Generalize quantization layers

Author: seanmor5

lib/axon.ex

@@ -3974,7 +3974,7 @@ defmodule Axon do
 
   """
   @doc type: :debug
-  def trace_init(model, template, params \\ %{}, opts \\ []) do
+  def trace_init(model, template, params \\ Axon.ModelState.empty(), opts \\ []) do
     {init_fn, _} = build(model, opts)
     Nx.Defn.jit(init_fn, compiler: Axon.Defn).(template, params)
   end

lib/axon/model_state.ex

@@ -171,12 +171,7 @@ defmodule Axon.ModelState do
   Returns an empty model state.
   """
   def empty() do
-    %Axon.ModelState{
-      data: %{},
-      parameters: %{},
-      state: %{},
-      frozen_parameters: %{}
-    }
+    new(%{})
   end
 
   @doc """
@@ -190,12 +185,40 @@ defmodule Axon.ModelState do
   def new(data) when is_map(data) do
     %Axon.ModelState{
       data: data,
-      parameters: get_paths(data),
+      parameters: transform_to_parameters(data),
       state: %{},
       frozen_parameters: %{}
     }
   end
 
+  defp transform_to_parameters(%Nx.Tensor{}), do: nil
+
+  defp transform_to_parameters(map) when is_map(map) do
+    map
+    |> Enum.map(fn {k, v} -> {k, transform_to_parameters(v)} end)
+    |> Enum.into(%{})
+  end
+
+  defp transform_to_parameters(list) when is_list(list) do
+    Enum.map(list, &transform_to_parameters/1)
+  end
+
+  defp transform_to_parameters(value) do
+    case value do
+      map when is_map(map) ->
+        keys = Map.keys(map)
+
+        if Enum.all?(keys, &(is_map(map[&1]) or match?(%Nx.Tensor{}, map[&1]))) do
+          keys
+        else
+          transform_to_parameters(map)
+        end
+
+      _ ->
+        value
+    end
+  end
+
   # Helpers
 
   defp get_paths(map) do

lib/axon/quantization/layers.ex

@@ -35,18 +35,20 @@ defmodule Axon.Quantization.Layers do
           bias,
           _opts
         ) do
-    x_shape = Nx.shape(x)
-    last_dim = Nx.axis_size(x, -1)
+    x_view = Nx.reshape(x, {:auto, Nx.axis_size(x, -1)})
 
-    x_view = Nx.reshape(x, {:auto, last_dim})
-
-    y = Nx.dot(x_view, Nx.as_type(Nx.transpose(w_int8), Nx.type(x)))
-    y = Nx.multiply(y, scales)
-    y = reshape_output(y, x_shape)
+    y = Nx.dot(x_view, Nx.as_type(w_int8, Nx.type(x)))
+    y = Nx.multiply(y, reshape_scales(scales, y))
+    y = reshape_output(y, Nx.shape(x))
 
     Nx.add(y, bias)
   end
 
+  deftransformp reshape_scales(scales, y) do
+    ones = List.to_tuple(List.duplicate(1, Nx.rank(y) - 1))
+    Nx.reshape(scales, Tuple.append(ones, :auto))
+  end
+
   deftransformp reshape_output(output, x_shape) do
     all_but_last = Tuple.delete_at(x_shape, tuple_size(x_shape) - 1)
     new_shape = Tuple.append(all_but_last, :auto)

lib/axon/quantization/q_tensor.ex

@@ -59,7 +59,7 @@ defmodule Axon.Quantization.QTensor do
         max: opts[:max]
       )
 
-    struct(__MODULE__, value: quantized_value, scale: scale, zero_point: zero_point)
+    struct(__MODULE__, value: Nx.transpose(quantized_value), scale: scale, zero_point: zero_point)
   end
 
   deftransformp quantize_affine(