diffusers 0.40: load LTX2 GGUF via LTX2VideoTransformer3DModel.from_single_file

app.py

@@ -80,12 +80,17 @@ def _download_gguf(dest: str) -> str:
 
 
 def _build_transformer(gguf_path: str):
-    from diffusers import AutoModel
+    try:
+        # diffusers 0.40 (pinned): the LTX-2 video transformer is loadable from
+        # a single GGUF file via FromOriginalModelMixin.
+        from diffusers import LTX2VideoTransformer3DModel as TransformerCls
+    except ImportError:  # older versions only expose the generic AutoModel
+        from diffusers import AutoModel as TransformerCls
 
     try:
-        # diffusers 0.33.x (pinned): GGUF config lives under quantizers.
+        # GGUF config lives in quantizers since 0.33; newer releases keep it there.
         from diffusers.quantizers.quantization_config import GGUFQuantizationConfig
-    except ImportError:  # later versions re-export it from diffusers.utils
+    except ImportError:  # legacy export path
         from diffusers.utils import GGUFQuantizationConfig
 
     kwargs = dict(
@@ -93,10 +98,10 @@ def _build_transformer(gguf_path: str):
         dtype=torch.bfloat16,
     )
     try:
-        return AutoModel.from_single_file(gguf_path, config=PIPELINE_ID, **kwargs)
+        return TransformerCls.from_single_file(gguf_path, config=PIPELINE_ID, **kwargs)
     except Exception as exc:  # config may live elsewhere; trust the GGUF KV metadata
         print(f"[ltx] from_single_file with config failed ({exc}), retrying without")
-        return AutoModel.from_single_file(gguf_path, **kwargs)
+        return TransformerCls.from_single_file(gguf_path, **kwargs)
 
 
 def _load_model() -> dict:

requirements.txt

@@ -1,8 +1,9 @@
 # For ZeroGPU, gradio/spaces/huggingface_hub/torch are preinstalled and
 # platform-managed; the Dockerfile installs them explicitly.
-# Pinned to 0.33: LTX2 pipelines + GGUFQuantizationConfig both live here;
-# newer diffusers moved GGUFQuantizationConfig out of diffusers.utils.
-diffusers==0.33.0
+# Pinned to 0.40: LTX2 video transformer + GGUF single-file loading both work
+# here (LTX2VideoTransformer3DModel + GGUFQuantizer). Earlier versions lack
+# AutoModel/from_single_file support for LTX-2 GGUF.
+diffusers==0.40.0
 transformers==4.49.0
 accelerate>=1.0
 gguf