zip(
             q.split(q_bucket_size, dim = -2),
@@ -63,7 +63,7 @@ def forward(ctx, q, k, v, mask, causal, q_bucket_size, k_bucket_size):
             for k_ind, (kc, vc) in enumerate(col_splits):
                 k_start_index = k_ind * k_bucket_size
 
-                attn_weights = einsum('... i d, ... j d -> ... i j', qc, kc)
+                attn_weights = einsum('... i d, ... j d -> ... i j', qc, kc) * scale
 
                 if exists(row_mask):
                     attn_weights.masked_fill_(~row_mask, max_neg_value)
@@ -73,7 +73,6 @@ def forward(ctx, q, k, v, mask, causal, q_bucket_size, k_bucket_size):
                     attn_weights.masked_fill_(causal_mask, max_neg_value)
 
                 block_row_maxes = attn_weights.amax(dim = -1, keepdims = True)
-
                 attn_weights -= block_row_maxes
                 exp_weights = torch.exp(attn_weights)
 
@@ -82,7 +81,7 @@ def forward(ctx, q, k, v, mask, causal, q_bucket_size, k_bucket_size):
 
                 block_row_sums = exp_weights.sum(dim = -1, keepdims = True).clamp(min = EPSILON)
 
-                new_row_maxes = torch.maximum(block_row_maxes, row_sums)
+                new_row_maxes = torch.maximum(block_row_maxes, row_maxes)
 
                 exp_values = einsum('... i j, ... j d -> ... i d', exp_weights, vc)
 
@@ -92,10 +91,11 @@ def forward(ctx, q, k, v, mask, causal, q_bucket_size, k_bucket_size):
                 new_row_sums = exp_row_max_diff * row_sums + exp_block_row_max_diff * block_row_sums
 
                 oc.mul_((row_sums / new_row_sums) * exp_row_max_diff).add_((exp_block_row_max_diff / new_row_sums) * exp_values)
+
                 row_maxes.copy_(new_row_maxes)
                 row_sums.copy_(new_row_sums)
 
-        ctx.args = (causal, mask, q_bucket_size, k_bucket_size)
+        ctx.args = (causal, scale, mask, q_bucket_size, k_bucket_size)
         ctx.save_for_backward(q, k, v, o, all_row_sums, all_row_maxes)
 
         return o
@@ -105,7 +105,7 @@ def forward(ctx, q, k, v, mask, causal, q_bucket_size, k_bucket_size):
     def backward(ctx, do):
         """ Algorithm 4 in the paper """
 
-        causal, mask, q_bucket_size, k_bucket_size = ctx.args
+        causal, scale, mask, q_bucket_size, k_bucket_size = ctx.args
         q, k, v, o, l, m = ctx.saved_tensors
 
         device = q.device
@@ -117,8 +117,6 @@ def backward(ctx, do):
         dk = torch.zeros_like(k)
         dv = torch.zeros_like(v)
 
-        scale = q.shape[-1] ** -0.5
-
         row_splits = zip(
             q.split(q_bucket_size, dim = -2),
             o.split(q_bucket_size, dim = -2),
@@ -142,8 +140,7 @@ def backward(ctx, do):
             for k_ind, (kc, vc, dkc, dvc) in enumerate(col_splits):
                 k_start_index = k_ind * k_bucket_size
 
-                qc_scaled = qc * scale
-                attn_weights = einsum('... i d, ... j d -> ... i j', qc_scaled, kc)
+                attn_weights = einsum('... i d, ... j d -> ... i j', qc, kc) * scale
 
                 if causal and q_start_index < (k_start_index + k_bucket_size - 1):
                     causal_mask = torch.ones((qc.shape[-2], kc.shape[-2]), dtype = torch.bool, device = device).triu(q_start_index - k_start_index + 1)
@@ -197,7 +194,7 @@ def __init__(
 
         self.to_q = nn.Linear(dim, inner_dim, bias = False)
         self.to_kv = nn.Linear(dim, inner_dim * 2, bias = False)
-        self.to_out = nn.Linear(inner_dim, dim)
+        self.to_out = nn.Linear(inner_dim, dim, bias = False)
 
         # memory efficient attention related parameters
         # can be overriden on forward

memory_efficient_attention_pytorch/memory_efficient_attention.py

@@ -180,7 +180,7 @@ def __init__(
 
         self.to_q = nn.Linear(dim, inner_dim, bias = False)
         self.to_kv = nn.Linear(dim, inner_dim * 2, bias = False)
-        self.to_out = nn.Linear(inner_dim, dim)
+        self.to_out = nn.Linear(inner_dim, dim, bias = False)
 
         # memory efficient attention related parameters
         # can be overriden on forward

memory_efficient_attention_pytorch/memory_efficient_cosine_sim_attention.py

@@ -164,7 +164,7 @@ def __init__(
 
         self.to_q = nn.Linear(dim, inner_dim, bias = False)
         self.to_kv = nn.Linear(dim, inner_dim * 2, bias = False)
-        self.to_out = nn.Linear(inner_dim, dim)
+        self.to_out = nn.Linear(inner_dim, dim, bias = False)
 
         # memory efficient attention related parameters
         # can be overriden on forward

setup.py

@@ -3,7 +3,7 @@
 setup(
   name = 'memory-efficient-attention-pytorch',
   packages = find_packages(exclude=[]),
-  version = '0.0.22',
+  version = '0.0.23',
   license='MIT',
   description = 'Memory Efficient Attention - Pytorch',
   long_description_content_type = 'text/markdown',

tests/test.py

@@ -1,6 +1,9 @@
 import torch
 from memory_efficient_attention_pytorch import Attention
 
+from memory_efficient_attention_pytorch.memory_efficient_attention import attention
+from memory_efficient_attention_pytorch.flash_attention import FlashAttention, FlashAttentionFunction
+
 # constants
 
 def isclose(a, b, atol = 1e-6):
@@ -53,3 +56,59 @@ def loss_fn(inp, **kwargs):
     mem_efficient_out_grad = x.grad.clone()
 
     assert isclose(out_grad, mem_efficient_out_grad, atol = 1e-5)
+
+# test flash attention
+
+def test_flash_attn_output_equal():
+    attn_kwargs = dict(
+        dim = 512,
+        dim_head = 64,
+        heads = 8,
+        q_bucket_size = 64,
+        k_bucket_size = 64,
+        causal = True
+    )
+
+    attn = Attention(**attn_kwargs)
+    flash_attn = FlashAttention(**attn_kwargs)
+
+    flash_attn.to_q = attn.to_q
+    flash_attn.to_kv = attn.to_kv
+    flash_attn.to_out = attn.to_out
+
+    x = torch.randn(2, 2048, 512)
+    mask = torch.ones(2, 2048).bool()
+
+    out = attn(x, mask = mask)
+    mem_efficient_out = flash_attn(x, mask = mask)
+
+    assert isclose(mem_efficient_out, out, atol = 1e-6)
+
+# test gradients equal
+
+def test_flash_attn_gradients_equal():
+    q = torch.randn(1, 8, 1024, 512).requires_grad_()
+    k = torch.randn(1, 8, 1024, 512).requires_grad_()
+    v = torch.randn(1, 8, 1024, 512).requires_grad_()
+
+    o = attention(q, k, v, causal = False)
+    o.sum().backward()
+
+    dq_grad = q.grad.clone()
+    dk_grad = k.grad.clone()
+    dv_grad = v.grad.clone()
+
+    q.grad.zero_()
+    k.grad.zero_()
+    v.grad.zero_()
+
+    flash_o = FlashAttentionFunction.apply(q, k, v, None, False, 64, 64)
+    flash_o.sum().backward()
+
+    flash_dq_grad = q.grad.clone()
+    flash_dk_grad = k.grad.clone()
+    flash_dv_grad = v.grad.clone()
+
+    assert isclose(flash_dq_grad, dq_grad, atol = 1e-5)
+    assert isclose(flash_dk_grad, dk_grad, atol = 1e-5)
+    assert isclose(flash_dv_grad, dv_grad, atol = 1e-5)