optimize cold start with local cache paths and font resolution

app.py

@@ -86,10 +86,12 @@ _cached_model_dir = None
 # ============================================================
 _preloaded_embedding = None
 _preloaded_tokenizer = None
+_cached_t5_dir = None
+_cached_clip_dir = None
 
 def preload_model_files():
     """Pre-download model files to cache at startup (no GPU needed)"""
-    global _preloaded_embedding, _preloaded_tokenizer
+    global _preloaded_embedding, _preloaded_tokenizer, _cached_t5_dir, _cached_clip_dir
     from huggingface_hub import snapshot_download, hf_hub_download
     
     hf_token = os.environ.get("HF_TOKEN", None)
@@ -108,21 +110,23 @@ def preload_model_files():
     
     # 2. T5 text encoder
     try:
-        snapshot_download(
+        _cached_t5_dir = snapshot_download(
             "xlabs-ai/xflux_text_encoders",
             token=hf_token
         )
-        print("✓ T5 text encoder cached")
+        os.environ["XFLUX_TEXT_ENCODER_PATH"] = _cached_t5_dir
+        print(f"✓ T5 text encoder cached at: {_cached_t5_dir}")
     except Exception as e:
         print(f"Warning: Could not pre-download T5: {e}")
     
     # 3. CLIP text encoder
     try:
-        snapshot_download(
+        _cached_clip_dir = snapshot_download(
             "openai/clip-vit-large-patch14",
             token=hf_token
         )
-        print("✓ CLIP text encoder cached")
+        os.environ["XFLUX_CLIP_ENCODER_PATH"] = _cached_clip_dir
+        print(f"✓ CLIP text encoder cached at: {_cached_clip_dir}")
     except Exception as e:
         print(f"Warning: Could not pre-download CLIP: {e}")
     

inference.py

@@ -288,7 +288,9 @@ class CalligraphyGenerator:
         )
 
         # Font for generating condition images
-        self.font_path = self._ensure_font_exists("./FangZhengKaiTiFanTi-1.ttf")
+        project_root = os.path.dirname(os.path.abspath(__file__))
+        local_font_path = os.path.join(project_root, "FangZhengKaiTiFanTi-1.ttf")
+        self.font_path = self._ensure_font_exists(local_font_path)
         self.default_font_size = 102  # 128 * 0.8
 
     def _ensure_font_exists(self, font_path: str) -> str:

src/flux/util.py

@@ -365,11 +365,13 @@ def load_controlnet(name, device, transformer=None):
 
 def load_t5(device: str | torch.device = "cuda", max_length: int = 512) -> HFEmbedder:
     # max length 64, 128, 256 and 512 should work (if your sequence is short enough)
-    return HFEmbedder("xlabs-ai/xflux_text_encoders", max_length=max_length, torch_dtype=torch.float32).to(device)
+    t5_source = os.environ.get("XFLUX_TEXT_ENCODER_PATH", "xlabs-ai/xflux_text_encoders")
+    return HFEmbedder(t5_source, max_length=max_length, torch_dtype=torch.float32).to(device)
     # return HFEmbedder("google/mt5-base", max_length=max_length, torch_dtype=torch.float32).to(device)
 
 def load_clip(device: str | torch.device = "cuda") -> HFEmbedder:
-    return HFEmbedder("openai/clip-vit-large-patch14", max_length=77, torch_dtype=torch.float32).to(device)
+    clip_source = os.environ.get("XFLUX_CLIP_ENCODER_PATH", "openai/clip-vit-large-patch14")
+    return HFEmbedder(clip_source, max_length=77, torch_dtype=torch.float32).to(device)
 
 
 def load_ae(name: str, device: str | torch.device = "cuda", hf_download: bool = True) -> AutoEncoder: