disabling stg+ rope type = split

Author: prishajain1

src/maxdiffusion/configs/ltx2_3_video.yml

@@ -31,8 +31,8 @@ guidance_scale: 3.0
 guidance_rescale: 0.7
 audio_guidance_scale: 7.0
 audio_guidance_rescale: 0.7
-stg_scale: 1.0
-audio_stg_scale: 1.0
+stg_scale: 0.0
+audio_stg_scale: 0.0
 modality_scale: 1.0
 audio_modality_scale: 1.0
 use_cross_timestep: true

src/maxdiffusion/models/ltx2/text_encoders/text_encoders_ltx2.py

@@ -53,7 +53,7 @@ def __init__(
       connector_rope_base_seq_len: int = 4096,
       rope_double_precision: bool = True,
       rope_theta: float = 10000.0,
-      rope_type: str = "interleaved",
+      rope_type: str = "split",
       causal_temporal_positioning: bool = False,
       dtype: DType = jnp.float32,
       attention_kernel: str = "flash",

src/maxdiffusion/pipelines/ltx2/ltx2_pipeline.py

@@ -141,7 +141,7 @@ def create_model(rngs: nnx.Rngs, ltx2_config: dict):
     ltx2_config = LTX2VideoTransformer3DModel.load_config(config.pretrained_model_name_or_path, subfolder=subfolder)
 
   # Align RoPE type with connectors
-  ltx2_config["rope_type"] = "interleaved"
+  ltx2_config["rope_type"] = "split"
 
   if ltx2_config.get("activation_fn") == "gelu-approximate":
     ltx2_config["activation_fn"] = "gelu"
@@ -388,7 +388,7 @@ def create_model(rngs: nnx.Rngs, config: HyperParameters):
                 "audio_gated_attn": True,
                 "per_modality_projections": True,
                 "proj_bias": True,
-                "rope_type": "interleaved",
+                "rope_type": "split",
             }
         )
       connector_repo = "Lightricks/LTX-2" if getattr(config, "model_name", "") == "ltx2.3" else config.pretrained_model_name_or_path