improve check for base case

* option to not concatenate during pretraining

* simplify conditional and add doc to config.qmd

docker/Dockerfile-cloud

@@ -20,7 +20,8 @@ RUN apt install --yes --no-install-recommends openssh-server tmux && \
     printf "\n[[ -z \"\$TMUX\"  ]] && { tmux attach-session -t ssh_tmux || tmux new-session -s ssh_tmux; exit; }\n" >> ~/.bashrc && \
     printf "[ ! -z \"\$TERM\" -a -r /etc/motd ] && cat /etc/motd\n" >> ~/.bashrc && \
     chmod +x /workspace/axolotl/scripts/cloud-entrypoint.sh && \
-    chmod +x /root/cloud-entrypoint.sh
+    chmod +x /root/cloud-entrypoint.sh && \
+    echo 'set-option -g history-limit 5000' >> ~/.tmux.conf
 
 ENTRYPOINT ["/root/cloud-entrypoint.sh"]
 CMD ["sleep", "infinity"]

docs/config.qmd

@@ -244,6 +244,8 @@ total_num_tokens:
 sample_packing_group_size: 100000
 # The number of samples which can be packed into one sequence. Increase if using a large sequence_len with many short samples.
 sample_packing_bin_size: 200
+# whether to concatenate samples during pretraining
+pretraining_sample_concatenation:
 
 # Use batch flattening for speedups when not using sample_packing
 batch_flattening:

scripts/chat_datasets.py

@@ -30,7 +30,7 @@ def parse_dataset(dataset=None, split="train"):
         )
     ds_cfg["field_messages"] = field_messages
 
-    message_fields = features["conversations"][0].keys()
+    message_fields = features[field_messages][0].keys()
     message_field_role = None
     for key in ["from", "role"]:
         if key in message_fields:

src/axolotl/common/datasets.py

@@ -11,7 +11,7 @@ from datasets import Dataset
 import axolotl.monkeypatch.data.batch_dataset_fetcher  # pylint: disable=unused-import  # noqa: F401
 from axolotl.cli.args import PreprocessCliArgs, TrainerCliArgs
 from axolotl.utils.data import prepare_dataset
-from axolotl.utils.data.rl import load_prepare_dpo_datasets
+from axolotl.utils.data.rl import load_prepare_preference_datasets
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_processor, load_tokenizer
 from axolotl.utils.tokenization import check_dataset_labels
@@ -103,9 +103,9 @@ def load_preference_datasets(
     cli_args: Union[PreprocessCliArgs, TrainerCliArgs],
 ) -> TrainDatasetMeta:
     """
-    Loads one or more training or evaluation datasets for DPO training, calling
+    Loads one or more training or evaluation datasets for RL training using paired
-    `axolotl.utils.data.rl.load_prepare_dpo_datasets`. Optionally, logs out debug
+    preference data, calling `axolotl.utils.data.rl.load_prepare_preference_datasets`.
-    information.
+    Optionally, logs out debug information.
 
     Args:
         cfg: Dictionary mapping `axolotl` config keys to values.
@@ -115,7 +115,7 @@ def load_preference_datasets(
         Dataclass with fields for training and evaluation datasets and the computed
         `total_num_steps`.
     """
-    train_dataset, eval_dataset = load_prepare_dpo_datasets(cfg)
+    train_dataset, eval_dataset = load_prepare_preference_datasets(cfg)
     total_num_steps = int(
         math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
     )

src/axolotl/core/trainer_builder.py

@@ -1877,6 +1877,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         self, training_args: AxolotlTrainingArguments, is_eval=False, **kwargs
     ):
         if training_args.pretraining:
+            if self.cfg.pretraining_sample_concatenation is False:
+                return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
             return None
 
         if self.cfg.model_config_type == "mamba":

src/axolotl/prompt_strategies/chat_template.py

@@ -223,7 +223,7 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
     def tokenize_prompt(self, prompt):
         # Old simple legacy behavior that works reliably.
         if (
-            not self.roles_to_train
+            (not self.roles_to_train or self.roles_to_train == ["assistant"])
             and not self.train_on_eos
             and not self.prompter.message_field_training
             and not self.prompter.message_field_training_detail

src/axolotl/utils/config/models/input/v0_4_1/__init__.py

@@ -706,6 +706,12 @@ class AxolotlInputConfig(
     pad_to_sequence_len: Optional[bool] = None
     curriculum_sampling: Optional[bool] = None
     multipack_real_batches: Optional[bool] = None
+    pretraining_sample_concatenation: Optional[bool] = Field(
+        default=None,
+        json_schema_extra={
+            "description": "whether to soft pack/concatenate samples during pretraining",
+        },
+    )
 
     batch_flattening: Optional[Union[Literal["auto"], bool]] = None
 

src/axolotl/utils/data/__init__.py

@@ -5,7 +5,7 @@ from axolotl.utils.data.pretraining import (  # noqa: F401
     encode_pretraining,
     wrap_pretraining_dataset,
 )
-from axolotl.utils.data.rl import load_prepare_dpo_datasets  # noqa: F401
+from axolotl.utils.data.rl import load_prepare_preference_datasets  # noqa: F401
 from axolotl.utils.data.sft import (  # noqa: F401
     get_dataset_wrapper,
     load_prepare_datasets,

src/axolotl/utils/data/pretraining.py

@@ -18,10 +18,14 @@ LOG = logging.getLogger("axolotl")
 
 
 def encode_pretraining(
-    tokenizer: PreTrainedTokenizerBase, max_tokens: int, examples: Dict[str, List]
+    tokenizer: PreTrainedTokenizerBase,
+    max_tokens: int,
+    examples: Dict[str, List],
+    text_column: str = "text",
+    concatenate: bool = True,
 ) -> Dict[str, List]:
     res = tokenizer(
-        examples["text"],
+        examples[text_column],
         truncation=True,
         max_length=max_tokens - 2,
         add_special_tokens=True,
@@ -30,6 +34,13 @@ def encode_pretraining(
     input_ids = [torch.tensor(seq) for seq in res["input_ids"]]
     targets = [torch.tensor(seq) for seq in res["input_ids"]]
     attention_mask = [torch.tensor(seq) for seq in res["attention_mask"]]
+    if not concatenate:
+        return {
+            "input_ids": [seq.tolist() for seq in input_ids],
+            "labels": [seq.tolist() for seq in targets],
+            "attention_mask": [seq.tolist() for seq in attention_mask],
+        }
+
     new_input_ids = []
     new_labels = []
     new_attention_mask = []
@@ -196,7 +207,13 @@ def wrap_pretraining_dataset(
         # set this to 1 so downstream data_loader doesn't try to increase the batch again
         cfg.micro_batch_size = 1
     else:
-        encode = functools.partial(encode_pretraining, tokenizer, max_tokens)
+        encode = functools.partial(
+            encode_pretraining,
+            tokenizer,
+            max_tokens,
+            text_column=cfg.pretraining_dataset[0].text_column or "text",
+            concatenate=cfg.pretraining_sample_concatenation is True,
+        )
 
     if cfg.shuffle_merged_datasets:
         dataset = dataset.shuffle(seed=seed, buffer_size=buffer_size)

src/axolotl/utils/data/rl.py

@@ -115,7 +115,7 @@ def drop_long_rl_seq(
     raise ValueError("Unknown RL type")
 
 
-def load_prepare_dpo_datasets(cfg):
+def load_prepare_preference_datasets(cfg):
     def load_split(dataset_cfgs, _cfg):
         split_datasets: List[Any] = []
         for i, ds_cfg in enumerate(dataset_cfgs):

src/axolotl/utils/models.py

@@ -1057,7 +1057,7 @@ class ModelLoader:
         )
         if (
             hasattr(self.model, "get_input_embeddings")
-            and self.model.get_input_embeddings().num_embeddings < embeddings_len
+            and self.model.get_input_embeddings().num_embeddings != embeddings_len
         ):
             resize_kwargs = {}
             if self.cfg.mean_resizing_embeddings is not None:

tests/e2e/test_llama_pretrain.py

@@ -4,7 +4,8 @@ E2E tests for llama pretrain
 
 import logging
 import os
-import unittest
+
+import pytest
 
 from axolotl.cli.args import TrainerCliArgs
 from axolotl.common.datasets import load_datasets
@@ -12,19 +13,22 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config
 from axolotl.utils.dict import DictDefault
 
-from .utils import check_model_output_exists, with_temp_dir
+from .utils import check_model_output_exists
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
 
 
-class TestPretrainLlama(unittest.TestCase):
+class TestPretrainLlama:
     """
     Test case for Llama models w pretraining
     """
 
-    @with_temp_dir
+    @pytest.mark.parametrize(
-    def test_pretrain_w_sample_packing(self, temp_dir):
+        "sample_packing",
+        [True, False],
+    )
+    def test_pretrain(self, temp_dir, sample_packing):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
@@ -32,7 +36,7 @@ class TestPretrainLlama(unittest.TestCase):
                 "tokenizer_type": "LlamaTokenizer",
                 "flash_attention": True,
                 "sequence_len": 1024,
-                "sample_packing": True,
+                "sample_packing": sample_packing,
                 "special_tokens": {
                     "unk_token": "<unk>",
                     "bos_token": "<s>",

tests/test_datasets.py

@@ -17,7 +17,7 @@ from huggingface_hub import snapshot_download
 from transformers import AutoTokenizer
 
 from axolotl.utils.data import load_tokenized_prepared_datasets
-from axolotl.utils.data.rl import load_prepare_dpo_datasets
+from axolotl.utils.data.rl import load_prepare_preference_datasets
 from axolotl.utils.dict import DictDefault
 
 
@@ -280,7 +280,7 @@ class TestDatasetPreparation(unittest.TestCase):
             }
         )
 
-        train_dataset, _ = load_prepare_dpo_datasets(cfg)
+        train_dataset, _ = load_prepare_preference_datasets(cfg)
 
         assert len(train_dataset) == 1800
         assert "conversation" in train_dataset.features
@@ -329,7 +329,7 @@ class TestDatasetPreparation(unittest.TestCase):
             }
         )
 
-        train_dataset, _ = load_prepare_dpo_datasets(cfg)
+        train_dataset, _ = load_prepare_preference_datasets(cfg)
 
         assert len(train_dataset) == 1800
         assert "conversation" in train_dataset.features

tests/test_exact_deduplication.py

@@ -12,7 +12,7 @@ from datasets import Dataset
 from transformers import AutoTokenizer
 
 from axolotl.utils.data import prepare_dataset
-from axolotl.utils.data.rl import load_prepare_dpo_datasets
+from axolotl.utils.data.rl import load_prepare_preference_datasets
 from axolotl.utils.data.utils import deduplicate_and_log_datasets
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_processor, load_tokenizer
@@ -236,7 +236,7 @@ class TestDeduplicateRLDataset(unittest.TestCase):
         """Verify that loading with deduplication removes duplicates."""
 
         # Load the dataset using the deduplication setting
-        train_dataset, _ = load_prepare_dpo_datasets(self.cfg)
+        train_dataset, _ = load_prepare_preference_datasets(self.cfg)
 
         # Verify that the dataset has been deduplicated
         assert len(train_dataset) == 1800, "Dataset was not properly deduplicated"
@@ -245,7 +245,7 @@ class TestDeduplicateRLDataset(unittest.TestCase):
         """Verify that loading without deduplication retains duplicates."""
         self.cfg.dataset_exact_deduplication = False
         # Load the dataset without deduplication
-        train_dataset, _ = load_prepare_dpo_datasets(self.cfg)
+        train_dataset, _ = load_prepare_preference_datasets(self.cfg)
 
         # Verify that the dataset retains duplicates
         assert (