Eval v4: --merged-repo, eos_token_id, deps without unsloth

eval_securecoder.py

@@ -71,8 +71,9 @@ TOOL_PARAM_PAT = re.compile(r"<parameter=([A-Za-z0-9_]+)>\s*(.*?)\s*</parameter>
 
 def parse_args() -> argparse.Namespace:
     p = argparse.ArgumentParser(description="SecureCoder evaluations")
-    p.add_argument("--adapter", default="Taimwe/securecoder-30b-pro")
-    p.add_argument("--base", default="unsloth/Qwen3-Coder-30B-A3B-Instruct")
+    p.add_argument("--adapter", default="Taimwe/securecoder-30b-pro", help="(unused when --merged-repo is set)")
+    p.add_argument("--base", default="unsloth/Qwen3-Coder-30B-A3B-Instruct", help="(unused when --merged-repo is set)")
+    p.add_argument("--merged-repo", default="Taimwe/securecoder-30b-pro-merged", help="merged safetensors repo")
     p.add_argument("--out-dir", default="/data/eval-out")
     p.add_argument("--tool-prompts", type=int, default=80)
     p.add_argument("--code-prompts", type=int, default=15)
@@ -372,16 +373,15 @@ def main() -> int:
         log.error("HF_TOKEN not set"); return 1
     os.makedirs(args.out_dir, exist_ok=True)
     import torch
-    from transformers import AutoTokenizer
-    from peft import PeftModel
-    from unsloth import FastLanguageModel
+    from transformers import AutoModelForCausalLM, AutoTokenizer
 
     random.seed(args.seed)
-    log.info("loading adapter %s on top of %s ...", args.adapter, args.base)
-    model, tokenizer = FastLanguageModel.from_pretrained(
-        model_name=args.base, max_seq_length=2048, dtype=torch.bfloat16, load_in_4bit=True)
-    model = PeftModel.from_pretrained(model, args.adapter, token=token)
-    log.info("adapter loaded")
+    log.info("loading merged model %s in 4-bit ...", args.merged_repo)
+    tokenizer = AutoTokenizer.from_pretrained(args.merged_repo, token=token)
+    model = AutoModelForCausalLM.from_pretrained(
+        args.merged_repo, token=token, dtype=torch.bfloat16, load_in_4bit=True,
+        device_map="auto")
+    log.info("model loaded")
 
     sections = [
         eval_tool_calls(tokenizer, model, args),