Enhance documentation and add batch size argument for improved GPU memory management

atlas-export.py

@@ -26,6 +26,20 @@ Example usage:
         --image-column image \
         --model openai/clip-vit-base-patch32 \
         --sample 10000
+    
+    # With custom batch size for GPU memory management
+    uv run atlas-export.py \
+        large-text-dataset \
+        --space-name large-text-viz \
+        --model sentence-transformers/all-mpnet-base-v2 \
+        --batch-size 64  # Increase for faster processing on powerful GPUs
+    
+    # Small batch size for limited GPU memory
+    uv run atlas-export.py \
+        image-dataset \
+        --space-name image-viz \
+        --image-column image \
+        --batch-size 8  # Reduce to avoid OOM errors
 
     # Run on HF Jobs with GPU (requires HF token for Space deployment)
     # Get your token: python -c "from huggingface_hub import get_token; print(get_token())"
@@ -93,7 +107,7 @@ def sample_dataset_to_parquet(
     temp_dir = Path(tempfile.mkdtemp(prefix="atlas_data_"))
     parquet_path = temp_dir / "data.parquet"
 
-    logger.info(f"Saving sampled data to temporary file...")
+    logger.info("Saving sampled data to temporary file...")
     sampled_ds.to_parquet(str(parquet_path))
 
     file_size = parquet_path.stat().st_size / (1024 * 1024)  # MB
@@ -130,6 +144,9 @@ def build_atlas_command(args) -> tuple[list, str, Optional[Path]]:
     if args.model:
         cmd.extend(["--model", args.model])
 
+    if args.batch_size:
+        cmd.extend(["--batch-size", str(args.batch_size)])
+
     # Always specify text column to avoid interactive prompt
     text_col = args.text_column or "text"  # Default to "text" if not specified
     cmd.extend(["--text", text_col])
@@ -346,6 +363,12 @@ def main():
         action="store_true",
         help="Trust remote code in dataset/model",
     )
+    parser.add_argument(
+        "--batch-size",
+        type=int,
+        help="Batch size for processing embeddings (default: 32 for text, 16 for images). "
+        "Larger values use more memory but may be faster on GPUs",
+    )
 
     # Pre-computed embeddings
     parser.add_argument(
@@ -408,29 +431,23 @@ def main():
         if hf_token:
             login(token=hf_token)
             logger.info("✅ Authenticated with Hugging Face")
+        elif is_interactive := sys.stdin.isatty():
+            logger.warning(
+                "No HF token provided. You may not be able to push to the Hub."
+            )
+            response = input("Continue anyway? (y/n): ")
+            if response.lower() != "y":
+                sys.exit(0)
         else:
-            # Check if running in non-interactive environment (HF Jobs, CI, etc.)
-            is_interactive = sys.stdin.isatty()
-
-            if is_interactive:
-                logger.warning(
-                    "No HF token provided. You may not be able to push to the Hub."
-                )
-                response = input("Continue anyway? (y/n): ")
-                if response.lower() != "y":
-                    sys.exit(0)
-            else:
-                # In non-interactive environments, fail immediately if no token
-                logger.error(
-                    "No HF token found. Cannot deploy to Space in non-interactive environment."
-                )
-                logger.error(
-                    "Please set HF_TOKEN environment variable or use --hf-token argument."
-                )
-                logger.error(
-                    "Checked: HF_TOKEN, HUGGING_FACE_HUB_TOKEN, and HF CLI login"
-                )
-                sys.exit(1)
+            # In non-interactive environments, fail immediately if no token
+            logger.error(
+                "No HF token found. Cannot deploy to Space in non-interactive environment."
+            )
+            logger.error(
+                "Please set HF_TOKEN environment variable or use --hf-token argument."
+            )
+            logger.error("Checked: HF_TOKEN, HUGGING_FACE_HUB_TOKEN, and HF CLI login")
+            sys.exit(1)
 
     # Set up output directory
     if args.output_dir: