Add a Causal LM model for Mistral

keras_nlp/models/__init__.py

@@ -93,6 +93,12 @@
 from keras_nlp.models.gpt_neo_x.gpt_neo_x_tokenizer import GPTNeoXTokenizer
 from keras_nlp.models.llama.llama_backbone import LlamaBackbone
 from keras_nlp.models.mistral.mistral_backbone import MistralBackbone
+from keras_nlp.models.mistral.mistral_causal_lm import MistralCausalLM
+from keras_nlp.models.mistral.mistral_causal_lm_preprocessor import (
+    MistralCausalLMPreprocessor,
+)
+from keras_nlp.models.mistral.mistral_preprocessor import MistralPreprocessor
+from keras_nlp.models.mistral.mistral_tokenizer import MistralTokenizer
 from keras_nlp.models.opt.opt_backbone import OPTBackbone
 from keras_nlp.models.opt.opt_causal_lm import OPTCausalLM
 from keras_nlp.models.opt.opt_causal_lm_preprocessor import (

keras_nlp/models/mistral/mistral_attention.py

@@ -136,7 +136,6 @@ def call(
         cache_update_index=None,
         training=None,
     ):
-        seq_len = ops.shape(hidden_states)[1]
         start_index = (
             cache_update_index if cache_update_index is not None else 0
         )
@@ -148,89 +147,34 @@ def call(
 
         query = self._query_dense(hidden_states)
 
-        # Note that the original PyTorch implementation uses
-        # view_as_complex/view_as_real while we use split/concatenate to
-        # convert to/from complex numbers. The transformations below make
-        # the rope computation numerically equivalent to the original
-        # implementation.
-        def _mistral_rope(x):
-            x = ops.concatenate([x[..., ::2], x[..., 1::2]], axis=-1)
-            x = self.rotary_embedding_layer(x, start_index=start_index)
-            x = ops.reshape(
-                ops.stack(ops.split(x, 2, axis=-1), axis=-1), ops.shape(x)
-            )
-            return x
-
         # Compute RoPE for queries
-        query = _mistral_rope(query)
+        query = self.rotary_embedding_layer(query, start_index=start_index)
 
         def _compute_key_value(x):
             key, value = self._key_dense(x), self._value_dense(x)
-            key = _mistral_rope(key)
+            # Compute RoPE for keys
+            key = self.rotary_embedding_layer(key, start_index=start_index)
             return key, value
 
         if cache is not None:
-            cache_k = cache[:, 0, ...]
-            cache_v = cache[:, 1, ...]
-
+            key_cache = cache[:, 0, ...]
+            value_cache = cache[:, 1, ...]
+            if cache_update_index is None:
+                key = key_cache
+                value = value_cache
+            else:
+                key_update, value_update = _compute_key_value(hidden_states)
+                start = [0, cache_update_index, 0, 0]
+                key = ops.slice_update(key_cache, start, key_update)
+                value = ops.slice_update(value_cache, start, value_update)
+                cache = ops.stack((key, value), axis=1)
+        else:
             if cache_update_index is not None:
-                # Compute the new keys and values
-                key, value = _compute_key_value(hidden_states)
-
-                # Cache is a rotating buffer, we want to warp around if
-                # the sequence length exceeds the sliding window.
-                update_end_index = (
-                    cache_update_index + seq_len - 1
-                ) % self._sliding_window + 1
-                update_end_index = ops.cast(update_end_index, "int32")
-                cache_update_index = cache_update_index % self._sliding_window
-                update_start_index = ops.cond(
-                    update_end_index > cache_update_index,
-                    lambda: ops.cast(cache_update_index, "int32"),
-                    lambda: ops.cast(0, "int32"),
-                )
-                # Also note that the update step below assumes that the
-                # sequence length is always one when `cache_update_index != 0`.
-                # This is necessary to support XLA compilation. Ideally, we
-                # would want to use
-                # `key[:, -(update_end_index - update_start_index):, ...]`
-                # as the update but updating using a dynamic slice gives an
-                # XLA compilation error in TensorFlow.
-                # Passing a sequence of length > 1 with cache update might give
-                # incorrect results (since there is no way to determine how
-                # many most recent tokens are to be saved if the tokens exceed
-                # the sliding window length).
-                cache_k = ops.slice_update(
-                    cache_k,
-                    [0, update_start_index, 0, 0],
-                    # We slice the keys and values since if the user has passed
-                    # a sequence of length > `self._sliding_window`. We want to
-                    # prefill the cache using just the most recent values in the
-                    # sliding window.
-                    ops.cast(
-                        key[:, -self._sliding_window :, ...], cache_k.dtype
-                    ),
+                raise ValueError(
+                    "`cache_update_index` should not be set if `cache` is "
+                    f"`None`. Received: cache={cache}, "
+                    f"cache_update_index={cache_update_index}"
                 )
-                cache_v = ops.slice_update(
-                    cache_v,
-                    [0, update_start_index, 0, 0],
-                    ops.cast(
-                        value[:, -self._sliding_window :, ...], cache_v.dtype
-                    ),
-                )
-                cache = ops.stack([cache_k, cache_v], axis=1)
-
-                # Get the required keys and values from the cache.
-                # Since we expect the user to pass a fixed-size cache, we just
-                # pick the first few slices up-to and including the newly computed
-                # keys and values.
-                cache_k = cache_k[:, :update_end_index, ...]
-                cache_v = cache_v[:, :update_end_index, ...]
-
-            key = ops.cast(cache_k, dtype=self.compute_dtype)
-            value = ops.cast(cache_v, dtype=self.compute_dtype)
-        else:
-            # Compute keys and values
             key, value = _compute_key_value(hidden_states)
 
         # [batch_shape, seq_len, num_key_value_heads, head_dim]
@@ -260,7 +204,7 @@ def _masked_softmax(self, attention_scores, attention_mask=None):
         return self._softmax(attention_scores)
 
     def _compute_attention(self, query, key, value, attention_mask=None):
-        attention_scores = ops.einsum(self._dot_product_equation, key, query)
+        attention_scores = ops.einsum(self._dot_product_equation, query, key)
 
         norm_factor = ops.sqrt(ops.cast(self._head_dim, self.compute_dtype))
 

keras_nlp/models/mistral/mistral_causal_lm.py

@@ -0,0 +1,213 @@
+# Copyright 2023 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.backend import ops
+from keras_nlp.models.generative_task import GenerativeTask
+from keras_nlp.models.mistral.mistral_backbone import MistralBackbone
+from keras_nlp.models.mistral.mistral_causal_lm_preprocessor import (
+    MistralCausalLMPreprocessor,
+)
+from keras_nlp.utils.python_utils import classproperty
+
+
+@keras_nlp_export("keras_nlp.models.MistralCausalLM")
+class MistralCausalLM(GenerativeTask):
+    """An end-to-end Mistral model for causal language modeling.
+
+    A causal language model (LM) predicts the next token based on previous
+    tokens. This task setup can be used to train the model unsupervised on
+    plain text input, or to autoregressively generate plain text similar to
+    the data used for training. This task can be used for pre-training or
+    fine-tuning a GPT-NeoX model, simply by calling `fit()`.
+
+    This model has a `generate()` method, which generates text based on a
+    prompt. The generation strategy used is controlled by an additional
+    `sampler` argument on `compile()`. You can recompile the model with
+    different `keras_nlp.samplers` objects to control the generation. By
+    default, `"top_k"` sampling will be used.
+
+    Args:
+        backbone: A `keras_nlp.models.MistralBackbone` instance.
+        preprocessor: A `keras_nlp.models.MistralCausalLMPreprocessor` or `None`.
+            If `None`, this model will not apply preprocessing, and inputs
+            should be preprocessed before calling the model.
+    """
+
+    def __init__(self, backbone, preprocessor=None, **kwargs):
+        # === Layers ===
+        self.backbone = backbone
+        self.preprocessor = preprocessor
+
+        # === Functional Model ===
+        inputs = backbone.inputs
+        hidden_states = backbone(inputs)
+        outputs = backbone.token_embedding(hidden_states, reverse=True)
+        super().__init__(
+            inputs=inputs,
+            outputs=outputs,
+            **kwargs,
+        )
+
+        # === Default compilation ===
+        self.compile(
+            loss=keras.losses.SparseCategoricalCrossentropy(from_logits=True),
+            optimizer=keras.optimizers.Adam(2e-5),
+            metrics=[keras.metrics.SparseCategoricalAccuracy()],
+            jit_compile=True,
+        )
+
+    @classproperty
+    def backbone_cls(cls):
+        return MistralBackbone
+
+    @classproperty
+    def preprocessor_cls(cls):
+        return MistralCausalLMPreprocessor
+
+    def call_with_cache(
+        self,
+        token_ids,
+        cache,
+        cache_update_index,
+    ):
+        """Forward pass of `MistralCausalLM` with cache.
+
+        `call_with_cache` adds an additional forward pass for the model for
+        autoregressive inference. Unlike calling the model directly, this method
+        allows caching previous key/value Tensors in multi-head attention layer,
+        and avoids recomputing the outputs of seen tokens.
+
+        Args:
+            token_ids: a dense int Tensor with shape `(batch_size, max_length)`.
+            cache: a dense float Tensor, the cache of key and value.
+            cache_update_index: int, or int Tensor. The index of current inputs
+            in the whole sequence.
+
+        Returns:
+            A (logits, hidden_states, cache) tuple. Where `logits` is the
+            language model logits for the input token_ids, `hidden_states` is
+            the final hidden representation of the input tokens, and `cache` is
+            the decoding cache.
+        """
+        x = self.backbone.token_embedding(token_ids)
+        # Each decoder layer has a cache; we update them separately.
+        updated_cache = []
+        for i in range(self.backbone.num_layers):
+            current_cache = cache[:, i, ...]
+            x, next_cache = self.backbone.transformer_layers[i](
+                x,
+                self_attention_cache=current_cache,
+                self_attention_cache_update_index=cache_update_index,
+            )
+            updated_cache.append(next_cache)
+        cache = ops.stack(updated_cache, axis=1)
+        hidden_states = x = self.backbone.layer_norm(x)
+        logits = self.backbone.token_embedding(x, reverse=True)
+        return logits, hidden_states, cache
+
+    def _build_cache(self, token_ids):
+        """Build an empty cache for use with `call_with_cache()`."""
+        batch_size = ops.shape(token_ids)[0]
+        max_length = ops.shape(token_ids)[1]
+        num_layers = self.backbone.num_layers
+        num_key_value_heads = self.backbone.num_key_value_heads
+        head_dim = self.backbone.hidden_dim // self.backbone.num_query_heads
+        shape = [
+            batch_size,
+            num_layers,
+            2,
+            max_length,
+            num_key_value_heads,
+            head_dim,
+        ]
+        cache = ops.zeros(shape, dtype=self.compute_dtype)
+        # Seed the cache.
+        _, hidden_states, cache = self.call_with_cache(token_ids, cache, 0)
+        return hidden_states, cache
+
+    def generate_step(
+        self,
+        inputs,
+        end_token_id=None,
+    ):
+        """A compilable generation function for a single batch of inputs.
+
+        This function represents the inner, XLA-compilable, generation function
+        for a single batch of inputs. Inputs should have the same structure as
+        model inputs, a dictionary with keys `"token_ids"` and `"padding_mask"`.
+
+        Args:
+            inputs: A dictionary with two keys `"token_ids"` and
+                `"padding_mask"` and batched tensor values.
+            end_token_id: The id of the end token to stop on. If all
+                sequences have produced a new `end_token_id`, generation
+                will stop.
+        """
+        token_ids, padding_mask = inputs["token_ids"], inputs["padding_mask"]
+        # Create and seed cache with a single forward pass.
+        hidden_states, cache = self._build_cache(token_ids)
+        # Compute the lengths of all user inputted tokens ids.
+        row_lengths = ops.sum(ops.cast(padding_mask, "int32"), axis=-1)
+        # Start at the first index that has no user inputted id.
+        index = ops.min(row_lengths)
+
+        def next(prompt, cache, index):
+            # The cache index is the index of our previous token.
+            cache_update_index = index - 1
+            batch_size = ops.shape(prompt)[0]
+            prompt = ops.slice(prompt, [0, cache_update_index], [batch_size, 1])
+            logits, hidden_states, cache = self.call_with_cache(
+                prompt,
+                cache,
+                cache_update_index,
+            )
+            return (
+                ops.squeeze(logits, axis=1),
+                ops.squeeze(hidden_states, axis=1),
+                cache,
+            )
+
+        token_ids = self._sampler(
+            next=next,
+            prompt=token_ids,
+            cache=cache,
+            index=index,
+            mask=padding_mask,
+            end_token_id=end_token_id,
+            hidden_states=hidden_states,
+        )
+
+        # Compute an output padding mask with the token ids we updated.
+        if end_token_id is not None:
+            # Build a mask of `end_token_id` locations not in the original
+            # prompt (not in locations where `padding_mask` is True).
+            end_locations = ops.logical_and(
+                ops.equal(token_ids, end_token_id),
+                ops.logical_not(padding_mask),
+            )
+            end_locations = ops.cast(end_locations, "int32")
+            # Use cumsum to get ones in all locations after end_locations.
+            cumsum = ops.cast(ops.cumsum(end_locations, axis=-1), "int32")
+            overflow = cumsum - end_locations
+            # Our padding mask is the inverse of these overflow locations.
+            padding_mask = ops.logical_not(ops.cast(overflow, "bool"))
+        else:
+            # Without early stopping, all locations will have been updated.
+            padding_mask = ops.ones_like(token_ids, dtype="bool")
+        return {
+            "token_ids": token_ids,
+            "padding_mask": padding_mask,
+        }