Enable pytest live log and show warning logs on GitHub Actions CI runs (huggingface#35912)

* fix

* remove

* fix

---------

Co-authored-by: ydshieh <[email protected]>

Author: ydshieh

pyproject.toml

@@ -52,3 +52,5 @@ markers = [
     "bitsandbytes: select (or deselect with `not`) bitsandbytes integration tests",
     "generate: marks tests that use the GenerationTesterMixin"
 ]
+log_cli = 1
+log_cli_level = "WARNING"

src/transformers/generation/configuration_utils.py

@@ -785,8 +785,7 @@ def validate(self, is_init=False):
             for arg_name in ("cache_implementation", "cache_config", "return_legacy_cache"):
                 if getattr(self, arg_name) is not None:
                     logger.warning_once(
-                        no_cache_warning.format(cache_arg=arg_name, cache_arg_value=getattr(self, arg_name)),
-                        UserWarning,
+                        no_cache_warning.format(cache_arg=arg_name, cache_arg_value=getattr(self, arg_name))
                     )
 
         # 6.  check watermarking arguments

src/transformers/utils/logging.py

@@ -101,7 +101,8 @@ def _configure_library_root_logger() -> None:
             formatter = logging.Formatter("[%(levelname)s|%(pathname)s:%(lineno)s] %(asctime)s >> %(message)s")
             _default_handler.setFormatter(formatter)
 
-        library_root_logger.propagate = False
+        is_ci = os.getenv("CI") is not None and os.getenv("CI").upper() in {"1", "ON", "YES", "TRUE"}
+        library_root_logger.propagate = True if is_ci else False
 
 
 def _reset_library_root_logger() -> None:

tests/generation/test_streamers.py

@@ -16,6 +16,7 @@
 import unittest
 from queue import Empty
 from threading import Thread
+from unittest.mock import patch
 
 import pytest
 
@@ -27,6 +28,7 @@
     is_torch_available,
 )
 from transformers.testing_utils import CaptureStdout, require_torch, torch_device
+from transformers.utils.logging import _get_library_root_logger
 
 from ..test_modeling_common import ids_tensor
 
@@ -102,9 +104,12 @@ def test_text_streamer_decode_kwargs(self):
         model.config.eos_token_id = -1
 
         input_ids = torch.ones((1, 5), device=torch_device).long() * model.config.bos_token_id
-        with CaptureStdout() as cs:
-            streamer = TextStreamer(tokenizer, skip_special_tokens=True)
-            model.generate(input_ids, max_new_tokens=1, do_sample=False, streamer=streamer)
+
+        root = _get_library_root_logger()
+        with patch.object(root, "propagate", False):
+            with CaptureStdout() as cs:
+                streamer = TextStreamer(tokenizer, skip_special_tokens=True)
+                model.generate(input_ids, max_new_tokens=1, do_sample=False, streamer=streamer)
 
         # The prompt contains a special token, so the streamer should not print it. As such, the output text, when
         # re-tokenized, must only contain one token