AI/diffusers
1
0
Fork 0
mirror of https://github.com/huggingface/diffusers.git synced 2026-10-23 09:09:01 +03:00
Code Activity

migrate wan vace lora tests to the pipeline-level mixins (#14614)

Browse Source 
Co-authored-by: Sayak Paul <spsayakpaul@gmail.com>
This commit is contained in:
Akshan Krithick
2026-08-26 21:41:34 -07:00
committed by GitHub
parent 276ac2f8c4
commit c14a1e80ec
2 changed files with 61 additions and 190 deletions
Download Patch File Download Diff File Expand all files Collapse all files

188
tests/lora/test_lora_layers_wanvace.py
View File

@@ -1,188 +0,0 @@
# Copyright 2026 HuggingFace Inc.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import os
import sys
import tempfile
import unittest
import numpy as np
import safetensors.torch
import torch
from PIL import Image
from transformers import AutoTokenizer, T5EncoderModel
from diffusers import AutoencoderKLWan, FlowMatchEulerDiscreteScheduler, WanVACEPipeline, WanVACETransformer3DModel
from diffusers.utils.import_utils import is_peft_available
from ..testing_utils import (
floats_tensor,
require_peft_backend,
require_peft_version_greater,
skip_mps,
torch_device,
)
if is_peft_available():
from peft.utils import get_peft_model_state_dict
sys.path.append(".")
from .utils import PeftLoraLoaderMixinTests # noqa: E402
@require_peft_backend
@skip_mps
class WanVACELoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
pipeline_class = WanVACEPipeline
scheduler_cls = FlowMatchEulerDiscreteScheduler
scheduler_kwargs = {}
transformer_kwargs = {
"patch_size": (1, 2, 2),
"num_attention_heads": 2,
"attention_head_dim": 8,
"in_channels": 4,
"out_channels": 4,
"text_dim": 32,
"freq_dim": 16,
"ffn_dim": 16,
"num_layers": 2,
"cross_attn_norm": True,
"qk_norm": "rms_norm_across_heads",
"rope_max_seq_len": 16,
"vace_layers": [0],
"vace_in_channels": 72,
}
transformer_cls = WanVACETransformer3DModel
vae_kwargs = {
"base_dim": 3,
"z_dim": 4,
"dim_mult": [1, 1, 1, 1],
"latents_mean": [-0.7571, -0.7089, -0.9113, -0.7245],
"latents_std": [2.8184, 1.4541, 2.3275, 2.6558],
"num_res_blocks": 1,
"temperal_downsample": [False, True, True],
}
vae_cls = AutoencoderKLWan
has_two_text_encoders = True
tokenizer_cls, tokenizer_id = AutoTokenizer, "hf-internal-testing/tiny-random-t5"
text_encoder_cls, text_encoder_id = T5EncoderModel, "hf-internal-testing/tiny-random-t5"
text_encoder_target_modules = ["q", "k", "v", "o"]
supports_text_encoder_loras = False
@property
def output_shape(self):
return (1, 9, 16, 16, 3)
def get_dummy_inputs(self, with_generator=True):
batch_size = 1
sequence_length = 16
num_channels = 4
num_frames = 9
num_latent_frames = 3 # (num_frames - 1) // temporal_compression_ratio + 1
sizes = (4, 4)
height, width = 16, 16
generator = torch.manual_seed(0)
noise = floats_tensor((batch_size, num_latent_frames, num_channels) + sizes)
input_ids = torch.randint(1, sequence_length, size=(batch_size, sequence_length), generator=generator)
video = [Image.new("RGB", (height, width))] * num_frames
mask = [Image.new("L", (height, width), 0)] * num_frames
pipeline_inputs = {
"video": video,
"mask": mask,
"prompt": "",
"num_frames": num_frames,
"num_inference_steps": 1,
"guidance_scale": 6.0,
"height": height,
"width": height,
"max_sequence_length": sequence_length,
"output_type": "np",
}
if with_generator:
pipeline_inputs.update({"generator": generator})
return noise, input_ids, pipeline_inputs
def test_simple_inference_with_text_lora_denoiser_fused_multi(self):
super().test_simple_inference_with_text_lora_denoiser_fused_multi(expected_atol=9e-3)
def test_simple_inference_with_text_denoiser_lora_unfused(self):
super().test_simple_inference_with_text_denoiser_lora_unfused(expected_atol=9e-3)
@unittest.skip("Not supported in Wan VACE.")
def test_simple_inference_with_text_denoiser_block_scale(self):
pass
@unittest.skip("Not supported in Wan VACE.")
def test_simple_inference_with_text_denoiser_block_scale_for_all_dict_options(self):
pass
@require_peft_version_greater("0.13.2")
def test_lora_exclude_modules_wanvace(self):
exclude_module_name = "vace_blocks.0.proj_out"
components, text_lora_config, denoiser_lora_config = self.get_dummy_components()
pipe = self.pipeline_class(**components).to(torch_device)
_, _, inputs = self.get_dummy_inputs(with_generator=False)
output_no_lora = self.get_base_pipe_output()
self.assertTrue(output_no_lora.shape == self.output_shape)
# only supported for `denoiser` now
denoiser_lora_config.target_modules = ["proj_out"]
denoiser_lora_config.exclude_modules = [exclude_module_name]
pipe, _ = self.add_adapters_to_pipeline(
pipe, text_lora_config=text_lora_config, denoiser_lora_config=denoiser_lora_config
)
# The state dict shouldn't contain the modules to be excluded from LoRA.
state_dict_from_model = get_peft_model_state_dict(pipe.transformer, adapter_name="default")
self.assertTrue(not any(exclude_module_name in k for k in state_dict_from_model))
self.assertTrue(any("proj_out" in k for k in state_dict_from_model))
output_lora_exclude_modules = pipe(**inputs, generator=torch.manual_seed(0))[0]
with tempfile.TemporaryDirectory() as tmpdir:
modules_to_save = self._get_modules_to_save(pipe, has_denoiser=True)
lora_state_dicts = self._get_lora_state_dicts(modules_to_save)
self.pipeline_class.save_lora_weights(save_directory=tmpdir, **lora_state_dicts)
pipe.unload_lora_weights()
# Check in the loaded state dict.
loaded_state_dict = safetensors.torch.load_file(os.path.join(tmpdir, "pytorch_lora_weights.safetensors"))
self.assertTrue(not any(exclude_module_name in k for k in loaded_state_dict))
self.assertTrue(any("proj_out" in k for k in loaded_state_dict))
# Check in the state dict obtained after loading LoRA.
pipe.load_lora_weights(tmpdir)
state_dict_from_model = get_peft_model_state_dict(pipe.transformer, adapter_name="default_0")
self.assertTrue(not any(exclude_module_name in k for k in state_dict_from_model))
self.assertTrue(any("proj_out" in k for k in state_dict_from_model))
output_lora_pretrained = pipe(**inputs, generator=torch.manual_seed(0))[0]
self.assertTrue(
not np.allclose(output_no_lora, output_lora_exclude_modules, atol=1e-3, rtol=1e-3),
"LoRA should change outputs.",
)
self.assertTrue(
np.allclose(output_lora_exclude_modules, output_lora_pretrained, atol=1e-3, rtol=1e-3),
"Lora outputs should match.",
)
def test_simple_inference_with_text_denoiser_lora_and_scale(self):
super().test_simple_inference_with_text_denoiser_lora_and_scale()

63
tests/pipelines/wan/test_wan_vace.py
View File

@@ -13,7 +13,10 @@
# limitations under the License.
import os
import pytest
import safetensors.torch
import torch
from PIL import Image
from transformers import AutoConfig, AutoTokenizer, T5EncoderModel
@@ -25,9 +28,20 @@ from diffusers import (
WanVACEPipeline,
WanVACETransformer3DModel,
)
from diffusers.utils.import_utils import is_peft_available
from ...testing_utils import assert_tensors_close, torch_device
from ..testing_utils import BasePipelineTesterConfig, MemoryTesterMixin, PipelineTesterMixin
from ...testing_utils import assert_tensors_close, require_peft_version_greater, torch_device
from ..testing_utils import (
BasePipelineTesterConfig,
LoraMemoryTesterMixin,
LoraTesterMixin,
MemoryTesterMixin,
PipelineTesterMixin,
)
if is_peft_available():
from peft.utils import get_peft_model_state_dict
class WanVACEPipelineTesterConfig(BasePipelineTesterConfig):
@@ -242,3 +256,48 @@ class TestWanVACEPipeline(WanVACEPipelineTesterConfig, PipelineTesterMixin):
class TestWanVACEPipelineMemory(WanVACEPipelineTesterConfig, MemoryTesterMixin):
pass
class TestWanVACEPipelineLoRA(WanVACEPipelineTesterConfig, LoraTesterMixin):
"""LoRA tests for the Wan VACE pipeline."""
@require_peft_version_greater("0.13.2")
def test_lora_exclude_modules(self, tmp_path, base_pipe_output):
exclude_module_name = "vace_blocks.0.proj_out"
pipe = self.get_pipeline().to(torch_device)
self.add_adapters_to_pipeline(
pipe, components=["transformer"], target_modules=["proj_out"], exclude_modules=[exclude_module_name]
)
# The state dict should not contain the modules excluded from LoRA.
state_dict_from_model = get_peft_model_state_dict(pipe.transformer, adapter_name="default")
assert not any(exclude_module_name in k for k in state_dict_from_model)
assert any("proj_out" in k for k in state_dict_from_model)
output_lora_exclude_modules = self.run_pipe(pipe)
denoiser_state_dict = get_peft_model_state_dict(pipe.transformer)
self.pipeline_class.save_lora_weights(tmp_path, transformer_lora_layers=denoiser_state_dict)
pipe.unload_lora_weights()
# Check in the saved state dict.
loaded_state_dict = safetensors.torch.load_file(os.path.join(tmp_path, "pytorch_lora_weights.safetensors"))
assert not any(exclude_module_name in k for k in loaded_state_dict)
assert any("proj_out" in k for k in loaded_state_dict)
# Check in the state dict obtained after loading LoRA.
pipe.load_lora_weights(tmp_path)
state_dict_from_model = get_peft_model_state_dict(pipe.transformer, adapter_name="default_0")
assert not any(exclude_module_name in k for k in state_dict_from_model)
assert any("proj_out" in k for k in state_dict_from_model)
output_lora_pretrained = self.run_pipe(pipe)
assert not torch.allclose(base_pipe_output, output_lora_exclude_modules, atol=1e-3, rtol=1e-3), (
"LoRA should change outputs."
)
assert torch.allclose(output_lora_exclude_modules, output_lora_pretrained, atol=1e-3, rtol=1e-3), (
"Lora outputs should match."
)
class TestWanVACEPipelineLoRAMemory(WanVACEPipelineTesterConfig, LoraMemoryTesterMixin):
"""LoRA offloading tests for the Wan VACE pipeline."""