fix attetion mask with packing

README.md

@@ -237,7 +237,7 @@ Have dataset(s) in one of the following format (JSONL recommended):
 #### How to add custom prompts
 
   1. Add your method to a file in [prompt_strategies](src/axolotl/prompt_strategies). Please see other files as example.
-  2. Use your custom file name as the dataset type `<prompt_strategies_file>.load_<load_fn>`.
+  2. Use your custom file name as the dataset type.
 
 Optionally, download some datasets, see [data/README.md](data/README.md)
 
@@ -255,18 +255,10 @@ See sample configs in [configs](configs) folder or [examples](examples) for quic
 
 - dataset
   ```yaml
-  sequence_len: 2048 # max token length for prompt
-
-  # huggingface repo
   datasets:
-    - path: vicgalle/alpaca-gpt4
-      type: alpaca # format from earlier
-
-  # local
-  datasets:
-    - path: json
-      data_files: data.jsonl # or json
+    - path: vicgalle/alpaca-gpt4 # local or huggingface repo
       type: alpaca # format from earlier
+  sequence_len: 2048 # max token length / prompt
   ```
 
 - loading
@@ -336,10 +328,10 @@ tf32: true # require >=ampere
 
 # a list of one or more datasets to finetune the model with
 datasets:
-  # hf dataset repo | "json" for local dataset, make sure to fill data_files
+  # this can be either a hf dataset, or relative path
   - path: vicgalle/alpaca-gpt4
   # The type of prompt to use for training. [alpaca, sharegpt, gpteacher, oasst, reflection]
-    type: alpaca # format | format:<prompt_style> (chat/instruct) | <prompt_strategies>.load_<load_fn>
+    type: alpaca # format OR format:prompt_style (chat/instruct)
     data_files: # path to source data files
     shards: # number of shards to split data into
 
@@ -349,7 +341,7 @@ dataset_prepared_path: data/last_run_prepared
 # push prepared dataset to hub
 push_dataset_to_hub: # repo path
 # push checkpoints to hub
-hub_model_id: # repo path
+push_to_hub_model_id: # repo path
 # whether to use hf `use_auth_token` for loading datasets. Useful for fetching private datasets
 # required to be true when used in combination with `push_dataset_to_hub`
 hf_use_auth_token: # boolean

docker/Dockerfile-base

@@ -97,4 +97,4 @@ RUN cd /workspace/builds/bitsandbytes && python3 setup.py install
 RUN git lfs install --skip-repo
 RUN pip3 install awscli && \
     # The base image ships with `pydantic==1.8.2` which is not working
-    pip3 install -U --no-cache-dir pydantic==1.10.10
+    pip3 install -U --no-cache-dir pydantic

src/axolotl/datasets.py

@@ -79,11 +79,13 @@ class ConstantLengthDataset(IterableDataset):
         buffer = {"input_ids": [], "attention_mask": [], "labels": []}
         buffer_len = 0
         for dataset in self.datasets:
+            idx = 0
             iterator = iter(dataset)
             more_examples = True
             while more_examples:
                 try:
                     example = next(iterator)
+                    idx += 1
                 except StopIteration:
                     more_examples = False
                     example = None
@@ -124,6 +126,7 @@ class ConstantLengthDataset(IterableDataset):
                         "labels": [],
                     }
                     buffer_len = 0
+                    idx = 1
 
                 if example:
                     # FIXME
@@ -132,11 +135,6 @@ class ConstantLengthDataset(IterableDataset):
                         input_ids = example["input_ids"]
                         attention_mask = example["attention_mask"]
                         labels = example["labels"]
-                        if (
-                            buffer["input_ids"]
-                            and input_ids[0] == self.tokenizer.bos_token_id
-                        ):
-                            attention_mask[0] = 0
 
                         if add_concat_token:
                             input_ids.append(self.concat_token_id)
@@ -147,7 +145,7 @@ class ConstantLengthDataset(IterableDataset):
                             input_ids, dtype=self.tokens_dtype
                         )
                         attention_mask_with_concat = torch.tensor(
-                            attention_mask, dtype=self.tokens_dtype
+                            [idx * m for m in attention_mask], dtype=torch.int16
                         )
                         labels_with_concat = torch.tensor(
                             labels, dtype=self.tokens_dtype

src/axolotl/utils/data.py

@@ -37,7 +37,7 @@ from axolotl.prompters import (
 
 
 def load_tokenized_prepared_datasets(
-    tokenizer, cfg, default_dataset_prepared_path
+    split, tokenizer, cfg, default_dataset_prepared_path
 ) -> DatasetDict:
     tokenizer_name = tokenizer.__class__.__name__
     ds_hash = str(
@@ -49,6 +49,8 @@ def load_tokenized_prepared_datasets(
                     sorted([f"{d.path}:{d.type}:{d.shards}" for d in cfg.datasets])
                 )
                 + "|"
+                + split
+                + "|"
                 + tokenizer_name
             ).encode("utf-8")
         ).hexdigest()
@@ -66,7 +68,7 @@ def load_tokenized_prepared_datasets(
                 f"{cfg.push_dataset_to_hub}/{ds_hash}",
                 use_auth_token=use_auth_token,
             )
-            dataset = dataset["train"]
+            dataset = dataset[split]
     except Exception:  # pylint: disable=broad-except # nosec
         pass
 
@@ -102,26 +104,13 @@ def load_tokenized_prepared_datasets(
                 pass
 
             # prefer local dataset, even if hub exists
-            local_path = Path(d.path)
-            if local_path.exists():
-                if local_path.is_dir():
-                    ds = load_dataset(
-                        d.path,
-                        data_files=d.data_files,
-                        streaming=False,
-                        split=None,
-                    )
-                elif local_path.is_file():
-                    ds = load_dataset(
-                        "json",
-                        data_files=d.path,
-                        streaming=False,
-                        split=None,
-                    )
-                else:
-                    raise ValueError(
-                        "unhandled dataset load: local path exists, but is neither a directory or a file"
-                    )
+            if Path(d.path).exists():
+                ds = load_dataset(
+                    "json",
+                    data_files=d.path,
+                    streaming=False,
+                    split=None,
+                )
             elif ds_from_hub:
                 if d.data_files:
                     ds = load_dataset(
@@ -147,8 +136,8 @@ def load_tokenized_prepared_datasets(
                 raise ValueError("unhandled dataset load")
             # support for using a subset of the data
             if d.shards:
-                if "train" in ds:
-                    ds = ds.shuffle(seed=seed)["train"].shard(
+                if split in ds:
+                    ds = ds.shuffle(seed=seed)[split].shard(
                         num_shards=d.shards, index=0
                     )
                 else:
@@ -157,8 +146,8 @@ def load_tokenized_prepared_datasets(
             d_type_split = d_type.split(":")
             d_base_type = d_type_split[0]
             d_prompt_style = d_type_split[1] if len(d_type_split) > 1 else None
-            if "train" in ds:
-                ds = ds["train"]
+            if split in ds:
+                ds = ds[split]
             if ds_strategy := load(d.type, tokenizer, cfg):
                 ds_wrapper = TokenizedPromptDataset(ds_strategy, ds)
                 datasets.append(ds_wrapper)
@@ -332,7 +321,6 @@ def load_prepare_datasets(
                     f"{cfg.push_dataset_to_hub}/{ds_hash}",
                     use_auth_token=use_auth_token,
                 )
-                dataset = dataset["train"]
         except Exception:  # pylint: disable=broad-except # nosec
             pass
 
@@ -352,28 +340,37 @@ def load_prepare_datasets(
                     f"{cfg.push_dataset_to_hub}/{ds_hash}", private=True
                 )
         else:
-            dataset = load_tokenized_prepared_datasets(
-                tokenizer, cfg, default_dataset_prepared_path
+            dataset_train = load_tokenized_prepared_datasets(
+                "train", tokenizer, cfg, default_dataset_prepared_path
             )
-
+            dataset_test = load_tokenized_prepared_datasets(
+                "test", tokenizer, cfg, default_dataset_prepared_path
+            )
+            dataset = DatasetDict({"train": dataset_train, "test": dataset_test})
             if cfg.seed:
                 dataset = dataset.shuffle(seed=cfg.seed)
 
-            constant_len_dataset = ConstantLengthDataset(
+            constant_len_dataset_train = ConstantLengthDataset(
                 tokenizer,
-                [dataset],
+                [dataset["train"]],
+                seq_length=max_packed_sequence_len,
+            )
+            constant_len_dataset_test = ConstantLengthDataset(
+                tokenizer,
+                [dataset["test"]],
                 seq_length=max_packed_sequence_len,
             )
             logging.info(
                 f"packing master dataset to len: {cfg.max_packed_sequence_len}"
             )
-            dataset = Dataset.from_list(list(constant_len_dataset))
+            dataset_train = Dataset.from_list(list(constant_len_dataset_train))
+            dataset_test = Dataset.from_list(list(constant_len_dataset_test))
 
             # filter out bad data
-            dataset = Dataset.from_list(
+            dataset_train = Dataset.from_list(
                 [
                     d
-                    for d in dataset
+                    for d in dataset_train
                     if len(d["input_ids"]) < cfg.sequence_len
                     and len(d["input_ids"]) > 0
                     and len(d["input_ids"]) == len(d["attention_mask"])
@@ -381,6 +378,19 @@ def load_prepare_datasets(
                 ]
             )
 
+            # filter out bad data
+            dataset_test = Dataset.from_list(
+                [
+                    d
+                    for d in dataset_test
+                    if len(d["input_ids"]) < cfg.sequence_len
+                    and len(d["input_ids"]) > 0
+                    and len(d["input_ids"]) == len(d["attention_mask"])
+                    and len(d["input_ids"]) == len(d["labels"])
+                ]
+            )
+            dataset = DatasetDict({"train": dataset_train, "test": dataset_test})
+
             if cfg.local_rank == 0:
                 logging.info(
                     f"Saving packed prepared dataset to disk... {prepared_ds_path}"
@@ -395,9 +405,13 @@ def load_prepare_datasets(
                         private=True,
                     )
     else:
-        dataset = load_tokenized_prepared_datasets(
-            tokenizer, cfg, default_dataset_prepared_path
+        dataset_train = load_tokenized_prepared_datasets(
+            "train", tokenizer, cfg, default_dataset_prepared_path
         )
+        dataset_test = load_tokenized_prepared_datasets(
+            "test", tokenizer, cfg, default_dataset_prepared_path
+        )
+        dataset = DatasetDict({"train": dataset_train, "test": dataset_test})
 
     if cfg.dataset_shard_num and cfg.dataset_shard_idx is not None:
         logging.info(
@@ -412,6 +426,9 @@ def load_prepare_datasets(
         dataset = dataset.train_test_split(test_size=cfg.val_set_size, shuffle=False)
         train_dataset = dataset["train"]
         eval_dataset = dataset["test"]
+    elif "train" in dataset:
+        train_dataset = dataset["train"]
+        eval_dataset = dataset["test"]
     else:
         train_dataset = dataset
         eval_dataset = None

src/axolotl/utils/models.py

@@ -202,7 +202,7 @@ def load_model(
                 else True,
             )
             load_in_8bit = False
-        elif cfg.is_llama_derived_model and not cfg.trust_remote_code:
+        elif cfg.is_llama_derived_model:
             from transformers import LlamaForCausalLM
 
             config = LlamaConfig.from_pretrained(base_model_config)
@@ -241,7 +241,7 @@ def load_model(
         #         device=cfg.device,
         #     )
         #     model.train() # sets to train instead of eval mode
-        elif model_type and not cfg.trust_remote_code:
+        elif model_type:
             model = getattr(transformers, model_type).from_pretrained(
                 base_model,
                 load_in_8bit=cfg.load_in_8bit and cfg.adapter is not None,

src/axolotl/utils/schedulers.py

@@ -1,9 +1,6 @@
 """Module for custom LRScheduler class"""
-import math
-from functools import partial
 
-from torch.optim import Optimizer
-from torch.optim.lr_scheduler import LambdaLR, LRScheduler
+from torch.optim.lr_scheduler import LRScheduler
 
 
 class InterpolatingLogScheduler(LRScheduler):
@@ -45,58 +42,3 @@ class InterpolatingLogScheduler(LRScheduler):
             lrs = [self.max_lr for base_lr in self.base_lrs]
 
         return lrs
-
-
-def _get_cosine_schedule_with_quadratic_warmup_lr_lambda(
-    current_step: int,
-    *,
-    num_warmup_steps: int,
-    num_training_steps: int,
-    num_cycles: float
-):
-    if current_step < num_warmup_steps:
-        return (float(current_step) / float(max(1, num_warmup_steps))) ** 2
-    progress = float(current_step - num_warmup_steps) / float(
-        max(1, num_training_steps - num_warmup_steps)
-    )
-    return max(
-        0.0, 0.5 * (1.0 + math.cos(math.pi * float(num_cycles) * 2.0 * progress))
-    )
-
-
-def get_cosine_schedule_with_quadratic_warmup(
-    optimizer: Optimizer,
-    num_warmup_steps: int,
-    num_training_steps: int,
-    num_cycles: float = 0.5,
-    last_epoch: int = -1,
-):
-    """
-    Create a schedule with a learning rate that decreases following the values of the cosine function between the
-    initial lr set in the optimizer to 0, after a warmup period during which it increases linearly between 0 and the
-    initial lr set in the optimizer.
-
-    Args:
-        optimizer ([`~torch.optim.Optimizer`]):
-            The optimizer for which to schedule the learning rate.
-        num_warmup_steps (`int`):
-            The number of steps for the warmup phase.
-        num_training_steps (`int`):
-            The total number of training steps.
-        num_cycles (`float`, *optional*, defaults to 0.5):
-            The number of waves in the cosine schedule (the defaults is to just decrease from the max value to 0
-            following a half-cosine).
-        last_epoch (`int`, *optional*, defaults to -1):
-            The index of the last epoch when resuming training.
-
-    Return:
-        `torch.optim.lr_scheduler.LambdaLR` with the appropriate schedule.
-    """
-
-    lr_lambda = partial(
-        _get_cosine_schedule_with_quadratic_warmup_lr_lambda,
-        num_warmup_steps=num_warmup_steps,
-        num_training_steps=num_training_steps,
-        num_cycles=num_cycles,
-    )
-    return LambdaLR(optimizer, lr_lambda, last_epoch)

src/axolotl/utils/trainer.py

@@ -5,7 +5,6 @@ import logging
 import math
 import os
 import sys
-from dataclasses import field
 from pathlib import Path
 from typing import Optional
 
@@ -14,67 +13,17 @@ import torch.cuda
 import transformers
 from torch import nn
 from torch.optim.lr_scheduler import OneCycleLR
-from transformers import EarlyStoppingCallback, Trainer, TrainingArguments
+from transformers import EarlyStoppingCallback, Trainer
 from transformers.trainer_pt_utils import get_parameter_names
 
 from axolotl.utils.callbacks import (
     SaveBetterTransformerModelCallback,
     SavePeftModelCallback,
 )
-from axolotl.utils.schedulers import (
-    InterpolatingLogScheduler,
-    get_cosine_schedule_with_quadratic_warmup,
-)
+from axolotl.utils.schedulers import InterpolatingLogScheduler
 
 
-class AxolotlTrainingArguments(TrainingArguments):
-    """
-    Extend the base TrainingArguments for axolotl helpers
-    """
-
-    lr_quadratic_warmup: bool = field(
-        default=False,
-        metadata={"help": "Use quadratic warmup for cosine scheduling."},
-    )
-
-
-class AxolotlTrainer(Trainer):
-    """
-    Extend the base Trainer for axolotl helpers
-    """
-
-    args = None  # type: AxolotlTrainingArguments
-
-    def create_scheduler(
-        self, num_training_steps: int, optimizer: torch.optim.Optimizer = None
-    ):
-        """
-        Setup the scheduler. The optimizer of the trainer must have been set up either before this method is called or
-        passed as an argument.
-
-        Args:
-            num_training_steps (int): The number of training steps to do.
-            optimizer (torch.optim.Optimizer): The training optimizer
-        """
-
-        # fmt: off
-        if self.lr_scheduler is None:  # type: ignore  # pylint: disable=access-member-before-definition
-            # fmt: on
-            if (
-                self.args.lr_scheduler_type == "cosine"
-                and self.args.lr_quadratic_warmup is True
-            ):
-                self.lr_scheduler = get_cosine_schedule_with_quadratic_warmup(  # pylint: disable=attribute-defined-outside-init
-                    optimizer,
-                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
-                    num_training_steps=num_training_steps,
-                )
-            else:
-                return super().create_scheduler(num_training_steps, optimizer)
-        return self.lr_scheduler
-
-
-class OneCycleLRSchedulerTrainer(AxolotlTrainer):
+class OneCycleLRSchedulerTrainer(Trainer):
     """
     Trainer subclass that uses the OneCycleLR scheduler
     """
@@ -154,9 +103,6 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
         if cfg.fsdp_config:
             training_arguments_kwargs["fsdp_config"] = dict(cfg.fsdp_config)
 
-    if cfg.lr_quadratic_warmup is not None:
-        training_arguments_kwargs["lr_quadratic_warmup"] = cfg.lr_quadratic_warmup
-
     # deepspeed
     if (
         os.environ.get("ACCELERATE_USE_DEEPSPEED") == "true"
@@ -178,11 +124,11 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
     if cfg.max_grad_norm:
         training_arguments_kwargs["max_grad_norm"] = cfg.max_grad_norm
 
-    if cfg.hub_model_id:
-        training_arguments_kwargs["hub_model_id"] = cfg.hub_model_id
+    if cfg.push_to_hub_model_id:
+        training_arguments_kwargs["push_to_hub_model_id"] = cfg.push_to_hub_model_id
         training_arguments_kwargs["push_to_hub"] = True
 
-    training_args = AxolotlTrainingArguments(
+    training_args = transformers.TrainingArguments(
         per_device_train_batch_size=cfg.micro_batch_size,
         per_device_eval_batch_size=cfg.eval_batch_size
         if cfg.eval_batch_size is not None
@@ -191,9 +137,9 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
         eval_accumulation_steps=cfg.gradient_accumulation_steps,
         num_train_epochs=cfg.num_epochs,
         learning_rate=cfg.learning_rate,
-        evaluation_strategy="steps" if cfg.val_set_size > 0 else "no",
+        evaluation_strategy="steps",
         save_strategy="steps" if cfg.save_steps else "epoch",
-        eval_steps=cfg.eval_steps if cfg.val_set_size > 0 else None,
+        eval_steps=cfg.eval_steps,
         save_steps=cfg.save_steps,
         output_dir=cfg.output_dir,
         save_total_limit=3,
@@ -332,7 +278,7 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer):
     trainer_cls = (
         OneCycleLRSchedulerTrainer
         if cfg.lr_scheduler == "one_cycle" and (cfg.fsdp or cfg.adapter == "qlora")
-        else AxolotlTrainer
+        else transformers.Trainer
     )
     trainer = trainer_cls(
         model=model,

src/axolotl/utils/validation.py

@@ -87,16 +87,11 @@ def validate_config(cfg):
             "You probably want to disable group_by_length as it will force a streamed dataset to download completely."
         )
 
-    if any([cfg.adam_beta1, cfg.adam_beta2, cfg.adam_epsilon]) and (
+    if any([cfg.adamw_beta1, cfg.adamw_beta2, cfg.adamw_epsilon]) and (
         not cfg.optimizer or "adamw" not in cfg.optimizer
     ):
         logging.warning("adamw hyperparameters found, but no adamw optimizer set")
 
-    if cfg.push_to_hub_model_id:
-        raise ValueError(
-            "push_to_hub_model_id is deprecated. Please use hub_model_id instead."
-        )
-
     # TODO
     # MPT 7b
     # https://github.com/facebookresearch/bitsandbytes/issues/25

tests/test_packed_dataset.py

@@ -27,7 +27,7 @@ class TestPacking(unittest.TestCase):
             }
         )
 
-    def test_resets_attention(self):
+    def test_increments_attention(self):
         prompter = AlpacaPrompter("chat")
         strat = AlpacaPromptTokenizingStrategy(
             prompter,
@@ -58,7 +58,7 @@ class TestPacking(unittest.TestCase):
 
         # but subsequent one does
         assert example["input_ids"][next_bos_index] == self.tokenizer.bos_token_id
-        assert example["attention_mask"][next_bos_index] == 0
+        assert example["attention_mask"][next_bos_index] == 2
 
 
 if __name__ == "__main__":

tests/test_validation.py

@@ -268,7 +268,7 @@ class ValidationTest(unittest.TestCase):
         cfg = DictDefault(
             {
                 "optimizer": None,
-                "adam_epsilon": 0.0001,
+                "adamw_epsilon": 0.0001,
             }
         )
 
@@ -283,7 +283,7 @@ class ValidationTest(unittest.TestCase):
         cfg = DictDefault(
             {
                 "optimizer": "adafactor",
-                "adam_beta1": 0.0001,
+                "adamw_beta1": 0.0001,
             }
         )
 
@@ -298,9 +298,9 @@ class ValidationTest(unittest.TestCase):
         cfg = DictDefault(
             {
                 "optimizer": "adamw_bnb_8bit",
-                "adam_beta1": 0.9,
-                "adam_beta2": 0.99,
-                "adam_epsilon": 0.0001,
+                "adamw_beta1": 0.0001,
+                "adamw_beta2": 0.0001,
+                "adamw_epsilon": 0.0001,
             }
         )