Add LoRA test for Z-Image

Author: CalamitousFelicitousness

tests/test_lora_layers_z_image.py

@@ -0,0 +1,177 @@
+# coding=utf-8
+# Copyright 2025 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 unittest
+
+import torch
+from transformers import Qwen2Tokenizer, Qwen3Config, Qwen3Model
+
+from diffusers import (
+    AutoencoderKL,
+    FlowMatchEulerDiscreteScheduler,
+    ZImagePipeline,
+    ZImageTransformer2DModel,
+)
+
+from ..testing_utils import floats_tensor, require_peft_backend
+
+
+# Z-Image requires torch.use_deterministic_algorithms(False) due to complex64 RoPE operations
+os.environ["CUDA_LAUNCH_BLOCKING"] = "1"
+os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":16:8"
+torch.use_deterministic_algorithms(False)
+torch.backends.cudnn.deterministic = True
+torch.backends.cudnn.benchmark = False
+if hasattr(torch.backends, "cuda"):
+    torch.backends.cuda.matmul.allow_tf32 = False
+
+
+sys.path.append(".")
+
+from .utils import PeftLoraLoaderMixinTests  # noqa: E402
+
+
+@require_peft_backend
+class ZImageLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
+    pipeline_class = ZImagePipeline
+    scheduler_cls = FlowMatchEulerDiscreteScheduler
+    scheduler_kwargs = {}
+
+    transformer_kwargs = {
+        "all_patch_size": (2,),
+        "all_f_patch_size": (1,),
+        "in_channels": 16,
+        "dim": 32,
+        "n_layers": 2,
+        "n_refiner_layers": 1,
+        "n_heads": 2,
+        "n_kv_heads": 2,
+        "norm_eps": 1e-5,
+        "qk_norm": True,
+        "cap_feat_dim": 16,
+        "rope_theta": 256.0,
+        "t_scale": 1000.0,
+        "axes_dims": [8, 4, 4],
+        "axes_lens": [256, 32, 32],
+    }
+    transformer_cls = ZImageTransformer2DModel
+    vae_kwargs = {
+        "in_channels": 3,
+        "out_channels": 3,
+        "down_block_types": ["DownEncoderBlock2D", "DownEncoderBlock2D"],
+        "up_block_types": ["UpDecoderBlock2D", "UpDecoderBlock2D"],
+        "block_out_channels": [32, 64],
+        "layers_per_block": 1,
+        "latent_channels": 16,
+        "norm_num_groups": 32,
+        "sample_size": 32,
+        "scaling_factor": 0.3611,
+        "shift_factor": 0.1159,
+    }
+    vae_cls = AutoencoderKL
+    tokenizer_cls, tokenizer_id = Qwen2Tokenizer, "hf-internal-testing/tiny-random-Qwen2VLForConditionalGeneration"
+    text_encoder_cls, text_encoder_id = Qwen3Model, None  # Will be created inline
+    denoiser_target_modules = ["to_q", "to_k", "to_v", "to_out.0"]
+
+    @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": 32,
+            "width": 32,
+            "max_sequence_length": 16,
+            "output_type": "np",
+        }
+        if with_generator:
+            pipeline_inputs.update({"generator": generator})
+
+        return noise, input_ids, pipeline_inputs
+
+    def get_dummy_components(self, scheduler_cls=None, use_dora=False, lora_alpha=None):
+        # Override to create Qwen3Model inline since it doesn't have a pretrained tiny model
+        torch.manual_seed(0)
+        config = Qwen3Config(
+            hidden_size=16,
+            intermediate_size=16,
+            num_hidden_layers=2,
+            num_attention_heads=2,
+            num_key_value_heads=2,
+            vocab_size=151936,
+            max_position_embeddings=512,
+        )
+        text_encoder = Qwen3Model(config)
+        tokenizer = Qwen2Tokenizer.from_pretrained(self.tokenizer_id)
+
+        transformer = self.transformer_cls(**self.transformer_kwargs)
+        vae = self.vae_cls(**self.vae_kwargs)
+
+        if scheduler_cls is None:
+            scheduler_cls = self.scheduler_cls
+        scheduler = scheduler_cls(**self.scheduler_kwargs)
+
+        return {
+            "transformer": transformer,
+            "vae": vae,
+            "scheduler": scheduler,
+            "text_encoder": text_encoder,
+            "tokenizer": tokenizer,
+        }
+
+    @unittest.skip("Not supported in ZImage.")
+    def test_simple_inference_with_text_denoiser_block_scale(self):
+        pass
+
+    @unittest.skip("Not supported in ZImage.")
+    def test_simple_inference_with_text_denoiser_block_scale_for_all_dict_options(self):
+        pass
+
+    @unittest.skip("Not supported in ZImage.")
+    def test_modify_padding_mode(self):
+        pass
+
+    @unittest.skip("Text encoder LoRA is not supported in ZImage.")
+    def test_simple_inference_with_partial_text_lora(self):
+        pass
+
+    @unittest.skip("Text encoder LoRA is not supported in ZImage.")
+    def test_simple_inference_with_text_lora(self):
+        pass
+
+    @unittest.skip("Text encoder LoRA is not supported in ZImage.")
+    def test_simple_inference_with_text_lora_and_scale(self):
+        pass
+
+    @unittest.skip("Text encoder LoRA is not supported in ZImage.")
+    def test_simple_inference_with_text_lora_fused(self):
+        pass
+
+    @unittest.skip("Text encoder LoRA is not supported in ZImage.")
+    def test_simple_inference_with_text_lora_save_load(self):
+        pass