argument to disable memory efficient for sdp

modules/sd_hijack.py

@@ -43,9 +43,14 @@ def apply_optimizations():
         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'
+        if cmd_opts.opt_sdp_no_mem_attention:
+            print("Applying scaled dot product cross attention optimization (without memory efficient attention).")
+            ldm.modules.attention.CrossAttention.forward = sd_hijack_optimizations.scaled_dot_product_no_mem_attention_forward
+            optimization_method = 'sdp-no-mem'
+        else:
+            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

@@ -388,6 +388,10 @@ def scaled_dot_product_attention_forward(self, x, context=None, mask=None):
     hidden_states = self.to_out[1](hidden_states)
     return hidden_states
 
+def scaled_dot_product_no_mem_attention_forward(self, x, context=None, mask=None):
+    with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=True, enable_mem_efficient=False):
+        return scaled_dot_product_attention_forward(self, x, context, mask)
+
 def cross_attention_attnblock_forward(self, x):
         h_ = x
         h_ = self.norm(h_)

modules/shared.py

@@ -70,6 +70,7 @@ parser.add_argument("--sub-quad-chunk-threshold", type=int, help="the percentage
 parser.add_argument("--opt-split-attention-invokeai", action='store_true', help="force-enables InvokeAI's cross-attention layer optimization. By default, it's on when cuda is unavailable.")
 parser.add_argument("--opt-split-attention-v1", action='store_true', help="enable older version of split attention optimization that does not consume all the VRAM it can find")
 parser.add_argument("--opt-sdp-attention", action='store_true', help="enable scaled dot product cross-attention layer optimization; requires PyTorch 2.*")
+parser.add_argument("--opt-sdp-no-mem-attention", action='store_true', help="disables memory efficient sdp, makes image generation deterministic; requires --opt-sdp-attention")
 parser.add_argument("--disable-opt-split-attention", action='store_true', help="force-disables cross-attention layer optimization")
 parser.add_argument("--disable-nan-check", action='store_true', help="do not check if produced images/latent spaces have nans; useful for running without a checkpoint in CI")
 parser.add_argument("--use-cpu", nargs='+', help="use CPU as torch device for specified modules", default=[], type=str.lower)