fix: select writable cache root at startup

training/space_runner.py

@@ -6,20 +6,46 @@ Unsloth is pre-installed in the Docker image.
 import logging
 import os
 import sys
+import tempfile
 import threading
 import time
 import traceback
 from pathlib import Path
 
-import gradio as gr
-from dotenv import load_dotenv
 
-load_dotenv()
+def _configure_writable_runtime_dirs():
+    """HF Spaces can run with HOME=/ and /app read-only; use a verified cache root."""
+    candidates = []
+    if os.getenv("SPACE_RUNTIME_DIR"):
+        candidates.append(Path(os.environ["SPACE_RUNTIME_DIR"]))
+
+    uid = getattr(os, "getuid", lambda: "user")()
+    candidates.extend(
+        [
+            Path(tempfile.gettempdir()) / f"context-corruption-training-{uid}",
+            Path("/dev/shm") / f"context-corruption-training-{uid}",
+        ]
+    )
+
+    last_error = None
+    for runtime_root in candidates:
+        try:
+            _apply_runtime_dirs(runtime_root)
+            return runtime_root
+        except OSError as exc:
+            last_error = exc
 
+    runtime_root = Path(tempfile.mkdtemp(prefix="context-corruption-training-"))
+    try:
+        _apply_runtime_dirs(runtime_root)
+    except OSError as exc:
+        raise RuntimeError(
+            f"Could not create writable runtime directories; last error: {last_error}"
+        ) from exc
+    return runtime_root
 
-def _configure_writable_runtime_dirs():
-    """HF Spaces can run with HOME=/ and /app read-only; use /tmp for caches."""
-    runtime_root = Path(os.getenv("SPACE_RUNTIME_DIR", "/tmp/context-corruption-training"))
+
+def _apply_runtime_dirs(runtime_root: Path):
     cache_root = runtime_root / "cache"
     env_dirs = {
         "HOME": runtime_root,
@@ -32,12 +58,24 @@ def _configure_writable_runtime_dirs():
         "MPLCONFIGDIR": cache_root / "matplotlib",
         "OUTPUT_DIR": runtime_root / "checkpoints" / "grpo-qwen-1.5b",
     }
+
+    for path in env_dirs.values():
+        path.mkdir(parents=True, exist_ok=True)
+
+    probe = runtime_root / ".write-test"
+    probe.write_text("ok")
+    probe.unlink(missing_ok=True)
+
     for key, path in env_dirs.items():
         os.environ[key] = str(path)
-        path.mkdir(parents=True, exist_ok=True)
 
 
-_configure_writable_runtime_dirs()
+_runtime_root = _configure_writable_runtime_dirs()
+
+import gradio as gr
+from dotenv import load_dotenv
+
+load_dotenv()
 
 _log_lines: list[str] = []
 _training_status = "idle"  # idle | installing | running | complete | failed