diff options
| author | Takuma Mori <takuma104@gmail.com> | 2023-01-21 17:42:04 +0900 |
|---|---|---|
| committer | Takuma Mori <takuma104@gmail.com> | 2023-01-21 17:42:04 +0900 |
| commit | 3262e825cc542ff634e6ba2e3a162eafdc6c1bba (patch) | |
| tree | 08f1be595157904a51ac206f5362ee4e0b51111d | |
| parent | 184e23eb89c198b42f351a4d5ff862ee64917619 (diff) | |
add --xformers-flash-attention option & impl
| -rw-r--r-- | modules/sd_hijack_optimizations.py | 26 | ||||
| -rw-r--r-- | modules/shared.py | 1 |
2 files changed, 25 insertions, 2 deletions
diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py index 4fa54329..9967359b 100644 --- a/modules/sd_hijack_optimizations.py +++ b/modules/sd_hijack_optimizations.py @@ -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
diff --git a/modules/shared.py b/modules/shared.py index 72fb1934..23328adf 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -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")
|