Fix decoder-only padding and completion slicing

evaluate_iapo_gsm8k.py

@@ -102,6 +102,7 @@ def load_model(spec: dict[str, str | None], dtype: str):
     tokenizer = AutoTokenizer.from_pretrained(spec["repo"], **kwargs)
     if tokenizer.pad_token_id is None:
         tokenizer.pad_token = tokenizer.eos_token
+    tokenizer.padding_side = "left"
     model = AutoModelForCausalLM.from_pretrained(
         spec["repo"],
         torch_dtype=torch_dtype,
@@ -186,8 +187,9 @@ def main() -> None:
                     )
                 gen_seconds += time.perf_counter() - t0
                 prompt_lens = encoded["attention_mask"].sum(dim=1).tolist()
+                input_width = int(encoded["input_ids"].shape[1])
                 expanded_prompt_lens = [
-                    prompt_lens[i]
+                    input_width
                     for i in range(len(prompt_lens))
                     for _ in range(args.num_return_sequences)
                 ]