Only support newest versions of lm-eval

Summary:
remove support for lm-eval 0.3 to reduce the options we have

Test Plan:
CI

Reviewers:

Subscribers:

Tasks:

Tags:

Author: jerryzh168

eval.py

@@ -29,25 +29,11 @@
 torch._inductor.config.triton.cudagraphs = True
 torch._dynamo.config.cache_size_limit = 100000
 
-try:
-    import lm_eval
+import lm_eval
 
-    lm_eval_available = True
-except:
-    lm_eval_available = False
-
-
-if lm_eval_available:
-    try:  # lm_eval version 0.4
-        from lm_eval.evaluator import evaluate
-        from lm_eval.models.huggingface import HFLM as eval_wrapper
-        from lm_eval.tasks import get_task_dict
-    except:  # lm_eval version 0.3
-        from lm_eval import base, evaluator, tasks
-
-        eval_wrapper = base.BaseLM
-        get_task_dict = tasks.get_task_dict
-        evaluate = evaluator.evaluate
+from lm_eval.evaluator import evaluate
+from lm_eval.models.huggingface import HFLM as eval_wrapper
+from lm_eval.tasks import get_task_dict
 
 
 def setup_cache_padded_seq_input_pos_max_seq_length_for_prefill(
@@ -193,11 +179,6 @@ def eval(
         model, tokenizer, max_seq_length, device=device
     )
 
-    try:
-        lm_eval.tasks.initialize_tasks()
-    except:
-        pass
-
     if "hendrycks_test" in tasks:
         tasks.remove("hendrycks_test")
         tasks += list(lm_eval.tasks.hendrycks_test.create_all_tasks().keys())