Fix adapter loading: use set_peft_model_state_dict (plain load_state_dict silently no-opped the LoRA)

app.py

@@ -12,7 +12,7 @@ import gradio as gr
 import safetensors.torch
 from huggingface_hub import hf_hub_download
 from transformers import AutoModelForCausalLM, AutoTokenizer
-from peft import PeftModel, LoraConfig
+from peft import PeftModel, LoraConfig, set_peft_model_state_dict
 
 # ZeroGPU decorator if available; no-op fallback so the app also runs locally
 try:
@@ -58,11 +58,20 @@ _model = PeftModel(_base, _peft_config)
 
 _weights_path = hf_hub_download(ADAPTER, "adapter_model.safetensors")
 _adapter_weights = safetensors.torch.load_file(_weights_path, device="cpu")
-_model.load_state_dict(_adapter_weights, strict=False)
+
+# Use PEFT's own loader, NOT _model.load_state_dict(). The safetensors keys are
+# named `....lora_A.weight`, but the PeftModel expects the adapter name baked in
+# (`....lora_A.default.weight`). A plain load_state_dict(strict=False) matches
+# none of them and silently leaves lora_B at its zero-init — i.e. the adapter
+# has zero effect. set_peft_model_state_dict performs the key remapping.
+_load_result = set_peft_model_state_dict(_model, _adapter_weights)
+assert not _load_result.unexpected_keys, (
+    f"Adapter weights did not load: {_load_result.unexpected_keys[:5]}"
+)
 _model.eval()
 
 
-@GPU(duration=120)
+@GPU(duration=20)
 def translate(text: str) -> str:
     """Tokenize, generate, decode — the GPU-heavy work lives here."""
     if not text.strip():