add pad_to_buckets in evaluation for hpu performance (#2011)

* add pad_to_buckets in evaluation for hpu performance
---------

Signed-off-by: xin3he <[email protected]>

examples/3.x_api/pytorch/nlp/huggingface_models/language-modeling/quantization/weight_only/run_clm_no_trainer.py

@@ -12,6 +12,7 @@
 from torch.nn.functional import pad
 from torch.utils.data import DataLoader
 from transformers import AutoModelForCausalLM, AutoConfig, AutoTokenizer
+from neural_compressor.torch.utils import is_hpex_available
 
 parser = argparse.ArgumentParser()
 parser.add_argument(
@@ -324,22 +325,26 @@ def run_fn_for_gptq(model, dataloader_for_calibration, *args):
     user_model, _ = get_user_model()
     tokenizer = AutoTokenizer.from_pretrained(args.model)
     config = AutoConfig.from_pretrained(args.model)
-    user_model = load(os.path.abspath(os.path.expanduser(args.output_dir)), user_model)
+    user_model = load(
+        os.path.abspath(os.path.expanduser(args.output_dir)),
+        user_model,
+        device="hpu" if is_hpex_available() else "cpu",
+    )
     setattr(user_model, "config", config)
 else:
     user_model, tokenizer = get_user_model()
 
 
 if args.accuracy:
     user_model.eval()
-    from intel_extension_for_transformers.transformers.llm.evaluation.lm_eval import evaluate, LMEvalParser
+    from neural_compressor.evaluation.lm_eval import evaluate, LMEvalParser
     eval_args = LMEvalParser(
         model="hf",
         user_model=user_model,
         tokenizer=tokenizer,
         batch_size=args.batch_size,
         tasks=args.tasks,
-        device="cpu",
+        device="hpu" if is_hpex_available() else "cpu",
     )
     results = evaluate(eval_args)
     for task_name in args.tasks.split(","):
@@ -352,7 +357,7 @@ def run_fn_for_gptq(model, dataloader_for_calibration, *args):
 
 if args.performance:
     user_model.eval()
-    from intel_extension_for_transformers.transformers.llm.evaluation.lm_eval import evaluate, LMEvalParser
+    from neural_compressor.evaluation.lm_eval import evaluate, LMEvalParser
     import time
 
     samples = args.iters * args.batch_size
@@ -363,7 +368,7 @@ def run_fn_for_gptq(model, dataloader_for_calibration, *args):
         batch_size=args.batch_size,
         tasks=args.tasks,
         limit=samples,
-        device="cpu",
+        device="hpu" if is_hpex_available() else "cpu",
     )
     start = time.time()
     results = evaluate(eval_args)

neural_compressor/evaluation/lm_eval/accuracy.py

@@ -36,18 +36,26 @@
 from pathlib import Path
 from typing import Union
 
+import lm_eval
 import numpy as np
-from lm_eval import utils
+from lm_eval import evaluator, utils
 from lm_eval.loggers import WandbLogger
 from lm_eval.tasks import TaskManager
 from lm_eval.utils import make_table, simple_parse_args_string
 
-from neural_compressor.evaluation.lm_eval import evaluator
-from neural_compressor.evaluation.lm_eval.evaluator import request_caching_arg_to_dict
-
 DEFAULT_RESULTS_FILE = "results.json"
 
 
+def request_caching_arg_to_dict(cache_requests: str) -> dict:
+    request_caching_args = {
+        "cache_requests": cache_requests in {"true", "refresh"},
+        "rewrite_requests_cache": cache_requests == "refresh",
+        "delete_requests_cache": cache_requests == "delete",
+    }
+
+    return request_caching_args
+
+
 def _handle_non_serializable(o):
     if isinstance(o, np.int64) or isinstance(o, np.int32):
         return int(o)
@@ -143,8 +151,57 @@ def cli_evaluate(args) -> None:
 
     request_caching_args = request_caching_arg_to_dict(cache_requests=args.cache_requests)
 
+    ### update model with user_model ###
+    if args.model_args is None:
+        args.model_args = ""
+    # replace HFLM.
+    from .models.huggingface import HFLM
+
+    lm_eval.api.registry.MODEL_REGISTRY["hf-auto"] = HFLM
+    lm_eval.api.registry.MODEL_REGISTRY["hf"] = HFLM
+    lm_eval.api.registry.MODEL_REGISTRY["huggingface"] = HFLM
+
+    if args.user_model is not None:
+        # use tiny model to built lm.
+        print(
+            "We use 'pretrained=Muennighoff/tiny-random-bert'"
+            + "to build `LM` instance, the actually run model is user_model you passed."
+        )
+        lm = lm_eval.api.registry.get_model(args.model).create_from_arg_string(
+            "pretrained=Muennighoff/tiny-random-bert",
+            {
+                "batch_size": args.batch_size,
+                "max_batch_size": args.max_batch_size,
+                "device": args.device,
+            },
+        )
+        lm._model = args.user_model
+        if args.tokenizer is not None:
+            lm.tokenizer = args.tokenizer
+        else:
+            assert False, "Please provide tokenizer in evaluation function"
+    elif isinstance(args.model_args, dict):
+        lm = lm_eval.api.registry.get_model(args.model).create_from_arg_obj(
+            args.model_args,
+            {
+                "batch_size": args.batch_size,
+                "max_batch_size": args.max_batch_size,
+                "device": args.device,
+            },
+        )
+    else:
+        lm = lm_eval.api.registry.get_model(args.model).create_from_arg_string(
+            args.model_args,
+            {
+                "batch_size": args.batch_size,
+                "max_batch_size": args.max_batch_size,
+                "device": args.device,
+            },
+        )
+    lm.pad_to_buckets = args.pad_to_buckets
+
     results = evaluator.simple_evaluate(
-        model=args.model,
+        model=lm,
         model_args=args.model_args,
         tasks=task_names,
         num_fewshot=args.num_fewshot,
@@ -163,8 +220,6 @@ def cli_evaluate(args) -> None:
         random_seed=args.seed[0],
         numpy_random_seed=args.seed[1],
         torch_random_seed=args.seed[2],
-        user_model=args.user_model,  # to validate the model in memory,
-        tokenizer=args.tokenizer,  # to use tokenizer in mem,
         **request_caching_args,
     )