keras-nightly 3.12.0.dev2025092403__py3-none-any.whl → 3.14.0.dev2026010104__py3-none-any.whl

keras/src/layers/core/dense.py

@@ -9,12 +9,11 @@ from keras.src import ops
 from keras.src import quantizers
 from keras.src import regularizers
 from keras.src.api_export import keras_export
-from keras.src.dtype_policies.dtype_policy import GPTQDTypePolicy
-from keras.src.dtype_policies.dtype_policy_map import DTypePolicyMap
 from keras.src.layers.input_spec import InputSpec
 from keras.src.layers.layer import Layer
-from keras.src.quantizers.gptq_config import GPTQConfig
+from keras.src.quantizers.quantization_config import QuantizationConfig
 from keras.src.quantizers.quantizers import dequantize_with_sz_map
+from keras.src.saving import serialization_lib
 
 
 @keras_export("keras.layers.Dense")
@@ -26,7 +25,9 @@ class Dense(Layer):
     where `activation` is the element-wise activation function
     passed as the `activation` argument, `kernel` is a weights matrix
     created by the layer, and `bias` is a bias vector created by the layer
-    (only applicable if `use_bias` is `True`).
+    (only applicable if `use_bias` is `True`). When this layer is
+    followed by a `BatchNormalization` layer, it is recommended to set
+    `use_bias=False` as `BatchNormalization` has its own bias term.
 
     Note: If the input to the layer has a rank greater than 2, `Dense`
     computes the dot product between the `inputs` and the `kernel` along the
@@ -93,8 +94,15 @@ class Dense(Layer):
         bias_constraint=None,
         lora_rank=None,
         lora_alpha=None,
+        quantization_config=None,
         **kwargs,
     ):
+        if not isinstance(units, int) or units <= 0:
+            raise ValueError(
+                "Received an invalid value for `units`, expected a positive "
+                f"integer. Received: units={units}"
+            )
+
         super().__init__(activity_regularizer=activity_regularizer, **kwargs)
         self.units = units
         self.activation = activations.get(activation)
@@ -108,13 +116,18 @@ class Dense(Layer):
         self.lora_rank = lora_rank
         self.lora_alpha = lora_alpha if lora_alpha is not None else lora_rank
         self.lora_enabled = False
+        self.quantization_config = quantization_config
         self.input_spec = InputSpec(min_ndim=2)
         self.supports_masking = True
 
     def build(self, input_shape):
         kernel_shape = (input_shape[-1], self.units)
         if self.quantization_mode:
-            self.quantized_build(kernel_shape, mode=self.quantization_mode)
+            self.quantized_build(
+                kernel_shape,
+                mode=self.quantization_mode,
+                config=self.quantization_config,
+            )
         if self.quantization_mode not in ("int8", "int4", "gptq"):
             # If the layer is quantized to int8 or int4, `self._kernel` will be
             # added in `self._int8_build` or `_int4_build`. Therefore, we skip
@@ -143,22 +156,47 @@ class Dense(Layer):
 
     @property
     def kernel(self):
+        from keras.src.quantizers import gptq_core
+
         if not self.built:
             raise AttributeError(
                 "You must build the layer before accessing `kernel`."
             )
-        if (
-            getattr(self, "is_gptq_calibrated", False)
-            and self.quantization_mode == "gptq"
-        ):
-            return self.quantized_kernel
-        kernel = self._kernel
-        if self.quantization_mode == "int4":
-            kernel = quantizers.unpack_int4(kernel, self._orig_input_dim)
+
+        mode = self.quantization_mode
+        is_gptq = mode == "gptq"
+        is_int4 = mode == "int4"
+        calibrated = bool(getattr(self, "is_gptq_calibrated", False))
+        gptq_bits = (
+            gptq_core.get_weight_bits_for_layer(self, None) if is_gptq else None
+        )
+
+        # Decide the source tensor first (packed vs already-quantized vs plain
+        # kernel)
+        if is_gptq and calibrated and gptq_bits != 4:
+            # calibrated GPTQ, not 4-bit, no unpacking needed
+            kernel = self.quantized_kernel
+        else:
+            # Start with the stored kernel
+            kernel = getattr(self, "_kernel", None)
+
+            # Handle int4 unpacking cases in one place
+            if is_int4:
+                kernel = quantizers.unpack_int4(kernel, self._orig_input_dim)
+            elif is_gptq and calibrated and gptq_bits == 4:
+                kernel = quantizers.unpack_int4(
+                    self.quantized_kernel,
+                    orig_len=self.units,
+                    axis=0,
+                    dtype="uint8",
+                )
+
+        # Apply LoRA once at the end.
         if self.lora_enabled:
-            return kernel + (self.lora_alpha / self.lora_rank) * ops.matmul(
+            kernel = kernel + (self.lora_alpha / self.lora_rank) * ops.matmul(
                 self.lora_kernel_a, self.lora_kernel_b
             )
+
         return kernel
 
     def call(self, inputs, training=None):
@@ -236,25 +274,25 @@ class Dense(Layer):
         if not self.built:
             return
         mode = self.quantization_mode
-        if mode not in self.quantization_variable_spec:
+        if mode not in self.variable_serialization_spec:
             raise self._quantization_mode_error(mode)
 
         # Kernel plus optional merged LoRA-aware scale (returns (kernel, None)
         # for None/gptq)
         kernel_value, merged_kernel_scale = self._get_kernel_with_merged_lora()
-
-        # Save the variables using the name as the key.
-        if mode != "gptq":
-            store["kernel"] = kernel_value
-        if self.bias is not None:
-            store["bias"] = self.bias
-        for name in self.quantization_variable_spec[mode]:
-            if name == "kernel_scale" and mode in ("int4", "int8"):
+        idx = 0
+        for name in self.variable_serialization_spec[mode]:
+            if name == "kernel":
+                store[str(idx)] = kernel_value
+            elif name == "bias" and self.bias is None:
+                continue
+            elif name == "kernel_scale" and mode in ("int4", "int8"):
                 # For int4/int8, the merged LoRA scale (if any) comes from
                 # `_get_kernel_with_merged_lora()`
-                store[name] = merged_kernel_scale
+                store[str(idx)] = merged_kernel_scale
             else:
-                store[name] = getattr(self, name)
+                store[str(idx)] = getattr(self, name)
+            idx += 1
 
     def load_own_variables(self, store):
         if not self.lora_enabled:
@@ -263,39 +301,21 @@ class Dense(Layer):
         if not self.built:
             return
         mode = self.quantization_mode
-        if mode not in self.quantization_variable_spec:
+        if mode not in self.variable_serialization_spec:
             raise self._quantization_mode_error(mode)
 
-        # Determine whether to use the legacy loading method.
-        if "0" in store:
-            return self._legacy_load_own_variables(store)
+        # A saved GPTQ quantized model will always be calibrated.
+        self.is_gptq_calibrated = mode == "gptq"
 
-        # Load the variables using the name as the key.
-        if mode != "gptq":
-            self._kernel.assign(store["kernel"])
-        if self.bias is not None:
-            self.bias.assign(store["bias"])
-        for name in self.quantization_variable_spec[mode]:
-            getattr(self, name).assign(store[name])
-        if self.lora_enabled:
-            self.lora_kernel_a.assign(ops.zeros(self.lora_kernel_a.shape))
-            self.lora_kernel_b.assign(ops.zeros(self.lora_kernel_b.shape))
-
-    def _legacy_load_own_variables(self, store):
-        # The keys of the `store` will be saved as determined because the
-        # default ordering will change after quantization
-        mode = self.quantization_mode
-        targets = []
-        if mode != "gptq":
-            targets.append(self._kernel)
-        if self.bias is not None:
-            targets.append(self.bias)
-        targets.extend(
-            getattr(self, name)
-            for name in self.quantization_variable_spec[mode]
-        )
-        for i, variable in enumerate(targets):
-            variable.assign(store[str(i)])
+        idx = 0
+        for name in self.variable_serialization_spec[mode]:
+            if name == "kernel":
+                self._kernel.assign(store[str(idx)])
+            elif name == "bias" and self.bias is None:
+                continue
+            else:
+                getattr(self, name).assign(store[str(idx)])
+            idx += 1
         if self.lora_enabled:
             self.lora_kernel_a.assign(ops.zeros(self.lora_kernel_a.shape))
             self.lora_kernel_b.assign(ops.zeros(self.lora_kernel_b.shape))
@@ -316,59 +336,51 @@ class Dense(Layer):
             "bias_regularizer": regularizers.serialize(self.bias_regularizer),
             "kernel_constraint": constraints.serialize(self.kernel_constraint),
             "bias_constraint": constraints.serialize(self.bias_constraint),
+            "quantization_config": serialization_lib.serialize_keras_object(
+                self.quantization_config
+            ),
         }
         if self.lora_rank:
             config["lora_rank"] = self.lora_rank
             config["lora_alpha"] = self.lora_alpha
         return {**base_config, **config}
 
-    def _check_load_own_variables(self, store):
-        all_vars = self._trainable_variables + self._non_trainable_variables
-        if len(store.keys()) != len(all_vars):
-            if len(all_vars) == 0 and not self.built:
-                raise ValueError(
-                    f"Layer '{self.name}' was never built "
-                    "and thus it doesn't have any variables. "
-                    f"However the weights file lists {len(store.keys())} "
-                    "variables for this layer.\n"
-                    "In most cases, this error indicates that either:\n\n"
-                    "1. The layer is owned by a parent layer that "
-                    "implements a `build()` method, but calling the "
-                    "parent's `build()` method did NOT create the state of "
-                    f"the child layer '{self.name}'. A `build()` method "
-                    "must create ALL state for the layer, including "
-                    "the state of any children layers.\n\n"
-                    "2. You need to implement "
-                    "the `def build_from_config(self, config)` method "
-                    f"on layer '{self.name}', to specify how to rebuild "
-                    "it during loading. "
-                    "In this case, you might also want to implement the "
-                    "method that generates the build config at saving time, "
-                    "`def get_build_config(self)`. "
-                    "The method `build_from_config()` is meant "
-                    "to create the state "
-                    "of the layer (i.e. its variables) upon deserialization.",
-                )
-            raise ValueError(
-                f"Layer '{self.name}' expected {len(all_vars)} variables, "
-                "but received "
-                f"{len(store.keys())} variables during loading. "
-                f"Expected: {[v.name for v in all_vars]}"
+    @classmethod
+    def from_config(cls, config):
+        config = config.copy()
+        config["quantization_config"] = (
+            serialization_lib.deserialize_keras_object(
+                config.get("quantization_config", None)
             )
+        )
+        return super().from_config(config)
 
     @property
-    def quantization_variable_spec(self):
-        """Returns a dict mapping quantization modes to variable names.
+    def variable_serialization_spec(self):
+        """Returns a dict mapping quantization modes to variable names in order.
 
         This spec is used by `save_own_variables` and `load_own_variables` to
-        determine which variables should be saved/loaded for each quantization
-        mode.
+        determine the correct ordering of variables during serialization for
+        each quantization mode. `None` means no quantization.
         """
         return {
-            None: [],
-            "int8": ["kernel_scale"],
-            "int4": ["kernel_scale"],
+            None: [
+                "kernel",
+                "bias",
+            ],
+            "int8": [
+                "kernel",
+                "bias",
+                "kernel_scale",
+            ],
+            "int4": [
+                "kernel",
+                "bias",
+                "kernel_scale",
+            ],
             "float8": [
+                "kernel",
+                "bias",
                 "inputs_scale",
                 "inputs_amax_history",
                 "kernel_scale",
@@ -377,6 +389,7 @@ class Dense(Layer):
                 "outputs_grad_amax_history",
             ],
             "gptq": [
+                "bias",
                 "quantized_kernel",
                 "kernel_scale",
                 "kernel_zero",
@@ -386,9 +399,9 @@ class Dense(Layer):
 
     def quantized_build(self, kernel_shape, mode, config=None):
         if mode == "int8":
-            self._int8_build(kernel_shape)
+            self._int8_build(kernel_shape, config)
         elif mode == "int4":
-            self._int4_build(kernel_shape)
+            self._int4_build(kernel_shape, config)
         elif mode == "float8":
             self._float8_build()
         elif mode == "gptq":
@@ -397,8 +410,13 @@ class Dense(Layer):
             raise self._quantization_mode_error(mode)
         self._is_quantized = True
 
-    def _int8_build(self, kernel_shape):
-        self.inputs_quantizer = quantizers.AbsMaxQuantizer(axis=-1)
+    def _int8_build(self, kernel_shape, config=None):
+        self.inputs_quantizer = (
+            QuantizationConfig.activation_quantizer_or_default(
+                config, quantizers.AbsMaxQuantizer()
+            )
+        )
+
         self._kernel = self.add_weight(
             name="kernel",
             shape=kernel_shape,
@@ -414,23 +432,33 @@ class Dense(Layer):
         )
 
     def _gptq_build(self, kernel_shape, config):
+        from keras.src.quantizers import gptq_core
+
         # Ensures the forward pass uses the original high-precision kernel
         # until calibration has been performed.
         self.is_gptq_calibrated = False
         self.kernel_shape = kernel_shape
+
+        weight_bits = gptq_core.get_weight_bits_for_layer(self, config)
+        # For 4-bit weights, we pack two values per byte.
+        units = (
+            (kernel_shape[1] + 1) // 2 if weight_bits == 4 else kernel_shape[1]
+        )
+
         self.quantized_kernel = self.add_weight(
             name="kernel",
-            shape=(kernel_shape[1], kernel_shape[0]),
+            shape=(units, kernel_shape[0]),
             initializer="zeros",
             dtype="uint8",
             trainable=False,
         )
 
-        group_size = self._get_gptq_group_size(config)
-        if group_size == -1:
-            n_groups = 1
-        else:
-            n_groups = math.ceil(self.kernel_shape[0] / group_size)
+        group_size = gptq_core.get_group_size_for_layer(self, config)
+        n_groups = (
+            1
+            if group_size == -1
+            else math.ceil(self.kernel_shape[0] / group_size)
+        )
         self.kernel_scale = self.add_weight(
             name="kernel_scale",
             shape=(self.units, n_groups),
@@ -453,18 +481,31 @@ class Dense(Layer):
         )
 
     def _gptq_call(self, inputs, training=False):
+        from keras.src.quantizers import gptq_core
+
         if not self.is_gptq_calibrated:
             W = self._kernel
         else:
+            should_unpack = (
+                gptq_core.get_weight_bits_for_layer(self, config=None) == 4
+            )
             W = (
-                ops.transpose(
-                    dequantize_with_sz_map(
-                        self.quantized_kernel,
-                        self.kernel_scale,
-                        self.kernel_zero,
-                        self.g_idx,
-                    )
-                ),
+                quantizers.unpack_int4(
+                    self.quantized_kernel,
+                    orig_len=self.units,
+                    axis=0,
+                    dtype="uint8",
+                )
+                if should_unpack
+                else self.quantized_kernel
+            )
+            W = ops.transpose(
+                dequantize_with_sz_map(
+                    W,
+                    self.kernel_scale,
+                    self.kernel_zero,
+                    self.g_idx,
+                )
             )
 
         y = ops.matmul(inputs, W)
@@ -474,7 +515,7 @@ class Dense(Layer):
             y = self.activation(y)
         return y
 
-    def _int4_build(self, kernel_shape):
+    def _int4_build(self, kernel_shape, config=None):
         """Build variables for int4 quantization.
 
         `kernel_shape` is the *original* float32 kernel shape
@@ -483,8 +524,10 @@ class Dense(Layer):
         int8 byte.
         """
         # Per-channel int8 quantizer for the last axis (features).
-        self.inputs_quantizer = quantizers.AbsMaxQuantizer(
-            axis=-1,
+        self.inputs_quantizer = (
+            QuantizationConfig.activation_quantizer_or_default(
+                config, quantizers.AbsMaxQuantizer()
+            )
         )
         input_dim, output_dim = kernel_shape
         packed_rows = (input_dim + 1) // 2  # ceil for odd dims
@@ -573,11 +616,15 @@ class Dense(Layer):
                 inputs_grad = ops.matmul(upstream, ops.transpose(float_kernel))
                 return (inputs_grad, None, None)
 
-            inputs, inputs_scale = self.inputs_quantizer(inputs)
+            output_scale = kernel_scale
+            if self.inputs_quantizer:
+                inputs, inputs_scale = self.inputs_quantizer(inputs, axis=-1)
+                output_scale = ops.multiply(output_scale, inputs_scale)
+
             x = ops.matmul(inputs, kernel)
             # De-scale outputs
             x = ops.cast(x, self.compute_dtype)
-            x = ops.divide(x, ops.multiply(inputs_scale, kernel_scale))
+            x = ops.divide(x, output_scale)
             return x, grad_fn
 
         x = matmul_with_inputs_gradient(
@@ -624,10 +671,15 @@ class Dense(Layer):
                 inputs_grad = ops.matmul(upstream, ops.transpose(float_kernel))
                 return (inputs_grad, None, None)
 
-            inputs, inputs_scale = self.inputs_quantizer(inputs)
+            output_scale = kernel_scale
+
+            if self.inputs_quantizer:
+                inputs, inputs_scale = self.inputs_quantizer(inputs, axis=-1)
+                output_scale = ops.multiply(output_scale, inputs_scale)
+
             x = ops.matmul(inputs, unpacked_kernel)
             x = ops.cast(x, self.compute_dtype)
-            x = ops.divide(x, ops.multiply(inputs_scale, kernel_scale))
+            x = ops.divide(x, output_scale)
             return x, grad_fn
 
         x = matmul_with_inputs_gradient(
@@ -739,30 +791,37 @@ class Dense(Layer):
             x = self.activation(x)
         return x
 
-    def quantize(self, mode, type_check=True, config=None):
+    def quantize(self, mode=None, type_check=True, config=None):
         # Prevent quantization of the subclasses
         if type_check and (type(self) is not Dense):
             raise self._not_implemented_error(self.quantize)
 
+        self.quantization_config = config
+
         kernel_shape = self._kernel.shape
         if mode == "int8":
-            kernel_value, kernel_scale = quantizers.abs_max_quantize(
-                self._kernel, axis=0, to_numpy=True
+            weight_quantizer = QuantizationConfig.weight_quantizer_or_default(
+                self.quantization_config, quantizers.AbsMaxQuantizer(axis=0)
+            )
+            kernel_value, kernel_scale = weight_quantizer(
+                self._kernel, to_numpy=True
             )
             kernel_scale = ops.squeeze(kernel_scale, axis=0)
             del self._kernel
             # Build variables for int8 mode
-            self.quantized_build(kernel_shape, mode)
+            self.quantized_build(kernel_shape, mode, self.quantization_config)
             self._kernel.assign(kernel_value)
             self.kernel_scale.assign(kernel_scale)
         elif mode == "int4":
             # 1. Quantize to int4 values (still int8 dtype, range [-8,7])
-            kernel_value_int4, kernel_scale = quantizers.abs_max_quantize(
-                self._kernel,
-                axis=0,
-                value_range=(-8, 7),
-                dtype="int8",
-                to_numpy=True,
+            weight_quantizer = QuantizationConfig.weight_quantizer_or_default(
+                self.quantization_config,
+                quantizers.AbsMaxQuantizer(
+                    axis=0, value_range=(-8, 7), output_dtype="int8"
+                ),
+            )
+            kernel_value_int4, kernel_scale = weight_quantizer(
+                self._kernel, to_numpy=True
             )
             kernel_scale = ops.squeeze(kernel_scale, axis=0)
             # 2. Pack two int4 values into a single int8 byte.
@@ -770,12 +829,12 @@ class Dense(Layer):
             del self._kernel
             # Build variables using the original kernel shape; _int4_build will
             # compute the packed shape internally.
-            self.quantized_build(kernel_shape, mode)
+            self.quantized_build(kernel_shape, mode, self.quantization_config)
             # Assign packed values.
             self._kernel.assign(packed_kernel_value)
             self.kernel_scale.assign(kernel_scale)
         elif mode == "gptq":
-            self.quantized_build(kernel_shape, mode, config)
+            self.quantized_build(kernel_shape, mode, self.quantization_config)
         elif mode == "float8":
             self.quantized_build(kernel_shape, mode)
         else:
@@ -787,7 +846,7 @@ class Dense(Layer):
 
             policy_name = mode
             if mode == "gptq":
-                policy_name = config.dtype_policy_string()
+                policy_name = self.quantization_config.dtype_policy_string()
             policy = dtype_policies.get(
                 f"{policy_name}_from_{self.dtype_policy.name}"
             )
@@ -875,43 +934,3 @@ class Dense(Layer):
         else:
             kernel_value = requantized_kernel
         return kernel_value, kernel_scale
-
-    def _get_gptq_group_size(self, config):
-        """Determine the group size for GPTQ quantization.
-
-        The group size can be specified either through the `config` argument
-        or through the `dtype_policy` if it is of type `GPTQDTypePolicy`.
-
-        The config argument is usually available when quantizing the layer
-        via the `quantize` method. If the layer was deserialized from a
-        saved model, the group size should be specified in the `dtype_policy`.
-
-        Args:
-            config: An optional configuration object that may contain the
-                `group_size` attribute.
-        Returns:
-            int. The determined group size for GPTQ quantization.
-        Raises:
-            ValueError: If the group size is not specified in either the
-                `config` or the `dtype_policy`.
-        """
-        if config and isinstance(config, GPTQConfig):
-            return config.group_size
-        elif isinstance(self.dtype_policy, GPTQDTypePolicy):
-            return self.dtype_policy.group_size
-        elif isinstance(self.dtype_policy, DTypePolicyMap):
-            policy = self.dtype_policy[self.path]
-            if not isinstance(policy, GPTQDTypePolicy):
-                # This should never happen based on how we set the
-                # quantization mode, but we check just in case.
-                raise ValueError(
-                    "Expected a `dtype_policy` of type `GPTQDTypePolicy`."
-                    f"Got: {type(policy)}"
-                )
-            return policy.group_size
-        else:
-            raise ValueError(
-                "For GPTQ quantization, the group_size must be specified"
-                "either through a `dtype_policy` of type "
-                "`GPTQDTypePolicy` or the `config` argument."
-            )