initial impl of streaming preprocessing

.coderabbit.yaml

@@ -12,6 +12,5 @@ reviews:
   auto_review:
     enabled: true
     drafts: false
-    auto_incremental_review: true
 chat:
   auto_reply: true

examples/gpt-oss/README.md

@@ -41,12 +41,6 @@ model, and final model output, you may need at least 3TB of free disk space to k
 axolotl train examples/gpt-oss/gpt-oss-120b-fft-fsdp2-offload.yaml
 ```
 
-To simplify fine-tuning across 2 nodes × 8x H100 (80GB) GPUs, we've partnered with [Baseten](https://baseten.co) to showcase multi-node
-training of the 120B model using Baseten Truss. You can read more about this recipe on
-[Baseten's blog](https://www.baseten.co/blog/how-to-fine-tune-gpt-oss-120b-with-baseten-and-axolotl/). The recipe can
-be found on their
-[GitHub](https://github.com/basetenlabs/ml-cookbook/tree/main/examples/oss-gpt-120b-axolotl/training).
-
 ERRATA: Transformers saves the model Architecture prefixed with `FSDP` which needs to be manually renamed in `config.json`.
 See https://github.com/huggingface/transformers/pull/40207 for the status of this issue.
 
@@ -67,23 +61,9 @@ mv ./outputs/gpt-oss-out/merged/* ./outputs/gpt-oss-out/
 
 ### Inferencing your fine-tuned model
 
-#### vLLM
-
 GPT-OSS support in vLLM does not exist in a stable release yet. See https://x.com/MaziyarPanahi/status/1955741905515323425
 for more information about using a special vllm-openai docker image for inferencing with vLLM.
 
-Optionally, vLLM can be installed from nightly:
-
-```bash
-pip install --no-build-isolation --pre -U vllm --extra-index-url https://wheels.vllm.ai/nightly
-```
-and the vLLM server can be started with the following command (modify `--tensor-parallel-size 8` to match your environment):
-```bash
-vllm serve ./outputs/gpt-oss-out/ --served-model-name axolotl/gpt-oss-20b --host 0.0.0.0 --port 8888  --tensor-parallel-size 8
-```
-
-#### SGLang
-
 SGLang has 0-day support in main, see https://github.com/sgl-project/sglang/issues/8833 for infomation on installing
 SGLang from source. Once you've installed SGLang, run the following command to launch a SGLang server:
 

examples/gpt-oss/gpt-oss-120b-fft-fsdp2-offload.yaml

@@ -44,7 +44,7 @@ bf16: true
 tf32: true
 
 flash_attention: true
-attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
+attn_implementation: kernels-community/vllm-flash-attn3
 
 gradient_checkpointing: true
 activation_offloading: true

examples/gpt-oss/gpt-oss-20b-fft-deepspeed-zero3.yaml

@@ -40,7 +40,7 @@ bf16: true
 tf32: true
 
 flash_attention: true
-attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
+attn_implementation: kernels-community/vllm-flash-attn3
 
 gradient_checkpointing: true
 activation_offloading: true

examples/gpt-oss/gpt-oss-20b-fft-fsdp2-offload.yaml

@@ -15,7 +15,7 @@ datasets:
     field_thinking: thinking
     template_thinking_key: thinking
 
-dataset_prepared_path: ./outputs/last_run_prepared
+dataset_prepared_path: last_run_prepared
 val_set_size: 0
 output_dir: ./outputs/gpt-oss-out/
 
@@ -41,7 +41,7 @@ bf16: true
 tf32: true
 
 flash_attention: true
-attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
+attn_implementation: kernels-community/vllm-flash-attn3
 
 gradient_checkpointing: true
 activation_offloading: true

examples/gpt-oss/gpt-oss-20b-fft-fsdp2.yaml

@@ -15,7 +15,7 @@ datasets:
     field_thinking: thinking
     template_thinking_key: thinking
 
-dataset_prepared_path: ./outputs/last_run_prepared
+dataset_prepared_path: last_run_prepared
 val_set_size: 0
 output_dir: ./outputs/gpt-oss-out/
 
@@ -40,7 +40,7 @@ bf16: true
 tf32: true
 
 flash_attention: true
-attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
+attn_implementation: kernels-community/vllm-flash-attn3
 
 gradient_checkpointing: true
 activation_offloading: true

examples/gpt-oss/gpt-oss-20b-sft-lora-singlegpu.yaml

@@ -53,7 +53,7 @@ bf16: true
 tf32: true
 
 flash_attention: true
-attn_implementation: kernels-community/vllm-flash-attn3  # this is not needed if using flash_attn >= 2.8.3
+attn_implementation: kernels-community/vllm-flash-attn3
 
 gradient_checkpointing: true
 activation_offloading: true

setup.py

@@ -118,9 +118,9 @@ def get_package_version():
 
 
 extras_require = {
-    "flash-attn": ["flash-attn==2.8.3"],
+    "flash-attn": ["flash-attn==2.8.2"],
     "ring-flash-attn": [
-        "flash-attn==2.8.3",
+        "flash-attn==2.8.2",
         "ring-flash-attn>=0.1.7",
         "yunchang==0.6.0",
     ],

src/axolotl/cli/cloud/modal_.py

@@ -82,7 +82,7 @@ class ModalCloud(Cloud):
         return res
 
     def get_image(self):
-        docker_tag = "main-py3.11-cu126-2.7.1"
+        docker_tag = "main-py3.11-cu124-2.6.0"
         if self.config.docker_tag:
             docker_tag = self.config.docker_tag
         docker_image = f"axolotlai/axolotl:{docker_tag}"
@@ -200,7 +200,7 @@ class ModalCloud(Cloud):
         if family in ["a10", "a10g"]:
             return modal.gpu.A10G(count=count)
         if family == "h100":
-            return f"H100:{count}"
+            return modal.gpu.H100(count=count)
         if family == "t4":
             return modal.gpu.T4(count=count)
         if family == "l4":

src/axolotl/cli/inference.py

@@ -64,7 +64,7 @@ def do_inference(
             importlib.import_module("axolotl.prompters"), prompter
         )
     elif cfg.chat_template:
-        chat_template_str = get_chat_template(cfg.chat_template, tokenizer=tokenizer)
+        chat_template_str = get_chat_template(cfg.chat_template)
     elif 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/preprocess.py

@@ -97,8 +97,7 @@ def do_cli(
     """
     # pylint: disable=duplicate-code
     os.environ["AXOLOTL_IS_PREPROCESS"] = "1"
-    is_preprocess = kwargs.pop("is_preprocess", True)
-    parsed_cfg = load_cfg(config, is_preprocess=is_preprocess, **kwargs)
+    parsed_cfg = load_cfg(config, **kwargs)
     parsed_cfg.is_preprocess = True
     parser = transformers.HfArgumentParser(PreprocessCliArgs)
     parsed_cli_args, _ = parser.parse_args_into_dataclasses(

src/axolotl/cli/utils/sweeps.py

@@ -3,12 +3,11 @@
 import random
 from copy import deepcopy
 from itertools import product
-from typing import Any
 
 
 def generate_sweep_configs(
     base_config: dict[str, list], sweeps_config: dict[str, list]
-) -> list[dict[str, Any]]:
+) -> list[dict[str, list]]:
     """
     Recursively generates all possible configurations by applying sweeps to the base config.
 

src/axolotl/cli/utils/train.py

@@ -4,7 +4,6 @@ import os
 import subprocess  # nosec
 import sys
 import tempfile
-from pathlib import Path
 from typing import Any, Iterator, Literal
 
 import yaml
@@ -89,12 +88,7 @@ def generate_config_files(config: str, sweep: str | None) -> Iterator[tuple[str,
     # Generate all possible configurations
     permutations = generate_sweep_configs(base_config, sweep_config)
     is_group = len(permutations) > 1
-    base_output_dir = base_config.get("output_dir", "./model-out")
-    for idx, permutation in enumerate(permutations, start=1):
-        permutation_dir = Path(permutation.get("output_dir", base_output_dir))
-        permutation_id = f"sweep{idx:04d}"
-        permutation["output_dir"] = str(permutation_dir / permutation_id)
-
+    for permutation in permutations:
         # pylint: disable=consider-using-with
         temp_file = tempfile.NamedTemporaryFile(
             mode="w",

src/axolotl/core/builders/causal.py

@@ -40,6 +40,7 @@ from axolotl.utils.collators import (
     BatchSamplerDataCollatorForSeq2Seq,
     DataCollatorForSeq2Seq,
     MambaDataCollator,
+    StreamingDataCollator,
     V2BatchSamplerDataCollatorForSeq2Seq,
 )
 from axolotl.utils.collators.mm_chat import MultiModalChatDataCollator
@@ -422,6 +423,17 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         is_eval=False,
         **kwargs,
     ):
+        from datasets import IterableDataset
+
+        if isinstance(self.train_dataset, IterableDataset) and not is_eval:
+            LOG.info("Using StreamingDataCollator")
+            return StreamingDataCollator(
+                tokenizer=self.tokenizer,
+                cfg=self.cfg,
+                prompter=None,
+                **kwargs,
+            )
+
         if training_args.pretraining:
             if (
                 self.cfg.pretraining_sample_concatenation is False

src/axolotl/datasets.py

@@ -43,7 +43,11 @@ class TokenizedPromptDataset(Dataset):
         )
 
     def process(self, dataset):
-        features = dataset.features.keys()
+        # For IterableDataset, we can't access features upfront
+        # We'll need to infer from the first batch
+        features = None
+        if hasattr(dataset, "features") and dataset.features:
+            features = dataset.features.keys()
 
         map_kwargs = {}
         if self.prompt_tokenizer.supports_batched:
@@ -54,18 +58,29 @@ class TokenizedPromptDataset(Dataset):
             hasattr(self.prompt_tokenizer, "filter_rows")
             and self.prompt_tokenizer.filter_rows
         ):
+            filter_kwargs = {"desc": "Strategy Filtering Rows"}
+            # Only add num_proc for regular datasets
+            if features is not None:
+                filter_kwargs["num_proc"] = self.process_count
+
             dataset = dataset.filter(
                 self.prompt_tokenizer.filter_rows,
-                num_proc=self.process_count,
-                desc="Strategy Filtering Rows",
+                **filter_kwargs,
             )
 
+        map_kwargs = {
+            **map_kwargs,
+            "desc": "Tokenizing Prompts",
+        }
+
+        # Only add remove_columns for regular datasets
+        if features is not None:
+            map_kwargs["remove_columns"] = features
+            map_kwargs["num_proc"] = self.process_count
+            map_kwargs["keep_in_memory"] = self.keep_in_memory
+
         return dataset.map(
             self.prompt_tokenizer.tokenize_prompt,
-            num_proc=self.process_count,
-            remove_columns=features,
-            keep_in_memory=self.keep_in_memory,
-            desc="Tokenizing Prompts",
             **map_kwargs,
         )
 

src/axolotl/monkeypatch/accelerate/fsdp2.py

@@ -187,7 +187,7 @@ def _process_lora_module_for_fsdp(module, fsdp2_kwargs):
 
     # Linear4Bit will keep it's bias term in fp32. If the weight dtype is in bf16 we are not able to
     # wrap this. Therefore we must ensure the bias has the same dtype as the weight
-    if hasattr(module.base_layer, "bias") and module.base_layer.bias is not None:
+    if module.base_layer.bias is not None:
         if module.base_layer.weight.dtype != module.base_layer.bias.dtype:
             log_bias_dtype_mismatch = True
             module.base_layer.bias.data = module.base_layer.bias.data.to(

src/axolotl/train.py

@@ -253,9 +253,7 @@ def save_trained_model(
             # final model weights have already been saved by `ReLoRACallback.on_train_end`
             return
 
-    if (  # pylint: disable=too-many-nested-blocks
-        trainer.is_fsdp_enabled or cfg.fsdp_config
-    ):
+    if trainer.is_fsdp_enabled or cfg.fsdp_config:
         if cfg.fsdp_config or cfg.fsdp:
             if cfg.fsdp_config.final_state_dict_type:
                 state_dict_type = cfg.fsdp_config.final_state_dict_type
@@ -287,8 +285,6 @@ def save_trained_model(
                 if trainer.accelerator.is_main_process:
                     # move all files in merged_path to cfg.output_dir
                     for merged_file in Path(merged_path).iterdir():
-                        if (Path(cfg.output_dir) / merged_file.name).exists():
-                            (Path(cfg.output_dir) / merged_file.name).unlink()
                         shutil.move(str(merged_file), cfg.output_dir)
                     shutil.rmtree(merged_path)  # remove what should be an empty dir
         # TODO(wing):see https://github.com/huggingface/transformers/pull/40207

src/axolotl/utils/collators/__init__.py

@@ -1,11 +1,19 @@
-"""
-shared axolotl collators for multipack, mamba, multimodal
-"""
+"""Shared axolotl collators for multipack, mamba, multimodal, etc."""
 
-from .batching import (  # noqa: F401
+from .batching import (
     BatchSamplerDataCollatorForSeq2Seq,
     DataCollatorForSeq2Seq,
     PretrainingBatchSamplerDataCollatorForSeq2Seq,
     V2BatchSamplerDataCollatorForSeq2Seq,
 )
-from .mamba import MambaDataCollator  # noqa: F401
+from .mamba import MambaDataCollator
+from .streaming import StreamingDataCollator
+
+__all__ = [
+    "BatchSamplerDataCollatorForSeq2Seq",
+    "DataCollatorForSeq2Seq",
+    "PretrainingBatchSamplerDataCollatorForSeq2Seq",
+    "V2BatchSamplerDataCollatorForSeq2Seq",
+    "MambaDataCollator",
+    "StreamingDataCollator",
+]

src/axolotl/utils/collators/streaming.py

@@ -0,0 +1,146 @@
+from dataclasses import dataclass
+from typing import Any, List
+
+import torch
+from transformers import PreTrainedTokenizerBase, default_data_collator
+from transformers.utils import PaddingStrategy
+
+from axolotl.prompters import Prompter
+from axolotl.utils.dict import DictDefault
+
+
+@dataclass
+class StreamingDataCollator:
+    tokenizer: PreTrainedTokenizerBase
+    cfg: DictDefault
+    prompter: Prompter | None = None
+    padding: bool | str | PaddingStrategy = True
+    max_length: int | None = None
+    pad_to_multiple_of: int | None = None
+    label_pad_token_id: int = -100
+    return_tensors: str = "pt"
+
+    def __post_init__(self):
+        if self.max_length is None:
+            self.max_length = self.cfg.sequence_len
+
+    def __call__(self, raw_batch: List[dict]) -> dict[str, Any]:
+        processed_samples = []
+
+        for raw_sample in raw_batch:
+            formatted_sample = raw_sample
+            if self.prompter:
+                formatted_sample = self._apply_prompt_formatting(raw_sample)
+
+            tokenized_sample = self._tokenize_sample(formatted_sample)
+
+            if len(tokenized_sample["input_ids"]) > self.max_length:
+                tokenized_sample = self._truncate_sample(tokenized_sample)
+
+            if tokenized_sample.get("input_ids"):
+                processed_samples.append(tokenized_sample)
+
+        return self._pad_and_batch(processed_samples)
+
+    def _apply_prompt_formatting(self, raw_sample: dict) -> dict:
+        formatted_text = self.prompter.build_prompt(
+            instruction=raw_sample.get("instruction", ""),
+            input=raw_sample.get("input", ""),
+            output=raw_sample.get("output", ""),
+        )
+        return {"text": formatted_text}
+
+    def _tokenize_sample(self, sample: dict) -> dict:
+        text = sample.get("text", sample.get("content", ""))
+
+        if not text:
+            instruction = sample.get("instruction", "")
+            input_text = sample.get("input", "")
+            output_text = sample.get("output", "")
+
+            parts = []
+            if instruction:
+                parts.append(f"Instruction: {instruction}")
+            if input_text:
+                parts.append(f"Input: {input_text}")
+            if output_text:
+                parts.append(f"Output: {output_text}")
+            text = "\n".join(parts)
+
+        if not text:
+            return {"input_ids": [], "attention_mask": [], "labels": []}
+
+        tokenized = self.tokenizer(
+            text,
+            truncation=False,
+            padding=False,
+            return_tensors=None,
+        )
+
+        tokenized["labels"] = tokenized["input_ids"].copy()
+        return tokenized
+
+    def _truncate_sample(self, tokenized_sample: dict) -> dict:
+        max_len = self.max_length
+        for key in ["input_ids", "attention_mask", "labels"]:
+            if key in tokenized_sample:
+                tokenized_sample[key] = tokenized_sample[key][:max_len]
+        return tokenized_sample
+
+    def _pad_and_batch(self, processed_samples: List[dict]) -> dict[str, Any]:
+        if not processed_samples:
+            processed_samples = [
+                {
+                    "input_ids": [self.tokenizer.eos_token_id],
+                    "attention_mask": [1],
+                    "labels": [self.tokenizer.eos_token_id],
+                }
+            ]
+
+        batch_samples = []
+        for sample in processed_samples:
+            batch_sample = {}
+            for key, value in sample.items():
+                if key in ["input_ids", "attention_mask", "labels"]:
+                    batch_sample[key] = torch.tensor(value, dtype=torch.long)
+            batch_samples.append(batch_sample)
+
+        if self.padding:
+            max_len_in_batch = max(len(sample["input_ids"]) for sample in batch_samples)
+
+            for sample in batch_samples:
+                current_len = len(sample["input_ids"])
+                pad_len = max_len_in_batch - current_len
+
+                if pad_len > 0:
+                    pad_token_id = (
+                        self.tokenizer.pad_token_id or self.tokenizer.eos_token_id
+                    )
+
+                    sample["input_ids"] = torch.cat(
+                        [
+                            sample["input_ids"],
+                            torch.full((pad_len,), pad_token_id, dtype=torch.long),
+                        ]
+                    )
+                    sample["attention_mask"] = torch.cat(
+                        [
+                            sample["attention_mask"],
+                            torch.zeros(pad_len, dtype=torch.long),
+                        ]
+                    )
+                    sample["labels"] = torch.cat(
+                        [
+                            sample["labels"],
+                            torch.full(
+                                (pad_len,), self.label_pad_token_id, dtype=torch.long
+                            ),
+                        ]
+                    )
+
+        batch = {}
+        for key in ["input_ids", "attention_mask", "labels"]:
+            if key in batch_samples[0]:
+                batch[key] = torch.stack([sample[key] for sample in batch_samples])
+
+        return batch

src/axolotl/utils/data/sft.py

@@ -9,6 +9,7 @@ from datasets import (
     Dataset,
     DatasetDict,
     IterableDataset,
+    IterableDatasetDict,
     load_dataset,
 )
 from transformers import PreTrainedTokenizer, ProcessorMixin
@@ -43,6 +44,18 @@ from axolotl.utils.trainer import (
 LOG = get_logger(__name__)
 
 
+def _determine_streaming_mode(cfg: DictDefault) -> bool:
+    """Determine if we should use streaming mode based on config."""
+    if cfg.streaming is not None:
+        return cfg.streaming
+
+    # Default to streaming for pretraining datasets
+    if cfg.pretraining_dataset:
+        return True
+
+    return False
+
+
 @retry_on_request_exceptions(max_retries=3, delay=5)
 def prepare_datasets(
     cfg: DictDefault,
@@ -61,11 +74,52 @@ def prepare_datasets(
     Returns:
         Tuple of (train_dataset, eval_dataset, total_steps, prompters).
     """
-    if cfg.pretraining_dataset:
-        return _prepare_pretraining_dataset(
-            cfg, tokenizer, processor, preprocess_iterable
+    streaming_mode = _determine_streaming_mode(cfg)
+
+    if streaming_mode:
+        if cfg.pretraining_dataset:
+            return _prepare_streaming_pretraining_dataset(cfg, tokenizer, processor)
+        else:
+            return _prepare_streaming_sft_dataset(cfg, tokenizer, processor)
+    else:
+        if cfg.pretraining_dataset:
+            return _prepare_pretraining_dataset(
+                cfg, tokenizer, processor, preprocess_iterable=False
+            )
+        else:
+            return _prepare_standard_dataset(
+                cfg, tokenizer, processor, preprocess_iterable=False
+            )
+
+
+def _prepare_streaming_sft_dataset(
+    cfg: DictDefault,
+    tokenizer: PreTrainedTokenizer,
+    processor: ProcessorMixin | None,
+) -> tuple[IterableDataset, Dataset | None, int, list[Prompter | None]]:
+    LOG.info("Loading streaming datasets")
+
+    raw_datasets = _load_raw_datasets_for_streaming(cfg, split="train")
+
+    eval_dataset = None
+    if cfg.test_datasets:
+        eval_raw_datasets = _load_raw_datasets_for_streaming(
+            cfg, split="test", dataset_configs=cfg.test_datasets
         )
-    return _prepare_standard_dataset(cfg, tokenizer, processor, preprocess_iterable)
+        eval_dataset = _process_eval_dataset_minimal(
+            eval_raw_datasets, cfg, tokenizer, processor
+        )
+    elif cfg.val_set_size:
+        LOG.info("Validation splits not supported for streaming datasets")
+
+    if not cfg.max_steps:
+        raise ValueError("max_steps must be set when using streaming datasets")
+
+    total_num_steps = cfg.max_steps
+    LOG.info(f"Maximum steps: {total_num_steps}")
+
+    prompters = [None] * len(cfg.datasets) if cfg.datasets else []
+    return raw_datasets, eval_dataset, total_num_steps, prompters
 
 
 def _prepare_standard_dataset(
@@ -373,7 +427,7 @@ def _load_and_process_single_dataset(
     d_base_type, d_prompt_style = _parse_dataset_type(dataset_config.type)
 
     # Select the appropriate split
-    if isinstance(dataset, DatasetDict):
+    if isinstance(dataset, (DatasetDict, IterableDatasetDict)):
         if dataset_config.split and dataset_config.split in dataset:
             dataset = dataset[dataset_config.split]
         elif split in dataset:
@@ -512,3 +566,78 @@ def _load_and_prepare_datasets(
         train_dataset, eval_dataset = _handle_test_dataset_split(dataset, cfg)
 
     return train_dataset, eval_dataset, prompters
+
+
+def _load_raw_datasets_for_streaming(
+    cfg: DictDefault, split: str = "train", dataset_configs: list | None = None
+) -> IterableDataset:
+    configs = (
+        dataset_configs
+        if dataset_configs is not None
+        else (cfg.datasets if split == "train" else cfg.test_datasets)
+    )
+
+    if not configs:
+        raise ValueError(f"No dataset configurations found for split '{split}'")
+
+    datasets = []
+    for dataset_config in datasets_with_name_generator(configs):
+        raw_dataset = load_dataset_with_config(
+            dataset_config, cfg.hf_use_auth_token, streaming=True
+        )
+
+        if isinstance(raw_dataset, (DatasetDict, IterableDatasetDict)):
+            if dataset_config.split and dataset_config.split in raw_dataset:
+                raw_dataset = raw_dataset[dataset_config.split]
+            elif split in raw_dataset:
+                raw_dataset = raw_dataset[split]
+            else:
+                raise ValueError(
+                    f"no {split} split found for dataset {dataset_config.path}, "
+                    "you may specify a split with 'split: ...'"
+                )
+
+        datasets.append(raw_dataset)
+
+    if len(datasets) == 1:
+        return datasets[0]
+    else:
+        return merge_datasets(datasets, cfg)
+
+
+def _process_eval_dataset_minimal(
+    raw_dataset: IterableDataset,
+    cfg: DictDefault,
+    tokenizer: PreTrainedTokenizer,
+    processor: ProcessorMixin | None,
+) -> Dataset | None:
+    LOG.info("Eval dataset processing skipped for streaming")
+    return None
+
+
+def _prepare_streaming_pretraining_dataset(
+    cfg: DictDefault,
+    tokenizer: PreTrainedTokenizer,
+    processor: ProcessorMixin | None,
+) -> tuple[IterableDataset, Dataset | None, int, list[Prompter | None]]:
+    pretraining_config = _extract_pretraining_config(cfg)
+
+    train_dataset = load_dataset_with_config(
+        pretraining_config, cfg.hf_use_auth_token, streaming=True
+    )
+
+    if isinstance(train_dataset, (DatasetDict, IterableDatasetDict)):
+        if pretraining_config.split and pretraining_config.split in train_dataset:
+            train_dataset = train_dataset[pretraining_config.split]
+        elif "train" in train_dataset:
+            train_dataset = train_dataset["train"]
+        else:
+            raise ValueError("no train split found for pretraining dataset")
+
+    if not cfg.max_steps:
+        raise ValueError("max_steps must be set when using streaming datasets")
+
+    total_num_steps = cfg.max_steps
+    LOG.info(f"Maximum steps: {total_num_steps}")
+
+    return train_dataset, None, total_num_steps, []

src/axolotl/utils/data/utils.py

@@ -190,12 +190,18 @@ def handle_long_seq_in_dataset(
     Returns:
         Filtered dataset with long sequences removed.
     """
-    if "input_ids" not in dataset.column_names:
-        LOG.warning(
-            "Dataset does not contain 'input_ids' column. Skip drop long seq. This is "
-            "expected for reward modeling."
-        )
-        return dataset
+    if hasattr(dataset, "column_names") and dataset.column_names:
+        if "input_ids" not in dataset.column_names:
+            LOG.warning(
+                "Dataset does not contain 'input_ids' column. Skip drop long seq. This is "
+                "expected for reward modeling."
+            )
+            return dataset
+    else:
+        # For IterableDataset, we can't check columns upfront, so skip for streaming
+        if isinstance(dataset, IterableDataset):
+            LOG.info("Skipping drop_long_seq for streaming datasets (not compatible)")
+            return dataset
 
     drop_long = functools.partial(
         drop_long_seq,

src/axolotl/utils/schemas/config.py

@@ -932,6 +932,34 @@ class AxolotlInputConfig(
 
     fix_untrained_tokens: int | list[int] | None = None
 
+    streaming: bool | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Whether to use streaming datasets (IterableDataset) for processing large datasets that don't fit in memory. When True, data is loaded on-demand during training without upfront preprocessing. Requires max_steps to be set. Pre-training datasets default to streaming unless explicitly set to False."
+        },
+    )
+
+    streaming_dataset_mixing_strategy: str | None = Field(
+        default="round_robin",
+        json_schema_extra={
+            "description": "Strategy for mixing multiple streaming datasets: 'round_robin' (equal sampling), 'weighted' (use streaming_mixing_weights), or 'random' (random sampling with equal probability)."
+        },
+    )
+
+    streaming_mixing_weights: list[float] | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Weights for weighted mixing strategy when using multiple streaming datasets. Must sum to 1.0 and have same length as datasets list. Only used when streaming_dataset_mixing_strategy='weighted'."
+        },
+    )
+
+    streaming_buffer_per_dataset: int | None = Field(
+        default=1000,
+        json_schema_extra={
+            "description": "Buffer size per dataset when mixing multiple streaming datasets. Higher values may improve mixing quality but use more memory."
+        },
+    )
+
     # INTERNALS - document for now, generally not set externally
     is_preprocess: bool | None = None
     preprocess_iterable: bool | None = None

src/axolotl/utils/schemas/validation.py

@@ -1337,6 +1337,30 @@ class GRPOVllmValidationMixin:
 
 
 # pylint: disable=too-many-ancestors
+class StreamingValidationMixin:
+    """Validation methods related to streaming datasets."""
+
+    @model_validator(mode="after")
+    def check_streaming_requires_max_steps(self):
+        """Ensure max_steps is set when using streaming datasets."""
+        # Check if streaming is explicitly enabled
+        streaming_enabled = getattr(self, "streaming", None) is True
+
+        # Check if pretraining dataset exists (defaults to streaming)
+        has_pretraining = getattr(self, "pretraining_dataset", None) is not None
+        streaming_default_for_pretraining = (
+            has_pretraining and getattr(self, "streaming", None) is None
+        )
+
+        # If streaming is enabled (explicitly or by default for pretraining)
+        if streaming_enabled or streaming_default_for_pretraining:
+            max_steps = getattr(self, "max_steps", None)
+            if not max_steps:
+                raise ValueError("max_steps must be set when using streaming datasets")
+
+        return self
+
+
 class ValidationMixin(
     DatasetValidationMixin,
     AttentionValidationMixin,
@@ -1347,6 +1371,7 @@ class ValidationMixin(
     SystemValidationMixin,
     ChatTemplateValidationMixin,
     PretrainingValidationMixin,
+    StreamingValidationMixin,
     ModelCompatibilityValidationMixin,
     ComplexValidationMixin,
     GRPOVllmValidationMixin,