cross attn dim and vocoder weights

Author: prishajain1

src/maxdiffusion/models/ltx2/ltx2_3_utils.py

@@ -389,7 +389,7 @@ def load_vocoder_weights_2_3(
 
     if flax_key[-1] == "kernel":
       if "upsamplers" in flax_key:
-        tensor = tensor.transpose(2, 0, 1)[::-1, :, :]
+        tensor = tensor.transpose(2, 0, 1)
       else:
         tensor = tensor.transpose(2, 1, 0)
 

src/maxdiffusion/pipelines/ltx2/ltx2_pipeline.py

@@ -119,7 +119,7 @@ def create_model(rngs: nnx.Rngs, ltx2_config: dict):
         "in_channels": 128,
         "num_attention_heads": 32,
         "attention_head_dim": 128,
-        "cross_attention_dim": 8192,
+        "cross_attention_dim": 4096,
         "audio_in_channels": 128,
         "audio_num_attention_heads": 32,
         "audio_attention_head_dim": 64,