Flax Remat for LongT5 (huggingface#17994)

* [Flax] Add remat (gradient checkpointing)

* fix variable naming in test

* flip: checkpoint using a method

* fix naming

* fix class naming

* apply PVP's suggestions from code review

* add gradient_checkpointing to examples

* Add gradient_checkpointing to run_mlm_flax

* Add remat to longt5

* Add gradient checkpointing test longt5

* Fix args errors

* Fix remaining tests

* Make fixup & quality fixes

* replace kwargs

* remove unecessary kwargs

* Make fixup changes

* revert long_t5_flax changes

* Remove return_dict and copy to LongT5

* Remove test_gradient_checkpointing

Co-authored-by: sanchit-gandhi <[email protected]>

Author: 2 people

examples/flax/language-modeling/run_mlm_flax.py

@@ -107,6 +107,12 @@ class TrainingArguments:
         default=None, metadata={"help": "The name of the repository to keep in sync with the local `output_dir`."}
     )
     hub_token: str = field(default=None, metadata={"help": "The token to use to push to the Model Hub."})
+    gradient_checkpointing: bool = field(
+        default=False,
+        metadata={
+            "help": "If True, use gradient checkpointing to save memory at the expense of slower backward pass."
+        },
+    )
 
     def __post_init__(self):
         if self.output_dir is not None:
@@ -640,6 +646,9 @@ def group_texts(examples):
             dtype=getattr(jnp, model_args.dtype),
         )
 
+    if training_args.gradient_checkpointing:
+        model.enable_gradient_checkpointing()
+
     # Store some constant
     num_epochs = int(training_args.num_train_epochs)
     train_batch_size = int(training_args.per_device_train_batch_size) * jax.device_count()

examples/flax/summarization/run_summarization_flax.py

@@ -121,6 +121,12 @@ class TrainingArguments:
         default=None, metadata={"help": "The name of the repository to keep in sync with the local `output_dir`."}
     )
     hub_token: str = field(default=None, metadata={"help": "The token to use to push to the Model Hub."})
+    gradient_checkpointing: bool = field(
+        default=False,
+        metadata={
+            "help": "If True, use gradient checkpointing to save memory at the expense of slower backward pass."
+        },
+    )
 
     def __post_init__(self):
         if self.output_dir is not None:
@@ -535,6 +541,9 @@ def main():
             dtype=getattr(jnp, model_args.dtype),
         )
 
+    if training_args.gradient_checkpointing:
+        model.enable_gradient_checkpointing()
+
     if model.config.decoder_start_token_id is None:
         raise ValueError("Make sure that `config.decoder_start_token_id` is correctly defined")
 

src/transformers/models/longt5/modeling_flax_longt5.py

@@ -25,6 +25,7 @@
 import jax.numpy as jnp
 from flax.core.frozen_dict import FrozenDict, freeze, unfreeze
 from flax.linen import combine_masks, make_causal_mask
+from flax.linen import partitioning as nn_partitioning
 from flax.linen.attention import dot_product_attention_weights
 from flax.traverse_util import flatten_dict, unflatten_dict
 from jax.random import PRNGKey
@@ -53,6 +54,8 @@
 _CONFIG_FOR_DOC = "LongT5Config"
 _TOKENIZER_FOR_DOC = "T5Tokenizer"
 
+remat = nn_partitioning.remat
+
 
 # Copied from transformers.models.bart.modeling_flax_bart.shift_tokens_right
 def shift_tokens_right(input_ids: np.array, pad_token_id: int, decoder_start_token_id: int) -> np.ndarray:
@@ -1356,7 +1359,6 @@ def __call__(
         encoder_attention_mask=None,
         encoder_decoder_position_bias=None,
         output_attentions=False,
-        return_dict=True,
         deterministic=True,
         init_cache=False,
     ):
@@ -1377,13 +1379,31 @@ def __call__(
 class FlaxLongT5BlockCollection(nn.Module):
     config: LongT5Config
     dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    gradient_checkpointing: bool = False
 
     def setup(self):
         self.causal = self.config.causal
-        self.blocks = [
-            FlaxLongT5LayerCollection(self.config, has_relative_attention_bias=(i == 0), dtype=self.dtype, name=str(i))
-            for i in range(self.config.num_layers)
-        ]
+        if self.gradient_checkpointing:
+            FlaxLongT5CheckpointLayer = remat(FlaxLongT5LayerCollection, static_argnums=(6, 7, 8))
+            self.blocks = [
+                FlaxLongT5CheckpointLayer(
+                    self.config,
+                    has_relative_attention_bias=(i == 0),
+                    dtype=self.dtype,
+                    name=str(i),
+                )
+                for i in range(self.config.num_layers)
+            ]
+        else:
+            self.blocks = [
+                FlaxLongT5LayerCollection(
+                    self.config,
+                    has_relative_attention_bias=(i == 0),
+                    dtype=self.dtype,
+                    name=str(i),
+                )
+                for i in range(self.config.num_layers)
+            ]
 
     def __call__(
         self,
@@ -1409,14 +1429,14 @@ def __call__(
 
             layer_outputs = layer_module(
                 hidden_states,
-                attention_mask=attention_mask,
-                position_bias=position_bias,
-                encoder_hidden_states=encoder_hidden_states,
-                encoder_attention_mask=encoder_attention_mask,
-                encoder_decoder_position_bias=encoder_decoder_position_bias,
-                output_attentions=output_attentions,
-                deterministic=deterministic,
-                init_cache=init_cache,
+                attention_mask,
+                position_bias,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                encoder_decoder_position_bias,
+                output_attentions,
+                deterministic,
+                init_cache,
             )
 
             hidden_states = layer_outputs[0]
@@ -1447,11 +1467,14 @@ class FlaxLongT5Stack(nn.Module):
     config: LongT5Config
     embed_tokens: nn.Embed
     dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    gradient_checkpointing: bool = False
 
     def setup(self):
         self.causal = self.config.causal
 
-        self.block = FlaxLongT5BlockCollection(self.config, dtype=self.dtype)
+        self.block = FlaxLongT5BlockCollection(
+            self.config, dtype=self.dtype, gradient_checkpointing=self.gradient_checkpointing
+        )
         self.final_layer_norm = FlaxLongT5LayerNorm(
             self.config.d_model, eps=self.config.layer_norm_epsilon, dtype=self.dtype
         )
@@ -1989,6 +2012,7 @@ def _decoder_forward(module, decoder_input_ids, decoder_attention_mask, **kwargs
 class FlaxLongT5Module(nn.Module):
     config: LongT5Config
     dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    gradient_checkpointing: bool = False
 
     def _get_encoder_module(self):
         return self.encoder
@@ -2005,12 +2029,22 @@ def setup(self):
 
         encoder_config = copy.deepcopy(self.config)
         encoder_config.causal = False
-        self.encoder = FlaxLongT5Stack(encoder_config, embed_tokens=self.shared, dtype=self.dtype)
+        self.encoder = FlaxLongT5Stack(
+            encoder_config,
+            embed_tokens=self.shared,
+            dtype=self.dtype,
+            gradient_checkpointing=self.gradient_checkpointing,
+        )
 
         decoder_config = copy.deepcopy(self.config)
         decoder_config.causal = True
         decoder_config.num_layers = self.config.num_decoder_layers
-        self.decoder = FlaxLongT5Stack(decoder_config, embed_tokens=self.shared, dtype=self.dtype)
+        self.decoder = FlaxLongT5Stack(
+            decoder_config,
+            embed_tokens=self.shared,
+            dtype=self.dtype,
+            gradient_checkpointing=self.gradient_checkpointing,
+        )
 
     def __call__(
         self,
@@ -2104,6 +2138,7 @@ class FlaxLongT5Model(FlaxLongT5PreTrainedModel):
 class FlaxLongT5ForConditionalGenerationModule(nn.Module):
     config: LongT5Config
     dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+    gradient_checkpointing: bool = False
 
     def _get_encoder_module(self):
         return self.encoder
@@ -2124,13 +2159,17 @@ def setup(self):
         encoder_config.causal = False
         encoder_config.use_cache = False
         encoder_config.is_encoder_decoder = False
-        self.encoder = FlaxLongT5Stack(encoder_config, self.shared, dtype=self.dtype)
+        self.encoder = FlaxLongT5Stack(
+            encoder_config, self.shared, dtype=self.dtype, gradient_checkpointing=self.gradient_checkpointing
+        )
 
         decoder_config = copy.deepcopy(self.config)
         decoder_config.causal = True
         decoder_config.is_encoder_decoder = False
         decoder_config.num_layers = self.config.num_decoder_layers
-        self.decoder = FlaxLongT5Stack(decoder_config, self.shared, dtype=self.dtype)
+        self.decoder = FlaxLongT5Stack(
+            decoder_config, self.shared, dtype=self.dtype, gradient_checkpointing=self.gradient_checkpointing
+        )
 
         self.lm_head = nn.Dense(
             self.config.vocab_size,