use latest

run_clm_flax.py

@@ -1,6 +1,3 @@
-'''
-    Using this to compare & make latest changes to train.py
-'''
 #!/usr/bin/env python
 # coding=utf-8
 # Copyright 2021 The HuggingFace Team All rights reserved.
@@ -401,7 +398,8 @@ def main():
         total_length = len(concatenated_examples[list(examples.keys())[0]])
         # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
         # customize this part to your needs.
-        total_length = (total_length // block_size) * block_size
+        if total_length >= block_size:
+            total_length = (total_length // block_size) * block_size
         # Split by chunks of max_len.
         result = {
             k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
@@ -492,17 +490,24 @@ def main():
         return traverse_util.unflatten_dict(flat_mask)
 
     # create adam optimizer
-    adamw = optax.adamw(
-        learning_rate=linear_decay_lr_schedule_fn,
-        b1=training_args.adam_beta1,
-        b2=training_args.adam_beta2,
-        eps=training_args.adam_epsilon,
-        weight_decay=training_args.weight_decay,
-        mask=decay_mask_fn,
-    )
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = optax.adamw(
+            learning_rate=linear_decay_lr_schedule_fn,
+            b1=training_args.adam_beta1,
+            b2=training_args.adam_beta2,
+            eps=training_args.adam_epsilon,
+            weight_decay=training_args.weight_decay,
+            mask=decay_mask_fn,
+        )
 
     # Setup train state
-    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw, dropout_rng=dropout_rng)
+    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer, dropout_rng=dropout_rng)
 
     def loss_fn(logits, labels):
         shift_logits = logits[..., :-1, :]

train.py

@@ -401,7 +401,8 @@ def main():
         total_length = len(concatenated_examples[list(examples.keys())[0]])
         # We drop the small remainder, we could add padding if the model supported it instead of this drop, you can
         # customize this part to your needs.
-        total_length = (total_length // block_size) * block_size
+        if total_length >= block_size:
+            total_length = (total_length // block_size) * block_size
         # Split by chunks of max_len.
         result = {
             k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
@@ -492,17 +493,24 @@ def main():
         return traverse_util.unflatten_dict(flat_mask)
 
     # create adam optimizer
-    adamw = optax.adamw(
-        learning_rate=linear_decay_lr_schedule_fn,
-        b1=training_args.adam_beta1,
-        b2=training_args.adam_beta2,
-        eps=training_args.adam_epsilon,
-        weight_decay=training_args.weight_decay,
-        mask=decay_mask_fn,
-    )
+    if training_args.adafactor:
+        # We use the default parameters here to initialize adafactor,
+        # For more details about the parameters please check https://github.com/deepmind/optax/blob/ed02befef9bf81cbbf236be3d2b0e032e9ed4a40/optax/_src/alias.py#L74
+        optimizer = optax.adafactor(
+            learning_rate=linear_decay_lr_schedule_fn,
+        )
+    else:
+        optimizer = optax.adamw(
+            learning_rate=linear_decay_lr_schedule_fn,
+            b1=training_args.adam_beta1,
+            b2=training_args.adam_beta2,
+            eps=training_args.adam_epsilon,
+            weight_decay=training_args.weight_decay,
+            mask=decay_mask_fn,
+        )
 
     # Setup train state
-    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=adamw, dropout_rng=dropout_rng)
+    state = TrainState.create(apply_fn=model.__call__, params=model.params, tx=optimizer, dropout_rng=dropout_rng)
 
     def compute_kernel(x, y):
         x_size = x.shape[0]