Use model state

lib/bumblebee.ex

@@ -299,7 +299,7 @@ defmodule Bumblebee do
   """
   @type model_info :: %{
           model: Axon.t(),
-          params: map(),
+          params: %Axon.ModelState{},
           spec: Bumblebee.ModelSpec.t()
         }
 

lib/bumblebee/conversion/pytorch_params.ex

@@ -29,7 +29,7 @@ defmodule Bumblebee.Conversion.PyTorchParams do
       `Bumblebee.Conversion.PyTorchLoader.load!/1`
 
   """
-  @spec load_params!(Axon.t(), map(), Path.t() | list(Path.t()), keyword()) :: map()
+  @spec load_params!(Axon.t(), map(), Path.t() | list(Path.t()), keyword()) :: %Axon.ModelState{}
   def load_params!(model, input_template, path, opts \\ []) do
     opts =
       opts
@@ -55,25 +55,27 @@ defmodule Bumblebee.Conversion.PyTorchParams do
         end)
         |> Enum.reduce(&Map.merge/2)
 
-      params_expr = Axon.trace_init(model, input_template)
+      model_state = Axon.trace_init(model, input_template)
 
+      params_expr = model_state.data
       {params, diff} = init_params(model, params_expr, pytorch_state, opts[:params_mapping])
+      model_state = %{model_state | data: params}
 
       params_complete? = diff.missing == [] and diff.mismatched == []
 
-      params =
+      model_state =
         if params_complete? do
-          params
+          model_state
         else
           {init_fun, _} = Axon.build(model, compiler: Nx.Defn.Evaluator)
-          init_fun.(input_template, params)
+          init_fun.(input_template, model_state)
         end
 
       if Keyword.get(opts, :log_params_diff, not params_complete?) do
         log_params_diff(diff)
       end
 
-      params
+      model_state
     end)
   end
 

mix.lock

@@ -1,5 +1,5 @@
 %{
-  "axon": {:git, "https://github.com/elixir-nx/axon.git", "7e0e5930ac4b8d2a89f48106b8121e103e597c89", []},
+  "axon": {:git, "https://github.com/elixir-nx/axon.git", "054eb4c1c224582528e8d1603ad08e7c4088f21c", []},
   "bypass": {:hex, :bypass, "2.1.0", "909782781bf8e20ee86a9cabde36b259d44af8b9f38756173e8f5e2e1fabb9b1", [:mix], [{:plug, "~> 1.7", [hex: :plug, repo: "hexpm", optional: false]}, {:plug_cowboy, "~> 2.0", [hex: :plug_cowboy, repo: "hexpm", optional: false]}, {:ranch, "~> 1.3", [hex: :ranch, repo: "hexpm", optional: false]}], "hexpm", "d9b5df8fa5b7a6efa08384e9bbecfe4ce61c77d28a4282f79e02f1ef78d96b80"},
   "castore": {:hex, :castore, "1.0.4", "ff4d0fb2e6411c0479b1d965a814ea6d00e51eb2f58697446e9c41a97d940b28", [:mix], [], "hexpm", "9418c1b8144e11656f0be99943db4caf04612e3eaecefb5dae9a2a87565584f8"},
   "cc_precompiler": {:hex, :cc_precompiler, "0.1.8", "933a5f4da3b19ee56539a076076ce4d7716d64efc8db46fd066996a7e46e2bfd", [:mix], [{:elixir_make, "~> 0.7.3", [hex: :elixir_make, repo: "hexpm", optional: false]}], "hexpm", "176bdf4366956e456bf761b54ad70bc4103d0269ca9558fd7cee93d1b3f116db"},

test/bumblebee/conversion/pytorch_params_test.exs

@@ -39,7 +39,7 @@ defmodule Bumblebee.Conversion.PyTorchParamsTest do
 
       log =
         ExUnit.CaptureLog.capture_log(fn ->
-          params =
+          %Axon.ModelState{data: params} =
             PyTorchParams.load_params!(model, input_template(), path,
               params_mapping: params_mapping()
             )
@@ -89,7 +89,7 @@ defmodule Bumblebee.Conversion.PyTorchParamsTest do
 
       log =
         ExUnit.CaptureLog.capture_log(fn ->
-          params =
+          %Axon.ModelState{data: params} =
             PyTorchParams.load_params!(model, input_template(), path,
               params_mapping: params_mapping()
             )
@@ -107,7 +107,7 @@ defmodule Bumblebee.Conversion.PyTorchParamsTest do
 
       log =
         ExUnit.CaptureLog.capture_log(fn ->
-          params =
+          %Axon.ModelState{data: params} =
             PyTorchParams.load_params!(model, input_template(), path,
               params_mapping: params_mapping()
             )

test/bumblebee/text/roberta_test.exs

@@ -33,7 +33,10 @@ defmodule Bumblebee.Text.RobertaTest do
     assert %Bumblebee.Text.Roberta{architecture: :for_masked_language_modeling} = spec
 
     # TODO: remove once we load tied embeddings
-    params = put_in(params["language_modeling_head.output"], params["embedder.token_embedding"])
+    params =
+      update_in(params, [Access.key!(:data)], fn data ->
+        put_in(data["language_modeling_head.output"], data["embedder.token_embedding"])
+      end)
 
     inputs = %{
       "input_ids" => Nx.tensor([[10, 20, 30, 40, 50, 60, 70, 80, 0, 0]]),
@@ -157,7 +160,10 @@ defmodule Bumblebee.Text.RobertaTest do
     assert %Bumblebee.Text.Roberta{architecture: :for_causal_language_modeling} = spec
 
     # TODO: remove once we load tied embeddings
-    params = put_in(params["language_modeling_head.output"], params["embedder.token_embedding"])
+    params =
+      update_in(params, [Access.key!(:data)], fn data ->
+        put_in(data["language_modeling_head.output"], data["embedder.token_embedding"])
+      end)
 
     inputs = %{
       "input_ids" => Nx.tensor([[10, 20, 30, 40, 50, 60, 70, 80, 0, 0]]),

test/bumblebee_test.exs

@@ -76,15 +76,15 @@ defmodule BumblebeeTest do
     end
 
     test "passing :type casts params accordingly" do
-      assert {:ok, %{params: params}} =
+      assert {:ok, %{params: %Axon.ModelState{data: params}}} =
                Bumblebee.load_model({:hf, "hf-internal-testing/tiny-random-GPT2Model"},
                  type: :bf16
                )
 
       assert Nx.type(params["decoder.blocks.0.ffn.output"]["kernel"]) == {:bf, 16}
       assert Nx.type(params["decoder.blocks.0.ffn.output"]["bias"]) == {:bf, 16}
 
-      assert {:ok, %{params: params}} =
+      assert {:ok, %{params: %Axon.ModelState{data: params}}} =
                Bumblebee.load_model({:hf, "hf-internal-testing/tiny-random-GPT2Model"},
                  type: Axon.MixedPrecision.create_policy(params: :f16)
                )