Fix error when context-resizing (#5860)

Harden local GGUF detection so a .gguf path is not misrouted to the transformers backend during the brief Windows lock window after llama-server is killed. Catch OSError from stat() and treat the path as the file, while a directory named *.gguf still falls through to the directory scan. Adds regression tests for the lock-window and directory cases.

studio/backend/tests/test_gguf_routing.py

@@ -0,0 +1,102 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""
+Tests for GGUF routing in detect_gguf_model.
+
+Regression test for the bug where a .gguf file temporarily appears
+inaccessible on Windows during llama-server process teardown, causing
+is_file() to return False and the model to be routed to the transformers
+backend instead of llama-server.
+"""
+
+import sys
+import os
+import types
+from pathlib import Path
+from unittest.mock import patch
+
+# Stub structlog before importing backend modules (mirrors other tests in this suite)
+if "structlog" not in sys.modules:
+
+    class _DummyLogger:
+        def __getattr__(self, _):
+            return lambda *a, **k: None
+
+    sys.modules["structlog"] = types.SimpleNamespace(
+        get_logger = lambda *a, **k: _DummyLogger(),
+        BoundLogger = _DummyLogger,
+    )
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
+
+from utils.models.model_config import detect_gguf_model
+
+
+def test_detects_gguf_file_normally(tmp_path):
+    """Normal case: .gguf file exists and is accessible."""
+    gguf = tmp_path / "gpt-oss-20b-MXFP4.gguf"
+    gguf.write_bytes(b"")
+    result = detect_gguf_model(str(gguf))
+    assert result is not None
+    assert result.endswith("gpt-oss-20b-MXFP4.gguf")
+
+
+def test_detects_gguf_when_stat_raises_oserror(tmp_path):
+    """
+    Regression: on Windows, both is_file() and exists() call stat() internally.
+    During the brief lock window after llama-server is killed, stat() raises
+    OSError, causing both to return False. detect_gguf_model must still route
+    to llama-server based on the file extension alone.
+    """
+    gguf = tmp_path / "gpt-oss-20b-MXFP4.gguf"
+    gguf.write_bytes(b"")
+
+    original_stat = Path.stat
+
+    def flaky_stat(self, **kwargs):
+        if self.suffix.lower() == ".gguf":
+            raise OSError("file temporarily inaccessible (Windows lock window)")
+        return original_stat(self, **kwargs)
+
+    with patch.object(Path, "stat", flaky_stat):
+        result = detect_gguf_model(str(gguf))
+
+    assert result is not None, (
+        "detect_gguf_model returned None when stat() raised OSError. "
+        "This causes the model to fall through to the transformers backend."
+    )
+
+
+def test_does_not_detect_mmproj_as_main_model(tmp_path):
+    """mmproj files must never be returned as the primary model."""
+    mmproj = tmp_path / "mmproj-model-f16.gguf"
+    mmproj.write_bytes(b"")
+    result = detect_gguf_model(str(mmproj))
+    assert result is None
+
+
+def test_detects_gguf_in_directory(tmp_path):
+    """Directory containing a .gguf file is resolved to that file."""
+    gguf = tmp_path / "model-Q4_K_M.gguf"
+    gguf.write_bytes(b"")
+    result = detect_gguf_model(str(tmp_path))
+    assert result is not None
+    assert result.endswith("model-Q4_K_M.gguf")
+
+
+def test_directory_named_like_gguf_scans_inside(tmp_path):
+    """A directory named *.gguf resolves the real .gguf inside, not itself."""
+    gguf_dir = tmp_path / "mymodel.gguf"
+    gguf_dir.mkdir()
+    inner = gguf_dir / "model-Q4_K_M.gguf"
+    inner.write_bytes(b"")
+    result = detect_gguf_model(str(gguf_dir))
+    assert result is not None
+    assert result.endswith("model-Q4_K_M.gguf")
+
+
+def test_returns_none_for_non_gguf_path(tmp_path):
+    """Non-.gguf paths with no .gguf files inside return None."""
+    result = detect_gguf_model(str(tmp_path))
+    assert result is None

studio/backend/utils/models/model_config.py

@@ -1156,13 +1156,18 @@ def detect_gguf_model(path: str) -> Optional[str]:
     p = Path(path)
 
     # Case 1: direct .gguf file
-    if p.suffix.lower() == ".gguf" and p.is_file():
+    if p.suffix.lower() == ".gguf":
         if _is_mmproj(p.name):
             return None
-        # Use absolute (not resolve) to preserve symlink names -- e.g.
-        # Ollama .studio_links/model.gguf -> blobs/sha256-... should
-        # keep the readable symlink name, not the opaque blob hash.
-        return str(p.absolute())
+        # Extension is authoritative: don't gate on is_file()/exists(), which
+        # can fail in the Windows lock window after llama-server is killed.
+        try:
+            is_dir = p.is_dir()
+        except OSError:
+            is_dir = False  # stat() unavailable in the lock window
+        if not is_dir:
+            return str(p.absolute())  # absolute() keeps symlink names readable
+        # Directory named "*.gguf": fall through to the dir scan below.
 
     # Case 2: directory containing .gguf files (skip mmproj)
     if p.is_dir():