Merge pull request #382 from AI-Hypercomputer:ltx2_bugfix

PiperOrigin-RevId: 903276594

Author: Google-ML-Automation

src/maxdiffusion/models/attention_flax.py

@@ -317,11 +317,11 @@ def _tpu_flash_attention(
 ) -> jax.Array:
   """TPU Flash Attention"""
 
-  block_sizes = _select_flash_block_sizes(query, key, flash_block_sizes, dtype, attention_kernel)
   num_context_shards = mesh.shape["context"]
   query, orig_q_seq_len = _reshape_data_for_flash(query, heads, num_context_shards)
   key, _ = _reshape_data_for_flash(key, heads, num_context_shards)
   value, _ = _reshape_data_for_flash(value, heads, num_context_shards)
+  block_sizes = _select_flash_block_sizes(query, key, flash_block_sizes, dtype, attention_kernel)
 
   q_axis_names = nn.logical_to_mesh_axes(axis_names_q)
   kv_axis_names = nn.logical_to_mesh_axes(axis_names_kv)

src/maxdiffusion/pipelines/ltx2/ltx2_pipeline.py

@@ -127,6 +127,8 @@ def create_model(rngs: nnx.Rngs, ltx2_config: dict):
   ltx2_config["dtype"] = config.activations_dtype
   ltx2_config["weights_dtype"] = config.weights_dtype
   ltx2_config["attention_kernel"] = config.attention
+  ltx2_config["a2v_attention_kernel"] = getattr(config, "a2v_attention_kernel", "flash")
+  ltx2_config["v2a_attention_kernel"] = getattr(config, "v2a_attention_kernel", "dot_product")
   ltx2_config["precision"] = get_precision(config)
   ltx2_config["flash_block_sizes"] = get_flash_block_sizes(config)
   ltx2_config["flash_min_seq_length"] = getattr(config, "flash_min_seq_length", 4096)