Fixing GHA Hangs (#2265)

* test commit

* update runners

* move pytest to transformers

* update import

* fix pathing

* don't automatically ifer reporters

* remove clearml

* my attempt to unclot GHA

* update model for commit tests

* fix export test

* can add back in clearml I think

* quality

* oops!

* remove clearml support

* PR comments

* QUALITY

* int tests

---------

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

setup.py

@@ -120,7 +120,6 @@
     "tensorboardX>=1.0",
     "evaluate>=0.4.1",
     "parameterized",
-    "clearml==1.14.4",
 ]
 
 _docs_deps = [

src/sparseml/transformers/finetune/text_generation.py

@@ -85,8 +85,12 @@ def apply(**kwargs):
     """
     CLI entrypoint for any of training, eval, predict or oneshot
     """
+    report_to = kwargs.get("report_to", None)
     model_args, data_args, training_args = parse_args(**kwargs)
     training_args.run_stages = True
+    if report_to is None:  # user didn't specify any reporters
+        # get rid of the reporters inferred from hugging face
+        training_args.report_to = []
     main(model_args, data_args, training_args)
 
 

tests/integrations/transformers/args.py

@@ -507,12 +507,6 @@ class _TransformersTrainArgs(BaseModel):
     push_to_hub_token: str = Field(
         default=None, description="The token to use to push to the Model Hub."
     )
-    n_gpu_: int = Field(init=False, repr=False, default=-1)
-    mp_parameters: Optional[str] = Field(
-        default=None,
-        description="Used by the SageMaker launcher to send mp-specific args. "
-        "Ignored in Trainer",
-    )
 
 
 class QuestionAnsweringArgs(_TransformersTrainArgs):

tests/sparseml/export/transformers/test_generative_transformers.py

@@ -42,13 +42,13 @@ def setup(self, tmp_path, stub, task):
 
         shutil.rmtree(tmp_path)
 
-    def test_export_initialized_model_no_source_path(self, setup):
+    def test_export_initialized_model_no_source_path(self, tmp_path, stub, task):
         # export the transformer model, that is being passed to the
         # `export` API directly as an object
-        source_path, target_path, task = setup
+        target_path = tmp_path / "target"
         export(
-            model=SparseAutoModelForCausalLM.from_pretrained(source_path),
-            tokenizer=SparseAutoTokenizer.from_pretrained(source_path),
+            model=SparseAutoModelForCausalLM.from_pretrained(stub),
+            tokenizer=SparseAutoTokenizer.from_pretrained(stub),
             target_path=target_path,
             sequence_length=384,
             task=task,
@@ -68,6 +68,8 @@ def test_export_initialized_model_no_source_path(self, setup):
             inp.name == "past_key_values.0.key" for inp in onnx_model.graph.input
         )
 
+        shutil.rmtree(tmp_path)
+
     def test_export_happy_path(self, setup):
         source_path, target_path, task = setup
         export(

tests/sparseml/pytorch/utils/test_logger.py

@@ -19,8 +19,7 @@
 
 import pytest
 
-from sparseml.pytorch.utils import (
-    ClearMLLogger,
+from sparseml.pytorch.utils import (  # ClearMLLogger,
     LambdaLogger,
     LoggerManager,
     PythonLogger,
@@ -46,7 +45,8 @@
             or True
         ),
         *([WANDBLogger()] if WANDBLogger.available() else []),
-        *([ClearMLLogger()] if ClearMLLogger.available() else []),
+        # TODO: uncomment this once the ClearML server hanging is fixed
+        # *([ClearMLLogger()] if ClearMLLogger.available() else []),
         SparsificationGroupLogger(
             lambda_func=lambda tag, value, values, step, wall_time, level: logging.info(
                 f"{tag}, {value}, {values}, {step}, {wall_time}, {level}"

tests/sparseml/pytorch/oneshot/oneshot_configs/tiny_stories_conf1.yaml → tests/sparseml/transformers/oneshot/oneshot_configs/tiny_stories_conf1.yaml

@@ -1,7 +1,7 @@
 cadence: "commit"
 test_type: "smoke"
 tokenize: False
-model: "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
+model: "Xenova/llama2.c-stories15M"
 dataset: open_platypus
 recipe: |
   test_stage:

tests/sparseml/transformers/oneshot/oneshot_configs/tiny_stories_conf2.yaml

@@ -0,0 +1,6 @@
+cadence: "commit"
+test_type: "smoke"
+tokenize: False
+model: "Xenova/llama2.c-stories15M"
+dataset: open_platypus
+recipe: "tests/sparseml/transformers/oneshot/oneshot_configs/recipes/recipe.yaml"

tests/sparseml/transformers/oneshot/oneshot_configs/tiny_stories_conf3.yaml

@@ -0,0 +1,7 @@
+cadence: "commit"
+test_type: "smoke"
+tokenize: False
+model: "Xenova/llama2.c-stories15M"
+dataset: "gsm8k"
+dataset_config_name: "main"
+recipe: "tests/sparseml/transformers/oneshot/oneshot_configs/recipes/recipe.yaml"

tests/sparseml/pytorch/oneshot/oneshot_configs/tiny_stories_conf4.yaml → tests/sparseml/transformers/oneshot/oneshot_configs/tiny_stories_conf4.yaml

@@ -1,7 +1,7 @@
 cadence: "commit"
 test_type: "smoke"
 tokenize: False
-model: "TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T"
+model: "Xenova/llama2.c-stories15M"
 dataset: "gsm8k"
 dataset_config_name: "main"
 recipe: |

tests/sparseml/transformers/oneshot/oneshot_configs/tiny_stories_conf5.yaml

@@ -0,0 +1,6 @@
+cadence: "commit"
+test_type: "smoke"
+tokenize: True
+model: "Xenova/llama2.c-stories15M"
+dataset: open_platypus
+recipe: "tests/sparseml/transformers/oneshot/oneshot_configs/recipes/recipe.yaml"

tests/sparseml/transformers/oneshot/oneshot_configs/tiny_stories_conf6.yaml

@@ -0,0 +1,6 @@
+cadence: "commit"
+test_type: "smoke"
+tokenize: True
+model: "Xenova/llama2.c-stories15M"
+dataset: "gsm8k"
+recipe: "tests/sparseml/transformers/oneshot/oneshot_configs/recipes/recipe.yaml"

tests/sparseml/pytorch/oneshot/test_api_inputs.py → tests/sparseml/transformers/oneshot/test_api_inputs.py

@@ -18,11 +18,11 @@
 import pytest
 
 from parameterized import parameterized_class
-from tests.sparseml.pytorch.oneshot.dataset_processing import get_data_utils
+from tests.sparseml.transformers.oneshot.dataset_processing import get_data_utils
 from tests.testing_utils import parse_params, requires_torch
 
 
-CONFIGS_DIRECTORY = "tests/sparseml/pytorch/oneshot/oneshot_configs"
+CONFIGS_DIRECTORY = "tests/sparseml/transformers/oneshot/oneshot_configs"
 
 # TODO: Seems better to mark test type (smoke, sanity, regression) as a marker as
 # opposed to using a field in the config file?

tests/sparseml/pytorch/oneshot/test_cli.py → tests/sparseml/transformers/oneshot/test_cli.py

@@ -21,7 +21,7 @@
 from tests.testing_utils import parse_params, requires_torch, run_cli_command
 
 
-CONFIGS_DIRECTORY = "tests/sparseml/pytorch/oneshot/oneshot_configs"
+CONFIGS_DIRECTORY = "tests/sparseml/transformers/oneshot/oneshot_configs"
 
 
 @pytest.mark.smoke