Speed up tests

test/bumblebee/text/bart_test.exs

@@ -144,21 +144,16 @@ defmodule Bumblebee.Text.BartTest do
 
   test "conditional generation" do
     {:ok, model_info} = Bumblebee.load_model({:hf, "facebook/bart-large-cnn"})
-    {:ok, tokenizer} = Bumblebee.load_tokenizer({:hf, "facebook/bart-large-cnn"})
     {:ok, generation_config} = Bumblebee.load_generation_config({:hf, "facebook/bart-large-cnn"})
 
     assert %Bumblebee.Text.Bart{architecture: :for_conditional_generation} = model_info.spec
 
-    article = """
-    PG&E stated it scheduled the blackouts in response to forecasts for high \
-    winds amid dry conditions. The aim is to reduce the risk of wildfires. \
-    Nearly 800 thousand customers were scheduled to be affected by the shutoffs \
-    which were expected to last through at least midday tomorrow.
-    """
+    inputs = %{
+      "input_ids" => Nx.tensor([[0, 133, 812, 9, 1470, 16, 2201, 2]]),
+      "attention_mask" => Nx.tensor([[1, 1, 1, 1, 1, 1, 1, 1]])
+    }
 
-    inputs = Bumblebee.apply_tokenizer(tokenizer, article)
-
-    generation_config = Bumblebee.configure(generation_config, max_length: 8)
+    generation_config = Bumblebee.configure(generation_config, max_new_tokens: 5)
 
     generate =
       Bumblebee.Text.Generation.build_generate(
@@ -167,8 +162,8 @@ defmodule Bumblebee.Text.BartTest do
         generation_config
       )
 
-    token_ids = EXLA.jit(generate).(model_info.params, inputs)
+    token_ids = generate.(model_info.params, inputs)
 
-    assert Bumblebee.Tokenizer.decode(tokenizer, token_ids) == ["PG&E scheduled the"]
+    assert_equal(token_ids, Nx.tensor([[2, 0, 133, 812, 9, 2]]))
   end
 end

test/bumblebee/text/blenderbot_test.exs

@@ -55,18 +55,18 @@ defmodule Bumblebee.Text.BlenderbotTest do
 
   test "conditional generation" do
     {:ok, model_info} = Bumblebee.load_model({:hf, "facebook/blenderbot-400M-distill"})
-    {:ok, tokenizer} = Bumblebee.load_tokenizer({:hf, "facebook/blenderbot-400M-distill"})
 
     {:ok, generation_config} =
       Bumblebee.load_generation_config({:hf, "facebook/blenderbot-400M-distill"})
 
     assert %Bumblebee.Text.Blenderbot{architecture: :for_conditional_generation} = model_info.spec
 
-    english_phrase = " Hey, how are you?"
+    inputs = %{
+      "input_ids" => Nx.tensor([[2675, 19, 544, 366, 304, 38, 2]]),
+      "attention_mask" => Nx.tensor([[1, 1, 1, 1, 1, 1, 1]])
+    }
 
-    inputs = Bumblebee.apply_tokenizer(tokenizer, english_phrase)
-
-    generation_config = Bumblebee.configure(generation_config, min_length: 0, max_length: 6)
+    generation_config = Bumblebee.configure(generation_config, max_new_tokens: 5)
 
     generate =
       Bumblebee.Text.Generation.build_generate(
@@ -75,8 +75,8 @@ defmodule Bumblebee.Text.BlenderbotTest do
         generation_config
       )
 
-    token_ids = EXLA.jit(generate).(model_info.params, inputs)
+    token_ids = generate.(model_info.params, inputs)
 
-    assert Bumblebee.Tokenizer.decode(tokenizer, token_ids) == [" I'm doing well"]
+    assert_equal(token_ids, Nx.tensor([[1, 281, 476, 929, 731, 2]]))
   end
 end

test/bumblebee/text/generation_test.exs

@@ -23,9 +23,7 @@ defmodule Bumblebee.Text.GenerationTest do
       generation_config = Bumblebee.configure(generation_config, max_new_tokens: 8)
 
       serving =
-        Bumblebee.Text.generation(model_info, tokenizer, generation_config,
-          defn_options: [compiler: EXLA]
-        )
+        Bumblebee.Text.generation(model_info, tokenizer, generation_config)
 
       assert %{results: [%{text: "PG&E scheduled the black"}]} = Nx.Serving.run(serving, article)
     end
@@ -39,9 +37,7 @@ defmodule Bumblebee.Text.GenerationTest do
         Bumblebee.configure(generation_config, max_new_tokens: 12, no_repeat_ngram_length: 2)
 
       serving =
-        Bumblebee.Text.generation(model_info, tokenizer, generation_config,
-          defn_options: [compiler: EXLA]
-        )
+        Bumblebee.Text.generation(model_info, tokenizer, generation_config)
 
       # Without :no_repeat_ngram_length we get
       # %{results: [%{text: "I was going to say, 'Well, I'm going to say,"}]}
@@ -62,10 +58,7 @@ defmodule Bumblebee.Text.GenerationTest do
         )
 
       serving =
-        Bumblebee.Text.generation(model_info, tokenizer, generation_config,
-          seed: 0,
-          defn_options: [compiler: EXLA]
-        )
+        Bumblebee.Text.generation(model_info, tokenizer, generation_config, seed: 0)
 
       # Note that this is just a snapshot test, we do not use any
       # reference value, because of PRNG difference
@@ -89,9 +82,7 @@ defmodule Bumblebee.Text.GenerationTest do
         )
 
       serving =
-        Bumblebee.Text.generation(model_info, tokenizer, generation_config,
-          defn_options: [compiler: EXLA]
-        )
+        Bumblebee.Text.generation(model_info, tokenizer, generation_config)
 
       assert %{results: [%{text: "I was going to say, 'Well, I don't know what you"}]} =
                Nx.Serving.run(serving, "I was going")

test/bumblebee/text/mbart_test.exs

@@ -158,18 +158,17 @@ defmodule Bumblebee.Text.MbartTest do
         module: Bumblebee.Text.Mbart
       )
 
-    {:ok, tokenizer} = Bumblebee.load_tokenizer({:hf, "facebook/mbart-large-en-ro"})
-
     {:ok, generation_config} =
       Bumblebee.load_generation_config({:hf, "facebook/mbart-large-en-ro"})
 
     assert %Bumblebee.Text.Mbart{architecture: :for_conditional_generation} = model_info.spec
 
-    english_phrase = "42 is the answer"
-
-    inputs = Bumblebee.apply_tokenizer(tokenizer, english_phrase)
+    inputs = %{
+      "input_ids" => Nx.tensor([[4828, 83, 70, 35166, 2, 250_004]]),
+      "attention_mask" => Nx.tensor([[1, 1, 1, 1, 1, 1]])
+    }
 
-    generation_config = Bumblebee.configure(generation_config, min_length: 0, max_length: 6)
+    generation_config = Bumblebee.configure(generation_config, max_new_tokens: 5)
 
     generate =
       Bumblebee.Text.Generation.build_generate(
@@ -178,8 +177,8 @@ defmodule Bumblebee.Text.MbartTest do
         generation_config
       )
 
-    token_ids = EXLA.jit(generate).(model_info.params, inputs)
+    token_ids = generate.(model_info.params, inputs)
 
-    assert Bumblebee.Tokenizer.decode(tokenizer, token_ids) == ["42 este răspunsul"]
+    assert_equal(token_ids, Nx.tensor([[250_020, 4828, 473, 54051, 202, 2]]))
   end
 end

test/bumblebee/text/t5_test.exs

@@ -146,15 +146,15 @@ defmodule Bumblebee.Text.T5Test do
     end
 
     test "text generation" do
-      assert {:ok, tokenizer} = Bumblebee.load_tokenizer({:hf, "t5-small"})
       assert {:ok, model_info} = Bumblebee.load_model({:hf, "t5-small"})
       assert {:ok, generation_config} = Bumblebee.load_generation_config({:hf, "t5-small"})
 
-      text = "translate English to German: How old are you?"
-
-      inputs = Bumblebee.apply_tokenizer(tokenizer, text)
+      inputs = %{
+        "input_ids" => Nx.tensor([[13959, 1566, 12, 2968, 10, 571, 625, 33, 25, 58, 1]]),
+        "attention_mask" => Nx.tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]])
+      }
 
-      generation_config = Bumblebee.configure(generation_config, min_length: 0, max_length: 10)
+      generation_config = Bumblebee.configure(generation_config, max_new_tokens: 5)
 
       generate =
         Bumblebee.Text.Generation.build_generate(
@@ -163,23 +163,23 @@ defmodule Bumblebee.Text.T5Test do
           generation_config
         )
 
-      token_ids = EXLA.jit(generate).(model_info.params, inputs)
+      token_ids = generate.(model_info.params, inputs)
 
-      assert Bumblebee.Tokenizer.decode(tokenizer, token_ids) == ["Wie alt sind Sie?"]
+      assert_equal(token_ids, Nx.tensor([[0, 2739, 4445, 436, 292, 58]]))
     end
 
     test "text generation (tied embeddings)" do
-      assert {:ok, tokenizer} = Bumblebee.load_tokenizer({:hf, "google/flan-t5-small"})
       assert {:ok, model_info} = Bumblebee.load_model({:hf, "google/flan-t5-small"})
 
       assert {:ok, generation_config} =
                Bumblebee.load_generation_config({:hf, "google/flan-t5-small"})
 
-      text = "translate English to German: How old are you?"
-
-      inputs = Bumblebee.apply_tokenizer(tokenizer, text)
+      inputs = %{
+        "input_ids" => Nx.tensor([[13959, 1566, 12, 2968, 10, 571, 625, 33, 25, 58, 1]]),
+        "attention_mask" => Nx.tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]])
+      }
 
-      generation_config = Bumblebee.configure(generation_config, min_length: 0, max_length: 10)
+      generation_config = Bumblebee.configure(generation_config, max_new_tokens: 5)
 
       generate =
         Bumblebee.Text.Generation.build_generate(
@@ -188,9 +188,9 @@ defmodule Bumblebee.Text.T5Test do
           generation_config
         )
 
-      token_ids = EXLA.jit(generate).(model_info.params, inputs)
+      token_ids = generate.(model_info.params, inputs)
 
-      assert Bumblebee.Tokenizer.decode(tokenizer, token_ids) == ["Wie ich er bitten?"]
+      assert_equal(token_ids, Nx.tensor([[0, 2739, 3, 362, 3, 49]]))
     end
   end
 end

test/bumblebee/vision/image_to_text_test.exs

@@ -20,9 +20,7 @@ defmodule Bumblebee.Vision.ImageToTextTest do
         Bumblebee.load_generation_config({:hf, "Salesforce/blip-image-captioning-base"})
 
       serving =
-        Bumblebee.Vision.ImageToText.image_to_text(blip, featurizer, tokenizer, generation_config,
-          defn_options: [compiler: EXLA]
-        )
+        Bumblebee.Vision.ImageToText.image_to_text(blip, featurizer, tokenizer, generation_config)
 
       image = StbImage.read_file!(Path.join(@images_dir, "coco/39769.jpeg"))