[DO NOT MERGE] All perf improvements bundle

configs/alt-diffusion-inference.yaml

@@ -40,7 +40,7 @@ model:
         use_spatial_transformer: True
         transformer_depth: 1
         context_dim: 768
-        use_checkpoint: True
+        use_checkpoint: False
         legacy: False
 
     first_stage_config:

configs/alt-diffusion-m18-inference.yaml

@@ -41,7 +41,7 @@ model:
         use_linear_in_transformer: True
         transformer_depth: 1
         context_dim: 1024
-        use_checkpoint: True
+        use_checkpoint: False
         legacy: False
 
     first_stage_config:

configs/instruct-pix2pix.yaml

@@ -45,7 +45,7 @@ model:
         use_spatial_transformer: True
         transformer_depth: 1
         context_dim: 768
-        use_checkpoint: True
+        use_checkpoint: False
         legacy: False
 
     first_stage_config:

configs/sd_xl_inpaint.yaml

@@ -21,7 +21,7 @@ model:
       params:
         adm_in_channels: 2816
         num_classes: sequential
-        use_checkpoint: True
+        use_checkpoint: False
         in_channels: 9
         out_channels: 4
         model_channels: 320

configs/v1-inference.yaml

@@ -40,7 +40,7 @@ model:
         use_spatial_transformer: True
         transformer_depth: 1
         context_dim: 768
-        use_checkpoint: True
+        use_checkpoint: False
         legacy: False
 
     first_stage_config:

configs/v1-inpainting-inference.yaml

@@ -40,7 +40,7 @@ model:
         use_spatial_transformer: True
         transformer_depth: 1
         context_dim: 768
-        use_checkpoint: True
+        use_checkpoint: False
         legacy: False
 
     first_stage_config:

extensions-builtin/Lora/networks.py

@@ -378,13 +378,18 @@ def network_apply_weights(self: Union[torch.nn.Conv2d, torch.nn.Linear, torch.nn
         self.network_weights_backup = weights_backup
 
     bias_backup = getattr(self, "network_bias_backup", None)
-    if bias_backup is None:
+    if bias_backup is None and wanted_names != ():
         if isinstance(self, torch.nn.MultiheadAttention) and self.out_proj.bias is not None:
             bias_backup = self.out_proj.bias.to(devices.cpu, copy=True)
         elif getattr(self, 'bias', None) is not None:
             bias_backup = self.bias.to(devices.cpu, copy=True)
         else:
             bias_backup = None
+
+        # Unlike weight which always has value, some modules don't have bias.
+        # Only report if bias is not None and current bias are not unchanged.
+        if bias_backup is not None and current_names != ():
+            raise RuntimeError("no backup bias found and current bias are not unchanged")
         self.network_bias_backup = bias_backup
 
     if current_names != wanted_names:

modules/cmd_args.py

@@ -41,7 +41,7 @@
 parser.add_argument("--lowram", action='store_true', help="load stable diffusion checkpoint weights to VRAM instead of RAM")
 parser.add_argument("--always-batch-cond-uncond", action='store_true', help="does not do anything")
 parser.add_argument("--unload-gfpgan", action='store_true', help="does not do anything.")
-parser.add_argument("--precision", type=str, help="evaluate at this precision", choices=["full", "autocast"], default="autocast")
+parser.add_argument("--precision", type=str, help="evaluate at this precision", choices=["full", "half", "autocast"], default="autocast")
 parser.add_argument("--upcast-sampling", action='store_true', help="upcast sampling. No effect with --no-half. Usually produces similar results to --no-half with better performance while using less memory.")
 parser.add_argument("--share", action='store_true', help="use share=True for gradio and make the UI accessible through their site")
 parser.add_argument("--ngrok", type=str, help="ngrok authtoken, alternative to gradio --share", default=None)
@@ -69,7 +69,8 @@
 parser.add_argument("--opt-sdp-attention", action='store_true', help="prefer scaled dot product cross-attention layer optimization for automatic choice of optimization; requires PyTorch 2.*")
 parser.add_argument("--opt-sdp-no-mem-attention", action='store_true', help="prefer scaled dot product cross-attention layer optimization without memory efficient attention for automatic choice of optimization, makes image generation deterministic; requires PyTorch 2.*")
 parser.add_argument("--disable-opt-split-attention", action='store_true', help="prefer no cross-attention layer optimization for automatic choice of 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("--disable-nan-check", action='store_true', help="[Deprecated] do not check if produced images/latent spaces have nans; useful for running without a checkpoint in CI")
+parser.add_argument("--enable-nan-check", action='store_true', help="Check if produced images/latent spaces have nans at extra performance cost. (~20ms/it)")
 parser.add_argument("--use-cpu", nargs='+', help="use CPU as torch device for specified modules", default=[], type=str.lower)
 parser.add_argument("--use-ipex", action="store_true", help="use Intel XPU as torch device")
 parser.add_argument("--disable-model-loading-ram-optimization", action='store_true', help="disable an optimization that reduces RAM use when loading a model")

modules/devices.py

@@ -114,6 +114,9 @@ def enable_tf32():
 
 cpu: torch.device = torch.device("cpu")
 fp8: bool = False
+# Force fp16 for all models in inference. No casting during inference.
+# This flag is controlled by "--precision half" command line arg.
+force_fp16: bool = False
 device: torch.device = None
 device_interrogate: torch.device = None
 device_gfpgan: torch.device = None
@@ -127,6 +130,8 @@ def enable_tf32():
 
 
 def cond_cast_unet(input):
+    if force_fp16:
+        return input.to(torch.float16)
     return input.to(dtype_unet) if unet_needs_upcast else input
 
 
@@ -206,6 +211,11 @@ def autocast(disable=False):
     if disable:
         return contextlib.nullcontext()
 
+    if force_fp16:
+        # No casting during inference if force_fp16 is enabled.
+        # All tensor dtype conversion happens before inference.
+        return contextlib.nullcontext()
+
     if fp8 and device==cpu:
         return torch.autocast("cpu", dtype=torch.bfloat16, enabled=True)
 
@@ -230,7 +240,7 @@ class NansException(Exception):
 
 
 def test_for_nans(x, where):
-    if shared.cmd_opts.disable_nan_check:
+    if not shared.cmd_opts.enable_nan_check:
         return
 
     if not torch.all(torch.isnan(x)).item():
@@ -250,8 +260,6 @@ def test_for_nans(x, where):
     else:
         message = "A tensor with all NaNs was produced."
 
-    message += " Use --disable-nan-check commandline argument to disable this check."
-
     raise NansException(message)
 
 
@@ -269,3 +277,17 @@ def first_time_calculation():
     x = torch.zeros((1, 1, 3, 3)).to(device, dtype)
     conv2d = torch.nn.Conv2d(1, 1, (3, 3)).to(device, dtype)
     conv2d(x)
+
+
+def force_model_fp16():
+    """
+    ldm and sgm has modules.diffusionmodules.util.GroupNorm32.forward, which
+    force conversion of input to float32. If force_fp16 is enabled, we need to
+    prevent this casting.
+    """
+    assert force_fp16
+    import sgm.modules.diffusionmodules.util as sgm_util
+    import ldm.modules.diffusionmodules.util as ldm_util
+    sgm_util.GroupNorm32 = torch.nn.GroupNorm
+    ldm_util.GroupNorm32 = torch.nn.GroupNorm
+    print("ldm/sgm GroupNorm32 replaced with normal torch.nn.GroupNorm due to `--precision half`.")

modules/launch_utils.py

@@ -440,6 +440,10 @@ def prepare_environment():
         git_pull_recursive(extensions_dir)
         startup_timer.record("update extensions")
 
+    if args.disable_nan_check:
+        print("Nan check disabled by default. --disable-nan-check argument is now ignored. "
+              "Use --enable-nan-check to re-enable nan check.")
+
     if "--exit" in sys.argv:
         print("Exiting because of --exit argument")
         exit(0)
@@ -454,8 +458,8 @@ def configure_for_tests():
         sys.argv.append(os.path.join(script_path, "test/test_files/empty.pt"))
     if "--skip-torch-cuda-test" not in sys.argv:
         sys.argv.append("--skip-torch-cuda-test")
-    if "--disable-nan-check" not in sys.argv:
-        sys.argv.append("--disable-nan-check")
+    if "--enable-nan-check" in sys.argv:
+        sys.argv.remove("--enable-nan-check")
 
     os.environ['COMMANDLINE_ARGS'] = ""
 

modules/processing.py

@@ -115,20 +115,17 @@ def txt2img_image_conditioning(sd_model, x, width, height):
         return x.new_zeros(x.shape[0], 2*sd_model.noise_augmentor.time_embed.dim, dtype=x.dtype, device=x.device)
 
     else:
-        sd = sd_model.model.state_dict()
-        diffusion_model_input = sd.get('diffusion_model.input_blocks.0.0.weight', None)
-        if diffusion_model_input is not None:
-            if diffusion_model_input.shape[1] == 9:
-                # The "masked-image" in this case will just be all 0.5 since the entire image is masked.
-                image_conditioning = torch.ones(x.shape[0], 3, height, width, device=x.device) * 0.5
-                image_conditioning = images_tensor_to_samples(image_conditioning,
-                                                              approximation_indexes.get(opts.sd_vae_encode_method))
+        if getattr(sd_model.model, "is_sdxl_inpaint", False):
+            # The "masked-image" in this case will just be all 0.5 since the entire image is masked.
+            image_conditioning = torch.ones(x.shape[0], 3, height, width, device=x.device) * 0.5
+            image_conditioning = images_tensor_to_samples(image_conditioning,
+                                                            approximation_indexes.get(opts.sd_vae_encode_method))
 
-                # Add the fake full 1s mask to the first dimension.
-                image_conditioning = torch.nn.functional.pad(image_conditioning, (0, 0, 0, 0, 1, 0), value=1.0)
-                image_conditioning = image_conditioning.to(x.dtype)
+            # Add the fake full 1s mask to the first dimension.
+            image_conditioning = torch.nn.functional.pad(image_conditioning, (0, 0, 0, 0, 1, 0), value=1.0)
+            image_conditioning = image_conditioning.to(x.dtype)
 
-                return image_conditioning
+            return image_conditioning
 
         # Dummy zero conditioning if we're not using inpainting or unclip models.
         # Still takes up a bit of memory, but no encoder call.
@@ -390,11 +387,8 @@ def img2img_image_conditioning(self, source_image, latent_image, image_mask=None
         if self.sampler.conditioning_key == "crossattn-adm":
             return self.unclip_image_conditioning(source_image)
 
-        sd = self.sampler.model_wrap.inner_model.model.state_dict()
-        diffusion_model_input = sd.get('diffusion_model.input_blocks.0.0.weight', None)
-        if diffusion_model_input is not None:
-            if diffusion_model_input.shape[1] == 9:
-                return self.inpainting_image_conditioning(source_image, latent_image, image_mask=image_mask)
+        if getattr(self.sampler.model_wrap.inner_model.model, "is_sdxl_inpaint", False):
+            return self.inpainting_image_conditioning(source_image, latent_image, image_mask=image_mask)
 
         # Dummy zero conditioning if we're not using inpainting or depth model.
         return latent_image.new_zeros(latent_image.shape[0], 5, 1, 1)

modules/sd_hijack_checkpoint.py

@@ -4,16 +4,19 @@
 import ldm.modules.diffusionmodules.openaimodel
 
 
+# Setting flag=False so that torch skips checking parameters.
+# parameters checking is expensive in frequent operations.
+
 def BasicTransformerBlock_forward(self, x, context=None):
-    return checkpoint(self._forward, x, context)
+    return checkpoint(self._forward, x, context, flag=False)
 
 
 def AttentionBlock_forward(self, x):
-    return checkpoint(self._forward, x)
+    return checkpoint(self._forward, x, flag=False)
 
 
 def ResBlock_forward(self, x, emb):
-    return checkpoint(self._forward, x, emb)
+    return checkpoint(self._forward, x, emb, flag=False)
 
 
 stored = []

modules/sd_hijack_optimizations.py

@@ -486,7 +486,19 @@ def xformers_attention_forward(self, x, context=None, mask=None, **kwargs):
     k_in = self.to_k(context_k)
     v_in = self.to_v(context_v)
 
-    q, k, v = (rearrange(t, 'b n (h d) -> b n h d', h=h) for t in (q_in, k_in, v_in))
+    def _reshape(t):
+        """rearrange(t, 'b n (h d) -> b n h d', h=h).
+        Using torch native operations to avoid overhead as this function is
+        called frequently. (70 times/it for SDXL)
+        """
+        b, n, _ = t.shape  # Get the batch size (b) and sequence length (n)
+        d = t.shape[2] // h  # Determine the depth per head
+        return t.reshape(b, n, h, d)
+
+    q = _reshape(q_in)
+    k = _reshape(k_in)
+    v = _reshape(v_in)
+
     del q_in, k_in, v_in
 
     dtype = q.dtype
@@ -497,7 +509,9 @@ def xformers_attention_forward(self, x, context=None, mask=None, **kwargs):
 
     out = out.to(dtype)
 
-    out = rearrange(out, 'b n h d -> b n (h d)', h=h)
+    # out = rearrange(out, 'b n h d -> b n (h d)', h=h)
+    b, n, h, d = out.shape
+    out = out.reshape(b, n, h * d)
     return self.to_out(out)
 
 

modules/sd_hijack_unet.py

@@ -36,7 +36,7 @@ def cat(self, tensors, *args, **kwargs):
 
 # Below are monkey patches to enable upcasting a float16 UNet for float32 sampling
 def apply_model(orig_func, self, x_noisy, t, cond, **kwargs):
-
+    """Always make sure inputs to unet are in correct dtype."""
     if isinstance(cond, dict):
         for y in cond.keys():
             if isinstance(cond[y], list):
@@ -45,7 +45,11 @@ def apply_model(orig_func, self, x_noisy, t, cond, **kwargs):
                 cond[y] = cond[y].to(devices.dtype_unet) if isinstance(cond[y], torch.Tensor) else cond[y]
 
     with devices.autocast():
-        return orig_func(self, x_noisy.to(devices.dtype_unet), t.to(devices.dtype_unet), cond, **kwargs).float()
+        result = orig_func(self, x_noisy.to(devices.dtype_unet), t.to(devices.dtype_unet), cond, **kwargs)
+        if devices.unet_needs_upcast:
+            return result.float()
+        else:
+            return result
 
 
 class GELUHijack(torch.nn.GELU, torch.nn.Module):
@@ -64,12 +68,11 @@ def hijack_ddpm_edit():
     if not ddpm_edit_hijack:
         CondFunc('modules.models.diffusion.ddpm_edit.LatentDiffusion.decode_first_stage', first_stage_sub, first_stage_cond)
         CondFunc('modules.models.diffusion.ddpm_edit.LatentDiffusion.encode_first_stage', first_stage_sub, first_stage_cond)
-        ddpm_edit_hijack = CondFunc('modules.models.diffusion.ddpm_edit.LatentDiffusion.apply_model', apply_model, unet_needs_upcast)
+        ddpm_edit_hijack = CondFunc('modules.models.diffusion.ddpm_edit.LatentDiffusion.apply_model', apply_model)
 
 
 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("1.13.2") 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)
@@ -81,5 +84,17 @@ def hijack_ddpm_edit():
 CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.encode_first_stage', first_stage_sub, first_stage_cond)
 CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.get_first_stage_encoding', lambda orig_func, *args, **kwargs: orig_func(*args, **kwargs).float(), first_stage_cond)
 
-CondFunc('sgm.modules.diffusionmodules.wrappers.OpenAIWrapper.forward', apply_model, unet_needs_upcast)
-CondFunc('sgm.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)
+CondFunc('ldm.models.diffusion.ddpm.LatentDiffusion.apply_model', apply_model)
+CondFunc('sgm.modules.diffusionmodules.wrappers.OpenAIWrapper.forward', apply_model)
+
+
+def timestep_embedding_cast_result(orig_func, timesteps, *args, **kwargs):
+    if devices.unet_needs_upcast and timesteps.dtype == torch.int64:
+        dtype = torch.float32
+    else:
+        dtype = devices.dtype_unet
+    return orig_func(timesteps, *args, **kwargs).to(dtype=dtype)
+
+
+CondFunc('ldm.modules.diffusionmodules.openaimodel.timestep_embedding', timestep_embedding_cast_result)
+CondFunc('sgm.modules.diffusionmodules.openaimodel.timestep_embedding', timestep_embedding_cast_result)

modules/sd_hijack_utils.py

@@ -1,7 +1,11 @@
 import importlib
 
+
+always_true_func = lambda *args, **kwargs: True
+
+
 class CondFunc:
-    def __new__(cls, orig_func, sub_func, cond_func):
+    def __new__(cls, orig_func, sub_func, cond_func=always_true_func):
         self = super(CondFunc, cls).__new__(cls)
         if isinstance(orig_func, str):
             func_path = orig_func.split('.')
@@ -20,13 +24,13 @@ def __new__(cls, orig_func, sub_func, cond_func):
                 print(f"Warning: Failed to resolve {orig_func} for CondFunc hijack")
                 pass
         self.__init__(orig_func, sub_func, cond_func)
-        return lambda *args, **kwargs: self(*args, **kwargs)
-    def __init__(self, orig_func, sub_func, cond_func):
-        self.__orig_func = orig_func
-        self.__sub_func = sub_func
-        self.__cond_func = cond_func
-    def __call__(self, *args, **kwargs):
-        if not self.__cond_func or self.__cond_func(self.__orig_func, *args, **kwargs):
-            return self.__sub_func(self.__orig_func, *args, **kwargs)
-        else:
-            return self.__orig_func(*args, **kwargs)
+        return lambda *args, **kwargs: self(*args, **kwargs)
+    def __init__(self, orig_func, sub_func, cond_func):
+        self.__orig_func = orig_func
+        self.__sub_func = sub_func
+        self.__cond_func = cond_func
+    def __call__(self, *args, **kwargs):
+        if not self.__cond_func or self.__cond_func(self.__orig_func, *args, **kwargs):
+            return self.__sub_func(self.__orig_func, *args, **kwargs)
+        else:
+            return self.__orig_func(*args, **kwargs)