Add FalconBackbone (keras-team#1475)

* Add Falcon backbone.

* Add docstring.

* Add dtype.

* Add checkpoint conversion script.

* Fix tests.

* Random fixes.

* Add cache.

* Cast cumsum to int32.

* Make sublayers public.

* Address backbone comments.

* Update attention computation to use einsum.

* Falcon only works with Keras3.

* Fix tests.

* Remove falcon_causal_lm file.

* Remove commented/unused codes.

keras_nlp/models/falcon/__init__.py

@@ -0,0 +1,13 @@
+# Copyright 2024 The KerasNLP Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.

keras_nlp/models/falcon/falcon_attention.py

@@ -0,0 +1,156 @@
+# Copyright 2024 The KerasNLP Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import math
+
+from keras_nlp.backend import keras
+from keras_nlp.backend import ops
+
+
+class FalconAttention(keras.layers.Layer):
+    def __init__(
+        self,
+        num_heads,
+        attention_dropout_rate,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.num_heads = num_heads
+        self.attention_dropout_rate = attention_dropout_rate
+
+    def build(self, inputs_shape):
+        # Einsum variables:
+        # b = batch size
+        # q = query length
+        # m = model dim
+        # n = num attention heads
+        # h = head dim
+        # k = key/value length
+
+        batch_size, seq_length, hidden_dim = inputs_shape
+
+        self.head_dim = hidden_dim // self.num_heads
+
+        # Layer-wise attention scaling
+        self.inv_norm_factor = 1.0 / math.sqrt(self.head_dim)
+
+        self.query_dense = keras.layers.EinsumDense(
+            equation="bqm,mnh->bqnh",
+            output_shape=(None, self.num_heads, self.head_dim),
+            bias_axes="nh",
+            dtype=self.dtype_policy,
+            name="query_dense",
+        )
+        self.query_dense.build(inputs_shape)
+
+        self.key_dense = keras.layers.EinsumDense(
+            equation="bkm,mnh->bknh",
+            output_shape=(None, self.num_heads, self.head_dim),
+            bias_axes="nh",
+            dtype=self.dtype_policy,
+            name="key_dense",
+        )
+        self.key_dense.build(inputs_shape)
+
+        self.value_dense = keras.layers.EinsumDense(
+            equation="bkm,mnh->bknh",
+            output_shape=(None, self.num_heads, self.head_dim),
+            bias_axes="nh",
+            dtype=self.dtype_policy,
+            name="value_dense",
+        )
+        self.value_dense.build(inputs_shape)
+
+        self.attention_dropout = keras.layers.Dropout(
+            rate=self.attention_dropout_rate,
+            dtype=self.dtype_policy,
+            name="attention_dropout",
+        )
+
+        self.output_dense = keras.layers.Dense(
+            hidden_dim,
+            dtype=self.dtype_policy,
+            name="output_dense",
+        )
+        self.output_dense.build(inputs_shape)
+
+        self.softmax = keras.layers.Softmax(dtype="float32", name="softmax")
+
+        self.built = True
+
+    def call(
+        self,
+        inputs,
+        alibi,
+        attention_mask=None,
+        cache=None,
+        cache_update_index=None,
+    ):
+        batch_size, seq_length, hidden_dim = ops.shape(inputs)
+
+        query = self.query_dense(inputs)
+        key = self.key_dense(inputs)
+        value = self.value_dense(inputs)
+
+        if cache is not None:
+            key_cache = cache[:, 0, ...]
+            value_cache = cache[:, 1, ...]
+            if cache_update_index is None:
+                key = key_cache
+                value = value_cache
+            else:
+                start = [0, cache_update_index, 0, 0]
+                key = ops.slice_update(key_cache, start, key)
+                value = ops.slice_update(value_cache, start, value)
+                cache = ops.stack((key, value), axis=1)
+        else:
+            if cache_update_index is not None:
+                raise ValueError(
+                    "`cache_update_index` should not be set if `cache` is "
+                    f"`None`. Received: cache={cache}, "
+                    f"cache_update_index={cache_update_index}"
+                )
+
+        attention_scores = ops.einsum("bqnh,bknh->bnqk", query, key)
+        attention_scores = ops.add(attention_scores, alibi)
+        attention_scores = (
+            attention_scores * self.inv_norm_factor
+        )  # [batch_size, num_heads, query_length, kv_length]
+        attention_scores = self.softmax(
+            attention_scores, ops.expand_dims(attention_mask, 1)
+        )
+        attention_scores = self.attention_dropout(attention_scores)
+        attention_output = ops.einsum(
+            "bnqk,bknh->bqnh", attention_scores, value
+        )
+        attention_output = ops.reshape(
+            attention_output,
+            [batch_size, seq_length, self.num_heads * self.head_dim],
+        )  # [batch_size, query_length, hidden_dim]
+
+        attention_output = self.output_dense(attention_output)
+
+        if cache is not None:
+            return attention_output, cache
+
+        return attention_output
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "num_heads": self.num_heads,
+                "attention_dropout_rate": self.attention_dropout_rate,
+            }
+        )
+        return config

keras_nlp/models/falcon/falcon_backbone.py

@@ -0,0 +1,160 @@
+# Copyright 2024 The KerasNLP Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from keras_nlp.api_export import keras_nlp_export
+from keras_nlp.backend import keras
+from keras_nlp.layers.modeling.reversible_embedding import ReversibleEmbedding
+from keras_nlp.models.backbone import Backbone
+from keras_nlp.models.falcon.falcon_transformer_decoder import (
+    FalconTransformerDecoder,
+)
+
+
+@keras_nlp_export("keras_nlp.models.FalconBackbone")
+class FalconBackbone(Backbone):
+    """The Falcon core architecure.
+
+    This network implements a Transformer-based decoder-only network,
+    [Falcon](https://arxiv.org/abs/2306.01116).
+
+    Args:
+        vocabulary_size: int. The size of the token vocabulary.
+        num_layers: int. The number of transformer layers.
+        num_attention_heads: int. The number of attention heads for each transformer.
+            The hidden size must be divisible by the number of attention heads.
+        hidden_dim: int. The dimensionality of the embeddings and hidden states.
+        intermediate_dim: int. The output dimension of the first Dense layer in
+            the MLP network of each transformer.
+        layer_norm_epsilon: float. Epsilon for the layer normalization layers in
+            the transformer decoder.
+        attention_dropout_rate: float. Dropout probability for the attention.
+        feedforward_dropout_rate: flaot. Dropout probability for the feedforward.
+        dtype: string or `keras.mixed_precision.DTypePolicy`. The dtype to use
+            for model computations and weights. Note that some computations,
+            such as softmax and layer normalization, will always be done at
+            float32 precision regardless of dtype.
+
+    Examples:
+    ```python
+    input_data = {
+        "token_ids": np.ones(shape=(1, 12), dtype="int32"),
+        "padding_mask": np.array([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]),
+    }
+
+    # Pretrained Falcon decoder.
+    # TODO: Update the preset.
+    model = keras_nlp.models.FalconBackbone.from_preset("falcon_preset")
+    model(input_data)
+
+    # Randomly initialized Falcon decoder with a custom config.
+    model = keras_nlp.models.FalconBackbone(
+        vocabulary_size=10,
+        num_layers=2,
+        num_attention_heads=2,
+        hidden_dim=32,
+        intermediate_dim=32*4,
+        layer_norm_epsilon=1e-5,
+        attention_dropout_rate=0,
+        feedforward_dropout_rate=0,
+        dtype="float32",
+    )
+    model(input_data)
+    ```
+    """
+
+    def __init__(
+        self,
+        vocabulary_size,
+        num_layers,
+        num_attention_heads,
+        hidden_dim,
+        intermediate_dim,
+        layer_norm_epsilon=1e-5,
+        attention_dropout_rate=0,
+        feedforward_dropout_rate=0,
+        dtype=None,
+        **kwargs,
+    ):
+        # === Layers ===
+        self.token_embedding = ReversibleEmbedding(
+            input_dim=vocabulary_size,
+            output_dim=hidden_dim,
+            dtype=dtype,
+            name="token_embedding",
+        )
+
+        self.transformer_layers = []
+        for i in range(num_layers):
+            layer = FalconTransformerDecoder(
+                num_attention_heads=num_attention_heads,
+                intermediate_dim=intermediate_dim,
+                attention_dropout_rate=attention_dropout_rate,
+                feedforward_dropout_rate=feedforward_dropout_rate,
+                dtype=dtype,
+                name=f"transformer_layer_{i}",
+            )
+            self.transformer_layers.append(layer)
+
+        self.final_layernorm = keras.layers.LayerNormalization(
+            epsilon=layer_norm_epsilon,
+            dtype=dtype,
+            name="final_layernorm",
+        )
+
+        # === Functional Model ===
+        token_ids = keras.Input(shape=(None,), dtype="int32", name="token_ids")
+        padding_mask = keras.Input(
+            shape=(None,), dtype="int32", name="padding_mask"
+        )
+        # Embed Tokens.
+        x = self.token_embedding(token_ids)
+
+        # Apply successive transformer decoder blocks.
+        for transformer_layer in self.transformer_layers:
+            x = transformer_layer(inputs=x, decoder_padding_mask=padding_mask)
+        sequence_output = self.final_layernorm(x)
+
+        super().__init__(
+            inputs={
+                "token_ids": token_ids,
+                "padding_mask": padding_mask,
+            },
+            outputs=sequence_output,
+            **kwargs,
+        )
+
+        # === Config ===
+        self.vocabulary_size = vocabulary_size
+        self.num_layers = num_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_dim = hidden_dim
+        self.intermediate_dim = intermediate_dim
+        self.attention_dropout_rate = attention_dropout_rate
+        self.feedforward_dropout_rate = feedforward_dropout_rate
+        self.layer_norm_epsilon = layer_norm_epsilon
+
+    def get_config(self):
+        config = super().get_config()
+        config.update(
+            {
+                "vocabulary_size": self.vocabulary_size,
+                "num_layers": self.num_layers,
+                "num_attention_heads": self.num_attention_heads,
+                "hidden_dim": self.hidden_dim,
+                "intermediate_dim": self.intermediate_dim,
+                "attention_dropout_rate": self.attention_dropout_rate,
+                "feedforward_dropout_rate": self.feedforward_dropout_rate,
+                "layer_norm_epsilon": self.layer_norm_epsilon,
+            }
+        )
+        return config

keras_nlp/models/falcon/falcon_backbone_test.py

@@ -0,0 +1,49 @@
+# Copyright 2024 The KerasNLP Authors
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     https://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import pytest
+
+from keras_nlp.backend import ops
+from keras_nlp.models.falcon.falcon_backbone import FalconBackbone
+from keras_nlp.tests.test_case import TestCase
+
+
+class FalconBackboneTest(TestCase):
+    def setUp(self):
+        self.init_kwargs = {
+            "vocabulary_size": 10,
+            "num_layers": 2,
+            "num_attention_heads": 8,
+            "hidden_dim": 16,
+            "intermediate_dim": 32,
+        }
+        self.input_data = {
+            "token_ids": ops.ones((2, 5), dtype="int32"),
+            "padding_mask": ops.ones((2, 5), dtype="int32"),
+        }
+
+    def test_backbone_basics(self):
+        self.run_backbone_test(
+            cls=FalconBackbone,
+            init_kwargs=self.init_kwargs,
+            input_data=self.input_data,
+            expected_output_shape=(2, 5, 16),
+        )
+
+    @pytest.mark.large
+    def test_saved_model(self):
+        self.run_model_saving_test(
+            cls=FalconBackbone,
+            init_kwargs=self.init_kwargs,
+            input_data=self.input_data,
+        )