Switch text encoder to HF Hub download instead of bucket mount

Uses hf_hub_download from ponpoke/flux2-klein-9b-uncensored-text-encoder
for faster download via CDN + automatic HF caching.

Co-Authored-By: Oz <oz-agent@warp.dev>

app.py

@@ -19,7 +19,8 @@ from config import (
     MAX_SEED,
     MAX_LORA_SLOTS,
     PERSISTENT_LORA_CATALOG_PATH,
-    UNCENSORED_TE_PATH,
+    UNCENSORED_TE_REPO,
+    UNCENSORED_TE_FILE,
 )
 from ui_theme import orange_red_theme
 from upscale import UPSCALE_MODELS, apply_realesrgan
@@ -326,15 +327,18 @@ pipe = _PipeClass.from_pretrained(MODEL_REPO, torch_dtype=torch.bfloat16).to(dev
 print(f"Model loaded successfully: FLUX.2 Klein {MODEL_VARIANT}")
 
 # ── Replace text encoder with abliterated (uncensored) version ───────────
-if os.path.isfile(UNCENSORED_TE_PATH):
+try:
+    from huggingface_hub import hf_hub_download
     from safetensors.torch import load_file
-    print(f"Loading abliterated text encoder from {UNCENSORED_TE_PATH}...")
-    state_dict = load_file(UNCENSORED_TE_PATH)
+    print(f"Downloading abliterated text encoder from {UNCENSORED_TE_REPO}...")
+    te_path = hf_hub_download(repo_id=UNCENSORED_TE_REPO, filename=UNCENSORED_TE_FILE)
+    print(f"Loading abliterated weights from {te_path}...")
+    state_dict = load_file(te_path)
     pipe.text_encoder.load_state_dict(state_dict, strict=True)
     pipe.text_encoder.to(device=device, dtype=torch.bfloat16)
     print("Abliterated text encoder loaded — safety filters removed.")
-else:
-    print(f"Abliterated text encoder not found at {UNCENSORED_TE_PATH} — using stock encoder.")
+except Exception as e:
+    print(f"Abliterated text encoder unavailable ({e}) — using stock encoder.")
 
 
 # ── UI helper callbacks ──────────────────────────────────────────────────────

config.py

@@ -31,8 +31,6 @@ PERSISTENT_LORA_CATALOG_PATH = os.environ.get(
 )
 
 # Abliterated (uncensored) text encoder — replaces the stock Qwen3 text encoder
-# weights after pipeline load. Set UNCENSORED_TE_PATH env var to override.
-UNCENSORED_TE_PATH = os.environ.get(
-    "UNCENSORED_TE_PATH",
-    "/loras-flux/fk9b-uncensored-text-encoder/model.safetensors",
-)
+# weights after pipeline load. Downloaded via HF Hub for speed + caching.
+UNCENSORED_TE_REPO = "ponpoke/flux2-klein-9b-uncensored-text-encoder"
+UNCENSORED_TE_FILE = "model.safetensors"