keras-nightly 3.12.0.dev2025083103__py3-none-any.whl → 3.14.0.dev2026011604__py3-none-any.whl

keras/src/layers/core/einsum_dense.py

@@ -1,3 +1,4 @@
+import math
 import re
 import string
 
@@ -5,6 +6,7 @@ import ml_dtypes
 import numpy as np
 
 from keras.src import activations
+from keras.src import backend
 from keras.src import constraints
 from keras.src import dtype_policies
 from keras.src import initializers
@@ -14,6 +16,9 @@ from keras.src import regularizers
 from keras.src.api_export import keras_export
 from keras.src.layers.input_spec import InputSpec
 from keras.src.layers.layer import Layer
+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.EinsumDense")
@@ -131,6 +136,8 @@ class EinsumDense(Layer):
         bias_constraint=None,
         lora_rank=None,
         lora_alpha=None,
+        gptq_unpacked_column_size=None,
+        quantization_config=None,
         **kwargs,
     ):
         super().__init__(**kwargs)
@@ -150,6 +157,8 @@ class EinsumDense(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.gptq_unpacked_column_size = gptq_unpacked_column_size
+        self.quantization_config = quantization_config
 
     def build(self, input_shape):
         shape_data = _analyze_einsum_string(
@@ -162,12 +171,16 @@ class EinsumDense(Layer):
         self.full_output_shape = tuple(full_output_shape)
         self.input_spec = InputSpec(ndim=len(input_shape))
         if self.quantization_mode is not None:
-            self.quantized_build(kernel_shape, mode=self.quantization_mode)
+            self.quantized_build(
+                kernel_shape,
+                mode=self.quantization_mode,
+                config=self.quantization_config,
+            )
         # Skip creating a duplicate kernel variable when the layer is already
         # quantized to int8 or int4, because `quantized_build` has created the
         # appropriate kernel variable. For other modes (e.g., float8 or no
         # quantization), we still need the floating-point kernel.
-        if self.quantization_mode not in ("int8", "int4"):
+        if self.quantization_mode not in ("int8", "int4", "gptq", "awq"):
             # If the layer is quantized to int8, `self._kernel` will be added
             # in `self._int8_build`. Therefore, we skip it here.
             self._kernel = self.add_weight(
@@ -197,15 +210,62 @@ class EinsumDense(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`."
             )
+
+        mode = self.quantization_mode
+        is_gptq = mode == "gptq"
+        is_awq = mode == "awq"
+        is_int4 = mode == "int4"
+        gptq_calibrated = bool(getattr(self, "is_gptq_calibrated", False))
+        awq_calibrated = bool(getattr(self, "is_awq_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 gptq_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_length_along_pack_axis,
+                    self._int4_pack_axis,
+                )
+            elif is_gptq and gptq_calibrated and gptq_bits == 4:
+                kernel = quantizers.unpack_int4(
+                    self.quantized_kernel,
+                    orig_len=self.gptq_unpacked_column_size,
+                    axis=0,
+                    dtype="uint8",
+                )
+            elif is_awq and awq_calibrated:
+                # AWQ always uses 4-bit quantization
+                kernel = quantizers.unpack_int4(
+                    self.quantized_kernel,
+                    orig_len=self.awq_unpacked_column_size,
+                    axis=0,
+                    dtype="uint8",
+                )
+
+        # Apply LoRA if enabled
         if self.lora_enabled:
-            return self._kernel + (
-                self.lora_alpha / self.lora_rank
-            ) * ops.matmul(self.lora_kernel_a, self.lora_kernel_b)
-        return self._kernel
+            kernel = kernel + (self.lora_alpha / self.lora_rank) * ops.matmul(
+                self.lora_kernel_a, self.lora_kernel_b
+            )
+
+        return kernel
 
     def compute_output_shape(self, _):
         return self.full_output_shape
@@ -239,6 +299,10 @@ class EinsumDense(Layer):
             raise ValueError(
                 "lora is already enabled. This can only be done once per layer."
             )
+        if self.quantization_mode == "gptq":
+            raise NotImplementedError(
+                "lora is not currently supported with GPTQ quantization."
+            )
         self._tracker.unlock()
         # Determine the appropriate (unpacked) kernel shape for LoRA.
         if self.quantization_mode == "int4":
@@ -277,26 +341,26 @@ class EinsumDense(Layer):
         # Do nothing if the layer isn't yet built
         if not self.built:
             return
-        # The keys of the `store` will be saved as determined because the
-        # default ordering will change after quantization
-        kernel_value, kernel_scale = self._get_kernel_with_merged_lora()
-        target_variables = [kernel_value]
-        if self.bias is not None:
-            target_variables.append(self.bias)
-        if self.quantization_mode is not None:
-            if self.quantization_mode in ("int8", "int4"):
-                target_variables.append(kernel_scale)
-            elif self.quantization_mode == "float8":
-                target_variables.append(self.inputs_scale)
-                target_variables.append(self.inputs_amax_history)
-                target_variables.append(self.kernel_scale)
-                target_variables.append(self.kernel_amax_history)
-                target_variables.append(self.outputs_grad_scale)
-                target_variables.append(self.outputs_grad_amax_history)
+        mode = self.quantization_mode
+        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()
+        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[str(idx)] = merged_kernel_scale
             else:
-                raise self._quantization_mode_error(self.quantization_mode)
-        for i, variable in enumerate(target_variables):
-            store[str(i)] = variable
+                store[str(idx)] = getattr(self, name)
+            idx += 1
 
     def load_own_variables(self, store):
         if not self.lora_enabled:
@@ -304,25 +368,23 @@ class EinsumDense(Layer):
         # Do nothing if the layer isn't yet built
         if not self.built:
             return
-        # The keys of the `store` will be saved as determined because the
-        # default ordering will change after quantization
-        target_variables = [self._kernel]
-        if self.bias is not None:
-            target_variables.append(self.bias)
-        if self.quantization_mode is not None:
-            if self.quantization_mode in ("int8", "int4"):
-                target_variables.append(self.kernel_scale)
-            elif self.quantization_mode == "float8":
-                target_variables.append(self.inputs_scale)
-                target_variables.append(self.inputs_amax_history)
-                target_variables.append(self.kernel_scale)
-                target_variables.append(self.kernel_amax_history)
-                target_variables.append(self.outputs_grad_scale)
-                target_variables.append(self.outputs_grad_amax_history)
+        mode = self.quantization_mode
+        if mode not in self.variable_serialization_spec:
+            raise self._quantization_mode_error(mode)
+
+        # A saved GPTQ/AWQ quantized model will always be calibrated.
+        self.is_gptq_calibrated = mode == "gptq"
+        self.is_awq_calibrated = mode == "awq"
+
+        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:
-                raise self._quantization_mode_error(self.quantization_mode)
-        for i, variable in enumerate(target_variables):
-            variable.assign(store[str(i)])
+                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))
@@ -347,64 +409,103 @@ class EinsumDense(Layer):
             ),
             "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
+        if self.gptq_unpacked_column_size:
+            config["gptq_unpacked_column_size"] = self.gptq_unpacked_column_size
         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 variable_serialization_spec(self):
+        """Returns a dict mapping quantization modes to variable names in order.
 
-    # Quantization-related (int8 and float8) methods
+        This spec is used by `save_own_variables` and `load_own_variables` to
+        determine the correct ordering of variables during serialization for
+        each quantization mode. `None` means no quantization.
+        """
+        return {
+            None: [
+                "kernel",
+                "bias",
+            ],
+            "int8": [
+                "kernel",
+                "bias",
+                "kernel_scale",
+            ],
+            "int4": [
+                "kernel",
+                "bias",
+                "kernel_scale",
+            ],
+            "float8": [
+                "kernel",
+                "bias",
+                "inputs_scale",
+                "inputs_amax_history",
+                "kernel_scale",
+                "kernel_amax_history",
+                "outputs_grad_scale",
+                "outputs_grad_amax_history",
+            ],
+            "gptq": [
+                "bias",
+                "quantized_kernel",
+                "kernel_scale",
+                "kernel_zero",
+                "g_idx",
+            ],
+            "awq": [
+                "bias",
+                "quantized_kernel",
+                "kernel_scale",
+                "kernel_zero",
+                "awq_scales",
+                "g_idx",
+            ],
+        }
 
-    def quantized_build(self, kernel_shape, mode):
+    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":
+            self._gptq_build(kernel_shape, config)
+        elif mode == "awq":
+            self._awq_build(kernel_shape, config)
         else:
             raise self._quantization_mode_error(mode)
         self._is_quantized = True
 
-    def _int8_build(self, kernel_shape):
+    def _int8_build(self, kernel_shape, config=None):
         self._set_quantization_info()
-        self.inputs_quantizer = quantizers.AbsMaxQuantizer(
-            axis=self._input_reduced_axes
+        self.inputs_quantizer = (
+            QuantizationConfig.activation_quantizer_or_default(
+                config,
+                quantizers.AbsMaxQuantizer(),
+            )
         )
+        # If the config provided a default AbsMaxQuantizer, we need to
+        # override the axis to match the equation's reduction axes.
+        self.quantization_axis = tuple(self._input_reduced_axes)
         self._kernel = self.add_weight(
             name="kernel",
             shape=kernel_shape,
@@ -420,7 +521,244 @@ class EinsumDense(Layer):
             trainable=False,
         )
 
-    def _int4_build(self, kernel_shape):
+    def _gptq_build(self, kernel_shape, config):
+        """
+        Allocate quantized kernel & params for EinsumDense.
+
+        Args:
+            kernel_shape: tuple/list; the layer's original kernel shape, e.g.
+                [in_features, out_features] or [in_features, heads, head_dim].
+            group_size: int; contiguous input-group size for quantization
+                (=-1 means per-output-channel with no grouping).
+        """
+        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.original_kernel_shape = kernel_shape
+        if len(kernel_shape) == 2:
+            rows = kernel_shape[0]
+            columns = kernel_shape[1]
+        elif len(kernel_shape) == 3:
+            shape = list(self.original_kernel_shape)
+            d_model_dim_index = shape.index(max(shape))
+
+            if d_model_dim_index == 0:  # QKV projection case
+                in_features, heads, head_dim = shape
+                rows, columns = (
+                    in_features,
+                    heads * head_dim,
+                )
+            elif d_model_dim_index in [1, 2]:  # Attention Output case
+                heads, head_dim, out_features = shape
+                rows, columns = (
+                    heads * head_dim,
+                    out_features,
+                )
+            else:
+                raise ValueError("Could not determine row/column split.")
+
+        group_size = gptq_core.get_group_size_for_layer(self, config)
+        n_groups = 1 if group_size == -1 else math.ceil(rows / group_size)
+
+        self.gptq_unpacked_column_size = columns
+
+        weight_bits = gptq_core.get_weight_bits_for_layer(self, config)
+        # For 4-bit weights, we pack two values per byte.
+        kernel_columns = (columns + 1) // 2 if weight_bits == 4 else columns
+
+        self._set_quantization_info()
+
+        self.quantized_kernel = self.add_weight(
+            name="kernel",
+            shape=(kernel_columns, rows),
+            initializer="zeros",
+            dtype="uint8",
+            trainable=False,
+        )
+
+        self.kernel_scale = self.add_weight(
+            name="kernel_scale",
+            shape=(columns, n_groups),
+            initializer="ones",
+            trainable=False,
+        )
+        self.kernel_zero = self.add_weight(
+            name="zero_point",
+            shape=(columns, n_groups),
+            initializer="zeros",
+            dtype="uint8",
+            trainable=False,
+        )
+
+        self.g_idx = self.add_weight(
+            name="g_idx",
+            shape=(rows,),
+            initializer="zeros",
+            dtype="float32",
+            trainable=False,
+        )
+
+    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 = (
+                quantizers.unpack_int4(
+                    self.quantized_kernel,
+                    orig_len=self.gptq_unpacked_column_size,
+                    axis=0,
+                    dtype="uint8",
+                )
+                if should_unpack
+                else self.quantized_kernel
+            )
+            W = dequantize_with_sz_map(
+                W,
+                self.kernel_scale,
+                self.kernel_zero,
+                self.g_idx,
+            )
+            W = ops.transpose(W)
+
+            W = ops.reshape(W, self.original_kernel_shape)
+
+        y = ops.einsum(self.equation, inputs, W)
+        if self.bias is not None:
+            y = ops.add(y, self.bias)
+        if self.activation is not None:
+            y = self.activation(y)
+        return y
+
+    def _awq_build(self, kernel_shape, config):
+        """Build variables for AWQ quantization.
+
+        AWQ uses 4-bit quantization with per-channel AWQ scales that protect
+        salient weights based on activation magnitudes.
+        """
+        from keras.src.quantizers import awq_core
+
+        # Ensures the forward pass uses the original high-precision kernel
+        # until calibration has been performed.
+        self.is_awq_calibrated = False
+
+        self.original_kernel_shape = kernel_shape
+        if len(kernel_shape) == 2:
+            rows = kernel_shape[0]
+            columns = kernel_shape[1]
+        elif len(kernel_shape) == 3:
+            shape = list(self.original_kernel_shape)
+            d_model_dim_index = shape.index(max(shape))
+
+            if d_model_dim_index == 0:  # QKV projection case
+                in_features, heads, head_dim = shape
+                rows, columns = (
+                    in_features,
+                    heads * head_dim,
+                )
+            elif d_model_dim_index in [1, 2]:  # Attention Output case
+                heads, head_dim, out_features = shape
+                rows, columns = (
+                    heads * head_dim,
+                    out_features,
+                )
+            else:
+                raise ValueError("Could not determine row/column split.")
+        else:
+            raise ValueError("AWQ quantization only supports 2D or 3D kernels.")
+
+        group_size = awq_core.get_group_size_for_layer(self, config)
+        num_groups = 1 if group_size == -1 else math.ceil(rows / group_size)
+
+        self.awq_unpacked_column_size = columns
+
+        # For 4-bit weights, we pack two values per byte.
+        kernel_columns = (columns + 1) // 2
+
+        self._set_quantization_info()
+
+        self.quantized_kernel = self.add_weight(
+            name="kernel",
+            shape=(kernel_columns, rows),
+            initializer="zeros",
+            dtype="uint8",
+            trainable=False,
+        )
+
+        self.kernel_scale = self.add_weight(
+            name="kernel_scale",
+            shape=(columns, num_groups),
+            initializer="ones",
+            trainable=False,
+        )
+        self.kernel_zero = self.add_weight(
+            name="zero_point",
+            shape=(columns, num_groups),
+            initializer="zeros",
+            dtype="uint8",
+            trainable=False,
+        )
+
+        # Per-channel AWQ scales from activation magnitudes
+        self.awq_scales = self.add_weight(
+            name="awq_scales",
+            shape=(rows,),
+            initializer="ones",
+            trainable=False,
+        )
+
+        self.g_idx = self.add_weight(
+            name="g_idx",
+            shape=(rows,),
+            initializer="zeros",
+            dtype="float32",
+            trainable=False,
+        )
+
+    def _awq_call(self, inputs, training=False):
+        """Forward pass for AWQ quantized layer."""
+        if not self.is_awq_calibrated:
+            W = self._kernel
+        else:
+            # Unpack 4-bit weights
+            W = quantizers.unpack_int4(
+                self.quantized_kernel,
+                orig_len=self.awq_unpacked_column_size,
+                axis=0,
+                dtype="uint8",
+            )
+            # Dequantize using scale/zero maps
+            W = dequantize_with_sz_map(
+                W,
+                self.kernel_scale,
+                self.kernel_zero,
+                self.g_idx,
+            )
+            W = ops.transpose(W)
+
+            # Apply AWQ scales by dividing to restore original magnitude
+            # (We multiplied by scales before quantization, so divide to undo)
+            # awq_scales has shape [input_dim], W has shape [input_dim, out_dim]
+            # Expand dims for proper broadcasting.
+            W = ops.divide(W, ops.expand_dims(self.awq_scales, -1))
+
+            W = ops.reshape(W, self.original_kernel_shape)
+
+        y = ops.einsum(self.equation, inputs, W)
+        if self.bias is not None:
+            y = ops.add(y, self.bias)
+        if self.activation is not None:
+            y = self.activation(y)
+        return y
+
+    def _int4_build(self, kernel_shape, config=None):
         """Build variables for int4 quantization.
 
         The packed int4 kernel stores two int4 values within a single int8
@@ -432,9 +770,15 @@ class EinsumDense(Layer):
         self._set_quantization_info()
 
         # Quantizer for the inputs (per the reduced axes)
-        self.inputs_quantizer = quantizers.AbsMaxQuantizer(
-            axis=self._input_reduced_axes
+        self.inputs_quantizer = (
+            QuantizationConfig.activation_quantizer_or_default(
+                config,
+                quantizers.AbsMaxQuantizer(),
+            )
         )
+        # If the config provided a default AbsMaxQuantizer, we need to
+        # override the axis to match the equation's reduction axes.
+        self.quantization_axis = tuple(self._input_reduced_axes)
 
         # Choose the axis to perform int4 packing - use the first reduced axis
         # for the kernel (analogous to the input dimension of a Dense layer).
@@ -556,13 +900,36 @@ class EinsumDense(Layer):
                 )
                 return (inputs_grad, None, None)
 
-            inputs, inputs_scale = self.inputs_quantizer(inputs)
-            x = ops.einsum(self.equation, inputs, kernel)
-            # Deal with `inputs_scale`
-            inputs_scale = self._adjust_scale_for_quant(inputs_scale, "input")
-            # De-scale outputs
-            x = ops.cast(x, self.compute_dtype)
-            x = ops.divide(x, ops.multiply(inputs_scale, kernel_scale))
+            if self.inputs_quantizer:
+                inputs, inputs_scale = self.inputs_quantizer(
+                    inputs, axis=self.quantization_axis
+                )
+                # Align `inputs_scale` axes with the output
+                # for correct broadcasting
+                inputs_scale = self._adjust_scale_for_quant(
+                    inputs_scale, "input"
+                )
+                x = ops.einsum(self.equation, inputs, kernel)
+                # De-scale outputs
+                x = ops.cast(x, self.compute_dtype)
+                x = ops.divide(x, ops.multiply(inputs_scale, kernel_scale))
+            else:
+                # Weight-only quantization: dequantize kernel and use float
+                # einsum. This is a workaround for PyTorch's einsum which
+                # doesn't support mixed-precision inputs (float input,
+                # int8 kernel).
+                if backend.backend() == "torch":
+                    kernel_scale = self._adjust_scale_for_dequant(kernel_scale)
+                    float_kernel = ops.divide(
+                        ops.cast(kernel, dtype=self.compute_dtype),
+                        kernel_scale,
+                    )
+                    x = ops.einsum(self.equation, inputs, float_kernel)
+                else:
+                    x = ops.einsum(self.equation, inputs, kernel)
+                    # De-scale outputs
+                    x = ops.cast(x, self.compute_dtype)
+                    x = ops.divide(x, kernel_scale)
             return x, grad_fn
 
         x = einsum_with_inputs_gradient(
@@ -632,17 +999,38 @@ class EinsumDense(Layer):
                 return (inputs_grad, None, None)
 
             # Quantize inputs per `self.inputs_quantizer`.
-            inputs_q, inputs_scale = self.inputs_quantizer(inputs)
-
-            # Compute einsum on quantized inputs and unpacked int4 kernel.
-            x = ops.einsum(self.equation, inputs_q, unpacked_kernel)
-
-            # Align `inputs_scale` axes with the output for correct broadcasting
-            inputs_scale = self._adjust_scale_for_quant(inputs_scale, "input")
-
-            # De-scale outputs.
-            x = ops.cast(x, self.compute_dtype)
-            x = ops.divide(x, ops.multiply(inputs_scale, kernel_scale))
+            if self.inputs_quantizer:
+                inputs_q, inputs_scale = self.inputs_quantizer(
+                    inputs, axis=self.quantization_axis
+                )
+                # Align `inputs_scale` axes with the output
+                # for correct broadcasting
+                inputs_scale = self._adjust_scale_for_quant(
+                    inputs_scale, "input"
+                )
+                x = ops.einsum(self.equation, inputs_q, unpacked_kernel)
+                # De-scale outputs
+                x = ops.cast(x, self.compute_dtype)
+                x = ops.divide(x, ops.multiply(inputs_scale, kernel_scale))
+            else:
+                # Weight-only quantization: dequantize kernel and use float
+                # einsum. This is a workaround for PyTorch's einsum which
+                # doesn't support mixed-precision inputs (float input,
+                # int4 kernel).
+                if backend.backend() == "torch":
+                    # Align `kernel_scale` to the same layout as
+                    # `unpacked_kernel`.
+                    kernel_scale = self._adjust_scale_for_dequant(kernel_scale)
+                    float_kernel = ops.divide(
+                        ops.cast(unpacked_kernel, dtype=self.compute_dtype),
+                        kernel_scale,
+                    )
+                    x = ops.einsum(self.equation, inputs, float_kernel)
+                else:
+                    x = ops.einsum(self.equation, inputs, unpacked_kernel)
+                    # De-scale outputs
+                    x = ops.cast(x, self.compute_dtype)
+                    x = ops.divide(x, kernel_scale)
             return x, grad_fn
 
         x = einsum_with_inputs_gradient(
@@ -756,30 +1144,40 @@ class EinsumDense(Layer):
             x = self.activation(x)
         return x
 
-    def quantize(self, mode, type_check=True):
+    def quantize(self, mode=None, type_check=True, config=None):
         # Prevent quantization of the subclasses
         if type_check and (type(self) is not EinsumDense):
             raise self._not_implemented_error(self.quantize)
 
+        self.quantization_config = config
+
         kernel_shape = self._kernel.shape
-        if mode in ("int8", "int4"):
+        if mode in ("int8", "int4", "gptq", "awq"):
             self._set_quantization_info()
 
         if mode == "int8":
             # Quantize `self._kernel` to int8 and compute corresponding scale
-            kernel_value, kernel_scale = quantizers.abs_max_quantize(
-                self._kernel, axis=self._kernel_reduced_axes, to_numpy=True
+            weight_quantizer = QuantizationConfig.weight_quantizer_or_default(
+                self.quantization_config,
+                quantizers.AbsMaxQuantizer(axis=self._kernel_reduced_axes),
+            )
+            kernel_value, kernel_scale = weight_quantizer(
+                self._kernel, to_numpy=True
             )
             kernel_scale = self._adjust_scale_for_quant(kernel_scale, "kernel")
             del self._kernel
         elif mode == "int4":
             # Quantize to int4 values (stored in int8 dtype, range [-8, 7])
-            kernel_value_int4, kernel_scale = quantizers.abs_max_quantize(
-                self._kernel,
-                axis=self._kernel_reduced_axes,
-                value_range=(-8, 7),
-                dtype="int8",
-                to_numpy=True,
+            weight_quantizer = QuantizationConfig.weight_quantizer_or_default(
+                self.quantization_config,
+                quantizers.AbsMaxQuantizer(
+                    axis=self._kernel_reduced_axes,
+                    value_range=(-8, 7),
+                    output_dtype="int8",
+                ),
+            )
+            kernel_value_int4, kernel_scale = weight_quantizer(
+                self._kernel, to_numpy=True
             )
             kernel_scale = self._adjust_scale_for_quant(kernel_scale, "kernel")
 
@@ -790,7 +1188,7 @@ class EinsumDense(Layer):
             )
             kernel_value = packed_kernel_value
             del self._kernel
-        self.quantized_build(kernel_shape, mode)
+        self.quantized_build(kernel_shape, mode, self.quantization_config)
 
         # Assign values to the newly created variables.
         if mode in ("int8", "int4"):
@@ -799,7 +1197,14 @@ class EinsumDense(Layer):
 
         # Set new dtype policy
         if self.dtype_policy.quantization_mode is None:
-            policy = dtype_policies.get(f"{mode}_from_{self.dtype_policy.name}")
+            policy_name = mode
+            if mode == "gptq":
+                policy_name = self.quantization_config.dtype_policy_string()
+            elif mode == "awq":
+                policy_name = self.quantization_config.dtype_policy_string()
+            policy = dtype_policies.get(
+                f"{policy_name}_from_{self.dtype_policy.name}"
+            )
             self.dtype_policy = policy
 
     def _get_kernel_scale_shape(self, kernel_shape):
@@ -860,7 +1265,7 @@ class EinsumDense(Layer):
                     This is `None` if the layer is not quantized.
         """
         # If not a quantized layer, return the full-precision kernel directly.
-        if self.dtype_policy.quantization_mode is None:
+        if self.dtype_policy.quantization_mode in (None, "gptq", "awq"):
             return self.kernel, None
 
         # If quantized but LoRA is not enabled, return the original quantized