Add modal launch hf-download utility

modal/cli/launch.py

@@ -6,7 +6,7 @@
 from pathlib import Path
 from typing import Any, Optional
 
-from typer import Typer
+from typer import Argument, Option, Typer
 
 from ..app import App
 from ..exception import _CliUserExecutionError
@@ -93,3 +93,49 @@ def vscode(
         "volume": volume,
     }
     _launch_program("vscode", "vscode.py", detach, args)
+
+
+@launch_cli.command(name="hf-download")
+def hf_download(
+    repo_id: str = Argument(help="The Hugging Face repository ID"),
+    volume: str = Argument(help="The name of the Modal volume to use for caching"),
+    secret: Optional[str] = Option(None, help="The name of a Modal secret with Hugging Face credentials"),
+    timeout: int = Option(600, help="Maximum time the download is allowed to run (in seconds)"),
+    type: Optional[str] = Option(None, help="The repository type(e.g. 'model' or 'dataset')"),
+    revision: Optional[str] = Option(None, help="A specific revision to download"),
+    ignore: list[str] = Option([], help="Ignore patterns to skip downloading matching files"),
+    allow: list[str] = Option([], help="Allow patterns to selectively download matching files"),
+    detatch: bool = Option(False, "--detach", help="Allow the download to continue if the local client disconnects"),
+):
+    """Download a snapshot from the Hugging Face Hub.
+
+    This command uses Hugging Face's `snapshot_download` function to download a snapshot from a repository
+    on Hugging Face's Hub and cache it in a Modal volume. In your Modal applications, if you mount the
+    Volume at a location corresponding to the `HF_HUB_CACHE` environment variable, Hugging Face will load
+    data from the cache instead of downloading it.
+
+    \b```
+    modal launch hf-download microsoft/Phi-3-mini hf-hub-cache --secret "hf-secret" --timeout 600 --detach
+    ```
+
+    Then in your Modal App:
+
+    \b```
+    volume = modal.Volume.from_name("hf-hub-cache")
+    @app.function(volumes={HF_HUB_CACHE: volume})
+    def f():
+        model = ModelClass.from_pretrained(model_name, cache_dir=HF_HUB_CACHE)
+    ```
+
+    """
+    args = {
+        "volume": volume,
+        "secret": secret,
+        "timeout": timeout,
+        "repo_id": repo_id,
+        "type": type,
+        "revision": revision,
+        "ignore": ignore,
+        "allow": allow,
+    }
+    _launch_program("hf-download", "hf_download.py", detatch, args)

modal/cli/programs/hf_download.py

@@ -0,0 +1,43 @@
+# Copyright Modal Labs 2025
+import json
+import os
+import time
+
+import modal
+
+# Passed by `modal launch` locally via CLI, plumbed to remote runner through secrets.
+args: dict = json.loads(os.environ.get("MODAL_LAUNCH_ARGS", "{}"))
+
+CACHE_DIR = "/hf-cache"
+
+image = (
+    modal.Image.debian_slim(python_version="3.12")
+    .pip_install("huggingface-hub[hf-transfer]==0.27.1")
+    .env({"HF_HUB_CACHE": CACHE_DIR, "HF_HUB_ENABLE_HF_TRANSFER": "1"})
+)
+volume = modal.Volume.from_name(str(args.get("volume")))
+secrets = [modal.Secret.from_dict({"MODAL_LAUNCH_ARGS": json.dumps(args)})]
+if user_secret := args.get("secret"):
+    secrets.append(modal.Secret.from_name(user_secret))
+app = modal.App("hf-download", image=image, secrets=secrets, volumes={CACHE_DIR: volume})
+
+
+@app.function(cpu=4, memory=1028, timeout=int(args.get("timeout", 600)))
+def run():
+    from huggingface_hub import snapshot_download
+
+    t0 = time.monotonic()
+    snapshot_download(
+        repo_id=args.get("repo_id"),
+        repo_type=args.get("type"),
+        revision=args.get("revision"),
+        ignore_patterns=args.get("ignore", []),
+        allow_patterns=args.get("allow", []),
+        cache_dir=CACHE_DIR,
+    )
+    print(f"Completed in {time.monotonic() - t0:.2f} seconds")
+
+
+@app.local_entrypoint()
+def main():
+    run.remote()