Add cross-attention optimization from InvokeAI

* Add cross-attention optimization from InvokeAI (~30% speed improvement on MPS)
* Add command line option for it
* Make it default when CUDA is unavailable

modules/sd_hijack.py

@@ -30,8 +30,11 @@ def apply_optimizations():
     elif cmd_opts.opt_split_attention_v1:
         print("Applying v1 cross attention optimization.")
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_v1
+    elif not cmd_opts.disable_opt_split_attention and (cmd_opts.opt_split_attention_invokeai or not torch.cuda.is_available()):
+        print("Applying cross attention optimization (InvokeAI).")
+        ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward_invokeAI
     elif not cmd_opts.disable_opt_split_attention and (cmd_opts.opt_split_attention or torch.cuda.is_available()):
-        print("Applying cross attention optimization.")
+        print("Applying cross attention optimization (Doggettx).")
         ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.split_cross_attention_forward
         ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.cross_attention_attnblock_forward
 

modules/sd_hijack_optimizations.py

@@ -1,6 +1,7 @@
 import math
 import sys
 import traceback
+import psutil
 
 import torch
 from torch import einsum
@@ -116,6 +117,84 @@ def split_cross_attention_forward(self, x, context=None, mask=None):
 
     return self.to_out(r2)
 
+# -- From https://github.com/invoke-ai/InvokeAI/blob/main/ldm/modules/attention.py (with hypernetworks support added) --
+
+mem_total_gb = psutil.virtual_memory().total // (1 << 30)
+
+def einsum_op_compvis(q, k, v):
+    s = einsum('b i d, b j d -> b i j', q, k)
+    s = s.softmax(dim=-1, dtype=s.dtype)
+    return einsum('b i j, b j d -> b i d', s, v)
+
+def einsum_op_slice_0(q, k, v, slice_size):
+    r = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
+    for i in range(0, q.shape[0], slice_size):
+        end = i + slice_size
+        r[i:end] = einsum_op_compvis(q[i:end], k[i:end], v[i:end])
+    return r
+
+def einsum_op_slice_1(q, k, v, slice_size):
+    r = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)
+    for i in range(0, q.shape[1], slice_size):
+        end = i + slice_size
+        r[:, i:end] = einsum_op_compvis(q[:, i:end], k, v)
+    return r
+
+def einsum_op_mps_v1(q, k, v):
+    if q.shape[1] <= 4096: # (512x512) max q.shape[1]: 4096
+        return einsum_op_compvis(q, k, v)
+    else:
+        slice_size = math.floor(2**30 / (q.shape[0] * q.shape[1]))
+        return einsum_op_slice_1(q, k, v, slice_size)
+
+def einsum_op_mps_v2(q, k, v):
+    if mem_total_gb > 8 and q.shape[1] <= 4096:
+        return einsum_op_compvis(q, k, v)
+    else:
+        return einsum_op_slice_0(q, k, v, 1)
+
+def einsum_op_tensor_mem(q, k, v, max_tensor_mb):
+    size_mb = q.shape[0] * q.shape[1] * k.shape[1] * q.element_size() // (1 << 20)
+    if size_mb <= max_tensor_mb:
+        return einsum_op_compvis(q, k, v)
+    div = 1 << int((size_mb - 1) / max_tensor_mb).bit_length()
+    if div <= q.shape[0]:
+        return einsum_op_slice_0(q, k, v, q.shape[0] // div)
+    return einsum_op_slice_1(q, k, v, max(q.shape[1] // div, 1))
+
+def einsum_op(q, k, v):
+    if q.device.type == 'mps':
+        if mem_total_gb >= 32:
+            return einsum_op_mps_v1(q, k, v)
+        return einsum_op_mps_v2(q, k, v)
+
+    # Smaller slices are faster due to L2/L3/SLC caches.
+    # Tested on i7 with 8MB L3 cache.
+    return einsum_op_tensor_mem(q, k, v, 32)
+
+def split_cross_attention_forward_invokeAI(self, x, context=None, mask=None):
+    h = self.heads
+
+    q = self.to_q(x)
+    context = default(context, x)
+
+    hypernetwork = shared.loaded_hypernetwork
+    hypernetwork_layers = (hypernetwork.layers if hypernetwork is not None else {}).get(context.shape[2], None)
+
+    if hypernetwork_layers is not None:
+        k = self.to_k(hypernetwork_layers[0](context)) * self.scale
+        v = self.to_v(hypernetwork_layers[1](context))
+    else:
+        k = self.to_k(context) * self.scale
+        v = self.to_v(context)
+    del context, x
+
+    q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q, k, v))
+    r = einsum_op(q, k, v)
+    return self.to_out(rearrange(r, '(b h) n d -> b n (h d)', h=h))
+
+# -- End of code from https://github.com/invoke-ai/InvokeAI/blob/main/ldm/modules/attention.py --
+
 def xformers_attention_forward(self, x, context=None, mask=None):
     h = self.heads
     q_in = self.to_q(x)

modules/shared.py

@@ -50,9 +50,10 @@ parser.add_argument("--ldsr-models-path", type=str, help="Path to directory with
 parser.add_argument("--xformers", action='store_true', help="enable xformers for cross attention layers")
 parser.add_argument("--force-enable-xformers", action='store_true', help="enable xformers for cross attention layers regardless of whether the checking code thinks you can run it; do not make bug reports if this fails to work")
 parser.add_argument("--deepdanbooru", action='store_true', help="enable deepdanbooru interrogator")
-parser.add_argument("--opt-split-attention", action='store_true', help="force-enables cross-attention layer optimization. By default, it's on for torch.cuda and off for other torch devices.")
-parser.add_argument("--disable-opt-split-attention", action='store_true', help="force-disables cross-attention layer optimization")
+parser.add_argument("--opt-split-attention", action='store_true', help="force-enables Doggettx's cross-attention layer optimization. By default, it's on for torch cuda.")
+parser.add_argument("--opt-split-attention-invokeai", action='store_true', help="force-enables InvokeAI's cross-attention layer optimization. By default, it's on when cuda is unavailable.")
 parser.add_argument("--opt-split-attention-v1", action='store_true', help="enable older version of split attention optimization that does not consume all the VRAM it can find")
+parser.add_argument("--disable-opt-split-attention", action='store_true', help="force-disables cross-attention layer optimization")
 parser.add_argument("--use-cpu", nargs='+',choices=['SD', 'GFPGAN', 'BSRGAN', 'ESRGAN', 'SCUNet', 'CodeFormer'], help="use CPU as torch device for specified modules", default=[])
 parser.add_argument("--listen", action='store_true', help="launch gradio with 0.0.0.0 as server name, allowing to respond to network requests")
 parser.add_argument("--port", type=int, help="launch gradio with given server port, you need root/admin rights for ports < 1024, defaults to 7860 if available", default=None)