support left padding and prefix post-processing for models like chatglm

bigcode_eval/utils.py

@@ -268,7 +268,7 @@ def complete_code(
                     batch["input_len"].max().item()
                 )
 
-            inputs = batch["ids"][:, : batch["input_len"]]
+            inputs = batch["ids"][:, : batch["input_len"]] if tokenizer.padding_side == "right" else batch["ids"]
             if "ids_encoder" in batch:
                 if is_wrapped:
                     generated_tokens = accelerator.unwrap_model(model).generate(
@@ -365,7 +365,7 @@ def update_code_gens(
     postprocess,
     code_gens,
     gen_token_dict,
-):
+):  
     for sample, generated_tokens in gen_token_dict.items():
         for s in generated_tokens:
             if INFILL_MODE or tokenizer.eos_token in task.stop_words:
@@ -378,6 +378,13 @@ def update_code_gens(
                 gen_code = tokenizer.decode(
                     s, skip_special_tokens=False, clean_up_tokenization_spaces=False
                 )
+                try:
+                    # some tokenizers add a multi-token prefix to the generation (e.g ChatGLM)
+                    tokenizer_prefix = tokenizer.decode(tokenizer.get_prefix_tokens())
+                    if gen_code.startswith(f"{tokenizer_prefix}"):
+                        gen_code = gen_code[len(tokenizer_prefix):].lstrip()
+                except:
+                    pass
                 if INFILL_MODE:
                     gen_code = _parse_infill(gen_code, tokenizer)
                 if INSTRUCTION_MODE:

main.py

@@ -109,6 +109,11 @@ def parse_args():
         action="store_true",
         help="Load model in 4bit",
     )
+    parser.add_argument(
+        "--left_padding",
+        action="store_true",
+        help="Force left padding, needed for models like chatglm3-6b",
+    )
     parser.add_argument(
         "--limit",
         type=int,
@@ -311,14 +316,25 @@ def main():
             model.merge_and_unload()
             print("Merge complete.")
 
-        tokenizer = AutoTokenizer.from_pretrained(
-            args.model,
-            revision=args.revision,
-            trust_remote_code=args.trust_remote_code,
-            use_auth_token=args.use_auth_token,
-            truncation_side="left",
-            padding_side="right",  # padding on the right is needed to cut off padding in `complete_code`
-        )
+        if args.left_padding:
+            # left padding is required for some models like chatglm3-6b
+            tokenizer = AutoTokenizer.from_pretrained(
+                args.model,
+                revision=args.revision,
+                trust_remote_code=args.trust_remote_code,
+                use_auth_token=args.use_auth_token,
+                padding_side="left",  
+            )
+        else:
+            # used by default for most models
+            tokenizer = AutoTokenizer.from_pretrained(
+                args.model,
+                revision=args.revision,
+                trust_remote_code=args.trust_remote_code,
+                use_auth_token=args.use_auth_token,
+                truncation_side="left",
+                padding_side="right",  
+            )
         if not tokenizer.eos_token:
             if tokenizer.bos_token:
                 tokenizer.eos_token = tokenizer.bos_token

tests/test_generation_evaluation.py

@@ -45,6 +45,7 @@ def update_args(args):
     args.limit_start = 0
     args.batch_size = 1
     args.max_length_generation = 300
+    args.left_padding = False
     args.do_sample = False
     args.top_p = 0
     args.n_samples = 1