moved eval import and added to docs (#1139)

configs/neox_arguments.md

@@ -976,7 +976,7 @@ Text Generation arguments
 
 - **prompt_end**: str
 
- Default = 
+ Default =
 
 
  a single prompt's end. Defaults to newline
@@ -1018,7 +1018,7 @@ Text Generation arguments
 
 - **eval_results_prefix**: str
 
- Default = 
+ Default =
 
  prefix to which to save evaluation results - final fp will be {eval_results_prefix}_eval_results_yy-mm-dd-HH-MM.json
 
@@ -1030,6 +1030,8 @@ Text Generation arguments
 
  Tasks to evaluate on using lm_eval_harness
 
+ NOTE: Requires internet connection
+
 
 
 ## NeoXArgsTokenizer
@@ -1768,7 +1770,7 @@ Args for deepspeed config
 
  Default = None
 
- 
+
 
 
 
@@ -2068,4 +2070,3 @@ Args for deepspeed runner (deepspeed.launcher.runner).
  Default = None
 
  Adds a `--account` to the DeepSpeed launch command. In DeeperSpeed this is passed on to the SlurmLauncher as well. Sometimes necessary for cluster rules, or so I've heard.
-

megatron/training.py

@@ -55,7 +55,6 @@
  CharCounter,
 )
 from megatron.model.gpt2_model import cross_entropy
-from eval_tasks import run_eval_harness
 
 
 def mup_weights_reinit(neox_args, model):
@@ -967,6 +966,8 @@ def evaluate(
  )
 
  if neox_args.eval_tasks:
+ from eval_tasks import run_eval_harness
+
  eval_results.update(
  run_eval_harness(
  model, forward_step_fn, neox_args, eval_tasks=neox_args.eval_tasks