fix for position_ids w packing

src/axolotl/utils/trainer.py

@@ -5,6 +5,7 @@ import math
 import os
 import sys
 from dataclasses import dataclass, field
+from functools import partial
 from pathlib import Path
 from typing import Optional, Union
 
@@ -166,10 +167,15 @@ def add_position_ids(sample):
     return sample
 
 
+def drop_long_seq(sample, sequence_len=2048):
+    return len(sample["input_ids"]) <= sequence_len
+
+
 def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
     if cfg.sample_packing:
-        # train_dataset = train_dataset.map(add_position_ids)
+        drop_long = partial(drop_long_seq, sequence_len=cfg.sequence_len)
-        # eval_dataset = eval_dataset.map(add_position_ids)
+        train_dataset = train_dataset.filter(drop_long).map(add_position_ids)
+        eval_dataset = eval_dataset.filter(drop_long).map(add_position_ids)
         if cfg.sample_packing_eff_est:
             total_num_tokens = sum(len(s["input_ids"]) for s in train_dataset)
             total_num_steps = (
@@ -417,8 +423,6 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
         data_collator_kwargs["pad_to_multiple_of"] = 8
 
     if cfg.is_llama_derived_model and cfg.landmark_attention:
-        from functools import partial
-
         from axolotl.monkeypatch.llama_landmark_attn import (
             add_mem_tokens,
             get_mem_id,