moving out all diff attn code to plugin repo

* basic evaluate CLI command / codepath

* tests for evaluate CLI command

* fixes and cleanup

* review comments; slightly DRYing up things

---------

Co-authored-by: Dan Saunders <danjsaund@gmail.com>

.github/workflows/main.yml

@@ -25,7 +25,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.3.1
             axolotl_extras: mamba-ssm
-            is_latest: true
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -36,6 +35,7 @@ jobs:
             python_version: "3.11"
             pytorch: 2.5.1
             axolotl_extras:
+            is_latest: true
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -92,7 +92,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.3.1
             axolotl_extras:
-            is_latest: true
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -103,6 +102,7 @@ jobs:
             python_version: "3.11"
             pytorch: 2.5.1
             axolotl_extras:
+            is_latest: true
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

.github/workflows/tests.yml

@@ -207,7 +207,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.1
+            pytorch: 2.4.1
             num_gpus: 1
             axolotl_extras:
     steps:
@@ -253,7 +253,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.4.1
+            pytorch: 2.5.1
             num_gpus: 1
             axolotl_extras:
     steps:

.gitignore

@@ -186,3 +186,6 @@ out/
 
 # vim
 *.swp
+
+# symlinked to axolotl-artifacts in docker containers
+outputs

cicd/cicd.sh

@@ -4,7 +4,6 @@ set -e
 python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
 
 pytest -v --durations=10 -n8 --ignore=tests/e2e/ --ignore=tests/patched/ /workspace/axolotl/tests/
-# pytest -v --durations=10 -n8 --dist loadfile /workspace/axolotl/tests/patched/
 pytest -v --durations=10 /workspace/axolotl/tests/e2e/patched/
 pytest -v --durations=10 /workspace/axolotl/tests/e2e/integrations/
 pytest -v --durations=10 --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ /workspace/axolotl/tests/e2e/

cicd/multigpu.py

@@ -1,6 +1,6 @@
 """
- modal application to run axolotl gpu tests in Modal
- """
+modal application to run axolotl gpu tests in Modal
+"""
 # pylint: disable=duplicate-code
 
 import os

docs/rlhf.qmd

@@ -29,7 +29,7 @@ datasets:
     type: chatml.intel
   - path: argilla/ultrafeedback-binarized-preferences
     split: train
-    type: chatml
+    type: chatml.argilla
 ```
 
 #### IPO

src/axolotl/cli/__init__.py

@@ -202,7 +202,7 @@ def do_inference(
         )
     elif cfg.chat_template:
         chat_template_str = get_chat_template(cfg.chat_template)
-    elif cfg.datasets[0].type == "chat_template":
+    elif cfg.datasets and cfg.datasets[0].type == "chat_template":
         chat_template_str = get_chat_template_from_config(
             cfg=cfg, ds_cfg=cfg.datasets[0], tokenizer=tokenizer
         )

src/axolotl/cli/evaluate.py

@@ -3,7 +3,7 @@ CLI to run training on a model
 """
 import logging
 from pathlib import Path
-from typing import Union
+from typing import Dict, Union
 
 import fire
 from dotenv import load_dotenv
@@ -23,7 +23,7 @@ from axolotl.evaluate import evaluate
 LOG = logging.getLogger("axolotl.cli.evaluate")
 
 
-def do_evaluate(cfg, cli_args) -> None:
+def do_evaluate(cfg, cli_args) -> Dict[str, float]:
     # pylint: disable=duplicate-code
     print_axolotl_text_art()
     check_accelerate_default_config()
@@ -34,7 +34,7 @@ def do_evaluate(cfg, cli_args) -> None:
     else:
         dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
 
-    evaluate(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+    return evaluate(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
 
 
 def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs) -> None:

src/axolotl/cli/main.py

@@ -1,11 +1,13 @@
 """CLI definition for various axolotl commands."""
 # pylint: disable=redefined-outer-name
+
 import subprocess  # nosec B404
 from typing import Optional
 
 import click
 
 import axolotl
+from axolotl.cli.plugins import setup_plugin_commands
 from axolotl.cli.utils import (
     add_options_from_config,
     add_options_from_dataclass,
@@ -25,20 +27,15 @@ def cli():
 
 @cli.command()
 @click.argument("config", type=click.Path(exists=True, path_type=str))
-@click.option(
-    "--iterable/--no-iterable",
-    default=False,
-    help="Use IterableDataset for streaming processing of large datasets",
-)
 @add_options_from_dataclass(PreprocessCliArgs)
 @add_options_from_config(AxolotlInputConfig)
-def preprocess(config: str, iterable: bool, **kwargs):
+def preprocess(config: str, **kwargs):
     """Preprocess datasets before training."""
     kwargs = {k: v for k, v in kwargs.items() if v is not None}
 
     from axolotl.cli.preprocess import do_cli
 
-    do_cli(config=config, iterable=iterable, **kwargs)
+    do_cli(config=config, **kwargs)
 
 
 @cli.command()
@@ -82,6 +79,9 @@ def evaluate(config: str, accelerate: bool, **kwargs):
     """Evaluate a model."""
     kwargs = {k: v for k, v in kwargs.items() if v is not None}
 
+    # Enable expandable segments for cuda allocation to improve VRAM usage
+    set_pytorch_cuda_alloc_conf()
+
     if accelerate:
         base_cmd = ["accelerate", "launch", "-m", "axolotl.cli.evaluate"]
         if config:
@@ -259,6 +259,9 @@ def fetch(directory: str, dest: Optional[str]):
     fetch_from_github(f"{directory}/", dest)
 
 
+setup_plugin_commands(cli)
+
+
 def main():
     cli()
 

src/axolotl/cli/plugins.py

@@ -0,0 +1,36 @@
+"""Module for adding click CLI commands from axolotl plugins."""
+
+import logging
+
+import click
+
+from axolotl.cli.utils import add_options_from_config, add_options_from_dataclass
+from axolotl.logging_config import configure_logging
+from axolotl.utils.config.models.input.v0_4_1 import AxolotlInputConfig
+
+configure_logging()
+LOG = logging.getLogger(__name__)
+
+
+def setup_plugin_commands(cli: click.core.Group) -> None:
+    """
+    Setup CLI commands for available plugins.
+
+    Args:
+        cli: Click CLI object to add plugin CLI options to.
+    """
+    try:
+        from axolotl_diff_transformer.convert_diff_transformer import do_cli
+        from axolotl_diff_transformer.plugin.cli import ConvertDiffTransformerCliArgs
+
+        @cli.command()
+        @click.argument("config", type=click.Path(exists=True, path_type=str))
+        @add_options_from_dataclass(ConvertDiffTransformerCliArgs)
+        @add_options_from_config(AxolotlInputConfig)
+        def convert_diff_transformer(config: str, **kwargs):
+            """Convert model attention layers to differential attention layers."""
+            kwargs = {k: v for k, v in kwargs.items() if v is not None}
+            do_cli(config=config, **kwargs)
+
+    except ImportError as exc:
+        LOG.debug("axolotl-diff-transformer not found: %s", exc)

src/axolotl/cli/preprocess.py

@@ -4,7 +4,7 @@ CLI to run training on a model
 import logging
 import warnings
 from pathlib import Path
-from typing import Optional, Union
+from typing import Union
 
 import fire
 import transformers
@@ -28,17 +28,11 @@ from axolotl.utils.trainer import disable_datasets_caching
 LOG = logging.getLogger("axolotl.cli.preprocess")
 
 
-def do_cli(
-    config: Union[Path, str] = Path("examples/"),
-    iterable: Optional[bool] = False,
-    **kwargs,
-):
+def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
     # pylint: disable=duplicate-code
     print_axolotl_text_art()
     parsed_cfg = load_cfg(config, **kwargs)
     parsed_cfg.is_preprocess = True
-    if iterable:
-        parsed_cfg.preprocess_iterable = iterable
     check_accelerate_default_config()
     check_user_token()
     parser = transformers.HfArgumentParser((PreprocessCliArgs))

src/axolotl/cli/utils.py

@@ -22,7 +22,6 @@ def add_options_from_dataclass(config_class: Type[Any]):
         # Process dataclass fields in reverse order for correct option ordering
         for field in reversed(dataclasses.fields(config_class)):
             field_type = field.type
-
             if get_origin(field_type) is Union and type(None) in get_args(field_type):
                 field_type = next(
                     t for t in get_args(field_type) if not isinstance(t, NoneType)
@@ -44,6 +43,7 @@ def add_options_from_dataclass(config_class: Type[Any]):
                     default=field.default,
                     help=field.metadata.get("description"),
                 )(function)
+
         return function
 
     return decorator
@@ -55,7 +55,14 @@ def add_options_from_config(config_class: Type[BaseModel]):
     def decorator(function):
         # Process model fields in reverse order for correct option ordering
         for name, field in reversed(config_class.model_fields.items()):
-            if field.annotation in (bool, Optional[bool]):
+            field_type = field.annotation
+            if get_origin(field_type) is Union and type(None) in get_args(field_type):
+                field_type = next(
+                    t for t in get_args(field_type) if not isinstance(t, NoneType)
+                )
+
+            # NOTE: defaults are handled by the pydantic model config classes.
+            if field_type == bool:
                 field_name = name.replace("_", "-")
                 option_name = f"--{field_name}/--no-{field_name}"
                 function = click.option(
@@ -66,6 +73,7 @@ def add_options_from_config(config_class: Type[BaseModel]):
                 function = click.option(
                     option_name, default=None, help=field.description
                 )(function)
+
         return function
 
     return decorator
@@ -84,6 +92,8 @@ def build_command(base_cmd: List[str], options: Dict[str, Any]) -> List[str]:
         if isinstance(value, bool):
             if value:
                 cmd.append(f"--{key}")
+            else:
+                cmd.append(f"--no{key}")
         else:
             cmd.extend([f"--{key}", str(value)])
 

src/axolotl/common/cli.py

@@ -4,22 +4,26 @@ shared module for cli specific things
 
 import logging
 from dataclasses import dataclass, field
-from typing import Optional
+from typing import TYPE_CHECKING, Optional, Union
 
 import axolotl.monkeypatch.data.batch_dataset_fetcher  # pylint: disable=unused-import  # noqa: F401
 from axolotl.logging_config import configure_logging
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_model, load_tokenizer
 
+if TYPE_CHECKING:
+    try:
+        from axolotl_diff_transformer.plugin.cli import ConvertDiffTransformerCliArgs
+    except:  # noqa: E722 # pylint: disable=bare-except # nosec B110
+        pass
+
 configure_logging()
-LOG = logging.getLogger("axolotl.common.cli")
+LOG = logging.getLogger(__name__)
 
 
 @dataclass
 class PreprocessCliArgs:
-    """
-    dataclass representing arguments for preprocessing only
-    """
+    """dataclass with arguments for preprocessing only"""
 
     debug: bool = field(default=False)
     debug_text_only: bool = field(default=False)
@@ -30,9 +34,7 @@ class PreprocessCliArgs:
 
 @dataclass
 class TrainerCliArgs:
-    """
-    dataclass representing the various non-training arguments
-    """
+    """dataclass with various non-training arguments"""
 
     debug: bool = field(default=False)
     debug_text_only: bool = field(default=False)
@@ -45,9 +47,7 @@ class TrainerCliArgs:
 
 @dataclass
 class EvaluateCliArgs:
-    """
-    dataclass representing the various evaluation arguments
-    """
+    """dataclass with various evaluation arguments"""
 
     debug: bool = field(default=False)
     debug_text_only: bool = field(default=False)
@@ -57,7 +57,7 @@ class EvaluateCliArgs:
 def load_model_and_tokenizer(
     *,
     cfg: DictDefault,
-    cli_args: TrainerCliArgs,
+    cli_args: Union[TrainerCliArgs, EvaluateCliArgs, "ConvertDiffTransformerCliArgs"],
 ):
     LOG.info(f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}")
     tokenizer = load_tokenizer(cfg)

src/axolotl/core/trainers/base.py

@@ -1,933 +0,0 @@
-"""
-module for customized trainers
-"""
-
-from __future__ import annotations
-
-# pylint: disable=too-many-lines
-import gc
-import logging
-import os
-from collections import defaultdict
-from functools import wraps
-from typing import Any, Dict, Literal, Optional, Union
-
-import torch
-from datasets import Dataset
-from peft.optimizers import create_loraplus_optimizer
-from torch import nn
-from torch.optim.lr_scheduler import OneCycleLR
-from torch.utils.data import BatchSampler, DataLoader, RandomSampler, SequentialSampler
-from transformers import Trainer
-from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR, seed_worker
-from transformers.utils import is_sagemaker_mp_enabled
-from trl import CPOTrainer, DPOTrainer, KTOTrainer, ORPOTrainer, RewardTrainer
-from trl.trainer.utils import pad_to_length
-
-from axolotl.monkeypatch.relora import ReLoRAScheduler
-from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
-from axolotl.utils.schedulers import (
-    get_cosine_schedule_with_min_lr,
-    get_cosine_schedule_with_quadratic_warmup,
-    get_cosine_schedule_with_warmup_decay_constant,
-)
-
-if is_sagemaker_mp_enabled():
-    import smdistributed.modelparallel.torch as smp
-
-LOG = logging.getLogger("axolotl.core.trainer_builder")
-
-
-def _sanitize_kwargs_for_tagging(tag_names, kwargs=None):
-    if isinstance(tag_names, str):
-        tag_names = [tag_names]
-
-    if kwargs is not None:
-        if "tags" not in kwargs:
-            kwargs["tags"] = tag_names
-        elif "tags" in kwargs and isinstance(kwargs["tags"], list):
-            kwargs["tags"].extend(tag_names)
-        elif "tags" in kwargs and isinstance(kwargs["tags"], str):
-            tag_names.append(kwargs["tags"])
-            kwargs["tags"] = tag_names
-
-    return kwargs
-
-
-def _sanitize_kwargs_for_ds_tagging(dataset_tags, kwargs=None):
-    if isinstance(dataset_tags, str):
-        dataset_tags = [dataset_tags]
-
-    if (dataset_tags is not None) and (kwargs is not None):
-        if "dataset_tags" not in kwargs:
-            kwargs["dataset_tags"] = dataset_tags
-        elif "dataset_tags" in kwargs and isinstance(kwargs["dataset_tags"], list):
-            kwargs["dataset_tags"].extend(dataset_tags)
-        elif "dataset_tags" in kwargs and isinstance(kwargs["dataset_tags"], str):
-            dataset_tags.append(kwargs["dataset_tags"])
-            kwargs["dataset_tags"] = dataset_tags
-
-    return kwargs
-
-
-class SchedulerMixin(Trainer):
-    """
-    Mixin class for scheduler setup in CausalTrainer.
-    """
-
-    args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
-
-    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
-        """
-        use_cosine_quadratic = (
-            self.args.lr_scheduler_type == "cosine"
-            and self.args.lr_quadratic_warmup is True
-        )
-
-        use_cosine_min_lr = (
-            self.args.lr_scheduler_type == "cosine"
-            and self.args.cosine_min_lr_ratio is not None
-        )
-
-        # fmt: off
-        if self.lr_scheduler is None:  # type: ignore  # pylint: disable=access-member-before-definition
-            # fmt: on
-            if self.args.alternate_lr_scheduler_type == "one_cycle":
-                num_warmup_steps = self.args.get_warmup_steps(num_training_steps)
-                pct_start = num_warmup_steps / num_training_steps
-                extra_lr_kwargs = {}
-                if "pct_start" not in self.args.lr_scheduler_kwargs:
-                    extra_lr_kwargs["pct_start"] = pct_start
-                if "anneal_strategy" not in self.args.lr_scheduler_kwargs:
-                    extra_lr_kwargs["anneal_strategy"] = "cos"
-
-                self.lr_scheduler = OneCycleLR(
-                    optimizer,
-                    max_lr=self.args.learning_rate,
-                    total_steps=num_training_steps,
-                    **extra_lr_kwargs,
-                    **self.args.lr_scheduler_kwargs,
-                )
-            elif use_cosine_quadratic:
-                if use_cosine_min_lr:
-                    LOG.warning("Both cosine quadratic warmup and min lr detected. Using quadratic warmup.")
-
-                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,
-                )
-            elif self.args.cosine_min_lr_ratio and self.args.cosine_constant_lr_ratio and use_cosine_min_lr:
-                assert 0 <= self.args.cosine_min_lr_ratio <= 1.0, "cosine_min_lr_ratio must be between 0.0 and 1.0"
-                assert 0 <= self.args.cosine_constant_lr_ratio <= 1.0, "cosine_constant_lr_ratio must be between 0.0 and 1.0"
-                self.lr_scheduler = get_cosine_schedule_with_warmup_decay_constant(  # pylint: disable=attribute-defined-outside-init
-                    optimizer,
-                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
-                    num_training_steps=num_training_steps,
-                    min_lr_ratio=self.args.cosine_min_lr_ratio,
-                    constant_lr_ratio=self.args.cosine_constant_lr_ratio,
-                )
-            elif self.args.cosine_min_lr_ratio and use_cosine_min_lr:
-                assert 0 <= self.args.cosine_min_lr_ratio <= 1.0, "cosine_min_lr_ratio must be between 0.0 and 1.0"
-                self.lr_scheduler = get_cosine_schedule_with_min_lr(  # pylint: disable=attribute-defined-outside-init
-                    optimizer,
-                    num_warmup_steps=self.args.get_warmup_steps(num_training_steps),
-                    num_training_steps=num_training_steps,
-                    min_lr_ratio=self.args.cosine_min_lr_ratio,
-                )
-            else:
-                return super().create_scheduler(num_training_steps, optimizer=optimizer)
-        else:
-            if use_cosine_quadratic:
-                LOG.warning("axolotl's cosine scheduler with quadratic warmup not used (e.g., because of deepspeed).")
-
-            if use_cosine_min_lr:
-                LOG.warning("axolotl's cosine scheduler with min lr not used (e.g., because of deepspeed).")
-
-        return self.lr_scheduler
-
-
-class AxolotlTrainer(SchedulerMixin, Trainer):
-    """
-    Extend the base Trainer for axolotl helpers
-    """
-
-    args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
-    tag_names = ["axolotl"]
-
-    def __init__(
-        self,
-        *_args,
-        bench_data_collator=None,
-        eval_data_collator=None,
-        dataset_tags=None,
-        **kwargs,
-    ):
-        self.bench_data_collator = bench_data_collator
-        self.eval_data_collator = eval_data_collator
-        self.dataset_tags = dataset_tags
-        self._signature_columns = None  # workaround for pylint
-        super().__init__(*_args, **kwargs)
-        self.train_data_collator = self.data_collator
-        self._stored_metrics = defaultdict(lambda: defaultdict(list))
-        if self.args.orpo_alpha:
-            self.loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
-
-    def _wrap_model(self, model, training=True, dataloader=None):
-        if self.args.torch_compile:
-            torch._dynamo.config.accumulated_cache_size_limit = (  # pylint: disable=protected-access
-                256
-            )
-            model = torch.compile(
-                model,
-                backend=self.args.torch_compile_backend,
-                mode=self.args.torch_compile_mode,
-            )
-        return super()._wrap_model(model, training=training, dataloader=dataloader)
-
-    def create_optimizer(self):
-        if (
-            self.args.loraplus_lr_ratio is None
-            and self.args.embedding_lr_scale is None
-            and self.args.embedding_lr is None
-            and self.args.alternate_optimizer
-            not in [
-                "optimi_adamw",
-                "ao_adamw_8bit",
-                "ao_adamw_4bit",
-                "ao_adamw_fp8",
-                "adopt_adamw",
-            ]
-        ):
-            return super().create_optimizer()
-
-        opt_model = self.model_wrapped if is_sagemaker_mp_enabled() else self.model
-        if self.optimizer is None:  # pylint: disable=access-member-before-definition
-            decay_parameters = self.get_decay_parameter_names(opt_model)
-            params = {
-                "to_weight_decay": {},  # LayerNorm and bias
-                "embeddings": {},  # lm_head, embed_tokens,
-                "no_weight_decay": {},
-            }
-
-            optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(
-                self.args,
-                opt_model,
-            )
-
-            for name, param in opt_model.named_parameters():
-                if not param.requires_grad:
-                    continue
-                if name.endswith("modules_to_save.default.weight") or any(
-                    embed_name in name for embed_name in ["embed_tokens", "lm_head"]
-                ):
-                    params["embeddings"][name] = param
-                elif name in decay_parameters:
-                    params["to_weight_decay"][name] = param
-                else:
-                    params["no_weight_decay"][name] = param
-            optimizer_grouped_parameters = []
-            if params["to_weight_decay"]:
-                optimizer_grouped_parameters.append(
-                    {
-                        "params": list(params["to_weight_decay"].values()),
-                        "weight_decay": self.args.weight_decay,
-                        "lr": optimizer_kwargs["lr"],
-                    }
-                )
-            if params["embeddings"]:
-                lr = optimizer_kwargs["lr"]  # pylint: disable=invalid-name
-                if self.args.embedding_lr_scale:
-                    lr *= self.args.embedding_lr_scale  # pylint: disable=invalid-name
-                elif self.args.embedding_lr:
-                    lr = self.args.embedding_lr  # pylint: disable=invalid-name
-                optimizer_grouped_parameters.append(
-                    {
-                        "params": list(params["embeddings"].values()),
-                        "weight_decay": 0.0,
-                        "lr": lr,
-                    }
-                )
-            if params["no_weight_decay"]:
-                optimizer_grouped_parameters.append(
-                    {
-                        "params": list(params["no_weight_decay"].values()),
-                        "weight_decay": 0.0,
-                        "lr": optimizer_kwargs["lr"],
-                    }
-                )
-
-            if self.args.loraplus_lr_ratio is not None:
-                loraplus_lr_ratio = getattr(self.args, "loraplus_lr_ratio", None)
-                loraplus_lr_embedding = getattr(
-                    self.args, "loraplus_lr_embedding", 1e-6
-                )
-                self.optimizer = create_loraplus_optimizer(  # pylint: disable=attribute-defined-outside-init
-                    opt_model,
-                    optimizer_cls,
-                    loraplus_lr_ratio=loraplus_lr_ratio,
-                    loraplus_lr_embedding=loraplus_lr_embedding,
-                    **optimizer_kwargs,
-                )
-            elif (
-                self.args.embedding_lr_scale is not None
-                or self.args.embedding_lr is not None
-            ):
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
-                )
-            elif self.args.alternate_optimizer == "optimi_adamw":
-                from optimi import AdamW
-
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    AdamW(
-                        optimizer_grouped_parameters, foreach=False, **optimizer_kwargs
-                    )
-                )
-            elif self.args.alternate_optimizer == "ao_adamw_4bit":
-                from torchao.prototype.low_bit_optim import AdamW4bit
-
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    AdamW4bit(optimizer_grouped_parameters, **optimizer_kwargs)
-                )
-            elif self.args.alternate_optimizer == "ao_adamw_8bit":
-                from torchao.prototype.low_bit_optim import AdamW8bit
-
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    AdamW8bit(optimizer_grouped_parameters, **optimizer_kwargs)
-                )
-            elif self.args.alternate_optimizer == "ao_adamw_fp8":
-                from torchao.prototype.low_bit_optim import AdamWFp8
-
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    AdamWFp8(optimizer_grouped_parameters, **optimizer_kwargs)
-                )
-            elif self.args.alternate_optimizer == "adopt_adamw":
-                from axolotl.utils.optimizers.adopt import ADOPT
-
-                self.optimizer = (  # pylint: disable=attribute-defined-outside-init
-                    ADOPT(
-                        optimizer_grouped_parameters,
-                        decouple=True,
-                        **optimizer_kwargs,
-                    )
-                )
-
-        if is_sagemaker_mp_enabled():
-            self.optimizer = smp.DistributedOptimizer(  # pylint: disable=attribute-defined-outside-init
-                self.optimizer
-            )
-
-        return self.optimizer
-
-    def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
-        if self.args.sample_packing and not self.args.pretraining:
-            if self.args.multipack_real_batches:
-                batch_size = self.args.per_device_train_batch_size
-                batch_max_len = self.args.max_seq_length
-            else:
-                batch_size = 1
-                train_batch_size = (
-                    self.state.train_batch_size or self.args.per_device_train_batch_size
-                )
-                batch_max_len = train_batch_size * self.args.max_seq_length
-
-            if self.args.curriculum_sampling:
-                sampler = SequentialSampler(self.train_dataset)
-            else:
-                sampler = RandomSampler(self.train_dataset)
-
-            return MultipackBatchSampler(
-                sampler,
-                lengths=get_dataset_lengths(self.train_dataset),
-                packing_efficiency_estimate=self.args.sample_packing_efficiency,
-                batch_max_len=batch_max_len,
-                batch_size=batch_size,
-                group_size=self.args.sample_packing_group_size,
-                bin_size=self.args.sample_packing_bin_size,
-                drop_last=True,
-            )
-        if self.args.curriculum_sampling:
-            return SequentialSampler(self.train_dataset)
-        return super()._get_train_sampler()
-
-    def _get_eval_sampler(
-        self, eval_dataset: Dataset
-    ) -> Optional[torch.utils.data.Sampler]:
-        if self.args.sample_packing and self.args.eval_sample_packing is not False:
-            if self.args.multipack_real_batches:
-                batch_size = self.args.per_device_eval_batch_size
-                batch_max_len = self.args.max_seq_length
-            else:
-                batch_size = 1
-                batch_max_len = (
-                    self.args.per_device_eval_batch_size * self.args.max_seq_length
-                )
-            return MultipackBatchSampler(
-                SequentialSampler(eval_dataset),
-                lengths=get_dataset_lengths(self.eval_dataset),
-                packing_efficiency_estimate=self.args.sample_packing_efficiency,
-                batch_max_len=batch_max_len,
-                batch_size=batch_size,
-                group_size=self.args.sample_packing_group_size,
-                bin_size=self.args.sample_packing_bin_size,
-                drop_last=True,
-            )
-        return super()._get_eval_sampler(eval_dataset)
-
-    def get_train_dataloader(self) -> DataLoader:
-        if self.args.sample_packing and not self.args.pretraining:
-            train_dataset = self.train_dataset
-            if "length" in train_dataset.features.keys():
-                train_dataset = train_dataset.remove_columns(["length"])
-            data_collator = self.data_collator
-            dataloader_params = {
-                "batch_size": self._train_batch_size,
-                "collate_fn": data_collator,
-                "num_workers": self.args.dataloader_num_workers,
-                "pin_memory": self.args.dataloader_pin_memory,
-            }
-            if self.args.dataloader_prefetch_factor:
-                dataloader_params[
-                    "prefetch_factor"
-                ] = self.args.dataloader_prefetch_factor
-
-            sampler = self._get_train_sampler()
-            if isinstance(sampler, BatchSampler):
-                dataloader_params["batch_sampler"] = sampler
-                del dataloader_params["batch_size"]
-            else:
-                dataloader_params["sampler"] = sampler
-                dataloader_params["drop_last"] = self.args.dataloader_drop_last
-            dataloader_params["worker_init_fn"] = seed_worker
-
-            self.accelerator.even_batches = False
-            return self.accelerator.prepare_data_loader(
-                DataLoader(train_dataset, **dataloader_params)
-            )
-        return super().get_train_dataloader()
-
-    def get_eval_dataloader(self, eval_dataset: Optional[Dataset] = None) -> DataLoader:
-        if self.args.sample_packing and self.args.eval_sample_packing is False:
-            self.data_collator = (  # pylint: disable=attribute-defined-outside-init
-                self.eval_data_collator
-            )
-            if eval_dataset:
-                eval_dataset = eval_dataset.remove_columns(["length"])
-            dataloader = super().get_eval_dataloader(eval_dataset)
-            self.data_collator = (  # pylint: disable=attribute-defined-outside-init
-                self.train_data_collator
-            )
-            return dataloader
-
-        if self.args.sample_packing and self.args.eval_sample_packing is not False:
-            eval_dataset = (
-                eval_dataset if eval_dataset is not None else self.eval_dataset
-            )
-
-            eval_sampler = self._get_eval_sampler(eval_dataset)
-            eval_dataset = eval_dataset.remove_columns(["length"])
-            data_collator = self.data_collator
-            dataloader_params = {
-                "batch_size": self.args.eval_batch_size,
-                "collate_fn": data_collator,
-                "num_workers": self.args.dataloader_num_workers,
-                "pin_memory": self.args.dataloader_pin_memory,
-            }
-            if self.args.dataloader_prefetch_factor:
-                dataloader_params[
-                    "prefetch_factor"
-                ] = self.args.dataloader_prefetch_factor
-
-            if isinstance(eval_sampler, BatchSampler):
-                dataloader_params["batch_sampler"] = eval_sampler
-                del dataloader_params["batch_size"]
-            else:
-                dataloader_params["sampler"] = eval_sampler
-                dataloader_params["drop_last"] = self.args.dataloader_drop_last
-
-            self.accelerator.even_batches = False
-            return self.accelerator.prepare_data_loader(
-                DataLoader(eval_dataset, **dataloader_params)
-            )
-
-        return super().get_eval_dataloader(eval_dataset)
-
-    def _get_bench_sampler(
-        self, bench_dataset: Dataset
-    ) -> Optional[torch.utils.data.Sampler]:
-        if self.args.world_size <= 1:
-            return SequentialSampler(bench_dataset)
-        return None
-
-    def get_bench_dataloader(
-        self,
-        bench_dataset: Dataset,
-    ) -> DataLoader:
-        dataloader_params = {
-            "batch_size": self.args.eval_batch_size,
-            "collate_fn": self.bench_data_collator,
-            "num_workers": self.args.dataloader_num_workers,
-            "pin_memory": self.args.dataloader_pin_memory,
-        }
-        if self.args.dataloader_prefetch_factor:
-            dataloader_params["prefetch_factor"] = self.args.dataloader_prefetch_factor
-
-        if not isinstance(bench_dataset, torch.utils.data.IterableDataset):
-            dataloader_params["sampler"] = self._get_bench_sampler(bench_dataset)
-            dataloader_params["drop_last"] = self.args.dataloader_drop_last
-
-        return DataLoader(bench_dataset, **dataloader_params)
-        # return self.accelerator.prepare(DataLoader(bench_dataset, **dataloader_params))
-
-    def compute_loss(
-        self, model, inputs, return_outputs=False, num_items_in_batch=None
-    ):
-        # use one's weighted cross entropy loss calc
-        # if self.args.sample_packing:
-        #     labels = inputs.pop("labels")
-        #     outputs = model(**inputs)
-        #     loss = trainer_weighted_loss(outputs, labels, shift_labels=True)
-        #     return (loss, outputs) if return_outputs else loss
-        if self.args.orpo_alpha:
-            return self.orpo_compute_loss(
-                model,
-                inputs,
-                return_outputs=return_outputs,
-                num_items_in_batch=num_items_in_batch,
-            )
-        return super().compute_loss(
-            model,
-            inputs,
-            return_outputs=return_outputs,
-            num_items_in_batch=num_items_in_batch,
-        )
-
-    @staticmethod
-    def orpo_concatenate_inputs(inputs, label_pad_token=-100, pad_token=0, device=None):
-        concatenated_batch = {}
-
-        max_length = max(
-            inputs["input_ids"].shape[1], inputs["rejected_input_ids"].shape[1]
-        )
-        # Concatenate positive and negative inputs
-        concatenated_batch["input_ids"] = pad_to_length(
-            inputs["input_ids"], max_length, pad_token
-        )
-        concatenated_batch["rejected_input_ids"] = pad_to_length(
-            inputs["rejected_input_ids"], max_length, pad_token
-        )
-        concatenated_batch["labels"] = pad_to_length(
-            inputs["labels"], max_length, label_pad_token
-        )
-        concatenated_batch["rejected_labels"] = pad_to_length(
-            inputs["rejected_labels"], max_length, label_pad_token
-        )
-        concatenated_batch["attention_mask"] = pad_to_length(
-            inputs["attention_mask"], max_length, 0
-        )
-        concatenated_batch["rejected_attention_mask"] = pad_to_length(
-            inputs["rejected_attention_mask"], max_length, 0
-        )
-        concatenated_batch["prompt_attention_mask"] = pad_to_length(
-            inputs["prompt_attention_mask"], max_length, 0
-        ).to(device=device)
-
-        input_ids = torch.cat(
-            [concatenated_batch["input_ids"], concatenated_batch["rejected_input_ids"]],
-            dim=0,
-        ).to(device=device)
-        attention_mask = torch.cat(
-            [
-                concatenated_batch["attention_mask"],
-                concatenated_batch["rejected_attention_mask"],
-            ],
-            dim=0,
-        ).to(device=device)
-        labels = torch.cat(
-            [concatenated_batch["labels"], concatenated_batch["rejected_labels"]], dim=0
-        ).to(device=device)
-
-        return {
-            "input_ids": input_ids,
-            "labels": labels,
-            "attention_mask": attention_mask,
-            "prompt_attention_mask": concatenated_batch["prompt_attention_mask"],
-        }
-
-    def orpo_compute_custom_loss(self, logits, labels):
-        logits = logits.contiguous()
-        loss = 0.0
-
-        if labels is not None:
-            # move labels to correct device to enable model parallelism
-            labels = labels.to(logits.device)
-            # Shift so that tokens < n predict n
-            shift_logits = logits[..., :-1, :].contiguous()
-            shift_labels = labels[..., 1:].contiguous()
-
-            # Flatten the tokens
-            loss = self.loss_fct(shift_logits.transpose(2, 1), shift_labels).mean(
-                dim=-1
-            )
-
-        return loss
-
-    def orpo_compute_logps(
-        self, prompt_attention_mask, chosen_inputs, chosen_attention_mask, logits
-    ):
-        # Get the shape of chosen_attention_mask[:, :-1]
-        chosen_shape = chosen_attention_mask[:, :-1].shape
-
-        # Calculate the padding size
-        pad_length = chosen_shape[1] - (prompt_attention_mask.shape[1] - 1)
-
-        # Pad prompt_attention_mask with zeros to match the desired shape
-        prompt_attention_mask_padded = torch.nn.functional.pad(
-            prompt_attention_mask[:, 1:], (0, pad_length), mode="constant", value=0
-        )
-
-        # Perform the subtraction operation
-        mask = chosen_attention_mask[:, :-1] > prompt_attention_mask_padded
-
-        per_token_logps = torch.gather(
-            logits[:, :-1, :].log_softmax(-1),
-            dim=2,
-            index=(mask * chosen_inputs[:, 1:]).unsqueeze(2),
-        ).squeeze(2)
-        return torch.mul(per_token_logps, mask).sum(dim=1) / mask.sum(dim=1)
-
-    def orpo_compute_loss(
-        self,
-        model,
-        inputs,
-        return_outputs=False,
-        num_items_in_batch=None,  # pylint: disable=unused-argument
-    ):
-        concat_inputs = AxolotlTrainer.orpo_concatenate_inputs(
-            inputs,
-            label_pad_token=-100,
-            pad_token=self.tokenizer.pad_token_id,
-            device=self.accelerator.device,
-        )
-
-        # Perform a single forward pass
-        outputs = model(
-            **{
-                "input_ids": concat_inputs["input_ids"],
-                "attention_mask": concat_inputs["attention_mask"],
-                "labels": concat_inputs["labels"],
-            },
-            output_hidden_states=True,
-        )
-
-        # Split the outputs for positive and negative examples
-        outputs_pos, outputs_neg = outputs.logits.chunk(2)
-
-        # Calculate NLL loss
-        pos_loss = self.orpo_compute_custom_loss(
-            logits=outputs_pos, labels=concat_inputs["input_ids"].chunk(2)[0]
-        )
-
-        # Calculate Log Probability
-        pos_prob = self.orpo_compute_logps(
-            prompt_attention_mask=concat_inputs["prompt_attention_mask"],
-            chosen_inputs=concat_inputs["input_ids"].chunk(2)[0],
-            chosen_attention_mask=concat_inputs["attention_mask"].chunk(2)[0],
-            logits=outputs_pos,
-        )
-        neg_prob = self.orpo_compute_logps(
-            prompt_attention_mask=concat_inputs["prompt_attention_mask"],
-            chosen_inputs=concat_inputs["input_ids"].chunk(2)[1],
-            chosen_attention_mask=concat_inputs["attention_mask"].chunk(2)[1],
-            logits=outputs_neg,
-        )
-
-        # Calculate log odds
-        log_odds = (pos_prob - neg_prob) - (
-            torch.log(1 - torch.exp(pos_prob)) - torch.log(1 - torch.exp(neg_prob))
-        )
-        sig_ratio = torch.nn.functional.sigmoid(log_odds)
-        ratio = torch.log(sig_ratio)
-
-        # Calculate the Final Loss
-        loss = torch.mean(pos_loss - self.args.orpo_alpha * ratio).to(
-            dtype=torch.bfloat16
-        )
-
-        metrics = {}
-        metrics["chosen_geometric_mean"] = torch.mean(pos_prob).cpu().item()
-        metrics["rejected_geometric_mean"] = torch.mean(neg_prob).cpu().item()
-        metrics["log_odds_ratio"] = torch.mean(ratio).cpu().item()
-        metrics["log_odds"] = torch.mean(log_odds).cpu().item()
-        self.store_metrics(metrics, train_eval="train")
-
-        return (loss, outputs_pos) if return_outputs else loss
-
-    @wraps(Trainer.push_to_hub)
-    def push_to_hub(self, *args, **kwargs) -> str:
-        """
-        Overwrite the `push_to_hub` method in order to force-add the tags when pushing the
-        model on the Hub. Please refer to `~transformers.Trainer.push_to_hub` for more details.
-        """
-        kwargs = _sanitize_kwargs_for_ds_tagging(
-            dataset_tags=self.dataset_tags, kwargs=kwargs
-        )
-        kwargs = _sanitize_kwargs_for_tagging(tag_names=self.tag_names, kwargs=kwargs)
-
-        return super().push_to_hub(*args, **kwargs)
-
-    @wraps(Trainer.create_accelerator_and_postprocess)
-    def create_accelerator_and_postprocess(self):
-        res = super().create_accelerator_and_postprocess()
-
-        if self.is_fsdp_enabled:
-            if (
-                "limit_all_gathers" in self.args.fsdp_config
-                and self.args.fsdp_config["limit_all_gathers"]
-            ):
-                self.accelerator.state.fsdp_plugin.limit_all_gathers = True
-
-        return res
-
-    def log(self, logs: Dict[str, float], start_time: Optional[float] = None) -> None:
-        """
-        Log `logs` on the various objects watching training, including stored metrics.
-
-        Args:
-            logs (`Dict[str, float]`):
-                The values to log.
-            start_time (`Optional[float]`):
-                The start of training.
-        """
-        # logs either has 'loss' or 'eval_loss'
-        train_eval = "train" if "loss" in logs else "eval"
-        # Add averaged stored metrics to logs
-        for key, metrics in self._stored_metrics[train_eval].items():
-            logs[key] = torch.tensor(metrics).mean().item()
-        del self._stored_metrics[train_eval]
-
-        return super().log(logs, start_time)
-
-    def store_metrics(
-        self, metrics: Dict[str, float], train_eval: Literal["train", "eval"] = "train"
-    ) -> None:
-        for key, value in metrics.items():
-            self._stored_metrics[train_eval][key].append(value)
-
-    def _save_checkpoint(self, model, trial, **kwargs):
-        # make sure the checkpoint dir exists, since trainer is flakey
-        checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}"
-        run_dir = self._get_output_dir(trial=trial)
-        output_dir = os.path.join(run_dir, checkpoint_folder)
-        os.makedirs(output_dir, exist_ok=True)
-        return super()._save_checkpoint(model, trial, **kwargs)
-
-
-class AxolotlMambaTrainer(AxolotlTrainer):
-    """
-    Mamba specific trainer to handle loss calculation
-    """
-
-    tag_names = ["axolotl", "mamba"]
-
-    def compute_loss(
-        self,
-        model,
-        inputs,
-        return_outputs=False,  # pylint: disable=unused-argument
-        num_items_in_batch=None,  # pylint: disable=unused-argument
-    ):
-        input_ids = inputs.pop("input_ids")
-        lm_logits = model(input_ids).logits
-
-        labels = input_ids.to(lm_logits.device)
-        shift_logits = lm_logits[:, :-1, :].contiguous()
-        labels = labels[:, 1:].contiguous()
-
-        loss_fct = torch.nn.CrossEntropyLoss()
-        lm_loss = loss_fct(
-            shift_logits.view(-1, shift_logits.size(-1)), labels.view(-1)
-        )
-
-        return lm_loss
-
-
-class ReLoRATrainer(AxolotlTrainer):
-    """
-    Trainer subclass that uses the OneCycleLR scheduler
-    """
-
-    tag_names = ["axolotl", "relora"]
-
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.lr_scheduler = None
-
-    def create_scheduler(
-        self,
-        num_training_steps: int,
-        optimizer: Optional[torch.optim.Optimizer] = None,
-    ):
-        optimizer = self.optimizer if optimizer is None else optimizer
-        lr_scheduler = super().create_scheduler(num_training_steps, optimizer)
-
-        if self.args.relora_steps:
-            warmup_steps = (
-                self.args.relora_warmup_steps if self.args.relora_warmup_steps else 10
-            )
-            anneal_steps = (
-                self.args.relora_anneal_steps if self.args.relora_anneal_steps else 1
-            )
-            self.lr_scheduler = ReLoRAScheduler(
-                optimizer,
-                lr_scheduler,
-                self.args.relora_steps,
-                anneal_steps,
-                warmup_steps,
-            )
-        else:
-            self.lr_scheduler = lr_scheduler
-
-        return self.lr_scheduler
-
-
-class AxolotlDPOTrainer(SchedulerMixin, DPOTrainer):
-    """
-    Extend the base DPOTrainer for axolotl helpers
-    """
-
-    tag_names = ["axolotl", "dpo"]
-
-    def __init__(self, *args, dataset_tags=None, **kwargs):
-        super().__init__(*args, **kwargs)
-        self.dataset_tags = dataset_tags
-        self.optimizer = None
-
-    def create_optimizer(self):
-        if self.args.loraplus_lr_ratio is None:
-            return super().create_optimizer()
-
-        opt_model = self.model_wrapped if is_sagemaker_mp_enabled() else self.model
-        if self.optimizer is None:  # pylint: disable=access-member-before-definition
-            optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(
-                self.args,
-                opt_model,
-            )
-
-            loraplus_lr_ratio = getattr(self.args, "loraplus_lr_ratio", None)
-            if loraplus_lr_ratio:
-                print("Using lora+")
-            loraplus_lr_embedding = getattr(self.args, "loraplus_lr_embedding", None)
-            self.optimizer = create_loraplus_optimizer(  # pylint: disable=attribute-defined-outside-init
-                opt_model,
-                optimizer_cls,
-                loraplus_lr_ratio=loraplus_lr_ratio,
-                loraplus_lr_embedding=loraplus_lr_embedding,
-                **optimizer_kwargs,
-            )
-
-        if is_sagemaker_mp_enabled():
-            self.optimizer = smp.DistributedOptimizer(  # pylint: disable=attribute-defined-outside-init
-                self.optimizer
-            )
-
-        return self.optimizer
-
-    @wraps(DPOTrainer.push_to_hub)
-    def push_to_hub(self, *args, **kwargs) -> str:
-        """
-        Overwrite the `push_to_hub` method in order to force-add the tags when pushing the
-        model on the Hub. Please refer to `~transformers.Trainer.push_to_hub` for more details.
-        """
-        kwargs = _sanitize_kwargs_for_ds_tagging(
-            dataset_tags=self.dataset_tags, kwargs=kwargs
-        )
-        kwargs = _sanitize_kwargs_for_tagging(tag_names=self.tag_names, kwargs=kwargs)
-
-        return super().push_to_hub(*args, **kwargs)
-
-    @staticmethod
-    def tokenize_row(
-        features,
-        processing_class,
-        max_prompt_length,
-        max_completion_length,
-        add_special_tokens,
-    ) -> Dict:
-        res = DPOTrainer.tokenize_row(
-            features,
-            processing_class,
-            max_prompt_length,
-            max_completion_length,
-            add_special_tokens,
-        )
-        # fix when the tokenizer doesn't have a bos_token_id, e.g. Qwen
-        if processing_class.bos_token is None and res["prompt_input_ids"][0] is None:
-            for key in res.keys():
-                res[key] = res[key][1:]
-
-        if processing_class.bos_token and processing_class.bos_token_id is not None:
-            # dpo trainer may incorrectly prepend the bos_token_id to the dpo outputs
-            if res["chosen_input_ids"][0] == processing_class.bos_token_id:
-                res["chosen_input_ids"] = res["chosen_input_ids"][1:]
-                res["chosen_labels"] = res["chosen_labels"][1:]
-                res["chosen_attention_mask"] = res["chosen_attention_mask"][1:]
-            if res["rejected_input_ids"][0] == processing_class.bos_token_id:
-                res["rejected_input_ids"] = res["rejected_input_ids"][1:]
-                res["rejected_labels"] = res["rejected_labels"][1:]
-                res["rejected_attention_mask"] = res["rejected_attention_mask"][1:]
-
-        return res
-
-    def training_step(
-        self,
-        model: nn.Module,
-        inputs: Dict[str, Union[torch.Tensor, Any]],
-        num_items_in_batch=None,
-    ) -> torch.Tensor:
-        loss: torch.Tensor = super().training_step(model, inputs, num_items_in_batch)
-        gc.collect()
-        torch.cuda.empty_cache()
-        return loss
-
-
-class AxolotlORPOTrainer(SchedulerMixin, ORPOTrainer):
-    """
-    Extend the base ORPOTrainer for axolotl helpers
-    """
-
-    tag_names = ["axolotl", "orpo"]
-
-
-class AxolotlKTOTrainer(SchedulerMixin, KTOTrainer):
-    """
-    Extend the base KTOTrainer for axolotl helpers
-    """
-
-    tag_names = ["axolotl", "kto"]
-
-
-class AxolotlCPOTrainer(SchedulerMixin, CPOTrainer):
-    """
-    Extend the base CPOTrainer for axolotl helpers
-    """
-
-    tag_names = ["axolotl", "cpo"]
-
-
-class AxolotlRewardTrainer(SchedulerMixin, RewardTrainer):
-    """
-    Extend the base RewardTrainer for axolotl helpers
-    """
-
-    tag_names = ["axolotl", "reward"]

src/axolotl/core/training_args.py

@@ -1,239 +0,0 @@
-"""
-extra axolotl specific training args
-"""
-from dataclasses import dataclass, field
-from typing import Optional
-
-from transformers import TrainingArguments
-from trl import CPOConfig, DPOConfig, KTOConfig, ORPOConfig, RewardConfig
-
-
-@dataclass
-class AxolotlTrainingMixins:
-    """
-    Mixin class for the Axolotl training args.
-    """
-
-    # pylint: disable=duplicate-code
-    model_type: Optional[str] = field(
-        default=None, metadata={"help": "HF model configuration model_type."}
-    )
-    lr_quadratic_warmup: bool = field(
-        default=False,
-        metadata={"help": "Use quadratic warmup for cosine scheduling."},
-    )
-    pretraining: bool = field(
-        default=False,
-        metadata={
-            "help": "Indicates to trainer whether we are doing continued pretraining."
-        },
-    )
-    sample_packing: bool = field(
-        default=False,
-        metadata={"help": "Use sample packing for efficient training."},
-    )
-    multipack_real_batches: bool = field(
-        default=False,
-        metadata={"help": "Use real batches for efficient training."},
-    )
-    eval_sample_packing: Optional[bool] = field(
-        default=None,
-        metadata={"help": "Use sample packing for efficient evals."},
-    )
-    sample_packing_efficiency: float = field(
-        default=1.0,
-        metadata={"help": "Sample packing efficiency for calculating batch length."},
-    )
-    sample_packing_bin_size: int = field(
-        default=200,
-        metadata={
-            "help": "The max number of samples that packed sample can contain after packing. Increase for better packing."
-        },
-    )
-    sample_packing_group_size: int = field(
-        default=100000,
-        metadata={
-            "help": "The number of samples to group together for packing. Increase for better packing."
-        },
-    )
-    max_seq_length: int = field(
-        default=2048,
-        metadata={"help": "The maximum sequence length the model can handle"},
-    )
-    relora_steps: Optional[int] = field(
-        default=None,
-        metadata={"help": "how often to reset for ReLoRA"},
-    )
-    relora_warmup_steps: Optional[int] = field(
-        default=None,
-        metadata={"help": "how many warmup steps to take after reset for ReLoRA"},
-    )
-    relora_anneal_steps: Optional[int] = field(
-        default=None,
-        metadata={"help": "how many warmup steps to take after reset for ReLoRA"},
-    )
-    relora_prune_ratio: Optional[float] = field(
-        default=0.9,
-        metadata={"help": "prune ratio for magnitude pruning of the optimizer"},
-    )
-    bench_split: Optional[str] = field(
-        default="eval", metadata={"help": "The benchmark split to run on"}
-    )
-    bench_dataset: Optional[str] = field(
-        default="pharaouk/dharma-1/dharma_1_mini.json",
-        metadata={
-            "help": "Benchmark dataset to use: options are `mmlu-zs`, `mmlu-fs`, or the full path to the dataset file"
-        },
-    )
-    do_bench_eval: Optional[bool] = field(
-        default=False, metadata={"help": "Whether to run the Benchmark evaluation."}
-    )
-    do_causal_lm_eval: Optional[bool] = field(
-        default=False, metadata={"help": "Whether to run the Causal LM evaluation."}
-    )
-    max_bench_samples: Optional[int] = field(
-        default=None,
-        metadata={
-            "help": "If set, only evaluates on `max_bench_samples` of the benchmark dataset."
-        },
-    )
-    bench_source_max_len: int = field(
-        default=2048, metadata={"help": "Maximum source sequence length for bench."}
-    )
-    dataloader_prefetch_factor: Optional[int] = field(
-        default=None,
-        metadata={"help": "prefetch_factor argument to the dataloader"},
-    )
-    cosine_min_lr_ratio: Optional[float] = field(
-        default=None,
-        metadata={"help": "Minimum learning rate is min_lr_ratio * learning_rate"},
-    )
-    cosine_constant_lr_ratio: Optional[float] = field(
-        default=None,
-        metadata={
-            "help": "Starting constant learning rate step is cosine_constant_lr_ratio * max_steps"
-        },
-    )
-    loraplus_lr_ratio: Optional[float] = field(
-        default=None, metadata={"help": "loraplus learning rate ratio lr_B / lr_A."}
-    )
-    loraplus_lr_embedding: Optional[float] = field(
-        default=1e-6,
-        metadata={"help": "loraplus learning rate for lora embedding layers."},
-    )
-    embedding_lr_scale: Optional[float] = field(
-        default=None,
-        metadata={"help": "Scale the learning rate for the embedding layers."},
-    )
-    embedding_lr: Optional[float] = field(
-        default=None,
-        metadata={"help": "absolute learning rate for the embedding layers."},
-    )
-    qlora: bool = field(
-        default=False,
-        metadata={"help": "whether this is a qlora training"},
-    )
-    orpo_alpha: Optional[float] = field(
-        default=None,
-    )
-    lisa_n_layers: Optional[int] = field(
-        default=None,
-        metadata={"help": "the number of activate layers in LISA"},
-    )
-    lisa_step_interval: Optional[int] = field(
-        default=None,
-        metadata={"help": "how often to switch layers in LISA"},
-    )
-    lisa_layers_attribute: Optional[str] = field(
-        default=None,
-        metadata={"help": "path under the model to access the layers"},
-    )
-    curriculum_sampling: Optional[bool] = field(
-        default=None,
-        metadata={"help": "whether to use sequential sampling for curriculum learning"},
-    )
-    alternate_optimizer: Optional[str] = field(
-        default=None,
-        metadata={
-            "help": "workaround to pass an alternate optimizer to the HF trainer"
-        },
-    )
-    alternate_lr_scheduler_type: Optional[str] = field(
-        default=None,
-        metadata={
-            "help": "workaround to pass an alternate lr scheduler to the HF trainer"
-        },
-    )
-    chat_template: Optional[str] = field(
-        default=None,
-        metadata={"help": "Chat template converting chat messages to text"},
-    )
-
-    kd_ce_alpha: Optional[float] = field(
-        default=None,
-        metadata={
-            "help": "The alpha scaling parameter for SFT cross entropy loss when using KD"
-        },
-    )
-
-    kd_alpha: Optional[float] = field(
-        default=1.0,
-        metadata={"help": "The alpha scaling parameter for KD loss"},
-    )
-
-    kd_temperature: Optional[float] = field(
-        default=1.0,
-        metadata={
-            "help": "the temperature parameter for KL divergence loss when using KD"
-        },
-    )
-
-
-@dataclass
-class AxolotlTrainingArguments(AxolotlTrainingMixins, TrainingArguments):
-    """
-    Training arguments for Causal trainer
-
-    This code is duplicated due to HF TrainingArguments not setting output_dir with a defaujlt value
-    so it can't be used as a mixin.
-    """
-
-
-@dataclass
-class AxolotlDPOConfig(AxolotlTrainingMixins, DPOConfig):
-    """
-    DPO config for DPO training
-    """
-
-
-@dataclass
-class AxolotlORPOConfig(AxolotlTrainingMixins, ORPOConfig):
-    """
-    ORPO config for ORPO training
-    """
-
-
-@dataclass
-class AxolotlKTOConfig(AxolotlTrainingMixins, KTOConfig):
-    """
-    KTO config for KTO training
-    """
-
-
-@dataclass
-class AxolotlCPOConfig(AxolotlTrainingMixins, CPOConfig):
-    """
-    CPO config for CPO training
-    """
-
-    simpo_gamma: Optional[float] = field(
-        default=None,
-        metadata={"help": "simpo gamma parameter"},
-    )
-
-
-@dataclass
-class AxolotlRewardConfig(AxolotlTrainingMixins, RewardConfig):
-    """
-    Reward config for Reward training
-    """

src/axolotl/datasets.py

@@ -2,7 +2,7 @@
 
 import logging
 import os
-from typing import List, Optional, Union
+from typing import List, Optional
 
 import torch
 from datasets import Dataset, IterableDataset
@@ -51,13 +51,7 @@ class TokenizedPromptDataset(Dataset):
         map_kwargs = {}
         if self.prompt_tokenizer.supports_batched:
             map_kwargs["batched"] = True
-            map_kwargs["batch_size"] = 1_000
-        if self.prompt_tokenizer.filter_rows:
-            dataset = dataset.filter(
-                self.prompt_tokenizer.filter_rows,
-                num_proc=num_proc,
-                desc="Strategy Filtering Rows",
-            )
+            map_kwargs["batch_size"] = 100
         return dataset.map(
             self.prompt_tokenizer.tokenize_prompt,
             num_proc=num_proc,
@@ -68,24 +62,6 @@ class TokenizedPromptDataset(Dataset):
         )
 
 
-def wrap_dataset_for_tokenized_prompt(
-    prompt_tokenizer: PromptTokenizingStrategy,
-    dataset: Union[Dataset, IterableDataset],
-    **kwargs,
-):
-    if isinstance(dataset, IterableDataset):
-        map_kwargs = {}
-        if prompt_tokenizer.supports_batched:
-            map_kwargs["batched"] = True
-        features = dataset.features.keys()
-        return dataset.map(
-            prompt_tokenizer.tokenize_prompt,
-            remove_columns=features,
-            **map_kwargs,
-        )
-    return TokenizedPromptDataset(prompt_tokenizer, dataset, **kwargs)
-
-
 # TODO this isn't the best since it can't interleave datasets
 class ConstantLengthDataset(IterableDataset):
     """

src/axolotl/evaluate.py

@@ -9,12 +9,11 @@ from typing import Dict, Optional
 import torch
 from accelerate.logging import get_logger
 
-from axolotl.common.cli import TrainerCliArgs
+from axolotl.common.cli import EvaluateCliArgs, load_model_and_tokenizer
 from axolotl.logging_config import configure_logging
 from axolotl.train import TrainDatasetMeta
-from axolotl.utils import set_pytorch_cuda_alloc_conf
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.models import load_model, load_processor, load_tokenizer
+from axolotl.utils.models import load_processor
 from axolotl.utils.trainer import setup_trainer
 
 project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
@@ -62,8 +61,9 @@ def evaluate_dataset(
     return metrics
 
 
+# pylint: disable=duplicate-code
 def evaluate(
-    *, cfg: DictDefault, cli_args: TrainerCliArgs, dataset_meta: TrainDatasetMeta
+    *, cfg: DictDefault, cli_args: EvaluateCliArgs, dataset_meta: TrainDatasetMeta
 ) -> Dict[str, float]:
     """
     Evaluate a model on training and validation datasets
@@ -79,16 +79,11 @@ def evaluate(
         - The tokenizer
         - Dictionary of evaluation metrics
     """
-    # pylint: disable=duplicate-code
-    # Enable expandable segments for cuda allocation to improve VRAM usage
-    set_pytorch_cuda_alloc_conf()
+    # Load model
+    LOG.debug("loading model for evaluation...")
 
-    # Load tokenizer
-    LOG.debug(
-        f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}",
-        main_process_only=True,
-    )
-    tokenizer = load_tokenizer(cfg)
+    model, tokenizer = load_model_and_tokenizer(cfg=cfg, cli_args=cli_args)
+    model = model.to(cfg.device, dtype=cfg.torch_dtype)
 
     # Load processor for multimodal models if needed
     processor = None
@@ -100,12 +95,6 @@ def evaluate(
     eval_dataset = dataset_meta.eval_dataset
     total_num_steps = dataset_meta.total_num_steps
 
-    # Load model
-    LOG.debug("loading model for evaluation...")
-    model, _ = load_model(
-        cfg, tokenizer, processor=processor, inference=cli_args.inference
-    )
-
     # Set up trainer
     trainer = setup_trainer(
         cfg,

src/axolotl/integrations/base.py

@@ -111,17 +111,6 @@ class BasePlugin:
         None
         """
 
-    def get_trainer_cls(self, cfg):  # pylint: disable=unused-argument):
-        """
-        Returns a custom class for the trainer.
-
-        Parameters:
-        cfg (dict): The global axolotl configuration.
-
-        Returns:
-        class: The class for the trainer.
-        """
-
     def create_optimizer(self, cfg, trainer):  # pylint: disable=unused-argument
         """
         Creates and returns an optimizer for training.
@@ -223,17 +212,7 @@ def load_plugin(plugin_name: str) -> BasePlugin:
     module_name, class_name = plugin_name.rsplit(".", 1)
 
     # import the module
-    try:
-        module = importlib.import_module(module_name)
-    except ModuleNotFoundError as orig_exc:
-        try:
-            if not module_name.startswith("axolotl.integrations."):
-                module = importlib.import_module("axolotl.integrations." + module_name)
-            else:
-                raise orig_exc
-        except ModuleNotFoundError as exc:
-            raise orig_exc from exc
-
+    module = importlib.import_module(module_name)
     # instantiate the class
     plugin_class = getattr(module, class_name)
     # create an instance of the class
@@ -293,10 +272,8 @@ class PluginManager:
         ImportError: If the plugin module cannot be imported.
         """
         try:
-            logging.info(f"Attempting to load plugin: {plugin_name}")
             plugin = load_plugin(plugin_name)
             self.plugins[plugin_name] = plugin
-            logging.info(f"Plugin loaded successfully: {plugin_name}")
         except ImportError:
             logging.error(f"Failed to load plugin: {plugin_name}")
 
@@ -369,22 +346,6 @@ class PluginManager:
         for plugin in self.plugins.values():
             plugin.post_lora_load(cfg, model)
 
-    def get_trainer_cls(self, cfg):
-        """
-        Calls the get_trainer_cls method of all registered plugins and returns the first non-None trainer class.
-
-        Parameters:
-        cfg (dict): The configuration for the plugins.
-
-        Returns:
-        object: The trainer class, or None if none was found.
-        """
-        for plugin in self.plugins.values():
-            trainer_cls = plugin.get_trainer_cls(cfg)
-            if trainer_cls is not None:
-                return trainer_cls
-        return None
-
     def create_optimizer(self, cfg, trainer):
         """
         Calls the create_optimizer method of all registered plugins and returns the first non-None optimizer.

src/axolotl/integrations/config.py

@@ -43,10 +43,12 @@ def merge_input_args():
     input_args: List[str] = plugin_manager.get_input_args()
     plugin_classes = []
     dynamic_input = ""
+
     for plugin_args in input_args:
         plugin_module, plugin_cls = plugin_args.rsplit(".", 1)
         dynamic_input += f"from {plugin_module} import {plugin_cls}\n"
         plugin_classes.append(plugin_cls)
+
     if dynamic_input:
         dynamic_input += f"class AxolotlConfigWCapabilities(AxolotlConfigWCapabilitiesBase, {', '.join(plugin_classes)}):\n    pass\n"
         dynamic_input += f"class AxolotlInputConfig(AxolotlInputConfigBase, {', '.join(plugin_classes)}):\n    pass\n"
@@ -62,4 +64,5 @@ def merge_input_args():
             "AxolotlConfigWCapabilities"
         ]
         return AxolotlConfigWCapabilities, AxolotlInputConfig
+
     return AxolotlConfigWCapabilitiesBase, AxolotlInputConfigBase

src/axolotl/integrations/kd/__init__.py

@@ -1,36 +0,0 @@
-# Copyright 2024 Axolotl AI. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-Plugin init to add KD support to Axolotl.
-"""
-from axolotl.integrations.base import BasePlugin
-
-from .args import KDArgs  # pylint: disable=unused-import. # noqa: F401
-
-
-class KDPlugin(BasePlugin):
-    """
-    Plugin for KD support in Axolotl.
-    """
-
-    def get_input_args(self):
-        return "axolotl.integrations.kd.KDArgs"
-
-    def get_trainer_cls(self, cfg):
-        if cfg.kd_trainer:
-            from .trainer import AxolotlKDTrainer
-
-            return AxolotlKDTrainer
-        return None

src/axolotl/integrations/kd/args.py

@@ -1,33 +0,0 @@
-# Copyright 2024 Axolotl AI. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-Plugin args for KD support.
-"""
-from typing import Optional
-
-from pydantic import BaseModel
-
-
-class KDArgs(BaseModel):
-    """
-    Input args for knowledge distillation.
-    """
-
-    kd_trainer: Optional[bool] = None  # whether to use KD trainer
-    kd_ce_alpha: Optional[
-        float
-    ] = None  # loss coefficient for cross-entropy loss during KD
-    kd_alpha: Optional[float] = None  # loss coefficient for KD loss
-    kd_temperature: Optional[float] = None  # temperature for sampling during KD

src/axolotl/integrations/kd/chat_template.py

@@ -1,164 +0,0 @@
-# Copyright 2024 Axolotl AI. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-Chat template prompt strategy loader with KD support
-"""
-from typing import Any, Dict
-
-import torch
-
-from axolotl.prompt_strategies.chat_template import ChatTemplateStrategy, StrategyLoader
-
-
-class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
-    """
-    Handle fields for logprob KD
-    """
-
-    def __init__(
-        self,
-        prompter,
-        tokenizer,
-        train_on_inputs,
-        sequence_len,
-        roles_to_train=None,
-        train_on_eos=None,
-        logprobs_field="logprobs",
-        gen_temperature=1.0,
-        kd_temperature=1.0,
-    ):
-        self.logprobs_field = logprobs_field
-        self.gen_temperature = gen_temperature
-        self.kd_temperature = kd_temperature
-
-        super().__init__(
-            prompter,
-            tokenizer,
-            train_on_inputs,
-            sequence_len,
-            roles_to_train=roles_to_train,
-            train_on_eos=train_on_eos,
-        )
-
-    def transform_logprobs(self, sample):
-        logprobs = sample.pop(self.logprobs_field)
-        target_seq_len = len(logprobs)
-        input_seq_len = len(sample["input_ids"])
-        input_padding_len = input_seq_len - target_seq_len
-        top_k = len(logprobs[0])
-        target_logprobs = []
-        target_token_ids = []
-        target_mask = []
-
-        # fill with -inf for padding_len tokens for top_k tokens
-        # extend target_logprobs with a padding_len x top_k 2D list filled with -inf
-        for _ in range(1, input_padding_len):  # start at 1 since this is causal
-            target_logprobs.append([-float("inf")] * top_k)
-            target_token_ids.append(list(range(top_k)))
-            target_mask.append([0] * top_k)
-
-        for _ in range(target_seq_len):
-            # TODO also check against sample["labels"]
-            target_mask.append([1] * top_k)
-
-        for _, token_pos_logprobs in enumerate(logprobs):
-            # Initialize collections for logprobs and token_ids
-            position_logprobs = []
-            position_token_ids = []
-
-            # Process each token probability entry
-            for entry in token_pos_logprobs:
-                # Extract logprob value
-                logprob = entry["logprob"]
-
-                # Parse token_id from the "token_id:###" format
-                token_id = int(entry["token"].split(":")[1])
-
-                # Append to our collections
-                position_logprobs.append(logprob)
-                position_token_ids.append(token_id)
-
-            # Convert to a tensor for easier manipulation
-            # Convert to tensor
-            position_logprobs_tensor = torch.tensor(
-                position_logprobs, dtype=torch.float
-            )
-
-            if self.kd_temperature != self.gen_temperature:
-                #
-                # Now we have distribution at T1 in log form, i.e. log p_{T1}(k).
-                # Next, re-scale to T2 = self.kd_temperature via exponent-based trick
-                # p_{T2}(k) = [p_{T1}(k)]^(T1 / T2) / Z
-                #
-                # Convert from log to probability
-                teacher_probs_t1 = position_logprobs_tensor.exp()
-                # Exponentiate by factor (T1 / T2)
-                exponent = self.gen_temperature / self.kd_temperature
-                teacher_probs_t2 = teacher_probs_t1**exponent
-                # Re-normalize
-                teacher_probs_t2 = teacher_probs_t2 / teacher_probs_t2.sum(
-                    dim=0, keepdim=True
-                )
-                # Convert back to log
-                position_logprobs_tensor = torch.log(teacher_probs_t2)
-
-            # Now we have log p_{teacher, T2}(k) stored in position_logprobs_tensor
-            position_logprobs_scaled = position_logprobs_tensor.tolist()
-
-            target_logprobs.append(position_logprobs_scaled)
-            target_token_ids.append(position_token_ids)
-
-        # since we started at index 1 for causal, we need one more padding token
-        target_logprobs.append([-float("inf")] * top_k)
-        target_token_ids.append(list(range(top_k)))
-        target_mask.append([0] * top_k)
-
-        # Update sample with transformed logprobs
-        sample["target_logprobs"] = target_logprobs
-        sample["target_token_ids"] = target_token_ids
-        sample["target_mask"] = target_mask
-
-        return sample
-
-    def _tokenize_single_prompt(self, prompt):
-        logprobs = prompt.pop(self.logprobs_field)
-        tokenized_prompt = super()._tokenize_single_prompt(prompt)
-        tokenized_prompt[self.logprobs_field] = logprobs
-        tokenized_prompt = self.transform_logprobs(tokenized_prompt)
-
-        return tokenized_prompt
-
-
-class KDStrategyLoader(StrategyLoader):
-    """
-    Load ChatTemplateStrategy with KD support using StrategyLoader.
-    """
-
-    def _get_strategy_cls(self):
-        return ChatTemplateStrategyWithKD
-
-    def _get_strategy_params(self, cfg, ds_cfg: Dict[str, Any]):
-        strategy_params = super()._get_strategy_params(cfg, ds_cfg)
-        if logprobs_field := ds_cfg.get("logprobs_field"):
-            strategy_params["logprobs_field"] = logprobs_field
-        if gen_temperature := ds_cfg.get("temperature"):
-            strategy_params["gen_temperature"] = gen_temperature
-        if kd_temperature := cfg.get("kd_temperature"):
-            strategy_params["kd_temperature"] = kd_temperature
-
-        return strategy_params
-
-
-load = KDStrategyLoader()

src/axolotl/integrations/kd/collator.py

@@ -1,255 +0,0 @@
-# Copyright 2024 Axolotl AI. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-DataCollator for axolotl to handle KD fields without using -inf for padding,
-and with a teacher_mask to identify padded positions.
-"""
-
-from dataclasses import dataclass
-from typing import Any, Optional, Union
-
-import numpy as np
-import torch
-from transformers import PreTrainedTokenizerBase
-from transformers.utils import PaddingStrategy
-
-from axolotl.utils.collators.batching import DataCollatorForSeq2Seq
-
-
-@dataclass
-class DataCollatorForKD(DataCollatorForSeq2Seq):
-    """
-    Data collator for KD, including handling KD-specific fields.
-
-    This version avoids using -inf and instead uses a large negative value for padding
-    target_logprobs. It also creates a teacher_mask to indicate which entries are valid.
-    """
-
-    # pylint: disable=duplicate-code
-    tokenizer: PreTrainedTokenizerBase
-    model: Optional[Any] = None
-    padding: Union[bool, str, PaddingStrategy] = True
-    max_length: Optional[int] = None
-    pad_to_multiple_of: Optional[int] = None
-    label_pad_token_id: int = -100
-    position_pad_token_id: int = 0
-    return_tensors: str = "pt"
-
-    def __call__(self, features, return_tensors=None):
-        if return_tensors is None:
-            return_tensors = self.return_tensors
-
-        padding_side = self.tokenizer.padding_side
-
-        # Pad labels and position_ids first
-        for feature_name, pad_token_id in [
-            ("labels", self.label_pad_token_id),
-            ("position_ids", self.position_pad_token_id),
-        ]:
-            if feature_name in features[0]:
-                feat = [f[feature_name] for f in features]
-                max_len = max(len(x) for x in feat)
-                if self.pad_to_multiple_of is not None:
-                    max_len = (
-                        (max_len + self.pad_to_multiple_of - 1)
-                        // self.pad_to_multiple_of
-                    ) * self.pad_to_multiple_of
-
-                for f in features:  # pylint: disable=invalid-name
-                    remainder = [pad_token_id] * (max_len - len(f[feature_name]))
-                    if isinstance(f[feature_name], list):
-                        f[feature_name] = (
-                            f[feature_name] + remainder
-                            if padding_side == "right"
-                            else remainder + f[feature_name]
-                        )
-                    else:
-                        # If they are numpy arrays
-                        if padding_side == "right":
-                            f[feature_name] = np.concatenate(
-                                [f[feature_name], remainder]
-                            ).astype(np.int64)
-                        else:
-                            f[feature_name] = np.concatenate(
-                                [remainder, f[feature_name]]
-                            ).astype(np.int64)
-
-        # Handle target_logprobs and target_token_ids manually
-        target_logprobs_list = []
-        target_token_ids_list = []
-        target_mask_list = []
-        has_teacher_data = ("target_logprobs" in features[0]) and (
-            "target_token_ids" in features[0]
-        )
-
-        if has_teacher_data:
-            # Extract and remove from features
-            for f in features:  # pylint: disable=invalid-name
-                target_logprobs_list.append(f.pop("target_logprobs"))
-                target_token_ids_list.append(f.pop("target_token_ids"))
-                target_mask_list.append(f.pop("target_mask"))
-
-            # Determine max lengths
-            max_teacher_seq_len = max(len(seq) for seq in target_logprobs_list)
-            max_k = max(len(seq_k) for seq in target_logprobs_list for seq_k in seq)
-
-            padded_target_logprobs = []
-            padded_target_token_ids = []
-            padded_teacher_mask_list = []
-
-            for t_logprobs, t_ids, t_mask in zip(
-                target_logprobs_list, target_token_ids_list, target_mask_list
-            ):
-                t_logprobs_padded = []
-                t_ids_padded = []
-                t_mask_padded = []
-
-                for lp, ids, mask in zip(  # pylint: disable=invalid-name
-                    t_logprobs, t_ids, t_mask
-                ):
-                    lp_len = len(lp)
-                    if lp_len < max_k:
-                        # Use -1e9 for padding logprobs and 0 for token_ids
-                        pad_len = max_k - lp_len
-                        lp = lp + [-1e9] * pad_len  # pylint: disable=invalid-name
-                        ids = ids + [0] * pad_len
-                        mask = mask + [0] * pad_len
-                    else:
-                        lp = lp[:max_k]  # pylint: disable=invalid-name
-                        ids = ids[:max_k]
-                        mask = mask[:max_k]
-
-                    t_logprobs_padded.append(lp)
-                    t_ids_padded.append(ids)
-                    t_mask_padded.append(mask)
-
-                seq_len_diff = max_teacher_seq_len - len(t_logprobs_padded)
-                if seq_len_diff > 0:
-                    # Pad sequences fully if needed
-                    t_logprobs_padded.extend(
-                        [[-1e9] * max_k for _ in range(seq_len_diff)]
-                    )
-                    t_ids_padded.extend([[0] * max_k for _ in range(seq_len_diff)])
-                    t_mask_padded.extend([[0] * max_k for _ in range(seq_len_diff)])
-
-                padded_target_logprobs.append(t_logprobs_padded)
-                padded_target_token_ids.append(t_ids_padded)
-                padded_teacher_mask_list.append(t_mask_padded)
-
-            # Convert to tensors
-            padded_target_logprobs = torch.tensor(
-                padded_target_logprobs, dtype=torch.float
-            )
-            padded_target_token_ids = torch.tensor(
-                padded_target_token_ids, dtype=torch.long
-            )
-            padded_teacher_mask_list = torch.tensor(
-                padded_teacher_mask_list, dtype=torch.int
-            )
-
-        # Pad using tokenizer for regular fields
-        features = self.tokenizer.pad(
-            features,
-            padding=self.padding,
-            max_length=self.max_length,
-            pad_to_multiple_of=self.pad_to_multiple_of,
-            return_tensors=return_tensors,
-        )
-
-        # Add back teacher data if present
-        if has_teacher_data:
-            features["target_logprobs"] = padded_target_logprobs
-            features["target_token_ids"] = padded_target_token_ids
-            features["target_mask"] = padded_teacher_mask_list
-
-        # Prepare decoder_input_ids if the model supports it
-        if (
-            "labels" in features
-            and self.model is not None
-            and hasattr(self.model, "prepare_decoder_input_ids_from_labels")
-        ):
-            decoder_input_ids = self.model.prepare_decoder_input_ids_from_labels(
-                labels=features["labels"]
-            )
-            features["decoder_input_ids"] = decoder_input_ids
-
-        return features
-
-
-class KDBatchSamplerDataCollatorForSeq2Seq(DataCollatorForKD):
-    """
-    Collator for multipack (batch of sub-batches) specifically for KD.
-    Adapts DataCollatorForKD so it can pack multiple sequences in a single batch item.
-    """
-
-    def __call__(self, features, return_tensors=None):
-        """
-        Expects that `features` could be either:
-          - a single list of dicts, OR
-          - a list of lists of dicts (the "sub-batches" to be packed).
-        """
-        # 1) If we are *not* dealing with multiple sequences per batch element,
-        #    just pass straight to parent.
-        if not isinstance(features[0], list):
-            return super().__call__(features, return_tensors=return_tensors)
-
-        # 2) Otherwise, we *are* dealing with multiple sequences in each batch item.
-        #    We want to produce a single "merged" feature dict for each sub-batch.
-        out_features = [{} for _ in features]
-
-        for i, sub_features in enumerate(features):
-            # sub_features is a list of dicts, each dict = one sequence’s features
-            # We'll merge them into out_features[i].
-            #
-            # NOTE: You can customize how you combine fields as needed (e.g. summation
-            # or offset for attention_mask). Below is a straightforward concatenation/extension.
-
-            for field_name in sub_features[0].keys():
-                # Some fields you might want to skip or treat specially:
-                if field_name == "length":
-                    continue
-
-                # If it’s a KD field that’s a list-of-lists (e.g. target_logprobs),
-                # you typically just want to flatten them by extending.
-                if field_name in ["target_logprobs", "target_token_ids", "target_mask"]:
-                    combined = []
-                    for feat in sub_features:
-                        combined.extend(feat[field_name])
-                    out_features[i][field_name] = combined
-
-                elif field_name == "attention_mask":
-                    # Here we apply the (j+1) factor to differentiate each sub-sample
-                    # within this merged batch item.
-                    arrays = []
-                    for j, feat in enumerate(sub_features):
-                        if field_name in feat:
-                            arrays.append((j + 1) * np.array(feat[field_name]))
-                    out_features[i][field_name] = np.concatenate(arrays)
-                else:
-                    # By default, just concatenate them if they are arrays
-                    # or extend them if they are lists.
-                    # For example, input_ids or labels are often arrays.
-                    arrays = []
-                    for feat in sub_features:
-                        if field_name in feat:
-                            arr = np.array(feat[field_name])
-                            arrays.append(arr)
-                    out_features[i][field_name] = np.concatenate(arrays)
-
-        # 3) Now call the parent collator, which will do:
-        #    - padding of labels/position_ids
-        #    - KD-specific padding for target_logprobs, target_token_ids, etc.
-        #    - final conversion to return_tensors
-        return super().__call__(out_features, return_tensors=return_tensors)

src/axolotl/integrations/kd/topk_logprob/LICENSE.md

@@ -1,58 +0,0 @@
-### AXOLOTL COMMUNITY LICENSE AGREEMENT
-
-This Axolotl Community License Agreement (“Agreement”) is entered into by and between Axolotl AI Corp. (“Axolotl”) and
-any individual or entity (“Licensee”) who wishes to use the Software (as defined below) in accordance with the terms
-and conditions set forth in this Agreement.
-
-1.  Definitions
-    1.1 “Licensee” refers to any individual or entity who has obtained a copy of the Software under this Agreement.
-    1.2 “Plugin Integration” means independent integration software modules which may or may not be offered by Axolotl,
-        which may be licensed separately by their respective  authors and/or licensors.
-    1.3 “Software” refers to the specific sub-directory of the Axolotl, Inc. software located at
-        https://github.com/axolotl-ai-cloud/axolotl/tree/main/src/axolotl/integrations and its subdirectories which
-        permits Plugin Integrations to integrate with the Axolotl service.
-2.  Grant of License
-    2.1	Axolotl hereby grants Licensee a worldwide, non-exclusive, royalty-free, license to use, copy, modify, merge,
-        publish, distribute, sublicense, and/or otherwise exploit the Software, subject to the following conditions:
-        - Licensee must comply with all the terms and conditions of this Agreement.
-        - Licensee must include the original copyright notice and disclaimer of warranty in all copies or substantial
-          portions of the Software.
-    2.2 Licensee may use the Software for any lawful purpose, except as restricted in Section 3.
-3.  Restrictions
-    3.1 Licensee shall not use the Software for any activity that constitutes a commercial activity of offering for
-        free or for sale any services, platform, or equivalent  to third parties for the purposes of allowing such
-        third parties to fine-tune artificial intelligence models.
-    3.2 Licensee shall not:
-        - Use the Software for any illegal or unauthorized purpose.
-        - Reverse engineer, decompile, or disassemble the Software.
-        - Remove or modify any copyright, trademark, or other proprietary notices contained in the Software.
-        - Use the Software in a way that could damage, disable, overburden, or impair the functionality of the
-          Software or interfere with any third-party use of the Software.
-    3.3 Axolotl reserves the right to restrict certain Plugin Integrations for use with the Software. To the extent Licensee integrates a permitted, applicable Plugin Integration with the Software, Licensee shall comply with any additional terms and conditions imposed by the licensors of such Plugin Integration for use of such Plugin Integrations. Licensee shall contact Axolotl if it has questions about whether its use of the Software falls beyond the scope of this Agreement.
-4.  Intellectual Property Rights
-    4.1 Axolotl and its contributors retain all intellectual property rights in and to the Software. Licensee
-        acknowledges that this Agreement does not transfer any ownership rights or intellectual property rights to
-        Licensee.
-5.  Disclaimer of Warranty
-    5.1 THE SOFTWARE IS PROVIDED “AS IS,” WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED
-        TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE, AND NON-INFRINGEMENT. IN NO EVENT SHALL
-        THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES, OR OTHER LIABILITY, WHETHER IN AN ACTION OF
-        CONTRACT, TORT, OR OTHERWISE, ARISING FROM, OUT OF, OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
-        DEALINGS IN THE SOFTWARE.
-6.  Termination
-    6.1 Axolotl may terminate this Agreement at any time if Licensee fails to comply with any of the terms and
-        conditions set forth herein. Upon termination, Licensee shall cease all use of the Software and destroy any
-        copies in its possession.
-7.  Governing Law
-    7.1 This Agreement shall be governed by and construed in accordance with the laws of the State of California,
-        without regards to conflicts of laws provisions thereof.
-8.  Entire Agreement
-    8.1 This Agreement constitutes the entire agreement between Axolotl and Licensee with respect to the subject matter
-        hereof and supersedes all prior or contemporaneous understandings or agreements between the parties concerning
-        the Software, whether written or oral. Axolotl may update the terms of this Agreement from time to time, and
-        Licensee’s continued use of the Software after any such updates shall constitute acceptance of updated terms
-        on a go-forward basis.  Axolotl will use commercially reasonable efforts to provide Licensee notice of any
-        material updates. By using the Software, Licensee acknowledges that it has read, understood, and agrees to be
-        bound by the terms and conditions of this Agreement.
-
-This Agreement was last updated on August 23, 2024.

src/axolotl/integrations/kd/topk_logprob/forward_kl.py

@@ -1,82 +0,0 @@
-# Copyright 2024 Axolotl AI. All rights reserved.
-#
-# This software may be used and distributed according to
-# the terms of the Axolotl Community License Agreement (the "License");
-# you may not use this file except in compliance with the License.
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
-# WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
-# License for the specific language governing permissions and limitations under
-# the License.
-
-"""
-loss for top_k KL divergence
-"""
-import torch
-
-
-@torch.jit.script
-def loss(
-    student_logits: torch.Tensor,
-    target_token_ids: torch.Tensor,
-    target_logprobs: torch.Tensor,
-    target_mask: torch.Tensor,
-    num_items_in_batch: int = -1,  # Use -1 to indicate "None"
-    kd_temperature: float = 1.0,
-) -> torch.Tensor:
-    """
-    A KD loss function that is TorchScript-friendly.
-    """
-
-    # Determine the teacher sequence length
-    # target_token_ids shape: [B, teacher_seq_len, K]
-    # student_logits shape:   [B, student_seq_len, vocab_size]
-    teacher_seq_len = target_token_ids.shape[1]
-
-    # Slice student logits to match teacher-provided sequence length
-    student_logits_for_kd = student_logits[
-        :, :teacher_seq_len, :
-    ]  # [B, teacher_seq_len, vocab_size]
-
-    # Gather student logits for teacher's top-K tokens
-    student_logits_topk = torch.gather(
-        student_logits_for_kd, dim=-1, index=target_token_ids
-    )  # [B, teacher_seq_len, K]
-
-    # Apply KD temperature to student’s logits
-    if kd_temperature != 1.0:
-        student_logits_topk = student_logits_topk / kd_temperature
-
-    # Convert student top-k logits to logprobs
-    student_logprobs_topk = student_logits_topk - torch.logsumexp(
-        student_logits_topk, dim=-1, keepdim=True
-    )  # [B, teacher_seq_len, K]
-
-    # Convert teacher_mask to boolean for indexing
-    # In TorchScript, .bool() is sometimes unsupported, so we do:
-    valid_mask = target_mask.to(torch.bool)
-
-    # Prune tensors to only keep valid tokens
-    student_logprobs_topk = student_logprobs_topk[valid_mask]
-    target_logprobs = target_logprobs[valid_mask]
-
-    # Convert teacher logprobs to probabilities
-    teacher_probs = target_logprobs.exp()
-
-    # Compute forward KL
-    kd_loss_per_token = teacher_probs * (target_logprobs - student_logprobs_topk)
-    kd_loss = kd_loss_per_token.sum()
-
-    # Multiply by T^2 (classical KD scaling)
-    if kd_temperature != 1.0:
-        kd_loss = kd_loss * (kd_temperature**2)
-
-    # Normalize by number of items (if provided) or by valid tokens
-    if num_items_in_batch > 0:
-        kd_loss = kd_loss / float(num_items_in_batch)
-    else:
-        # Fall back to average over valid tokens
-        kd_loss = kd_loss / float(kd_loss_per_token.size(0))
-
-    return kd_loss

src/axolotl/integrations/kd/trainer.py

@@ -1,107 +0,0 @@
-# Copyright 2024 Axolotl AI. All rights reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-"""
-KD trainer
-"""
-
-from axolotl.core.trainers.base import AxolotlTrainer
-
-from .topk_logprob.forward_kl import loss as topk_kd_loss
-
-
-class AxolotlKDTrainer(AxolotlTrainer):
-    """
-    Custom trainer subclass for Knowledge Distillation (KD)
-    """
-
-    def _set_signature_columns_if_needed(self):
-        super()._set_signature_columns_if_needed()
-        columns_to_add = []
-        if self._signature_columns:
-            if "target_logprobs" not in self._signature_columns:
-                columns_to_add.append("target_logprobs")
-            if "target_token_ids" not in self._signature_columns:
-                columns_to_add.append("target_token_ids")
-            if "target_mask" not in self._signature_columns:
-                columns_to_add.append("target_mask")
-            if columns_to_add:
-                self._signature_columns += columns_to_add
-
-    def compute_loss(
-        self,
-        model,
-        inputs,
-        return_outputs=False,
-        num_items_in_batch=None,
-        shift_targets=False,
-    ):
-        """
-        How the loss is computed by Trainer. By default, all models return the loss in the first element.
-
-        Subclass and override for custom behavior.
-        """
-
-        target_logprobs = inputs.pop("target_logprobs")
-        target_token_ids = inputs.pop("target_token_ids")
-        target_mask = inputs.pop("target_mask")
-
-        seq_len = target_token_ids.shape[1]
-
-        if self.model_accepts_loss_kwargs:
-            loss_kwargs = {}
-            if num_items_in_batch is not None:
-                loss_kwargs["num_items_in_batch"] = num_items_in_batch
-            inputs = {**inputs, **loss_kwargs}
-        outputs = model(**inputs)
-
-        # FIXME: account for tokenizer.padding_side
-        student_logits = outputs["logits"][:, :seq_len, :].contiguous()
-
-        if shift_targets:
-            shift_logits = student_logits[..., :-1, :].contiguous()
-            target_logprobs_for_loss = target_logprobs[..., 1:, :].contiguous()
-            target_token_ids_for_loss = target_token_ids[..., 1:, :].contiguous()
-            target_mask_for_loss = target_mask[..., 1:, :].contiguous()
-        else:
-            shift_logits = student_logits.contiguous()
-            target_logprobs_for_loss = target_logprobs.contiguous()
-            target_token_ids_for_loss = target_token_ids.contiguous()
-            target_mask_for_loss = target_mask.contiguous()
-
-        loss_kd = topk_kd_loss(
-            shift_logits,
-            target_token_ids_for_loss,
-            target_logprobs_for_loss,
-            target_mask_for_loss,
-            num_items_in_batch=num_items_in_batch,
-            kd_temperature=self.args.kd_temperature,
-        )
-
-        if self.args.kd_ce_alpha > 0:
-            kd_alpha = self.args.kd_alpha
-            loss = self.args.kd_ce_alpha * outputs["loss"] + kd_alpha * loss_kd
-        else:
-            loss = loss_kd
-        # Save past state if it exists
-        # TODO: this needs to be fixed and made cleaner later.
-        if self.args.past_index >= 0:
-            self._past = outputs[  # pylint: disable=attribute-defined-outside-init
-                self.args.past_index
-            ]
-
-        if self.args.average_tokens_across_devices and self.model_accepts_loss_kwargs:
-            loss *= self.accelerator.num_processes
-
-        return (loss, outputs) if return_outputs else loss

src/axolotl/prompt_strategies/__init__.py

@@ -16,21 +16,10 @@ def load(strategy, tokenizer, cfg, ds_cfg, processor=None):
 
             return messages_load(tokenizer, cfg, ds_cfg, processor=processor)
         load_fn = "load"
-        package = "axolotl.prompt_strategies"
         if strategy.split(".")[-1].startswith("load_"):
             load_fn = strategy.split(".")[-1]
             strategy = ".".join(strategy.split(".")[:-1])
-        elif len(strategy.split(".")) > 1:
-            try:
-                importlib.import_module(
-                    "." + strategy.split(".")[-1],
-                    ".".join(strategy.split(".")[:-1]),
-                )
-                package = ".".join(strategy.split(".")[:-1])
-                strategy = strategy.split(".")[-1]
-            except ModuleNotFoundError:
-                pass
-        mod = importlib.import_module(f".{strategy}", package)
+        mod = importlib.import_module(f".{strategy}", "axolotl.prompt_strategies")
         func = getattr(mod, load_fn)
         load_kwargs = {}
         if strategy == "user_defined":

src/axolotl/prompt_strategies/base.py

@@ -10,8 +10,6 @@ LOG = logging.getLogger("axolotl")
 
 def load(strategy, cfg, module_base=None, **kwargs):
     try:
-        if len(strategy.split(".")) == 1:
-            strategy = strategy + ".default"
         load_fn = strategy.split(".")[-1]
         strategy = ".".join(strategy.split(".")[:-1])
         mod = importlib.import_module(f".{strategy}", module_base)

src/axolotl/prompt_strategies/bradley_terry/chat_template.py

@@ -21,11 +21,7 @@ class BTChatTemplateStrategy(ChatTemplateStrategy):
     Bradley-Terry reward model pairwise chat template prompt strategy.
     """
 
-    @property
-    def supports_batched(self) -> bool:
-        return False
-
-    def _tokenize_single_prompt(self, prompt):
+    def tokenize_prompt(self, prompt):
         """
 
         :param prompt: the actual row of data from the underlying dataset
@@ -43,7 +39,7 @@ class BTChatTemplateStrategy(ChatTemplateStrategy):
             )
         prompt[self.messages].append({"role": "user", "content": prompt["input"]})
         prompt[self.messages].append({"role": "assistant", "content": prompt["chosen"]})
-        chosen_tokenized = super()._tokenize_single_prompt(prompt)
+        chosen_tokenized = super().tokenize_prompt(prompt)
 
         if len(chosen_tokenized["input_ids"]) > max_length:
             LOG.warning(
@@ -66,7 +62,7 @@ class BTChatTemplateStrategy(ChatTemplateStrategy):
         prompt[self.messages].append(
             {"role": "assistant", "content": prompt["rejected"]}
         )
-        rejected_tokenized = super()._tokenize_single_prompt(prompt)
+        rejected_tokenized = super().tokenize_prompt(prompt)
 
         if len(rejected_tokenized["input_ids"]) > max_length:
             LOG.warning(

src/axolotl/prompt_strategies/chat_template.py

@@ -3,7 +3,6 @@ HF Chat Templates prompt strategy
 """
 
 import logging
-from collections import defaultdict
 from typing import Any, Dict, List, Optional
 
 from transformers import ProcessorMixin
@@ -194,7 +193,7 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
 
     def __init__(
         self,
-        prompter: ChatTemplatePrompter,
+        prompter,
         tokenizer,
         train_on_inputs,
         sequence_len,
@@ -221,61 +220,22 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
     def messages(self, messages):
         self._messages = messages
 
-    @property
-    def supports_batched(self) -> bool:
-        # Let calling code know we can handle lists of examples
-        return True
-
-    def is_prompt_batched(self, prompt: dict[str, Any]) -> bool:
-        try:
-            return all(isinstance(v, list) for v in prompt.values()) and all(
-                isinstance(v, list) for v in prompt[self.messages]
-            )
-        except KeyError:
-            return False
-
-    def tokenize_prompt(self, prompt: dict[str, Any]):
-        """
-        Public method that can handle either a single prompt or a batch of prompts.
-        """
-
-        if not self.is_prompt_batched(prompt) or not self.supports_batched:
-            return self._tokenize_single_prompt(prompt)
-
-        res = defaultdict(lambda: [])
-        feature_names = list(prompt.keys())
-
-        # Process each prompt individually
-        for row in zip(*prompt.values()):
-            tokenized_prompt = self._tokenize_single_prompt(
-                dict(zip(feature_names, row))
-            )
-            for key, val in tokenized_prompt.items():
-                for i in range(0, len(val), self.sequence_len):
-                    res[key].append(val[i : i + self.sequence_len])
-
-        # If there are no examples left, return an empty dictionary
-        if not res:
-            return {}
-
-        return dict(res)
-
-    def _tokenize_single_prompt(self, prompt: dict) -> Dict[str, List[int]]:
+    def tokenize_prompt(self, prompt):
         # Old simple legacy behavior that works reliably.
         if (
             not self.roles_to_train
             and not self.train_on_eos
-            and not self.prompter.message_field_training  # type: ignore
-            and not self.prompter.message_field_training_detail  # type: ignore
+            and not self.prompter.message_field_training
+            and not self.prompter.message_field_training_detail
         ):
             turns = self.get_conversation_thread(prompt)
             images = self.get_images(prompt)
-            prompt_ids = self.prompter.build_prompt(  # type: ignore
+            prompt_ids = self.prompter.build_prompt(
                 turns[:-1],
                 add_generation_prompt=True,
                 images=images,
             )
-            tokenized_res = self.prompter.build_prompt(turns, images=images)  # type: ignore
+            tokenized_res = self.prompter.build_prompt(turns, images=images)
             tokenized_prompt = {}
             if isinstance(tokenized_res, list):
                 input_ids = prompt_ids + tokenized_res[len(prompt_ids) :]
@@ -296,7 +256,7 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
             return tokenized_prompt
 
         turns = self.get_conversation_thread(prompt)
-        input_ids = self.prompter.build_prompt(turns)  # type: ignore
+        input_ids = self.prompter.build_prompt(turns)
         labels = [IGNORE_TOKEN_ID] * len(input_ids)
 
         last_eos_idx = -1
@@ -326,7 +286,7 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
 
             if should_train and turn_start_idx != -1 and turn_end_idx != -1:
                 if train_detail:
-                    token_offsets = self.prompter.get_offsets_for_train_detail(  # type: ignore
+                    token_offsets = self.prompter.get_offsets_for_train_detail(
                         content, train_detail
                     )
                     LOG.debug(f"Token offsets: {token_offsets}")
@@ -499,62 +459,43 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
         return prompt.get(self.images, None)
 
 
-class StrategyLoader:
-    """
-    Load chat template strategy based on configuration.
-    """
+def load(tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None, processor=None):
+    # pylint: disable=duplicate-code
+    ds_cfg = ds_cfg or {}
+    chat_template_string = get_chat_template_from_config(
+        cfg=cfg, ds_cfg=ds_cfg, tokenizer=tokenizer
+    )
+    LOG.info(f"Using chat template:\n---\n{chat_template_string!s}\n---")
 
-    def _get_strategy_cls(self):
-        return ChatTemplateStrategy
+    prompter_params = {
+        "tokenizer": tokenizer,
+        "chat_template": chat_template_string,
+        "message_field_role": ds_cfg.get("message_field_role", "role"),
+        "message_field_content": ds_cfg.get("message_field_content", "content"),
+        "message_field_training": ds_cfg.get("message_field_training", None),
+        "message_field_training_detail": ds_cfg.get(
+            "message_field_training_detail",
+            None,
+        ),
+        "roles": ds_cfg.get("roles"),
+        "drop_system_message": ds_cfg.get("drop_system_message", False),
+        # we need to add one for detecting sequences with exceeding the `sequence_len` limit.
+        "max_length": cfg.sequence_len + 1,
+        "processor": processor,
+    }
 
-    def _get_strategy_params(self, cfg, ds_cfg: Dict[str, Any]):
-        return {
-            "train_on_inputs": cfg.train_on_inputs,
-            "sequence_len": cfg.sequence_len,
-            "roles_to_train": ds_cfg.get("roles_to_train", ["assistant"]),
-            "train_on_eos": ds_cfg.get("train_on_eos", "turn"),
-        }
+    strategy_params = {
+        "train_on_inputs": cfg.train_on_inputs,
+        "sequence_len": cfg.sequence_len,
+        "roles_to_train": ds_cfg.get("roles_to_train", ["assistant"]),
+        "train_on_eos": ds_cfg.get("train_on_eos", "turn"),
+    }
 
-    def __call__(
-        self, tokenizer, cfg, ds_cfg: Optional[Dict[str, Any]] = None, processor=None
-    ):
-        # pylint: disable=duplicate-code
-        ds_cfg = ds_cfg or {}
-        chat_template_string = get_chat_template_from_config(
-            cfg=cfg, ds_cfg=ds_cfg, tokenizer=tokenizer
-        )
-        LOG.info(f"Using chat template:\n---\n{chat_template_string!s}\n---")
+    strategy = ChatTemplateStrategy(
+        ChatTemplatePrompter(**prompter_params), tokenizer=tokenizer, **strategy_params
+    )
 
-        prompter_params = {
-            "tokenizer": tokenizer,
-            "chat_template": chat_template_string,
-            "message_field_role": ds_cfg.get("message_field_role", "role"),
-            "message_field_content": ds_cfg.get("message_field_content", "content"),
-            "message_field_training": ds_cfg.get("message_field_training", None),
-            "message_field_training_detail": ds_cfg.get(
-                "message_field_training_detail",
-                None,
-            ),
-            "roles": ds_cfg.get("roles"),
-            "drop_system_message": ds_cfg.get("drop_system_message", False),
-            # we need to add one for detecting sequences with exceeding the `sequence_len` limit.
-            "max_length": cfg.sequence_len + 1,
-            "processor": processor,
-        }
+    if "field_messages" in ds_cfg and hasattr(strategy, "messages"):
+        strategy.messages = ds_cfg["field_messages"]
 
-        strategy_params = self._get_strategy_params(cfg, ds_cfg)
-        strategy_cls = self._get_strategy_cls()
-
-        strategy = strategy_cls(
-            ChatTemplatePrompter(**prompter_params),
-            tokenizer=tokenizer,
-            **strategy_params,
-        )
-
-        if "field_messages" in ds_cfg and hasattr(strategy, "messages"):
-            strategy.messages = ds_cfg["field_messages"]
-
-        return strategy
-
-
-load = StrategyLoader()
+    return strategy

src/axolotl/prompt_strategies/dpo/chatml.py

@@ -3,41 +3,22 @@ DPO strategies for chatml
 """
 
 
-def default(
+def argilla(
     cfg,
     **kwargs,
 ):  # pylint: disable=possibly-unused-variable,unused-argument
     def transform_fn(sample):
-        if "prompt" in sample.keys():
-            prompt_key = "prompt"
-        elif "input" in sample.keys():
-            prompt_key = "input"
-        elif "question" in sample.keys():
-            prompt_key = "question"
-        else:
-            prompt_key = "instruction"
-
-        if "chosen" in sample.keys():
-            chosen_key = "chosen"
-        else:
-            chosen_key = "chosen_response"
-
-        if "rejected" in sample.keys():
-            rejected_key = "rejected"
-        else:
-            rejected_key = "rejected_response"
-
         if "system" in sample and sample["system"]:
             sample["prompt"] = (
                 f"<|im_start|>system\n{sample['system']}<|im_end|>\n"
-                f"<|im_start|>user\n{sample[prompt_key]}<|im_end|>\n<|im_start|>assistant\n"
+                f"<|im_start|>user\n{sample['instruction']}<|im_end|>\n<|im_start|>assistant\n"
             )
         else:
             sample[
                 "prompt"
-            ] = f"<|im_start|>user\n{sample[prompt_key]}<|im_end|>\n<|im_start|>assistant\n"
-        sample["chosen"] = f"{sample[chosen_key]}<|im_end|>"
-        sample["rejected"] = f"{sample[rejected_key]}<|im_end|>"
+            ] = f"<|im_start|>user\n{sample['instruction']}<|im_end|>\n<|im_start|>assistant\n"
+        sample["chosen"] = f"{sample['chosen_response']}<|im_end|>"
+        sample["rejected"] = f"{sample['rejected_response']}<|im_end|>"
         return sample
 
     return transform_fn

src/axolotl/prompt_strategies/dpo/llama3.py

@@ -3,42 +3,22 @@ DPO strategies for llama-3 chat template
 """
 
 
-def default(
+def argilla(
     cfg,
     **kwargs,
 ):  # pylint: disable=possibly-unused-variable,unused-argument
     def transform_fn(sample):
-        # pylint: disable=duplicate-code
-        if "prompt" in sample.keys():
-            prompt_key = "prompt"
-        elif "input" in sample.keys():
-            prompt_key = "input"
-        elif "question" in sample.keys():
-            prompt_key = "question"
-        else:
-            prompt_key = "instruction"
-
-        if "chosen" in sample.keys():
-            chosen_key = "chosen"
-        else:
-            chosen_key = "chosen_response"
-
-        if "rejected" in sample.keys():
-            rejected_key = "rejected"
-        else:
-            rejected_key = "rejected_response"
-
         if "system" in sample and sample["system"]:
             sample["prompt"] = (
                 f"<|start_header_id|>system<|end_header_id|>\n\n{sample['system']}<|eot_id|>"
-                f"<|start_header_id|>user<|end_header_id|>\n\n{sample[prompt_key]}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+                f"<|start_header_id|>user<|end_header_id|>\n\n{sample['instruction']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
             )
         else:
             sample[
                 "prompt"
-            ] = f"<|start_header_id|>user<|end_header_id|>\n\n{sample[prompt_key]}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
-        sample["chosen"] = f"{sample[chosen_key]}<|eot_id|>"
-        sample["rejected"] = f"{sample[rejected_key]}<|eot_id|>"
+            ] = f"<|start_header_id|>user<|end_header_id|>\n\n{sample['instruction']}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n"
+        sample["chosen"] = f"{sample['chosen_response']}<|eot_id|>"
+        sample["rejected"] = f"{sample['rejected_response']}<|eot_id|>"
         return sample
 
     return transform_fn

src/axolotl/prompt_tokenizers.py

@@ -2,7 +2,7 @@
 
 import abc
 import logging
-from typing import Callable, Dict, List, Optional, Tuple, Union
+from typing import Dict, List, Tuple, Union
 
 from transformers import BatchEncoding, PreTrainedTokenizer
 
@@ -34,8 +34,6 @@ class PromptTokenizingStrategy(abc.ABC):
     Abstract class for tokenizing strategies
     """
 
-    filter_rows: Optional[Callable] = None
-
     def __init__(
         self,
         prompter: Prompter,

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

@@ -154,7 +154,6 @@ class SFTDataset(BaseModel):
     type: Optional[Union[str, UserDefinedPrompterType]] = None
     input_transform: Optional[str] = None
     shards: Optional[int] = None
-    preprocess_shards: Optional[int] = None
     conversation: Optional[str] = None
     # Do not make this too strict or it will break the validator to choose different dataset class
     chat_template: Optional[
@@ -177,8 +176,6 @@ class SFTDataset(BaseModel):
     message_field_content: Optional[str] = None
     message_field_training: Optional[str] = None
     message_field_training_detail: Optional[str] = None
-    logprobs_field: Optional[str] = None
-    temperature: Optional[float] = None
     roles_to_train: Optional[List[str]] = None
     train_on_eos: Optional[str] = None
     roles: Optional[Dict[str, List[str]]] = None
@@ -802,7 +799,6 @@ class AxolotlInputConfig(
 
     # INTERNALS - document for now, generally not set externally
     is_preprocess: Optional[bool] = None
-    preprocess_iterable: Optional[bool] = None
 
     total_num_tokens: Optional[int] = None
     total_supervised_tokens: Optional[int] = None

src/axolotl/utils/data/sft.py

@@ -8,7 +8,6 @@ from typing import List, Tuple, Union
 from datasets import (
     Dataset,
     DatasetDict,
-    IterableDataset,
     concatenate_datasets,
     load_dataset,
     load_from_disk,
@@ -16,7 +15,7 @@ from datasets import (
 from transformers import PreTrainedTokenizerBase
 
 from axolotl.common.const import DEFAULT_DATASET_PREPARED_PATH
-from axolotl.datasets import wrap_dataset_for_tokenized_prompt
+from axolotl.datasets import TokenizedPromptDataset
 from axolotl.prompt_strategies import load
 from axolotl.prompt_strategies.bradley_terry import load as bradley_terry_load
 from axolotl.prompt_tokenizers import (
@@ -177,11 +176,10 @@ def load_tokenized_prepared_datasets(
                 + "@"
                 + str(cfg.group_by_length)
                 + "@"
-                + str(cfg.kd_temperature or 1.0)
                 + "|".join(
                     sorted(
                         [
-                            f"{d.path}:{d.type}:{d.shards}:{d.conversation}:{d.split}:{d.temperature or 1.0}"
+                            f"{d.path}:{d.type}:{d.shards}:{d.conversation}{d.split}"
                             for d in cfg_datasets
                         ]
                     )
@@ -256,25 +254,13 @@ def load_tokenized_prepared_datasets(
                     # at the same time for a given dataset
                     for name in dataset.name:
                         yield DictDefault({**dataset, "name": name})
-                elif dataset.preprocess_shards and not dataset.shards:
-                    for shard in range(dataset.preprocess_shards):
-                        yield DictDefault(
-                            {
-                                **dataset,
-                                "shards": dataset.preprocess_shards,
-                                "shards_idx": shard,
-                            }
-                        )
                 else:
                     yield dataset
 
-        streaming_ds = False
-        if cfg.preprocess_iterable:
-            streaming_ds = True
         # pylint: disable=invalid-name
         for config_dataset in for_d_in_datasets(cfg_datasets):
             ds: Union[Dataset, DatasetDict] = load_dataset_w_config(
-                config_dataset, use_auth_token, streaming=streaming_ds
+                config_dataset, use_auth_token
             )
 
             d_base_type = d_prompt_style = None
@@ -331,21 +317,7 @@ def load_tokenized_prepared_datasets(
 
         if cfg.local_rank == 0 and not cfg.skip_prepare_dataset:
             LOG.info(f"Saving merged prepared dataset to disk... {prepared_ds_path}")
-            if isinstance(dataset, IterableDataset):
-
-                def gen_from_iter_ds(_ds, _=None):
-                    yield from _ds
-
-                ds_from_iter = Dataset.from_generator(
-                    functools.partial(gen_from_iter_ds, dataset),
-                    features=dataset.features,
-                    num_proc=cfg.dataset_processes,
-                    split=split,
-                    gen_kwargs={"_": list(range(cfg.dataset_processes))},
-                )
-                ds_from_iter.save_to_disk(str(prepared_ds_path))
-            else:
-                dataset.save_to_disk(str(prepared_ds_path))
+            dataset.save_to_disk(str(prepared_ds_path))
             if cfg.push_dataset_to_hub:
                 LOG.info(
                     f"Pushing merged prepared dataset to Huggingface hub at {cfg.push_dataset_to_hub} (version {ds_hash})..."
@@ -471,7 +443,7 @@ def get_dataset_wrapper(
             "user_defined", tokenizer, cfg, config_dataset.type.to_dict()
         )
         dataset_prompter = UnsupportedPrompter()
-        dataset_wrapper = wrap_dataset_for_tokenized_prompt(
+        dataset_wrapper = TokenizedPromptDataset(
             ds_strategy,
             dataset,
             **ds_kwargs,
@@ -484,7 +456,7 @@ def get_dataset_wrapper(
         config_dataset.type.split(".", 1)[1], tokenizer, cfg, config_dataset
     ):
         dataset_prompter = UnsupportedPrompter()
-        dataset_wrapper = wrap_dataset_for_tokenized_prompt(
+        dataset_wrapper = TokenizedPromptDataset(
             ds_strategy,
             dataset,
             **ds_kwargs,
@@ -496,7 +468,7 @@ def get_dataset_wrapper(
             dataset_wrapper = ds_strategy.wrap_dataset(dataset, **ds_kwargs)
         else:
             dataset_prompter = UnsupportedPrompter()
-            dataset_wrapper = wrap_dataset_for_tokenized_prompt(
+            dataset_wrapper = TokenizedPromptDataset(
                 ds_strategy,
                 dataset,
                 **ds_kwargs,
@@ -509,7 +481,7 @@ def get_dataset_wrapper(
             cfg.train_on_inputs,
             cfg.sequence_len,
         )
-        ds_wrapper = wrap_dataset_for_tokenized_prompt(
+        ds_wrapper = TokenizedPromptDataset(
             ds_strategy,
             dataset,
             **ds_kwargs,
@@ -523,7 +495,7 @@ def get_dataset_wrapper(
             cfg.train_on_inputs,
             cfg.sequence_len,
         )
-        ds_wrapper = wrap_dataset_for_tokenized_prompt(
+        ds_wrapper = TokenizedPromptDataset(
             ds_strategy,
             dataset,
             **ds_kwargs,
@@ -537,7 +509,7 @@ def get_dataset_wrapper(
             cfg.train_on_inputs,
             cfg.sequence_len,
         )
-        ds_wrapper = wrap_dataset_for_tokenized_prompt(
+        ds_wrapper = TokenizedPromptDataset(
             ds_strategy,
             dataset,
             **ds_kwargs,
@@ -551,7 +523,7 @@ def get_dataset_wrapper(
             cfg.train_on_inputs,
             cfg.sequence_len,
         )
-        ds_wrapper = wrap_dataset_for_tokenized_prompt(
+        ds_wrapper = TokenizedPromptDataset(
             ds_strategy,
             dataset,
             **ds_kwargs,
@@ -565,7 +537,7 @@ def get_dataset_wrapper(
             cfg.train_on_inputs,
             cfg.sequence_len,
         )
-        ds_wrapper = wrap_dataset_for_tokenized_prompt(
+        ds_wrapper = TokenizedPromptDataset(
             ds_strategy,
             dataset,
             **ds_kwargs,
@@ -579,7 +551,7 @@ def get_dataset_wrapper(
             cfg.train_on_inputs,
             cfg.sequence_len,
         )
-        ds_wrapper = wrap_dataset_for_tokenized_prompt(
+        ds_wrapper = TokenizedPromptDataset(
             ds_strategy,
             dataset,
             **ds_kwargs,
@@ -593,7 +565,7 @@ def get_dataset_wrapper(
             cfg.train_on_inputs,
             cfg.sequence_len,
         )
-        ds_wrapper = wrap_dataset_for_tokenized_prompt(
+        ds_wrapper = TokenizedPromptDataset(
             ds_strategy,
             dataset,
             **ds_kwargs,
@@ -607,7 +579,7 @@ def get_dataset_wrapper(
             cfg.train_on_inputs,
             cfg.sequence_len,
         )
-        ds_wrapper = wrap_dataset_for_tokenized_prompt(
+        ds_wrapper = TokenizedPromptDataset(
             ds_strategy,
             dataset,
             **ds_kwargs,

src/axolotl/utils/data/shared.py

@@ -29,9 +29,7 @@ def get_ds_type(config_dataset: DictDefault):
     return ds_type
 
 
-def load_dataset_w_config(
-    config_dataset, auth_token, streaming=False
-) -> Union[Dataset, DatasetDict]:
+def load_dataset_w_config(config_dataset, auth_token):
     # pylint: disable=invalid-name
     ds: Optional[Union[Dataset, DatasetDict]] = None  # pylint: disable=invalid-name
     ds_from_hub = False
@@ -119,7 +117,7 @@ def load_dataset_w_config(
                     ds_type,
                     name=config_dataset.name,
                     data_files=config_dataset.data_files,
-                    streaming=streaming,
+                    streaming=False,
                     split=None,
                 )
             else:
@@ -155,7 +153,7 @@ def load_dataset_w_config(
         ds = load_dataset(
             config_dataset.path,
             name=config_dataset.name,
-            streaming=streaming,
+            streaming=False,
             data_files=config_dataset.data_files,
             token=auth_token,
             revision=config_dataset.revision,
@@ -174,7 +172,7 @@ def load_dataset_w_config(
                 ds_type,
                 name=config_dataset.name,
                 data_files=config_dataset.path,
-                streaming=streaming,
+                streaming=False,
                 split=None,
                 storage_options=storage_options,
                 trust_remote_code=config_dataset.trust_remote_code,
@@ -185,7 +183,7 @@ def load_dataset_w_config(
             ds_type,
             name=config_dataset.name,
             data_files=config_dataset.path,
-            streaming=streaming,
+            streaming=False,
             split=None,
             storage_options=storage_options,
             trust_remote_code=config_dataset.trust_remote_code,
@@ -215,7 +213,7 @@ def load_dataset_w_config(
             "json",
             name=config_dataset.name,
             data_files=fp,
-            streaming=streaming,
+            streaming=False,
             split=None,
         )
     if not ds:

src/axolotl/utils/models.py

@@ -713,19 +713,45 @@ class ModelLoader:
         if self.cfg.flash_attention:
             if not self.cfg.sample_packing and self.cfg.s2_attention:
                 pass
-            self.model_kwargs["attn_implementation"] = "flash_attention_2"
-            self.model_config._attn_implementation = (  # pylint: disable=protected-access
-                "flash_attention_2"
-            )
+
+            if self.cfg.diff_attention:
+                self.model_kwargs[
+                    "attn_implementation"
+                ] = "differential_flash_attention_2"
+                self.model_config._attn_implementation = (  # pylint: disable=protected-access
+                    "differential_flash_attention_2"
+                )
+            else:
+                self.model_kwargs["attn_implementation"] = "flash_attention_2"
+                self.model_config._attn_implementation = (  # pylint: disable=protected-access
+                    "flash_attention_2"
+                )
         elif self.cfg.sdp_attention:
-            self.model_kwargs["attn_implementation"] = "sdpa"
-            self.model_config._attn_implementation = (  # pylint: disable=protected-access
-                "sdpa"
-            )
+            if self.cfg.diff_attention:
+                self.model_kwargs["attn_implementation"] = "differential_sdpa"
+                self.model_config._attn_implementation = (  # pylint: disable=protected-access
+                    "differential_sdpa"
+                )
+            else:
+                self.model_kwargs["attn_implementation"] = "sdpa"
+                self.model_config._attn_implementation = (  # pylint: disable=protected-access
+                    "sdpa"
+                )
         elif self.cfg.eager_attention:
-            self.model_kwargs["attn_implementation"] = "eager"
+            if self.cfg.diff_attention:
+                self.model_kwargs["attn_implementation"] = "differential_eager"
+                self.model_config._attn_implementation = (  # pylint: disable=protected-access
+                    "differential_eager"
+                )
+            else:
+                self.model_kwargs["attn_implementation"] = "eager"
+                self.model_config._attn_implementation = (  # pylint: disable=protected-access
+                    "eager"
+                )
+        elif self.cfg.diff_attention:
+            self.model_kwargs["attn_implementation"] = "differential_eager"
             self.model_config._attn_implementation = (  # pylint: disable=protected-access
-                "eager"
+                "differential_eager"
             )
 
         if self.cfg.low_cpu_mem_usage:
@@ -816,6 +842,7 @@ class ModelLoader:
 
             if self.cfg.is_multimodal:
                 self.model_config.text_config = self.text_model_config
+
             self.model = self.AutoModelLoader.from_pretrained(
                 self.base_model,
                 config=self.model_config,

src/axolotl/utils/tokenization.py

@@ -26,7 +26,6 @@ def check_example_labels(example, tokenizer, text_only=False):
     # Get the input_ids, labels, and attention_mask from the dataset
     input_ids = example["input_ids"]
     labels = example["labels"]
-    target_mask = example.pop("target_mask", None)
 
     # You can compare the input_ids and labels element-wise
     # Remember to ignore positions with IGNORE_TOKEN_ID (if you use it) or attention_mask equal to 0
@@ -43,13 +42,6 @@ def check_example_labels(example, tokenizer, text_only=False):
     delimiter = "" if text_only else " "
     LOG.info(delimiter.join(colored_tokens))
     LOG.info("\n\n\n")
-    target_labels_count = sum(label_id != -100 for label_id in labels)
-    total_len = len(input_ids)
-    LOG.info(f"Total input len: {total_len}")
-    LOG.info(f"Count of labels: {target_labels_count}")
-    if target_mask:
-        target_mask_positions = sum(m[0] for m in target_mask)
-        LOG.info(f"Number of positions in target_mask: {target_mask_positions}")
 
     return " ".join(colored_tokens)
 

src/axolotl/utils/trainer.py

@@ -11,7 +11,7 @@ import numpy as np
 import torch
 import torch.cuda
 from accelerate.logging import get_logger
-from datasets import IterableDataset, disable_caching, enable_caching
+from datasets import disable_caching, enable_caching
 from torch.utils.data import DataLoader, RandomSampler
 from transformers.utils import is_torch_bf16_gpu_available
 
@@ -95,46 +95,9 @@ def disable_datasets_caching():
 
 
 def add_position_ids(sample):
-    """
-    Handle both single-example and batched data.
-    - single example: sample['input_ids'] is a list[int]
-    - batched data: sample['input_ids'] is a list[list[int]]
-    """
-    if "input_ids" not in sample:
-        # If there's no "input_ids", just return sample unchanged
-        return sample
-
-    input_ids = sample["input_ids"]
-
-    # Detect if it's a single example or a batch
-    if not input_ids:
-        # Edge case: empty
-        return sample
-
-    # If first element is an int, it’s a single example
-    # If first element is a list, it’s a batch
-    if isinstance(input_ids[0], int):
-        # ---- SINGLE EXAMPLE ----
-        seq_len = len(input_ids)
-        # Position IDs for a single example
-        # As a list
-        sample["position_ids"] = list(range(seq_len))
-        sample["length"] = seq_len
-
-    else:
-        # ---- BATCHED EXAMPLES ----
-        # input_ids is a list of lists
-        position_ids_batch = []
-        lengths_batch = []
-        for seq in input_ids:
-            seq_len = len(seq)
-            position_ids_batch.append(list(range(seq_len)))
-            lengths_batch.append(seq_len)
-
-        # Now store them back
-        sample["position_ids"] = position_ids_batch
-        sample["length"] = lengths_batch
-
+    sample_len = len(sample["input_ids"])
+    sample["position_ids"] = torch.arange(len(sample["input_ids"]))
+    sample["length"] = sample_len
     return sample
 
 
@@ -209,31 +172,10 @@ def add_length(sample):
 
 
 def drop_long_seq(sample, sequence_len=2048, min_sequence_len=2):
-    """
-    Drop samples whose sequence length is either too long (> sequence_len)
-    or too short (< min_sequence_len).
-
-    Works for both single-example (list[int]) or batched (list[list[int]]).
-    """
-    input_ids = sample["input_ids"]
-
-    # Edge case: if input_ids is empty
-    if not input_ids:
-        # Decide if you want to drop or keep empty. Let's drop.
-        return False
-
-    # Check if single example or batched by looking at the first element
-    if isinstance(input_ids[0], int):
-        # Single example (input_ids is a list of int)
-        length = len(input_ids)
-        return min_sequence_len <= length <= sequence_len
-
-    # Batched (input_ids is a list of lists)
-    results = []
-    for seq in input_ids:
-        length = len(seq)
-        results.append(min_sequence_len <= length <= sequence_len)
-    return results
+    return (
+        len(sample["input_ids"]) <= sequence_len
+        and len(sample["input_ids"]) >= min_sequence_len
+    )
 
 
 def process_datasets_for_packing(cfg, train_dataset, eval_dataset):
@@ -243,13 +185,10 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset):
         min_sequence_len=cfg.min_sample_len or 2,
     )
 
-    try:
-        min_input_len = np.min(get_dataset_lengths(train_dataset))
-        LOG.debug(f"min_input_len: {min_input_len}", main_process_only=True)
-        max_input_len = np.max(get_dataset_lengths(train_dataset))
-        LOG.debug(f"max_input_len: {max_input_len}", main_process_only=True)
-    except AttributeError:
-        pass
+    min_input_len = np.min(get_dataset_lengths(train_dataset))
+    LOG.debug(f"min_input_len: {min_input_len}", main_process_only=True)
+    max_input_len = np.max(get_dataset_lengths(train_dataset))
+    LOG.debug(f"max_input_len: {max_input_len}", main_process_only=True)
 
     if cfg.model_config_type == "mamba":
         LOG.info("dropping attention_mask column")
@@ -264,108 +203,59 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset):
         if eval_dataset and "token_type_ids" in eval_dataset.column_names:
             eval_dataset = eval_dataset.remove_columns("token_type_ids")
 
-    filter_map_kwargs = {}
-    if not isinstance(train_dataset, IterableDataset):
-        filter_map_kwargs["num_proc"] = cfg.dataset_processes
-        filter_map_kwargs["load_from_cache_file"] = not cfg.is_preprocess
-
-    try:
-        prior_len = len(train_dataset)
-    except TypeError:
-        # handle iterable datasets case
-        prior_len = None
-    drop_long_kwargs = {}
-    if filter_map_kwargs:
-        drop_long_kwargs["desc"] = "Dropping Long Sequences"
+    prior_len = len(train_dataset)
     train_dataset = train_dataset.filter(
         drop_long,
-        **filter_map_kwargs,
-        **drop_long_kwargs,
+        num_proc=cfg.dataset_processes,
+        load_from_cache_file=not cfg.is_preprocess,
+        desc="Dropping Long Sequences",
     )
-    if prior_len:
-        dropped = prior_len - len(train_dataset)
-        if dropped:
-            LOG.warning(f"Dropped {dropped} long samples from train dataset")
+    dropped = prior_len - len(train_dataset)
+    if dropped:
+        LOG.warning(f"Dropped {dropped} long samples from train dataset")
 
     if eval_dataset:
-        try:
-            prior_len = len(eval_dataset)
-        except TypeError:
-            # handle iterable datasets case
-            prior_len = None
+        prior_len = len(eval_dataset)
         eval_dataset = eval_dataset.filter(
             drop_long,
-            **filter_map_kwargs,
-            **drop_long_kwargs,
+            num_proc=cfg.dataset_processes,
+            load_from_cache_file=not cfg.is_preprocess,
+            desc="Dropping Long Sequences",
         )
-        if prior_len:
-            dropped = prior_len - len(eval_dataset)
-            if dropped:
-                LOG.warning(f"Dropped {dropped} long samples from eval dataset")
+        dropped = prior_len - len(eval_dataset)
+        if dropped:
+            LOG.warning(f"Dropped {dropped} long samples from eval dataset")
 
+    # drop samples with where the number of elements with labels not equal to -100 is zero
     def drop_no_trainable_tokens(sample):
-        """
-        Drop samples if all labels are -100 (i.e., zero trainable tokens).
-        Works for both single-example or batched input.
-        """
-        labels = sample["labels"]
-        if not labels:
-            # Edge case: if labels is empty, decide if you want to keep or drop
-            return True  # or False
+        return np.sum(np.array(sample["labels"]) != -100) > 0
 
-        # Check if single example or batch
-        # If first element is an int, we assume a single example
-        # If it's a list, we assume we're dealing with a batch
-        if isinstance(labels[0], int):
-            # Single example: return a single bool
-            return np.sum(np.array(labels) != -100) > 0
-
-        # Batched: 'labels' is a list of lists
-        # Return a list of booleans, one per sub-list
-        results = []
-        for row_labels in labels:
-            # Each row_labels is a list[int]
-            results.append(np.sum(np.array(row_labels) != -100) > 0)
-        return results
-
-    try:
-        prior_len = len(train_dataset)
-    except TypeError:
-        # handle iterable datasets case
-        prior_len = None
-    drop_long_kwargs = {}
-    if filter_map_kwargs:
-        drop_long_kwargs["desc"] = "Drop Samples with Zero Trainable Tokens"
+    prior_len = len(train_dataset)
     train_dataset = train_dataset.filter(
         drop_no_trainable_tokens,
-        batched=True,
-        **filter_map_kwargs,
-        **drop_long_kwargs,
+        num_proc=cfg.dataset_processes,
+        load_from_cache_file=not cfg.is_preprocess,
+        desc="Drop Samples with Zero Trainable Tokens",
     )
-    if prior_len:
-        dropped = prior_len - len(train_dataset)
-        if dropped:
-            LOG.warning(
-                f"Dropped {dropped} samples with no trainable tokens from train dataset"
-            )
+    dropped = prior_len - len(train_dataset)
+    if dropped:
+        LOG.warning(
+            f"Dropped {dropped} samples with no trainable tokens from train dataset"
+        )
 
     if eval_dataset:
-        try:
-            prior_len = len(eval_dataset)
-        except TypeError:
-            # handle iterable datasets case
-            prior_len = None
+        prior_len = len(eval_dataset)
         eval_dataset = eval_dataset.filter(
             drop_no_trainable_tokens,
-            **filter_map_kwargs,
-            **drop_long_kwargs,
+            num_proc=cfg.dataset_processes,
+            load_from_cache_file=not cfg.is_preprocess,
+            desc="Drop Samples with Zero Trainable Tokens",
         )
-        if prior_len:
-            dropped = prior_len - len(eval_dataset)
-            if dropped:
-                LOG.warning(
-                    f"Dropped {dropped} samples with no trainable tokens from eval dataset"
-                )
+        dropped = prior_len - len(eval_dataset)
+        if dropped:
+            LOG.warning(
+                f"Dropped {dropped} samples with no trainable tokens from eval dataset"
+            )
 
     if cfg.group_by_length:
         train_dataset = train_dataset.map(
@@ -401,21 +291,19 @@ def process_datasets_for_packing(cfg, train_dataset, eval_dataset):
                     desc="Add position_id column (PoSE)",
                 )
     elif cfg.sample_packing:
-        drop_long_kwargs = {}
-        if filter_map_kwargs:
-            drop_long_kwargs["desc"] = "Add position_id column (Sample Packing)"
         train_dataset = train_dataset.map(
             add_position_ids,
-            batched=True,
-            **filter_map_kwargs,
-            **drop_long_kwargs,
+            num_proc=cfg.dataset_processes,
+            load_from_cache_file=not cfg.is_preprocess,
+            desc="Add position_id column (Sample Packing)",
         )
         if cfg.eval_sample_packing is not False:
             if eval_dataset:
                 eval_dataset = eval_dataset.map(
                     add_position_ids,
-                    **filter_map_kwargs,
-                    **drop_long_kwargs,
+                    num_proc=cfg.dataset_processes,
+                    load_from_cache_file=not cfg.is_preprocess,
+                    desc="Add position_id column (Sample Packing)",
                 )
 
     return train_dataset, eval_dataset
@@ -446,7 +334,7 @@ def calculate_total_num_steps(cfg, train_dataset, update=True):
         and not cfg.reward_model
     ):
         total_num_tokens = np.sum(
-            train_dataset.select_columns("input_ids")
+            train_dataset.data.column("input_ids")
             .to_pandas()
             .apply(lambda x: len(x))  # pylint: disable=unnecessary-lambda
             .values

src/axolotl/utils/yaml.py

@@ -0,0 +1,157 @@
+"""Utilities for YAML files."""
+
+from collections import OrderedDict
+from typing import Any, Dict, List, Set, Tuple, Union
+
+import yaml
+
+
+class YAMLOrderTracker:
+    """Tracks the order of keys and section breaks in YAML files."""
+
+    def __init__(self, yaml_path: str):
+        self.yaml_path = yaml_path
+        self.structure, self.needs_break = self._parse_yaml_structure()
+
+    def _get_indentation_level(self, line: str) -> int:
+        """Get the indentation level of a line."""
+        return len(line) - len(line.lstrip())
+
+    def _parse_yaml_structure(
+        self,
+    ) -> Tuple[Dict[str, Union[List[str], Dict]], Set[str]]:
+        """Parse the YAML file to extract structure and identify section breaks."""
+        with open(self.yaml_path, "r", encoding="utf-8") as file:
+            contents = file.readlines()
+
+        structure: OrderedDict = OrderedDict()
+        needs_break = set()  # Track which keys should have a break before them
+        current_path = []
+        last_indentation = -1
+        had_empty_line = False
+
+        for line in contents:
+            # Track empty lines and comments
+            if not line.strip() or line.strip().startswith("#"):
+                had_empty_line = True
+                continue
+
+            # Get indentation level and content
+            indentation = self._get_indentation_level(line)
+            content = line.strip()
+
+            # Skip lines that don't define keys
+            if ":" not in content:
+                continue
+
+            # Extract key
+            key = content.split(":")[0].strip()
+
+            # If this is a top-level key and we had an empty line, mark it
+            if indentation == 0:
+                if had_empty_line:
+                    needs_break.add(key)
+                had_empty_line = False
+
+            # Handle indentation changes
+            if indentation > last_indentation:
+                current_path.append(key)
+            elif indentation < last_indentation:
+                levels_up = (last_indentation - indentation) // 2
+                current_path = current_path[:-levels_up]
+                current_path[-1] = key
+            else:
+                if current_path:
+                    current_path[-1] = key
+
+            # Update structure
+            current_dict = structure
+            for path_key in current_path[:-1]:
+                if path_key not in current_dict:
+                    current_dict[path_key] = OrderedDict()
+                current_dict = current_dict[path_key]
+
+            if current_path:
+                if current_path[-1] not in current_dict:
+                    current_dict[current_path[-1]] = OrderedDict()
+
+            last_indentation = indentation
+
+        return structure, needs_break
+
+
+class OrderedDumper(yaml.SafeDumper):
+    """Custom YAML dumper that maintains dictionary order."""
+
+
+def represent_none(self, _):
+    """Represent None values as empty fields."""
+    return self.represent_scalar("tag:yaml.org,2002:null", "")
+
+
+def ordered_dict_representer(dumper: OrderedDumper, data: Dict) -> Any:
+    """Custom representer for dictionaries that maintains order."""
+    return dumper.represent_mapping("tag:yaml.org,2002:map", data.items())
+
+
+def reorder_dict(data: Dict, reference_structure: Dict) -> OrderedDict:
+    """Reorder a dictionary based on a reference structure."""
+    ordered = OrderedDict()
+
+    # First add keys that are in the reference order
+    for key in reference_structure:
+        if key in data:
+            if isinstance(reference_structure[key], dict) and isinstance(
+                data[key], dict
+            ):
+                ordered[key] = reorder_dict(data[key], reference_structure[key])
+            else:
+                ordered[key] = data[key]
+
+    # Then add any remaining keys that weren't in the reference
+    for key in data:
+        if key not in ordered:
+            ordered[key] = data[key]
+
+    return ordered
+
+
+def dump_yaml_preserved_order(
+    data: Dict, reference_yaml_path: str, output_path: str
+) -> None:
+    """Dump YAML file while preserving nested order and normalized spacing."""
+    # Get reference structure and spacing
+    tracker = YAMLOrderTracker(reference_yaml_path)
+
+    # Reorder the data
+    ordered_data = reorder_dict(data, tracker.structure)
+
+    # Register the custom representers
+    OrderedDumper.add_representer(type(None), represent_none)
+    OrderedDumper.add_representer(dict, ordered_dict_representer)
+    OrderedDumper.add_representer(OrderedDict, ordered_dict_representer)
+
+    # First dump to string
+    yaml_str = yaml.dump(
+        ordered_data, Dumper=OrderedDumper, sort_keys=False, default_flow_style=False
+    )
+
+    # Add spacing according to reference
+    lines = yaml_str.split("\n")
+    result_lines: List[str] = []
+    current_line = 0
+
+    while current_line < len(lines):
+        line = lines[current_line]
+        if line.strip() and ":" in line and not line.startswith(" "):  # Top-level key
+            key = line.split(":")[0].strip()
+            if key in tracker.needs_break:
+                # Add single empty line before this key
+                if result_lines and result_lines[-1] != "":
+                    result_lines.append("")
+        result_lines.append(line)
+        current_line += 1
+
+    # Write the final result
+    with open(output_path, "w", encoding="utf-8") as file:
+        file.write("\n".join(result_lines))

tests/cli/conftest.py

@@ -1,4 +1,5 @@
 """Shared pytest fixtures for cli module."""
+
 import pytest
 from click.testing import CliRunner
 

tests/cli/test_cli_base.py

@@ -43,14 +43,12 @@ class BaseCliTest:
             result = cli_runner.invoke(cli, [command, str(config_path)])
 
             assert mock.called
-            assert mock.call_args.args[0] == [
+            assert mock.call_args.args[0][:5] == [
                 "accelerate",
                 "launch",
                 "-m",
                 f"axolotl.cli.{command}",
                 str(config_path),
-                "--debug-num-examples",
-                "0",
             ]
             assert mock.call_args.kwargs == {"check": True}
             assert result.exit_code == 0

tests/cli/test_cli_fetch.py

@@ -1,4 +1,5 @@
 """pytest tests for axolotl CLI fetch command."""
+
 from unittest.mock import patch
 
 from axolotl.cli.main import fetch

tests/cli/test_cli_inference.py

@@ -1,4 +1,5 @@
 """pytest tests for axolotl CLI inference command."""
+
 from unittest.mock import patch
 
 from axolotl.cli.main import cli

tests/cli/test_cli_interface.py

@@ -1,4 +1,5 @@
 """General pytest tests for axolotl.cli.main interface."""
+
 from axolotl.cli.main import build_command, cli
 
 
@@ -22,6 +23,7 @@ def test_build_command():
         "--batch-size",
         "8",
         "--debug",
+        "--nouse-fp16",
     ]
 
 

tests/cli/test_cli_merge_lora.py

@@ -1,4 +1,5 @@
 """pytest tests for axolotl CLI merge_lora command."""
+
 from unittest.mock import patch
 
 from axolotl.cli.main import cli

tests/cli/test_cli_merge_sharded_fsdp_weights.py

@@ -1,5 +1,6 @@
 """pytest tests for axolotl CLI merge_sharded_fsdp_weights command."""
 # pylint: disable=duplicate-code
+
 from unittest.mock import patch
 
 from axolotl.cli.main import cli

tests/cli/test_cli_preprocess.py

@@ -1,4 +1,5 @@
 """pytest tests for axolotl CLI preprocess command."""
+
 import shutil
 from pathlib import Path
 from unittest.mock import patch

tests/cli/test_cli_shard.py

@@ -1,5 +1,6 @@
 """pytest tests for axolotl CLI shard command."""
 # pylint: disable=duplicate-code
+
 from unittest.mock import patch
 
 from axolotl.cli.main import cli
@@ -11,14 +12,12 @@ def test_shard_with_accelerate(cli_runner, config_path):
         result = cli_runner.invoke(cli, ["shard", str(config_path), "--accelerate"])
 
         assert mock.called
-        assert mock.call_args.args[0] == [
+        assert mock.call_args.args[0][:5] == [
             "accelerate",
             "launch",
             "-m",
             "axolotl.cli.shard",
             str(config_path),
-            "--debug-num-examples",
-            "0",
         ]
         assert mock.call_args.kwargs == {"check": True}
         assert result.exit_code == 0

tests/cli/test_cli_version.py

@@ -1,4 +1,5 @@
 """pytest tests for axolotl CLI --version"""
+
 from axolotl.cli.main import cli
 
 

tests/cli/test_utils.py

@@ -1,5 +1,6 @@
 """pytest tests for axolotl CLI utils."""
 # pylint: disable=redefined-outer-name
+
 import json
 from unittest.mock import Mock, patch
 

tests/e2e/integrations/test_kd.py

@@ -1,121 +0,0 @@
-"""
-e2e tests for kd trainer support in Axolotl
-"""
-from pathlib import Path
-
-import pytest
-from e2e.utils import check_tensorboard, require_torch_2_5_1
-
-from axolotl.cli import load_datasets
-from axolotl.common.cli import TrainerCliArgs
-from axolotl.train import train
-from axolotl.utils.config import normalize_config, prepare_plugins
-from axolotl.utils.dict import DictDefault
-
-
-@pytest.fixture(name="kd_min_cfg")
-def min_cfg(temp_dir):
-    return {
-        "base_model": "osllmai-community/Llama-3.2-1B",
-        "tokenizer_config": "axolotl-ai-co/Llama-3.3-70B-Instruct-tokenizer",
-        "plugins": [
-            "axolotl.integrations.kd.KDPlugin",
-            "axolotl.integrations.liger.LigerPlugin",
-        ],
-        "liger_rms_norm": True,
-        "liger_glu_activation": True,
-        "torch_compile": True,
-        "chat_template": "llama3",
-        "kd_trainer": True,
-        "kd_ce_alpha": 0.1,
-        "kd_alpha": 0.9,
-        "kd_temperature": 2.0,
-        "dataloader_prefetch_factor": 8,
-        "dataloader_num_workers": 4,
-        "dataloader_pin_memory": True,
-        "datasets": [
-            {
-                "path": "axolotl-ai-co/evolkit-logprobs-pipeline-75k-v2-sample",
-                "type": "axolotl.integrations.kd.chat_template",
-                "field_messages": "messages_combined",
-                "split": "train",
-                "logprobs_field": "llm_text_generation_vllm_logprobs",
-                "temperature": 1.0,
-                "preprocess_shards": 2,
-            },
-        ],
-        "val_set_size": 0.0,
-        "sequence_len": 2048,
-        "sample_packing": True,
-        "pad_to_sequence_len": True,
-        "gradient_accumulation_steps": 2,
-        "micro_batch_size": 1,
-        "num_epochs": 1,
-        "optimizer": "adamw_8bit",
-        "lr_scheduler": "cosine",
-        "learning_rate": 0.00001,
-        "bf16": "auto",
-        "gradient_checkpointing": True,
-        "flash_attention": True,
-        "special_tokens": {
-            "pad_token": "<|end_of_text|>",
-            "eos_token": "<|eot_id|>",
-        },
-        "max_steps": 5,
-        "output_dir": temp_dir,
-        "save_safetensors": True,
-        "use_tensorboard": True,
-    }
-
-
-class TestKnowledgeDistillation:
-    """
-    Test case for Knowledge Distillation
-    """
-
-    # While this will run on torch 2.4.x without torch_compile enabled
-    # the VRAM requirement is higher than what is available in CI
-    @require_torch_2_5_1
-    def test_llama_kd(self, temp_dir, kd_min_cfg):
-        cfg = DictDefault(kd_min_cfg)
-        # pylint: disable=duplicate-code
-        prepare_plugins(cfg)
-        normalize_config(cfg)
-        cli_args = TrainerCliArgs()
-        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
-
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "model.safetensors").exists()
-        check_tensorboard(
-            temp_dir + "/runs", "train/loss", 1.0, "Train Loss is too high"
-        )
-
-    @pytest.mark.parametrize(
-        "load_in_8bit",
-        [True, False],
-    )
-    def test_llama_lora_kd(self, temp_dir, kd_min_cfg, load_in_8bit):
-        cfg = DictDefault(
-            {
-                "load_in_8bit": load_in_8bit,
-                "torch_compile": False,
-                "adapter": "lora",
-                "peft_use_dora": True,
-                "lora_target_linear": True,
-                "lora_r": 16,
-                "lora_alpha": 32,
-                "lora_dropout": 0.0,
-            }
-            | kd_min_cfg
-        )
-        # pylint: disable=duplicate-code
-        prepare_plugins(cfg)
-        normalize_config(cfg)
-        cli_args = TrainerCliArgs()
-        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
-
-        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
-        assert (Path(temp_dir) / "adapter_model.safetensors").exists()
-        check_tensorboard(
-            temp_dir + "/runs", "train/loss", 1.0, "Train Loss is too high"
-        )

tests/e2e/integrations/test_liger.py

@@ -54,7 +54,6 @@ class LigerIntegrationTestCase:
                 "max_steps": 5,
             }
         )
-        # pylint: disable=duplicate-code
         prepare_plugins(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()
@@ -100,7 +99,6 @@ class LigerIntegrationTestCase:
                 "max_steps": 5,
             }
         )
-        # pylint: disable=duplicate-code
         prepare_plugins(cfg)
         normalize_config(cfg)
         cli_args = TrainerCliArgs()