Upload pipeline.py with huggingface_hub

pipeline.py

@@ -99,18 +99,27 @@ def load_pipeline(
 
     if progress_cb:
         progress_cb("Loading transformer…")
-    # fp8_e4m3fn weights are loaded into bfloat16 compute precision to maximise
-    # diffusers compatibility; the weight file on disk is fp8-quantised.
-    transformer = LTX2VideoTransformer3DModel.from_single_file(
-        paths["transformer"],
-        torch_dtype=torch.bfloat16,
-        # Sulphur 2 (22B) was trained with cross-attn adaln enabled, giving
-        # 9-row (not 6-row) scale_shift tables.
-        original_config={
-            "video_cross_attn_adaln": True,
-            "audio_cross_attn_adaln": True,
-        },
-    ).to(device)
+    # Sulphur 2 (22B) was trained with cross-attn modulation enabled.
+    # `from_single_file` can't infer this, so construct the model with the
+    # right config first, then load the converted state dict.
+    from diffusers.loaders.single_file_utils import (
+        convert_ltx_transformer_checkpoint_to_diffusers,
+    )
+    from safetensors import safe_open
+
+    transformer = LTX2VideoTransformer3DModel(
+        cross_attn_mod=True,
+        audio_cross_attn_mod=True,
+    )
+    with safe_open(paths["transformer"], framework="pt") as f:
+        # FP8 weights -> BF16 for diffusers compatibility
+        checkpoint = {k: f.get_slice(k)[:].to(torch.bfloat16) for k in f.keys()}
+    converted = convert_ltx_transformer_checkpoint_to_diffusers(checkpoint)
+    # Filter to keys that exist in the model (drops any extras)
+    model_keys = set(transformer.state_dict().keys())
+    filtered = {k: v for k, v in converted.items() if k in model_keys}
+    transformer.load_state_dict(filtered, strict=False)
+    transformer = transformer.to(torch.bfloat16).to(device)
 
     if progress_cb:
         progress_cb("Loading video VAE…")