From 948533950c9db5069a874d925fadd50bac00fdb5 Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Tue, 11 Oct 2022 11:09:51 +0300
Subject: replace duplicate code with a function

---
 modules/sd_hijack_optimizations.py | 44 +++++++++++++-------------------------
 1 file changed, 15 insertions(+), 29 deletions(-)

(limited to 'modules/sd_hijack_optimizations.py')

diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py
index 18408e62..25cb67a4 100644
--- a/modules/sd_hijack_optimizations.py
+++ b/modules/sd_hijack_optimizations.py
@@ -8,7 +8,8 @@ from torch import einsum
 from ldm.util import default
 from einops import rearrange
 
-from modules import shared
+from modules import shared, hypernetwork
+
 
 if shared.cmd_opts.xformers or shared.cmd_opts.force_enable_xformers:
     try:
@@ -26,16 +27,10 @@ def split_cross_attention_forward_v1(self, x, context=None, mask=None):
     q_in = 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_in = self.to_k(hypernetwork_layers[0](context))
-        v_in = self.to_v(hypernetwork_layers[1](context))
-    else:
-        k_in = self.to_k(context)
-        v_in = self.to_v(context)
-    del context, x
+    context_k, context_v = hypernetwork.apply_hypernetwork(shared.loaded_hypernetwork, context)
+    k_in = self.to_k(context_k)
+    v_in = self.to_v(context_v)
+    del context, context_k, context_v, x
 
     q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q_in, k_in, v_in))
     del q_in, k_in, v_in
@@ -59,22 +54,16 @@ def split_cross_attention_forward_v1(self, x, context=None, mask=None):
     return self.to_out(r2)
 
 
-# taken from https://github.com/Doggettx/stable-diffusion
+# taken from https://github.com/Doggettx/stable-diffusion and modified
 def split_cross_attention_forward(self, x, context=None, mask=None):
     h = self.heads
 
     q_in = 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_in = self.to_k(hypernetwork_layers[0](context))
-        v_in = self.to_v(hypernetwork_layers[1](context))
-    else:
-        k_in = self.to_k(context)
-        v_in = self.to_v(context)
+    context_k, context_v = hypernetwork.apply_hypernetwork(shared.loaded_hypernetwork, context)
+    k_in = self.to_k(context_k)
+    v_in = self.to_v(context_v)
 
     k_in *= self.scale
 
@@ -130,14 +119,11 @@ def xformers_attention_forward(self, x, context=None, mask=None):
     h = self.heads
     q_in = 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_in = self.to_k(hypernetwork_layers[0](context))
-        v_in = self.to_v(hypernetwork_layers[1](context))
-    else:
-        k_in = self.to_k(context)
-        v_in = self.to_v(context)
+
+    context_k, context_v = hypernetwork.apply_hypernetwork(shared.loaded_hypernetwork, context)
+    k_in = self.to_k(context_k)
+    v_in = self.to_v(context_v)
+
     q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> b n h d', h=h), (q_in, k_in, v_in))
     del q_in, k_in, v_in
     out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None)
-- 
cgit v1.2.1


From 530103b586109c11fd068eb70ef09503ec6a4caf Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Tue, 11 Oct 2022 14:53:02 +0300
Subject: fixes related to merge

---
 modules/sd_hijack_optimizations.py | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

(limited to 'modules/sd_hijack_optimizations.py')

diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py
index 25cb67a4..27e571fc 100644
--- a/modules/sd_hijack_optimizations.py
+++ b/modules/sd_hijack_optimizations.py
@@ -8,7 +8,8 @@ from torch import einsum
 from ldm.util import default
 from einops import rearrange
 
-from modules import shared, hypernetwork
+from modules import shared
+from modules.hypernetwork import hypernetwork
 
 
 if shared.cmd_opts.xformers or shared.cmd_opts.force_enable_xformers:
-- 
cgit v1.2.1


From 873efeed49bb5197a42da18272115b326c5d68f3 Mon Sep 17 00:00:00 2001
From: AUTOMATIC <16777216c@gmail.com>
Date: Tue, 11 Oct 2022 15:51:22 +0300
Subject: rename hypernetwork dir to hypernetworks to prevent clash with an old
 filename that people who use zip instead of git clone will have

---
 modules/sd_hijack_optimizations.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

(limited to 'modules/sd_hijack_optimizations.py')

diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py
index 27e571fc..3349b9c3 100644
--- a/modules/sd_hijack_optimizations.py
+++ b/modules/sd_hijack_optimizations.py
@@ -9,7 +9,7 @@ from ldm.util import default
 from einops import rearrange
 
 from modules import shared
-from modules.hypernetwork import hypernetwork
+from modules.hypernetworks import hypernetwork
 
 
 if shared.cmd_opts.xformers or shared.cmd_opts.force_enable_xformers:
-- 
cgit v1.2.1


From c0484f1b986ce7acb0e3596f6089a191279f5442 Mon Sep 17 00:00:00 2001
From: brkirch <brkirch@users.noreply.github.com>
Date: Mon, 10 Oct 2022 22:48:54 -0400
Subject: 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_optimizations.py | 79 ++++++++++++++++++++++++++++++++++++++
 1 file changed, 79 insertions(+)

(limited to 'modules/sd_hijack_optimizations.py')

diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py
index 3349b9c3..870226c5 100644
--- a/modules/sd_hijack_optimizations.py
+++ b/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)
-- 
cgit v1.2.1


From 98fd5cde72d5bda1620ab78416c7828fdc3dc10b Mon Sep 17 00:00:00 2001
From: brkirch <brkirch@users.noreply.github.com>
Date: Mon, 10 Oct 2022 23:55:48 -0400
Subject: Add check for psutil

---
 modules/sd_hijack_optimizations.py | 19 +++++++++++++++----
 1 file changed, 15 insertions(+), 4 deletions(-)

(limited to 'modules/sd_hijack_optimizations.py')

diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py
index 870226c5..2a4ac7e0 100644
--- a/modules/sd_hijack_optimizations.py
+++ b/modules/sd_hijack_optimizations.py
@@ -1,7 +1,7 @@
 import math
 import sys
 import traceback
-import psutil
+import importlib
 
 import torch
 from torch import einsum
@@ -117,9 +117,20 @@ 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 check_for_psutil():
+    try:
+        spec = importlib.util.find_spec('psutil')
+        return spec is not None
+    except ModuleNotFoundError:
+        return False
+
+invokeAI_mps_available = check_for_psutil()
+
+# -- Taken from https://github.com/invoke-ai/InvokeAI --
+if invokeAI_mps_available:
+    import psutil
+    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)
@@ -193,7 +204,7 @@ def split_cross_attention_forward_invokeAI(self, x, context=None, mask=None):
     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 --
+# -- End of code from https://github.com/invoke-ai/InvokeAI --
 
 def xformers_attention_forward(self, x, context=None, mask=None):
     h = self.heads
-- 
cgit v1.2.1


From 574c8e554a5371eca2cbf344764cb241c6ec4efc Mon Sep 17 00:00:00 2001
From: brkirch <brkirch@users.noreply.github.com>
Date: Tue, 11 Oct 2022 03:32:11 -0400
Subject: Add InvokeAI and lstein to credits, add back CUDA support

---
 modules/sd_hijack_optimizations.py | 13 +++++++++++++
 1 file changed, 13 insertions(+)

(limited to 'modules/sd_hijack_optimizations.py')

diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py
index 2a4ac7e0..f006427f 100644
--- a/modules/sd_hijack_optimizations.py
+++ b/modules/sd_hijack_optimizations.py
@@ -173,7 +173,20 @@ def einsum_op_tensor_mem(q, k, v, max_tensor_mb):
         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_cuda(q, k, v):
+    stats = torch.cuda.memory_stats(q.device)
+    mem_active = stats['active_bytes.all.current']
+    mem_reserved = stats['reserved_bytes.all.current']
+    mem_free_cuda, _ = torch.cuda.mem_get_info(q.device)
+    mem_free_torch = mem_reserved - mem_active
+    mem_free_total = mem_free_cuda + mem_free_torch
+    # Divide factor of safety as there's copying and fragmentation
+    return self.einsum_op_tensor_mem(q, k, v, mem_free_total / 3.3 / (1 << 20))
+
 def einsum_op(q, k, v):
+    if q.device.type == 'cuda':
+        return einsum_op_cuda(q, k, v)
+
     if q.device.type == 'mps':
         if mem_total_gb >= 32:
             return einsum_op_mps_v1(q, k, v)
-- 
cgit v1.2.1


From 861db783c7acfcb93cf0b5191db3d50f9a9bc531 Mon Sep 17 00:00:00 2001
From: brkirch <brkirch@users.noreply.github.com>
Date: Tue, 11 Oct 2022 05:13:17 -0400
Subject: Use apply_hypernetwork function

---
 modules/sd_hijack_optimizations.py | 14 ++++----------
 1 file changed, 4 insertions(+), 10 deletions(-)

(limited to 'modules/sd_hijack_optimizations.py')

diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py
index f006427f..79405525 100644
--- a/modules/sd_hijack_optimizations.py
+++ b/modules/sd_hijack_optimizations.py
@@ -202,16 +202,10 @@ def split_cross_attention_forward_invokeAI(self, x, context=None, mask=None):
     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
+    context_k, context_v = hypernetwork.apply_hypernetwork(shared.loaded_hypernetwork, context)
+    k = self.to_k(context_k) * self.scale
+    v = self.to_v(context_v)
+    del context, context_k, context_v, 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)
-- 
cgit v1.2.1