Bisect: base+adapter vs merged, to locate the escaped-newline defect

bisect_adapter.py

@@ -0,0 +1,79 @@
+# /// script
+# requires-python = ">=3.10"
+# dependencies = [
+#     "torch",
+#     "transformers[torch]",
+#     "peft",
+#     "huggingface_hub",
+#     "bitsandbytes",
+#     "accelerate",
+# ]
+# ///
+"""Bisect: base+adapter (no merge) vs merged checkpoint.
+
+If base+adapter emits real newlines, the MERGE is what breaks decoding.
+If it also emits literal backslash-n, the ADAPTER (i.e. the training) is at fault.
+"""
+import os
+import re
+import sys
+
+import torch
+from peft import PeftModel
+from transformers import (AutoModelForCausalLM, AutoTokenizer,
+                          BitsAndBytesConfig)
+
+BASE = "unsloth/Qwen3-Coder-30B-A3B-Instruct"
+ADAPTER = "Taimwe/securecoder-30b-pro-v2"
+
+PROMPTS = [
+    "Write a Python function `def is_palindrome(s: str) -> bool:` that checks whether a string reads the same forwards and backwards, ignoring case and non-alphanumeric characters.",
+    "Write a Python function `def two_sum(nums: list[int], target: int) -> list[int]:` returning the indices of the two numbers adding to target, or an empty list.",
+    "Write a Python function `def flatten(nested: list) -> list:` that flattens an arbitrarily nested list.",
+]
+
+token = os.environ.get("HF_TOKEN")
+tok = AutoTokenizer.from_pretrained(BASE, token=token)
+bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
+                         bnb_4bit_compute_dtype=torch.bfloat16,
+                         bnb_4bit_use_double_quant=True)
+model = AutoModelForCausalLM.from_pretrained(
+    BASE, token=token, quantization_config=bnb,
+    dtype=torch.bfloat16, device_map="auto")
+model = PeftModel.from_pretrained(model, ADAPTER, token=token)
+model.eval()
+
+ok = 0
+for i, p in enumerate(PROMPTS, 1):
+    text = tok.apply_chat_template([{"role": "user", "content": p}],
+                                   tokenize=False, add_generation_prompt=True)
+    ids = tok(text, return_tensors="pt", add_special_tokens=False).input_ids.to(model.device)
+    with torch.no_grad():
+        gen = model.generate(ids, max_new_tokens=1024, do_sample=False,
+                             eos_token_id=tok.eos_token_id, pad_token_id=tok.eos_token_id)
+    new_ids = gen[0, ids.shape[-1]:]
+    reply = tok.decode(new_ids, skip_special_tokens=True)
+
+    has_real_nl = "\n" in reply
+    has_lit = "\\n" in reply
+    blocks = re.findall(r"```(?:python)?\s*\n(.*?)```", reply, re.S)
+    block = max(blocks, key=len) if blocks else None
+    if block is None:
+        s = reply.find("def ")
+        block = reply[s:] if s >= 0 else None
+    err = None
+    if block:
+        try:
+            compile(block, "<t>", "exec")
+        except Exception as exc:
+            err = "%s: %s" % (type(exc).__name__, exc)
+    if not err:
+        ok += 1
+    print("\n" + "=" * 72)
+    print("PROMPT %d  new_tokens=%d  fenced=%d  real_nl=%s  literal_backslash_n=%s"
+          % (i, len(new_ids), len(blocks), has_real_nl, has_lit))
+    print("COMPILE:", err or "OK")
+    print("-" * 72)
+    print(reply[:1200])
+
+print("\n\nSUMMARY base+adapter: %d/%d compiled cleanly" % (ok, len(PROMPTS)))