check if user has read/write access to HF_HOME and warn them if not

fa8825fae243eb3681257373f9cd781a2590c533 · 2024-11-18 20:22:04 +0400 · Alex Cheema

Files touched

M exo/download/hf/hf_helpers.py
M exo/main.py

Diff

commit fa8825fae243eb3681257373f9cd781a2590c533
Author: Alex Cheema <alexcheema123@gmail.com>
Date:   Mon Nov 18 20:22:04 2024 +0400

    check if user has read/write access to HF_HOME and warn them if not
---
 exo/download/hf/hf_helpers.py | 10 ++++++++++
 exo/main.py                   | 13 +++++++++++++
 2 files changed, 23 insertions(+)

diff --git a/exo/download/hf/hf_helpers.py b/exo/download/hf/hf_helpers.py
index e0087da0..4729e5f6 100644
--- a/exo/download/hf/hf_helpers.py
+++ b/exo/download/hf/hf_helpers.py
@@ -412,3 +412,13 @@ def get_allow_patterns(weight_map: Dict[str, str], shard: Shard) -> List[str]:
     shard_specific_patterns = set("*.safetensors")
   if DEBUG >= 2: print(f"get_allow_patterns {weight_map=} {shard=} {shard_specific_patterns=}")
   return list(default_patterns | shard_specific_patterns)
+
+async def has_hf_home_read_access() -> bool:
+  hf_home = get_hf_home()
+  try: return await aios.access(hf_home, os.R_OK)
+  except OSError: return False
+
+async def has_hf_home_write_access() -> bool:
+  hf_home = get_hf_home()
+  try: return await aios.access(hf_home, os.W_OK)
+  except OSError: return False
diff --git a/exo/main.py b/exo/main.py
index 470fb8ea..b6701c6d 100644
--- a/exo/main.py
+++ b/exo/main.py
@@ -24,6 +24,7 @@ from exo.inference.tokenizers import resolve_tokenizer
 from exo.orchestration.node import Node
 from exo.models import build_base_shard, get_repo
 from exo.viz.topology_viz import TopologyViz
+from exo.download.hf.hf_helpers import has_hf_home_read_access, has_hf_home_write_access, get_hf_home
 
 # parse args
 parser = argparse.ArgumentParser(description="Initialize GRPC Discovery")
@@ -207,6 +208,18 @@ async def run_model_cli(node: Node, inference_engine: InferenceEngine, model_nam
 async def main():
   loop = asyncio.get_running_loop()
 
+  # Check HuggingFace directory permissions
+  hf_home, has_read, has_write = get_hf_home(), await has_hf_home_read_access(), await has_hf_home_write_access()
+  if DEBUG >= 1: print(f"Model storage directory: {hf_home}")
+  print(f"{has_read=}, {has_write=}")
+  if not has_read or not has_write:
+    print(f"""
+          WARNING: Limited permissions for model storage directory: {hf_home}.
+          This may prevent model downloads from working correctly.
+          {"❌ No read access" if not has_read else ""}
+          {"❌ No write access" if not has_write else ""}
+          """)
+
   # Use a more direct approach to handle signals
   def handle_exit():
     asyncio.ensure_future(shutdown(signal.SIGTERM, loop))

← fd84201b remove redundant dummy import · back to Exo · clean branch fea1c0fc →