Public/automatic
1
0
Fork 0
mirror of https://github.com/vladmandic/automatic synced 2026-09-18 16:54:33 +02:00
Code Issues Packages Projects Releases Wiki Activity

cleanup before merge

Browse Source
This commit is contained in:
Vladimir Mandic
2023-03-10 09:57:10 -05:00
parent e8766b4975
commit b48d05ca5a
4 changed files with 102 additions and 66 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/sd_hijack.py
-4
View File
@@ -42,10 +42,6 @@ def apply_optimizations():
ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.xformers_attention_forward
ldm.modules.diffusionmodules.model.AttnBlock.forward = sd_hijack_optimizations.xformers_attnblock_forward
optimization_method = 'xformers'
elif cmd_opts.opt_sdp_attention and (hasattr(torch.nn.functional, "scaled_dot_product_attention") and callable(getattr(torch.nn.functional, "scaled_dot_product_attention"))):
print("Applying scaled dot product cross attention optimization.")
ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.scaled_dot_product_attention_forward
optimization_method = 'sdp'
elif cmd_opts.opt_sub_quad_attention:
print("Applying sub-quadratic cross attention optimization.")
ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.sub_quad_attention_forward
modules/sd_hijack_optimizations.py
-42
View File
@@ -346,48 +346,6 @@ def xformers_attention_forward(self, x, context=None, mask=None):
out = rearrange(out, 'b n h d -> b n (h d)', h=h)
return self.to_out(out)
# Based on Diffusers usage of scaled dot product attention from https://github.com/huggingface/diffusers/blob/c7da8fd23359a22d0df2741688b5b4f33c26df21/src/diffusers/models/cross_attention.py
# The scaled_dot_product_attention_forward function contains parts of code under Apache-2.0 license listed under Scaled Dot Product Attention in the Licenses section of the web UI interface
def scaled_dot_product_attention_forward(self, x, context=None, mask=None):
batch_size, sequence_length, inner_dim = x.shape
if mask is not None:
mask = self.prepare_attention_mask(mask, sequence_length, batch_size)
mask = mask.view(batch_size, self.heads, -1, mask.shape[-1])
h = self.heads
q_in = self.to_q(x)
context = default(context, x)
context_k, context_v = hypernetwork.apply_hypernetworks(shared.loaded_hypernetworks, context)
k_in = self.to_k(context_k)
v_in = self.to_v(context_v)
head_dim = inner_dim // h
q = q_in.view(batch_size, -1, h, head_dim).transpose(1, 2)
k = k_in.view(batch_size, -1, h, head_dim).transpose(1, 2)
v = v_in.view(batch_size, -1, h, head_dim).transpose(1, 2)
del q_in, k_in, v_in
dtype = q.dtype
if shared.opts.upcast_attn:
q, k = q.float(), k.float()
# the output of sdp = (batch, num_heads, seq_len, head_dim)
hidden_states = torch.nn.functional.scaled_dot_product_attention(
q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False
)
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, h * head_dim)
hidden_states = hidden_states.to(dtype)
# linear proj
hidden_states = self.to_out[0](hidden_states)
# dropout
hidden_states = self.to_out[1](hidden_states)
return hidden_states
def cross_attention_attnblock_forward(self, x):
h_ = x
h_ = self.norm(h_)