Add quantization support for Gemma, Gemma2 and PaliGemma (#1670)

* Introduce quantization support to `Gemma`

* Revert `SentencePieceTokenizer`

* Add tests for `PaliGemma`

* Add quantization support for Gemma2

* Address comments

keras_nlp/src/layers/modeling/reversible_embedding.py

@@ -107,8 +107,7 @@ def __init__(
 
     def build(self, inputs_shape=None):
         super().build(inputs_shape)
-
-        if not self.tie_weights:
+        if not self.tie_weights and self.quantization_mode != "int8":
             self.reverse_embeddings = self.add_weight(
                 name="reverse_embeddings",
                 shape=(self.output_dim, self.input_dim),
@@ -143,20 +142,28 @@ def save_own_variables(self, store):
         if not self.built:
             return
         super().save_own_variables(store)
-        # Before Keras 3.2, the reverse weight is saved in the super() call.
-        # After Keras 3.2, the reverse weight must be saved manually.
-        if len(store.keys()) < len(self.weights):
-            # Store the reverse embedding as the last weight.
-            store[str(len(store.keys()))] = self.reverse_embeddings
+        target_variables = []
+        if not self.tie_weights:
+            # Store the reverse embedding weights as the last weights.
+            target_variables.append(self.reverse_embeddings)
+            if self.quantization_mode == "int8":
+                target_variables.append(self.reverse_embeddings_scale)
+            for i, variable in enumerate(target_variables, start=len(store)):
+                store[str(i)] = variable
 
     def load_own_variables(self, store):
         if not self.built:
             self.build()
         super().load_own_variables(store)
         if not self.tie_weights:
-            # Last weight in the store is the reverse embedding weights.
-            key = str(len(store.keys()) - 1)
-            self.reverse_embeddings.assign(store[key])
+            # Last weights in the stores are the reverse embedding weights.
+            target_variables = [self.reverse_embeddings]
+            if self.quantization_mode == "int8":
+                target_variables.append(self.reverse_embeddings_scale)
+            for i, variable in enumerate(
+                target_variables, start=len(store) - len(target_variables)
+            ):
+                variable.assign(store[str(i)])
 
     def compute_output_spec(self, inputs, reverse=False):
         output_shape = list(inputs.shape)
@@ -165,3 +172,105 @@ def compute_output_spec(self, inputs, reverse=False):
         else:
             output_shape += [self.output_dim]
         return keras.KerasTensor(output_shape, dtype=self.dtype)
+
+    # Quantization-related (int8) methods
+
+    def quantized_call(self, inputs, reverse=False):
+        # TODO (hongyu): This function could be removed once we add `*args` and
+        # `**kwargs` for `Embedding.quantized_call`
+        if self.quantization_mode == "int8":
+            return self._int8_call(inputs, reverse=reverse)
+        else:
+            self._quantization_mode_error(self.quantization_mode)
+
+    def _int8_build(
+        self,
+        embeddings_initializer="zeros",
+        embeddings_scale_initializer="ones",
+        reverse_embeddings_initializer="zeros",
+        reverse_embeddings_scale_initializer="ones",
+    ):
+        super()._int8_build(
+            embeddings_initializer, embeddings_scale_initializer
+        )
+        self.inputs_quantizer = keras.quantizers.AbsMaxQuantizer(axis=-1)
+        if not self.tie_weights:
+            self.reverse_embeddings = self.add_weight(
+                name="reverse_embeddings",
+                shape=(self.output_dim, self.input_dim),
+                initializer=reverse_embeddings_initializer,
+                dtype="int8",
+                trainable=False,
+            )
+            self.reverse_embeddings_scale = self.add_weight(
+                name="reverse_embeddings_scale",
+                shape=(self.input_dim,),
+                initializer=reverse_embeddings_scale_initializer,
+                trainable=False,
+            )
+
+    def _int8_call(self, inputs, reverse=False):
+        if reverse:
+            if self.tie_weights:
+                kernel = ops.transpose(self._embeddings)
+                scale = ops.transpose(self.embeddings_scale)
+            else:
+                kernel = self.reverse_embeddings
+                scale = self.reverse_embeddings_scale
+            inputs, inputs_scale = self.inputs_quantizer(inputs)
+            outputs = ops.matmul(inputs, kernel)
+            # De-scale outputs
+            outputs = ops.cast(outputs, self.compute_dtype)
+            outputs = ops.divide(outputs, ops.multiply(inputs_scale, scale))
+            return outputs
+
+        return super()._int8_call(inputs)
+
+    def quantize(self, mode):
+        import gc
+
+        if type(self) is not ReversibleEmbedding:
+            raise NotImplementedError(
+                f"Layer {self.__class__.__name__} does not have a `quantize()` "
+                "method implemented."
+            )
+        self._check_quantize_args(mode, self.compute_dtype)
+
+        self._tracker.unlock()
+        if mode == "int8":
+            embeddings, embeddings_scale = keras.quantizers.abs_max_quantize(
+                self._embeddings, axis=-1
+            )
+            embeddings_scale = ops.squeeze(embeddings_scale, axis=-1)
+            self._untrack_variable(self._embeddings)
+            del self._embeddings
+            if not self.tie_weights:
+                reverse_embeddings, reverse_embeddings_scale = (
+                    keras.quantizers.abs_max_quantize(
+                        self.reverse_embeddings, axis=0
+                    )
+                )
+                reverse_embeddings_scale = ops.squeeze(
+                    reverse_embeddings_scale, axis=0
+                )
+                self._untrack_variable(self.reverse_embeddings)
+                del self.reverse_embeddings
+            else:
+                reverse_embeddings = None
+                reverse_embeddings_scale = None
+            self._int8_build(
+                lambda shape, dtype: embeddings,
+                lambda shape, dtype: embeddings_scale,
+                lambda shape, dtype: reverse_embeddings,
+                lambda shape, dtype: reverse_embeddings_scale,
+            )
+        else:
+            raise self._quantization_mode_error(mode)
+        self._tracker.lock()
+
+        if self.dtype_policy.quantization_mode is None:
+            policy = keras.dtype_policies.get(
+                f"{mode}_from_{self.dtype_policy.name}"
+            )
+            self.dtype_policy = policy
+        gc.collect()

keras_nlp/src/layers/modeling/reversible_embedding_test.py

@@ -98,3 +98,71 @@ def test_reverse_dtype(self):
         output_data = embedding(input_data, reverse=True)
         self.assertEqual(output_data.shape, (4, 10, 100))
         self.assertDTypeEqual(output_data, "float16")
+
+    @parameterized.named_parameters(
+        ("tie_weights", True), ("untie_weights", False)
+    )
+    def test_quantize_int8(self, tie_weights):
+        layer_config = dict(
+            input_dim=100, output_dim=32, tie_weights=tie_weights
+        )
+        layer = ReversibleEmbedding(**layer_config)
+        layer.build()
+        x = random.randint(shape=(64, 100), minval=0, maxval=9)
+        x_reverse = random.uniform(shape=(64, 32))
+        y_float = layer(x)
+        y_reverse_float = layer(x_reverse, reverse=True)
+        layer.quantize("int8")
+
+        # Verify weights dtype
+        if not tie_weights:
+            self.assertEqual(
+                keras.backend.standardize_dtype(layer.reverse_embeddings.dtype),
+                "int8",
+            )
+            self.assertEqual(
+                keras.backend.standardize_dtype(
+                    layer.reverse_embeddings_scale.dtype
+                ),
+                layer.variable_dtype,
+            )
+
+        # Try eager call and verify output correctness
+        y_quantized = layer(x)
+        y_reverse_quantized = layer(x_reverse, reverse=True)
+        mse = ops.mean(ops.square(y_float - y_quantized))
+        mse_reverse = ops.mean(
+            ops.square(y_reverse_float - y_reverse_quantized)
+        )
+        self.assertLess(mse, 1e-3)  # A weak correctness test
+        self.assertLess(mse_reverse, 1e-3)  # A weak correctness test
+
+        # Try saving and reloading the model
+        model = keras.models.Sequential([layer])
+        temp_filepath = os.path.join(
+            self.get_temp_dir(), "quantized_model.keras"
+        )
+        model.save(temp_filepath)
+        new_model = keras.models.load_model(temp_filepath)
+        self.assertAllClose(model.predict(x), new_model.predict(x))
+
+    @parameterized.named_parameters(
+        ("tie_weights", True),
+        ("untie_weights", False),
+    )
+    def test_quantize_dtype_argument(self, tie_weights):
+        self.run_layer_test(
+            cls=ReversibleEmbedding,
+            init_kwargs={
+                "input_dim": 100,
+                "output_dim": 32,
+                "tie_weights": tie_weights,
+                "embeddings_initializer": "HeNormal",
+                "dtype": "int8_from_float32",
+            },
+            input_data=random.randint(minval=0, maxval=100, shape=(4, 10)),
+            expected_output_shape=(4, 10, 32),
+            expected_num_trainable_weights=0,
+            expected_num_non_trainable_weights=2 if tie_weights else 4,
+            expected_num_non_trainable_variables=2 if tie_weights else 4,
+        )

keras_nlp/src/models/backbone.py

@@ -75,11 +75,7 @@ def __init__(self, *args, dtype=None, **kwargs):
             id(layer) for layer in self._flatten_layers()
         )
         self._initialized = True
-        if dtype is not None:
-            if isinstance(dtype, keras.DTypePolicy):
-                self.dtype_policy = dtype
-            else:
-                self.dtype_policy = keras.DTypePolicy(dtype)
+        self.dtype_policy = keras.dtype_policies.get(dtype)
 
     def __setattr__(self, name, value):
         # Work around setattr issues for Keras 2 and Keras 3 torch backend.
@@ -107,11 +103,20 @@ def token_embedding(self, value):
     def get_config(self):
         # Don't chain to super here. `get_config()` for functional models is
         # a nested layer config and cannot be passed to Backbone constructors.
-        return {
+        config = {
             "name": self.name,
             "trainable": self.trainable,
         }
 
+        # Add quantization support by utilizing `DTypePolicyMap`
+        policy_map = keras.dtype_policies.DTypePolicyMap()
+        for layer in self._flatten_layers():
+            if layer.quantization_mode is not None:
+                policy_map[layer.path] = layer.dtype_policy
+        if len(policy_map) > 0:
+            config.update({"dtype": policy_map})
+        return config
+
     @classmethod
     def from_config(cls, config):
         # The default `from_config()` for functional models will return a

keras_nlp/src/models/bloom/bloom_backbone_test.py

@@ -39,6 +39,9 @@ def test_backbone_basics(self):
             init_kwargs=self.init_kwargs,
             input_data=self.input_data,
             expected_output_shape=(2, 5, 8),
+            # TODO: Set to `True`. Error msg: Layer LayerNormalization does not
+            # have a `quantized_call()` method implemented.
+            run_quantization_check=False,
         )
 
     @pytest.mark.large

keras_nlp/src/models/gemma/gemma_backbone_test.py

@@ -171,7 +171,6 @@ def test_distribution_with_lora(self):
                 self.assertEqual(tuple(w.value.sharding.spec), (None, None))
 
 
-@pytest.mark.keras_3_only
 class Gemma2BackboneTest(TestCase):
     def setUp(self):
         self.init_kwargs = {

keras_nlp/src/models/opt/opt_backbone_test.py

@@ -40,6 +40,10 @@ def test_backbone_basics(self):
             init_kwargs=self.init_kwargs,
             input_data=self.input_data,
             expected_output_shape=(2, 5, 2),
+            # TODO: Set to `True`. Error msg: Layer 'token_embedding' expected 1
+            # variables, but received 0 variables during loading. Expected:
+            # ['embeddings']
+            run_quantization_check=False,
         )
 
     @pytest.mark.large