Merge branch 'main' into test_update

setup.py

@@ -55,6 +55,7 @@
     "GPUtil>=1.4.0",
     "protobuf>=3.12.2,<=3.20.3",
     "click>=7.1.2,!=8.0.0",  # latest version < 8.0 + blocked version with reported bug
+    "clearml==1.14.4",
 ]
 _nm_deps = [f"{'sparsezoo' if is_release else 'sparsezoo-nightly'}~={version_nm_deps}"]
 _deepsparse_deps = [

src/sparseml/pytorch/utils/logger.py

@@ -45,11 +45,21 @@
     wandb = None
     wandb_err = err
 
+
+try:
+    from clearml import Task
+
+    clearml_err = None
+except Exception as err:
+    clearml = None
+    clearml_err = err
+
 from sparseml.utils import ALL_TOKEN, create_dirs
 
 
 __all__ = [
     "BaseLogger",
+    "ClearMLLogger",
     "LambdaLogger",
     "PythonLogger",
     "TensorBoardLogger",
@@ -628,6 +638,101 @@ def save(
         return True
 
 
+class ClearMLLogger(LambdaLogger):
+    @staticmethod
+    def available() -> bool:
+        """
+        :return: True if wandb is available and installed, False, otherwise
+        """
+        return not clearml_err
+
+    def __init__(
+        self,
+        name: str = "clearml",
+        enabled: bool = True,
+        project_name: str = "sparseml",
+        task_name: str = "",
+    ):
+        if task_name == "":
+            now = datetime.now()
+            task_name = now.strftime("%d-%m-%Y_%H.%M.%S")
+
+        self.task = Task.init(project_name=project_name, task_name=task_name)
+
+        super().__init__(
+            lambda_func=self.log_scalar,
+            name=name,
+            enabled=enabled,
+        )
+
+    def log_hyperparams(
+        self,
+        params: Dict,
+        level: Optional[int] = None,
+    ) -> bool:
+        """
+        :param params: Each key-value pair in the dictionary is the name of the
+            hyper parameter and it's corresponding value.
+        :return: True if logged, False otherwise.
+        """
+        if not self.enabled:
+            return False
+
+        self.task.connect(params)
+        return True
+
+    def log_scalar(
+        self,
+        tag: str,
+        value: float,
+        step: Optional[int] = None,
+        wall_time: Optional[float] = None,
+        level: Optional[int] = None,
+    ) -> bool:
+        """
+        :param tag: identifying tag to log the value with
+        :param value: value to save
+        :param step: global step for when the value was taken
+        :param wall_time: global wall time for when the value was taken,
+            defaults to time.time()
+        :param kwargs: additional logging arguments to support Python and custom loggers
+        :return: True if logged, False otherwise.
+        """
+        logger = self.task.get_logger()
+        # each series is superimposed on the same plot on title
+        logger.report_scalar(
+            title=tag, series=str(level) or tag, value=value, iteration=step
+        )
+        return True
+
+    def log_scalars(
+        self,
+        tag: str,
+        values: Dict[str, float],
+        step: Optional[int] = None,
+        wall_time: Optional[float] = None,
+        level: Optional[int] = None,
+    ) -> bool:
+        """
+        :param tag: identifying tag to log the values with
+        :param values: values to save
+        :param step: global step for when the values were taken
+        :param wall_time: global wall time for when the values were taken,
+            defaults to time.time()
+        :param kwargs: additional logging arguments to support Python and custom loggers
+        :return: True if logged, False otherwise.
+        """
+        for k, v in values.items():
+            self.log_scalar(
+                tag=f"{tag}.{k}",
+                value=v,
+                step=step,
+                wall_time=wall_time,
+                level=level,
+            )
+        return True
+
+
 class SparsificationGroupLogger(BaseLogger):
     """
     Modifier logger that handles outputting values to other supported systems.

src/sparseml/transformers/sparsification/modification/base.py

@@ -24,7 +24,7 @@
 __all__ = ["check_transformers_version"]
 
 _TRANSFORMERS_MIN_VERSION = "4.39.0"
-_TRANSFORMERS_MAX_VERSION = "4.39.2"
+_TRANSFORMERS_MAX_VERSION = "4.39.3"
 
 
 def check_transformers_version(
@@ -56,7 +56,7 @@ def check_transformers_version(
         _LOGGER.warning(
             "Attempting to modify the transformers model to support "
             "the SparseML-specific functionalities. However, the detected "
-            f"transformers version ({current_version}) does not fall within the"
+            f"transformers version ({current_version}) does not fall within the "
             f"supported version range ({min_version} - {max_version}). "
             "This may lead to unexpected behavior. Please ensure that the "
             "correct transformers version is installed."

src/sparseml/transformers/sparsification/sparse_model.py

@@ -40,9 +40,7 @@
     modify_save_pretrained,
 )
 from sparseml.transformers.sparsification.modification import modify_model
-from sparseml.transformers.utils.helpers import resolve_recipe
-from sparseml.utils import download_zoo_training_dir
-from sparseml.utils.fsdp.context import main_process_first_context
+from sparseml.transformers.utils.helpers import download_model_directory, resolve_recipe
 
 
 __all__ = ["SparseAutoModel", "SparseAutoModelForCausalLM", "get_shared_tokenizer_src"]
@@ -101,15 +99,9 @@ def skip(*args, **kwargs):
             else pretrained_model_name_or_path
         )
 
-        if pretrained_model_name_or_path.startswith("zoo:"):
-            _LOGGER.debug(
-                "Passed zoo stub to SparseAutoModelForCausalLM object. "
-                "Loading model from SparseZoo training files..."
-            )
-            with main_process_first_context():
-                pretrained_model_name_or_path = download_zoo_training_dir(
-                    zoo_stub=pretrained_model_name_or_path
-                )
+        pretrained_model_name_or_path = download_model_directory(
+            pretrained_model_name_or_path, **kwargs
+        )
 
         # determine compression format, if any, from the model config
         compressor = infer_compressor_from_model_config(pretrained_model_name_or_path)

src/sparseml/transformers/utils/helpers.py

@@ -34,8 +34,10 @@
 from transformers.trainer_utils import get_last_checkpoint
 from transformers.utils import PaddingStrategy
 
-from huggingface_hub import HUGGINGFACE_CO_URL_HOME, hf_hub_download
+from huggingface_hub import HUGGINGFACE_CO_URL_HOME, HfFileSystem, hf_hub_download
 from sparseml.export.helpers import ONNX_MODEL_NAME
+from sparseml.utils import download_zoo_training_dir
+from sparseml.utils.fsdp.context import main_process_first_context
 from sparsezoo import Model, setup_model
 
 
@@ -52,6 +54,8 @@
     "ALL_TASK_NAMES",
     "create_fake_dataloader",
     "POSSIBLE_TOKENIZER_FILES",
+    "download_repo_from_huggingface_hub",
+    "download_model_directory",
 ]
 
 
@@ -92,6 +96,7 @@ class TaskNames(Enum):
     "special_tokens_map.json",
     "tokenizer_config.json",
 }
+RELEVANT_HF_SUFFIXES = ["json", "md", "bin", "safetensors", "yaml", "yml", "py"]
 
 
 def remove_past_key_value_support_from_config(config: AutoConfig) -> AutoConfig:
@@ -553,3 +558,94 @@ def fetch_recipe_path(target: str):
         recipe_path = hf_hub_download(repo_id=target, filename=DEFAULT_RECIPE_NAME)
 
     return recipe_path
+
+
+def download_repo_from_huggingface_hub(repo_id, **kwargs):
+    """
+    Download relevant model files from the Hugging Face Hub
+    using the huggingface_hub.hf_hub_download function
+
+    Note(s):
+    - Does not download the entire repo, only the relevant files
+    for the model, such as the model weights, tokenizer files, etc.
+    - Does not re-download files that already exist locally, unless
+    the force_download flag is set to True
+
+    :pre-condition: the repo_id must be a valid Hugging Face Hub repo id
+    :param repo_id: the repo id to download
+    :param kwargs: additional keyword arguments to pass to hf_hub_download
+    """
+    hf_filesystem = HfFileSystem()
+    files = hf_filesystem.ls(repo_id)
+
+    if not files:
+        raise ValueError(f"Could not find any files in HF repo {repo_id}")
+
+    # All file(s) from hf_filesystem have "name" key
+    # Extract the file names from the files
+    relevant_file_names = (
+        Path(file["name"]).name
+        for file in files
+        if any(file["name"].endswith(suffix) for suffix in RELEVANT_HF_SUFFIXES)
+    )
+
+    hub_kwargs_names = (
+        "subfolder",
+        "repo_type",
+        "revision",
+        "library_name",
+        "library_version",
+        "cache_dir",
+        "local_dir",
+        "local_dir_use_symlinks",
+        "user_agent",
+        "force_download",
+        "force_filename",
+        "proxies",
+        "etag_timeout",
+        "resume_download",
+        "token",
+        "local_files_only",
+        "headers",
+        "legacy_cache_layout",
+        "endpoint",
+    )
+    hub_kwargs = {name: kwargs[name] for name in hub_kwargs_names if name in kwargs}
+
+    for file_name in relevant_file_names:
+        last_file = hf_hub_download(repo_id=repo_id, filename=file_name, **hub_kwargs)
+
+    # parent directory of the last file is the model directory
+    return str(Path(last_file).parent.resolve().absolute())
+
+
+def download_model_directory(pretrained_model_name_or_path: str, **kwargs):
+    """
+    Download the model directory from the HF hub or SparseZoo if the model
+    is not found locally
+
+    :param pretrained_model_name_or_path: the name of or path to the model to load
+        can be a SparseZoo/HuggingFace model stub
+    :param kwargs: additional keyword arguments to pass to the download function
+    :return: the path to the downloaded model directory
+    """
+    pretrained_model_path: Path = Path(pretrained_model_name_or_path)
+
+    if pretrained_model_path.exists():
+        _LOGGER.debug(
+            "Model directory already exists locally.",
+        )
+        return pretrained_model_name_or_path
+
+    with main_process_first_context():
+        if pretrained_model_name_or_path.startswith("zoo:"):
+            _LOGGER.debug(
+                "Passed zoo stub to SparseAutoModelForCausalLM object. "
+                "Loading model from SparseZoo training files..."
+            )
+            return download_zoo_training_dir(zoo_stub=pretrained_model_name_or_path)
+
+        _LOGGER.debug("Downloading model from HuggingFace Hub.")
+        return download_repo_from_huggingface_hub(
+            repo_id=pretrained_model_name_or_path, **kwargs
+        )

tests/sparseml/pytorch/utils/test_logger.py

@@ -20,6 +20,7 @@
 import pytest
 
 from sparseml.pytorch.utils import (
+    ClearMLLogger,
     LambdaLogger,
     LoggerManager,
     PythonLogger,
@@ -45,6 +46,7 @@
             or True
         ),
         *([WANDBLogger()] if WANDBLogger.available() else []),
+        *([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}"
@@ -79,12 +81,12 @@ def test_log_scalar(self, logger):
 
     def test_log_scalars(self, logger):
         logger.log_scalars("test-scalars-tag", {"scalar1": 0.0, "scalar2": 1.0})
-        logger.log_scalars("test-scalars-tag", {"scalar1": 0.0, "scalar2": 1.0}, 1)
+        logger.log_scalars("test-scalars-tag2", {"scalar1": 0.0, "scalar2": 1.0}, 1)
         logger.log_scalars(
-            "test-scalars-tag", {"scalar1": 0.0, "scalar2": 1.0}, 2, time.time() - 1
+            "test-scalars-tag3", {"scalar1": 0.0, "scalar2": 1.0}, 2, time.time() - 1
         )
         logger.log_scalars(
-            "test-scalars-tag",
+            "test-scalars-tag4",
             {"scalar1": 0.0, "scalar2": 1.0},
             2,
             time.time() - 1,

tests/sparseml/test_clear_ml.py

@@ -0,0 +1,63 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from pathlib import Path
+
+from clearml import Task
+from sparseml.transformers import apply
+from sparseml.utils import is_package_available
+
+
+is_torch_available = is_package_available("torch")
+if is_torch_available:
+    import torch
+
+    torch_err = None
+else:
+    torch = object
+    torch_err = ModuleNotFoundError(
+        "`torch` is not installed, use `pip install torch` to log to Weights and Biases"
+    )
+
+
+def test_oneshot_and_finetune(tmp_path: Path):
+    recipe_str = "tests/sparseml/transformers/finetune/test_alternate_recipe.yaml"
+    model = "Xenova/llama2.c-stories15M"
+    device = "cuda:0"
+    if is_torch_available and not torch.cuda.is_available():
+        device = "cpu"
+    dataset = "wikitext"
+    dataset_config_name = "wikitext-2-raw-v1"
+    concatenate_data = True
+    run_stages = True
+    output_dir = tmp_path
+    max_steps = 50
+    splits = {"train": "train[:50%]", "calibration": "train[50%:60%]"}
+
+    # clearML will automatically log default capturing entries without
+    # explicitly calling logger. Logs accessible in https://app.clear.ml/
+    Task.init(project_name="test", task_name="test_oneshot_and_finetune")
+
+    apply(
+        model=model,
+        dataset=dataset,
+        dataset_config_name=dataset_config_name,
+        run_stages=run_stages,
+        output_dir=output_dir,
+        recipe=recipe_str,
+        max_steps=max_steps,
+        concatenate_data=concatenate_data,
+        splits=splits,
+        oneshot_device=device,
+    )