Fix cloud compatibility issues in Text Components by using fsspec and posixpath

nemo_curator/stages/text/deduplication/semantic.py

@@ -21,7 +21,8 @@
 3. Optional duplicate removal based on identified duplicates
 """
 
 import os
+import posixpath
 import time
 from dataclasses import dataclass, field
 from typing import Any, Literal
@@ -179,14 +180,14 @@
         self.cache_path = self.cache_path or self.output_path
 
         # Intermediate paths
-        self.embeddings_path = os.path.join(self.cache_path, "embeddings")
-        self.semantic_dedup_path = os.path.join(self.cache_path, "semantic_dedup")
+        self.embeddings_path = posixpath.join(self.cache_path, "embeddings")
+        self.semantic_dedup_path = posixpath.join(self.cache_path, "semantic_dedup")
         # Output paths
-        self.duplicates_path = None if self.eps is None else os.path.join(self.output_path, "duplicates")
+        self.duplicates_path = None if self.eps is None else posixpath.join(self.output_path, "duplicates")
         self.deduplicated_output_path = (
-            None if not self.perform_removal else os.path.join(self.output_path, "deduplicated")
+            None if not self.perform_removal else posixpath.join(self.output_path, "deduplicated")
         )
-        self.id_generator_state_file = os.path.join(self.output_path, "semantic_id_generator.json")
+        self.id_generator_state_file = posixpath.join(self.output_path, "semantic_id_generator.json")
 
         self._validate_config()
 

nemo_curator/stages/text/download/base/download.py

@@ -13,11 +13,13 @@
 # limitations under the License.
 
 import os
+import posixpath
 import subprocess
 from abc import ABC, abstractmethod
 from dataclasses import dataclass
 from typing import Any
 
+import fsspec
 from loguru import logger
 
 from nemo_curator.stages.base import ProcessingStage
@@ -37,7 +39,9 @@
         """
         self._download_dir = download_dir
         self._verbose = verbose
-        os.makedirs(download_dir, exist_ok=True)
+        # Use fsspec for cloud-compatible directory creation
+        fs, _ = fsspec.core.url_to_fs(download_dir)
+        fs.makedirs(download_dir, exist_ok=True)
 
     def _check_s5cmd_installed(self) -> bool:
         """Check if s5cmd is installed."""
@@ -87,14 +91,24 @@
         """
         # Generate output filename
         output_name = self._get_output_filename(url)
-        output_file = os.path.join(self._download_dir, output_name)
+        output_file = posixpath.join(self._download_dir, output_name)
         temp_file = output_file + ".tmp"
 
+        # Use fsspec for cloud-compatible file operations
+        fs, _ = fsspec.core.url_to_fs(output_file)
+
         # If final file exists and is non-empty, assume it's complete
-        if os.path.exists(output_file) and os.path.getsize(output_file) > 0:
-            if self._verbose:
-                logger.info(f"File: {output_file} exists. Not downloading")
-            return output_file
+        if fs.exists(output_file):
+            try:
+                file_info = fs.info(output_file)
+                file_size = file_info.get("size", 0)
+                if file_size > 0:
+                    if self._verbose:
+                        logger.info(f"File: {output_file} exists. Not downloading")
+                    return output_file
+            except Exception:
+                # If we can't get file info, proceed with download
+                pass
 
         # Download to temporary file
         success, error_message = self._download_to_path(url, temp_file)
@@ -103,8 +117,13 @@
             # Download successful, atomically move temp file to final location
             os.rename(temp_file, output_file)
             if self._verbose:
-                file_size = os.path.getsize(output_file)
-                logger.info(f"Successfully downloaded to {output_file} ({file_size} bytes)")
+                try:
+                    fs, _ = fsspec.core.url_to_fs(output_file)
+                    file_info = fs.info(output_file)
+                    file_size = file_info.get("size", 0)
+                    logger.info(f"Successfully downloaded to {output_file} ({file_size} bytes)")
+                except Exception:
+                    logger.info(f"Successfully downloaded to {output_file}")
             return output_file
         else:
             # Download failed

nemo_curator/stages/text/download/base/iterator.py

@@ -12,7 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
 import os
+import posixpath
 from abc import ABC, abstractmethod
 from collections.abc import Iterator
 from dataclasses import dataclass
@@ -89,8 +90,8 @@
                         if self.record_limit and record_count >= self.record_limit:
                             break
                         if self.add_filename_column:
-                            # TODO: Support cloud storage https://github.com/NVIDIA-NeMo/Curator/issues/779
-                            record_dict[self.filename_col] = os.path.basename(file_path)  # type: ignore[reportReturnType]
+                            # Use posixpath for cloud storage compatibility
+                            record_dict[self.filename_col] = posixpath.basename(file_path)  # type: ignore[reportReturnType]
                         records.append(record_dict)
                         record_count += 1
 

nemo_curator/stages/text/filters/fasttext_filter.py

@@ -12,9 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
 import os
 
 import fasttext
+import fsspec
 import numpy as np
 
 from nemo_curator.stages.text.filters.doc_filter import DocumentFilter
@@ -32,7 +33,8 @@
         self._name = "fasttext_quality_filter"
 
     def model_check_or_download(self) -> None:
-        if not os.path.exists(self._model_path):
+        fs, _ = fsspec.core.url_to_fs(self._model_path)
+        if not fs.exists(self._model_path):
             msg = f"Model file {self._model_path} not found"
             raise FileNotFoundError(msg)
 
@@ -66,7 +68,8 @@
         self._name = "lang_id"
 
     def model_check_or_download(self) -> None:
-        if not os.path.exists(self._model_path):
+        fs, _ = fsspec.core.url_to_fs(self._model_path)
+        if not fs.exists(self._model_path):
             msg = f"Model file {self._model_path} not found"
             raise FileNotFoundError(msg)
 

nemo_curator/stages/text/filters/heuristic_filter.py

@@ -16,6 +16,7 @@
 import tarfile
 from typing import Literal
 
+import fsspec
 import huggingface_hub
 import requests
 from platformdirs import user_cache_dir
@@ -789,7 +790,8 @@ def _download_histograms(self) -> None:
             raise requests.exceptions.RequestException(msg)
 
         # Open a file to write the content
-        os.makedirs(self._cache_dir, exist_ok=True)
+        fs, _ = fsspec.core.url_to_fs(self._cache_dir)
+        fs.makedirs(self._cache_dir, exist_ok=True)
         download_dest_path = os.path.join(self._cache_dir, "histograms.tar.gz")
         with open(download_dest_path, "wb") as file:
             file.write(response.content)

nemo_curator/stages/text/utils/text_utils.py

@@ -13,7 +13,8 @@
 # limitations under the License.
 
 import ast
 import os
+import posixpath
 import string
 import tokenize
 import warnings
@@ -167,7 +168,7 @@
     """Parse Python source code from file or string and print docstrings."""
     if hasattr(source, "read"):
         filename = getattr(source, "name", module)
-        module = os.path.splitext(os.path.basename(filename))[0]
+        module = posixpath.splitext(posixpath.basename(filename))[0]
         source = source.read()
 
     docstrings = sorted(parse_docstrings(source), key=lambda x: (NODE_TYPES.get(type(x[0])), x[1]))