Fix HfFolder shim ordering + cache mmcv wheel on the Hub

- spaces itself imports HfFolder, and we moved `import spaces` to the top, so
the shim must run before it. Move the HfFolder shim above `import spaces`
(it only needs huggingface_hub, no CUDA).
- Cache the compiled mmcv-full wheel to the Hub (keyed by torch/cuda/python)
and reuse it on later cold starts, so the ~9 min compile happens only once.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

app.py

@@ -9,6 +9,32 @@ import subprocess
 import tempfile
 import traceback
 
+# ── HfFolder shim (must precede `import spaces` and `import gradio`) ──────────
+# gradio 4.44's oauth.py AND the `spaces` package both do
+# `from huggingface_hub import HfFolder`, but the container ships
+# huggingface_hub 1.x which removed HfFolder. Inject a minimal stand-in. This
+# only imports huggingface_hub (no CUDA), so it is safe to run before `spaces`.
+import huggingface_hub as _hfh
+if not hasattr(_hfh, "HfFolder"):
+    class _HfFolderShim:
+        path_token = None
+
+        @staticmethod
+        def get_token():
+            return (os.environ.get("HF_TOKEN")
+                    or os.environ.get("HUGGING_FACE_HUB_TOKEN"))
+
+        @classmethod
+        def save_token(cls, token):
+            pass
+
+        @classmethod
+        def delete_token(cls):
+            pass
+
+    _hfh.HfFolder = _HfFolderShim
+del _hfh
+
 # ZeroGPU: `spaces` MUST be imported before torch / any CUDA-related package,
 # otherwise it raises "CUDA has been initialized before importing the spaces
 # package". The mmcv bootstrap below imports torch, so import spaces first.
@@ -26,48 +52,87 @@ else:
 sys.path.insert(0, REPO)
 sys.path.insert(0, os.path.join(REPO, "demo"))  # for `import configs` inside helpers.py
 
-# ── ensure mmcv-full (with CUDA ops) is importable; build at runtime if not ──
-# HF's build phase can't compile mmcv (isolated pip, no Space Variables). We do
-# it here, where Variables are present and we control pip flags. Diagnostics are
-# printed so the container "run" logs reveal the toolchain state.
+# ── ensure mmcv-full (with CUDA ops) is importable ───────────────────────────
+# HF's build phase can't compile mmcv (isolated pip, no Space Variables), so we
+# handle it at runtime. Compiling from source takes ~9 min, so we cache the
+# built wheel on the Hub keyed by torch/cuda/python and reuse it on later cold
+# starts (download ~1 min). MMCV_CACHE_REPO must allow writes via HF_TOKEN.
+MMCV_GIT = "git+https://github.com/open-mmlab/mmcv.git@v1.7.2"
+MMCV_CACHE_REPO = os.environ.get("MMCV_CACHE_REPO", "ahmaddarkhalil/hoi-detr")
+
+
 def _ensure_mmcv():
     import shutil
+    import glob
+
     def run(cmd, env=None, check=True):
         print("[bootstrap] $", " ".join(cmd), flush=True)
         return subprocess.run(cmd, env=env, check=check)
 
-    try:
-        import mmcv
-        from mmcv.ops import RoIAlign  # noqa: F401 — proves CUDA ops compiled
-        print(f"[bootstrap] mmcv {mmcv.__version__} (with ops) already present",
-              flush=True)
+    def have_mmcv():
+        try:
+            import mmcv  # noqa: F811
+            from mmcv.ops import RoIAlign  # noqa: F401 — proves CUDA ops present
+            print(f"[bootstrap] mmcv {mmcv.__version__} (with ops) ready",
+                  flush=True)
+            return True
+        except Exception:
+            return False
+
+    if have_mmcv():
         return
-    except Exception as e:
-        print(f"[bootstrap] mmcv/ops unavailable ({e!r}); building from source",
-              flush=True)
 
     import torch
-    print(f"[bootstrap] torch={torch.__version__} "
-          f"torch.version.cuda={torch.version.cuda}", flush=True)
-    nvcc = shutil.which("nvcc")
-    print(f"[bootstrap] system nvcc: {nvcc}", flush=True)
-    run(["bash", "-lc", "nvcc --version || true"], check=False)
+    tver = torch.__version__
+    cuver = (torch.version.cuda or "none")
+    pytag = f"cp{sys.version_info.major}{sys.version_info.minor}"
+    wheel_name = (f"mmcv_full-1.7.2-torch{tver.replace('+', '_')}-"
+                  f"cu{cuver.replace('.', '')}-{pytag}-linux_x86_64.whl")
+    token = os.environ.get("HF_TOKEN")
+    print(f"[bootstrap] torch={tver} cuda={cuver}; cache wheel={wheel_name}",
+          flush=True)
+
+    # 1) Try a cached prebuilt wheel from the Hub.
+    try:
+        from huggingface_hub import hf_hub_download
+        whl = hf_hub_download(repo_id=MMCV_CACHE_REPO, filename=wheel_name,
+                              token=token)
+        run([sys.executable, "-m", "pip", "install", whl])
+        if have_mmcv():
+            print("[bootstrap] installed cached mmcv wheel", flush=True)
+            return
+    except Exception as e:
+        print(f"[bootstrap] no usable cached wheel ({e!r}); building", flush=True)
+
+    # 2) Build from source.
+    print(f"[bootstrap] system nvcc: {shutil.which('nvcc')}", flush=True)
     run(["bash", "-lc", "gcc --version | head -1 || true"], check=False)
     print(f"[bootstrap] CUDA_HOME={os.environ.get('CUDA_HOME')}", flush=True)
-
     env = dict(os.environ)
     env["MMCV_WITH_OPS"] = "1"
     env["FORCE_CUDA"] = "1"
-    # RTX Pro 6000 Blackwell == sm_120; +PTX for forward-compat JIT.
-    env.setdefault("TORCH_CUDA_ARCH_LIST", "12.0+PTX")
+    env.setdefault("TORCH_CUDA_ARCH_LIST", "12.0+PTX")  # RTX Pro 6000 = sm_120
     env.setdefault("MAX_JOBS", "4")
-    # Bare git URL (no "name @"): mmcv builds as "mmcv-full" with ops, and the
-    # name@ form makes pip reject it for an inconsistent project name.
-    run([sys.executable, "-m", "pip", "install", "--no-build-isolation", "-v",
-         "git+https://github.com/open-mmlab/mmcv.git@v1.7.2"], env=env)
-
-    import mmcv  # noqa: F811
-    print(f"[bootstrap] built mmcv {mmcv.__version__}", flush=True)
+    outdir = "/tmp/mmcv_wheel"
+    run([sys.executable, "-m", "pip", "wheel", "--no-build-isolation",
+         "--no-deps", "-w", outdir, MMCV_GIT], env=env)
+    built = sorted(glob.glob(os.path.join(outdir, "mmcv_full-*.whl")))[0]
+    run([sys.executable, "-m", "pip", "install", built])  # also pulls addict
+    if not have_mmcv():
+        raise RuntimeError("mmcv built but import still fails")
+    print("[bootstrap] built mmcv from source", flush=True)
+
+    # 3) Cache the wheel for future cold starts (best-effort).
+    try:
+        if token:
+            from huggingface_hub import upload_file
+            upload_file(path_or_fileobj=built, path_in_repo=wheel_name,
+                        repo_id=MMCV_CACHE_REPO, token=token,
+                        commit_message="cache mmcv-full wheel")
+            print(f"[bootstrap] cached wheel -> {MMCV_CACHE_REPO}/{wheel_name}",
+                  flush=True)
+    except Exception as e:
+        print(f"[bootstrap] wheel cache upload skipped ({e!r})", flush=True)
 
 
 _ensure_mmcv()
@@ -84,34 +149,6 @@ del _mmcv_mod, _real_mmcv_ver
 
 import mmcv
 import numpy as np
-
-# ── HfFolder shim ─────────────────────────────────────────────────────────────
-# gradio 4.44 (pinned by the Space sdk_version, with the [oauth] extra forced by
-# HF) does `from huggingface_hub import HfFolder`, but the container ships
-# huggingface_hub 1.x which removed HfFolder. Restore a minimal stand-in so the
-# gradio.oauth import succeeds. We don't use OAuth, so behaviour is irrelevant.
-import huggingface_hub as _hfh
-if not hasattr(_hfh, "HfFolder"):
-    class _HfFolderShim:
-        path_token = None
-
-        @staticmethod
-        def get_token():
-            return (os.environ.get("HF_TOKEN")
-                    or os.environ.get("HUGGING_FACE_HUB_TOKEN"))
-
-        @classmethod
-        def save_token(cls, token):
-            pass
-
-        @classmethod
-        def delete_token(cls):
-            pass
-
-    _hfh.HfFolder = _HfFolderShim
-del _hfh
-# ─────────────────────────────────────────────────────────────────────────────
-
 import gradio as gr
 from huggingface_hub import hf_hub_download