fix: harden upstream-protection so the Space survives upstream deletion

The audit revealed two paths where a runtime fetch could still hit
upstream sources (Wan-AI, lightx2v, github.com/enesmsahin) even though
all the weights are mirrored at
JackIsNotInTheBox/Video_Watermark_Remover_Checkpoints. Both closed:

pipeline/lama.py
- LAMA_MODEL_URL now defaults to the mirror URL directly (was None
unless the Space variable was set). If the Space variable is dropped
for any reason, the prefetch still works — no silent fallback to the
hardcoded GitHub release that simple_lama_inpainting would otherwise
reach for.

pipeline/vace.py
- New VACE_LOCAL_ONLY knob (defaults to True when VACE_PREWARM=1, which
is the default). Forces every from_pretrained / load_lora_weights call
inside _get_pipe() to pass local_files_only=True. Combined with the
prewarm step populating the cache from the mirror, this guarantees no
upstream HF Hub fetch is ever attempted at inference time — even if
the mirror itself goes offline mid-session, the cached files keep
working. If anything's missing from the cache, from_pretrained raises
loudly instead of silently downloading from upstream.

After this change the Space's runtime is fully self-contained:
✓ LaMa weights — mirror (lama/big-lama.pt)
✓ VACE-14B base — mirror (vace-14b/) loaded with local_files_only=True
✓ Distill LoRA — mirror (loras/...) loaded with local_files_only=True
✓ Prewarm fetches everything at startup from the mirror
✗ Upstream Wan-AI / lightx2v / GitHub releases — never touched
at runtime regardless of their state

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

pipeline/lama.py

@@ -36,11 +36,18 @@ from pipeline.crop import CropRegion
 # ---------------------------------------------------------------------------
 # simple-lama-inpainting hardcodes its big-lama.pt download URL to a GitHub
 # release. If that release ever vanishes, every Space cold start fails.
-# When LAMA_MODEL_URL is set, we pre-populate the file in the cache path
-# SimpleLama() will look at, so its own download logic finds it pre-existing
-# and never reaches GitHub. This closes the upstream-protection loop opened
-# by the JackIsNotInTheBox/Video_Watermark_Remover_Checkpoints mirror.
-LAMA_MODEL_URL = os.environ.get("LAMA_MODEL_URL")
+# We pre-populate the file in the cache path SimpleLama() will look at, so
+# its own download logic finds it pre-existing and never reaches GitHub.
+#
+# Default URL points at the project's HF mirror so the Space remains
+# self-contained even if the Space variable LAMA_MODEL_URL is dropped or
+# the upstream GitHub release disappears. Override via env var if you want
+# to test a different mirror.
+LAMA_MODEL_URL = os.environ.get(
+    "LAMA_MODEL_URL",
+    "https://huggingface.co/JackIsNotInTheBox/"
+    "Video_Watermark_Remover_Checkpoints/resolve/main/lama/big-lama.pt",
+)
 
 
 def _ensure_lama_weights_cached() -> None:

pipeline/vace.py

@@ -48,7 +48,10 @@ Configuration knobs (all read at module import via env vars)
 - VACE_SUBFOLDER       : subfolder within the repo (default: ``vace-14b``)
 - VACE_LORA_REPO_ID    : HF repo holding the distill LoRA (default: mirror)
 - VACE_LORA_FILE       : LoRA filename (default: lightx2v rank-64 4-step)
-- VACE_PREWARM         : "1" (default) prewarms checkpoint cache; "0" skips
+- VACE_PREWARM         : "1" (default) prewarms checkpoint cache; "0" skips.
+                         Also controls whether _get_pipe() forces
+                         ``local_files_only=True`` on its from_pretrained calls
+                         (yes when prewarmed, no when prewarm is disabled).
 
 License: Apache-2.0 (Wan2.1 base) + Apache-2.0 (lightx2v distill LoRA).
 """
@@ -85,6 +88,15 @@ VACE_LORA_FILE = os.environ.get(
     "loras/wan2.1_t2v_14b_lora_rank64_lightx2v_4step.safetensors",
 )
 
+# Force ``local_files_only=True`` on every from_pretrained / load_lora_weights
+# call inside _get_pipe(). Combined with prewarm_vace_cache() populating the
+# cache from the mirror at startup, this guarantees no upstream HF Hub fetch
+# is ever attempted at runtime — even if the mirror itself goes offline
+# mid-session, what's already cached keeps working. If you opt out of
+# prewarm (VACE_PREWARM=0) you almost certainly want to opt out of this too,
+# so the toggle defaults follow each other.
+VACE_LOCAL_ONLY = os.environ.get("VACE_PREWARM", "1") != "0"
+
 # VACE requires num_frames = 4n+1. 81 = 16*5+1 is the documented sweet spot.
 CHUNK_FRAMES = 81
 # Frames shared between consecutive chunks for temporal continuity at seams.
@@ -184,16 +196,20 @@ def _get_pipe():
     )
 
     # VAE in fp32 (per the official diffusers example) for numerical stability.
+    # local_files_only=True (default) prevents silent fallback to upstream HF
+    # if anything's missing from the prewarmed cache.
     vae = AutoencoderKLWan.from_pretrained(
         VACE_REPO_ID,
         subfolder=f"{VACE_SUBFOLDER}/vae",
         torch_dtype=torch.float32,
+        local_files_only=VACE_LOCAL_ONLY,
     )
     pipe = WanVACEPipeline.from_pretrained(
         VACE_REPO_ID,
         subfolder=VACE_SUBFOLDER,
         vae=vae,
         torch_dtype=torch.bfloat16,
+        local_files_only=VACE_LOCAL_ONLY,
     )
 
     # flow_shift = 3.0 → 480P-friendly. 5.0 would be 720P-friendly.
@@ -211,6 +227,7 @@ def _get_pipe():
             VACE_LORA_REPO_ID,
             weight_name=VACE_LORA_FILE,
             adapter_name="distill",
+            local_files_only=VACE_LOCAL_ONLY,
         )
         pipe.set_adapters(["distill"], adapter_weights=[1.0])
         pipe.fuse_lora(adapter_names=["distill"], lora_scale=1.0)