initial commit

app.py

@@ -79,10 +79,13 @@ def generate(
     generate_kwargs = {
         "base": {"input_ids": prompt["input_ids"], "attention_mask": prompt["attention_mask"]},
         "unit_locations": {"sources->base": (None, [[[base_unit_location]]])},
+        "max_new_tokens": max_new_tokens,
         "intervene_on_prompt": True,
         "streamer": streamer,
         "eos_token_id": tokenizer.eos_token_id,
         "early_stopping": True,
+        "do_sample": False,
+        "repetition_penalty": 1.1,
     }
 
     t = Thread(target=reft_model.generate, kwargs=generate_kwargs)