Merge pull request #189 from stanfordnlp/gemma

[P1] Update the version requirement for transformers and its dependencies.

requirements.txt

@@ -1,12 +1,13 @@
 torch>=2.0.0
-transformers==4.40.2
-datasets>=2.16.1
+transformers==4.45.1
+tokenizers>=0.20.0
+datasets>=3.0.1
 protobuf>=3.20.0
 matplotlib>=3.7.4
 ipywidgets>=8.1.1
 plotnine>=0.12.4
-huggingface-hub>=0.24.0
+huggingface-hub>=0.25.1
 numpy>=1.23.5
 fsspec>=2023.6.0
-accelerate>=0.29.1
+accelerate>=0.34.2
 sentencepiece>=0.1.96

tests/integration_tests/ComplexInterventionWithGPT2TestCase.py

@@ -63,10 +63,11 @@ def test_clean_run_positive(self):
         base = {"input_ids": torch.randint(0, 10, (10, 5)).to(self.device)}
         golden_out = self.gpt2(**base).logits
         our_output = intervenable(base, output_original_output=True)[0][0]
-        self.assertTrue(torch.allclose(golden_out, our_output))
+        self.assertTrue(torch.allclose(golden_out, our_output, rtol=1e-05, atol=1e-06))
         # make sure the toolkit also works
         self.assertTrue(
-            torch.allclose(GPT2_RUN(self.gpt2, base["input_ids"], {}, {}), golden_out)
+            torch.allclose(GPT2_RUN(self.gpt2, base["input_ids"], {}, {}), golden_out,
+                           rtol=1e-05, atol=1e-06)
         )
 
     def _test_subspace_partition_in_forward(self, intervention_type):
@@ -134,7 +135,8 @@ def _test_subspace_partition_in_forward(self, intervention_type):
         # make sure the toolkit also works
         self.assertTrue(
             torch.allclose(
-                with_partition_our_output[0], without_partition_our_output[0]
+                with_partition_our_output[0], without_partition_our_output[0],
+                rtol=1e-05, atol=1e-06
             )
         )
 

tests/integration_tests/IntervenableBasicTestCase.py

@@ -609,6 +609,7 @@ def test_customized_intervention_function_get(self):
 
         _, tokenizer, gpt2 = pv.create_gpt2()
 
+        gpt2.config.output_attentions = True
         pv_gpt2 = pv.IntervenableModel({
             "layer": 10,
             "component": "attention_weight",
@@ -668,4 +669,4 @@ def test_customized_intervention_function_zeroout(self):
     def tearDownClass(self):
         print(f"Removing testing dir {self._test_dir}")
         if os.path.exists(self._test_dir) and os.path.isdir(self._test_dir):
-            shutil.rmtree(self._test_dir)
+            shutil.rmtree(self._test_dir)

tests/integration_tests/InterventionWithGPT2TestCase.py

@@ -132,8 +132,9 @@ def _test_with_head_position_intervention(
                     )
                 },
             )
-
-        self.assertTrue(torch.allclose(out_output[0], golden_out))
+        # Relax the atol to 1e-6 to accommodate for different Transformers versions.
+        # The max of the absolute diff is usually between 1e-8 to 1e-7.
+        self.assertTrue(torch.allclose(out_output[0], golden_out, rtol=1e-05, atol=1e-06))
 
 
     def test_with_multiple_heads_positions_vanilla_intervention_positive(self):