Move offline-mode patch inside each GPU function's own body

Live repro (direct HTTP call to /gradio_api/queue/join for MMAudio's
_run_mmaudio, bypassing the browser) showed the previous fix (78108d3,
patching inside _catch_oom's wrapper) still didn't stop the live ~4GB
open_clip re-download. _catch_oom's wrapper sits *outside* the raw
function — it's the code that dispatches the call, not necessarily the
code that runs inside the actual ZeroGPU worker. This codebase's own
docstrings already documented that @spaces.GPU functions run in a
genuinely separate "GPU worker process" (kwargs are silently dropped,
paths must exist in the worker's own filesystem), which points at the
same boundary.

Added _ensure_offline_in_worker() (with a diagnostic print of pid +
the resulting is_offline_mode()) and call it as literally the first
line of all six raw GPU function bodies (_taro_gpu_infer,
_mmaudio_gpu_infer, _hunyuan_gpu_infer, and the three _regen_*_gpu
variants), so it runs wherever each function's own bytecode actually
executes, regardless of how @spaces.GPU dispatches the call.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>

app.py

@@ -228,6 +228,7 @@ cavp_ckpt_path, onset_ckpt_path, taro_ckpt_path = _fut_taro.result()
 mmaudio_model_path, mmaudio_vae_path, mmaudio_synchformer_path = _fut_mmaudio.result()
 bigvgan_local_dir = _fut_bigvgan.result()
 print(f"[startup] All downloads done in {time.perf_counter() - _t_dl_start:.1f}s")
+print(f"[startup] main process pid={os.getpid()}")
 
 # Force any later HF Hub call to be cache-only. With every encoder cache
 # pre-populated above from the user's mirror, downstream from_pretrained()
@@ -252,6 +253,26 @@ os.environ["HF_HUB_OFFLINE"] = "1"
 import huggingface_hub.constants as _hf_constants
 _hf_constants.HF_HUB_OFFLINE = True
 
+def _ensure_offline_in_worker():
+    """Re-assert offline mode from inside a @spaces.GPU function's own body.
+
+    The module-level patch above (and an earlier attempt inside _catch_oom's
+    wrapper, which sits *outside* the raw function) both failed to stop
+    MMAudio's open_clip CLIP loader from re-downloading its ~4GB checkpoint
+    live during GPU inference — confirmed via a direct HTTP repro against the
+    deployed Space. This codebase's own docstrings already document that
+    @spaces.GPU functions run in a genuinely separate "GPU worker process"
+    (see e.g. _mmaudio_gpu_infer: "kwargs are silently dropped" / paths must
+    "exist in the GPU worker's process"), so any patch applied by code that
+    merely *wraps* the call — rather than running inside the function body
+    that's actually shipped to and executed by that worker — never reaches
+    it. Calling this as literally the first line of each raw GPU function
+    guarantees it runs wherever that function's own bytecode runs."""
+    os.environ["HF_HUB_OFFLINE"] = "1"
+    _hf_constants.HF_HUB_OFFLINE = True
+    print(f"[_ensure_offline_in_worker] pid={os.getpid()} HF_HUB_OFFLINE={_hf_constants.HF_HUB_OFFLINE} "
+          f"is_offline_mode()={_hf_constants.is_offline_mode()}")
+
 # ================================================================== #
 #                     SHARED CONSTANTS / HELPERS                      #
 # ================================================================== #
@@ -1109,6 +1130,7 @@ def _taro_gpu_infer(video_file, seed_val, cfg_scale, num_steps, mode,
                     crossfade_s, crossfade_db, num_samples):
     """GPU-only TARO inference — model loading + feature extraction + diffusion.
     Returns list of (wavs_list, onset_feats) per sample."""
+    _ensure_offline_in_worker()
     seed_val     = _resolve_seed(seed_val)
     crossfade_s  = float(crossfade_s)
     num_samples  = int(num_samples)
@@ -1294,6 +1316,7 @@ def _mmaudio_gpu_infer(video_file, prompt, negative_prompt, seed_val,
     tmp files that don't exist in the GPU worker's process.
     """
     print(f"[_mmaudio_gpu_infer] START video={video_file!r} silent={silent_video!r} clip_start={clip_start_s} clip_dur={clip_dur_s} num_samples={num_samples}")
+    _ensure_offline_in_worker()
     _ensure_syspath("MMAudio")
     from mmaudio.eval_utils        import generate, load_video
     from mmaudio.model.flow_matching   import FlowMatching
@@ -1465,6 +1488,7 @@ def _hunyuan_gpu_infer(video_file, prompt, negative_prompt, seed_val,
     import traceback as _tb
     print(f"[_hunyuan_gpu_infer] START video={video_file!r} model_size={model_size!r} "
           f"num_steps={num_steps!r} clip_start={clip_start_s} clip_dur={clip_dur_s}")
+    _ensure_offline_in_worker()
     try:
         _ensure_syspath("HunyuanVideo-Foley")
         from hunyuanvideo_foley.utils.model_utils   import denoise_process
@@ -1675,6 +1699,7 @@ def _regen_taro_gpu(video_file, seg_idx, seg_meta_json,
                     seed_val, cfg_scale, num_steps, mode,
                     crossfade_s, crossfade_db, slot_id=None):
     """GPU-only TARO regen — returns new_wav for a single segment."""
+    _ensure_offline_in_worker()
     meta    = json.loads(seg_meta_json)
     seg_idx = int(seg_idx)
     seg_start_s, seg_end_s = meta["segments"][seg_idx]
@@ -1753,6 +1778,7 @@ def _regen_mmaudio_gpu(video_file, seg_idx, seg_meta_json,
                        cfg_strength, num_steps, crossfade_s, crossfade_db,
                        slot_id=None):
     """GPU-only MMAudio regen — returns (new_wav, sr) for a single segment."""
+    _ensure_offline_in_worker()
     meta    = json.loads(seg_meta_json)
     seg_idx = int(seg_idx)
     seg_start, seg_end = meta["segments"][seg_idx]
@@ -1847,6 +1873,7 @@ def _regen_hunyuan_gpu(video_file, seg_idx, seg_meta_json,
                        guidance_scale, num_steps, model_size,
                        crossfade_s, crossfade_db, slot_id=None):
     """GPU-only HunyuanFoley regen — returns (new_wav, sr) for a single segment."""
+    _ensure_offline_in_worker()
     meta    = json.loads(seg_meta_json)
     seg_idx = int(seg_idx)
     seg_start, seg_end = meta["segments"][seg_idx]