AI/diffusers
1
0
Fork 0
mirror of https://github.com/huggingface/diffusers.git synced 2026-01-27 17:22:53 +03:00
Code Activity

Add option to use fixed latents for UniDiffuserPipelineSlowTests and fix issue in modeling_text_decoder.py.

Browse Source
This commit is contained in:
Daniel Gu
2023-05-15 02:10:38 -07:00
parent be4abffe88
commit 848b7e6c2e
2 changed files with 28 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/diffusers/pipelines/unidiffuser/modeling_text_decoder.py
View File

@@ -234,7 +234,7 @@ class UniDiffuserTextDecoder(ModelMixin, ConfigMixin, ModuleUtilsMixin):
tokens = None
scores = None
seq_lengths = torch.ones(beam_size, device=device)
is_stopped = torch.zeros(beam_size, device=device, dtype=bool)
is_stopped = torch.zeros(beam_size, device=device, dtype=torch.bool)
if embed is not None:
generated = embed

28
tests/pipelines/unidiffuser/test_unidiffuser.py
View File

@@ -522,7 +522,7 @@ class UniDiffuserPipelineSlowTests(unittest.TestCase):
gc.collect()
torch.cuda.empty_cache()
def get_inputs(self, seed=0):
def get_inputs(self, device, seed=0, generate_latents=False):
generator = torch.manual_seed(seed)
image = load_image(
"https://huggingface.co/datasets/hf-internal-testing/diffusers-images/resolve/main/unidiffuser/unidiffuser_example_image.jpg"
@@ -535,7 +535,33 @@ class UniDiffuserPipelineSlowTests(unittest.TestCase):
"guidance_scale": 8.0,
"output_type": "numpy",
}
if generate_latents:
latents = self.get_fixed_latents(device, seed=seed)
for latent_name, latent_tensor in latents.items():
inputs[latent_name] = latent_tensor
return inputs
def get_fixed_latents(self, device, seed=0):
if type(device) == str:
device = torch.device(device)
latent_device = torch.device("cpu")
generator = torch.Generator(device=latent_device).manual_seed(seed)
# Hardcode the shapes for now.
prompt_latents = randn_tensor((1, 77, 32), generator=generator, device=device, dtype=torch.float32)
vae_latents = randn_tensor((1, 4, 16, 16), generator=generator, device=device, dtype=torch.float32)
clip_latents = randn_tensor((1, 1, 32), generator=generator, device=device, dtype=torch.float32)
# Move latents onto desired device.
prompt_latents = prompt_latents.to(device)
vae_latents = vae_latents.to(device)
clip_latents = clip_latents.to(device)
latents = {
"prompt_latents": prompt_latents,
"vae_latents": vae_latents,
"clip_latents": clip_latents,
}
return latents
def test_unidiffuser_default_joint_v1(self):
pipe = UniDiffuserPipeline.from_pretrained("dg845/unidiffuser-diffusers")