log warning re: logged losses / gradient scaling per rank

* fix: remove unused config

* feat: add doc on dataset loading

* feat: enable azure and oci remote file system

* feat: add adlfs and ocifs to requirements

* fix: add links between dataset formats and dataset loading

* fix: remove unused condition

* Revert "fix: remove unused condition"

This reverts commit 5fe13be73e.

.github/workflows/multi-gpu-e2e.yml

@@ -24,6 +24,13 @@ jobs:
       fail-fast: false
       matrix:
         include:
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.6.0
+            axolotl_extras: vllm
+            num_gpus: 2
+            nightly_build: "true"
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -38,13 +45,6 @@ jobs:
             axolotl_extras: vllm
             num_gpus: 2
             nightly_build: "true"
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.6.0
-            axolotl_extras: vllm
-            num_gpus: 2
-            nightly_build: "true"
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     steps:

.github/workflows/tests.yml

@@ -211,7 +211,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.1
+            pytorch: 2.6.0
             num_gpus: 1
             axolotl_extras: vllm
     steps:
@@ -258,7 +258,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.5.1
             num_gpus: 1
             axolotl_extras: vllm
     steps:

_quarto.yml

@@ -231,6 +231,7 @@ website:
             - docs/reward_modelling.qmd
             - docs/lr_groups.qmd
             - docs/lora_optims.qmd
+            - docs/dataset_loading.qmd
 
         - section: "Core Concepts"
           contents:

docker/Dockerfile-base

@@ -29,7 +29,7 @@ ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
 WORKDIR /workspace
 
 RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel && \
-    python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} --extra-index-url https://download.pytorch.org/whl/cu$CUDA && \
+    python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} torchvision --extra-index-url https://download.pytorch.org/whl/cu$CUDA && \
     python3 -m pip install --no-cache-dir "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" && \
     python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"
 

docs/config.qmd

@@ -109,7 +109,7 @@ datasets:
     preprocess_shards: # Optional[int] process dataset in N sequential chunks for memory efficiency (exclusive with `shards`)
 
     name: # Optional[str] name of dataset configuration to load
-    train_on_split: train # Optional[str] name of dataset split to load from
+    split: train # Optional[str] name of dataset split to load from
     revision: # Optional[str] The specific revision of the dataset to use when loading from the Hugging Face Hub. This can be a commit hash, tag, or branch name. If not specified, the latest version will be used. This parameter is ignored for local datasets.
     trust_remote_code: # Optional[bool] Trust remote code for untrusted source
 
@@ -165,7 +165,9 @@ datasets:
       content: value
       # ...
 
-    # Optional[Dict[str, List]]. Roles mapping in the messages. The default is:
+    # Optional[Dict[str, List]]. Roles mapping in the messages.
+    # The format is {target_role: [source_roles]}. All source roles will be mapped to the target role.
+    # The default is:
     roles:
       user: ["human", "user"]
       assistant: ["gpt", "assistant"]

docs/dataset-formats/index.qmd

@@ -13,6 +13,13 @@ As there are a lot of available options in Axolotl, this guide aims to provide a
 
 Axolotl supports 3 kinds of training methods: pre-training, supervised fine-tuning, and preference-based post-training (e.g. DPO, ORPO, PRMs). Each method has their own dataset format which are described below.
 
+::: {.callout-tip}
+
+This guide will mainly use JSONL as an introduction. Please refer to the [dataset loading docs](../dataset_loading.qmd) to understand how to load datasets from other sources.
+
+For `pretraining_dataset:` specifically, please refer to the [Pre-training section](#pre-training).
+:::
+
 ## Pre-training
 
 When aiming to train on large corpora of text datasets, pre-training is your go-to choice. Due to the size of these datasets, downloading the entire-datasets before beginning training would be prohibitively time-consuming. Axolotl supports [streaming](https://huggingface.co/docs/datasets/en/stream) to only load batches into memory at a time.

docs/dataset_loading.qmd

@@ -0,0 +1,276 @@
+---
+title: Dataset Loading
+description: Understanding how to load datasets from different sources
+back-to-top-navigation: true
+toc: true
+toc-depth: 5
+---
+
+## Overview
+
+Datasets can be loaded in a number of different ways depending on the how it is saved (the extension of the file) and where it is stored.
+
+## Loading Datasets
+
+We use the `datasets` library to load datasets and a mix of `load_dataset` and `load_from_disk` to load them.
+
+You may recognize the similar named configs between `load_dataset` and the `datasets` section of the config file.
+
+```yaml
+datasets:
+  - path:
+    name:
+    data_files:
+    split:
+    revision:
+    trust_remote_code:
+```
+
+::: {.callout-tip}
+
+Do not feel overwhelmed by the number of options here. A lot of them are optional. In fact, the most common config to use would be `path` and sometimes `data_files`.
+
+:::
+
+This matches the API of [`datasets.load_dataset`](https://github.com/huggingface/datasets/blob/0b5998ac62f08e358f8dcc17ec6e2f2a5e9450b6/src/datasets/load.py#L1838-L1858), so if you're familiar with that, you will feel right at home.
+
+For HuggingFace's guide to load different dataset types, see [here](https://huggingface.co/docs/datasets/loading).
+
+For full details on the config, see [config.qmd](config.qmd).
+
+::: {.callout-note}
+
+You can set multiple datasets in the config file by more than one entry under `datasets`.
+
+```yaml
+datasets:
+  - path: /path/to/your/dataset
+  - path: /path/to/your/other/dataset
+```
+
+:::
+
+### Local dataset
+
+#### Files
+
+Usually, to load a JSON file, you would do something like this:
+
+```python
+from datasets import load_dataset
+
+dataset = load_dataset("json", data_files="data.json")
+```
+
+Which translates to the following config:
+
+```yaml
+datasets:
+  - path: json
+    data_files: /path/to/your/file.jsonl
+```
+
+However, to make things easier, we have added a few shortcuts for loading local dataset files.
+
+You can just point the `path` to the file or directory along with the `ds_type` to load the dataset. The below example shows for a JSON file:
+
+```yaml
+datasets:
+  - path: /path/to/your/file.jsonl
+    ds_type: json
+```
+
+This works for CSV, JSON, Parquet, and Arrow files.
+
+::: {.callout-tip}
+
+If `path` points to a file and `ds_type` is not specified, we will automatically infer the dataset type from the file extension, so you could omit `ds_type` if you'd like.
+
+:::
+
+#### Directory
+
+If you're loading a directory, you can point the `path` to the directory.
+
+Then, you have two options:
+
+##### Loading entire directory
+
+You do not need any additional configs.
+
+We will attempt to load in the following order:
+- datasets saved with `datasets.save_to_disk`
+- loading entire directory of files (such as with parquet/arrow files)
+
+```yaml
+datasets:
+  - path: /path/to/your/directory
+```
+
+##### Loading specific files in directory
+
+Provide `data_files` with a list of files to load.
+
+```yaml
+datasets:
+    # single file
+  - path: /path/to/your/directory
+    ds_type: csv
+    data_files: file1.csv
+
+    # multiple files
+  - path: /path/to/your/directory
+    ds_type: json
+    data_files:
+      - file1.jsonl
+      - file2.jsonl
+
+    # multiple files for parquet
+  - path: /path/to/your/directory
+    ds_type: parquet
+    data_files:
+      - file1.parquet
+      - file2.parquet
+
+```
+
+### HuggingFace Hub
+
+The method you use to load the dataset depends on how the dataset was created, whether a folder was uploaded directly or a HuggingFace Dataset was pushed.
+
+::: {.callout-note}
+
+If you're using a private dataset, you will need to enable the `hf_use_auth_token` flag in the root-level of the config file.
+
+:::
+
+#### Folder uploaded
+
+This would mean that the dataset is a single file or file(s) uploaded to the Hub.
+
+```yaml
+datasets:
+  - path: org/dataset-name
+    data_files:
+      - file1.jsonl
+      - file2.jsonl
+```
+
+#### HuggingFace Dataset
+
+This means that the dataset is created as a HuggingFace Dataset and pushed to the Hub via `datasets.push_to_hub`.
+
+```yaml
+datasets:
+  - path: org/dataset-name
+```
+
+::: {.callout-note}
+
+There are some other configs which may be required like `name`, `split`, `revision`, `trust_remote_code`, etc depending on the dataset.
+
+:::
+
+### Remote Filesystems
+
+Via the `storage_options` config under `load_dataset`, you can load datasets from remote filesystems like S3, GCS, Azure, and OCI.
+
+::: {.callout-warning}
+
+This is currently experimental. Please let us know if you run into any issues!
+
+:::
+
+The only difference between the providers is that you need to prepend the path with the respective protocols.
+
+```yaml
+datasets:
+    # Single file
+  - path: s3://bucket-name/path/to/your/file.jsonl
+
+    # Directory
+  - path: s3://bucket-name/path/to/your/directory
+```
+
+For directory, we load via `load_from_disk`.
+
+#### S3
+
+Prepend the path with `s3://`.
+
+The credentials are pulled in the following order:
+
+- `AWS_ACCESS_KEY_ID`, `AWS_SECRET_ACCESS_KEY`, and `AWS_SESSION_TOKEN` environment variables
+- from the `~/.aws/credentials` file
+- for nodes on EC2, the IAM metadata provider
+
+::: {.callout-note}
+
+We assume you have credentials setup and not using anonymous access. If you want to use anonymous access, let us know! We may have to open a config option for this.
+
+:::
+
+Other environment variables that can be set can be found in [boto3 docs](https://boto3.amazonaws.com/v1/documentation/api/latest/guide/configuration.html#using-environment-variables)
+
+#### GCS
+
+Prepend the path with `gs://` or `gcs://`.
+
+The credentials are loaded in the following order:
+
+- gcloud credentials
+- for nodes on GCP, the google metadata service
+- anonymous access
+
+#### Azure
+
+##### Gen 1
+
+Prepend the path with `adl://`.
+
+Ensure you have the following environment variables set:
+
+- `AZURE_STORAGE_TENANT_ID`
+- `AZURE_STORAGE_CLIENT_ID`
+- `AZURE_STORAGE_CLIENT_SECRET`
+
+##### Gen 2
+
+Prepend the path with `abfs://` or `az://`.
+
+Ensure you have the following environment variables set:
+
+- `AZURE_STORAGE_ACCOUNT_NAME`
+- `AZURE_STORAGE_ACCOUNT_KEY`
+
+Other environment variables that can be set can be found in [adlfs docs](https://github.com/fsspec/adlfs?tab=readme-ov-file#setting-credentials)
+
+#### OCI
+
+Prepend the path with `oci://`.
+
+It would attempt to read in the following order:
+
+- `OCIFS_IAM_TYPE`, `OCIFS_CONFIG_LOCATION`, and `OCIFS_CONFIG_PROFILE` environment variables
+- when on OCI resource, resource principal
+
+Other environment variables:
+
+- `OCI_REGION_METADATA`
+
+Please see the [ocifs docs](https://ocifs.readthedocs.io/en/latest/getting-connected.html#Using-Environment-Variables).
+
+### HTTPS
+
+The path should start with `https://`.
+
+```yaml
+datasets:
+  - path: https://path/to/your/dataset/file.jsonl
+```
+
+This must be publically accessible.
+
+## Next steps
+
+Now that you know how to load datasets, you can learn more on how to load your specific dataset format into your target output format [dataset formats docs](dataset-formats).

docs/multimodal.qmd

@@ -9,6 +9,7 @@ format:
 ## Supported Models
 
 - [Mllama](#sec-mllama)
+- [Llama4](#sec-llama4)
 - [Pixtral](#sec-pixtral)
 - [Llava-1.5](#sec-llava-15)
 - [Mistral-Small-3.1](#sec-mistral-small-31)
@@ -63,6 +64,14 @@ base_model: meta-llama/Llama-3.2-11B-Vision-Instruct
 chat_template: llama3_2_vision
 ```
 
+### Llama4 {#sec-llama4}
+
+```yaml
+base_model: meta-llama/Llama-4-Scout-17B-16E-Instruct
+
+chat_template: llama4
+```
+
 ### Pixtral {#sec-pixtral}
 
 ```yaml

examples/llama4/scout-lora.yaml

@@ -0,0 +1,75 @@
+base_model: meta-llama/Llama-4-Scout-17B-16E
+model_type: Llama4ForConditionalGeneration
+  # Automatically upload checkpoint and final model to HF
+  # hub_model_id: username/custom_model_name
+
+strict: false
+
+  # torch_compile: true
+
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_target_modules:
+  - self_attn.q_proj
+  - self_attn.k_proj
+  - self_attn.v_proj
+  - self_attn.o_proj
+lora_modules_to_save:
+  - lm_head
+  - embed_tokens
+
+chat_template: llama4
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_8bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+# gradient_checkpointing: true
+# gradient_checkpointing_kwargs:
+#   use_reentrant: false
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 100
+evals_per_epoch: 2
+saves_per_epoch: 1
+weight_decay: 0.0
+fsdp:
+  - auto_wrap
+  - full_shard
+fsdp_config:
+  fsdp_version: 2
+  fsdp_offload_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: Llama4TextDecoderLayer
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_reshard_after_forward: true
+  fsdp_activation_checkpointing: true
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot|>

requirements.txt

@@ -6,18 +6,19 @@ triton>=3.0.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
 autoawq==0.2.7.post3
-liger-kernel==0.5.5
+liger-kernel==0.5.6
 # END section
 
 packaging==23.2
 
-peft==0.15.0
+peft==0.15.1
 transformers==4.51.0
 tokenizers>=0.21.1
 accelerate==1.6.0
 datasets==3.5.0
 deepspeed>=0.15.4
 trl==0.16.1
+hf_xet==1.0.0
 
 optimum==1.16.2
 hf_transfer
@@ -48,7 +49,8 @@ python-dotenv==1.0.1
 # remote filesystems
 s3fs>=2024.5.0
 gcsfs>=2024.5.0
-# adlfs
+adlfs>=2024.5.0
+ocifs==1.3.2
 
 zstandard==0.22.0
 fastcore

src/axolotl/core/trainers/base.py

@@ -235,6 +235,9 @@ class AxolotlTrainer(
             self.accelerator.even_batches = False
 
         # Return unprepared dataloader if using sequence parallelism
+        # TODO(djsaunde): We might be able to use `accelerate`'s dataloader preparation
+        # if we use `dispatch_batches` and `slice_fn_for_dispatch` properly (i.e.,
+        # slice each batch along the sequence dimension).
         if self.args.sequence_parallel_degree > 1:
             return dataloader
 
@@ -562,6 +565,19 @@ class AxolotlTrainer(
 
         return res
 
+    def additional_accelerator_args(
+        self, fp8=None, **kwargs
+    ):  # pylint: disable=unused-argument
+        ret_kwargs = {}
+        if fp8:
+            from accelerate.utils import AORecipeKwargs
+
+            ret_kwargs["mixed_precision"] = "fp8"
+            ret_kwargs["kwargs_handlers"] = [AORecipeKwargs()]
+            os.environ["ACCELERATE_MIXED_PRECISION"] = "fp8"
+
+        return ret_kwargs
+
     def log(self, logs: dict[str, float], start_time: float | None = None) -> None:
         """
         Log `logs` on the various objects watching training, including stored metrics.

src/axolotl/core/trainers/mixins/sequence_parallel.py

@@ -1,34 +1,22 @@
 """Module for Axolotl trainer sequence parallelism mixin"""
 
 import logging
-from typing import Any
 
-import torch
 import torch.distributed as dist
-import torch.nn.functional as F
 from datasets import Dataset
-from torch import nn
 from torch.utils.data import DistributedSampler, Sampler
 
 from axolotl.monkeypatch.attention.ring_attn import get_ring_attn_group
 
 LOG = logging.getLogger(__name__)
 
-try:
-    from ring_flash_attn import update_ring_flash_attn_params
-except ImportError:
-    # We pass silently here, but raise an ImportError in our Axolotl config validation
-    # if cfg.sequence_parallel_degree > 1 and `ring-flash-attn` is not installed.
-    pass
-
 
 class SequenceParallelMixin:
     """
     Mixin class for sequence parallelism support in trainers.
 
     This mixin provides functionality for handling sequence parallelism,
-    including creating appropriate samplers, managing data partitioning,
-    and updating ring flash attention parameters during training.
+    specifically for creating appropriate data samplers.
     """
 
     args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
@@ -99,84 +87,3 @@ class SequenceParallelMixin:
         return self._create_sequence_parallel_sampler(
             eval_dataset, shuffle=False, is_eval=True
         )
-
-    def _update_ring_flash_attn_params(self, inputs: dict[str, torch.Tensor | Any]):
-        """
-        Calculate the cu_seqlens for the current forward pass and pass the value to
-        the substituted ring_flash_attn. This is accomplished by using the passed
-        `input_ids`.
-
-        Args:
-            inputs: Current batch of inputs.
-        """
-        # At this point, inputs should already be partitioned by the sequence
-        # parallel data collator
-        batch_size = inputs["input_ids"].shape[0]
-        seq_len = inputs["input_ids"].shape[1]
-        packed_seq_lens = [seq_len] * batch_size
-
-        # Calculate the full sequence length across all GPUs in this SP group
-        total_seq_len = seq_len * self.args.sequence_parallel_degree
-
-        cu_seqlens = torch.cumsum(
-            torch.tensor(
-                packed_seq_lens, device=torch.cuda.current_device(), dtype=torch.int32
-            ),
-            dim=-1,
-            dtype=torch.int32,
-        )
-        cu_seqlens = F.pad(
-            F.pad(cu_seqlens, (1, 0), value=0), (0, 1), value=total_seq_len
-        )
-
-        update_ring_flash_attn_params(cu_seqlens, self.ring_attn_group)
-
-    def training_step(
-        self,
-        model: nn.Module,
-        inputs: dict[str, torch.Tensor | Any],
-        num_items_in_batch: int | None = None,
-    ) -> torch.Tensor:
-        """
-        Perform a training step on a batch of inputs. Overrides the
-        `transformers.trainer.Trainer` method to handle sequence parallelism if
-        enabled.
-
-        Args:
-            model: Model to perform training step for.
-            inputs: Dictionary mapping.
-        """
-        # Set up sequence parallelism for this step if enabled
-        if self.args.sequence_parallel_degree > 1:
-            self._update_ring_flash_attn_params(inputs)
-
-        # Proceed with normal training step
-        return super().training_step(model, inputs, num_items_in_batch)  # type: ignore
-
-    def prediction_step(
-        self,
-        model: nn.Module,
-        inputs: dict[str, torch.Tensor | Any],
-        prediction_loss_only: bool,
-        ignore_keys: list[str] | None = None,
-    ) -> tuple[torch.Tensor | None, torch.Tensor | None, torch.Tensor | None]:
-        """
-        Perform a prediction step on a batch of inputs. Overrides the
-        `transformers.trainer.Trainer` method to handle sequence parallelism if
-        enabled.
-
-        Args:
-            model: Model to perform prediction step for.
-            inputs: Dictionary mapping of inputs.
-            prediction_loss_only: Whether to return only the loss.
-            ignore_keys: Keys to ignore in the inputs.
-
-        Returns:
-            Tuple of (loss, logits, labels).
-        """
-        # Set up sequence parallelism for this prediction step if enabled
-        if self.args.sequence_parallel_degree > 1:
-            self._update_ring_flash_attn_params(inputs)
-
-        # Proceed with normal prediction step
-        return super().prediction_step(model, inputs, prediction_loss_only, ignore_keys)  # type: ignore

src/axolotl/integrations/liger/__init__.py

@@ -173,5 +173,17 @@ class LigerPlugin(BasePlugin):
                 raise NotImplementedError(
                     "Fused linear cross entropy is not yet supported for Gemma3."
                 )
+        elif cfg.model_config_type == "llama4":
+            from axolotl.integrations.liger.models.llama4 import (
+                apply_liger_kernel_to_llama4,
+            )
+
+            apply_liger_kernel_to_llama4(
+                cross_entropy=cfg.liger_cross_entropy,
+                fused_linear_cross_entropy=cfg.liger_fused_linear_cross_entropy,
+                glu_activation=cfg.liger_glu_activation,
+                rms_norm=cfg.liger_rms_norm,
+                layer_norm=cfg.liger_layer_norm,
+            )
         elif cfg.model_config_type in ["deepseek_v3"]:
             raise ValueError(f"Unsupported model config type: {cfg.model_config_type}")

src/axolotl/integrations/liger/models/llama4.py

@@ -0,0 +1,171 @@
+"""
+Liger FLCE for llama4
+"""
+
+import sys
+from typing import List, Optional, Tuple, Union
+
+import torch
+from liger_kernel.transformers.model.loss_utils import LigerForCausalLMLoss
+from transformers.modeling_outputs import CausalLMOutputWithPast
+
+
+def lce_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[
+        Union["Cache", List[torch.FloatTensor]]  # noqa: F821
+    ] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **loss_kwargs,
+) -> Union[Tuple, CausalLMOutputWithPast]:
+    r"""
+    Args:
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+    """
+
+    # pylint: disable=duplicate-code
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+    outputs = self.model(
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+    )
+
+    hidden_states = outputs[0]
+
+    if hasattr(self.config, "pretraining_tp") and self.config.pretraining_tp > 1:
+        raise Exception(  # pylint: disable=broad-exception-raised
+            "Liger Kernel does not support pretraining_tp!!"
+        )
+
+    logits = None
+    loss = None
+    # if in training mode, don't materialize logits
+    if self.training and (labels is not None):
+        loss = LigerForCausalLMLoss(
+            hidden_states=hidden_states,
+            lm_head_weight=self.lm_head.weight,
+            labels=labels,
+            hidden_size=self.config.hidden_size,
+            **loss_kwargs,
+        )
+
+    else:  # if in inference mode materialize logits
+        slice_indices = (
+            slice(-logits_to_keep, None)
+            if isinstance(logits_to_keep, int)
+            else logits_to_keep
+        )
+        logits = self.lm_head(hidden_states[:, slice_indices, :])
+        if labels is not None:
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.vocab_size,
+                **loss_kwargs,
+            )
+
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+    )
+
+
+def apply_liger_kernel_to_llama4(
+    cross_entropy: bool = False,
+    fused_linear_cross_entropy: bool = False,
+    rms_norm: bool = False,
+    glu_activation: bool = False,
+    layer_norm: bool = False,
+    **kwargs,  # pylint: disable=unused-argument
+) -> None:
+    """
+    Apply Liger kernels to replace original implementation in HuggingFace Llama models (2 and 3)
+
+    Args:
+        cross_entropy (bool): Whether to apply Liger's cross entropy loss. Default is False.
+        fused_linear_cross_entropy (bool):
+            Whether to apply Liger's fused linear cross entropy loss. Default is False.
+            `cross_entropy` and `fused_linear_cross_entropy` cannot both be False.
+            If `fused_linear_cross_entropy` is True, the logits will not be materialized but more memory efficient.
+        rms_norm (bool): Whether to apply Liger's RMSNorm. Default is False.
+        glu_activation (bool): Whether to apply Liger's SwiGLU MLP. Default is False.
+        layer_norm (bool): Whether to apply Liger's LayerNorm. Default is False.
+    """
+
+    import transformers.models.llama4.modeling_llama4  # noqa: F401  # pylint: disable=unused-import
+    from liger_kernel.transformers.functional import liger_cross_entropy
+    from liger_kernel.transformers.layer_norm import LigerLayerNorm
+    from liger_kernel.transformers.rms_norm import LigerRMSNorm
+    from liger_kernel.transformers.swiglu import LigerSwiGLUMLP
+
+    assert not (
+        cross_entropy and fused_linear_cross_entropy
+    ), "cross_entropy and fused_linear_cross_entropy cannot both be True."
+
+    modeling_llama4 = sys.modules["transformers.models.llama4.modeling_llama4"]
+
+    if rms_norm:
+        modeling_llama4.Llama4TextRMSNorm = LigerRMSNorm
+    if glu_activation:
+        modeling_llama4.Llama4TextMLP = LigerSwiGLUMLP
+    if layer_norm:
+        modeling_llama4.nn.LayerNorm = LigerLayerNorm
+
+    if cross_entropy:
+        from transformers.loss.loss_utils import nn
+
+        nn.functional.cross_entropy = liger_cross_entropy
+
+    if fused_linear_cross_entropy:
+        modeling_llama4.Llama4ForCausalLM.forward = lce_forward

src/axolotl/monkeypatch/attention/flex_attn.py

@@ -162,7 +162,6 @@ def patch_flex_make_mask():
     for n in tuple(sys.modules):
         if ".modeling_" in n and "llama4" not in n:
             if hasattr(sys.modules[n], "make_flex_block_causal_mask"):
-                print(n)
                 sys.modules[n].make_flex_block_causal_mask = (
                     patched_make_flex_block_causal_mask
                 )

src/axolotl/monkeypatch/attention/ring_attn.py

@@ -6,10 +6,12 @@ package, specifically the `hf_adapter.substitute_hf_flash_attn` function to patc
 their sequence parallel version of Flash Attention 2.
 """
 
+import torch
 import torch.distributed as dist
 from accelerate.logging import get_logger
 
 from axolotl.logging_config import configure_logging
+from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
 
 configure_logging()
 LOG = get_logger(__name__)
@@ -98,3 +100,27 @@ def register_ring_attn(sequence_parallel_degree: int, heads_k_stride: int | None
     substitute_hf_flash_attn(
         process_group=get_ring_attn_group(), heads_k_stride=heads_k_stride
     )
+
+
+def update_ring_attn_params(batch: dict[str, torch.Tensor]):
+    """
+    Calculate the cumulative sequence lengths for the current forward pass and pass the
+    value to the substituted `ring_flash_attn`.
+
+    Args:
+        batch: A dictionary with a batch of data. May or may not contain `position_ids`
+            data; if not, we compute it.
+    """
+    from ring_flash_attn import update_ring_flash_attn_params
+
+    input_ids = batch["input_ids"]
+    position_ids = batch.get("position_ids")
+    if position_ids is None:
+        seq_len = input_ids.shape[1]
+        position_ids = torch.arange(
+            0, seq_len, dtype=torch.long, device=input_ids.device
+        ).unsqueeze(0)
+
+    cu_seqlens, _ = get_cu_seqlens_from_pos_ids(position_ids)
+    cu_seqlens = cu_seqlens.squeeze().to(device=torch.cuda.current_device())
+    update_ring_flash_attn_params(cu_seqlens, get_ring_attn_group())

src/axolotl/monkeypatch/multipack.py

@@ -13,6 +13,7 @@ from axolotl.monkeypatch.utils import get_unpad_data
 SUPPORTED_MULTIPACK_MODEL_TYPES = [
     "mllama_text_model",
     "llama",
+    "llama4",
     "mistral",
     "mixtral",
     "qwen2",

src/axolotl/monkeypatch/trainer_accelerator_args.py

@@ -0,0 +1,80 @@
+"""
+allow adding additional kwargs to Accelerator init
+"""
+
+import inspect
+import logging
+
+from transformers import Trainer
+
+from axolotl.monkeypatch.utils import detab_code
+
+LOG = logging.getLogger(__name__)
+
+ORIGINAL_TRAINER_CODE = """
+    # create accelerator object
+    self.accelerator = Accelerator(**args)
+"""
+
+PATCHED_TRAINER_CODE = """
+    if hasattr(self, "additional_accelerator_args"):
+        additional_args = self.additional_accelerator_args(fp8=True, **args)
+        if additional_args:
+            args.update(additional_args)
+
+    # create accelerator object
+    self.accelerator = Accelerator(**args)
+"""
+
+
+def get_create_accelerate_code() -> str:
+    training_loop = inspect.getsource(Trainer.create_accelerator_and_postprocess)
+    return training_loop
+
+
+def check_create_accelerate_code_is_patchable() -> bool:
+    create_code = get_create_accelerate_code()
+    create_code, _ = detab_code(create_code)
+    return ORIGINAL_TRAINER_CODE in create_code
+
+
+def patch_create_accelerate_code_for_fp8():
+    """
+    monkeypatch create_accelerator_and_postprocess so it checks for additional kwargs
+    """
+
+    try:
+        create_code = get_create_accelerate_code()
+    except OSError:
+        return
+    Trainer._original_create_accelerator_and_postprocess = (  # pylint: disable=protected-access
+        create_code
+    )
+    create_code, _ = detab_code(create_code)
+    if ORIGINAL_TRAINER_CODE not in create_code:
+        return
+
+    create_code = create_code.replace(ORIGINAL_TRAINER_CODE, PATCHED_TRAINER_CODE)
+    create_code = create_code.replace(
+        "def create_accelerator_and_postprocess(",
+        "def fixed_create_accelerator_and_postprocess(",
+        1,
+    )
+
+    # load imports necessary
+    import transformers.trainer
+
+    items_to_import = []
+    for item in dir(transformers.trainer):
+        if item in create_code:
+            items_to_import.append(item)
+
+    exec(  # pylint: disable=exec-used  # nosec B102
+        "from transformers.trainer import ("
+        + ", ".join(x for x in items_to_import)
+        + ")",
+        globals(),
+    )
+    exec(create_code, globals())  # pylint: disable=exec-used  # nosec B102
+    LOG.info("patching create_accelerator_and_postprocess to allow for overrides")
+    Trainer.create_accelerator_and_postprocess = fixed_create_accelerator_and_postprocess  # pylint: disable=protected-access  # pylint: disable=undefined-variable  # noqa: F821

src/axolotl/monkeypatch/utils.py

@@ -96,7 +96,9 @@ def get_cu_seqlens(attn_mask):
     return torch.stack(results).to(dtype=torch.int32), torch.stack(max_seq_lens)
 
 
-def get_cu_seqlens_from_pos_ids(position_ids):
+def get_cu_seqlens_from_pos_ids(
+    position_ids: torch.Tensor,
+) -> tuple[torch.Tensor, torch.Tensor]:
     """generate a cumulative sequence length mask for flash attention using pos ids"""
     if len(position_ids.shape) == 1:
         position_ids = position_ids.unsqueeze(0)

src/axolotl/processing_strategies.py

@@ -268,6 +268,7 @@ def get_processing_strategy(
         )
     if chat_template_type in [
         "llama3_2_vision",
+        "llama4",
         "llava",
         "mistral_v7_tekken",
         "pixtral",

src/axolotl/utils/collators/batching.py

@@ -3,7 +3,6 @@ Data collators for axolotl to pad labels and position_ids for packed sequences.
 includes logic for handling sequence parallelism collation.
 """
 
-import logging
 from dataclasses import dataclass
 from typing import Any, Optional, Union
 
@@ -13,46 +12,7 @@ import torch.distributed as dist
 from transformers import PreTrainedTokenizerBase
 from transformers.utils import PaddingStrategy
 
-logger = logging.getLogger(__name__)
-
-
-def adjust_position_ids_for_slice(
-    position_ids: torch.Tensor, start_idx: int
-) -> torch.Tensor:
-    """
-    Adjust position IDs for a sliced sequence to maintain proper relative positions.
-    This handles the case where position IDs might not be contiguous due to sample
-    packing.
-    """
-    # Convert to tensor if not already
-    # Find the boundaries between samples (where position_ids reset)
-    adjusted_pos_ids = position_ids.clone()
-
-    # Process each sequence in the batch
-    for i in range(position_ids.shape[0]):
-        seq = position_ids[i]
-
-        # Find sample boundaries
-        boundaries = []
-        for j in range(1, len(seq)):
-            if seq[j] < seq[j - 1]:
-                boundaries.append(j)
-
-        # No need to adjust if there are no boundaries or this is a single sample
-        if not boundaries:
-            adjusted_pos_ids[i] = seq - start_idx
-            continue
-
-        # Adjust each segment separately
-        prev_boundary = 0
-        for boundary in boundaries:
-            adjusted_pos_ids[i, prev_boundary:boundary] -= start_idx
-            prev_boundary = boundary
-
-        # Last segment
-        adjusted_pos_ids[i, prev_boundary:] -= start_idx
-
-    return adjusted_pos_ids
+from axolotl.monkeypatch.attention.ring_attn import update_ring_attn_params
 
 
 @dataclass
@@ -196,23 +156,20 @@ class DataCollatorForSeq2Seq:
         Returns:
             Sliced batch dictionary.
         """
-        keys_to_slice = ["input_ids", "attention_mask", "labels", "position_ids"]
+        # Get local (start, end) for sequence parallelism slicing
+        total_seq_len = batch["input_ids"].shape[1]
+        slice_size = total_seq_len // self.local_world_size
+        start = self.local_rank * slice_size
+        end = start + slice_size
 
+        # Update params for ring attention calculation
+        update_ring_attn_params(batch=batch)
+
+        # Slice batch for sequence parallel processing
+        keys_to_slice = ["input_ids", "attention_mask", "labels", "position_ids"]
         for key in keys_to_slice:
             if key in batch:
-                seq_len = batch[key].shape[1]
-                slice_size = seq_len // self.local_world_size
-                start_idx = self.local_rank * slice_size
-                end_idx = (
-                    start_idx + slice_size
-                    if self.local_rank < self.local_world_size - 1
-                    else seq_len
-                )
-                batch[key] = batch[key][:, start_idx:end_idx]
-
-                # Special handling for position_ids
-                if key == "position_ids" and self.local_rank > 0:
-                    batch[key] = adjust_position_ids_for_slice(batch[key], start_idx)
+                batch[key] = batch[key][:, start:end]
 
         return batch
 

src/axolotl/utils/data/shared.py

@@ -96,20 +96,17 @@ def load_dataset_w_config(
         pass
 
     ds_from_cloud = False
-    storage_options = {}
+    storage_options: dict = {}
     remote_file_system = None
     if config_dataset.path.startswith("s3://"):
         try:
-            import aiobotocore.session  # type: ignore
             import s3fs  # type: ignore
         except ImportError as exc:
-            raise ImportError(
-                "s3:// paths require aiobotocore and s3fs to be installed"
-            ) from exc
+            raise ImportError("s3:// paths require s3fs to be installed") from exc
 
-        # Takes credentials from ~/.aws/credentials for default profile
-        s3_session = aiobotocore.session.AioSession(profile="default")
-        storage_options = {"session": s3_session}
+        # Reads env, credentials from ~/.aws/credentials, or IAM metadata provider
+        # https://s3fs.readthedocs.io/en/latest/index.html?highlight=storage_options#credentials
+        storage_options = {"anon": False}
         remote_file_system = s3fs.S3FileSystem(**storage_options)
     elif config_dataset.path.startswith("gs://") or config_dataset.path.startswith(
         "gcs://"
@@ -125,28 +122,44 @@ def load_dataset_w_config(
         # https://gcsfs.readthedocs.io/en/latest/#credentials
         storage_options = {"token": None}
         remote_file_system = gcsfs.GCSFileSystem(**storage_options)
-    # TODO: Figure out how to get auth creds passed
-    # elif config_dataset.path.startswith("adl://") or config_dataset.path.startswith("abfs://"):
-    #     try:
-    #         import adlfs
-    #     except ImportError as exc:
-    #        raise ImportError(
-    #            "adl:// or abfs:// paths require adlfs to be installed"
-    #        ) from exc
+    elif (
+        config_dataset.path.startswith("adl://")
+        or config_dataset.path.startswith("abfs://")
+        or config_dataset.path.startswith("az://")
+    ):
+        try:
+            import adlfs
+        except ImportError as exc:
+            raise ImportError(
+                "adl:// or abfs:// paths require adlfs to be installed"
+            ) from exc
 
-    #     # Gen 1
-    #     storage_options = {
-    #         "tenant_id": TENANT_ID,
-    #         "client_id": CLIENT_ID,
-    #         "client_secret": CLIENT_SECRET,
-    #     }
-    #     # Gen 2
-    #     storage_options = {
-    #         "account_name": ACCOUNT_NAME,
-    #         "account_key": ACCOUNT_KEY,
-    #     }
+        # # Ensure you have the following environment variables set:
+        # # Gen 1
+        # storage_options = {
+        #     "tenant_id": AZURE_STORAGE_TENANT_ID,
+        #     "client_id": AZURE_STORAGE_CLIENT_ID,
+        #     "client_secret": AZURE_STORAGE_CLIENT_SECRET,
+        # }
+        # # Gen 2
+        # storage_options = {
+        #     "account_name": AZURE_STORAGE_ACCOUNT_NAME,
+        #     "account_key": AZURE_STORAGE_ACCOUNT_KEY,
+        # }
+
+        # Reads env
+        # https://github.com/fsspec/adlfs?tab=readme-ov-file#setting-credentials
+        storage_options = {"anon": False}
+        remote_file_system = adlfs.AzureBlobFileSystem(**storage_options)
+    elif config_dataset.path.startswith("oci://"):
+        try:
+            import ocifs
+        except ImportError as exc:
+            raise ImportError("oci:// paths require ocifs to be installed") from exc
+
+        # https://ocifs.readthedocs.io/en/latest/getting-connected.html#Using-Environment-Variables
+        remote_file_system = ocifs.OCIFileSystem(**storage_options)
 
-    #     remote_file_system = adlfs.AzureBlobFileSystem(**storage_options)
     try:
         if remote_file_system and remote_file_system.exists(config_dataset.path):
             ds_from_cloud = True

src/axolotl/utils/models.py

@@ -36,6 +36,7 @@ from transformers import (
     BitsAndBytesConfig,
     Gemma3ForConditionalGeneration,
     GPTQConfig,
+    Llama4ForConditionalGeneration,
     LlavaForConditionalGeneration,
     Mistral3ForConditionalGeneration,
     MllamaForConditionalGeneration,
@@ -76,6 +77,7 @@ LOG = logging.getLogger(__name__)
 
 MULTIMODAL_AUTO_MODEL_MAPPING = {
     "mllama": MllamaForConditionalGeneration,
+    "llama4": Llama4ForConditionalGeneration,
     "llava": LlavaForConditionalGeneration,
     "qwen2_vl": Qwen2VLForConditionalGeneration,
     "qwen2_5_vl": Qwen2_5_VLForConditionalGeneration,
@@ -557,6 +559,14 @@ class ModelLoader:
         plugin_manager = PluginManager.get_instance()
         plugin_manager.pre_model_load(self.cfg)
 
+        # monkey patch to allow additional Accelerator init kwargs
+        if self.cfg.fp8:
+            from axolotl.monkeypatch.trainer_accelerator_args import (
+                patch_create_accelerate_code_for_fp8,
+            )
+
+            patch_create_accelerate_code_for_fp8()
+
         if self.cfg.adapter:
             from axolotl.monkeypatch.transformers_fa_utils import (
                 patch_fa_peft_integration,
@@ -988,10 +998,11 @@ class ModelLoader:
             )
             skip_move_to_device = True
         elif (
-            self.model_config.model_type == "llama"
+            self.model_config.model_type in ["llama", "llama4"]
             and not self.cfg.trust_remote_code
             and not self.cfg.gptq
         ):
+            # TODO do we need to open this up for all models?
             if self.cfg.fsdp and self.cfg.fsdp_config.fsdp_cpu_ram_efficient_loading:
                 skip_move_to_device = True
                 if "device_map" in self.model_kwargs:

src/axolotl/utils/schemas/config.py

@@ -169,6 +169,7 @@ class AxolotlInputConfig(
 
     bf16: Literal["auto"] | bool | None = "auto"
     fp16: bool | None = None
+    fp8: bool | None = None
     bfloat16: bool | None = None  # for non-AMP cases
     float16: bool | None = None  # for non-AMP cases
     tf32: bool | None = None
@@ -464,9 +465,10 @@ class AxolotlInputConfig(
             data.get("sample_packing")
             and not data.get("flash_attention")
             and not data.get("sdp_attention")
+            and not data.get("flex_attention")
         ):
             LOG.warning(
-                "sample_packing without flash_attention or sdp_attention does not handle cross-attention."
+                "sample_packing without flash, sdp or flex attention does not handle cross sample decontamination."
             )
 
         return data
@@ -1154,6 +1156,12 @@ class AxolotlInputConfig(
                     "flash_attention: true must be set with sequence_parallel_degree > 1"
                 )
 
+            if not info.data["micro_batch_size"] == 1:
+                raise ValueError(
+                    "micro_batch_size must be set to 1 "
+                    "due to a `ring-flash-attn` requirement"
+                )
+
             try:
                 import ring_flash_attn  # noqa: F401 # pylint:disable=unused-import
             except ImportError as exception:
@@ -1163,6 +1171,18 @@ class AxolotlInputConfig(
                     "or `pip install ring-flash-attn>=0.1.4`."
                 ) from exception
 
+            # TODO: monkeypatch / callback to average losses correctly across SP ranks
+            # / fix gradient scaling across SP ranks. Losses, grads should be scaled
+            # according to the proportion of non-padding tokens per rank.
+            LOG.warning(
+                "Sequence parallelism (SP) is enabled with "
+                f"sequence_parallel_degree={value}. Please note that logged losses may "
+                "differ slightly to the non-SP losses due to transformers Trainer "
+                "implementation details. Please see "
+                "https://github.com/axolotl-ai-cloud/axolotl/pull/2495#issuecomment-2784022042 "
+                "for more details."
+            )
+
         return value
 
     @model_validator(mode="before")

src/axolotl/utils/schemas/datasets.py

@@ -39,7 +39,6 @@ class SFTDataset(BaseModel):
     input_format: str | None = None
     name: str | None = None
     ds_type: str | None = None
-    train_on_split: str | None = None
     field: str | None = None
     field_human: str | None = None
     field_model: str | None = None

src/axolotl/utils/schemas/enums.py

@@ -27,6 +27,7 @@ class ChatTemplate(str, Enum):
     cohere = "cohere"  # pylint: disable=invalid-name
     llama3 = "llama3"  # pylint: disable=invalid-name
     llama3_2_vision = "llama3_2_vision"  # pylint: disable=invalid-name
+    llama4 = "llama4"  # pylint: disable=invalid-name
     phi_3 = "phi_3"  # pylint: disable=invalid-name
     phi_35 = "phi_35"  # pylint: disable=invalid-name
     deepseek_v2 = "deepseek_v2"  # pylint: disable=invalid-name

src/axolotl/utils/trainer.py

@@ -582,7 +582,9 @@ def prepare_optim_env(cfg):
 
     setup_torch_compile_env(cfg)
 
-    if (cfg.bf16 == "auto" and is_torch_bf16_gpu_available()) or cfg.bf16 is True:
+    if cfg.fp8:
+        os.environ["ACCELERATE_MIXED_PRECISION"] = "fp8"
+    elif (cfg.bf16 == "auto" and is_torch_bf16_gpu_available()) or cfg.bf16 is True:
         os.environ["ACCELERATE_MIXED_PRECISION"] = "bf16"
     elif cfg.fp16:
         os.environ["ACCELERATE_MIXED_PRECISION"] = "fp16"

tests/e2e/multigpu/test_llama.py

@@ -7,9 +7,11 @@ import os
 from pathlib import Path
 
 import pytest
+import transformers
 import yaml
 from accelerate.test_utils import execute_subprocess_async
 from huggingface_hub import snapshot_download
+from packaging import version
 from transformers.testing_utils import get_torch_dist_unique_port
 
 from axolotl.utils.dict import DictDefault
@@ -28,6 +30,10 @@ def download_model():
     snapshot_download("HuggingFaceTB/SmolLM2-135M")
 
 
+def transformers_version_eq(required_version):
+    return version.parse(transformers.__version__) == version.parse(required_version)
+
+
 class TestMultiGPULlama:
     """
     Test case for Llama models using LoRA
@@ -56,7 +62,7 @@ class TestMultiGPULlama:
                 ],
                 "num_epochs": 1,
                 "max_steps": 2,
-                "micro_batch_size": 4,
+                "micro_batch_size": 1,
                 "gradient_accumulation_steps": 4,
                 # "gradient_checkpointing": True,
                 "output_dir": temp_dir,
@@ -108,7 +114,7 @@ class TestMultiGPULlama:
                 "lora_alpha": 16,
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
-                "val_set_size": 0.01,
+                "val_set_size": 0.05,
                 "special_tokens": {
                     "pad_token": "<|endoftext|>",
                 },
@@ -116,6 +122,7 @@ class TestMultiGPULlama:
                     {
                         "path": "tatsu-lab/alpaca",
                         "type": "alpaca",
+                        "split": "train[:20%]",
                     },
                 ],
                 "num_epochs": 1,
@@ -193,7 +200,7 @@ class TestMultiGPULlama:
                 ],
                 "num_epochs": 1,
                 "max_steps": 2,
-                "micro_batch_size": 4,
+                "micro_batch_size": 2,
                 "gradient_accumulation_steps": 4,
                 # "gradient_checkpointing": True,
                 "output_dir": temp_dir,
@@ -390,7 +397,7 @@ class TestMultiGPULlama:
                 "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "sample_packing": True,
                 "pad_to_sequence_len": True,
-                "sequence_len": 2048,
+                "sequence_len": 1024,
                 "val_set_size": 0.01,
                 "special_tokens": {
                     "pad_token": "<|endoftext|>",
@@ -403,7 +410,7 @@ class TestMultiGPULlama:
                 ],
                 "num_epochs": 1,
                 "max_steps": 2,
-                "micro_batch_size": 4,
+                "micro_batch_size": 2,
                 "gradient_accumulation_steps": 2,
                 # "gradient_checkpointing": True,
                 "output_dir": temp_dir,
@@ -493,9 +500,7 @@ class TestMultiGPULlama:
                 ],
                 "fsdp_config": {
                     "fsdp_version": 2,
-                    "fsdp_forward_prefetch": True,
-                    "fsdp_sync_module_states": True,
-                    "fsdp_use_orig_params": True,
+                    # "fsdp_forward_prefetch": True,  # not yet implemented in accelerate
                     "fsdp_offload_params": False,
                     "fsdp_cpu_ram_efficient_loading": False,
                     "fsdp_transformer_layer_cls_to_wrap": "LlamaDecoderLayer",
@@ -551,7 +556,7 @@ class TestMultiGPULlama:
                 "sample_packing": True,
                 "eval_sample_packing": False,
                 "pad_to_sequence_len": True,
-                "sequence_len": 2048,
+                "sequence_len": 1024,
                 "val_set_size": 0.01,
                 "special_tokens": {
                     "pad_token": "<|endoftext|>",
@@ -565,7 +570,7 @@ class TestMultiGPULlama:
                 ],
                 "num_epochs": 1,
                 "max_steps": 2,
-                "micro_batch_size": 4,
+                "micro_batch_size": 2,
                 "gradient_accumulation_steps": 2,
                 # "gradient_checkpointing": True,
                 "output_dir": temp_dir,
@@ -612,8 +617,11 @@ class TestMultiGPULlama:
             temp_dir + "/runs", "train/train_loss", 2.3, "Train Loss is too high"
         )
 
-    @pytest.mark.skip(
-        reason="ds-zero3 broken in main until transformers#37281 resolved"
+    # TODO: remove skip once deepspeed regression is fixed
+    # see https://github.com/huggingface/transformers/pull/37324
+    @pytest.mark.skipif(
+        transformers_version_eq("4.51.0"),
+        reason="zero3 is not supported with transformers==4.51.0",
     )
     @pytest.mark.parametrize(
         "gradient_accumulation_steps",
@@ -651,7 +659,7 @@ class TestMultiGPULlama:
                 "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "sample_packing": True,
                 "pad_to_sequence_len": True,
-                "sequence_len": 2048,
+                "sequence_len": 1024,
                 "val_set_size": 0.01,
                 "special_tokens": {
                     "pad_token": "<|endoftext|>",
@@ -724,7 +732,7 @@ class TestMultiGPULlama:
                 "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "sample_packing": True,
                 "pad_to_sequence_len": True,
-                "sequence_len": 2048,
+                "sequence_len": 1024,
                 "val_set_size": 0.01,
                 "special_tokens": {
                     "pad_token": "<|endoftext|>",
@@ -797,7 +805,7 @@ class TestMultiGPULlama:
                 "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "sample_packing": True,
                 "pad_to_sequence_len": True,
-                "sequence_len": 2048,
+                "sequence_len": 1024,
                 "val_set_size": 0.01,
                 "special_tokens": {
                     "pad_token": "<|endoftext|>",
@@ -885,7 +893,7 @@ class TestMultiGPULlama:
                 "sample_packing": True,
                 "bf16": True,
                 "save_safetensors": True,
-                "deepspeed": str(AXOLOTL_ROOT / "deepspeed_configs/zero1.json"),
+                # "deepspeed": str(AXOLOTL_ROOT / "deepspeed_configs/zero1.json"),
                 "use_tensorboard": True,
             }
         )

tests/e2e/multigpu/test_ray.py

@@ -31,7 +31,7 @@ class TestMultiGPURay:
         cfg = DictDefault(
             {
                 "base_model": "HuggingFaceTB/SmolLM2-135M",
-                "sequence_len": 2048,
+                "sequence_len": 1024,
                 "adapter": "lora",
                 "lora_r": 8,
                 "lora_alpha": 16,
@@ -94,8 +94,8 @@ class TestMultiGPURay:
                 "base_model": "HuggingFaceTB/SmolLM2-135M",
                 "sample_packing": True,
                 "pad_to_sequence_len": True,
-                "sequence_len": 2048,
-                "val_set_size": 0.05,
+                "sequence_len": 1024,
+                "val_set_size": 0.01,
                 "special_tokens": {
                     "pad_token": "<|endoftext|>",
                 },

tests/e2e/multigpu/test_sp.py

@@ -0,0 +1,87 @@
+"""E2E tests for sequence parallelism"""
+
+import os
+from pathlib import Path
+
+import yaml
+from accelerate.test_utils import execute_subprocess_async
+from transformers.testing_utils import get_torch_dist_unique_port
+
+from axolotl.utils.dict import DictDefault
+
+from ..utils import check_tensorboard
+
+os.environ["WANDB_DISABLED"] = "true"
+
+
+class TestSequenceParallelism:
+    """Test case for training with sequence parallelism enabled"""
+
+    def test_sequence_parallel_training(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "load_in_8bit": False,
+                "load_in_4bit": True,
+                "strict": False,
+                "sequence_len": 2048,
+                "adapter": "qlora",
+                "sample_packing": True,
+                "eval_sample_packing": True,
+                "pad_to_sequence_len": True,
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "lora_modules_to_save": ["embed_tokens", "lm_head"],
+                "special_tokens": {"pad_token": "<|endoftext|>"},
+                "datasets": [
+                    {
+                        "path": "tatsu-lab/alpaca",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 8,
+                "micro_batch_size": 1,
+                "gradient_accumulation_steps": 2,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_8bit",
+                "lr_scheduler": "cosine",
+                "flash_attention": True,
+                "loss_watchdog_threshold": 5.0,
+                "loss_watchdog_patience": 3,
+                "bf16": "auto",
+                "warmup_steps": 1,
+                "saves_per_epoch": 1,
+                "logging_steps": 1,
+                "weight_decay": 0.0,
+                "use_tensorboard": True,
+                "sequence_parallel_degree": 2,
+            }
+        )
+
+        # write cfg to yaml file
+        Path(temp_dir).mkdir(parents=True, exist_ok=True)
+        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
+            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
+
+        execute_subprocess_async(
+            [
+                "accelerate",
+                "launch",
+                "--num-processes",
+                "2",
+                "--main_process_port",
+                f"{get_torch_dist_unique_port()}",
+                "-m",
+                "axolotl.cli.train",
+                str(Path(temp_dir) / "config.yaml"),
+            ]
+        )
+
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.6, "Train Loss is too high"
+        )

tests/e2e/patched/test_sp.py

@@ -12,7 +12,6 @@ from axolotl.monkeypatch.attention.ring_attn import (
     get_ring_attn_group,
     set_ring_attn_group,
 )
-from axolotl.utils.collators.batching import adjust_position_ids_for_slice
 from axolotl.utils.dict import DictDefault
 
 
@@ -48,33 +47,6 @@ def fixture_cfg():
     return cfg
 
 
-class TestSequenceParallelHelpers:
-    """Test helper functions used in sequence parallelism."""
-
-    def test_adjust_position_ids_for_slice(self, partial_state):
-        """Test position_ids adjustment for sequence slices."""
-        # Create sample position_ids with multiple sequences
-        position_ids = torch.tensor(
-            [
-                # First sequence with 2 samples
-                [0, 1, 2, 3, 4, 0, 1, 2, 3],
-                # Second sequence with 3 samples
-                [0, 1, 2, 0, 1, 2, 3, 0, 1],
-            ]
-        )
-
-        # Adjust as if this was the second slice (start_idx = 4)
-        adjusted = adjust_position_ids_for_slice(position_ids, start_idx=4)
-
-        # For first sequence: [0,1,2,3,4,0,1,2,3] -> [-4,-3,-2,-1,0,-4,-3,-2,-1]
-        # For second sequence: [0,1,2,0,1,2,3,0,1] -> [-4,-3,-2,-4,-3,-2,-1,-4,-3]
-        expected_first_seq = torch.tensor([0, 1, 2, 3, 4, 0, 1, 2, 3]) - 4
-        expected_second_seq = torch.tensor([0, 1, 2, 0, 1, 2, 3, 0, 1]) - 4
-
-        assert torch.all(adjusted[0] == expected_first_seq)
-        assert torch.all(adjusted[1] == expected_second_seq)
-
-
 class TestRingAttention:
     """Tests for the ring attention functionality."""