Merge pull request #8782 from FNSpd/master

--upcast-sampling support for CUDA

3 files changed, 4 insertions, 3 deletions

diff --git a/extensions-builtin/Lora/lora.py b/extensions-builtin/Lora/lora.py
index 8937b585..7c371deb 100644
--- a/extensions-builtin/Lora/lora.py
+++ b/extensions-builtin/Lora/lora.py
@@ -178,6 +178,7 @@ def load_loras(names, multipliers=None):
 

 

 def lora_forward(module, input, res):

+    input = devices.cond_cast_unet(input)

     if len(loaded_loras) == 0:

         return res

 

diff --git a/modules/sd_hijack_optimizations.py b/modules/sd_hijack_optimizations.py
index 2e307b5d..372555ff 100644
--- a/modules/sd_hijack_optimizations.py
+++ b/modules/sd_hijack_optimizations.py
@@ -337,7 +337,7 @@ def xformers_attention_forward(self, x, context=None, mask=None):
 

     dtype = q.dtype

     if shared.opts.upcast_attn:

-        q, k = q.float(), k.float()

+        q, k, v = q.float(), k.float(), v.float()

 

     out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None, op=get_xformers_flash_attention_op(q, k, v))

 

@@ -372,7 +372,7 @@ def scaled_dot_product_attention_forward(self, x, context=None, mask=None):
 

     dtype = q.dtype

     if shared.opts.upcast_attn:

-        q, k = q.float(), k.float()

+        q, k, v = q.float(), k.float(), v.float()

 

     # the output of sdp = (batch, num_heads, seq_len, head_dim)

     hidden_states = torch.nn.functional.scaled_dot_product_attention(

diff --git a/modules/sd_hijack_unet.py b/modules/sd_hijack_unet.py
index 843ab66c..15858263 100644
--- a/modules/sd_hijack_unet.py
+++ b/modules/sd_hijack_unet.py
@@ -67,7 +67,7 @@ def hijack_ddpm_edit():
 unet_needs_upcast = lambda *args, **kwargs: devices.unet_needs_upcast

 CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.apply_model', apply_model, unet_needs_upcast)

 CondFunc('ldm.modules.diffusionmodules.openaimodel.timestep_embedding', lambda orig_func, timesteps, *args, **kwargs: orig_func(timesteps, *args, **kwargs).to(torch.float32 if timesteps.dtype == torch.int64 else devices.dtype_unet), unet_needs_upcast)

-if version.parse(torch.__version__) <= version.parse("1.13.1"):

+if version.parse(torch.__version__) <= version.parse("1.13.2") or torch.cuda.is_available():

     CondFunc('ldm.modules.diffusionmodules.util.GroupNorm32.forward', lambda orig_func, self, *args, **kwargs: orig_func(self.float(), *args, **kwargs), unet_needs_upcast)

     CondFunc('ldm.modules.attention.GEGLU.forward', lambda orig_func, self, x: orig_func(self.float(), x.float()).to(devices.dtype_unet), unet_needs_upcast)

     CondFunc('open_clip.transformer.ResidualAttentionBlock.__init__', lambda orig_func, *args, **kwargs: kwargs.update({'act_layer': GELUHijack}) and False or orig_func(*args, **kwargs), lambda _, *args, **kwargs: kwargs.get('act_layer') is None or kwargs['act_layer'] == torch.nn.GELU)