fix: restrict cache validation behaviors to CI only (#14849)

python/sglang/srt/model_loader/weight_utils.py

@@ -47,6 +47,7 @@ from sglang.srt.model_loader.weight_validation import (
     _validate_sharded_model,
 )
 from sglang.srt.utils import find_local_repo_dir, log_info_on_rank0, print_warning_once
+from sglang.utils import is_in_ci
 
 logger = logging.getLogger(__name__)
 
@@ -326,8 +327,13 @@ def _find_local_hf_snapshot_dir_unlocked(
     if not os.path.isdir(found_local_snapshot_dir):
         return None
 
+    # Only perform cache validation and cleanup in CI to avoid
+    # unnecessary overhead for regular users
+    if is_in_ci():
         # Check for incomplete files and clean up if found
-    repo_folder = os.path.abspath(os.path.join(found_local_snapshot_dir, "..", ".."))
+        repo_folder = os.path.abspath(
+            os.path.join(found_local_snapshot_dir, "..", "..")
+        )
         blobs_dir = os.path.join(repo_folder, "blobs")
 
         # Check for incomplete download markers
@@ -366,6 +372,8 @@ def _find_local_hf_snapshot_dir_unlocked(
         )
         local_weight_files = []
 
+    # Only perform cache validation and cleanup in CI
+    if is_in_ci():
         # Validate sharded models and check for corruption
         if local_weight_files:
             is_valid, error_msg, corrupted_files = _validate_sharded_model(
@@ -380,7 +388,9 @@ def _find_local_hf_snapshot_dir_unlocked(
                         f"{model_name_or_path}: {error_msg}. "
                         "Will selectively clean and re-download only these files.",
                     )
-                _cleanup_corrupted_files_selective(model_name_or_path, corrupted_files)
+                    _cleanup_corrupted_files_selective(
+                        model_name_or_path, corrupted_files
+                    )
                     return None
                 else:
                     # Cannot selectively clean (e.g., missing shards) - remove entire cache
@@ -569,6 +579,8 @@ def download_weights_from_hf(
 
         log_info_on_rank0(logger, f"Using model weights format {allow_patterns}")
 
+        # Only perform validation and retry in CI to avoid overhead for regular users
+        if is_in_ci():
             # Retry loop for handling corrupted downloads
             for attempt in range(max_retries):
                 hf_folder = snapshot_download(
@@ -606,6 +618,18 @@ def download_weights_from_hf(
 
             # This should never be reached, but just in case
             return hf_folder
+        else:
+            # Simple download without validation for non-CI environments
+            hf_folder = snapshot_download(
+                model_name_or_path,
+                allow_patterns=allow_patterns,
+                ignore_patterns=ignore_patterns,
+                cache_dir=cache_dir,
+                tqdm_class=DisabledTqdm,
+                revision=revision,
+                local_files_only=huggingface_hub.constants.HF_HUB_OFFLINE,
+            )
+            return hf_folder
 
 
 def download_safetensors_index_file_from_hf(