Remove consolidate_results.py

language/llama3-405b/README.md

@@ -152,17 +152,6 @@ if [ -e ${ACCURACY_LOG_FILE} ]; then
         python evaluate-accuracy.py --checkpoint-path ${CHECKPOINT_PATH} \
                 --mlperf-accuracy-file ${ACCURACY_LOG_FILE} --dataset-file ${DATASET_PATH} --dtype int32
 fi
-
-# Optional: Create a pickled pandas DataFrame that is the original dataset with extra columns with output data from the
-# accuracy run. The following columns will be added:
-# - "gen_output_tok_id": A list of ints representing the tokenized output sequence.
-# - "gen_output_text": A str representing the untokenized output sequence.
-# - "gen_output_tok_len": An int representing the number of output tokens.
-# - "rouge1": The rouge1 score for this sample
-# - "rouge2": The rouge2 score for this sample
-# - "rougeL": The rougeL score for this sample
-# This file will by default be saved to 'full_output.pkl'. You can modify this with --output-pkl-path.
-python consolidate_results.py --dataset-path ${DATASET_PATH} --model-dir ${CHECKPOINT_PATH}
 ```
 
 For the GPU run - The above steps have been automated in `run_accuracy.sh`. You can also modify this script to use

language/llama3-405b/dataset.py

@@ -80,16 +80,6 @@ def postProcess(
         output_seq = out_tokens
         assert len(query_id_list) == len(output_seq)
 
-        # Save outputs
-        if not os.path.exists("run_outputs"):
-            os.makedirs("run_outputs")
-        fname = "q" + "_".join([str(i) for i in query_id_list])
-        fname = f"run_outputs/{fname}.pkl"
-        with open(fname, mode="wb") as f:
-            d = {"query_ids": query_id_list, "outputs": output_seq}
-            log.info(f"Saving outputs to {fname}")
-            pickle.dump(d, f)
-
         return np.asarray(output_seq, dtype=np.int32)
 
     def LoadSamplesToRam(self, sample_list):