openvinotoolkit/stable-diffusion-webui
1
1
Fork 0
mirror of https://github.com/openvinotoolkit/stable-diffusion-webui.git synced 2024-12-14 14:45:06 +03:00
Code Issues Packages Projects Releases Wiki Activity

Merge remote-tracking branch 'takuma104/xformers-flash-attention'

Browse Source
This commit is contained in:
AUTOMATIC 2023-01-23 16:01:53 +03:00
commit 3fa482076a
2 changed files with 25 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
modules/sd_hijack_optimizations.py
View File

@ -290,7 +290,19 @@ def xformers_attention_forward(self, x, context=None, mask=None):
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)
if shared.cmd_opts.xformers_flash_attention:
op = xformers.ops.MemoryEfficientAttentionFlashAttentionOp
fw, bw = op
if not fw.supports(xformers.ops.fmha.Inputs(query=q, key=k, value=v, attn_bias=None)):
# print('xformers_attention_forward', q.shape, k.shape, v.shape)
# Flash Attention is not availabe for the input arguments.
# Fallback to default xFormers' backend.
op = None
else:
op = None
out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=op)
out = rearrange(out, 'b n h d -> b n (h d)', h=h)
return self.to_out(out)
@ -365,7 +377,17 @@ def xformers_attnblock_forward(self, x):
q = q.contiguous()
k = k.contiguous()
v = v.contiguous()
out = xformers.ops.memory_efficient_attention(q, k, v)
if shared.cmd_opts.xformers_flash_attention:
op = xformers.ops.MemoryEfficientAttentionFlashAttentionOp
fw, bw = op
if not fw.supports(xformers.ops.fmha.Inputs(query=q, key=k, value=v)):
# print('xformers_attnblock_forward', q.shape, k.shape, v.shape)
# Flash Attention is not availabe for the input arguments.
# Fallback to default xFormers' backend.
op = None
else:
op = None
out = xformers.ops.memory_efficient_attention(q, k, v, op=op)
out = rearrange(out, 'b (h w) c -> b c h w', h=h)
out = self.proj_out(out)
return x + out

1
modules/shared.py
View File

@ -57,6 +57,7 @@ parser.add_argument("--realesrgan-models-path", type=str, help="Path to director
parser.add_argument("--clip-models-path", type=str, help="Path to directory with CLIP model file(s).", default=None)
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("--xformers-flash-attention", action='store_true', help="enable xformers with Flash Attention to improve reproducibility (supported for SD2.x or variant only)")
parser.add_argument("--deepdanbooru", action='store_true', help="does not do anything")
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-sub-quad-attention", action='store_true', help="enable memory efficient sub-quadratic cross-attention layer optimization")