Source sync: Add Flax configuration updates and improve WAN trainer profiling

src/maxdiffusion/__init__.py

@@ -64,6 +64,11 @@
     ],
 }
 
+if is_flax_available():
+  from flax import config as flax_config
+
+  flax_config.update("flax_always_shard_variable", False)
+
 try:
   if not is_onnx_available():
     raise OptionalDependencyNotAvailable()

src/maxdiffusion/configuration_utils.py

@@ -376,11 +376,11 @@ def load_config(
     if os.path.isfile(pretrained_model_name_or_path):
       config_file = pretrained_model_name_or_path
     elif os.path.isdir(pretrained_model_name_or_path):
-      if os.path.isfile(os.path.join(pretrained_model_name_or_path, cls.config_name)):
+      if subfolder is not None and os.path.isfile(os.path.join(pretrained_model_name_or_path, subfolder, cls.config_name)):
+        config_file = os.path.join(pretrained_model_name_or_path, subfolder, cls.config_name)
+      elif os.path.isfile(os.path.join(pretrained_model_name_or_path, cls.config_name)):
         # Load from a PyTorch checkpoint
         config_file = os.path.join(pretrained_model_name_or_path, cls.config_name)
-      elif subfolder is not None and os.path.isfile(os.path.join(pretrained_model_name_or_path, subfolder, cls.config_name)):
-        config_file = os.path.join(pretrained_model_name_or_path, subfolder, cls.config_name)
       else:
         raise EnvironmentError(f"Error no file named {cls.config_name} found in directory {pretrained_model_name_or_path}.")
     else:

src/maxdiffusion/trainers/wan_trainer.py

@@ -444,19 +444,21 @@ def loss_fn(params):
     noise = jax.random.normal(key=new_rng, shape=latents.shape, dtype=latents.dtype)
     noisy_latents = scheduler.add_noise(scheduler_state, latents, noise, timesteps)
 
-    model_pred = model(
-        hidden_states=noisy_latents,
-        timestep=timesteps,
-        encoder_hidden_states=encoder_hidden_states,
-        deterministic=False,
-        rngs=nnx.Rngs(dropout_rng),
-    )
+    with jax.named_scope("forward_pass"):
+      model_pred = model(
+          hidden_states=noisy_latents,
+          timestep=timesteps,
+          encoder_hidden_states=encoder_hidden_states,
+          deterministic=False,
+          rngs=nnx.Rngs(dropout_rng),
+      )
 
-    training_target = scheduler.training_target(latents, noise, timesteps)
-    training_weight = jnp.expand_dims(scheduler.training_weight(scheduler_state, timesteps), axis=(1, 2, 3, 4))
-    loss = (training_target - model_pred) ** 2
-    loss = loss * training_weight
-    loss = jnp.mean(loss)
+    with jax.named_scope("loss"):
+      training_target = scheduler.training_target(latents, noise, timesteps)
+      training_weight = jnp.expand_dims(scheduler.training_weight(scheduler_state, timesteps), axis=(1, 2, 3, 4))
+      loss = (training_target - model_pred) ** 2
+      loss = loss * training_weight
+      loss = jnp.mean(loss)
 
     return loss