Address PR feedback

Signed-off-by: SimJeg <[email protected]>

Author: SimJeg

evaluation/evaluate.py

@@ -329,9 +329,7 @@ def _load_and_prepare_dataset(self):
             # FinchPress uses a delimiter token to separate context and question
             # So we need to update the tokenizer and the model embeddings.
             logger.info("FinchPress detected, updating model and tokenizer with delimiter token.")
-            self.press.update_model_and_tokenizer(
-                self.pipeline.model, self.pipeline.tokenizer
-            )  # type: ignore[attr-defined]
+            self.press.update_model_and_tokenizer(self.pipeline.model, self.pipeline.tokenizer)  # type: ignore[attr-defined]
             df["context"] = df["context"] + self.press.delimiter_token  # type: ignore[attr-defined, index]
 
         if self.config.compress_questions:

tests/test_decoding_compression.py

@@ -239,8 +239,6 @@ def test_all_presses_work_with_decoding_press(press_config):
         # CompactorPress -> Meant for prefill scenario.
         logger.info(f"Press {press_cls.__name__} is not supported, skipping test")
         return
-    if hasattr(base_press, "post_init_from_model"):
-        base_press.post_init_from_model(pipe.model)
 
     # Create DecodingPress with this base press
     decoding_press = DecodingPress(base_press=base_press, compression_interval=3, target_size=48)

tests/test_pipeline.py

@@ -17,7 +17,7 @@
 from tests.fixtures import kv_press_llama3_2_flash_attn_pipeline, kv_press_unit_test_pipeline  # noqa: F401
 
 
-@pytest.mark.skip(reason="Non kvpress related error")
+@pytest.mark.xfail(reason="Known issue not related to kvpress", strict=False)
 def test_pipeline(kv_press_unit_test_pipeline, caplog):  # noqa: F811
     with caplog.at_level(logging.DEBUG):
         context = "This is a test article. It was written on 2022-01-01."