up

* skip invalid test case for helio pipeline

Signed-off-by: Liu, Kaixuan <kaixuan.liu@intel.com>

* update skip reason

Signed-off-by: Liu, Kaixuan <kaixuan.liu@intel.com>

---------

Signed-off-by: Liu, Kaixuan <kaixuan.liu@intel.com>

docs/source/en/optimization/attention_backends.md

@@ -143,6 +143,7 @@ Refer to the table below for a complete list of available attention backends and
 | `flash_varlen` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | Variable length FlashAttention |
 | `flash_varlen_hub` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | Variable length FlashAttention from kernels |
 | `aiter` | [AI Tensor Engine for ROCm](https://github.com/ROCm/aiter) | FlashAttention for AMD ROCm |
+| `flash_4_hub` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | FlashAttention-4 |
 | `_flash_3` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | FlashAttention-3 |
 | `_flash_varlen_3` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | Variable length FlashAttention-3 |
 | `_flash_3_hub` | [FlashAttention](https://github.com/Dao-AILab/flash-attention) | FlashAttention-3 from kernels |

scripts/convert_sana_video_to_diffusers.py

@@ -12,6 +12,7 @@ from termcolor import colored
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 from diffusers import (
+    AutoencoderKLLTX2Video,
     AutoencoderKLWan,
     DPMSolverMultistepScheduler,
     FlowMatchEulerDiscreteScheduler,
@@ -24,7 +25,10 @@ from diffusers.utils.import_utils import is_accelerate_available
 
 CTX = init_empty_weights if is_accelerate_available else nullcontext
 
-ckpt_ids = ["Efficient-Large-Model/SANA-Video_2B_480p/checkpoints/SANA_Video_2B_480p.pth"]
+ckpt_ids = [
+    "Efficient-Large-Model/SANA-Video_2B_480p/checkpoints/SANA_Video_2B_480p.pth",
+    "Efficient-Large-Model/SANA-Video_2B_720p/checkpoints/SANA_Video_2B_720p_LTXVAE.pth",
+]
 # https://github.com/NVlabs/Sana/blob/main/inference_video_scripts/inference_sana_video.py
 
 
@@ -92,12 +96,22 @@ def main(args):
     if args.video_size == 480:
         sample_size = 30  # Wan-VAE: 8xp2 downsample factor
         patch_size = (1, 2, 2)
+        in_channels = 16
+        out_channels = 16
     elif args.video_size == 720:
-        sample_size = 22  # Wan-VAE: 32xp1 downsample factor
+        sample_size = 22  # DC-AE-V: 32xp1 downsample factor
         patch_size = (1, 1, 1)
+        in_channels = 32
+        out_channels = 32
     else:
         raise ValueError(f"Video size {args.video_size} is not supported.")
 
+    if args.vae_type == "ltx2":
+        sample_size = 22
+        patch_size = (1, 1, 1)
+        in_channels = 128
+        out_channels = 128
+
     for depth in range(layer_num):
         # Transformer blocks.
         converted_state_dict[f"transformer_blocks.{depth}.scale_shift_table"] = state_dict.pop(
@@ -182,8 +196,8 @@ def main(args):
     # Transformer
     with CTX():
         transformer_kwargs = {
-            "in_channels": 16,
-            "out_channels": 16,
+            "in_channels": in_channels,
+            "out_channels": out_channels,
             "num_attention_heads": 20,
             "attention_head_dim": 112,
             "num_layers": 20,
@@ -235,9 +249,12 @@ def main(args):
     else:
         print(colored(f"Saving the whole Pipeline containing {args.model_type}", "green", attrs=["bold"]))
         # VAE
-        vae = AutoencoderKLWan.from_pretrained(
-            "Wan-AI/Wan2.1-T2V-1.3B-Diffusers", subfolder="vae", torch_dtype=torch.float32
-        )
+        if args.vae_type == "ltx2":
+            vae_path = args.vae_path or "Lightricks/LTX-2"
+            vae = AutoencoderKLLTX2Video.from_pretrained(vae_path, subfolder="vae", torch_dtype=torch.float32)
+        else:
+            vae_path = args.vae_path or "Wan-AI/Wan2.1-T2V-1.3B-Diffusers"
+            vae = AutoencoderKLWan.from_pretrained(vae_path, subfolder="vae", torch_dtype=torch.float32)
 
         # Text Encoder
         text_encoder_model_path = "Efficient-Large-Model/gemma-2-2b-it"
@@ -314,7 +331,23 @@ if __name__ == "__main__":
         choices=["flow-dpm_solver", "flow-euler", "uni-pc"],
         help="Scheduler type to use.",
     )
-    parser.add_argument("--task", default="t2v", type=str, required=True, help="Task to convert, t2v or i2v.")
+    parser.add_argument(
+        "--vae_type",
+        default="wan",
+        type=str,
+        choices=["wan", "ltx2"],
+        help="VAE type to use for saving full pipeline (ltx2 uses patchify 1x1x1).",
+    )
+    parser.add_argument(
+        "--vae_path",
+        default=None,
+        type=str,
+        required=False,
+        help="Optional VAE path or repo id. If not set, a default is used per VAE type.",
+    )
+    parser.add_argument(
+        "--task", default="t2v", type=str, required=True, choices=["t2v", "i2v"], help="Task to convert, t2v or i2v."
+    )
     parser.add_argument("--dump_path", default=None, type=str, required=True, help="Path to the output pipeline.")
     parser.add_argument("--save_full_pipeline", action="store_true", help="save all the pipeline elements in one.")
     parser.add_argument("--dtype", default="fp32", type=str, choices=["fp32", "fp16", "bf16"], help="Weight dtype.")

src/diffusers/models/attention_dispatch.py

@@ -229,6 +229,7 @@ class AttentionBackendName(str, Enum):
     FLASH_HUB = "flash_hub"
     FLASH_VARLEN = "flash_varlen"
     FLASH_VARLEN_HUB = "flash_varlen_hub"
+    FLASH_4_HUB = "flash_4_hub"
     _FLASH_3 = "_flash_3"
     _FLASH_VARLEN_3 = "_flash_varlen_3"
     _FLASH_3_HUB = "_flash_3_hub"
@@ -358,6 +359,11 @@ _HUB_KERNELS_REGISTRY: dict["AttentionBackendName", _HubKernelConfig] = {
         function_attr="sageattn",
         version=1,
     ),
+    AttentionBackendName.FLASH_4_HUB: _HubKernelConfig(
+        repo_id="kernels-staging/flash-attn4",
+        function_attr="flash_attn_func",
+        version=0,
+    ),
 }
 
 
@@ -521,6 +527,7 @@ def _check_attention_backend_requirements(backend: AttentionBackendName) -> None
         AttentionBackendName._FLASH_3_HUB,
         AttentionBackendName._FLASH_3_VARLEN_HUB,
         AttentionBackendName.SAGE_HUB,
+        AttentionBackendName.FLASH_4_HUB,
     ]:
         if not is_kernels_available():
             raise RuntimeError(
@@ -2676,6 +2683,37 @@ def _flash_attention_3_varlen_hub(
     return (out, lse) if return_lse else out
 
 
+@_AttentionBackendRegistry.register(
+    AttentionBackendName.FLASH_4_HUB,
+    constraints=[_check_device, _check_qkv_dtype_bf16_or_fp16, _check_shape],
+    supports_context_parallel=False,
+)
+def _flash_attention_4_hub(
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attn_mask: torch.Tensor | None = None,
+    scale: float | None = None,
+    is_causal: bool = False,
+    return_lse: bool = False,
+    _parallel_config: "ParallelConfig" | None = None,
+) -> torch.Tensor:
+    if attn_mask is not None:
+        raise ValueError("`attn_mask` is not supported for flash-attn 4.")
+
+    func = _HUB_KERNELS_REGISTRY[AttentionBackendName.FLASH_4_HUB].kernel_fn
+    out = func(
+        q=query,
+        k=key,
+        v=value,
+        softmax_scale=scale,
+        causal=is_causal,
+    )
+    if isinstance(out, tuple):
+        return (out[0], out[1]) if return_lse else out[0]
+    return out
+
+
 @_AttentionBackendRegistry.register(
     AttentionBackendName._FLASH_VARLEN_3,
     constraints=[_check_device, _check_qkv_dtype_bf16_or_fp16, _check_shape],

src/diffusers/pipelines/sana_video/pipeline_sana_video.py

@@ -24,7 +24,7 @@ from transformers import Gemma2PreTrainedModel, GemmaTokenizer, GemmaTokenizerFa
 
 from ...callbacks import MultiPipelineCallbacks, PipelineCallback
 from ...loaders import SanaLoraLoaderMixin
-from ...models import AutoencoderDC, AutoencoderKLWan, SanaVideoTransformer3DModel
+from ...models import AutoencoderDC, AutoencoderKLLTX2Video, AutoencoderKLWan, SanaVideoTransformer3DModel
 from ...schedulers import DPMSolverMultistepScheduler
 from ...utils import (
     BACKENDS_MAPPING,
@@ -194,7 +194,7 @@ class SanaVideoPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
             The tokenizer used to tokenize the prompt.
         text_encoder ([`Gemma2PreTrainedModel`]):
             Text encoder model to encode the input prompts.
-        vae ([`AutoencoderKLWan` or `AutoencoderDCAEV`]):
+        vae ([`AutoencoderKLWan`, `AutoencoderDC`, or `AutoencoderKLLTX2Video`]):
             Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
         transformer ([`SanaVideoTransformer3DModel`]):
             Conditional Transformer to denoise the input latents.
@@ -213,7 +213,7 @@ class SanaVideoPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
         self,
         tokenizer: GemmaTokenizer | GemmaTokenizerFast,
         text_encoder: Gemma2PreTrainedModel,
-        vae: AutoencoderDC | AutoencoderKLWan,
+        vae: AutoencoderDC | AutoencoderKLLTX2Video | AutoencoderKLWan,
         transformer: SanaVideoTransformer3DModel,
         scheduler: DPMSolverMultistepScheduler,
     ):
@@ -223,8 +223,19 @@ class SanaVideoPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
             tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
         )
 
-        self.vae_scale_factor_temporal = self.vae.config.scale_factor_temporal if getattr(self, "vae", None) else 4
-        self.vae_scale_factor_spatial = self.vae.config.scale_factor_spatial if getattr(self, "vae", None) else 8
+        if getattr(self, "vae", None):
+            if isinstance(self.vae, AutoencoderKLLTX2Video):
+                self.vae_scale_factor_temporal = self.vae.config.temporal_compression_ratio
+                self.vae_scale_factor_spatial = self.vae.config.spatial_compression_ratio
+            elif isinstance(self.vae, (AutoencoderDC, AutoencoderKLWan)):
+                self.vae_scale_factor_temporal = self.vae.config.scale_factor_temporal
+                self.vae_scale_factor_spatial = self.vae.config.scale_factor_spatial
+            else:
+                self.vae_scale_factor_temporal = 4
+                self.vae_scale_factor_spatial = 8
+        else:
+            self.vae_scale_factor_temporal = 4
+            self.vae_scale_factor_spatial = 8
 
         self.vae_scale_factor = self.vae_scale_factor_spatial
 
@@ -985,14 +996,21 @@ class SanaVideoPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
                 if is_torch_version(">=", "2.5.0")
                 else torch_accelerator_module.OutOfMemoryError
             )
-            latents_mean = (
-                torch.tensor(self.vae.config.latents_mean)
-                .view(1, self.vae.config.z_dim, 1, 1, 1)
-                .to(latents.device, latents.dtype)
-            )
-            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
-                latents.device, latents.dtype
-            )
+            if isinstance(self.vae, AutoencoderKLLTX2Video):
+                latents_mean = self.vae.latents_mean
+                latents_std = self.vae.latents_std
+                z_dim = self.vae.config.latent_channels
+            elif isinstance(self.vae, AutoencoderKLWan):
+                latents_mean = torch.tensor(self.vae.config.latents_mean)
+                latents_std = torch.tensor(self.vae.config.latents_std)
+                z_dim = self.vae.config.z_dim
+            else:
+                latents_mean = torch.zeros(latents.shape[1], device=latents.device, dtype=latents.dtype)
+                latents_std = torch.ones(latents.shape[1], device=latents.device, dtype=latents.dtype)
+                z_dim = latents.shape[1]
+
+            latents_mean = latents_mean.view(1, z_dim, 1, 1, 1).to(latents.device, latents.dtype)
+            latents_std = 1.0 / latents_std.view(1, z_dim, 1, 1, 1).to(latents.device, latents.dtype)
             latents = latents / latents_std + latents_mean
             try:
                 video = self.vae.decode(latents, return_dict=False)[0]

src/diffusers/pipelines/sana_video/pipeline_sana_video_i2v.py

@@ -26,7 +26,7 @@ from transformers import Gemma2PreTrainedModel, GemmaTokenizer, GemmaTokenizerFa
 from ...callbacks import MultiPipelineCallbacks, PipelineCallback
 from ...image_processor import PipelineImageInput
 from ...loaders import SanaLoraLoaderMixin
-from ...models import AutoencoderDC, AutoencoderKLWan, SanaVideoTransformer3DModel
+from ...models import AutoencoderDC, AutoencoderKLLTX2Video, AutoencoderKLWan, SanaVideoTransformer3DModel
 from ...schedulers import FlowMatchEulerDiscreteScheduler
 from ...utils import (
     BACKENDS_MAPPING,
@@ -184,7 +184,7 @@ class SanaImageToVideoPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
             The tokenizer used to tokenize the prompt.
         text_encoder ([`Gemma2PreTrainedModel`]):
             Text encoder model to encode the input prompts.
-        vae ([`AutoencoderKLWan` or `AutoencoderDCAEV`]):
+        vae ([`AutoencoderKLWan`, `AutoencoderDC`, or `AutoencoderKLLTX2Video`]):
             Variational Auto-Encoder (VAE) Model to encode and decode videos to and from latent representations.
         transformer ([`SanaVideoTransformer3DModel`]):
             Conditional Transformer to denoise the input latents.
@@ -203,7 +203,7 @@ class SanaImageToVideoPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
         self,
         tokenizer: GemmaTokenizer | GemmaTokenizerFast,
         text_encoder: Gemma2PreTrainedModel,
-        vae: AutoencoderDC | AutoencoderKLWan,
+        vae: AutoencoderDC | AutoencoderKLLTX2Video | AutoencoderKLWan,
         transformer: SanaVideoTransformer3DModel,
         scheduler: FlowMatchEulerDiscreteScheduler,
     ):
@@ -213,8 +213,19 @@ class SanaImageToVideoPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
             tokenizer=tokenizer, text_encoder=text_encoder, vae=vae, transformer=transformer, scheduler=scheduler
         )
 
-        self.vae_scale_factor_temporal = self.vae.config.scale_factor_temporal if getattr(self, "vae", None) else 4
-        self.vae_scale_factor_spatial = self.vae.config.scale_factor_spatial if getattr(self, "vae", None) else 8
+        if getattr(self, "vae", None):
+            if isinstance(self.vae, AutoencoderKLLTX2Video):
+                self.vae_scale_factor_temporal = self.vae.config.temporal_compression_ratio
+                self.vae_scale_factor_spatial = self.vae.config.spatial_compression_ratio
+            elif isinstance(self.vae, (AutoencoderDC, AutoencoderKLWan)):
+                self.vae_scale_factor_temporal = self.vae.config.scale_factor_temporal
+                self.vae_scale_factor_spatial = self.vae.config.scale_factor_spatial
+            else:
+                self.vae_scale_factor_temporal = 4
+                self.vae_scale_factor_spatial = 8
+        else:
+            self.vae_scale_factor_temporal = 4
+            self.vae_scale_factor_spatial = 8
 
         self.vae_scale_factor = self.vae_scale_factor_spatial
 
@@ -687,14 +698,18 @@ class SanaImageToVideoPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
             image_latents = retrieve_latents(self.vae.encode(image), sample_mode="argmax")
             image_latents = image_latents.repeat(batch_size, 1, 1, 1, 1)
 
-        latents_mean = (
-            torch.tensor(self.vae.config.latents_mean)
-            .view(1, -1, 1, 1, 1)
-            .to(image_latents.device, image_latents.dtype)
-        )
-        latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, -1, 1, 1, 1).to(
-            image_latents.device, image_latents.dtype
-        )
+        if isinstance(self.vae, AutoencoderKLLTX2Video):
+            _latents_mean = self.vae.latents_mean
+            _latents_std = self.vae.latents_std
+        elif isinstance(self.vae, AutoencoderKLWan):
+            _latents_mean = torch.tensor(self.vae.config.latents_mean)
+            _latents_std = torch.tensor(self.vae.config.latents_std)
+        else:
+            _latents_mean = torch.zeros(image_latents.shape[1], device=image_latents.device, dtype=image_latents.dtype)
+            _latents_std = torch.ones(image_latents.shape[1], device=image_latents.device, dtype=image_latents.dtype)
+
+        latents_mean = _latents_mean.view(1, -1, 1, 1, 1).to(image_latents.device, image_latents.dtype)
+        latents_std = 1.0 / _latents_std.view(1, -1, 1, 1, 1).to(image_latents.device, image_latents.dtype)
         image_latents = (image_latents - latents_mean) * latents_std
 
         latents[:, :, 0:1] = image_latents.to(dtype)
@@ -1034,14 +1049,21 @@ class SanaImageToVideoPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
                 if is_torch_version(">=", "2.5.0")
                 else torch_accelerator_module.OutOfMemoryError
             )
-            latents_mean = (
-                torch.tensor(self.vae.config.latents_mean)
-                .view(1, self.vae.config.z_dim, 1, 1, 1)
-                .to(latents.device, latents.dtype)
-            )
-            latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
-                latents.device, latents.dtype
-            )
+            if isinstance(self.vae, AutoencoderKLLTX2Video):
+                latents_mean = self.vae.latents_mean
+                latents_std = self.vae.latents_std
+                z_dim = self.vae.config.latent_channels
+            elif isinstance(self.vae, AutoencoderKLWan):
+                latents_mean = torch.tensor(self.vae.config.latents_mean)
+                latents_std = torch.tensor(self.vae.config.latents_std)
+                z_dim = self.vae.config.z_dim
+            else:
+                latents_mean = torch.zeros(latents.shape[1], device=latents.device, dtype=latents.dtype)
+                latents_std = torch.ones(latents.shape[1], device=latents.device, dtype=latents.dtype)
+                z_dim = latents.shape[1]
+
+            latents_mean = latents_mean.view(1, z_dim, 1, 1, 1).to(latents.device, latents.dtype)
+            latents_std = 1.0 / latents_std.view(1, z_dim, 1, 1, 1).to(latents.device, latents.dtype)
             latents = latents / latents_std + latents_mean
             try:
                 video = self.vae.decode(latents, return_dict=False)[0]

tests/models/transformers/test_models_transformer_qwenimage.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2025 HuggingFace Inc.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -13,49 +12,84 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-import unittest
+import warnings
 
 import torch
 
 from diffusers import QwenImageTransformer2DModel
 from diffusers.models.transformers.transformer_qwenimage import compute_text_seq_len_from_mask
+from diffusers.utils.torch_utils import randn_tensor
 
 from ...testing_utils import enable_full_determinism, torch_device
-from ..test_modeling_common import ModelTesterMixin, TorchCompileTesterMixin
+from ..testing_utils import (
+    AttentionTesterMixin,
+    BaseModelTesterConfig,
+    BitsAndBytesTesterMixin,
+    ContextParallelTesterMixin,
+    LoraHotSwappingForModelTesterMixin,
+    LoraTesterMixin,
+    MemoryTesterMixin,
+    ModelTesterMixin,
+    TorchAoTesterMixin,
+    TorchCompileTesterMixin,
+    TrainingTesterMixin,
+)
 
 
 enable_full_determinism()
 
 
-class QwenImageTransformerTests(ModelTesterMixin, unittest.TestCase):
-    model_class = QwenImageTransformer2DModel
-    main_input_name = "hidden_states"
-    # We override the items here because the transformer under consideration is small.
-    model_split_percents = [0.7, 0.6, 0.6]
-
-    # Skip setting testing with default: AttnProcessor
-    uses_custom_attn_processor = True
+class QwenImageTransformerTesterConfig(BaseModelTesterConfig):
+    @property
+    def model_class(self):
+        return QwenImageTransformer2DModel
 
     @property
-    def dummy_input(self):
-        return self.prepare_dummy_input()
-
-    @property
-    def input_shape(self):
+    def output_shape(self) -> tuple[int, int]:
         return (16, 16)
 
     @property
-    def output_shape(self):
+    def input_shape(self) -> tuple[int, int]:
         return (16, 16)
 
-    def prepare_dummy_input(self, height=4, width=4):
+    @property
+    def model_split_percents(self) -> list:
+        return [0.7, 0.6, 0.6]
+
+    @property
+    def main_input_name(self) -> str:
+        return "hidden_states"
+
+    @property
+    def generator(self):
+        return torch.Generator("cpu").manual_seed(0)
+
+    def get_init_dict(self) -> dict[str, int | list[int]]:
+        return {
+            "patch_size": 2,
+            "in_channels": 16,
+            "out_channels": 4,
+            "num_layers": 2,
+            "attention_head_dim": 16,
+            "num_attention_heads": 4,
+            "joint_attention_dim": 16,
+            "guidance_embeds": False,
+            "axes_dims_rope": (8, 4, 4),
+        }
+
+    def get_dummy_inputs(self) -> dict[str, torch.Tensor]:
         batch_size = 1
         num_latent_channels = embedding_dim = 16
-        sequence_length = 7
+        height = width = 4
+        sequence_length = 8
         vae_scale_factor = 4
 
-        hidden_states = torch.randn((batch_size, height * width, num_latent_channels)).to(torch_device)
-        encoder_hidden_states = torch.randn((batch_size, sequence_length, embedding_dim)).to(torch_device)
+        hidden_states = randn_tensor(
+            (batch_size, height * width, num_latent_channels), generator=self.generator, device=torch_device
+        )
+        encoder_hidden_states = randn_tensor(
+            (batch_size, sequence_length, embedding_dim), generator=self.generator, device=torch_device
+        )
         encoder_hidden_states_mask = torch.ones((batch_size, sequence_length)).to(torch_device, torch.long)
         timestep = torch.tensor([1.0]).to(torch_device).expand(batch_size)
         orig_height = height * 2 * vae_scale_factor
@@ -70,89 +104,57 @@ class QwenImageTransformerTests(ModelTesterMixin, unittest.TestCase):
             "img_shapes": img_shapes,
         }
 
-    def prepare_init_args_and_inputs_for_common(self):
-        init_dict = {
-            "patch_size": 2,
-            "in_channels": 16,
-            "out_channels": 4,
-            "num_layers": 2,
-            "attention_head_dim": 16,
-            "num_attention_heads": 3,
-            "joint_attention_dim": 16,
-            "guidance_embeds": False,
-            "axes_dims_rope": (8, 4, 4),
-        }
-
-        inputs_dict = self.dummy_input
-        return init_dict, inputs_dict
-
-    def test_gradient_checkpointing_is_applied(self):
-        expected_set = {"QwenImageTransformer2DModel"}
-        super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
 
+class TestQwenImageTransformer(QwenImageTransformerTesterConfig, ModelTesterMixin):
     def test_infers_text_seq_len_from_mask(self):
-        """Test that compute_text_seq_len_from_mask correctly infers sequence lengths and returns tensors."""
-        init_dict, inputs = self.prepare_init_args_and_inputs_for_common()
+        init_dict = self.get_init_dict()
+        inputs = self.get_dummy_inputs()
         model = self.model_class(**init_dict).to(torch_device)
 
-        # Test 1: Contiguous mask with padding at the end (only first 2 tokens valid)
         encoder_hidden_states_mask = inputs["encoder_hidden_states_mask"].clone()
-        encoder_hidden_states_mask[:, 2:] = 0  # Only first 2 tokens are valid
+        encoder_hidden_states_mask[:, 2:] = 0
 
         rope_text_seq_len, per_sample_len, normalized_mask = compute_text_seq_len_from_mask(
             inputs["encoder_hidden_states"], encoder_hidden_states_mask
         )
 
-        # Verify rope_text_seq_len is returned as an int (for torch.compile compatibility)
-        self.assertIsInstance(rope_text_seq_len, int)
+        assert isinstance(rope_text_seq_len, int)
+        assert isinstance(per_sample_len, torch.Tensor)
+        assert int(per_sample_len.max().item()) == 2
+        assert normalized_mask.dtype == torch.bool
+        assert normalized_mask.sum().item() == 2
+        assert rope_text_seq_len >= inputs["encoder_hidden_states"].shape[1]
 
-        # Verify per_sample_len is computed correctly (max valid position + 1 = 2)
-        self.assertIsInstance(per_sample_len, torch.Tensor)
-        self.assertEqual(int(per_sample_len.max().item()), 2)
-
-        # Verify mask is normalized to bool dtype
-        self.assertTrue(normalized_mask.dtype == torch.bool)
-        self.assertEqual(normalized_mask.sum().item(), 2)  # Only 2 True values
-
-        # Verify rope_text_seq_len is at least the sequence length
-        self.assertGreaterEqual(rope_text_seq_len, inputs["encoder_hidden_states"].shape[1])
-
-        # Test 2: Verify model runs successfully with inferred values
         inputs["encoder_hidden_states_mask"] = normalized_mask
         with torch.no_grad():
             output = model(**inputs)
-        self.assertEqual(output.sample.shape[1], inputs["hidden_states"].shape[1])
+        assert output.sample.shape[1] == inputs["hidden_states"].shape[1]
 
-        # Test 3: Different mask pattern (padding at beginning)
         encoder_hidden_states_mask2 = inputs["encoder_hidden_states_mask"].clone()
-        encoder_hidden_states_mask2[:, :3] = 0  # First 3 tokens are padding
-        encoder_hidden_states_mask2[:, 3:] = 1  # Last 4 tokens are valid
+        encoder_hidden_states_mask2[:, :3] = 0
+        encoder_hidden_states_mask2[:, 3:] = 1
 
         rope_text_seq_len2, per_sample_len2, normalized_mask2 = compute_text_seq_len_from_mask(
             inputs["encoder_hidden_states"], encoder_hidden_states_mask2
         )
 
-        # Max valid position is 6 (last token), so per_sample_len should be 7
-        self.assertEqual(int(per_sample_len2.max().item()), 7)
-        self.assertEqual(normalized_mask2.sum().item(), 4)  # 4 True values
+        assert int(per_sample_len2.max().item()) == 8
+        assert normalized_mask2.sum().item() == 5
 
-        # Test 4: No mask provided (None case)
         rope_text_seq_len_none, per_sample_len_none, normalized_mask_none = compute_text_seq_len_from_mask(
             inputs["encoder_hidden_states"], None
         )
-        self.assertEqual(rope_text_seq_len_none, inputs["encoder_hidden_states"].shape[1])
-        self.assertIsInstance(rope_text_seq_len_none, int)
-        self.assertIsNone(per_sample_len_none)
-        self.assertIsNone(normalized_mask_none)
+        assert rope_text_seq_len_none == inputs["encoder_hidden_states"].shape[1]
+        assert isinstance(rope_text_seq_len_none, int)
+        assert per_sample_len_none is None
+        assert normalized_mask_none is None
 
     def test_non_contiguous_attention_mask(self):
-        """Test that non-contiguous masks work correctly (e.g., [1, 0, 1, 0, 1, 0, 0])"""
-        init_dict, inputs = self.prepare_init_args_and_inputs_for_common()
+        init_dict = self.get_init_dict()
+        inputs = self.get_dummy_inputs()
         model = self.model_class(**init_dict).to(torch_device)
 
-        # Create a non-contiguous mask pattern: valid, padding, valid, padding, etc.
         encoder_hidden_states_mask = inputs["encoder_hidden_states_mask"].clone()
-        # Pattern: [True, False, True, False, True, False, False]
         encoder_hidden_states_mask[:, 1] = 0
         encoder_hidden_states_mask[:, 3] = 0
         encoder_hidden_states_mask[:, 5:] = 0
@@ -160,95 +162,85 @@ class QwenImageTransformerTests(ModelTesterMixin, unittest.TestCase):
         inferred_rope_len, per_sample_len, normalized_mask = compute_text_seq_len_from_mask(
             inputs["encoder_hidden_states"], encoder_hidden_states_mask
         )
-        self.assertEqual(int(per_sample_len.max().item()), 5)
-        self.assertEqual(inferred_rope_len, inputs["encoder_hidden_states"].shape[1])
-        self.assertIsInstance(inferred_rope_len, int)
-        self.assertTrue(normalized_mask.dtype == torch.bool)
+        assert int(per_sample_len.max().item()) == 5
+        assert inferred_rope_len == inputs["encoder_hidden_states"].shape[1]
+        assert isinstance(inferred_rope_len, int)
+        assert normalized_mask.dtype == torch.bool
 
         inputs["encoder_hidden_states_mask"] = normalized_mask
 
         with torch.no_grad():
             output = model(**inputs)
 
-        self.assertEqual(output.sample.shape[1], inputs["hidden_states"].shape[1])
+        assert output.sample.shape[1] == inputs["hidden_states"].shape[1]
 
     def test_txt_seq_lens_deprecation(self):
-        """Test that passing txt_seq_lens raises a deprecation warning."""
-        init_dict, inputs = self.prepare_init_args_and_inputs_for_common()
+        init_dict = self.get_init_dict()
+        inputs = self.get_dummy_inputs()
         model = self.model_class(**init_dict).to(torch_device)
 
-        # Prepare inputs with txt_seq_lens (deprecated parameter)
         txt_seq_lens = [inputs["encoder_hidden_states"].shape[1]]
 
-        # Remove encoder_hidden_states_mask to use the deprecated path
         inputs_with_deprecated = inputs.copy()
         inputs_with_deprecated.pop("encoder_hidden_states_mask")
         inputs_with_deprecated["txt_seq_lens"] = txt_seq_lens
 
-        # Test that deprecation warning is raised
-        with self.assertWarns(FutureWarning) as warning_context:
+        with warnings.catch_warnings(record=True) as w:
+            warnings.simplefilter("always")
             with torch.no_grad():
                 output = model(**inputs_with_deprecated)
 
-        # Verify the warning message mentions the deprecation
-        warning_message = str(warning_context.warning)
-        self.assertIn("txt_seq_lens", warning_message)
-        self.assertIn("deprecated", warning_message)
-        self.assertIn("encoder_hidden_states_mask", warning_message)
+            future_warnings = [x for x in w if issubclass(x.category, FutureWarning)]
+            assert len(future_warnings) > 0, "Expected FutureWarning to be raised"
 
-        # Verify the model still works correctly despite the deprecation
-        self.assertEqual(output.sample.shape[1], inputs["hidden_states"].shape[1])
+            warning_message = str(future_warnings[0].message)
+            assert "txt_seq_lens" in warning_message
+            assert "deprecated" in warning_message
+
+        assert output.sample.shape[1] == inputs["hidden_states"].shape[1]
 
     def test_layered_model_with_mask(self):
-        """Test QwenImageTransformer2DModel with use_layer3d_rope=True (layered model)."""
-        # Create layered model config
         init_dict = {
             "patch_size": 2,
             "in_channels": 16,
             "out_channels": 4,
             "num_layers": 2,
             "attention_head_dim": 16,
-            "num_attention_heads": 3,
+            "num_attention_heads": 4,
             "joint_attention_dim": 16,
-            "axes_dims_rope": (8, 4, 4),  # Must match attention_head_dim (8+4+4=16)
-            "use_layer3d_rope": True,  # Enable layered RoPE
-            "use_additional_t_cond": True,  # Enable additional time conditioning
+            "axes_dims_rope": (8, 4, 4),
+            "use_layer3d_rope": True,
+            "use_additional_t_cond": True,
         }
 
         model = self.model_class(**init_dict).to(torch_device)
 
-        # Verify the model uses QwenEmbedLayer3DRope
         from diffusers.models.transformers.transformer_qwenimage import QwenEmbedLayer3DRope
 
-        self.assertIsInstance(model.pos_embed, QwenEmbedLayer3DRope)
+        assert isinstance(model.pos_embed, QwenEmbedLayer3DRope)
 
-        # Test single generation with layered structure
         batch_size = 1
-        text_seq_len = 7
+        text_seq_len = 8
         img_h, img_w = 4, 4
         layers = 4
 
-        # For layered model: (layers + 1) because we have N layers + 1 combined image
         hidden_states = torch.randn(batch_size, (layers + 1) * img_h * img_w, 16).to(torch_device)
         encoder_hidden_states = torch.randn(batch_size, text_seq_len, 16).to(torch_device)
 
-        # Create mask with some padding
         encoder_hidden_states_mask = torch.ones(batch_size, text_seq_len).to(torch_device)
-        encoder_hidden_states_mask[0, 5:] = 0  # Only 5 valid tokens
+        encoder_hidden_states_mask[0, 5:] = 0
 
         timestep = torch.tensor([1.0]).to(torch_device)
 
-        # additional_t_cond for use_additional_t_cond=True (0 or 1 index for embedding)
         addition_t_cond = torch.tensor([0], dtype=torch.long).to(torch_device)
 
-        # Layer structure: 4 layers + 1 condition image
         img_shapes = [
             [
-                (1, img_h, img_w),  # layer 0
-                (1, img_h, img_w),  # layer 1
-                (1, img_h, img_w),  # layer 2
-                (1, img_h, img_w),  # layer 3
-                (1, img_h, img_w),  # condition image (last one gets special treatment)
+                (1, img_h, img_w),
+                (1, img_h, img_w),
+                (1, img_h, img_w),
+                (1, img_h, img_w),
+                (1, img_h, img_w),
             ]
         ]
 
@@ -262,37 +254,113 @@ class QwenImageTransformerTests(ModelTesterMixin, unittest.TestCase):
                 additional_t_cond=addition_t_cond,
             )
 
-        self.assertEqual(output.sample.shape[1], hidden_states.shape[1])
+        assert output.sample.shape[1] == hidden_states.shape[1]
 
 
-class QwenImageTransformerCompileTests(TorchCompileTesterMixin, unittest.TestCase):
-    model_class = QwenImageTransformer2DModel
+class TestQwenImageTransformerMemory(QwenImageTransformerTesterConfig, MemoryTesterMixin):
+    """Memory optimization tests for QwenImage Transformer."""
 
-    def prepare_init_args_and_inputs_for_common(self):
-        return QwenImageTransformerTests().prepare_init_args_and_inputs_for_common()
 
-    def prepare_dummy_input(self, height, width):
-        return QwenImageTransformerTests().prepare_dummy_input(height=height, width=width)
+class TestQwenImageTransformerTraining(QwenImageTransformerTesterConfig, TrainingTesterMixin):
+    """Training tests for QwenImage Transformer."""
 
-    def test_torch_compile_recompilation_and_graph_break(self):
-        super().test_torch_compile_recompilation_and_graph_break()
+    def test_gradient_checkpointing_is_applied(self):
+        expected_set = {"QwenImageTransformer2DModel"}
+        super().test_gradient_checkpointing_is_applied(expected_set=expected_set)
+
+
+class TestQwenImageTransformerAttention(QwenImageTransformerTesterConfig, AttentionTesterMixin):
+    """Attention processor tests for QwenImage Transformer."""
+
+
+class TestQwenImageTransformerContextParallel(QwenImageTransformerTesterConfig, ContextParallelTesterMixin):
+    """Context Parallel inference tests for QwenImage Transformer."""
+
+
+class TestQwenImageTransformerLoRA(QwenImageTransformerTesterConfig, LoraTesterMixin):
+    """LoRA adapter tests for QwenImage Transformer."""
+
+
+class TestQwenImageTransformerLoRAHotSwap(QwenImageTransformerTesterConfig, LoraHotSwappingForModelTesterMixin):
+    """LoRA hot-swapping tests for QwenImage Transformer."""
+
+    @property
+    def different_shapes_for_compilation(self):
+        return [(4, 4), (4, 8), (8, 8)]
+
+    def get_dummy_inputs(self, height: int = 4, width: int = 4) -> dict[str, torch.Tensor]:
+        batch_size = 1
+        num_latent_channels = embedding_dim = 16
+        sequence_length = 8
+        vae_scale_factor = 4
+
+        hidden_states = randn_tensor(
+            (batch_size, height * width, num_latent_channels), generator=self.generator, device=torch_device
+        )
+        encoder_hidden_states = randn_tensor(
+            (batch_size, sequence_length, embedding_dim), generator=self.generator, device=torch_device
+        )
+        encoder_hidden_states_mask = torch.ones((batch_size, sequence_length)).to(torch_device, torch.long)
+        timestep = torch.tensor([1.0]).to(torch_device).expand(batch_size)
+        orig_height = height * 2 * vae_scale_factor
+        orig_width = width * 2 * vae_scale_factor
+        img_shapes = [(1, orig_height // vae_scale_factor // 2, orig_width // vae_scale_factor // 2)] * batch_size
+
+        return {
+            "hidden_states": hidden_states,
+            "encoder_hidden_states": encoder_hidden_states,
+            "encoder_hidden_states_mask": encoder_hidden_states_mask,
+            "timestep": timestep,
+            "img_shapes": img_shapes,
+        }
+
+
+class TestQwenImageTransformerCompile(QwenImageTransformerTesterConfig, TorchCompileTesterMixin):
+    """Torch compile tests for QwenImage Transformer."""
+
+    @property
+    def different_shapes_for_compilation(self):
+        return [(4, 4), (4, 8), (8, 8)]
+
+    def get_dummy_inputs(self, height: int = 4, width: int = 4) -> dict[str, torch.Tensor]:
+        batch_size = 1
+        num_latent_channels = embedding_dim = 16
+        sequence_length = 8
+        vae_scale_factor = 4
+
+        hidden_states = randn_tensor(
+            (batch_size, height * width, num_latent_channels), generator=self.generator, device=torch_device
+        )
+        encoder_hidden_states = randn_tensor(
+            (batch_size, sequence_length, embedding_dim), generator=self.generator, device=torch_device
+        )
+        encoder_hidden_states_mask = torch.ones((batch_size, sequence_length)).to(torch_device, torch.long)
+        timestep = torch.tensor([1.0]).to(torch_device).expand(batch_size)
+        orig_height = height * 2 * vae_scale_factor
+        orig_width = width * 2 * vae_scale_factor
+        img_shapes = [(1, orig_height // vae_scale_factor // 2, orig_width // vae_scale_factor // 2)] * batch_size
+
+        return {
+            "hidden_states": hidden_states,
+            "encoder_hidden_states": encoder_hidden_states,
+            "encoder_hidden_states_mask": encoder_hidden_states_mask,
+            "timestep": timestep,
+            "img_shapes": img_shapes,
+        }
 
     def test_torch_compile_with_and_without_mask(self):
-        """Test that torch.compile works with both None mask and padding mask."""
-        init_dict, inputs = self.prepare_init_args_and_inputs_for_common()
+        init_dict = self.get_init_dict()
+        inputs = self.get_dummy_inputs()
         model = self.model_class(**init_dict).to(torch_device)
         model.eval()
         model.compile(mode="default", fullgraph=True)
 
-        # Test 1: Run with None mask (no padding, all tokens are valid)
         inputs_no_mask = inputs.copy()
         inputs_no_mask["encoder_hidden_states_mask"] = None
 
-        # First run to allow compilation
         with torch.no_grad():
             output_no_mask = model(**inputs_no_mask)
 
-        # Second run to verify no recompilation
         with (
             torch._inductor.utils.fresh_inductor_cache(),
             torch._dynamo.config.patch(error_on_recompile=True),
@@ -300,19 +368,15 @@ class QwenImageTransformerCompileTests(TorchCompileTesterMixin, unittest.TestCas
         ):
             output_no_mask_2 = model(**inputs_no_mask)
 
-        self.assertEqual(output_no_mask.sample.shape[1], inputs["hidden_states"].shape[1])
-        self.assertEqual(output_no_mask_2.sample.shape[1], inputs["hidden_states"].shape[1])
+        assert output_no_mask.sample.shape[1] == inputs["hidden_states"].shape[1]
+        assert output_no_mask_2.sample.shape[1] == inputs["hidden_states"].shape[1]
 
-        # Test 2: Run with all-ones mask (should behave like None)
         inputs_all_ones = inputs.copy()
-        # Keep the all-ones mask
-        self.assertTrue(inputs_all_ones["encoder_hidden_states_mask"].all().item())
+        assert inputs_all_ones["encoder_hidden_states_mask"].all().item()
 
-        # First run to allow compilation
         with torch.no_grad():
             output_all_ones = model(**inputs_all_ones)
 
-        # Second run to verify no recompilation
         with (
             torch._inductor.utils.fresh_inductor_cache(),
             torch._dynamo.config.patch(error_on_recompile=True),
@@ -320,21 +384,18 @@ class QwenImageTransformerCompileTests(TorchCompileTesterMixin, unittest.TestCas
         ):
             output_all_ones_2 = model(**inputs_all_ones)
 
-        self.assertEqual(output_all_ones.sample.shape[1], inputs["hidden_states"].shape[1])
-        self.assertEqual(output_all_ones_2.sample.shape[1], inputs["hidden_states"].shape[1])
+        assert output_all_ones.sample.shape[1] == inputs["hidden_states"].shape[1]
+        assert output_all_ones_2.sample.shape[1] == inputs["hidden_states"].shape[1]
 
-        # Test 3: Run with actual padding mask (has zeros)
         inputs_with_padding = inputs.copy()
         mask_with_padding = inputs["encoder_hidden_states_mask"].clone()
-        mask_with_padding[:, 4:] = 0  # Last 3 tokens are padding
+        mask_with_padding[:, 4:] = 0
 
         inputs_with_padding["encoder_hidden_states_mask"] = mask_with_padding
 
-        # First run to allow compilation
         with torch.no_grad():
             output_with_padding = model(**inputs_with_padding)
 
-        # Second run to verify no recompilation
         with (
             torch._inductor.utils.fresh_inductor_cache(),
             torch._dynamo.config.patch(error_on_recompile=True),
@@ -342,8 +403,15 @@ class QwenImageTransformerCompileTests(TorchCompileTesterMixin, unittest.TestCas
         ):
             output_with_padding_2 = model(**inputs_with_padding)
 
-        self.assertEqual(output_with_padding.sample.shape[1], inputs["hidden_states"].shape[1])
-        self.assertEqual(output_with_padding_2.sample.shape[1], inputs["hidden_states"].shape[1])
+        assert output_with_padding.sample.shape[1] == inputs["hidden_states"].shape[1]
+        assert output_with_padding_2.sample.shape[1] == inputs["hidden_states"].shape[1]
 
-        # Verify that outputs are different (mask should affect results)
-        self.assertFalse(torch.allclose(output_no_mask.sample, output_with_padding.sample, atol=1e-3))
+        assert not torch.allclose(output_no_mask.sample, output_with_padding.sample, atol=1e-3)
+
+
+class TestQwenImageTransformerBitsAndBytes(QwenImageTransformerTesterConfig, BitsAndBytesTesterMixin):
+    """BitsAndBytes quantization tests for QwenImage Transformer."""
+
+
+class TestQwenImageTransformerTorchAo(QwenImageTransformerTesterConfig, TorchAoTesterMixin):
+    """TorchAO quantization tests for QwenImage Transformer."""

tests/pipelines/helios/test_helios.py

@@ -139,9 +139,9 @@ class HeliosPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         generated_slice = torch.cat([generated_slice[:8], generated_slice[-8:]])
         self.assertTrue(torch.allclose(generated_slice, expected_slice, atol=1e-3))
 
-    # Override to set a more lenient max diff threshold.
+    @unittest.skip("Helios uses a lot of mixed precision internally, which is not suitable for this test case")
     def test_save_load_float16(self):
-        super().test_save_load_float16(expected_max_diff=0.03)
+        pass
 
     @unittest.skip("Test not supported")
     def test_attention_slicing_forward_pass(self):