Mirror/stable-diffusion-webui
1
0
Fork 0
mirror of https://github.com/AUTOMATIC1111/stable-diffusion-webui.git synced 2024-06-07 21:20:49 +00:00
Code Issues Packages Projects Releases Wiki Activity

switch to the proper way of calling xformers

Browse Source
This commit is contained in:
C43H66N12O12S2 2022-10-08 04:09:18 +03:00 committed by GitHub
parent 5e3ff846c5
commit c9cc65b201
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 3 additions and 25 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

28
modules/sd_hijack_optimizations.py
View File

@ -94,39 +94,17 @@ def split_cross_attention_forward(self, x, context=None, mask=None):
return self.to_out(r2) return self.to_out(r2)
def _maybe_init(self, x):
"""
Initialize the attention operator, if required We expect the head dimension to be exposed here, meaning that x
: B, Head, Length
"""
if self.attention_op is not None:
return
_, M, K = x.shape
try:
self.attention_op = xformers.ops.AttentionOpDispatch(
dtype=x.dtype,
device=x.device,
k=K,
attn_bias_type=type(None),
has_dropout=False,
kv_len=M,
q_len=M,
).op
except NotImplementedError as err:
raise NotImplementedError(f"Please install xformers with the flash attention / cutlass components.\n{err}")
def xformers_attention_forward(self, x, context=None, mask=None): def xformers_attention_forward(self, x, context=None, mask=None):
h = self.heads h = self.heads
q_in = self.to_q(x) q_in = self.to_q(x)
context = default(context, x) context = default(context, x)
k_in = self.to_k(context) k_in = self.to_k(context)
v_in = self.to_v(context) v_in = self.to_v(context)
q, k, v = map(lambda t: rearrange(t, 'b n (h d) -> (b h) n d', h=h), (q_in, k_in, v_in)) 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 del q_in, k_in, v_in
self._maybe_init(q) out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None)
out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=self.attention_op)
out = rearrange(out, '(b h) n d -> b n (h d)', h=h) out = rearrange(out, 'b n h d -> b n (h d)', h=h)
return self.to_out(out) return self.to_out(out)
def cross_attention_attnblock_forward(self, x): def cross_attention_attnblock_forward(self, x):