File size: 6,026 Bytes
dde5d93 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 |
# coding=utf-8
# Copyright 2024 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 transformers import AutoTokenizer, CLIPTextModel, CLIPTokenizer, T5EncoderModel
from diffusers import FlowMatchEulerDiscreteScheduler, FluxPipeline, FluxTransformer2DModel
from diffusers.utils.testing_utils import floats_tensor, is_peft_available, require_peft_backend, torch_device
if is_peft_available():
from peft.utils import get_peft_model_state_dict
sys.path.append(".")
from utils import PeftLoraLoaderMixinTests, check_if_lora_correctly_set # noqa: E402
@require_peft_backend
class FluxLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
pipeline_class = FluxPipeline
scheduler_cls = FlowMatchEulerDiscreteScheduler()
scheduler_kwargs = {}
uses_flow_matching = True
transformer_kwargs = {
"patch_size": 1,
"in_channels": 4,
"num_layers": 1,
"num_single_layers": 1,
"attention_head_dim": 16,
"num_attention_heads": 2,
"joint_attention_dim": 32,
"pooled_projection_dim": 32,
"axes_dims_rope": [4, 4, 8],
}
transformer_cls = FluxTransformer2DModel
vae_kwargs = {
"sample_size": 32,
"in_channels": 3,
"out_channels": 3,
"block_out_channels": (4,),
"layers_per_block": 1,
"latent_channels": 1,
"norm_num_groups": 1,
"use_quant_conv": False,
"use_post_quant_conv": False,
"shift_factor": 0.0609,
"scaling_factor": 1.5035,
}
has_two_text_encoders = True
tokenizer_cls, tokenizer_id = CLIPTokenizer, "peft-internal-testing/tiny-clip-text-2"
tokenizer_2_cls, tokenizer_2_id = AutoTokenizer, "hf-internal-testing/tiny-random-t5"
text_encoder_cls, text_encoder_id = CLIPTextModel, "peft-internal-testing/tiny-clip-text-2"
text_encoder_2_cls, text_encoder_2_id = T5EncoderModel, "hf-internal-testing/tiny-random-t5"
@property
def output_shape(self):
return (1, 8, 8, 3)
def get_dummy_inputs(self, with_generator=True):
batch_size = 1
sequence_length = 10
num_channels = 4
sizes = (32, 32)
generator = torch.manual_seed(0)
noise = floats_tensor((batch_size, num_channels) + sizes)
input_ids = torch.randint(1, sequence_length, size=(batch_size, sequence_length), generator=generator)
pipeline_inputs = {
"prompt": "A painting of a squirrel eating a burger",
"num_inference_steps": 4,
"guidance_scale": 0.0,
"height": 8,
"width": 8,
"output_type": "np",
}
if with_generator:
pipeline_inputs.update({"generator": generator})
return noise, input_ids, pipeline_inputs
def test_with_alpha_in_state_dict(self):
components, _, denoiser_lora_config = self.get_dummy_components(FlowMatchEulerDiscreteScheduler)
pipe = self.pipeline_class(**components)
pipe = pipe.to(torch_device)
pipe.set_progress_bar_config(disable=None)
_, _, inputs = self.get_dummy_inputs(with_generator=False)
output_no_lora = pipe(**inputs, generator=torch.manual_seed(0)).images
self.assertTrue(output_no_lora.shape == self.output_shape)
pipe.transformer.add_adapter(denoiser_lora_config)
self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in transformer")
images_lora = pipe(**inputs, generator=torch.manual_seed(0)).images
with tempfile.TemporaryDirectory() as tmpdirname:
denoiser_state_dict = get_peft_model_state_dict(pipe.transformer)
self.pipeline_class.save_lora_weights(tmpdirname, transformer_lora_layers=denoiser_state_dict)
self.assertTrue(os.path.isfile(os.path.join(tmpdirname, "pytorch_lora_weights.safetensors")))
pipe.unload_lora_weights()
pipe.load_lora_weights(os.path.join(tmpdirname, "pytorch_lora_weights.safetensors"))
# modify the state dict to have alpha values following
# https://huggingface.co/TheLastBen/Jon_Snow_Flux_LoRA/blob/main/jon_snow.safetensors
state_dict_with_alpha = safetensors.torch.load_file(
os.path.join(tmpdirname, "pytorch_lora_weights.safetensors")
)
alpha_dict = {}
for k, v in state_dict_with_alpha.items():
# only do for `transformer` and for the k projections -- should be enough to test.
if "transformer" in k and "to_k" in k and "lora_A" in k:
alpha_dict[f"{k}.alpha"] = float(torch.randint(10, 100, size=()))
state_dict_with_alpha.update(alpha_dict)
images_lora_from_pretrained = pipe(**inputs, generator=torch.manual_seed(0)).images
self.assertTrue(check_if_lora_correctly_set(pipe.transformer), "Lora not correctly set in denoiser")
pipe.unload_lora_weights()
pipe.load_lora_weights(state_dict_with_alpha)
images_lora_with_alpha = pipe(**inputs, generator=torch.manual_seed(0)).images
self.assertTrue(
np.allclose(images_lora, images_lora_from_pretrained, atol=1e-3, rtol=1e-3),
"Loading from saved checkpoints should give same results.",
)
self.assertFalse(np.allclose(images_lora_with_alpha, images_lora, atol=1e-3, rtol=1e-3))
|