update

* fixing text encoder lora loading

* following Cursor's review

src/diffusers/loaders/lora_conversion_utils.py

@@ -856,7 +856,7 @@ def _convert_kohya_flux_lora_to_diffusers(state_dict):
                 )
             state_dict = {k: v for k, v in state_dict.items() if not k.startswith("text_encoders.t5xxl.transformer.")}
 
-        has_diffb = any("diff_b" in k and k.startswith(("lora_unet_", "lora_te_")) for k in state_dict)
+        has_diffb = any("diff_b" in k and k.startswith(("lora_unet_", "lora_te_", "lora_te1_")) for k in state_dict)
         if has_diffb:
             zero_status_diff_b = state_dict_all_zero(state_dict, ".diff_b")
             if zero_status_diff_b:
@@ -895,7 +895,7 @@ def _convert_kohya_flux_lora_to_diffusers(state_dict):
         state_dict = {
             _custom_replace(k, limit_substrings): v
             for k, v in state_dict.items()
-            if k.startswith(("lora_unet_", "lora_te_"))
+            if k.startswith(("lora_unet_", "lora_te_", "lora_te1_"))
         }
 
         if any("text_projection" in k for k in state_dict):

src/diffusers/models/attention_dispatch.py

@@ -733,7 +733,7 @@ def _wrapped_flash_attn_3(
 ) -> tuple[torch.Tensor, torch.Tensor]:
     # Hardcoded for now because pytorch does not support tuple/int type hints
     window_size = (-1, -1)
-    out, lse, *_ = flash_attn_3_func(
+    result = flash_attn_3_func(
         q=q,
         k=k,
         v=v,
@@ -750,7 +750,9 @@ def _wrapped_flash_attn_3(
         pack_gqa=pack_gqa,
         deterministic=deterministic,
         sm_margin=sm_margin,
+        return_attn_probs=True,
     )
+    out, lse, *_ = result
     lse = lse.permute(0, 2, 1)
     return out, lse
 
@@ -2701,7 +2703,7 @@ def _flash_varlen_attention_3(
     key_packed = torch.cat(key_valid, dim=0)
     value_packed = torch.cat(value_valid, dim=0)
 
-    out, lse, *_ = flash_attn_3_varlen_func(
+    result = flash_attn_3_varlen_func(
         q=query_packed,
         k=key_packed,
         v=value_packed,
@@ -2711,7 +2713,13 @@ def _flash_varlen_attention_3(
         max_seqlen_k=max_seqlen_k,
         softmax_scale=scale,
         causal=is_causal,
+        return_attn_probs=return_lse,
     )
+    if isinstance(result, tuple):
+        out, lse, *_ = result
+    else:
+        out = result
+        lse = None
     out = out.unflatten(0, (batch_size, -1))
 
     return (out, lse) if return_lse else out

src/diffusers/pipelines/ltx2/pipeline_ltx2_image2video.py

@@ -699,9 +699,13 @@ class LTX2ImageToVideoPipeline(DiffusionPipeline, FromSingleFileMixin, LTX2LoraL
         mask_shape = (batch_size, 1, num_frames, height, width)
 
         if latents is not None:
-            conditioning_mask = latents.new_zeros(mask_shape)
-            conditioning_mask[:, :, 0] = 1.0
             if latents.ndim == 5:
+                # conditioning_mask needs to the same shape as latents in two stages generation.
+                batch_size, _, num_frames, height, width = latents.shape
+                mask_shape = (batch_size, 1, num_frames, height, width)
+                conditioning_mask = latents.new_zeros(mask_shape)
+                conditioning_mask[:, :, 0] = 1.0
+
                 latents = self._normalize_latents(
                     latents, self.vae.latents_mean, self.vae.latents_std, self.vae.config.scaling_factor
                 )
@@ -710,6 +714,9 @@ class LTX2ImageToVideoPipeline(DiffusionPipeline, FromSingleFileMixin, LTX2LoraL
                 latents = self._pack_latents(
                     latents, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
                 )
+            else:
+                conditioning_mask = latents.new_zeros(mask_shape)
+                conditioning_mask[:, :, 0] = 1.0
             conditioning_mask = self._pack_latents(
                 conditioning_mask, self.transformer_spatial_patch_size, self.transformer_temporal_patch_size
             ).squeeze(-1)

src/diffusers/utils/dynamic_modules_utils.py

@@ -299,7 +299,10 @@ def get_cached_module_file(
     # Download and cache module_file from the repo `pretrained_model_name_or_path` of grab it if it's a local file.
     pretrained_model_name_or_path = str(pretrained_model_name_or_path)
 
-    module_file_or_url = os.path.join(pretrained_model_name_or_path, module_file)
+    if subfolder is not None:
+        module_file_or_url = os.path.join(pretrained_model_name_or_path, subfolder, module_file)
+    else:
+        module_file_or_url = os.path.join(pretrained_model_name_or_path, module_file)
 
     if os.path.isfile(module_file_or_url):
         resolved_module_file = module_file_or_url
@@ -384,7 +387,11 @@ def get_cached_module_file(
                 if not os.path.exists(submodule_path / module_folder):
                     os.makedirs(submodule_path / module_folder)
             module_needed = f"{module_needed}.py"
-            shutil.copyfile(os.path.join(pretrained_model_name_or_path, module_needed), submodule_path / module_needed)
+            if subfolder is not None:
+                source_path = os.path.join(pretrained_model_name_or_path, subfolder, module_needed)
+            else:
+                source_path = os.path.join(pretrained_model_name_or_path, module_needed)
+            shutil.copyfile(source_path, submodule_path / module_needed)
     else:
         # Get the commit hash
         # TODO: we will get this info in the etag soon, so retrieve it from there and not here.

tests/models/test_models_auto.py

@@ -1,6 +1,10 @@
+import json
+import os
+import tempfile
 import unittest
 from unittest.mock import MagicMock, patch
 
+import torch
 from transformers import CLIPTextModel, LongformerModel
 
 from diffusers.models import AutoModel, UNet2DConditionModel
@@ -35,6 +39,45 @@ class TestAutoModel(unittest.TestCase):
         )
         assert isinstance(model, CLIPTextModel)
 
+    def test_load_dynamic_module_from_local_path_with_subfolder(self):
+        CUSTOM_MODEL_CODE = (
+            "import torch\n"
+            "from diffusers import ModelMixin, ConfigMixin\n"
+            "from diffusers.configuration_utils import register_to_config\n"
+            "\n"
+            "class CustomModel(ModelMixin, ConfigMixin):\n"
+            "    @register_to_config\n"
+            "    def __init__(self, hidden_size=8):\n"
+            "        super().__init__()\n"
+            "        self.linear = torch.nn.Linear(hidden_size, hidden_size)\n"
+            "\n"
+            "    def forward(self, x):\n"
+            "        return self.linear(x)\n"
+        )
+
+        with tempfile.TemporaryDirectory() as tmpdir:
+            subfolder = "custom_model"
+            model_dir = os.path.join(tmpdir, subfolder)
+            os.makedirs(model_dir)
+
+            with open(os.path.join(model_dir, "modeling.py"), "w") as f:
+                f.write(CUSTOM_MODEL_CODE)
+
+            config = {
+                "_class_name": "CustomModel",
+                "_diffusers_version": "0.0.0",
+                "auto_map": {"AutoModel": "modeling.CustomModel"},
+                "hidden_size": 8,
+            }
+            with open(os.path.join(model_dir, "config.json"), "w") as f:
+                json.dump(config, f)
+
+            torch.save({}, os.path.join(model_dir, "diffusion_pytorch_model.bin"))
+
+            model = AutoModel.from_pretrained(tmpdir, subfolder=subfolder, trust_remote_code=True)
+            assert model.__class__.__name__ == "CustomModel"
+            assert model.config["hidden_size"] == 8
+
 
 class TestAutoModelFromConfig(unittest.TestCase):
     @patch(

tests/pipelines/ltx2/test_ltx2_image2video.py

@@ -24,7 +24,8 @@ from diffusers import (
     LTX2ImageToVideoPipeline,
     LTX2VideoTransformer3DModel,
 )
-from diffusers.pipelines.ltx2 import LTX2TextConnectors
+from diffusers.pipelines.ltx2 import LTX2LatentUpsamplePipeline, LTX2TextConnectors
+from diffusers.pipelines.ltx2.latent_upsampler import LTX2LatentUpsamplerModel
 from diffusers.pipelines.ltx2.vocoder import LTX2Vocoder
 
 from ...testing_utils import enable_full_determinism
@@ -174,6 +175,15 @@ class LTX2ImageToVideoPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
 
         return components
 
+    def get_dummy_upsample_component(self, in_channels=4, mid_channels=32, num_blocks_per_stage=1):
+        upsampler = LTX2LatentUpsamplerModel(
+            in_channels=in_channels,
+            mid_channels=mid_channels,
+            num_blocks_per_stage=num_blocks_per_stage,
+        )
+
+        return upsampler
+
     def get_dummy_inputs(self, device, seed=0):
         if str(device).startswith("mps"):
             generator = torch.manual_seed(seed)
@@ -287,5 +297,60 @@ class LTX2ImageToVideoPipelineFastTests(PipelineTesterMixin, unittest.TestCase):
         assert torch.allclose(expected_video_slice, generated_video_slice, atol=1e-4, rtol=1e-4)
         assert torch.allclose(expected_audio_slice, generated_audio_slice, atol=1e-4, rtol=1e-4)
 
+    def test_two_stages_inference_with_upsampler(self):
+        device = "cpu"
+
+        components = self.get_dummy_components()
+        pipe = self.pipeline_class(**components)
+        pipe.to(device)
+        pipe.set_progress_bar_config(disable=None)
+
+        inputs = self.get_dummy_inputs(device)
+        inputs["output_type"] = "latent"
+        first_stage_output = pipe(**inputs)
+        video_latent = first_stage_output.frames
+        audio_latent = first_stage_output.audio
+
+        self.assertEqual(video_latent.shape, (1, 4, 3, 16, 16))
+        self.assertEqual(audio_latent.shape, (1, 2, 5, 2))
+        self.assertEqual(audio_latent.shape[1], components["vocoder"].config.out_channels)
+
+        upsampler = self.get_dummy_upsample_component(in_channels=video_latent.shape[1])
+        upsample_pipe = LTX2LatentUpsamplePipeline(vae=pipe.vae, latent_upsampler=upsampler)
+        upscaled_video_latent = upsample_pipe(latents=video_latent, output_type="latent", return_dict=False)[0]
+        self.assertEqual(upscaled_video_latent.shape, (1, 4, 3, 32, 32))
+
+        inputs["latents"] = upscaled_video_latent
+        inputs["audio_latents"] = audio_latent
+        inputs["output_type"] = "pt"
+        second_stage_output = pipe(**inputs)
+        video = second_stage_output.frames
+        audio = second_stage_output.audio
+
+        self.assertEqual(video.shape, (1, 5, 3, 64, 64))
+        self.assertEqual(audio.shape[0], 1)
+        self.assertEqual(audio.shape[1], components["vocoder"].config.out_channels)
+
+        # fmt: off
+        expected_video_slice = torch.tensor(
+            [
+                0.4497, 0.6757, 0.4219, 0.7686, 0.4525, 0.6483, 0.3969, 0.7404, 0.3541, 0.3039, 0.4592, 0.3521, 0.3665, 0.2785, 0.3336, 0.3079
+            ]
+        )
+        expected_audio_slice = torch.tensor(
+            [
+                0.0271, 0.0492, 0.1249, 0.1126, 0.1661, 0.1060, 0.1717, 0.0944, 0.0672, -0.0069, 0.0688, 0.0097, 0.0808, 0.1231, 0.0986, 0.0739
+            ]
+        )
+        # fmt: on
+
+        video = video.flatten()
+        audio = audio.flatten()
+        generated_video_slice = torch.cat([video[:8], video[-8:]])
+        generated_audio_slice = torch.cat([audio[:8], audio[-8:]])
+
+        assert torch.allclose(expected_video_slice, generated_video_slice, atol=1e-4, rtol=1e-4)
+        assert torch.allclose(expected_audio_slice, generated_audio_slice, atol=1e-4, rtol=1e-4)
+
     def test_inference_batch_single_identical(self):
         self._test_inference_batch_single_identical(batch_size=2, expected_max_diff=2e-2)