(requested) support AMD 8GB GPUs via Windows DirectML

this update is requested by users

modules/ops.py

@@ -0,0 +1,19 @@
+import torch
+import contextlib
+
+
+@contextlib.contextmanager
+def use_patched_ops(operations):
+    op_names = ['Linear', 'Conv2d', 'Conv3d', 'GroupNorm', 'LayerNorm']
+    backups = {op_name: getattr(torch.nn, op_name) for op_name in op_names}
+
+    try:
+        for op_name in op_names:
+            setattr(torch.nn, op_name, getattr(operations, op_name))
+
+        yield
+
+    finally:
+        for op_name in op_names:
+            setattr(torch.nn, op_name, backups[op_name])
+    return

modules/patch.py

@@ -218,7 +218,7 @@ def compute_cfg(uncond, cond, cfg_scale, t):
 def patched_sampling_function(model, x, timestep, uncond, cond, cond_scale, model_options=None, seed=None):
     global eps_record
 
-    if math.isclose(cond_scale, 1.0):
+    if math.isclose(cond_scale, 1.0) and not model_options.get("disable_cfg1_optimization", False):
         final_x0 = calc_cond_uncond_batch(model, cond, None, x, timestep, model_options)[0]
 
         if eps_record is not None:
@@ -480,6 +480,10 @@ def build_loaded(module, loader_name):
 
 
 def patch_all():
+    if ldm_patched.modules.model_management.directml_enabled:
+        ldm_patched.modules.model_management.lowvram_available = True
+        ldm_patched.modules.model_management.OOM_EXCEPTION = Exception
+    
     patch_all_precision()
     patch_all_clip()
 

modules/patch_clip.py

@@ -16,30 +16,12 @@ import ldm_patched.modules.samplers
 import ldm_patched.modules.sd
 import ldm_patched.modules.sd1_clip
 import ldm_patched.modules.clip_vision
-import ldm_patched.modules.model_management as model_management
 import ldm_patched.modules.ops as ops
-import contextlib
 
+from modules.ops import use_patched_ops
 from transformers import CLIPTextModel, CLIPTextConfig, modeling_utils, CLIPVisionConfig, CLIPVisionModelWithProjection
 
 
-@contextlib.contextmanager
-def use_patched_ops(operations):
-    op_names = ['Linear', 'Conv2d', 'Conv3d', 'GroupNorm', 'LayerNorm']
-    backups = {op_name: getattr(torch.nn, op_name) for op_name in op_names}
-
-    try:
-        for op_name in op_names:
-            setattr(torch.nn, op_name, getattr(operations, op_name))
-
-        yield
-
-    finally:
-        for op_name in op_names:
-            setattr(torch.nn, op_name, backups[op_name])
-    return
-
-
 def patched_encode_token_weights(self, token_weight_pairs):
     to_encode = list()
     max_token_len = 0

modules/sample_hijack.py

@@ -99,6 +99,13 @@ def sample_hacked(model, noise, positive, negative, cfg, device, sampler, sigmas
     calculate_start_end_timesteps(model, negative)
     calculate_start_end_timesteps(model, positive)
 
+    if latent_image is not None:
+        latent_image = model.process_latent_in(latent_image)
+
+    if hasattr(model, 'extra_conds'):
+        positive = encode_model_conds(model.extra_conds, positive, noise, device, "positive", latent_image=latent_image, denoise_mask=denoise_mask)
+        negative = encode_model_conds(model.extra_conds, negative, noise, device, "negative", latent_image=latent_image, denoise_mask=denoise_mask)
+
     #make sure each cond area has an opposite one with the same area
     for c in positive:
         create_cond_with_same_area_if_none(negative, c)
@@ -111,13 +118,6 @@ def sample_hacked(model, noise, positive, negative, cfg, device, sampler, sigmas
     apply_empty_x_to_equal_area(list(filter(lambda c: c.get('control_apply_to_uncond', False) == True, positive)), negative, 'control', lambda cond_cnets, x: cond_cnets[x])
     apply_empty_x_to_equal_area(positive, negative, 'gligen', lambda cond_cnets, x: cond_cnets[x])
 
-    if latent_image is not None:
-        latent_image = model.process_latent_in(latent_image)
-
-    if hasattr(model, 'extra_conds'):
-        positive = encode_model_conds(model.extra_conds, positive, noise, device, "positive", latent_image=latent_image, denoise_mask=denoise_mask)
-        negative = encode_model_conds(model.extra_conds, negative, noise, device, "negative", latent_image=latent_image, denoise_mask=denoise_mask)
-
     extra_args = {"cond":positive, "uncond":negative, "cond_scale": cfg, "model_options": model_options, "seed":seed}
 
     if current_refiner is not None and hasattr(current_refiner.model, 'extra_conds'):
@@ -174,7 +174,7 @@ def calculate_sigmas_scheduler_hacked(model, scheduler_name, steps):
     elif scheduler_name == "sgm_uniform":
         sigmas = normal_scheduler(model, steps, sgm=True)
     elif scheduler_name == "turbo":
-        sigmas = SDTurboScheduler().get_sigmas(namedtuple('Patcher', ['model'])(model=model), steps)[0]
+        sigmas = SDTurboScheduler().get_sigmas(namedtuple('Patcher', ['model'])(model=model), steps=steps, denoise=1.0)[0]
     else:
         raise TypeError("error invalid scheduler")
     return sigmas