automatically switch to 32-bit float VAE if the generated picture has NaNs.

2 files changed, 37 insertions, 5 deletions

diff --git a/modules/processing.py b/modules/processing.py
index e028bf9e..a74a5302 100644
--- a/modules/processing.py
+++ b/modules/processing.py
@@ -14,7 +14,7 @@ from skimage import exposure
 from typing import Any, Dict, List

 

 import modules.sd_hijack

-from modules import devices, prompt_parser, masking, sd_samplers, lowvram, generation_parameters_copypaste, extra_networks, sd_vae_approx, scripts, sd_samplers_common, sd_unet

+from modules import devices, prompt_parser, masking, sd_samplers, lowvram, generation_parameters_copypaste, extra_networks, sd_vae_approx, scripts, sd_samplers_common, sd_unet, errors

 from modules.sd_hijack import model_hijack

 from modules.shared import opts, cmd_opts, state

 import modules.shared as shared

@@ -538,6 +538,40 @@ def create_random_tensors(shape, seeds, subseeds=None, subseed_strength=0.0, see
     return x

 

 

+def decode_latent_batch(model, batch, target_device=None, check_for_nans=False):

+    samples = []

+

+    for i in range(batch.shape[0]):

+        sample = decode_first_stage(model, batch[i:i + 1])[0]

+

+        if check_for_nans:

+            try:

+                devices.test_for_nans(sample, "vae")

+            except devices.NansException as e:

+                if devices.dtype_vae == torch.float32 or not shared.opts.auto_vae_precision:

+                    raise e

+

+                errors.print_error_explanation(

+                    "A tensor with all NaNs was produced in VAE.\n"

+                    "Web UI will now convert VAE into 32-bit float and retry.\n"

+                    "To disable this behavior, disable the 'Automaticlly revert VAE to 32-bit floats' setting.\n"

+                    "To always start with 32-bit VAE, use --no-half-vae commandline flag."

+                )

+

+                devices.dtype_vae = torch.float32

+                model.first_stage_model.to(devices.dtype_vae)

+                batch = batch.to(devices.dtype_vae)

+

+                sample = decode_first_stage(model, batch[i:i + 1])[0]

+

+        if target_device is not None:

+            sample = sample.to(target_device)

+

+        samples.append(sample)

+

+    return samples

+

+

 def decode_first_stage(model, x):

     x = model.decode_first_stage(x.to(devices.dtype_vae))

 

@@ -758,10 +792,7 @@ def process_images_inner(p: StableDiffusionProcessing) -> Processed:
             with devices.without_autocast() if devices.unet_needs_upcast else devices.autocast():

                 samples_ddim = p.sample(conditioning=p.c, unconditional_conditioning=p.uc, seeds=p.seeds, subseeds=p.subseeds, subseed_strength=p.subseed_strength, prompts=p.prompts)

 

-            x_samples_ddim = [decode_first_stage(p.sd_model, samples_ddim[i:i+1].to(dtype=devices.dtype_vae))[0].cpu() for i in range(samples_ddim.size(0))]

-            for x in x_samples_ddim:

-                devices.test_for_nans(x, "vae")

-

+            x_samples_ddim = decode_latent_batch(p.sd_model, samples_ddim, target_device=devices.cpu, check_for_nans=True)

             x_samples_ddim = torch.stack(x_samples_ddim).float()

             x_samples_ddim = torch.clamp((x_samples_ddim + 1.0) / 2.0, min=0.0, max=1.0)

 

diff --git a/modules/shared.py b/modules/shared.py
index 1ce7b49e..aa72c9c8 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -427,6 +427,7 @@ options_templates.update(options_section(('sd', "Stable Diffusion"), {
     "comma_padding_backtrack": OptionInfo(20, "Prompt word wrap length limit", gr.Slider, {"minimum": 0, "maximum": 74, "step": 1}).info("in tokens - for texts shorter than specified, if they don't fit into 75 token limit, move them to the next 75 token chunk"),

     "CLIP_stop_at_last_layers": OptionInfo(1, "Clip skip", gr.Slider, {"minimum": 1, "maximum": 12, "step": 1}).link("wiki", "https://github.com/AUTOMATIC1111/stable-diffusion-webui/wiki/Features#clip-skip").info("ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer"),

     "upcast_attn": OptionInfo(False, "Upcast cross attention layer to float32"),

+    "auto_vae_precision": OptionInfo(True, "Automaticlly revert VAE to 32-bit floats").info("triggers when a tensor with NaNs is produced in VAE; disabling the option in this case will result in a black square image"),

     "randn_source": OptionInfo("GPU", "Random number generator source.", gr.Radio, {"choices": ["GPU", "CPU"]}).info("changes seeds drastically; use CPU to produce the same picture across different videocard vendors"),

 }))