New Eval Pipeline as the MLPerf (#254)

Author: susanbao

src/maxdiffusion/configs/base_wan_14b.yml

@@ -234,6 +234,9 @@ global_batch_size: 0
 # For creating tfrecords from dataset
 tfrecords_dir: ''
 no_records_per_shard: 0
+enable_eval_timesteps: False
+considered_timesteps_list: [125, 250, 375, 500, 625, 750, 875]
+num_eval_samples: 420
 
 warmup_steps_fraction: 0.1
 learning_rate_schedule_steps: -1 # By default the length of the schedule is set to the number of steps.
@@ -315,3 +318,6 @@ quantization_calibration_method: "absmax"
 eval_every: -1
 eval_data_dir: ""
 enable_generate_video_for_eval: False # This will increase the used TPU memory.
+eval_max_number_of_samples_in_bucket: 60 # The number of samples per bucket for evaluation. This is calculated by num_eval_samples / len(considered_timesteps_list).
+
+enable_ssim: True

src/maxdiffusion/data_preprocessing/wan_pusav1_to_tfrecords.py

@@ -55,13 +55,17 @@ def float_feature_list(value):
   return tf.train.Feature(float_list=tf.train.FloatList(value=value))
 
 
-def create_example(latent, hidden_states):
+def create_example(latent, hidden_states, timestep=None):
   latent = tf.io.serialize_tensor(latent)
   hidden_states = tf.io.serialize_tensor(hidden_states)
   feature = {
       "latents": bytes_feature(latent),
       "encoder_hidden_states": bytes_feature(hidden_states),
   }
+  # Add timestep feature if it is provided
+  if timestep is not None:
+    feature["timesteps"] = int64_feature(timestep)
+
   example = tf.train.Example(features=tf.train.Features(feature=feature))
   return example.SerializeToString()
 
@@ -80,6 +84,12 @@ def generate_dataset(config):
   )
   shard_record_count = 0
 
+  # Define timesteps and bucket configuration
+  num_eval_samples = config.num_eval_samples
+  timesteps_list = config.timesteps_list
+  assert num_eval_samples % len(timesteps_list) == 0
+  bucket_size = num_eval_samples // len(timesteps_list)
+
   # Load dataset
   metadata_path = os.path.join(config.train_data_dir, "metadata.csv")
   with open(metadata_path, "r", newline="") as file:
@@ -102,7 +112,20 @@ def generate_dataset(config):
       # Save them as float32 because numpy cannot read bfloat16.
       latent = jnp.array(latent.float().numpy(), dtype=jnp.float32)
       prompt_embeds = jnp.array(prompt_embeds.float().numpy(), dtype=jnp.float32)
-      writer.write(create_example(latent, prompt_embeds))
+
+      current_timestep = None
+      # Determine the timestep for the first 420 samples
+      if config.enable_eval_timesteps:
+        if global_record_count < num_eval_samples:
+          print(f"global_record_count: {global_record_count}")
+          bucket_index = global_record_count // bucket_size
+          current_timestep = timesteps_list[bucket_index]
+        else:
+          print(f"value {global_record_count} is greater than or equal to {num_eval_samples}")
+          return
+
+      # Write the example, including the timestep if applicable
+      writer.write(create_example(latent, prompt_embeds, timestep=current_timestep))
       shard_record_count += 1
       global_record_count += 1
 

src/maxdiffusion/trainers/wan_trainer.py

@@ -38,6 +38,7 @@
 from skimage.metrics import structural_similarity as ssim
 from flax.training import train_state
 from maxdiffusion.pipelines.wan.wan_pipeline import WanPipeline
+from jax.experimental import multihost_utils
 
 
 class TrainState(train_state.TrainState):
@@ -156,6 +157,11 @@ def get_data_shardings(self, mesh):
     data_sharding = {"latents": data_sharding, "encoder_hidden_states": data_sharding}
     return data_sharding
 
+  def get_eval_data_shardings(self, mesh):
+    data_sharding = jax.sharding.NamedSharding(mesh, P(*self.config.data_sharding))
+    data_sharding = {"latents": data_sharding, "encoder_hidden_states": data_sharding, "timesteps": data_sharding}
+    return data_sharding
+
   def load_dataset(self, mesh, is_training=True):
     # Stages of training as described in the Wan 2.1 paper - https://arxiv.org/pdf/2503.20314
     # Image pre-training - txt2img 256px
@@ -170,34 +176,43 @@ def load_dataset(self, mesh, is_training=True):
       raise ValueError(
           "Wan 2.1 training only supports config.dataset_type set to tfrecords and config.cache_latents_text_encoder_outputs set to True"
       )
-
     feature_description = {
         "latents": tf.io.FixedLenFeature([], tf.string),
         "encoder_hidden_states": tf.io.FixedLenFeature([], tf.string),
     }
 
-    def prepare_sample(features):
+    if not is_training:
+      feature_description["timesteps"] = tf.io.FixedLenFeature([], tf.int64)
+
+    def prepare_sample_train(features):
       latents = tf.io.parse_tensor(features["latents"], out_type=tf.float32)
       encoder_hidden_states = tf.io.parse_tensor(features["encoder_hidden_states"], out_type=tf.float32)
       return {"latents": latents, "encoder_hidden_states": encoder_hidden_states}
 
+    def prepare_sample_eval(features):
+      latents = tf.io.parse_tensor(features["latents"], out_type=tf.float32)
+      encoder_hidden_states = tf.io.parse_tensor(features["encoder_hidden_states"], out_type=tf.float32)
+      timesteps = features["timesteps"]
+      return {"latents": latents, "encoder_hidden_states": encoder_hidden_states, "timesteps": timesteps}
+
     data_iterator = make_data_iterator(
         config,
         jax.process_index(),
         jax.process_count(),
         mesh,
         config.global_batch_size_to_load,
         feature_description=feature_description,
-        prepare_sample_fn=prepare_sample,
+        prepare_sample_fn=prepare_sample_train if is_training else prepare_sample_eval,
         is_training=is_training,
     )
     return data_iterator
 
   def start_training(self):
 
     pipeline = self.load_checkpoint()
-    # Generate a sample before training to compare against generated sample after training.
-    pretrained_video_path = generate_sample(self.config, pipeline, filename_prefix="pre-training-")
+    if self.config.enable_ssim:
+      # Generate a sample before training to compare against generated sample after training.
+      pretrained_video_path = generate_sample(self.config, pipeline, filename_prefix="pre-training-")
 
     if self.config.eval_every == -1 or (not self.config.enable_generate_video_for_eval):
       # save some memory.
@@ -215,8 +230,57 @@ def start_training(self):
     # Returns pipeline with trained transformer state
     pipeline = self.training_loop(pipeline, optimizer, learning_rate_scheduler, train_data_iterator)
 
-    posttrained_video_path = generate_sample(self.config, pipeline, filename_prefix="post-training-")
-    print_ssim(pretrained_video_path, posttrained_video_path)
+    if self.config.enable_ssim:
+      posttrained_video_path = generate_sample(self.config, pipeline, filename_prefix="post-training-")
+      print_ssim(pretrained_video_path, posttrained_video_path)
+
+  def eval(self, mesh, eval_rng_key, step, p_eval_step, state, scheduler_state, writer):
+    eval_data_iterator = self.load_dataset(mesh, is_training=False)
+    eval_rng = eval_rng_key
+    eval_losses_by_timestep = {}
+    # Loop indefinitely until the iterator is exhausted
+    while True:
+      try:
+        eval_start_time = datetime.datetime.now()
+        eval_batch = load_next_batch(eval_data_iterator, None, self.config)
+        with mesh, nn_partitioning.axis_rules(
+            self.config.logical_axis_rules
+        ):
+          metrics, eval_rng = p_eval_step(state, eval_batch, eval_rng, scheduler_state)
+          metrics["scalar"]["learning/eval_loss"].block_until_ready()
+        losses = metrics["scalar"]["learning/eval_loss"]
+        timesteps = eval_batch["timesteps"]
+        gathered_losses = multihost_utils.process_allgather(losses)
+        gathered_losses = jax.device_get(gathered_losses)
+        gathered_timesteps = multihost_utils.process_allgather(timesteps)
+        gathered_timesteps = jax.device_get(gathered_timesteps)
+        if jax.process_index() == 0:
+          for t, l in zip(gathered_timesteps.flatten(), gathered_losses.flatten()):
+            timestep = int(t)
+            if timestep not in eval_losses_by_timestep:
+                eval_losses_by_timestep[timestep] = []
+            eval_losses_by_timestep[timestep].append(l)
+          eval_end_time = datetime.datetime.now()
+          eval_duration = eval_end_time - eval_start_time
+          max_logging.log(f"Eval time: {eval_duration.total_seconds():.2f} seconds.")
+      except StopIteration:
+        # This block is executed when the iterator has no more data
+        break
+    # Check if any evaluation was actually performed
+    if eval_losses_by_timestep and jax.process_index() == 0:
+      mean_per_timestep = []
+      if jax.process_index() == 0:
+        max_logging.log(f"Step {step}, calculating mean loss per timestep...")
+      for timestep, losses in sorted(eval_losses_by_timestep.items()):
+          losses = jnp.array(losses)
+          losses = losses[: min(self.config.eval_max_number_of_samples_in_bucket, len(losses))]
+          mean_loss = jnp.mean(losses)
+          max_logging.log(f"  Mean eval loss for timestep {timestep}: {mean_loss:.4f}")
+          mean_per_timestep.append(mean_loss)
+      final_eval_loss = jnp.mean(jnp.array(mean_per_timestep))
+      max_logging.log(f"Step {step}, Final Average Eval loss: {final_eval_loss:.4f}")
+      if writer:
+        writer.add_scalar("learning/eval_loss", final_eval_loss, step)
 
   def training_loop(self, pipeline, optimizer, learning_rate_scheduler, train_data_iterator):
     mesh = pipeline.mesh
@@ -231,6 +295,7 @@ def training_loop(self, pipeline, optimizer, learning_rate_scheduler, train_data
       state = jax.lax.with_sharding_constraint(state, state_spec)
       state_shardings = nnx.get_named_sharding(state, mesh)
     data_shardings = self.get_data_shardings(mesh)
+    eval_data_shardings = self.get_eval_data_shardings(mesh)
 
     writer = max_utils.initialize_summary_writer(self.config)
     writer_thread = threading.Thread(target=_tensorboard_writer_worker, args=(writer, self.config), daemon=True)
@@ -255,11 +320,12 @@ def training_loop(self, pipeline, optimizer, learning_rate_scheduler, train_data
     )
     p_eval_step = jax.jit(
         functools.partial(eval_step, scheduler=pipeline.scheduler, config=self.config),
-        in_shardings=(state_shardings, data_shardings, None, None),
+        in_shardings=(state_shardings, eval_data_shardings, None, None),
         out_shardings=(None, None),
     )
 
     rng = jax.random.key(self.config.seed)
+    rng, eval_rng_key = jax.random.split(rng)
     start_step = 0
     last_step_completion = datetime.datetime.now()
     local_metrics_file = open(self.config.metrics_file, "a", encoding="utf8") if self.config.metrics_file else None
@@ -304,27 +370,8 @@ def training_loop(self, pipeline, optimizer, learning_rate_scheduler, train_data
             inference_generate_video(self.config, pipeline, filename_prefix=f"{step+1}-train_steps-")
           # Re-create the iterator each time you start evaluation to reset it
           # This assumes your data loading logic can be called to get a fresh iterator.
-          eval_data_iterator = self.load_dataset(mesh, is_training=False)
-          eval_rng = jax.random.key(self.config.seed + step)
-          eval_metrics = []
-          # Loop indefinitely until the iterator is exhausted
-          while True:
-            try:
-              with mesh:
-                eval_batch = load_next_batch(eval_data_iterator, None, self.config)
-                metrics, eval_rng = p_eval_step(state, eval_batch, eval_rng, scheduler_state)
-                eval_metrics.append(metrics["scalar"]["learning/eval_loss"])
-            except StopIteration:
-              # This block is executed when the iterator has no more data
-              break
-          # Check if any evaluation was actually performed
-          if eval_metrics:
-            eval_loss = jnp.mean(jnp.array(eval_metrics))
-            max_logging.log(f"Step {step}, Eval loss: {eval_loss:.4f}")
-            if writer:
-              writer.add_scalar("learning/eval_loss", eval_loss, step)
-          else:
-            max_logging.log(f"Step {step}, evaluation dataset was empty.")
+          self.eval(mesh, eval_rng_key, step, p_eval_step, state, scheduler_state, writer)
+
         example_batch = next_batch_future.result()
         if step != 0 and self.config.checkpoint_every != -1 and step % self.config.checkpoint_every == 0:
           max_logging.log(f"Saving checkpoint for step {step}")
@@ -394,57 +441,54 @@ def eval_step(state, data, rng, scheduler_state, scheduler, config):
   """
   Computes the evaluation loss for a single batch without updating model weights.
   """
-  _, new_rng, timestep_rng = jax.random.split(rng, num=3)
-
-  # This ensures the batch size is consistent, though it might be redundant
-  # if the evaluation dataloader is already configured correctly.
-  for k, v in data.items():
-    data[k] = v[: config.global_batch_size_to_train_on, :]
 
   # The loss function logic is identical to training. We are evaluating the model's
   # ability to perform its core training objective (e.g., denoising).
-  def loss_fn(params):
+  @jax.jit
+  def loss_fn(params, latents, encoder_hidden_states, timesteps, rng):
     # Reconstruct the model from its definition and parameters
     model = nnx.merge(state.graphdef, params, state.rest_of_state)
 
-    # Prepare inputs
-    latents = data["latents"].astype(config.weights_dtype)
-    encoder_hidden_states = data["encoder_hidden_states"].astype(config.weights_dtype)
-    bsz = latents.shape[0]
-
-    # Sample random timesteps and noise, just as in a training step
-    timesteps = jax.random.randint(
-        timestep_rng,
-        (bsz,),
-        0,
-        scheduler.config.num_train_timesteps,
-    )
-    noise = jax.random.normal(key=new_rng, shape=latents.shape, dtype=latents.dtype)
+    noise = jax.random.normal(key=rng, shape=latents.shape, dtype=latents.dtype)
     noisy_latents = scheduler.add_noise(scheduler_state, latents, noise, timesteps)
 
     # Get the model's prediction
     model_pred = model(
         hidden_states=noisy_latents,
         timestep=timesteps,
         encoder_hidden_states=encoder_hidden_states,
+        deterministic=True,
     )
 
     # Calculate the loss against the target
     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)
+    # Calculate the mean loss per sample across all non-batch dimensions.
+    loss = loss.reshape(loss.shape[0], -1).mean(axis=1)
 
     return loss
 
   # --- Key Difference from train_step ---
   # Directly compute the loss without calculating gradients.
   # The model's state.params are used but not updated.
-  loss = loss_fn(state.params)
+  # TODO(coolkp): Explore optimizing the creation of PRNGs in a vmap or statically outside of the loop
+  bs = len(data["latents"])
+  single_batch_size = config.global_batch_size_to_train_on
+  losses = jnp.zeros(bs)
+  for i in range(0, bs, single_batch_size):
+    start = i
+    end = min(i + single_batch_size, bs)
+    latents= data["latents"][start:end, :].astype(config.weights_dtype)
+    encoder_hidden_states = data["encoder_hidden_states"][start:end, :].astype(config.weights_dtype)
+    timesteps = data["timesteps"][start:end].astype("int64")
+    _, new_rng = jax.random.split(rng, num=2)
+    loss = loss_fn(state.params, latents, encoder_hidden_states, timesteps, new_rng)
+    losses = losses.at[start:end].set(loss)
 
   # Structure the metrics for logging and aggregation
-  metrics = {"scalar": {"learning/eval_loss": loss}}
+  metrics = {"scalar": {"learning/eval_loss": losses}}
 
   # Return the computed metrics and the new RNG key for the next eval step
   return metrics, new_rng