Add HF_TOKEN support for accessing private/gated models like Shako-4B-it-v4

app.py

@@ -11,12 +11,18 @@ import time
 # model config
 model_12b_name = "anaspro/Shako-4B-it-v4"
 model_4b_name = "anaspro/Shako-4B-it"
+
+# Load token from environment if available
+hf_token = os.getenv("HF_TOKEN")
+
 model_12b = Gemma3ForConditionalGeneration.from_pretrained(
     model_12b_name,
     device_map="auto",
-    torch_dtype=torch.bfloat16
+    torch_dtype=torch.bfloat16,
+    token=hf_token
 ).eval()
-processor_12b = AutoProcessor.from_pretrained(model_12b_name)
+processor_12b = AutoProcessor.from_pretrained(model_12b_name, token=hf_token)
+
 model_4b = Gemma3ForConditionalGeneration.from_pretrained(
     model_4b_name,
     device_map="auto",