Make sure the initial decoding cache has the proper types (#346)

lib/bumblebee/text/generation.ex

@@ -249,7 +249,7 @@ defmodule Bumblebee.Text.Generation do
           })
 
         max_length = max_length_fun.(1)
-        inputs = prepare_decoder_inputs(inputs, "decoder_", spec, max_length)
+        inputs = prepare_decoder_inputs(inputs, "decoder_", spec, model, max_length)
         {inputs, inputs["decoder_input_ids"], max_length}
       end
 
@@ -260,7 +260,7 @@ defmodule Bumblebee.Text.Generation do
       prepare_inputs_fun = fn inputs, _params ->
         sequence_length = Nx.axis_size(inputs["input_ids"], 1)
         max_length = max_length_fun.(sequence_length)
-        inputs = prepare_decoder_inputs(inputs, "", spec, max_length)
+        inputs = prepare_decoder_inputs(inputs, "", spec, model, max_length)
         {inputs, inputs["input_ids"], max_length}
       end
 
@@ -279,7 +279,7 @@ defmodule Bumblebee.Text.Generation do
     inputs["input_ids"] || inputs["input_features"] || inputs["pixel_values"]
   end
 
-  defp prepare_decoder_inputs(inputs, prefix, spec, max_length) do
+  defp prepare_decoder_inputs(inputs, prefix, spec, model, max_length) do
     input_ids = inputs[prefix <> "input_ids"]
     attention_mask = inputs[prefix <> "attention_mask"] || Nx.broadcast(1, input_ids)
 
@@ -295,9 +295,32 @@ defmodule Bumblebee.Text.Generation do
 
     batch_size = Nx.axis_size(input_ids, 0)
     cache = init_cache(spec, batch_size, max_length, inputs)
+
+    output_policy = model_output_policy(model)
+
+    # TODO: fix Axon.MixedPrecision.cast/2 to not cast integers, to
+    # match Axon compiler
+
+    # Cast all float cache tensors to match the model output. This way
+    # we make sure the cache we pass as input has the same types as
+    # the updated cache returned from the model
+    cache =
+      Bumblebee.Utils.Nx.map(cache, fn tensor ->
+        if Nx.Type.integer?(Nx.type(tensor)) do
+          tensor
+        else
+          Axon.MixedPrecision.cast(output_policy, tensor, :output)
+        end
+      end)
+
     Map.put(inputs, "cache", cache)
   end
 
+  defp model_output_policy(model) do
+    {node, _} = Axon.pop_node(model)
+    node.policy
+  end
+
   defp update_decoder_inputs(prefix, inputs, cache, token_ids) do
     inputs
     |> Map.replace!(prefix <> "input_ids", token_ids)

test/bumblebee/text/bart_test.exs

@@ -154,4 +154,32 @@ defmodule Bumblebee.Text.BartTest do
 
     assert_equal(token_ids, Nx.tensor([[988, 988, 988]]))
   end
+
+  test "generation with :for_conditional_generation and lower precision" do
+    assert {:ok, %{model: model, params: params, spec: spec}} =
+             Bumblebee.load_model(
+               {:hf, "hf-internal-testing/tiny-random-BartForConditionalGeneration"},
+               type: :f16
+             )
+
+    {:ok, generation_config} =
+      Bumblebee.load_generation_config(
+        {:hf, "hf-internal-testing/tiny-random-BartForConditionalGeneration"}
+      )
+
+    assert %Bumblebee.Text.Bart{architecture: :for_conditional_generation} = spec
+
+    inputs = %{
+      "input_ids" => Nx.tensor([[10, 20, 30, 40, 50, 60, 70, 80, 0, 0]]),
+      "attention_mask" => Nx.tensor([[1, 1, 1, 1, 1, 1, 1, 1, 0, 0]]),
+      "seed" => Nx.tensor([0])
+    }
+
+    generation_config = Bumblebee.configure(generation_config, max_new_tokens: 3)
+
+    generate = Bumblebee.Text.Generation.build_generate(model, spec, generation_config)
+    %{token_ids: token_ids} = generate.(params, inputs)
+
+    assert_equal(token_ids, Nx.tensor([[988, 988, 988]]))
+  end
 end