set v0.9.2 version for tag

* offload activations to disk instead of CPU RAM

* add prefetch

* Disco :dance:

* include offload_disk in e2e test for AC

* document and make sure to cleanup

* fix annotation to match docs

* fix docs build

* address PR feedback

.github/workflows/tests.yml

@@ -347,12 +347,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.6.0
-            num_gpus: 1
-            axolotl_extras: llmcompressor
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"

_quarto.yml

@@ -48,23 +48,8 @@ quartodoc:
       contents:
         - core.trainers.base
         - core.trainers.trl
-        - core.trainers.mamba
-        - core.trainers.relora
         - core.trainers.dpo.trainer
         - core.trainers.grpo.trainer
-        - core.trainers.grpo.sampler
-        - core.trainers.utils
-    - title: Mixins
-      desc: Mixin classes for augmenting trainers
-      contents:
-        - core.trainers.mixins.optimizer
-        - core.trainers.mixins.rng_state_loader
-        - core.trainers.mixins.scheduler
-        - core.trainers.mixins.sequence_parallel
-    - title: Context Managers
-      desc: Context managers for altering trainer behaviors
-      contents:
-        - utils.ctx_managers.sequence_parallel
     - title: Prompt Strategies
       desc: Prompt formatting strategies
       contents:
@@ -101,7 +86,7 @@ quartodoc:
         - kernels.swiglu
         - kernels.quantize
         - kernels.utils
-    - title: Monkey Patches
+    - title: MonkeyPatches
       desc: Runtime patches for model optimizations
       contents:
         - monkeypatch.llama_attn_hijack_flash
@@ -139,7 +124,8 @@ quartodoc:
         - utils.optimizers.adopt
         - utils.data.pretraining
         - utils.data.sft
-        - utils.gradient_checkpointing.unsloth
+        - utils.gradient_checkpointing.offload_cpu
+        - utils.gradient_checkpointing.offload_disk
     - title: Schemas
       desc: Pydantic data models for Axolotl config
       contents:

docs/config.qmd

@@ -539,7 +539,7 @@ train_on_inputs: false
 # Note that training loss may have an oscillating pattern with this enabled.
 group_by_length: false
 
-# Whether to use gradient checkpointing. Available options are: true, false, "offload".
+# Whether to use gradient checkpointing. Available options are: true, false, "offload", "offload_disk".
 # https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
 gradient_checkpointing: false
 # additional kwargs to pass to the trainer for gradient checkpointing

docs/custom_integrations.qmd

@@ -49,8 +49,7 @@ sections = [
     ("Knowledge Distillation (KD)", "kd"),
     ("Liger Kernels", "liger"),
     ("Language Model Evaluation Harness (LM Eval)", "lm_eval"),
-    ("Spectrum", "spectrum"),
+    ("Spectrum", "spectrum")
-    ("LLMCompressor", "llm_compressor")
 ]
 
 for section_name, folder_name in sections:

docs/sequence_parallelism.qmd

@@ -3,6 +3,8 @@ title: Sequence Parallelism
 description: Train with long sequences split across multiple GPUs.
 ---
 
+# Sequence Parallelism
+
 Sequence parallelism is a technique that splits sequences across multiple GPUs,
 allowing you to train with very long sequences that wouldn't fit on a single GPU. Each
 GPU processes a different portion of the sequence, and the results are aggregated
@@ -25,7 +27,7 @@ To enable sequence parallelism, add the following to your configuration file:
 sequence_parallel_degree: 4  # Split sequences across 4 GPUs
 # Optional; strides across the key dimension. Larger values use more memory but should make training faster.
 heads_k_stride: 1
-# Optional; one of "varlen_llama3" or "batch_ring". Defaults to
+# Optional; one of "varlen_llama3", "batch_ring", "batch_zigzag", "batch_stripe". Defaults to
 # "varlen_llama3" when `sample_packing: true`, and "batch_ring" otherwise.
 ring_attn_func:
 ```

examples/llama-3/sparse-finetuning.yaml

@@ -1,77 +0,0 @@
-base_model: neuralmagic/Sparse-Llama-3.1-8B-2of4
-
-plugins:
-  - axolotl.integrations.llm_compressor.LLMCompressorPlugin
-
-load_in_8bit: false
-load_in_4bit: false
-strict: false
-
-datasets:
-  - path: tatsu-lab/alpaca
-    type: alpaca
-dataset_prepared_path: last_run_prepared
-val_set_size: 0.05
-output_dir: ./outputs/out
-
-sequence_len: 4096
-sample_packing: true
-pad_to_sequence_len: true
-eval_sample_packing: false
-
-wandb_project:
-wandb_entity:
-wandb_watch:
-wandb_name:
-wandb_log_model:
-
-gradient_accumulation_steps: 8
-micro_batch_size: 1
-num_epochs: 1
-optimizer: paged_adamw_8bit
-lr_scheduler: cosine
-learning_rate: 2e-5
-
-train_on_inputs: false
-group_by_length: false
-bf16: auto
-fp16:
-tf32: false
-
-gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
-early_stopping_patience:
-resume_from_checkpoint:
-logging_steps: 1
-xformers_attention:
-flash_attention: true
-
-warmup_steps: 100
-evals_per_epoch: 2
-eval_table_size:
-saves_per_epoch: 1
-debug:
-deepspeed:
-weight_decay: 0.0
-fsdp:
-fsdp_config:
-special_tokens:
-  pad_token: <|end_of_text|>
-
-llmcompressor:
-  recipe:
-    finetuning_stage:
-      finetuning_modifiers:
-        ConstantPruningModifier:
-          targets: [
-            're:.*q_proj.weight',
-            're:.*k_proj.weight',
-            're:.*v_proj.weight',
-            're:.*o_proj.weight',
-            're:.*gate_proj.weight',
-            're:.*up_proj.weight',
-            're:.*down_proj.weight',
-          ]
-          start: 0
-  save_compressed: true

setup.py

@@ -150,9 +150,6 @@ extras_require = {
     "vllm": [
         "vllm==0.7.2",
     ],
-    "llmcompressor": [
-        "llmcompressor==0.5.1",
-    ],
 }
 
 install_requires, dependency_links, extras_require_build = parse_requirements(

src/axolotl/__init__.py

@@ -4,4 +4,4 @@ import pkgutil
 
 __path__ = pkgutil.extend_path(__path__, __name__)  # Make this a namespace package
 
-__version__ = "0.10.0.dev0"
+__version__ = "0.9.2"

src/axolotl/cli/main.py

@@ -342,13 +342,6 @@ def delinearize_llama4(model: str, output: str) -> None:
     do_delinearize_llama4(model, output)
 
 
-@cli.command()
-def wizard():
-    from axolotl.cli.wizard import do_wizard
-
-    do_wizard()
-
-
 cli.add_command(lm_eval)
 
 

src/axolotl/cli/wizard.py

@@ -1,429 +0,0 @@
-"""Wizard for creating yaml configs."""
-
-import click
-import torch
-import yaml
-from packaging import version
-from transformers.training_args import OptimizerNames
-
-from axolotl.cli.art import print_axolotl_text_art
-from axolotl.utils.dict import DictDefault
-from axolotl.utils.models import load_model_config
-from axolotl.utils.schemas.enums import CustomSupportedOptimizers
-
-
-def do_wizard():
-    print_axolotl_text_art()
-
-    # Ask where to save the config
-    cfg = DictDefault({})
-    config_path = click.prompt(
-        "Where do you want to save the config?", type=str, default="config.yaml"
-    )
-
-    # Ask base model
-    base_model = click.prompt("What base model do you want to use?", type=str)
-    cfg["base_model"] = base_model.strip()
-
-    # Ask whether want to enable Vision model
-    # TODO: check if model has vision layers instead of asking user
-    train_vision_model = click.confirm(
-        "If this model has vision layers, do you want to train them?", default=False
-    )
-
-    if train_vision_model:
-        cfg["processor_type"] = "AutoProcessor"
-        cfg["skip_prepare_dataset"] = True
-        cfg["remove_unused_columns"] = False
-        cfg["sample_packing"] = False
-
-    # Ask whether they want to set any advanced model features (custom tokenizer, custom config, etc)
-    advanced_model_features = click.confirm(
-        "Do you want to set any advanced model features? (custom tokenizer, custom config, remote code etc)",
-        default=False,
-    )
-
-    if advanced_model_features:
-        # Ask whether they want to use a custom config
-        base_model_config = click.prompt(
-            "What model config do you want to use? (leave blank for default)",
-            type=str,
-            default="",
-        )
-
-        if base_model_config:
-            cfg["base_model_config"] = base_model_config
-
-        # Ask whether they want to use a specific revision of the model
-        revision_of_model = click.prompt(
-            "What revision of the model do you want to use? (leave blank for default)",
-            type=str,
-            default="",
-        )
-
-        if revision_of_model:
-            cfg["revision_of_model"] = revision_of_model
-
-        # Ask whether they want to use a custom tokenizer
-        tokenizer_config = click.prompt(
-            "What tokenizer do you want to use? (leave blank for default)",
-            type=str,
-            default="",
-        )
-
-        if tokenizer_config:
-            cfg["tokenizer_config"] = tokenizer_config
-
-        # Ask whether they want to use remote code
-        trust_remote_code = click.confirm(
-            "Do you want to use remote code?", default=False
-        )
-
-        if trust_remote_code:
-            cfg["trust_remote_code"] = trust_remote_code
-
-        # Whether to resize token embeddings
-        resize_token_embeddings_to_32x = click.confirm(
-            "Do you want to resize token embeddings to 32x?", default=False
-        )
-
-        if resize_token_embeddings_to_32x:
-            cfg["resize_token_embeddings_to_32x"] = resize_token_embeddings_to_32x
-
-        # Whether to shrink embeddings to len(tokenizer)
-        shrink_embeddings = click.confirm(
-            "Do you want to shrink embeddings to len(tokenizer)?", default=False
-        )
-
-        if shrink_embeddings:
-            cfg["shrink_embeddings"] = shrink_embeddings
-
-        # Whether to skip upcast embeddings
-        embeddings_skip_upcast = click.confirm(
-            "Do you want to skip upcast embeddings?", default=False
-        )
-
-        if embeddings_skip_upcast:
-            cfg["embeddings_skip_upcast"] = embeddings_skip_upcast
-
-        # Whether to random init weights
-        random_init_weights = click.confirm(
-            "Do you want to random init weights?", default=False
-        )
-
-        if random_init_weights:
-            cfg["random_init_weights"] = random_init_weights
-
-    # Get model type
-    config = load_model_config(cfg)
-    model_type = config.model_type
-
-    # Ask sequence length
-    sequence_length = click.prompt("What sequence length do you want to use?", type=int)
-    cfg["sequence_length"] = sequence_length
-
-    # Whether to turn on sample packing
-    if cfg["sample_packing"] is None:
-        cfg["sample_packing"] = click.confirm(
-            "Do you want to turn on sample packing? This will speed up training by packing multiple samples into a single batch.",
-            default=True,
-        )
-
-        if cfg["sample_packing"]:
-            cfg["pad_to_sequence_len"] = True
-
-        # Whether to turn off eval sample packing
-        no_eval_sample_packing = click.confirm(
-            "Do you want to turn off eval sample packing? This will slow down evaluation but is recommended if you are using a small validation set.",
-            default=False,
-        )
-
-        if no_eval_sample_packing:
-            cfg["eval_sample_packing"] = False
-
-    # Hardware check
-    try:
-        is_ampere_or_newer = torch.cuda.get_device_capability()[0] >= 8
-    except RuntimeError:
-        is_ampere_or_newer = False
-    except AssertionError:  # this is raised if no cuda is available
-        is_ampere_or_newer = False
-
-    # Get num gpus
-    try:
-        num_gpus = torch.cuda.device_count()
-    except RuntimeError:
-        num_gpus = 0
-
-    # Get torch version
-    torch_version = str(torch.__version__).split("+", maxsplit=1)[0]
-
-    is_torch_2_6_or_newer = version.parse(torch_version) >= version.parse("2.6.0")
-
-    # Whether to turn on attention
-    opt = ["xformers", "sdp"]
-
-    if is_ampere_or_newer:
-        opt.append("flash")
-
-    if is_torch_2_6_or_newer:
-        opt.append("flex")
-
-    if cfg["sample_packing"]:
-        if "flash" in opt:
-            default_opt = "flash"
-        elif "flex" in opt:
-            default_opt = "flex"
-        else:
-            default_opt = opt[0]
-
-        attention = click.prompt(
-            "Which attention backend do you want to use? Sample packing requires an attention backend to be set.",
-            type=click.Choice(opt),
-            default=default_opt,
-        )
-    else:
-        # non-sample packing supports no attention and S2
-        opt.extend(["none", "s2"])
-
-        attention = click.prompt(
-            "Which attention backend do you want to use?",
-            type=click.Choice(opt),
-            default="none",
-        )
-
-        if attention == "none":
-            attention = None
-
-    # TODO: if xformers, check if FA is installed
-    # TODO: flex doc mentioned requiring seq len to be divisible by 128. Unclear if limitation still exists
-
-    # TODO: requires #2489
-    cfg["attention"] = attention
-
-    # Whether to turn on gradient checkpointing
-    # TODO: need to wait for offload_disk PR to be merged
-    gradient_checkpointing = click.prompt(
-        "Which gradient checkpointing strategy do you want to use?",
-        type=click.Choice(["none", "true", "offload", "offload_disk"]),
-        default="true",
-    )
-
-    if gradient_checkpointing == "none":
-        gradient_checkpointing = False
-    elif gradient_checkpointing == "true":
-        gradient_checkpointing = True
-
-    # Ask whether to set use_reentrant
-    # TODO: get correct defaults based on SFT/RL mode and single/multigpu
-    # use_reentrant = click.confirm(
-    #     "Do you want to set use_reentrant?",
-    #     default=True,
-    # )
-
-    # if use_reentrant:
-    #     cfg["use_reentrant"] = use_reentrant
-
-    # Optimizer
-    cfg["optimizer"] = click.prompt(
-        "Which optimizer do you want to use?",
-        type=click.Choice((OptimizerNames | CustomSupportedOptimizers)),
-        default=OptimizerNames.ADAMW_TORCH_FUSED,
-    )
-
-    cfg["lr_scheduler"] = click.prompt(
-        "Which learning rate scheduler do you want to use?",
-        type=click.Choice(
-            [
-                "cosine",
-                "one_cycle",
-                "rex",
-                "log_sweep",
-                "linear",
-                "cosine_with_restarts",
-                "polynomial",
-                "constant",
-                "constant_with_warmup",
-                "inverse_sqrt",
-                "reduce_lr_on_plateau",
-                "cosine_with_min_lr",
-                "warmup_stable_decay",
-            ]
-        ),
-        default="cosine",
-    )
-
-    # Plugins
-
-    cfg["plugins"] = []
-
-    # Whether to turn on cut cross entropy
-    if is_ampere_or_newer:
-        # Note: This may error if users don't have CCE installed
-        from axolotl.integrations.cut_cross_entropy.monkeypatch.patch import (
-            CUT_CROSS_ENTROPY_MODEL_MAPPING,
-        )
-
-        if model_type in CUT_CROSS_ENTROPY_MODEL_MAPPING:
-            cut_cross_entropy = click.confirm(
-                "Do you want to turn on cut cross entropy? This will save VRAM if the model has a large vocab size.",
-                default=True,
-            )
-
-            if cut_cross_entropy:
-                cfg["plugins"].append(
-                    "axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin"
-                )
-
-                cfg["cut_cross_entropy"] = True
-
-    use_liger_kernel = click.confirm(
-        "Do you want to use the liger kernel? This will speed up training and save VRAM.",
-        default=True,
-    )
-
-    if use_liger_kernel:
-        cfg["plugins"].append("axolotl.integrations.liger.LigerPlugin")
-
-        cfg["liger_rope"] = click.confirm(
-            "Do you want to enable liger rope?",
-            default=True,
-        )
-
-        cfg["liger_rms_norm"] = click.confirm(
-            "Do you want to enable liger rms norm?",
-            default=True,
-        )
-
-        cfg["liger_glu_activation"] = click.confirm(
-            "Do you want to enable liger glu activation?",
-            default=True,
-        )
-
-        cfg["liger_layer_norm"] = click.confirm(
-            "Do you want to enable liger layer norm?",
-            default=True,
-        )
-
-        if cfg["cut_cross_entropy"] is not True:
-            cfg["liger_fused_linear_cross_entropy"] = click.confirm(
-                "Do you want to enable liger fused linear cross entropy?",
-                default=True,
-            )
-
-    # TODO: lora kernels (but they auto enable via validator already)
-
-    # TODO: is there incompat between torch compile and liger?
-    cfg["torch_compile"] = click.confirm(
-        "Do you want to enable torch compile?",
-        default=True,
-    )
-
-    # Multi-gpu
-    if num_gpus > 1:
-        # Ask whether to use DDP/Deepspeed/FSDP
-        multi_gpu_mode = click.prompt(
-            "Which multi-gpu mode do you want to use?",
-            type=click.Choice(["ddp", "deepspeed", "fsdp"]),
-            default="ddp",
-        )
-
-        if multi_gpu_mode == "deepspeed":
-            # Ask which deepspeed config to use
-            cfg["deepspeed"] = click.prompt(
-                "Which deepspeed config do you want to use? The higher the number, the more VRAM you will save, but the slower it will run.",
-                type=click.Choice(
-                    [
-                        "zero1.json",
-                        "zero1_torch_compile.json",
-                        "zero2.json",
-                        "zero3.json",
-                        "zero3_bf16.json",
-                        "zero3_bf16_cpuoffload_all.json",
-                        "zero3_bf16_cpuoffload_params.json",
-                    ]
-                ),
-                default="zero1.json",
-            )
-        elif multi_gpu_mode == "fsdp":
-            fsdp_version = click.prompt(
-                "Which fsdp version do you want to use?",
-                type=click.Choice([1, 2]),
-                default=2,
-            )
-
-            # TODO: Handle FSDP config
-
-            if fsdp_version == 1:
-                cfg["fsdp"] = ["full_shard", "auto_wrap"]
-
-                # Ask which state dict type to use
-                fsdp_state_dict_type = click.prompt(
-                    "Which fsdp state dict type do you want to use?",
-                    type=click.Choice(["FULL_STATE_DICT", "SHARDED_STATE_DICT"]),
-                    default="FULL_STATE_DICT",
-                )
-
-                fsdp_offload_params = click.confirm(
-                    "Do you want to offload parameters?",
-                    default=True,
-                )
-
-                # TODO: can we load the model class and auto pull a default for this?
-                fsdp_transformer_layer_cls_to_wrap = click.prompt(
-                    "Which transformer layer class to wrap? It is usually the Decoder layer class.",
-                    type=str,
-                )
-
-                # TODO: add other options
-
-                cfg["fsdp_config"] = {
-                    "state_dict_type": fsdp_state_dict_type,
-                    "offload_params": fsdp_offload_params,
-                    "transformer_layer_cls_to_wrap": fsdp_transformer_layer_cls_to_wrap,
-                }
-
-            elif fsdp_version == 2:
-                raise NotImplementedError()
-
-    # Training mode (sft or rl)
-    training_mode = click.prompt(
-        "Which training mode do you want to use?",
-        type=click.Choice(["sft", "rl"]),
-        default="sft",
-    )
-
-    if training_mode == "rl":
-        cfg["rl"] = click.prompt(
-            "Which rl mode do you want to use?",
-            type=click.Choice(["dpo", "ipo", "orpo", "kto", "grpo", "simpo"]),
-        )
-
-        # TODO: handle RL options
-
-    # Whether to use adapter
-
-    # Get batch/grad accu
-
-    # Get learning rate
-
-    # Get weight decay
-
-    # Get max grad norm
-
-    # Get num train epochs
-
-    # Get warmup ratio
-
-    # Get save ratio
-
-    # Get eval ratio
-
-    # Get dataset config
-
-    # Load metric tracker
-
-    # Save config to yaml
-    # TODO: improve output yaml formatting. Need to add comments to help separate sections
-    with open(config_path, "w", encoding="utf-8") as f:
-        yaml.dump(cfg.to_dict(), f, sort_keys=False)

src/axolotl/common/datasets.py

@@ -14,7 +14,6 @@ from axolotl.utils.data import prepare_dataset
 from axolotl.utils.data.rl import load_prepare_preference_datasets
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.models import load_processor, load_tokenizer
-from axolotl.utils.schemas.enums import RLType
 from axolotl.utils.tokenization import check_dataset_labels
 
 LOG = logging.getLogger(__name__)
@@ -134,7 +133,7 @@ def load_preference_datasets(
     total_num_steps: Optional[int] = int(
         math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
     )
-    if cfg.rl is RLType.GRPO:
+    if cfg.rl == "grpo":
         total_num_steps = None
 
     if cli_args.debug or cfg.debug:

src/axolotl/core/trainer_builder.py

@@ -87,7 +87,7 @@ from axolotl.utils.collators import (
 )
 from axolotl.utils.collators.mm_chat import MultiModalChatDataCollator
 from axolotl.utils.models import ensure_dtype
-from axolotl.utils.schemas.enums import CustomSupportedOptimizers, RLType
+from axolotl.utils.schemas.enums import CustomSupportedOptimizers
 
 try:
     import torch._dynamo  # pylint: disable=ungrouped-imports
@@ -353,7 +353,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["warmup_steps"] = warmup_steps
         training_arguments_kwargs["logging_steps"] = logging_steps
 
-        if self.cfg.seed is not None:
+        if self.cfg.seed:
             training_arguments_kwargs["seed"] = self.cfg.seed
 
         if self.cfg.gradient_checkpointing:
@@ -547,6 +547,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         report_to = []
         if self.cfg.use_wandb:
             report_to.append("wandb")
+            if self.cfg.wandb_name:
+                training_arguments_kwargs["run_name"] = self.cfg.wandb_name
         if self.cfg.use_mlflow:
             report_to.append("mlflow")
         if self.cfg.use_tensorboard:
@@ -819,15 +821,14 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         data_collator_kwargs = {
             "padding": True,  # True/"longest" is the default
         }
-        multiple = 64
         if self.cfg.pad_to_sequence_len:
-            data_collator_kwargs["pad_to_multiple_of"] = multiple * math.ceil(
+            data_collator_kwargs["pad_to_multiple_of"] = 64 * math.ceil(
-                self.cfg.sequence_len / multiple
+                self.cfg.sequence_len / 64
             )
         else:
             # A100 is best at 64, while others at 8. Let's use the larger so we don't have to check
             # https://docs.nvidia.com/deeplearning/performance/dl-performance-matrix-multiplication/index.html
-            data_collator_kwargs["pad_to_multiple_of"] = multiple
+            data_collator_kwargs["pad_to_multiple_of"] = 64
 
         if self.cfg.reward_model:
             data_collator_kwargs["max_length"] = self.cfg.sequence_len
@@ -1033,10 +1034,6 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             training_args_kwargs["dataloader_prefetch_factor"] = (
                 self.cfg.dataloader_prefetch_factor
             )
-
-        if self.cfg.seed is not None:
-            training_args_kwargs["seed"] = self.cfg.seed
-
         if self.cfg.gradient_checkpointing:
             training_args_kwargs["gradient_checkpointing"] = (
                 self.cfg.gradient_checkpointing
@@ -1079,13 +1076,9 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         if self.cfg.use_wandb:
             training_args_kwargs["run_name"] = self.cfg.wandb_name
 
-        training_args_kwargs["sequence_parallel_degree"] = (
-            self.cfg.sequence_parallel_degree
-        )
-
         training_args_cls = None
         blocklist_args_kwargs = []
-        if self.cfg.rl is RLType.SIMPO:
+        if self.cfg.rl == "simpo":
             training_args_cls = AxolotlCPOConfig
             training_args_kwargs["loss_type"] = "simpo"
             training_args_kwargs["max_length"] = self.cfg.sequence_len
@@ -1093,13 +1086,13 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             if self.cfg.cpo_alpha is not None:
                 training_args_kwargs["cpo_alpha"] = self.cfg.cpo_alpha
 
-        elif self.cfg.rl is RLType.ORPO:
+        elif self.cfg.rl == "orpo":
             training_args_cls = AxolotlORPOConfig
             training_args_kwargs["max_length"] = self.cfg.sequence_len
             if self.cfg.max_prompt_len:
                 training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
 
-        elif self.cfg.rl is RLType.KTO:
+        elif self.cfg.rl == "kto":
             training_args_cls = AxolotlKTOConfig
 
             training_args_kwargs["desirable_weight"] = (
@@ -1113,14 +1106,14 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             if self.cfg.max_prompt_len:
                 training_args_kwargs["max_prompt_length"] = self.cfg.max_prompt_len
 
-        elif self.cfg.rl is RLType.GRPO:
+        elif self.cfg.rl == "grpo":
             training_args_cls = GRPOStrategy.get_training_args_class()
             training_args_kwargs.update(GRPOStrategy.set_training_args_kwargs(self.cfg))
             blocklist_args_kwargs = GRPOStrategy.get_blocklist_args_kwargs()
 
         else:
             training_args_cls = AxolotlDPOConfig
-            if self.cfg.rl is RLType.IPO:
+            if self.cfg.rl == "ipo":
                 training_args_kwargs["loss_type"] = "ipo"
             training_args_kwargs["max_length"] = self.cfg.sequence_len
             training_args_kwargs["max_completion_length"] = None
@@ -1163,35 +1156,33 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
 
     def build(self, total_num_steps):
         training_args = self.build_training_arguments(total_num_steps)
-        trainer_kwargs = {}
+        dpo_trainer_kwargs = {}
-        if self.cfg.rl is RLType.IPO:
+        if self.cfg.rl == "ipo":
             if self.cfg.dpo_label_smoothing:
-                trainer_kwargs["label_smoothing"] = self.cfg.dpo_label_smoothing
+                dpo_trainer_kwargs["label_smoothing"] = self.cfg.dpo_label_smoothing
         if self.eval_dataset:
-            trainer_kwargs["eval_dataset"] = self.eval_dataset
+            dpo_trainer_kwargs["eval_dataset"] = self.eval_dataset
         if self.cfg.adapter and self.peft_config:
-            trainer_kwargs["peft_config"] = self.peft_config
+            dpo_trainer_kwargs["peft_config"] = self.peft_config
         if self.cfg.precompute_ref_log_probs is not None:
-            trainer_kwargs["precompute_ref_log_probs"] = (
+            dpo_trainer_kwargs["precompute_ref_log_probs"] = (
                 self.cfg.precompute_ref_log_probs
             )
-        if self.cfg.rl is RLType.GRPO:
+        if self.cfg.rl == "grpo":
-            trainer_cls = GRPOStrategy.get_trainer_class(
+            trainer_cls = GRPOStrategy.get_trainer_class()
-                sequence_parallel=self.cfg.sequence_parallel_degree > 1
-            )
             trainer_cls_args = [self.model]
             trainer_cls_args.extend(GRPOStrategy.set_trainer_args(self.cfg))
-            trainer_kwargs.update(GRPOStrategy.set_trainer_kwargs(self.cfg))
+            dpo_trainer_kwargs.update(GRPOStrategy.set_trainer_kwargs(self.cfg))
-        elif self.cfg.rl in [RLType.DPO, RLType.IPO]:
+        elif self.cfg.rl in ["dpo", "ipo"]:
             trainer_cls = DPOStrategy.get_trainer_class()
             trainer_cls_args = [self.model, self.model_ref]
-        elif self.cfg.rl is RLType.ORPO:
+        elif self.cfg.rl == "orpo":
             trainer_cls = AxolotlORPOTrainer
             trainer_cls_args = [self.model]
-        elif self.cfg.rl is RLType.KTO:
+        elif self.cfg.rl in ["kto"]:
             trainer_cls = AxolotlKTOTrainer
             trainer_cls_args = [self.model]
-        elif self.cfg.rl is RLType.SIMPO:
+        elif self.cfg.rl in ["simpo"]:
             trainer_cls = AxolotlCPOTrainer
             trainer_cls_args = [self.model]
         else:
@@ -1203,33 +1194,33 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
 
         sig = inspect.signature(trainer_cls)
         if "tokenizer" in sig.parameters.keys():
-            trainer_kwargs["tokenizer"] = self.tokenizer
+            dpo_trainer_kwargs["tokenizer"] = self.tokenizer
         else:
-            trainer_kwargs["processing_class"] = self.tokenizer
+            dpo_trainer_kwargs["processing_class"] = self.tokenizer
 
         if self.cfg.datasets is not None and (
             trainer_cls is DPOStrategy.get_trainer_class()
         ):
-            trainer_kwargs["dataset_tags"] = [
+            dpo_trainer_kwargs["dataset_tags"] = [
                 d["path"] for d in self.cfg.datasets if not Path(d["path"]).is_dir()
             ]
-        trainer = trainer_cls(
+        dpo_trainer = trainer_cls(
             *trainer_cls_args,
             args=training_args,
             train_dataset=self.train_dataset,
             callbacks=self.get_callbacks(),
-            **trainer_kwargs,
+            **dpo_trainer_kwargs,
         )
         if self.cfg.fsdp:
-            ensure_dtype(trainer.model, dtype=self.cfg.torch_dtype)
+            ensure_dtype(dpo_trainer.model, dtype=self.cfg.torch_dtype)
-            if self.cfg.rl in [RLType.DPO, RLType.IPO] and trainer.ref_model:
+            if self.cfg.rl in ["dpo", "ipo"] and dpo_trainer.ref_model:
-                ensure_dtype(trainer.ref_model, dtype=self.cfg.torch_dtype)
+                ensure_dtype(dpo_trainer.ref_model, dtype=self.cfg.torch_dtype)
 
-        trainer = self.hook_post_create_trainer(trainer)
+        dpo_trainer = self.hook_post_create_trainer(dpo_trainer)
-        for callback in self.get_post_trainer_create_callbacks(trainer):
+        for callback in self.get_post_trainer_create_callbacks(dpo_trainer):
-            trainer.add_callback(callback)
+            dpo_trainer.add_callback(callback)
 
-        return trainer
+        return dpo_trainer
 
 
 class HFPPOTrainerBuilder(TrainerBuilderBase):

src/axolotl/core/trainers/__init__.py

@@ -5,7 +5,7 @@
 
 from .base import AxolotlTrainer
 from .dpo.trainer import AxolotlDPOTrainer
-from .grpo.trainer import AxolotlGRPOSequenceParallelTrainer, AxolotlGRPOTrainer
+from .grpo.trainer import AxolotlGRPOTrainer
 from .mamba import AxolotlMambaTrainer
 from .relora import ReLoRATrainer
 from .trl import (

src/axolotl/core/trainers/base.py

@@ -373,13 +373,15 @@ class AxolotlTrainer(
                 num_items_in_batch=num_items_in_batch,
             )
 
-        return super().compute_loss(
+        loss = super().compute_loss(
             model,
             inputs,
             return_outputs=return_outputs,
             num_items_in_batch=num_items_in_batch,
         )
 
+        return loss
+
     @staticmethod
     def orpo_concatenate_inputs(inputs, label_pad_token=-100, pad_token=0, device=None):
         concatenated_batch = {}

src/axolotl/core/trainers/dpo/__init__.py

@@ -1,11 +1,14 @@
-"""DPO Specific Strategy for training"""
+"""
+DPO Specific Strategy for training
+"""
 
 from axolotl.core.trainers.dpo.trainer import AxolotlDPOTrainer
-from axolotl.utils.schemas.enums import RLType
 
 
 class DPOStrategy:
-    """Strategy for DPO training"""
+    """
+    Strategy for DPO training
+    """
 
     @classmethod
     def get_trainer_class(cls):
@@ -20,7 +23,7 @@ class DPOStrategy:
     @classmethod
     def set_training_args_kwargs(cls, cfg):
         training_args_kwargs = {}
-        if cfg.rl is RLType.IPO:
+        if cfg.rl == "ipo":
             training_args_kwargs["loss_type"] = "ipo"
         training_args_kwargs["max_length"] = cfg.sequence_len
         training_args_kwargs["max_completion_length"] = None

src/axolotl/core/trainers/grpo/__init__.py

@@ -1,41 +1,37 @@
-"""GRPO Specific Strategy for training"""
+"""
+GRPO Specific Strategy for training
+"""
 
 import importlib
 import inspect
 import logging
-from typing import Any
 
 from trl.trainer.grpo_trainer import RewardFunc
 
-from axolotl.core.trainers.grpo.args import AxolotlGRPOConfig
+from axolotl.core.trainers.grpo.trainer import AxolotlGRPOTrainer
-from axolotl.core.trainers.grpo.trainer import (
-    AxolotlGRPOSequenceParallelTrainer,
-    AxolotlGRPOTrainer,
-)
-from axolotl.utils.dict import DictDefault
 from axolotl.utils.schemas.trl import TRLConfig
 
-LOG = logging.getLogger(__name__)
+LOG = logging.getLogger("axolotl")
 
 
 class GRPOStrategy:
-    """Strategy for GRPO training"""
+    """
+    Strategy for GRPO training
+    """
 
     @classmethod
-    def get_trainer_class(
+    def get_trainer_class(cls):
-        cls, sequence_parallel: bool
-    ) -> type[AxolotlGRPOTrainer] | type[AxolotlGRPOSequenceParallelTrainer]:
-        if sequence_parallel:
-            return AxolotlGRPOSequenceParallelTrainer
         return AxolotlGRPOTrainer
 
     @classmethod
-    def get_training_args_class(cls) -> type[AxolotlGRPOConfig]:
+    def get_training_args_class(cls):
+        from axolotl.core.trainers.grpo.args import AxolotlGRPOConfig
+
         return AxolotlGRPOConfig
 
     @classmethod
-    def set_training_args_kwargs(cls, cfg: DictDefault) -> dict[str, Any]:
+    def set_training_args_kwargs(cls, cfg):
-        grpo_args_kwargs: dict[str, Any] = {}
+        grpo_args_kwargs = {}
 
         if not hasattr(cfg, "trl") or not cfg.trl:
             return grpo_args_kwargs
@@ -44,8 +40,8 @@ class GRPOStrategy:
 
         if trl.use_vllm:
             grpo_args_kwargs["use_vllm"] = trl.use_vllm
-            grpo_args_kwargs["vllm_server_host"] = trl.vllm_server_host or trl.vllm.host  # type: ignore[attr-defined]
+            grpo_args_kwargs["vllm_server_host"] = trl.vllm_server_host or trl.vllm.host
-            grpo_args_kwargs["vllm_server_port"] = trl.vllm_server_port or trl.vllm.port  # type: ignore[attr-defined]
+            grpo_args_kwargs["vllm_server_port"] = trl.vllm_server_port or trl.vllm.port
             if trl.vllm_server_timeout:
                 grpo_args_kwargs["vllm_server_timeout"] = trl.vllm_server_timeout
             if trl.vllm_guided_decoding_regex:
@@ -106,18 +102,17 @@ class GRPOStrategy:
         return grpo_args_kwargs
 
     @classmethod
-    def set_trainer_args(cls, cfg: DictDefault) -> list[Any]:
+    def set_trainer_args(cls, cfg):
         trainer_args = []
         if cfg.trl and cfg.trl.reward_funcs:
             reward_funcs = []
             for reward_func_fqn in cfg.trl.reward_funcs:
                 reward_funcs.append(cls.get_reward_func(reward_func_fqn))
             trainer_args.append(reward_funcs)
-
         return trainer_args
 
     @classmethod
-    def set_trainer_kwargs(cls, cfg: DictDefault) -> dict[str, Any]:
+    def set_trainer_kwargs(cls, cfg):
         trainer_kwargs = {}
         if cfg.trl and cfg.trl.reward_processing_classes:
             trainer_kwargs["reward_processing_classes"] = (
@@ -131,7 +126,7 @@ class GRPOStrategy:
         return None
 
     @classmethod
-    def get_blocklist_args_kwargs(cls) -> list[str]:
+    def get_blocklist_args_kwargs(cls):
         return ["dataset_num_proc"]
 
     @classmethod
@@ -142,13 +137,13 @@ class GRPOStrategy:
         Args:
             reward_func_fqn (str): Fully qualified name of the reward function (e.g. r1_grpo.gsm8k_transform),
                 or a HF hub path to the reward model.
+        Raises:
+            ValueError: If the reward function does not accept at least two arguments.
 
         Returns:
             RewardFunc: A callable that accepts prompts and completions and returns rewards,
                 or a path to a reward model.
 
-        Raises:
-            ValueError: If the reward function does not accept at least two arguments.
         """
         try:
             # use importlib to dynamically load the reward function from the module

src/axolotl/core/trainers/grpo/args.py

@@ -11,4 +11,6 @@ from axolotl.core.training_args import AxolotlTrainingMixins
 
 @dataclass
 class AxolotlGRPOConfig(AxolotlTrainingMixins, GRPOConfig):
-    """Axolotl GRPO Config for GRPO training"""
+    """
+    Axolotl GRPO Config for GRPO training
+    """

src/axolotl/core/trainers/grpo/sampler.py

@@ -1,172 +0,0 @@
-"""Repeat random sampler (similar to the one implemented in
-https://github.com/huggingface/trl/blob/main/trl/trainer/grpo_trainer.py) that adds
-sequence parallelism functionality; i.e., duplicating data across ranks in the same
-sequence parallel group.
-"""
-
-from typing import Iterator, Sized
-
-import torch
-from torch.utils.data import Sampler
-
-
-class SequenceParallelRepeatRandomSampler(Sampler):
-    """Sampler for GRPO training with sequence parallelism.
-
-    This sampler ensures:
-    - Ranks in the same sequence parallel (SP) group receive identical data.
-    - Each index is repeated multiple times for sampling different completions.
-    - Entire batches are repeated for reuse in multiple updates.
-    - Data is properly distributed across SP groups.
-
-    In the table below, the values represent dataset indices. Each SP group has
-    `sequence_parallel_degree = 2` GPUs working together on the same data. There are 2
-    SP groups (SP0 and SP1), with `world_size = 4` total GPUs.
-
-                                               Sequence Parallel Groups
-                                        |       SP0        |       SP1        |
-                                        |  GPU 0  |  GPU 1 |  GPU 2  |  GPU 3 |
-                    global_step  step    <---> mini_repeat_count=3
-                                            <----------> batch_size=2 per SP group
-    grad_accum=2   ▲  ▲  0       0         [0 0 0  1 1 1]     [2 2 2  3 3 3]   <- SP groups get different data
-                   ▼  |  0       1         [0 0 0  1 1 1]     [2 2 2  3 3 3]   <- Same data for each SP group GPU
-                      |
-                      |  1       2         [0 0 0  1 1 1]     [2 2 2  3 3 3]   <- Repeat same indices for iterations
-    num_iterations=2  ▼  1       3         [0 0 0  1 1 1]     [2 2 2  3 3 3]   <- When using gradient accumulation
-
-                         2       4         [4 4 4  5 5 5]     [6 6 6  7 7 7]   <- New batch of data indices
-                         2       5         [4 4 4  5 5 5]     [6 6 6  7 7 7]
-                                            ...
-
-    Args:
-        dataset: Dataset to sample from.
-        mini_repeat_count: How many times to repeat each sample immediately.
-        world_size: Total number of processes.
-        rank: Rank of current process.
-        batch_size: Number of samples per batch.
-        repeat_count: How many times to repeat the full sampling process.
-        sequence_parallel_degree: Number of ranks in a sequence parallel group.
-        shuffle: Whether to shuffle the dataset.
-        seed: Random seed for shuffling.
-        drop_last: Whether to drop the last incomplete batch.
-    """
-
-    def __init__(
-        self,
-        dataset: Sized,
-        mini_repeat_count: int,
-        world_size: int,
-        rank: int,
-        batch_size: int = 1,
-        repeat_count: int = 1,
-        sequence_parallel_degree: int = 1,
-        shuffle: bool = True,
-        seed: int = 0,
-        drop_last: bool = False,
-    ):
-        self.dataset = dataset
-        self.mini_repeat_count = mini_repeat_count
-        self.batch_size = batch_size
-        self.repeat_count = repeat_count
-        self.shuffle = shuffle
-        self.seed = seed
-        self.drop_last = drop_last
-        self.epoch = 0
-
-        self.world_size = world_size
-        self.rank = rank
-
-        # Sequence parallelism parameters
-        self.sequence_parallel_degree = sequence_parallel_degree
-        self.num_sp_groups = world_size // sequence_parallel_degree
-        self.sp_group_id = rank // sequence_parallel_degree
-
-        # Adjust dataset size for distributed sampling
-        self.num_samples = len(self.dataset)
-        self.total_size = self.num_samples
-
-        # Calculate effective number of samples per SP group
-        if (
-            self.drop_last
-            and self.total_size % (self.num_sp_groups * self.batch_size) != 0
-        ):
-            # Drop last incomplete batch if drop_last is True
-            self.num_samples_per_sp_group = (
-                self.total_size // self.batch_size // self.num_sp_groups
-            ) * self.batch_size
-        else:
-            # Round up to include last batch if drop_last is False
-            self.num_samples_per_sp_group = (
-                (self.total_size + self.batch_size * self.num_sp_groups - 1)
-                // (self.batch_size * self.num_sp_groups)
-                * self.batch_size
-            )
-
-        if shuffle:
-            self.generator = torch.Generator()
-            self.generator.manual_seed(seed)
-
-    def __iter__(self) -> Iterator[int]:
-        """Creates iterator over dataset indices.
-
-        Returns:
-            Iterator that yields indices into the dataset.
-        """
-        # Deterministically shuffle based on epoch and seed
-        if self.shuffle:
-            indices = torch.randperm(
-                self.num_samples, generator=self.generator
-            ).tolist()
-        else:
-            indices = list(range(self.num_samples))
-
-        # Add extra samples to make it evenly divisible by batch_size
-        if len(indices) % self.batch_size != 0:
-            padding = indices[: self.batch_size - len(indices) % self.batch_size]
-            indices += padding
-
-        # Subsample based on SP group ID
-        # Each SP group gets distinct batches of data
-        batch_indices = []
-        for i in range(0, len(indices), self.batch_size * self.num_sp_groups):
-            start_idx = i + self.sp_group_id * self.batch_size
-            end_idx = min(start_idx + self.batch_size, len(indices))
-            if start_idx < len(indices):
-                for j in range(self.batch_size):
-                    if start_idx + j < end_idx:
-                        batch_indices.append(indices[start_idx + j])
-
-        # Make sure batch_indices is exactly batch_size * num_batches_per_sp_group
-        if self.drop_last:
-            num_batches_per_sp_group = self.num_samples_per_sp_group // self.batch_size
-            target_len = self.batch_size * num_batches_per_sp_group
-            if len(batch_indices) > target_len:
-                batch_indices = batch_indices[:target_len]
-
-        # Apply the GRPO repeat pattern
-        final_indices = []
-        for _ in range(self.repeat_count):
-            for idx in batch_indices:
-                for _ in range(self.mini_repeat_count):
-                    final_indices.append(idx)
-
-        return iter(final_indices)
-
-    def __len__(self) -> int:
-        """Returns the total length of the iterable including repetitions.
-
-        Returns:
-            Total number of samples.
-        """
-        # Total length including all repetitions
-        return (
-            self.num_samples_per_sp_group * self.mini_repeat_count * self.repeat_count
-        )
-
-    def set_epoch(self, epoch: int) -> None:
-        """Sets the epoch for this sampler.
-
-        Args:
-            epoch: Epoch number to use for shuffling.
-        """
-        self.epoch = epoch

src/axolotl/core/trainers/grpo/trainer.py

@@ -1,63 +1,23 @@
-"""Axolotl GRPO trainers (with and without sequence parallelism handling)"""
+"""
+Axolotl GRPO trainer
+"""
 
-# pylint: disable=too-many-lines,duplicate-code,protected-access,no-member
-
-import warnings
 from contextlib import nullcontext
-from typing import Any
 
-import datasets
+from accelerate.utils import is_deepspeed_available, is_peft_model
-import torch
-import torch.distributed as dist
-import torch.utils.data
-from accelerate.utils import (
-    broadcast_object_list,
-    gather,
-    gather_object,
-    is_peft_model,
-)
-from datasets import Dataset, IterableDataset
-from torch import nn
-from torch.utils.data import (
-    BatchSampler,
-    DataLoader,
-    Sampler,
-)
-from transformers import (
-    PreTrainedModel,
-    PreTrainedTokenizerBase,
-    Trainer,
-    TrainerCallback,
-)
-from transformers.trainer_utils import seed_worker
-from transformers.utils import is_peft_available
 from trl import GRPOTrainer
-from trl.data_utils import (
+from trl.extras.profiling import profiling_decorator
-    apply_chat_template,
-    is_conversational,
-    maybe_apply_chat_template,
-)
-from trl.extras.profiling import profiling_context, profiling_decorator
-from trl.import_utils import is_deepspeed_available
-from trl.models import unwrap_model_for_generation
-from trl.trainer.grpo_config import GRPOConfig
-from trl.trainer.grpo_trainer import RewardFunc, nanstd
-from trl.trainer.utils import pad
 
-from axolotl.core.trainers.grpo.sampler import SequenceParallelRepeatRandomSampler
 from axolotl.core.trainers.mixins import RngLoaderMixin, SchedulerMixin
-from axolotl.monkeypatch.attention.ring_attn.patch import get_ring_attn_group
-
-if is_peft_available():
-    # pylint: disable=unused-import
-    from peft import PeftConfig
 
 if is_deepspeed_available():
     import deepspeed
 
 
 class AxolotlGRPOTrainer(RngLoaderMixin, SchedulerMixin, GRPOTrainer):
-    """Extend the base GRPOTrainer for axolotl helpers"""
+    """
+    Extend the base GRPOTrainer for axolotl helpers
+    """
 
     _tag_names = ["trl", "grpo", "axolotl"]
 
@@ -107,600 +67,3 @@ class AxolotlGRPOTrainer(RngLoaderMixin, SchedulerMixin, GRPOTrainer):
         # Reset cache on main process
         if self.accelerator.is_main_process:
             self.vllm_client.reset_prefix_cache()
-
-
-class AxolotlGRPOSequenceParallelTrainer(AxolotlGRPOTrainer):
-    """Extend the base GRPOTrainer for sequence parallelism handling"""
-
-    def __init__(
-        self,
-        model: str | PreTrainedModel,
-        reward_funcs: RewardFunc | list[RewardFunc],
-        args: GRPOConfig | None = None,
-        train_dataset: Dataset | IterableDataset | None = None,
-        eval_dataset: (
-            Dataset | IterableDataset | dict[str, Dataset | IterableDataset] | None
-        ) = None,
-        processing_class: PreTrainedTokenizerBase | None = None,
-        reward_processing_classes: (
-            PreTrainedTokenizerBase | list[PreTrainedTokenizerBase] | None
-        ) = None,
-        callbacks: list[TrainerCallback] | None = None,
-        optimizers: tuple[
-            torch.optim.Optimizer | None, torch.optim.lr_scheduler.LambdaLR | None
-        ] = (None, None),
-        peft_config: "PeftConfig | None" = None,
-    ):
-        # First call the superclass constructor with all arguments
-        super().__init__(
-            model=model,
-            reward_funcs=reward_funcs,
-            args=args,
-            train_dataset=train_dataset,
-            eval_dataset=eval_dataset,
-            processing_class=processing_class,
-            reward_processing_classes=reward_processing_classes,
-            callbacks=callbacks,
-            optimizers=optimizers,
-            peft_config=peft_config,
-        )
-
-        # Get number of SP groups (number of processes divided by SP degree)
-        num_processes = self.accelerator.num_processes
-        num_sp_groups = num_processes // self.args.sequence_parallel_degree
-
-        # Calculate batch size per SP group (not per process)
-        sp_group_batch_size = self.args.per_device_train_batch_size * num_sp_groups
-        possible_values = [
-            n_gen
-            for n_gen in range(2, sp_group_batch_size + 1)
-            if (sp_group_batch_size) % n_gen == 0
-        ]
-
-        if self.num_generations not in possible_values:
-            raise ValueError(
-                f"The batch size per SP group ({num_sp_groups} x "
-                f"{self.args.per_device_train_batch_size}) must be evenly divisible by "
-                f"the number of generations per prompt ({self.num_generations}). Given "
-                "the current configuration, the valid values for the number of "
-                f"generations are: {possible_values}."
-            )
-
-        if self.args.eval_strategy != "no":
-            # If sequence parallelism is enabled, calculate batch size per SP group
-            sp_group_eval_batch_size = args.per_device_eval_batch_size * num_sp_groups  # type: ignore[union-attr]
-            possible_values = [
-                n_gen
-                for n_gen in range(2, sp_group_eval_batch_size + 1)
-                if (sp_group_eval_batch_size) % n_gen == 0
-            ]
-
-            if self.num_generations not in possible_values:
-                raise ValueError(
-                    f"With sequence parallelism (degree {self.args.sequence_parallel_degree}), "
-                    f"the eval batch size per SP group ({num_sp_groups} x {self.args.per_device_eval_batch_size}) "
-                    f"must be evenly divisible by the number of generations per prompt "
-                    f"({self.num_generations}). Given the current eval batch size, "
-                    f"the valid values for the number of generations are: {possible_values}."
-                )
-
-        # Initialize the SP group
-        self.sp_group = get_ring_attn_group()
-        self.rank = dist.get_rank()
-        self.world_size = dist.get_world_size()
-        self.local_rank = dist.get_rank(group=self.sp_group)
-        self.local_world_size = dist.get_world_size(group=self.sp_group)
-
-    def _get_train_sampler(self) -> Sampler:
-        effective_batch_size = (
-            self.args.per_device_train_batch_size
-            * self.world_size
-            * self.args.gradient_accumulation_steps
-        )
-
-        return SequenceParallelRepeatRandomSampler(
-            dataset=self.train_dataset,
-            mini_repeat_count=self.num_generations,
-            world_size=self.world_size,
-            rank=self.rank,
-            batch_size=effective_batch_size
-            // self.num_generations
-            // self.args.sequence_parallel_degree,
-            repeat_count=self.num_iterations * self.args.gradient_accumulation_steps,
-            sequence_parallel_degree=self.args.sequence_parallel_degree,
-            shuffle=True,
-            seed=self.args.seed,
-            drop_last=True,
-        )
-
-    def _create_dataloader_params(self, is_eval=False, custom_batch_size=None):
-        """Create common dataloader parameters for train or eval."""
-        batch_size = custom_batch_size or (
-            self.args.eval_batch_size if is_eval else self._train_batch_size
-        )
-
-        params = {
-            "batch_size": batch_size,
-            "collate_fn": self.data_collator,
-            "num_workers": self.args.dataloader_num_workers,
-            "pin_memory": self.args.dataloader_pin_memory,
-        }
-
-        # Add persistent workers only for training
-        if not is_eval and hasattr(self.args, "dataloader_persistent_workers"):
-            params["persistent_workers"] = self.args.dataloader_persistent_workers
-
-        # Add prefetch factor if specified
-        if self.args.dataloader_prefetch_factor:
-            params["prefetch_factor"] = self.args.dataloader_prefetch_factor
-
-        return params
-
-    def _prepare_dataloader(
-        self, dataset, sampler, is_eval=False, custom_batch_size=None
-    ):
-        """Prepare a dataloader with the given dataset and sampler."""
-        # Get base parameters
-        dataloader_params = self._create_dataloader_params(is_eval, custom_batch_size)
-
-        # Add sampler configuration
-        if not isinstance(dataset, torch.utils.data.IterableDataset):
-            if isinstance(sampler, BatchSampler):
-                # batch_size and batch_sampler are mutually exclusive
-                dataloader_params["batch_sampler"] = sampler
-                del dataloader_params["batch_size"]
-            else:
-                dataloader_params["sampler"] = sampler
-                dataloader_params["drop_last"] = self.args.dataloader_drop_last
-
-            if not is_eval:
-                dataloader_params["worker_init_fn"] = seed_worker
-
-        # Create the dataloader
-        dataloader = DataLoader(dataset, **dataloader_params)
-
-        if self.args.sample_packing and (
-            (not is_eval and not self.args.pretraining)
-            or (is_eval and self.args.eval_sample_packing is not False)
-        ):
-            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
-
-        # Otherwise prepare with accelerator
-        return self.accelerator.prepare_data_loader(dataloader)
-
-    def get_train_dataloader(self) -> DataLoader:
-        """Get dataloader for training"""
-        train_dataset = self.train_dataset
-        # pylint: disable=access-member-before-definition
-        data_collator = self.data_collator  # type: ignore
-
-        # Handle dataset preprocessing
-        if isinstance(train_dataset, datasets.Dataset):
-            # Add debug print before any modifications
-            if self.args.sample_packing and not self.args.pretraining:
-                train_dataset = train_dataset.remove_columns(["length"])
-            if not self.args.sample_packing or self.args.pretraining:
-                train_dataset = self._remove_unused_columns(
-                    train_dataset, description="training"
-                )
-        else:
-            self.data_collator = self._get_collator_with_removed_columns(  # pylint: disable=attribute-defined-outside-init
-                data_collator,
-                description="training",
-            )
-
-        # Get sampler and create dataloader
-        sampler = self._get_train_sampler()
-        dataloader = self._prepare_dataloader(train_dataset, sampler, is_eval=False)
-
-        return dataloader
-
-    def _generate_and_score_completions(
-        self, inputs: list[dict[str, torch.Tensor | Any]]
-    ) -> dict[str, torch.Tensor | Any]:
-        device = self.accelerator.device
-        mode = "eval" if self.control.should_evaluate else "train"
-
-        prompts = [x["prompt"] for x in inputs]
-        prompts_text = [
-            maybe_apply_chat_template(example, self.processing_class)["prompt"]
-            for example in inputs
-        ]
-        prompt_inputs = self.processing_class(
-            text=prompts_text,
-            return_tensors="pt",
-            padding=True,
-            padding_side="left",
-            add_special_tokens=False,
-        )
-        prompt_inputs = Trainer._prepare_inputs(self, prompt_inputs)
-        prompt_ids, prompt_mask = (
-            prompt_inputs["input_ids"],
-            prompt_inputs["attention_mask"],
-        )
-
-        if self.max_prompt_length is not None:
-            prompt_ids = prompt_ids[:, -self.max_prompt_length :]
-            prompt_mask = prompt_mask[:, -self.max_prompt_length :]
-
-        # Generate completions using either vLLM or regular generation
-        if self.args.use_vllm:
-            # First, have main process load weights if needed
-            # pylint: disable=access-member-before-definition
-            if self.state.global_step != self._last_loaded_step:  # type: ignore[has-type]
-                self._move_model_to_vllm()
-                # pylint: disable=attribute-defined-outside-init
-                self._last_loaded_step = self.state.global_step
-
-            # Generate completions using vLLM: gather all prompts and use them in a single call in the main process
-            all_prompts_text = gather_object(prompts_text)
-            if self.accelerator.is_main_process:
-                if self.args.sequence_parallel_degree > 1:
-                    # Calculate sequence parallel group information
-                    world_size = self.accelerator.num_processes
-                    sequence_parallel_degree = self.args.sequence_parallel_degree
-                    num_sp_groups = world_size // sequence_parallel_degree
-
-                    # Since processes in the same SP group have the same prompts, we need to ensure
-                    # we only take one copy of each prompt from each SP group
-                    ordered_set_of_prompts = []
-                    for sp_group_id in range(num_sp_groups):
-                        # Get the first process from each SP group (typically the group leader)
-                        group_leader_rank = sp_group_id * sequence_parallel_degree
-
-                        # Extract prompts from this SP group, accounting for num_generations duplicates
-                        # We only need prompts from one rank in each SP group
-                        group_prompts = all_prompts_text[
-                            group_leader_rank
-                            * len(prompts_text) : (group_leader_rank + 1)
-                            * len(prompts_text) : self.num_generations
-                        ]
-
-                        ordered_set_of_prompts.extend(group_prompts)
-                else:
-                    # Since 'prompts' contains 'num_generations' duplicates, we first take unique prompts, and generate
-                    # num_generations outputs for each one. This is faster than generating outputs for each duplicate
-                    # prompt individually.
-                    ordered_set_of_prompts = all_prompts_text[
-                        :: self.num_generations * self.args.sequence_parallel_degree
-                    ]
-
-                with profiling_context(self, "vLLM.generate"):
-                    completion_ids = self.vllm_client.generate(
-                        prompts=ordered_set_of_prompts,
-                        n=self.num_generations,
-                        repetition_penalty=self.repetition_penalty,
-                        temperature=self.temperature,
-                        top_p=self.top_p,
-                        top_k=-1 if self.top_k is None else self.top_k,
-                        min_p=0.0 if self.min_p is None else self.min_p,
-                        max_tokens=self.max_completion_length,
-                        guided_decoding_regex=self.guided_decoding_regex,
-                    )
-            else:
-                completion_ids = [None] * (
-                    len(all_prompts_text) // self.args.sequence_parallel_degree
-                )
-
-            # Broadcast the completions from the main process to all processes
-            completion_ids = broadcast_object_list(completion_ids, from_process=0)
-
-            # Determine the appropriate slice based on sequence parallelism
-            if self.args.sequence_parallel_degree > 1:
-                # Calculate SP group ID (which group of ranks this rank belongs to)
-                sp_group_id = self.accelerator.process_index // self.local_world_size
-
-                # Calculate the start index for this SP group
-                sp_group_start = sp_group_id * len(prompts) * self.local_world_size
-
-                # All ranks in the same SP group get the same data slice
-                process_slice = slice(
-                    sp_group_start,
-                    sp_group_start + len(prompts),
-                )
-                completion_ids = completion_ids[process_slice]
-            else:
-                # Original behavior for non-sequence parallel case
-                process_slice = slice(
-                    self.accelerator.process_index * len(prompts),
-                    (self.accelerator.process_index + 1) * len(prompts),
-                )
-                completion_ids = completion_ids[process_slice]
-
-            # Pad the completions, and concatenate them with the prompts
-            completion_ids = [
-                torch.tensor(ids, device=device) for ids in completion_ids
-            ]
-            completion_ids = pad(
-                completion_ids, padding_value=self.processing_class.pad_token_id
-            )
-            prompt_completion_ids = torch.cat([prompt_ids, completion_ids], dim=1)
-        else:
-            # Regular generation path
-            with unwrap_model_for_generation(
-                self.model_wrapped,
-                self.accelerator,
-                gather_deepspeed3_params=self.args.ds3_gather_for_generation,
-            ) as unwrapped_model:
-                prompt_completion_ids = unwrapped_model.generate(
-                    prompt_ids,
-                    attention_mask=prompt_mask,
-                    generation_config=self.generation_config,
-                )
-
-            # Compute prompt length and extract completion ids
-            prompt_length = prompt_ids.size(1)
-            prompt_ids = prompt_completion_ids[:, :prompt_length]
-            completion_ids = prompt_completion_ids[:, prompt_length:]
-
-        # Mask everything after the first EOS token
-        is_eos = completion_ids == self.processing_class.eos_token_id
-        eos_idx = torch.full(
-            (is_eos.size(0),), is_eos.size(1), dtype=torch.long, device=device
-        )
-        eos_idx[is_eos.any(dim=1)] = is_eos.int().argmax(dim=1)[is_eos.any(dim=1)]
-        sequence_indices = torch.arange(is_eos.size(1), device=device).expand(
-            is_eos.size(0), -1
-        )
-        completion_mask = (sequence_indices <= eos_idx.unsqueeze(1)).int()
-
-        # If mask_truncated_completions is enabled, zero out truncated completions in completion_mask
-        if self.args.mask_truncated_completions:
-            truncated_completions = ~is_eos.any(dim=1)
-            completion_mask = (
-                completion_mask * (~truncated_completions).unsqueeze(1).int()
-            )
-
-        # Concatenate prompt_mask with completion_mask for logit computation
-        attention_mask = torch.cat([prompt_mask, completion_mask], dim=1)  # (B, P+C)
-
-        logits_to_keep = completion_ids.size(
-            1
-        )  # we only need to compute the logits for the completion tokens
-        batch_size = (
-            self.args.per_device_train_batch_size
-            if mode == "train"
-            else self.args.per_device_eval_batch_size
-        )
-
-        with torch.no_grad():
-            # When using num_iterations == 1, old_per_token_logps == per_token_logps, so we can skip it's
-            # computation here, and use per_token_logps.detach() instead.
-            if self.num_iterations > 1:
-                old_per_token_logps = self._get_per_token_logps(
-                    self.model,
-                    prompt_completion_ids,
-                    attention_mask,
-                    logits_to_keep,
-                    batch_size,
-                )
-            else:
-                old_per_token_logps = None
-
-            if self.beta == 0.0:
-                ref_per_token_logps = None
-            elif self.ref_model is not None:
-                ref_per_token_logps = self._get_per_token_logps(
-                    self.ref_model,
-                    prompt_completion_ids,
-                    attention_mask,
-                    logits_to_keep,
-                    batch_size,
-                )
-            else:
-                with self.accelerator.unwrap_model(self.model).disable_adapter():
-                    ref_per_token_logps = self._get_per_token_logps(
-                        self.model,
-                        prompt_completion_ids,
-                        attention_mask,
-                        logits_to_keep,
-                        batch_size,
-                    )
-
-        # Decode the generated completions
-        completions_text = self.processing_class.batch_decode(
-            completion_ids, skip_special_tokens=True
-        )
-        if is_conversational(inputs[0]):
-            completions = []
-            for prompt, completion in zip(prompts, completions_text):
-                bootstrap = (
-                    prompt.pop()["content"] if prompt[-1]["role"] == "assistant" else ""
-                )
-                completions.append(
-                    [{"role": "assistant", "content": bootstrap + completion}]
-                )
-        else:
-            completions = completions_text
-
-        rewards_per_func = torch.zeros(
-            len(prompts), len(self.reward_funcs), device=device
-        )
-        for i, (reward_func, reward_processing_class, reward_func_name) in enumerate(
-            zip(
-                self.reward_funcs,
-                self.reward_processing_classes,
-                self.reward_func_names,
-            )
-        ):
-            with profiling_context(self, reward_func_name):
-                if isinstance(
-                    reward_func, nn.Module
-                ):  # Module instead of PretrainedModel for compat with compiled models
-                    if is_conversational(inputs[0]):
-                        messages = [
-                            {"messages": p + c} for p, c in zip(prompts, completions)
-                        ]
-                        texts = [
-                            apply_chat_template(x, reward_processing_class)["text"]
-                            for x in messages
-                        ]
-                    else:
-                        texts = [p + c for p, c in zip(prompts, completions)]
-                    reward_inputs = reward_processing_class(
-                        text=texts,
-                        return_tensors="pt",
-                        padding=True,
-                        padding_side="right",
-                        add_special_tokens=False,
-                    )
-                    reward_inputs = Trainer._prepare_inputs(self, reward_inputs)
-                    with torch.inference_mode():
-                        rewards_per_func[:, i] = reward_func(**reward_inputs).logits[
-                            :, 0
-                        ]  # Shape (B*G,)
-                else:
-                    # Repeat all input columns (but "prompt" and "completion") to match the number of generations
-                    keys = [
-                        key for key in inputs[0] if key not in ["prompt", "completion"]
-                    ]
-                    reward_kwargs = {
-                        key: [example[key] for example in inputs] for key in keys
-                    }
-                    output_reward_func = reward_func(
-                        prompts=prompts, completions=completions, **reward_kwargs
-                    )
-                    # Convert None values to NaN
-                    output_reward_func = [
-                        reward if reward is not None else torch.nan
-                        for reward in output_reward_func
-                    ]
-
-                    rewards_per_func[:, i] = torch.tensor(
-                        output_reward_func, dtype=torch.float32, device=device
-                    )
-
-        # If all reward functions return None for a given row, issue a detailed warning
-        if torch.isnan(rewards_per_func).all(dim=1).any():
-            nan_row_idx = (
-                torch.isnan(rewards_per_func).all(dim=1).nonzero(as_tuple=True)[0][0]
-            )
-            row_reward_kwargs = {
-                key: value[nan_row_idx] for key, value in reward_kwargs.items()
-            }
-            row_reward_kwargs["prompt"] = prompts[nan_row_idx]
-            row_reward_kwargs["completion"] = completions[nan_row_idx]
-            warnings.warn(
-                f"All reward functions returned None for the following kwargs: {row_reward_kwargs}. "
-                "Please ensure that at least one reward function returns a valid reward."
-            )
-
-        # Gather the reward per function: this part is crucial, because the rewards are normalized per group and the
-        # completions may be distributed across processes
-        rewards_per_func = gather(rewards_per_func)
-
-        # Apply weights to each reward function's output and sum
-        rewards = (
-            rewards_per_func * self.reward_weights.to(device).unsqueeze(0)
-        ).nansum(dim=1)
-
-        # Compute grouped-wise rewards
-        mean_grouped_rewards = rewards.view(-1, self.num_generations).mean(dim=1)
-        std_grouped_rewards = rewards.view(-1, self.num_generations).std(dim=1)
-
-        # Normalize the rewards to compute the advantages
-        mean_grouped_rewards = mean_grouped_rewards.repeat_interleave(
-            self.num_generations, dim=0
-        )
-        std_grouped_rewards = std_grouped_rewards.repeat_interleave(
-            self.num_generations, dim=0
-        )
-        advantages = rewards - mean_grouped_rewards
-        if self.args.scale_rewards:
-            advantages = advantages / (std_grouped_rewards + 1e-4)
-
-        # Slice to keep only the local part of the data
-        if self.args.sequence_parallel_degree > 1:
-            # Calculate SP group ID (which group of ranks this rank belongs to)
-            sp_group_id = self.accelerator.process_index // self.local_world_size
-
-            # Calculate the start index for this SP group
-            sp_group_start = sp_group_id * len(prompts) * self.local_world_size
-
-            # All ranks in the same SP group get the same data slice
-            process_slice = slice(
-                sp_group_start,
-                sp_group_start + len(prompts),
-            )
-        else:
-            # Original behavior for non-sequence parallel case
-            process_slice = slice(
-                self.accelerator.process_index * len(prompts),
-                (self.accelerator.process_index + 1) * len(prompts),
-            )
-        advantages = advantages[process_slice]
-
-        # Log the metrics
-        if mode == "train":
-            self._total_train_tokens += (
-                self.accelerator.gather_for_metrics(attention_mask.sum()).sum().item()
-            )
-        self._metrics[mode]["num_tokens"] = [self._total_train_tokens]
-
-        # log completion lengths, mean, min, max
-        agg_completion_mask = self.accelerator.gather_for_metrics(
-            completion_mask.sum(1)
-        )
-        self._metrics[mode]["completions/mean_length"].append(
-            agg_completion_mask.float().mean().item()
-        )
-        self._metrics[mode]["completions/min_length"].append(
-            agg_completion_mask.float().min().item()
-        )
-        self._metrics[mode]["completions/max_length"].append(
-            agg_completion_mask.float().max().item()
-        )
-
-        # identify sequences that terminated with EOS and log their lengths
-        agg_terminated_with_eos = self.accelerator.gather_for_metrics(is_eos.any(dim=1))
-        term_completion_mask = agg_completion_mask[agg_terminated_with_eos]
-        clipped_completions_ratio = 1 - len(term_completion_mask) / len(
-            agg_completion_mask
-        )
-        self._metrics[mode]["completions/clipped_ratio"].append(
-            clipped_completions_ratio
-        )
-        if len(term_completion_mask) == 0:
-            # edge case where no completed sequences are found
-            term_completion_mask = torch.zeros(1, device=device)
-        self._metrics[mode]["completions/mean_terminated_length"].append(
-            term_completion_mask.float().mean().item()
-        )
-        self._metrics[mode]["completions/min_terminated_length"].append(
-            term_completion_mask.float().min().item()
-        )
-        self._metrics[mode]["completions/max_terminated_length"].append(
-            term_completion_mask.float().max().item()
-        )
-
-        # Calculate mean reward per function, but only for samples where the function was applied (non-NaN values)
-        for i, reward_func_name in enumerate(self.reward_func_names):
-            mean_rewards = torch.nanmean(rewards_per_func[:, i]).item()
-            self._metrics[mode][f"rewards/{reward_func_name}/mean"].append(mean_rewards)
-            std_rewards = nanstd(rewards_per_func[:, i]).item()
-            self._metrics[mode][f"rewards/{reward_func_name}/std"].append(std_rewards)
-        self._metrics[mode]["reward"].append(mean_grouped_rewards.mean().item())
-        self._metrics[mode]["reward_std"].append(std_grouped_rewards.mean().item())
-
-        # Log prompt and completion texts
-        self._textual_logs["prompt"].extend(gather_object(prompts_text))
-        self._textual_logs["completion"].extend(gather_object(completions_text))
-        for i, name in enumerate(self.reward_func_names):
-            self._textual_logs["rewards"][name].extend(rewards_per_func[:, i].tolist())
-
-        return {
-            "prompt_ids": prompt_ids,
-            "prompt_mask": prompt_mask,
-            "completion_ids": completion_ids,
-            "completion_mask": completion_mask,
-            "advantages": advantages,
-            "old_per_token_logps": old_per_token_logps,
-            "ref_per_token_logps": ref_per_token_logps,
-        }

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

@@ -6,4 +6,4 @@
 from .optimizer import OptimizerMixin
 from .rng_state_loader import RngLoaderMixin
 from .scheduler import SchedulerMixin
-from .sequence_parallel import SequenceParallelMixin
+from .sequence_parallel import SequenceParallelContextManager, SequenceParallelMixin

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

@@ -1,13 +1,85 @@
-"""Module for Axolotl trainer sequence parallelism mixin"""
+"""
+Module for Axolotl trainer sequence parallelism mixin and training context manager
+"""
 
+import functools
+import logging
+
+import torch
 import torch.distributed as dist
 from datasets import Dataset
+from torch import nn
 from torch.utils.data import DistributedSampler, Sampler
+from torch.utils.hooks import RemovableHandle
 
 from axolotl.monkeypatch.attention.ring_attn import (
+    RingAttnFunc,
     get_ring_attn_group,
+    update_ring_attn_params,
 )
 
+LOG = logging.getLogger(__name__)
+
+
+def apply_sequence_parallelism(
+    batch: dict[str, torch.Tensor],
+    local_rank: int,
+    local_world_size: int,
+    ring_attn_func: RingAttnFunc,
+) -> dict[str, torch.Tensor]:
+    """
+    Apply sequence parallelism slicing to a batch.
+
+    Args:
+        batch: Batch dictionary (e.g., input_ids, attention_mask, etc.)
+        local_rank: Local rank in the sequence parallel group
+        local_world_size: World size of the sequence parallel group
+        ring_attn_func: The ring attention function to use
+
+    Returns:
+        Sliced batch dictionary.
+    """
+    # Update ring attention params if needed
+    if batch.get("position_ids") is not None:
+        update_ring_attn_params(position_ids=batch["position_ids"])
+
+    # Slice batch for sequence parallel processing
+    total_seq_len = batch["input_ids"].size(1)
+    for key in batch:
+        if (
+            key in batch
+            and isinstance(batch[key], torch.Tensor)
+            and batch[key].dim() > 1
+            and batch[key].size(1) == total_seq_len
+        ):
+
+            if ring_attn_func in [
+                RingAttnFunc.VARLEN_LLAMA3,
+                RingAttnFunc.BATCH_RING,
+            ]:
+                # Split in sequential fashion and grab this rank's chunk
+                batch[key] = (
+                    batch[key].chunk(local_world_size, dim=1)[local_rank].contiguous()
+                )
+            elif ring_attn_func is RingAttnFunc.BATCH_ZIGZAG:
+                chunks = batch[key].chunk(2 * local_world_size, dim=1)
+
+                # Take rank's chunk and opposing chunk for zigzag pattern
+                selected_chunks = [
+                    chunks[local_rank],
+                    chunks[2 * local_world_size - local_rank - 1],
+                ]
+                batch[key] = torch.cat(selected_chunks, dim=1).contiguous()
+            elif ring_attn_func is RingAttnFunc.BATCH_STRIPE:
+                # Split into striped data and stack
+                tensor = torch.stack(
+                    batch[key].split(local_world_size, dim=1),
+                    dim=1,
+                ).transpose(1, 2)
+                batch[key] = tensor[:, local_rank].contiguous()
+
+    return batch
+
 
 class SequenceParallelMixin:
     """
@@ -85,3 +157,157 @@ class SequenceParallelMixin:
         return self._create_sequence_parallel_sampler(
             eval_dataset, shuffle=False, is_eval=True
         )
+
+
+class SequenceParallelContextManager:
+    """
+    Context manager for sequence parallelism operations.
+
+    This class provides a context that will automatically apply sequence parallelism
+    during model forward passes using a pre-forward hook, and gather outputs from
+    across the sequence parallelism group using a post-forward hook.
+    """
+
+    def __init__(
+        self,
+        model: nn.Module,
+        sequence_parallel_degree: int,
+        ring_attn_func: RingAttnFunc,
+    ):
+        self.model = model
+        self.sequence_parallel_degree = sequence_parallel_degree
+        self.ring_attn_func = ring_attn_func
+        self.process_group = get_ring_attn_group()
+
+        # Initialize sequence parallel group details
+        self.local_rank = dist.get_rank(self.process_group)
+        self.local_world_size = dist.get_world_size(self.process_group)
+
+        # Will store hook handles for removal
+        self.hook_handles: list[RemovableHandle] = []
+
+        # Create a partially applied version of the apply_sequence_parallelism function
+        # with pre-configured params
+        self.apply_sequence_parallelism = functools.partial(
+            apply_sequence_parallelism,
+            local_rank=self.local_rank,
+            local_world_size=self.local_world_size,
+            ring_attn_func=self.ring_attn_func,
+        )
+
+    def __enter__(self):
+        # Forward pre-hook to apply sequence parallelism
+        def sequence_parallel_pre_hook(_, args, kwargs):
+            # Apply sequence parallelism to kwargs
+            kwargs = self.apply_sequence_parallelism(batch=kwargs)
+            return args, kwargs
+
+        # Forward post-hook to gather outputs
+        def sequence_parallel_post_hook(_, __, output):
+            # Gather the sharded outputs
+            return self.gather_outputs(output)
+
+        # Register both hooks
+        self.hook_handles.append(
+            self.model.register_forward_pre_hook(
+                sequence_parallel_pre_hook, with_kwargs=True
+            )
+        )
+        self.hook_handles.append(
+            self.model.register_forward_hook(sequence_parallel_post_hook)
+        )
+
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        # Remove all hooks
+        for handle in self.hook_handles:
+            handle.remove()
+        self.hook_handles = []
+
+    def gather_outputs(self, output):
+        """Gather sharded outputs from all ranks and reconstruct the full tensor."""
+        # Handle different output formats (dict, tensor, etc.)
+        if isinstance(output, dict):
+            gathered_output = {}
+            for key, value in output.items():
+                if isinstance(value, torch.Tensor) and value.dim() > 1:
+                    # Gather logits or other sequence-sharded tensors
+                    gathered_value = self.gather_tensor(value)
+                    gathered_output[key] = gathered_value
+                else:
+                    gathered_value = value.clone()
+                    dist.all_reduce(
+                        gathered_value, op=dist.ReduceOp.SUM, group=self.process_group
+                    )
+                    gathered_output[key] = gathered_value
+            return gathered_output
+        if isinstance(output, torch.Tensor):
+            return self.gather_tensor(output)
+
+        return output
+
+    def gather_tensor(self, tensor):
+        """Gather a sharded tensor from all ranks."""
+        # Prepare tensors for all_gather
+        world_size = self.local_world_size
+
+        # Create list to store tensors from all ranks
+        gathered_tensors = [torch.zeros_like(tensor) for _ in range(world_size)]
+
+        # All-gather operation
+        dist.all_gather(gathered_tensors, tensor, group=self.process_group)
+
+        # Concatenate along sequence dimension (typically dim=1)
+        if self.ring_attn_func in [RingAttnFunc.VARLEN_LLAMA3, RingAttnFunc.BATCH_RING]:
+            # Simple concatenation for standard sharding
+            return torch.cat(gathered_tensors, dim=1)
+
+        if self.ring_attn_func is RingAttnFunc.BATCH_ZIGZAG:
+            # Each rank has a pattern of (rank, world_size*2-rank-1)
+            reconstituted_tensors = [None] * (world_size * 2)
+
+            # First, split each gathered tensor into its two chunks
+            for rank, gathered_tensor in enumerate(gathered_tensors):
+                # Each tensor contains two chunks in the sequence dimension
+                chunk_size = gathered_tensor.size(1) // 2
+                chunk1, chunk2 = gathered_tensor.split(chunk_size, dim=1)
+
+                # Place chunks in their original positions
+                reconstituted_tensors[rank] = chunk1
+                reconstituted_tensors[world_size * 2 - rank - 1] = chunk2
+
+            # Concatenate the reconstituted tensors in the correct order
+            return torch.cat(reconstituted_tensors, dim=1)
+
+        # Otherwise, RingAttnFunc.BATCH_STRIPE
+        # In striping, each rank has every world_size-th slice
+        batch_size = tensor.size(0)
+        hidden_dim = tensor.size(-1)
+
+        # First, determine the full sequence length
+        total_seq_len = 0
+        for t in gathered_tensors:
+            total_seq_len += t.size(1)
+
+        # Create a tensor to hold the unstriped result
+        result = torch.zeros(
+            batch_size,
+            total_seq_len,
+            hidden_dim,
+            dtype=tensor.dtype,
+            device=tensor.device,
+        )
+
+        # For each rank's tensor, distribute its slices to the correct positions
+        for rank, gathered_tensor in enumerate(gathered_tensors):
+            # The rank's tensor contains every world_size-th slice
+            # starting from its rank position
+            seq_len = gathered_tensor.size(1)
+            for i in range(seq_len):
+                # Calculate the position in the full tensor
+                pos = i * world_size + rank
+                if pos < total_seq_len:
+                    result[:, pos] = gathered_tensor[:, i]
+
+        return result

src/axolotl/core/training_args.py

@@ -9,7 +9,7 @@ from PIL.Image import Resampling
 from transformers import TrainingArguments
 from trl import CPOConfig, KTOConfig, ORPOConfig, PRMConfig, RewardConfig
 
-from axolotl.utils.schemas.enums import RingAttnFunc
+from axolotl.monkeypatch.attention.ring_attn.patch import RingAttnFunc
 
 
 @dataclass

src/axolotl/integrations/llm_compressor/README.md

@@ -1,108 +0,0 @@
-# LLMCompressor Integration
-
-Fine-tune sparsified models in Axolotl using Neural Magic's [LLMCompressor](https://github.com/vllm-project/llm-compressor).
-
-This integration enables fine-tuning of models sparsified using LLMCompressor within the Axolotl training framework. By combining LLMCompressor's model compression capabilities with Axolotl's distributed training pipelines, users can efficiently fine-tune sparse models at scale.
-
-It uses Axolotl’s plugin system to hook into the fine-tuning flows while maintaining sparsity throughout training.
-
----
-
-## Requirements
-
-- Axolotl with `llmcompressor` extras:
-
-  ```bash
-  pip install "axolotl[llmcompressor]"
-  ```
-
-- Requires `llmcompressor >= 0.5.1`
-
-This will install all necessary dependencies to fine-tune sparsified models using the integration.
-
----
-
-## Usage
-
-To enable sparse fine-tuning with this integration, include the plugin in your Axolotl config:
-
-```yaml
-plugins:
-  - axolotl.integrations.llm_compressor.LLMCompressorPlugin
-
-llmcompressor:
-  recipe:
-    finetuning_stage:
-      finetuning_modifiers:
-        ConstantPruningModifier:
-          targets: [
-            're:.*q_proj.weight',
-            're:.*k_proj.weight',
-            're:.*v_proj.weight',
-            're:.*o_proj.weight',
-            're:.*gate_proj.weight',
-            're:.*up_proj.weight',
-            're:.*down_proj.weight',
-          ]
-          start: 0
-  save_compressed: true
-# ... (other training arguments)
-```
-
-This plugin **does not apply pruning or sparsification itself** — it is intended for **fine-tuning models that have already been sparsified**.
-
-Pre-sparsified checkpoints can be:
-- Generated using [LLMCompressor](https://github.com/vllm-project/llm-compressor)
-- Downloaded from [Neural Magic's Hugging Face page](https://huggingface.co/neuralmagic)
-- Any custom LLM with compatible sparsity patterns that you've created yourself
-
-To learn more about writing and customizing LLMCompressor recipes, refer to the official documentation:
-[https://github.com/vllm-project/llm-compressor/blob/main/README.md](https://github.com/vllm-project/llm-compressor/blob/main/README.md)
-
-### Storage Optimization with save_compressed
-
-Setting `save_compressed: true` in your configuration enables saving models in a compressed format, which:
-- Reduces disk space usage by approximately 40%
-- Maintains compatibility with vLLM for accelerated inference
-- Maintains compatibility with llmcompressor for further optimization (example: quantization)
-
-This option is highly recommended when working with sparse models to maximize the benefits of model compression.
-
-### Example Config
-
-See [`examples/llama-3/sparse-finetuning.yaml`](examples/llama-3/sparse-finetuning.yaml) for a complete example.
-
----
-
-## Inference with vLLM
-
-After fine-tuning your sparse model, you can leverage vLLM for efficient inference.
-You can also use LLMCompressor to apply additional quantization to your fine-tuned
-sparse model before inference for even greater performance benefits.:
-
-```python
-from vllm import LLM, SamplingParams
-
-prompts = [
-    "Hello, my name is",
-    "The president of the United States is",
-    "The capital of France is",
-    "The future of AI is",
-]
-sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
-llm = LLM("path/to/your/sparse/model")
-outputs = llm.generate(prompts, sampling_params)
-
-for output in outputs:
-    prompt = output.prompt
-    generated_text = output.outputs[0].text
-    print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
-```
-
-For more details on vLLM's capabilities and advanced configuration options, see the [official vLLM documentation](https://docs.vllm.ai/).
-
-## Learn More
-
-For details on available sparsity and quantization schemes, fine-tuning recipes, and usage examples, visit the official LLMCompressor repository:
-
-[https://github.com/vllm-project/llm-compressor](https://github.com/vllm-project/llm-compressor)

src/axolotl/integrations/llm_compressor/__init__.py

@@ -1,5 +0,0 @@
-"""Integration entry point for the LLMCompressor plugin."""
-
-from .plugin import LLMCompressorPlugin
-
-__all__ = ["LLMCompressorPlugin"]

src/axolotl/integrations/llm_compressor/args.py

@@ -1,40 +0,0 @@
-"""
-LLMCompressor and Sparse Finetuning config models.
-"""
-
-from typing import Any
-
-from pydantic import BaseModel, Field
-from typing_extensions import Annotated
-
-
-class CompressionArgs(BaseModel):
-    """Sparse Finetuning config for LLMCompressor."""
-
-    # Typing for recipe is set to Any due to:
-    # https://github.com/vllm-project/llm-compressor/issues/1319
-    recipe: Annotated[
-        Any,
-        Field(
-            description="The recipe containing the compression algorithms and hyperparameters to apply."
-        ),
-    ]
-
-    save_compressed: Annotated[
-        bool,
-        Field(
-            default=False,
-            description="Whether to save the compressed model after training.",
-        ),
-    ]
-
-
-class LLMCompressorArgs(BaseModel):
-    """LLMCompressor configuration BaseModel."""
-
-    llmcompressor: Annotated[
-        CompressionArgs,
-        Field(
-            description="Arguments enabling compression pathways through the LLM Compressor plugins"
-        ),
-    ]

src/axolotl/integrations/llm_compressor/plugin.py

@@ -1,171 +0,0 @@
-"""
-Sparse Finetuning plugin for Axolotl — enables handling of sparse neural networks
-by maintaining masks for zero weights during training.
-"""
-
-import logging
-from functools import wraps
-from typing import Any, Callable, Concatenate, ParamSpec, TypeVar
-
-from llmcompressor import active_session, create_session
-from llmcompressor.core import callbacks as session_callbacks
-from llmcompressor.recipe import Recipe
-from torch.nn import Module
-from transformers.trainer import Trainer
-from transformers.trainer_callback import TrainerCallback, TrainerControl, TrainerState
-from transformers.training_args import TrainingArguments
-
-from axolotl.integrations.base import BasePlugin
-
-P = ParamSpec("P")  # Params for generic function signatures
-R = TypeVar("R")  # Return type for generic function signatures
-
-LOG = logging.getLogger("axolotl.integrations.llm_compressor")
-
-
-class LLMCompressorCallbackHandler(TrainerCallback):
-    """
-    Trainer callback for Sparse Finetuning.
-    Maintains sparsity patterns during training by applying masks after optimization steps,
-    ensuring zero-weight updates are canceled out.
-    """
-
-    def __init__(self, trainer: Trainer, recipe: Any):
-        """
-        Initialize the Sparse Finetuning callback handler.
-
-        Args:
-            trainer (Trainer): Huggingface Trainer instance.
-            recipe (Recipe | dict): Sparse finetuning recipe to apply.
-        """
-        super().__init__()
-        self.trainer = trainer
-        self.recipe = (
-            Recipe.model_validate(recipe) if not isinstance(recipe, Recipe) else recipe
-        )
-        self.original_compute_loss = trainer.compute_loss
-        self.trainer.compute_loss = compute_loss_wrapper(self.trainer.compute_loss)
-        create_session()
-
-    def on_train_begin(
-        self,
-        args: TrainingArguments,
-        state: TrainerState,
-        control: TrainerControl,
-        **kwargs,
-    ) -> None:
-        """
-        Called at the beginning of training. Initializes the compression session.
-
-        Args:
-            args (TrainingArguments): Training arguments.
-            state (TrainerState): Trainer state.
-            control (TrainerControl): Trainer control.
-        """
-        super().on_train_begin(args, state, control, **kwargs)
-        self.trainer.accelerator.wait_for_everyone()
-        active_session().initialize(
-            model=self.trainer.model,
-            optimizer=self.trainer.optimizer,
-            start=state.epoch,
-            recipe=self.recipe,
-        )
-        self.trainer.accelerator.wait_for_everyone()
-
-    def on_step_begin(
-        self,
-        args: TrainingArguments,
-        state: TrainerState,
-        control: TrainerControl,
-        **kwargs,
-    ) -> None:
-        """
-        Called at the beginning of a training step. Triggers batch_start callback.
-        """
-        super().on_step_begin(args, state, control, **kwargs)
-        session_callbacks.batch_start()
-
-    def on_step_end(
-        self,
-        args: TrainingArguments,
-        state: TrainerState,
-        control: TrainerControl,
-        **kwargs,
-    ) -> None:
-        """
-        Called at the end of a training step. Triggers optimizer and batch_end callbacks.
-        """
-        super().on_step_end(args, state, control, **kwargs)
-        session_callbacks.optim_pre_step()
-        session_callbacks.optim_post_step()
-        session_callbacks.batch_end()
-
-    def on_train_end(
-        self,
-        args: TrainingArguments,
-        state: TrainerState,
-        control: TrainerControl,
-        **kwargs,
-    ) -> None:
-        """
-        Called at the end of training. Finalizes the compression session.
-        """
-        super().on_train_end(args, state, control, **kwargs)
-        active_session().finalize()
-        self.trainer.compute_loss_func = self.original_compute_loss
-
-
-class LLMCompressorPlugin(BasePlugin):
-    """
-    Sparse Finetuning plugin for Axolotl integration.
-    """
-
-    def get_input_args(self) -> str:
-        """
-        Returns the path to the plugin's argument definition.
-
-        Returns:
-            str: Dotted path to the LLMCompressorArgs class.
-        """
-        return "axolotl.integrations.llm_compressor.args.LLMCompressorArgs"
-
-    def add_callbacks_post_trainer(self, cfg: Any, trainer: Trainer) -> list:
-        """
-        Adds Sparse Finetuning callback to the Trainer instance.
-
-        Args:
-            cfg (Any): Configuration object containing the sparse recipe.
-            trainer (Trainer): Huggingface Trainer instance.
-
-        Returns:
-            list: List containing the configured callback instances.
-        """
-        LOG.info("Adding Sparse Finetuning callback to the trainer")
-        callback = LLMCompressorCallbackHandler(
-            trainer=trainer,
-            recipe=cfg.llmcompressor.recipe,
-        )
-        return [callback]
-
-
-def compute_loss_wrapper(
-    compute_loss_func: Callable[Concatenate[Module, P], R],
-) -> Callable[Concatenate[Module, P], R]:
-    """
-    Wraps the loss computation function to trigger the loss_calculated callback.
-
-    Args:
-        compute_loss_func (Callable): Original loss computation function.
-
-    Returns:
-        Callable: Wrapped function that also invokes the loss_calculated callback.
-    """
-
-    @wraps(compute_loss_func)
-    def compute_and_notify(model: Module, *args: P.args, **kwargs: P.kwargs) -> R:
-        loss = compute_loss_func(model, *args, **kwargs)
-        if active_session().lifecycle.initialized_ and model.training:
-            session_callbacks.loss_calculated(loss=loss)
-        return loss
-
-    return compute_and_notify

src/axolotl/integrations/llm_compressor/utils.py

@@ -1,40 +0,0 @@
-"""Utilities for llmcompressor integration with axolotl."""
-
-from typing import Union
-
-from llmcompressor.transformers.sparsification.compressed_tensors_utils import (
-    modify_save_pretrained,
-)
-from transformers import PreTrainedModel, Trainer
-
-
-def save_compressed_model(
-    model: PreTrainedModel,
-    output_dir: Union[str, bytes],
-    trainer: Trainer,
-    safe_serialization: bool = False,
-    save_compressed: bool = False,
-) -> None:
-    """
-    Synchronize processes, apply compression hooks, and save the model.
-
-    Args:
-        model (PreTrainedModel): The model to be saved.
-        output_dir (str or bytes): Path where the model files will be written.
-        trainer (Trainer): Hugging Face Trainer for process synchronization.
-        safe_serialization (bool): Use safe serialization if True.
-        save_compressed (bool): Write compressed tensors if True.
-    """
-    trainer.accelerator.wait_for_everyone()
-
-    # Only the main process writes the files
-    if not trainer.accelerator.is_main_process:
-        return
-
-    modify_save_pretrained(model)
-    model.save_pretrained(
-        output_dir,
-        safe_serialization=safe_serialization,
-        save_compressed=save_compressed,
-        skip_sparsity_compression_stats=not save_compressed,
-    )

src/axolotl/monkeypatch/attention/ring_attn/__init__.py

@@ -4,6 +4,7 @@
 # flake8: noqa
 
 from .patch import (
+    RingAttnFunc,
     get_ring_attn_group,
     register_ring_attn,
     set_ring_attn_group,

src/axolotl/monkeypatch/attention/ring_attn/adapters/batch.py

@@ -16,7 +16,11 @@ import torch
 import torch.distributed as dist
 import transformers
 import transformers.modeling_flash_attention_utils
-from ring_flash_attn import ring_flash_attn_func
+from ring_flash_attn import (
+    ring_flash_attn_func,
+    stripe_flash_attn_func,
+    zigzag_ring_flash_attn_func,
+)
 from ring_flash_attn.adapters.hf_adapter import check_params
 from transformers.modeling_flash_attention_utils import (
     _flash_supports_window_size,
@@ -24,12 +28,12 @@ from transformers.modeling_flash_attention_utils import (
 )
 from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
 
-from axolotl.utils.schemas.enums import RingAttnFunc
+from axolotl.monkeypatch.attention.ring_attn.patch import RingAttnFunc
 
 RING_ATTN_FUNC_MAPPING = {
-    RingAttnFunc.BATCH_RING: torch.compile(ring_flash_attn_func),
+    RingAttnFunc.BATCH_RING: ring_flash_attn_func,
-    # RingAttnFunc.BATCH_ZIGZAG: torch.compile(zigzag_ring_flash_attn_func),
+    RingAttnFunc.BATCH_ZIGZAG: zigzag_ring_flash_attn_func,
-    # RingAttnFunc.BATCH_STRIPE: torch.compile(stripe_flash_attn_func),
+    RingAttnFunc.BATCH_STRIPE: stripe_flash_attn_func,
 }
 
 

src/axolotl/monkeypatch/attention/ring_attn/patch.py

@@ -6,12 +6,13 @@ package, specifically the `hf_adapter.substitute_hf_flash_attn` function to patc
 their sequence parallel version of Flash Attention 2.
 """
 
+from enum import Enum
+
 import torch
 import torch.distributed as dist
 from accelerate.logging import get_logger
 
 from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
-from axolotl.utils.schemas.enums import RingAttnFunc
 
 LOG = get_logger(__name__)
 
@@ -40,6 +41,17 @@ def set_ring_attn_group(ring_attn_group: dist.ProcessGroup | None):
     RING_ATTN_GROUP = ring_attn_group
 
 
+class RingAttnFunc(str, Enum):
+    """Enum class for supported `ring-flash-attn` implementations"""
+
+    # VARLEN_RING = "varlen_ring"
+    # VARLEN_ZIGZAG = "varlen_zigzag"
+    VARLEN_LLAMA3 = "varlen_llama3"
+    BATCH_RING = "batch_ring"
+    BATCH_ZIGZAG = "batch_zigzag"
+    BATCH_STRIPE = "batch_stripe"
+
+
 def register_ring_attn(
     sequence_parallel_degree: int,
     heads_k_stride: int | None,
@@ -105,7 +117,11 @@ def register_ring_attn(
         substitute_hf_flash_attn(
             process_group=get_ring_attn_group(), heads_k_stride=heads_k_stride or 1
         )
-    elif ring_attn_func is RingAttnFunc.BATCH_RING:
+    elif ring_attn_func in [
+        RingAttnFunc.BATCH_RING,
+        RingAttnFunc.BATCH_ZIGZAG,
+        RingAttnFunc.BATCH_STRIPE,
+    ]:
         from axolotl.monkeypatch.attention.ring_attn.adapters.batch import (
             substitute_hf_flash_attn,
         )

src/axolotl/train.py

@@ -7,7 +7,7 @@ import os
 import signal
 import sys
 import weakref
-from contextlib import ExitStack
+from contextlib import nullcontext
 from pathlib import Path
 from typing import Any, Dict
 
@@ -27,13 +27,14 @@ from axolotl.contribs.lgpl import (  # pylint: disable = no-name-in-module
     fix_untrained_tokens,
 )
 from axolotl.core.trainer_builder import HFCausalTrainerBuilder, HFRLTrainerBuilder
+from axolotl.core.trainers.mixins.sequence_parallel import (
+    SequenceParallelContextManager,
+)
 from axolotl.integrations.base import PluginManager
-from axolotl.utils.ctx_managers.sequence_parallel import SequenceParallelContextManager
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import cleanup_distributed
 from axolotl.utils.freeze import freeze_layers_except
 from axolotl.utils.models import load_model, load_processor, load_tokenizer
-from axolotl.utils.schemas.enums import RLType
 from axolotl.utils.trainer import setup_trainer
 
 try:
@@ -106,7 +107,7 @@ def setup_reference_model(
         Reference model if needed for RL training, `None` otherwise.
     """
     model_ref = None
-    if cfg.rl and cfg.rl != RLType.ORPO:
+    if cfg.rl and cfg.rl != "orpo":
         if cfg.adapter and not cfg.rl_adapter_ref_model:
             # use built-in trl autounwrap
             LOG.debug("Passing model_ref: None to RL trainer")
@@ -187,32 +188,28 @@ def execute_training(
         trainer: The configured trainer object.
         resume_from_checkpoint: Path to checkpoint to resume from, if applicable.
     """
-    with ExitStack() as stack:
+    # Define the context managers to use
-        # Define the context managers to use
+    flash_context = (
-        if cfg.flash_optimum:
+        torch.backends.cuda.sdp_kernel(
-            stack.enter_context(
+            enable_flash=True,
-                torch.backends.cuda.sdp_kernel(
+            enable_math=True,
-                    enable_flash=True,
+            enable_mem_efficient=True,
-                    enable_math=True,
+        )
-                    enable_mem_efficient=True,
+        if cfg.flash_optimum
-                )
+        else nullcontext()
-            )
+    )
+    sequence_parallel_context = (
+        SequenceParallelContextManager(
+            model=trainer.model,
+            sequence_parallel_degree=cfg.sequence_parallel_degree,
+            ring_attn_func=cfg.ring_attn_func,
+        )
+        if cfg.sequence_parallel_degree > 1
+        else nullcontext()
+    )
 
-        if cfg.sequence_parallel_degree > 1:
+    LOG.info("Starting trainer...")
-            models = [trainer.model]
+    with flash_context, sequence_parallel_context:
-            if hasattr(trainer, "ref_model"):
-                models.append(trainer.ref_model)
-
-            stack.enter_context(
-                SequenceParallelContextManager(
-                    models=models,
-                    sequence_parallel_degree=cfg.sequence_parallel_degree,
-                    gradient_accumulation_steps=cfg.gradient_accumulation_steps,
-                    ring_attn_func=cfg.ring_attn_func,
-                )
-            )
-
-        LOG.info("Starting trainer...")
         trainer.train(resume_from_checkpoint=resume_from_checkpoint)
 
 
@@ -297,23 +294,8 @@ def save_trained_model(
             trainer.model.save_pretrained(
                 cfg.output_dir, safe_serialization=safe_serialization
             )
-
         model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
 
-    if hasattr(cfg, "llmcompressor") and cfg.llmcompressor:
-        # TODO: add integration support so this can be implemented completely within the plugin
-        from axolotl.integrations.llm_compressor.utils import (
-            save_compressed_model,
-        )
-
-        save_compressed_model(
-            model=model,
-            output_dir=cfg.output_dir,
-            trainer=trainer,
-            safe_serialization=safe_serialization,
-            save_compressed=cfg.llmcompressor.save_compressed,
-        )
-
 
 def create_model_card(cfg: DictDefault, trainer: Trainer):
     """

src/axolotl/utils/ctx_managers/__init__.py

@@ -1,6 +0,0 @@
-"""Init for context manager submodule"""
-
-# pylint: disable=unused-import
-# flake8: noqa
-
-from .sequence_parallel import SequenceParallelContextManager

src/axolotl/utils/ctx_managers/sequence_parallel.py

@@ -1,335 +0,0 @@
-"""Module for Axolotl trainer sequence parallelism manager and utilities"""
-
-import functools
-
-import torch
-import torch.distributed as dist
-from torch import nn
-from torch.utils.hooks import RemovableHandle
-from transformers.modeling_outputs import CausalLMOutputWithPast
-from transformers.utils import ModelOutput
-
-from axolotl.monkeypatch.attention.ring_attn.patch import (
-    get_ring_attn_group,
-    update_ring_attn_params,
-)
-from axolotl.utils.schemas.enums import RingAttnFunc
-
-
-# TODO(djsaunde): implement zigzag, stripe patterns here (and elsewhere) in this
-# module. Currently, we just focus on batch ring and varlen llama3 for simplicity.
-def apply_sequence_parallelism(
-    batch: dict[str, torch.Tensor],
-    local_rank: int,
-    local_world_size: int,
-    gradient_accumulation_steps: int,
-    ring_attn_func: RingAttnFunc,  # pylint: disable=unused-argument
-) -> tuple[dict[str, torch.Tensor], int, int]:
-    """
-    Apply sequence parallelism slicing to a batch.
-
-    Special handling is implemented for integer logits_to_keep, which indicates
-    to only keep the last N tokens in the sequence during generation.
-
-    Args:
-        batch: Batch dictionary (e.g., input_ids, attention_mask, etc.).
-        local_rank: Local rank in the sequence parallel group.
-        local_world_size: World size of the sequence parallel group.
-        gradient_accumulation_steps: Number of steps to accumulate gradients over.
-        ring_attn_func: Which ring attention function to use. Currently unused, but
-            related to above TODO.
-
-    Returns:
-        tuple of:
-            - Batch dictionary with sliced tensors.
-            - The original sequence length before padding.
-            - The number of padding tokens added.
-    """
-    original_seq_len = batch["input_ids"].size(1)
-
-    # Update ring attention params if needed
-    if batch.get("position_ids") is not None:
-        update_ring_attn_params(position_ids=batch["position_ids"])
-    else:
-        # If position_ids aren't already in the batch, create them
-        batch["position_ids"] = torch.arange(
-            0,
-            original_seq_len,
-            dtype=torch.long,
-            device=batch["input_ids"].device,
-        ).expand(batch["input_ids"].size(0), -1)
-
-    if "logits_to_keep" in batch and isinstance(batch["logits_to_keep"], int):
-        logits_to_keep = batch["logits_to_keep"]
-
-        # Calculate which positions in the full sequence contain the last N tokens
-        start_position = max(0, original_seq_len - logits_to_keep)
-        chunk_size = original_seq_len // local_world_size
-        rank_start = local_rank * chunk_size
-        rank_end = rank_start + chunk_size
-
-        # Create a boolean mask tensor for this rank's chunk
-        mask = torch.zeros(
-            chunk_size,
-            dtype=torch.bool,
-            device=batch["input_ids"].device,
-        )
-
-        if rank_end > start_position:
-            # Calculate how many of the last N tokens fall within this rank's range
-            tokens_in_rank = min(rank_end, original_seq_len) - max(
-                rank_start, start_position
-            )
-
-            # Calculate where these tokens start in the local chunk
-            local_start_idx = max(0, start_position - rank_start)
-
-            # Set the appropriate positions in the mask to True
-            mask[local_start_idx : local_start_idx + tokens_in_rank] = True
-
-        # Replace the integer with the boolean mask
-        batch["logits_to_keep"] = mask
-
-    # Add padding to make sequence length divisible by local_world_size
-    total_seq_len = original_seq_len
-    pad_len = 0
-    divisor = min(local_world_size, 64)
-    if total_seq_len % divisor != 0:
-        pad_len = divisor - (total_seq_len % divisor)
-
-        # Apply padding to all relevant tensors
-        for key in batch:
-            if (
-                isinstance(batch[key], torch.Tensor)
-                and batch[key].dim() > 1
-                and batch[key].size(1) == total_seq_len
-            ):
-                # Create padding tensor
-                pad_value = -100 if key == "labels" else 0
-                padding = torch.full(
-                    (batch[key].size(0), pad_len, *batch[key].shape[2:]),
-                    pad_value,
-                    dtype=batch[key].dtype,
-                    device=batch[key].device,
-                )
-
-                # Concatenate padding to the right side of the tensor
-                batch[key] = torch.cat([batch[key], padding], dim=1)
-            if key == "logits_to_keep":
-                # Create padding tensor
-                padding = torch.ones(
-                    1,
-                    dtype=batch[key].dtype,
-                    device=batch[key].device,
-                )
-
-                # Concatenate padding to the right side of the tensor
-                batch[key] = torch.cat([batch[key], padding], dim=0)
-
-        # Update the total sequence length after padding
-        total_seq_len = batch["input_ids"].size(1)
-
-    # Slice batch for sequence parallel
-    for key in batch:
-        if not isinstance(batch[key], torch.Tensor) or batch[key].dim() <= 1:
-            continue
-
-        # Split in sequential fashion and grab this rank's chunk
-        if batch[key].size(1) == total_seq_len:
-            batch[key] = (
-                batch[key].chunk(local_world_size, dim=1)[local_rank].contiguous()
-            )
-        elif key == "logits_to_keep":
-            batch[key] = (
-                batch[key].chunk(local_world_size, dim=0)[local_rank].contiguous()
-            )
-
-        # Handle num_items_in_batch
-        if "num_items_in_batch" in batch:
-            # Approximation; this needed since num_items_in_batch may be counted across
-            # all samples in a gradient accumulated batch, not on a per-step basis.
-            batch["num_items_in_batch"] = (
-                batch["labels"] != -100
-            ).sum() * gradient_accumulation_steps
-
-    return batch, original_seq_len, pad_len
-
-
-class SequenceParallelContextManager:
-    """Context manager for sequence parallelism operations.
-
-    This class provides a context that will automatically apply sequence parallelism
-    during model forward passes using a pre-forward hook, and gather outputs from
-    across the sequence parallelism group using a post-forward hook.
-
-    Args:
-        models: List of models to apply sequence parallelism to pre- and post- forward
-            hooks.
-        sequence_parallel_degree: Number of processes to split sequences over.
-        gradient_accumulation_steps: Number of steps to accumulate gradients over.
-        ring_attn_func: Which ring attention function to use. Currently unused.
-    """
-
-    def __init__(
-        self,
-        models: list[nn.Module],
-        sequence_parallel_degree: int,
-        gradient_accumulation_steps: int,
-        ring_attn_func: RingAttnFunc,
-    ):
-        self.models = models
-        self.sequence_parallel_degree = sequence_parallel_degree
-        self.gradient_accumulation_steps = gradient_accumulation_steps
-        self.ring_attn_func = ring_attn_func
-        self.process_group = get_ring_attn_group()
-
-        # Initialize sequence parallel group details
-        self.local_rank = dist.get_rank(self.process_group)
-        self.local_world_size = dist.get_world_size(self.process_group)
-
-        # Will store hook handles for removal
-        self.hook_handles: list[RemovableHandle] = []
-
-        # Store original sequence length and padding information
-        self.original_seq_len = 0
-        self.pad_len = 0
-
-        # Create a partially applied version of the apply_sequence_parallelism function
-        self.apply_sequence_parallelism = functools.partial(
-            apply_sequence_parallelism,
-            local_rank=self.local_rank,
-            local_world_size=self.local_world_size,
-            gradient_accumulation_steps=self.gradient_accumulation_steps,
-            ring_attn_func=self.ring_attn_func,
-        )
-
-    def __enter__(self):
-        # Forward pre-hook to apply sequence parallelism
-        def sequence_parallel_pre_hook(_, args, kwargs):
-            # Apply sequence parallelism to kwargs and get original sequence length and padding info
-            kwargs, self.original_seq_len, self.pad_len = (
-                self.apply_sequence_parallelism(batch=kwargs)
-            )
-
-            return args, kwargs
-
-        # Forward post-hook to gather outputs
-        def sequence_parallel_post_hook(_, __, output: ModelOutput) -> ModelOutput:
-            # Gather the sharded outputs
-            output = self.gather_outputs(output)
-
-            # Remove padding if it was added
-            if self.pad_len > 0:
-                for key, value in output.items():
-                    if isinstance(value, torch.Tensor) and value.dim() > 1:
-                        if value.size(1) == self.original_seq_len + self.pad_len:
-                            # Slice to remove padding
-                            output[key] = value[:, : self.original_seq_len].contiguous()
-
-            return output
-
-        # Register both hooks
-        for model in self.models:
-            self.hook_handles.append(
-                model.register_forward_pre_hook(
-                    sequence_parallel_pre_hook, with_kwargs=True
-                )
-            )
-            self.hook_handles.append(
-                model.register_forward_hook(sequence_parallel_post_hook)
-            )
-
-        return self
-
-    def __exit__(self, exc_type, exc_val, exc_tb):
-        # Remove all hooks
-        for handle in self.hook_handles:
-            handle.remove()
-        self.hook_handles = []
-
-    def gather_outputs(self, output: CausalLMOutputWithPast) -> CausalLMOutputWithPast:
-        """Gather sharded outputs from all ranks and reconstruct the full tensor."""
-        for key, value in output.items():
-            if isinstance(value, torch.Tensor) and value.dim() > 1:
-                output[key] = AllGatherWithGrad.apply(value, self.process_group)
-
-        return output
-
-
-class AllGatherWithGrad(torch.autograd.Function):
-    """Custom autograd function for all-gather to preserve gradients."""
-
-    @staticmethod
-    def forward(
-        ctx: torch.autograd.function.FunctionCtx,
-        input_tensor: torch.Tensor,
-        group: dist.ProcessGroup,
-    ) -> torch.Tensor:
-        """
-        Forward pass of all-gather of data with sequence dimension.
-
-        Args:
-            ctx: `torch.autograd` function context.
-            input_tensor: Tensor from model output with sequence dimension.
-            group: `torch.distributed` process group.
-
-        Returns:
-            Tensor from gathering the `input_tensor` from across the process group and
-                concatenating along the sequence dimension.
-        """
-        ctx.group = group
-        ctx.rank = dist.get_rank(group)
-        world_size = dist.get_world_size(group)
-
-        # Gather shape metadata
-        local_shape = torch.tensor(list(input_tensor.shape), device=input_tensor.device)
-        all_shapes = [torch.zeros_like(local_shape) for _ in range(world_size)]
-        dist.all_gather(all_shapes, local_shape, group=group)
-
-        # Store sequence lengths for backward pass
-        seq_lens = [int(shape[1].item()) for shape in all_shapes]
-        ctx.seq_lens = seq_lens
-
-        # Perform all_gather operation
-        gathered = [
-            torch.zeros(
-                tuple(shape.tolist()),
-                dtype=input_tensor.dtype,
-                device=input_tensor.device,
-            )
-            for shape in all_shapes
-        ]
-        dist.all_gather(gathered, input_tensor, group=group)
-
-        # Concatenate tensors along sequence dimension
-        result = torch.cat(gathered, dim=1)
-
-        return result
-
-    @staticmethod
-    def backward(
-        ctx: torch.autograd.function.FunctionCtx, grad_output: torch.Tensor
-    ) -> tuple[torch.Tensor, None]:
-        """
-        Backward pass for all-gather operation.
-
-        Extracts the gradient slice corresponding to this rank's original input
-        from the full gradient tensor.
-
-        Args:
-            ctx: `torch.autograd` function context.
-            grad_output: Gradient from subsequent layers with respect to the
-                concatenated output tensor.
-
-        Returns:
-            Tuple containing the gradient slice for this rank's input tensor and `None`
-                for the process group parameter which doesn't require gradients.
-        """
-        rank = ctx.rank
-        seq_lens = ctx.seq_lens
-
-        # Extract gradient for this rank's chunk
-        offset = sum(seq_lens[:rank])
-        grad_slice = grad_output[:, offset : offset + seq_lens[rank]].contiguous()
-
-        return grad_slice, None

src/axolotl/utils/data/rl.py

@@ -18,9 +18,8 @@ from axolotl.utils.data.utils import deduplicate_and_log_datasets, md5
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import is_main_process, zero_first
 from axolotl.utils.models import load_tokenizer
-from axolotl.utils.schemas.enums import RLType
 
-LOG = logging.getLogger(__name__)
+LOG = logging.getLogger("axolotl")
 
 
 def _get_path(ds_hash, cfg):
@@ -81,7 +80,7 @@ def map_dataset(cfg, data_set, ds_transform_fn, tokenizer, **map_kwargs):
 def drop_long_rl_seq(
     sample, rl, tokenizer, sequence_len  # pylint: disable=invalid-name
 ):
-    if rl in (RLType.DPO, RLType.IPO, RLType.ORPO, RLType.SIMPO):
+    if rl in ("dpo", "ipo", "orpo", "simpo"):
         if not (
             sample.get("prompt") and sample.get("chosen") and sample.get("rejected")
         ):
@@ -101,7 +100,7 @@ def drop_long_rl_seq(
             len_prompt + len_rejected
         ) <= sequence_len
 
-    if rl is RLType.KTO:
+    if rl == "kto":
         if not (sample.get("prompt") and sample.get("completion")):
             raise ValueError("Prompt and completion keys are required for KTO datasets")
 
@@ -115,7 +114,7 @@ def drop_long_rl_seq(
 
         return (len_prompt + len_completion) <= sequence_len
 
-    if rl is RLType.GRPO:
+    if rl == "grpo":
         return True
 
     raise ValueError("Unknown RL type")
@@ -138,9 +137,9 @@ def load_prepare_preference_datasets(cfg):
             if _type:
                 if isinstance(_type, DictDefault):
                     _type = "user_defined.default"
-                if _cfg.rl is RLType.ORPO:
+                if _cfg.rl == "orpo":
                     ds_transform_fn = load_orpo(_type, _cfg, dataset_idx=i)
-                elif _cfg.rl is RLType.KTO:
+                elif _cfg.rl == "kto":
                     ds_transform_fn = load_kto(_type, _cfg, dataset_idx=i)
                 else:
                     ds_transform_fn = load_dpo(_type, _cfg, dataset_idx=i)
@@ -151,7 +150,7 @@ def load_prepare_preference_datasets(cfg):
                 split_datasets[i] = map_dataset(
                     cfg, data_set, ds_transform_fn, tokenizer, **map_kwargs
                 )
-            elif _cfg.rl is RLType.KTO:
+            elif _cfg.rl == "kto":
                 ds_transform_fn = load_kto(_type, _cfg, dataset_idx=i)
                 map_kwargs = {}
                 if isinstance(ds_transform_fn, tuple):
@@ -186,7 +185,7 @@ def load_prepare_preference_datasets(cfg):
                     )
 
         combined_datasets = concatenate_datasets(split_datasets)
-        combined_datasets = combined_datasets.shuffle(seed=cfg.seed or 42)
+        combined_datasets = combined_datasets.shuffle(seed=cfg.seed)
 
         return combined_datasets
 
@@ -206,8 +205,6 @@ def load_prepare_preference_datasets(cfg):
                 eval_dataset = load_split(cfg.test_datasets, cfg)
         if not eval_dataset:
             if cfg.val_set_size:
-                seed = cfg.seed if cfg.seed is not None else 42
-
                 # ensure we end up with the same fingerprint by doing rank0 first and being able to cache
                 to_hash_train = (
                     train_dataset._fingerprint  # pylint: disable=protected-access
@@ -216,7 +213,7 @@ def load_prepare_preference_datasets(cfg):
                     + "|"
                     + "train"
                     + "|"
-                    + str(seed)
+                    + str(cfg.seed or 42)
                 )
                 to_hash_test = (
                     train_dataset._fingerprint  # pylint: disable=protected-access
@@ -225,13 +222,13 @@ def load_prepare_preference_datasets(cfg):
                     + "|"
                     + "test"
                     + "|"
-                    + str(seed)
+                    + str(cfg.seed or 42)
                 )
                 train_fingerprint = md5(to_hash_train)
                 test_fingerprint = md5(to_hash_test)
                 ds_w_test_split = train_dataset.train_test_split(
                     test_size=cfg.val_set_size,
-                    seed=seed,
+                    seed=cfg.seed,
                     shuffle=False,
                     train_new_fingerprint=train_fingerprint,
                     test_new_fingerprint=test_fingerprint,

src/axolotl/utils/data/sft.py

@@ -148,7 +148,7 @@ def prepare_dataset(cfg, tokenizer, processor=None, preprocess_iterable=None):
             ds_wrapper_partial,
             max_tokens=cfg.sequence_len,
             batch_size=cfg.micro_batch_size,
-            seed=cfg.seed if cfg.seed is not None else 42,
+            seed=cfg.seed or 42,
             buffer_size=cfg.pretrain_multipack_buffer_size or 10_000,
         )
         # https://discuss.huggingface.co/t/how-to-use-huggingface-trainer-streaming-datasets-without-wrapping-it-with-torchdatas-iterablewrapper/25230
@@ -416,8 +416,6 @@ def load_prepare_datasets(
     )
 
     if split == "train" and val_set_size:
-        seed = cfg.seed if cfg.seed is not None else 42
-
         # ensure we end up with the same fingerprint by doing rank0 first and being able to cache
         to_hash_train = (
             dataset._fingerprint  # pylint: disable=protected-access
@@ -426,7 +424,7 @@ def load_prepare_datasets(
             + "|"
             + "train"
             + "|"
-            + str(seed)
+            + str(cfg.seed or 42)
         )
         to_hash_test = (
             dataset._fingerprint  # pylint: disable=protected-access
@@ -435,7 +433,7 @@ def load_prepare_datasets(
             + "|"
             + "test"
             + "|"
-            + str(seed)
+            + str(cfg.seed or 42)
         )
         train_fingerprint = md5(to_hash_train)
         test_fingerprint = md5(to_hash_test)
@@ -444,7 +442,7 @@ def load_prepare_datasets(
         dataset = dataset.train_test_split(
             test_size=val_set_size,
             shuffle=False,
-            seed=seed,
+            seed=cfg.seed or 42,
             train_new_fingerprint=train_fingerprint,
             test_new_fingerprint=test_fingerprint,
         )

src/axolotl/utils/gradient_checkpointing/__init__.py

@@ -5,8 +5,11 @@ from functools import partial
 
 from packaging import version
 
-from axolotl.utils.gradient_checkpointing.unsloth import (
+from axolotl.utils.gradient_checkpointing.offload_cpu import (
-    Unsloth_Offloaded_Gradient_Checkpointer,
+    CPU_Offloaded_Gradient_Checkpointer,
+)
+from axolotl.utils.gradient_checkpointing.offload_disk import (
+    Disco,
 )
 
 transformers_version = version.parse(importlib.metadata.version("transformers"))
@@ -26,12 +29,31 @@ def hf_grad_checkpoint_offload_wrapper(
     decoder_layer, *args, use_reentrant=None
 ):  # pylint: disable=unused-argument
     if uses_gc_layers(decoder_layer):
-        return Unsloth_Offloaded_Gradient_Checkpointer.apply(
+        return CPU_Offloaded_Gradient_Checkpointer.apply(
             decoder_layer,
             *args,
         )
 
-    return Unsloth_Offloaded_Gradient_Checkpointer.apply(
+    return CPU_Offloaded_Gradient_Checkpointer.apply(
+        (
+            decoder_layer.func.__self__
+            if isinstance(decoder_layer, partial)
+            else decoder_layer.__self__
+        ),
+        *args,
+    )
+
+
+def hf_grad_checkpoint_disk_offload_wrapper(
+    decoder_layer, *args, use_reentrant=None
+):  # pylint: disable=unused-argument
+    if uses_gc_layers(decoder_layer):
+        return Disco.apply(
+            decoder_layer,
+            *args,
+        )
+
+    return Disco.apply(
         (
             decoder_layer.func.__self__
             if isinstance(decoder_layer, partial)

src/axolotl/utils/gradient_checkpointing/offload_cpu.py

@@ -1,4 +1,4 @@
-"""Unsloth checkpointing"""
+"""CPU offloaded checkpointing"""
 
 # Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
 #
@@ -26,7 +26,7 @@ else:
     torch_cuda_amp_custom_bwd = torch.amp.custom_bwd(device_type="cuda")
 
 
-class Unsloth_Offloaded_Gradient_Checkpointer(  # pylint: disable=invalid-name
+class CPU_Offloaded_Gradient_Checkpointer(  # pylint: disable=invalid-name
     torch.autograd.Function
 ):
     """

src/axolotl/utils/gradient_checkpointing/offload_disk.py

@@ -0,0 +1,531 @@
+"""
+DISCO - DIsk-based Storage and Checkpointing with Optimized prefetching
+"""
+
+# Copyright 2025 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.
+
+import atexit
+import concurrent.futures
+import logging
+import os
+import queue
+import shutil
+import tempfile
+import threading
+import time
+import uuid
+from collections import deque
+from concurrent.futures import Future
+from typing import Dict
+
+import torch
+
+torch_cuda_amp_custom_fwd = torch.amp.custom_fwd(device_type="cuda")
+torch_cuda_amp_custom_bwd = torch.amp.custom_bwd(device_type="cuda")
+
+# Setup logger
+logger = logging.getLogger(__name__)
+
+
+class DiskOffloadManager:
+    """
+    Manages offloaded tensors and handles prefetching in a separate thread.
+    Includes synchronization to prevent race conditions.
+    """
+
+    def __init__(
+        self,
+        prefetch_size: int = 3,
+        prefetch_to_gpu: bool = True,
+        save_workers: int = 4,
+    ):
+        """
+        Args:
+            prefetch_size: Maximum number of tensors to prefetch in the background.
+            prefetch_to_gpu: Whether to prefetch tensors directly to GPU memory.
+            save_workers: Maximum number of concurrent save operations.
+        """
+        self.temp_dir = tempfile.mkdtemp(prefix="disco_")
+
+        # Track tensor paths and their status
+        self.tensor_paths: deque = deque()  # Ordered history of tensor paths (LIFO)
+        self.file_locks: Dict[str, threading.Lock] = (
+            {}
+        )  # Maps file_path -> threading.Lock()
+        # Maps file_path -> status ("saving", "ready", "prefetching", "loaded", "deleted")
+        self.file_status: Dict[str, str] = {}
+
+        self.max_prefetch = prefetch_size
+        self.prefetch_to_gpu = prefetch_to_gpu
+
+        # Thread synchronization
+        self.manager_lock = threading.RLock()  # Used for thread-safe operations
+
+        # Prefetch queue and cache
+        self.prefetch_queue: queue.Queue = queue.Queue()
+        self.prefetch_cache: Dict[str, torch.Tensor] = {}  # Maps file_path -> tensor
+
+        # Save queue and thread pool
+        self.save_queue: queue.Queue = queue.Queue()
+        self.save_pool = concurrent.futures.ThreadPoolExecutor(max_workers=save_workers)
+        self.save_futures: Dict[str, Future] = {}
+        self.save_semaphore = threading.Semaphore(
+            save_workers * 2
+        )  # Limit concurrent save operations
+
+        # Start prefetch worker thread
+        self.stop_event = threading.Event()
+        # start multiple threads for prefetching
+        self.prefetch_worker_count = 2
+        self.prefetch_workers = []
+        for _ in range(self.prefetch_worker_count):
+            worker = threading.Thread(target=self._prefetch_worker, daemon=True)
+            worker.start()
+            self.prefetch_workers.append(worker)
+
+        # Start save worker thread
+        self.save_worker = threading.Thread(target=self._save_worker, daemon=True)
+        self.save_worker.start()
+        self.idx = 0
+
+        atexit.register(self.cleanup)
+
+    def _save_worker(self):
+        """Background thread that processes the save queue"""
+        while not self.stop_event.is_set():
+            try:
+                save_item = self.save_queue.get(timeout=0.5)
+                if save_item is None:
+                    continue
+
+                tensor, file_path = save_item
+
+                # Submit the save task to the thread pool
+                future = self.save_pool.submit(
+                    self._save_tensor_to_disk, tensor, file_path
+                )
+                with self.manager_lock:
+                    self.save_futures[file_path] = future
+
+                self.save_queue.task_done()
+
+            except queue.Empty:
+                time.sleep(0.01)  # Small sleep to prevent CPU spinning
+                continue
+
+    def _save_tensor_to_disk(self, tensor: torch.Tensor, file_path: str):
+        """Actually save the tensor to disk"""
+        try:
+            # Save tensor to disk
+            cpu_tensor = tensor.detach().cpu()
+            torch.save(cpu_tensor, file_path)
+            del cpu_tensor
+
+            with self.manager_lock:
+                # Mark file as ready
+                self.file_status[file_path] = "ready"
+
+            # Release semaphore
+            self.save_semaphore.release()
+
+            return True
+        except FileNotFoundError as e:
+            logger.error(f"Error saving tensor to {file_path}: {e}")
+            with self.manager_lock:
+                self.file_status[file_path] = "error"
+
+            # Release semaphore
+            self.save_semaphore.release()
+
+            return False
+
+    def _prefetch_worker(self):
+        """Background thread that loads tensors from disk ahead of time"""
+        while not self.stop_event.is_set():
+            try:
+                file_path = self.prefetch_queue.get(timeout=0.5)
+                if file_path is None:
+                    continue
+
+                # Check if file is available and not already in cache
+                with self.manager_lock:
+                    if (
+                        file_path not in self.file_status
+                        or self.file_status[file_path] == "deleted"
+                    ):
+                        self.prefetch_queue.task_done()
+                    if file_path in self.prefetch_cache:
+                        self.prefetch_queue.task_done()
+                        continue
+
+                    # If file is still being saved, wait for it
+                    if (
+                        self.file_status[file_path] == "saving"
+                        and file_path in self.save_futures
+                    ):
+                        # Re-queue this prefetch request with a little delay
+                        self.prefetch_queue.task_done()
+                        time.sleep(0.1)
+                        self.prefetch_queue.put(file_path)
+                        continue
+
+                    # Mark file as being prefetched
+                    self.file_status[file_path] = "prefetching"
+
+                # Load tensor from disk and store in cache
+                try:
+                    if os.path.exists(file_path):
+                        if self.prefetch_to_gpu:
+                            tensor = torch.load(
+                                file_path,
+                                map_location=torch.device("cuda"),
+                                weights_only=True,
+                            )
+                        else:
+                            tensor = torch.load(file_path, weights_only=True)
+
+                        with self.manager_lock:
+                            self.prefetch_cache[file_path] = tensor
+                            self.file_status[file_path] = "ready"
+                    else:
+                        with self.manager_lock:
+                            if self.file_status.get(file_path) != "deleted":
+                                logger.warning(
+                                    f"Prefetch error: File not found {file_path}"
+                                )
+                                self.file_status[file_path] = "missing"
+
+                except FileNotFoundError as e:
+                    with self.manager_lock:
+                        if self.file_status.get(file_path) != "deleted":
+                            logger.warning(f"Prefetch error for {file_path}: {e}")
+                            self.file_status[file_path] = "error"
+
+                self.prefetch_queue.task_done()
+
+            except queue.Empty:
+                time.sleep(0.01)  # Small sleep to prevent CPU spinning
+                continue
+
+    def save_tensor(self, tensor: torch.Tensor):
+        """Save tensor to disk asynchronously and return file path with thread-safe operations"""
+        # Generate unique file path
+        self.idx += 1
+        file_path: str = os.path.join(
+            self.temp_dir, f"{self.idx:06d}-{uuid.uuid4()}.pt"
+        )
+
+        with self.manager_lock:
+            # Mark file as being saved
+            self.file_locks[file_path] = threading.Lock()
+            self.file_status[file_path] = "saving"
+            # Add to history
+            self.tensor_paths.append(file_path)
+
+        # Acquire semaphore to limit concurrent save operations
+        self.save_semaphore.acquire()  # pylint: disable=consider-using-with
+        # Queue tensor for saving in background
+        self.save_queue.put((tensor.detach(), file_path))
+
+        return file_path
+
+    def wait_for_save(self, file_path, timeout=None) -> None:
+        """Wait for a tensor to be saved to disk"""
+        start_time = time.time()
+        while timeout is None or time.time() - start_time < timeout:
+            with self.manager_lock:
+                if self.file_status.get(file_path) == "ready":
+                    return
+                if self.file_status.get(file_path) in ["error", "missing", "deleted"]:
+                    return
+
+                if file_path in self.save_futures:
+                    future = self.save_futures[file_path]
+                    if future.done():
+                        return
+
+            # Small sleep to prevent CPU spinning
+            time.sleep(0.01)
+
+        # Timeout
+        logger.warning(f"Timeout waiting for tensor to be saved: {file_path}")
+        return
+
+    def load_tensor(self, file_path, target_device="cuda"):
+        """Load tensor from disk or prefetch cache with proper synchronization"""
+        # Wait for tensor to be saved if it's still in progress
+        self.wait_for_save(file_path)
+
+        tensor = None
+
+        # Try to get from cache first
+        with self.manager_lock:
+            # Check if tensor is already in cache
+            if file_path in self.prefetch_cache:
+                tensor = self.prefetch_cache[file_path]
+                del self.prefetch_cache[file_path]
+                self.file_status[file_path] = "loaded"
+
+        if tensor is not None:
+            # Ensure tensor is on correct device
+            if target_device != "cpu" and tensor.device.type == "cpu":
+                tensor = tensor.to(target_device, non_blocking=True)
+            return tensor
+
+        # If not in cache, load directly from disk
+        try:
+            if not os.path.exists(file_path):
+                logger.error(f"File not found for loading: {file_path}")
+                raise FileNotFoundError(f"File not found: {file_path}")
+
+            tensor = torch.load(file_path, weights_only=True)
+
+            with self.manager_lock:
+                self.file_status[file_path] = "loaded"
+
+            if target_device != "cpu":
+                tensor = tensor.to(target_device, non_blocking=True)
+
+            return tensor
+
+        except Exception as e:
+            logger.error(f"Error loading tensor from {file_path}: {e}")
+            raise
+
+    def _safe_delete_file(self, file_path):
+        """Safely delete a file with proper synchronization"""
+        with self.manager_lock:
+            # Make sure any save operation is completed
+            if file_path in self.save_futures:
+                future = self.save_futures[file_path]
+                try:
+                    if not future.done():
+                        future.cancel()
+                    del self.save_futures[file_path]
+                except FileNotFoundError as e:
+                    logger.warning(
+                        f"Error canceling save operation for {file_path}: {e}"
+                    )
+
+            # Only delete if file exists and is not being prefetched
+            status = self.file_status.get(file_path)
+            if status in ["ready", "loaded", "error", "missing"]:
+                try:
+                    if os.path.exists(file_path):
+                        os.remove(file_path)
+                    self.file_status[file_path] = "deleted"
+                    return True
+                except FileNotFoundError as e:
+                    logger.warning(f"Error deleting file {file_path}: {e}")
+            return False
+
+    def trigger_prefetch(self, n=None):
+        """Trigger prefetching of the next N tensors with proper synchronization"""
+        if n is None:
+            n = self.max_prefetch
+
+        prefetch_paths = []
+        with self.manager_lock:
+            # Find files that are ready to be prefetched (not already in cache or being prefetched)
+            for path in reversed(self.tensor_paths):
+                if (
+                    path not in self.prefetch_cache
+                    and self.file_status.get(path) == "ready"
+                ):
+                    prefetch_paths.append(path)
+                    if len(prefetch_paths) >= n:
+                        break
+
+        # Queue files for prefetching
+        for path in prefetch_paths:
+            self.prefetch_queue.put(path)
+
+    def cleanup_tensor(self, file_path: str):
+        """Clean up a specific tensor file after it's been used"""
+        with self.manager_lock:
+            if file_path in self.tensor_paths:
+                self.tensor_paths.remove(file_path)
+
+            # Remove from prefetch cache if present
+            if file_path in self.prefetch_cache:
+                del self.prefetch_cache[file_path]
+
+            # Remove from save futures if present
+            if file_path in self.save_futures:
+                future = self.save_futures[file_path]
+                if not future.done():
+                    future.cancel()
+                del self.save_futures[file_path]
+
+        # Try to delete the file
+        self._safe_delete_file(file_path)
+
+    def cleanup(self):
+        """Clean up all temp files and stop prefetch thread with proper synchronization"""
+        self.stop_event.set()
+
+        # Cancel all pending save operations
+        with self.manager_lock:
+            for _, future in self.save_futures.items():
+                if not future.done():
+                    future.cancel()
+            self.save_futures.clear()
+
+        # Drain the save queue
+        while not self.save_queue.empty():
+            try:
+                self.save_queue.get_nowait()
+                self.save_queue.task_done()
+            except queue.Empty:
+                break
+
+        # Shutdown the save pool
+        self.save_pool.shutdown(wait=False)
+
+        # Join the save worker thread
+        if self.save_worker.is_alive():
+            self.save_worker.join(timeout=2.0)
+
+        # Join the prefetch worker threads
+        for thread in self.prefetch_workers:
+            if thread.is_alive():
+                thread.join(timeout=2.0)
+
+        # Clear cache and remove all temporary files
+        with self.manager_lock:
+            self.prefetch_cache.clear()
+            paths_to_delete = list(self.tensor_paths)
+            self.tensor_paths.clear()
+
+        # Delete all temporary files
+        for path in paths_to_delete:
+            self._safe_delete_file(path)
+
+        # Remove temp directory
+        try:
+            if os.path.exists(self.temp_dir):
+                shutil.rmtree(self.temp_dir, ignore_errors=True)
+        except FileNotFoundError as e:
+            logger.warning(f"Error removing temporary directory {self.temp_dir}: {e}")
+
+
+class Disco(torch.autograd.Function):
+    """
+    Disco: DIsk-based Storage and Checkpointing with Optimized prefetching
+    Advanced disk-based gradient checkpointer with prefetching.
+    """
+
+    # Shared manager instance across all checkpointing operations
+    _manager = None
+
+    @staticmethod
+    def get_instance(prefetch_size=1, prefetch_to_gpu=True, save_workers=4):
+        """Get or create the offload manager"""
+        if Disco._manager is None:
+            Disco._manager = DiskOffloadManager(
+                prefetch_size=prefetch_size,
+                prefetch_to_gpu=prefetch_to_gpu,
+                save_workers=save_workers,
+            )
+        return Disco._manager
+
+    @staticmethod
+    @torch_cuda_amp_custom_fwd
+    def forward(
+        ctx,
+        forward_function,
+        hidden_states,
+        *args,
+        prefetch_size=1,
+        prefetch_to_gpu=True,
+        save_workers=4,
+    ):
+        """Forward pass that offloads activations to disk asynchronously"""
+        # Get or create the manager
+        manager = Disco.get_instance(
+            prefetch_size=prefetch_size,
+            prefetch_to_gpu=prefetch_to_gpu,
+            save_workers=save_workers,
+        )
+
+        # Save tensor to disk asynchronously
+        file_path = manager.save_tensor(hidden_states)
+
+        # Run forward pass immediately without waiting for save to complete
+        with torch.no_grad():
+            output = forward_function(hidden_states, *args)
+
+        # Store what we need for backward
+        ctx.save_for_backward(torch.tensor([0]))  # Dummy tensor
+        ctx.file_path = file_path
+        ctx.forward_function = forward_function
+        ctx.args = args
+
+        return output
+
+    @staticmethod
+    @torch_cuda_amp_custom_bwd
+    def backward(ctx, *grad_outputs):
+        """Backward pass that loads activations from disk with prefetching"""
+        # Get the manager
+        manager = Disco._manager
+
+        # Trigger prefetching for future tensors
+        # This happens at the start of backward, so should have time to complete
+        manager.trigger_prefetch()
+
+        # Load hidden states from disk or prefetch cache
+        file_path = ctx.file_path
+        try:
+            # Ensure the file is saved before we try to load it
+            manager.wait_for_save(file_path)
+
+            hidden_states = manager.load_tensor(file_path)
+            hidden_states.requires_grad = True
+
+            # Compute gradients
+            with torch.enable_grad():
+                output = ctx.forward_function(hidden_states, *ctx.args)
+
+                # Handle tuple outputs properly
+                if isinstance(output, tuple):
+                    if len(grad_outputs) == len(output):
+                        torch.autograd.backward(output, grad_outputs)
+                    else:
+                        torch.autograd.backward(output, grad_outputs[0])
+                else:
+                    torch.autograd.backward(output, grad_outputs[0])
+
+            # Clean up the file after we're done with it
+            manager.cleanup_tensor(file_path)
+
+            return (
+                (
+                    None,  # forward_function
+                    hidden_states.grad,  # hidden_states grad
+                )
+                + (None,) * len(ctx.args)  # for each arg
+                + (
+                    None,  # prefetch_size
+                    None,  # prefetch_to_gpu
+                    None,  # save_workers
+                )
+            )
+
+        except Exception as e:
+            logger.error(f"Error in backward pass: {e}")
+            # Clean up the file even on error
+            manager.cleanup_tensor(file_path)
+            raise

src/axolotl/utils/models.py

@@ -70,10 +70,12 @@ from axolotl.utils.distributed import (
     is_local_main_process,
     is_main_process,
 )
-from axolotl.utils.gradient_checkpointing import hf_grad_checkpoint_offload_wrapper
+from axolotl.utils.gradient_checkpointing import (
+    hf_grad_checkpoint_disk_offload_wrapper,
+    hf_grad_checkpoint_offload_wrapper,
+)
 from axolotl.utils.lora_embeddings import get_linear_embedding_layers
 from axolotl.utils.model_shard_quant import load_sharded_model, load_sharded_model_quant
-from axolotl.utils.schemas.enums import RLType
 
 LOG = logging.getLogger(__name__)
 PLUGIN_MANAGER = PluginManager.get_instance()
@@ -142,22 +144,6 @@ def check_model_config(cfg: DictDefault, model_config: PretrainedConfig):
         hasattr(model_config, "quantization_config")
         and model_config.quantization_config
     )
-
-    # Detect compressed-tensors config
-    is_compressed_tensors_config = (
-        quant_config_exists
-        and model_config.quantization_config.get("quant_method") == "compressed-tensors"
-    )
-
-    if is_compressed_tensors_config:
-        if model_config.quantization_config.get("config_groups"):
-            LOG.warning(
-                "Found `config_groups` in a compressed-tensors config. "
-                "QAT integration with llmcompressor is not tested."
-            )
-        # Skip further quant checks for compressed-tensors
-        return
-
     quant_config_method_is_gptq = (
         quant_config_exists
         and "quant_method" in model_config.quantization_config
@@ -620,6 +606,10 @@ class ModelLoader:
 
         if self.cfg.gradient_checkpointing in ["unsloth", "offload"]:
             transformers.modeling_utils.checkpoint = hf_grad_checkpoint_offload_wrapper
+        if self.cfg.gradient_checkpointing == "offload_disk":
+            transformers.modeling_utils.checkpoint = (
+                hf_grad_checkpoint_disk_offload_wrapper
+            )
 
         if self.cfg.flash_attention:
             self.patch_attention()
@@ -1373,7 +1363,7 @@ class ModelLoader:
             # then the dpo trainer doesn't want the peft model loaded over it, it just wants the lora/peft config
             if (
                 self.cfg.adapter
-                and self.cfg.rl in [RLType.DPO, RLType.IPO, RLType.KTO]
+                and self.cfg.rl in ["dpo", "ipo", "kto"]
                 and not self.cfg.merge_lora
             ):
                 _, lora_config = load_lora(

src/axolotl/utils/schemas/config.py

@@ -27,7 +27,7 @@ from axolotl.utils.schemas.datasets import (
     StepwiseSupervisedDataset,
 )
 from axolotl.utils.schemas.deprecated import DeprecatedParameters, RemappedParameters
-from axolotl.utils.schemas.enums import ChatTemplate, RingAttnFunc, RLType
+from axolotl.utils.schemas.enums import ChatTemplate, RLType
 from axolotl.utils.schemas.integrations import (
     CometConfig,
     GradioConfig,
@@ -83,7 +83,6 @@ class AxolotlInputConfig(
     # optionally shrink the embeddings when the tokenizer vocab size is smaller
     shrink_embeddings: bool | None = None
     embeddings_skip_upcast: bool | None = None
-    random_init_weights: bool | None = None
 
     rl: RLType | None = None
     trl: TRLConfig | None = Field(
@@ -179,7 +178,7 @@ class AxolotlInputConfig(
 
     # torch_dtype: torch.dtype | None
 
-    gradient_checkpointing: Literal["unsloth", "offload"] | bool | None = Field(
+    gradient_checkpointing: Literal["offload", "offload_disk"] | bool | None = Field(
         default=False
     )
     gradient_checkpointing_kwargs: dict[str, Any] | None = None
@@ -261,7 +260,7 @@ class AxolotlInputConfig(
 
     sequence_parallel_degree: int | None = None
     heads_k_stride: int | None = None
-    ring_attn_func: RingAttnFunc | None = None
+    ring_attn_func: str | None = None
 
     special_tokens: SpecialTokensConfig | None = None
     tokens: list[str] | None = None
@@ -783,7 +782,7 @@ class AxolotlInputConfig(
 
     @model_validator(mode="after")
     def check_simpo_warmup(self):
-        if self.rl is RLType.SIMPO and self.warmup_ratio:
+        if self.rl == "simpo" and self.warmup_ratio:
             raise ValueError(
                 "warmup_ratio is not supported with the simpo trainer. Please use `warmup_steps` instead"
             )
@@ -1186,7 +1185,7 @@ class AxolotlInputConfig(
 
             if self.sample_packing and self.micro_batch_size > 1:
                 raise ValueError(
-                    "micro_batch_size must be set to 1 when sample_packing is enabled "
+                    "micro_batch_size must be set to 1 when sample_packing is enabled"
                     "due to a `ring-flash-attn` requirement"
                 )
 
@@ -1218,8 +1217,16 @@ class AxolotlInputConfig(
         if getattr(self, "sequence_parallel_degree", 1) == 1:
             return self
 
+        from axolotl.monkeypatch.attention.ring_attn.patch import RingAttnFunc
+
         if self.ring_attn_func is not None:
-            self.ring_attn_func = RingAttnFunc(self.ring_attn_func)
+            valid_funcs = list(RingAttnFunc)
+            if self.ring_attn_func in valid_funcs:
+                self.ring_attn_func = RingAttnFunc(self.ring_attn_func)
+            else:
+                raise ValueError(
+                    f"ring_attn_func: {self.ring_attn_func} must be in {valid_funcs}"
+                )
         else:
             # Default ring attention function selection
             sample_packing = getattr(self, "sample_packing", False)

src/axolotl/utils/schemas/enums.py

@@ -6,12 +6,12 @@ from enum import Enum
 class RLType(str, Enum):
     """RL trainer type configuration subset"""
 
-    DPO = "dpo"  # pylint: disable=invalid-name
+    dpo = "dpo"  # pylint: disable=invalid-name
-    GRPO = "grpo"  # pylint: disable=invalid-name
+    grpo = "grpo"  # pylint: disable=invalid-name
-    IPO = "ipo"  # pylint: disable=invalid-name
+    ipo = "ipo"  # pylint: disable=invalid-name
-    ORPO = "orpo"  # pylint: disable=invalid-name
+    orpo = "orpo"  # pylint: disable=invalid-name
-    KTO = "kto"  # pylint: disable=invalid-name
+    kto = "kto"  # pylint: disable=invalid-name
-    SIMPO = "simpo"  # pylint: disable=invalid-name
+    simpo = "simpo"  # pylint: disable=invalid-name
 
 
 class ChatTemplate(str, Enum):
@@ -55,14 +55,3 @@ class CustomSupportedOptimizers(str, Enum):
     adopt_adamw = "adopt_adamw"  # pylint: disable=invalid-name
     came_pytorch = "came_pytorch"  # pylint: disable=invalid-name
     muon = "muon"  # pylint: disable=invalid-name
-
-
-class RingAttnFunc(str, Enum):
-    """Enum class for supported `ring-flash-attn` implementations"""
-
-    # VARLEN_RING = "varlen_ring"
-    # VARLEN_ZIGZAG = "varlen_zigzag"
-    VARLEN_LLAMA3 = "varlen_llama3"
-    BATCH_RING = "batch_ring"
-    # BATCH_ZIGZAG = "batch_zigzag"
-    # BATCH_STRIPE = "batch_stripe"

tests/e2e/integrations/test_llm_compressor.py

@@ -1,111 +0,0 @@
-"""
-E2E smoke tests for LLMCompressorPlugin integration
-"""
-
-from pathlib import Path
-
-import pytest
-
-from axolotl.cli.args import TrainerCliArgs
-from axolotl.common.datasets import load_datasets
-from axolotl.train import train
-from axolotl.utils.config import normalize_config, prepare_plugins, validate_config
-from axolotl.utils.dict import DictDefault
-
-from tests.e2e.utils import (
-    check_model_output_exists,
-    require_llmcompressor,
-    require_torch_2_4_1,
-)
-
-MODELS = [
-    "nm-testing/llama2.c-stories42M-pruned2.4-compressed",
-    "nm-testing/llama2.c-stories42M-gsm8k-sparse-only-compressed",
-]
-
-
-@pytest.mark.parametrize(
-    "base_model", MODELS, ids=["no-checkpoint-recipe", "with-checkpoint-recipe"]
-)
-@pytest.mark.parametrize(
-    "save_compressed", [True, False], ids=["save_compressed", "save_uncompressed"]
-)
-class TestLLMCompressorIntegration:
-    """
-    e2e tests for axolotl.integrations.llm_compressor.LLMCompressorPlugin
-    """
-
-    @require_llmcompressor
-    @require_torch_2_4_1
-    def test_llmcompressor_plugin(
-        self, temp_dir, base_model: str, save_compressed: bool
-    ):
-        from llmcompressor import active_session
-
-        # core cfg
-        cfg = DictDefault(
-            {
-                "base_model": base_model,
-                "plugins": ["axolotl.integrations.llm_compressor.LLMCompressorPlugin"],
-                "sequence_len": 1024,
-                "val_set_size": 0.05,
-                "special_tokens": {"pad_token": "<|endoftext|>"},
-                "datasets": [{"path": "mhenrichsen/alpaca_2k_test", "type": "alpaca"}],
-                "num_epochs": 1,
-                "micro_batch_size": 2,
-                "gradient_accumulation_steps": 2,
-                "output_dir": temp_dir,
-                "learning_rate": 1e-5,
-                "optimizer": "adamw_torch_fused",
-                "lr_scheduler": "cosine",
-                "save_safetensors": True,
-                "bf16": "auto",
-                "max_steps": 5,
-                "llmcompressor": {
-                    "recipe": {
-                        "finetuning_stage": {
-                            "finetuning_modifiers": {
-                                "ConstantPruningModifier": {
-                                    "targets": [
-                                        "re:.*q_proj.weight",
-                                        "re:.*k_proj.weight",
-                                        "re:.*v_proj.weight",
-                                        "re:.*o_proj.weight",
-                                        "re:.*gate_proj.weight",
-                                        "re:.*up_proj.weight",
-                                        "re:.*down_proj.weight",
-                                    ],
-                                    "start": 0,
-                                },
-                            },
-                        },
-                    },
-                    "save_compressed": save_compressed,
-                },
-            }
-        )
-
-        prepare_plugins(cfg)
-        cfg = validate_config(cfg)
-        normalize_config(cfg)
-        cli_args = TrainerCliArgs()
-        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
-
-        try:
-            train(cfg=cfg, dataset_meta=dataset_meta)
-            check_model_output_exists(temp_dir, cfg)
-            _check_llmcompressor_model_outputs(temp_dir, save_compressed)
-        finally:
-            active_session().reset()
-
-
-def _check_llmcompressor_model_outputs(temp_dir, save_compressed):
-    if save_compressed:
-        assert (Path(temp_dir) / "recipe.yaml").exists()
-
-        from compressed_tensors import ModelCompressor
-        from compressed_tensors.config import Sparse24BitMaskConfig
-
-        compressor = ModelCompressor.from_pretrained(temp_dir)
-        assert compressor is not None
-        assert isinstance(compressor.sparsity_config, Sparse24BitMaskConfig)

tests/e2e/multigpu/patched/test_sp.py

@@ -25,7 +25,6 @@ class TestSequenceParallelism:
         micro_batch_size=1,
         pad_to_sequence_len=True,
         ring_attn_func=None,
-        threshold=2.0,
     ):
         """Helper method to run sequence parallel tests with different configurations"""
         cfg = DictDefault(
@@ -94,22 +93,22 @@ class TestSequenceParallelism:
         )
 
         check_tensorboard(
-            temp_dir + "/runs", "train/train_loss", threshold, "Train Loss is too high"
+            temp_dir + "/runs", "train/train_loss", 2.6, "Train Loss is too high"
         )
 
     @pytest.mark.parametrize(
-        "sample_packing, micro_batch_size, pad_to_sequence_len, ring_attn_func, threshold",
+        "sample_packing, micro_batch_size, pad_to_sequence_len, ring_attn_func",
         [
-            (True, 1, True, None, 2.5),  # defaults to varlen_llama3 ring_attn_func
+            (True, 1, True, None),  # defaults to varlen_llama3 ring_attn_func
-            (False, 2, True, None, 2.5),  # defaults to batch_ring ring_attn_func
+            (False, 2, True, None),  # defaults to batch_ring ring_attn_func
-            # (False, 2, True, "batch_zigzag", 2.5),
+            (False, 2, True, "batch_zigzag"),
-            (False, 2, False, None, 2.5),  # defaults to batch_ring ring_attn_func
+            # (False, 2, False),  # not yet working
         ],
         ids=[
             "sample_packing, varlen_llama3 ring_attn_func",
-            "no sample_packing, pad_to_sequence_len, batch_ring ring_attn_func",
-            # "no sample_packing, no pad_to_sequence_len, batch_zigzag ring_attn_func",
             "no sample_packing, no pad_to_sequence_len, batch_ring ring_attn_func",
+            "no sample_packing, no pad_to_sequence_len, batch_zigzag ring_attn_func",
+            # "no sample_packing, pad_to_sequence_len",  # not yet working
         ],
     )
     def test_sequence_parallel_training(
@@ -119,7 +118,6 @@ class TestSequenceParallelism:
         micro_batch_size,
         pad_to_sequence_len,
         ring_attn_func,
-        threshold,
     ):
         """Test sequence parallel training with different configurations"""
         self._run_sequence_parallel_test(
@@ -128,5 +126,4 @@ class TestSequenceParallelism:
             micro_batch_size=micro_batch_size,
             pad_to_sequence_len=pad_to_sequence_len,
             ring_attn_func=ring_attn_func,
-            threshold=threshold,
         )

tests/e2e/patched/test_activation_checkpointing.py

@@ -26,10 +26,15 @@ class TestActivationCheckpointing:
     E2E tests for activation checkpointing
     """
 
+    @pytest.mark.parametrize(
+        "gradient_checkpointing",
+        ["offload", "offload_disk"],
+    )
     def test_activation_checkpointing_offload(
         self,
         temp_dir,
         fix_checkpoint_after_test,  # pylint: disable=unused-argument,redefined-outer-name
+        gradient_checkpointing,
     ):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
@@ -64,7 +69,7 @@ class TestActivationCheckpointing:
                 "sample_packing": True,
                 "bf16": True,
                 "save_safetensors": True,
-                "gradient_checkpointing": "offload",
+                "gradient_checkpointing": gradient_checkpointing,
             }
         )
 

tests/e2e/patched/test_sp.py

@@ -10,15 +10,14 @@ import pytest
 import torch
 from accelerate.state import PartialState
 
+from axolotl.core.trainers.mixins.sequence_parallel import apply_sequence_parallelism
 from axolotl.monkeypatch.attention.ring_attn import (
+    RingAttnFunc,
     get_ring_attn_group,
     register_ring_attn,
     set_ring_attn_group,
 )
-from axolotl.utils.ctx_managers.sequence_parallel import apply_sequence_parallelism
 from axolotl.utils.dict import DictDefault
-from axolotl.utils.schemas.enums import RingAttnFunc
-from axolotl.utils.schemas.trl import TRLConfig
 
 
 @pytest.fixture
@@ -63,14 +62,12 @@ def sequence_parallel_batch():
     input_ids = torch.arange(batch_size * seq_len).reshape(batch_size, seq_len)
     attention_mask = torch.ones(batch_size, seq_len)
     position_ids = torch.arange(seq_len).expand(batch_size, seq_len)
-    labels = input_ids.clone()
 
     # Create test batch
     batch = {
         "input_ids": input_ids,
         "attention_mask": attention_mask,
         "position_ids": position_ids,
-        "labels": labels,
     }
 
     return batch
@@ -182,44 +179,12 @@ class TestConfigValidation:
                 False,
                 "micro_batch_size must be set to 1",
             ),
-            # Valid: Basic GRPO config
-            (
-                {
-                    "sequence_parallel_degree": 2,
-                    "flash_attention": True,
-                    "micro_batch_size": 2,
-                    "trl": {"use_liger_loss": True},
-                },
-                {
-                    "sequence_parallel_degree": 2,
-                    "flash_attention": True,
-                    "micro_batch_size": 2,
-                    "trl": TRLConfig(use_liger_loss=True),
-                },
-                True,
-                "GRPO + SP + Liger not currently supported",
-            ),
-            # Invalid: GRPO config with Liger loss
-            (
-                {
-                    "rl": "grpo",
-                    "sequence_parallel_degree": 2,
-                    "flash_attention": True,
-                    "micro_batch_size": 2,
-                    "trl": {"use_liger_loss": True},
-                },
-                None,
-                False,
-                "GRPO + SP + Liger not currently supported",
-            ),
         ],
         ids=[
             "valid_config",
             "default_sp_degree",
             "without_flash_attention",
             "sample_packing_with_large_batch",
-            "valid_grpo",
-            "grpo_with_liger_loss",
         ],
     )
     def test_sequence_parallel_config_validation(
@@ -291,7 +256,7 @@ class TestConfigValidation:
             AxolotlInputConfig(**cfg)
 
         # Verify error message
-        assert "Input should be 'varlen_llama3' or 'batch_ring'" in str(excinfo.value)
+        assert "ring_attn_func: INVALID_FUNC must be in" in str(excinfo.value)
 
 
 class TestApplySequenceParallelism:
@@ -325,11 +290,10 @@ class TestApplySequenceParallelism:
 
     def test_world_size_one(self, sequence_parallel_batch):
         """Test that function returns original batch when world size is 1."""
-        result, _, _ = apply_sequence_parallelism(
+        result = apply_sequence_parallelism(
             batch=sequence_parallel_batch,
             local_rank=0,
             local_world_size=1,
-            gradient_accumulation_steps=1,
             ring_attn_func=RingAttnFunc.BATCH_RING,
         )
 
@@ -341,11 +305,10 @@ class TestApplySequenceParallelism:
         batch = sequence_parallel_batch
         seq_len = batch["input_ids"].size(1)
 
-        result, _, _ = apply_sequence_parallelism(
+        result = apply_sequence_parallelism(
             batch=batch,
             local_rank=0,
             local_world_size=2,
-            gradient_accumulation_steps=1,
             ring_attn_func=RingAttnFunc.BATCH_RING,
         )
 
@@ -365,59 +328,57 @@ class TestApplySequenceParallelism:
         seq_len = batch["input_ids"].size(1)
         original_input_ids = batch["input_ids"].clone()
 
-        result, _, _ = apply_sequence_parallelism(
+        result = apply_sequence_parallelism(
             batch=batch,
             local_rank=1,
             local_world_size=2,
-            gradient_accumulation_steps=1,
             ring_attn_func=RingAttnFunc.BATCH_RING,
         )
 
         # Verify content: rank 1 should get the second half of the sequence
         assert torch.equal(result["input_ids"], original_input_ids[:, seq_len // 2 :])
 
-    # TODO(djsaunde): add back once implemented.
+    def test_batch_zigzag(self, sequence_parallel_batch):
-    # def test_batch_zigzag(self, sequence_parallel_batch):
+        """Test BATCH_ZIGZAG sharding pattern."""
-    #     """Test BATCH_ZIGZAG sharding pattern."""
+        batch = sequence_parallel_batch
-    #     batch = sequence_parallel_batch
+        original_input_ids = batch["input_ids"].clone()
-    #     original_input_ids = batch["input_ids"].clone()
+        seq_len = batch["input_ids"].size(1)
-    #     seq_len = batch["input_ids"].size(1)
 
-    #     # Test rank 0
+        # Test rank 0
-    #     result_rank0 = apply_sequence_parallelism(
+        result_rank0 = apply_sequence_parallelism(
-    #         batch={k: v.clone() for k, v in batch.items()},
+            batch={k: v.clone() for k, v in batch.items()},
-    #         local_rank=0,
+            local_rank=0,
-    #         local_world_size=2,
+            local_world_size=2,
-    #         ring_attn_func=RingAttnFunc.BATCH_ZIGZAG,
+            ring_attn_func=RingAttnFunc.BATCH_ZIGZAG,
-    #     )
+        )
 
-    #     # Test rank 1
+        # Test rank 1
-    #     result_rank1 = apply_sequence_parallelism(
+        result_rank1 = apply_sequence_parallelism(
-    #         batch={k: v.clone() for k, v in batch.items()},
+            batch={k: v.clone() for k, v in batch.items()},
-    #         local_rank=1,
+            local_rank=1,
-    #         local_world_size=2,
+            local_world_size=2,
-    #         ring_attn_func=RingAttnFunc.BATCH_ZIGZAG,
+            ring_attn_func=RingAttnFunc.BATCH_ZIGZAG,
-    #     )
+        )
 
-    #     # Checks for both ranks
+        # Checks for both ranks
-    #     assert result_rank0["input_ids"].shape[1] == seq_len // 2
+        assert result_rank0["input_ids"].shape[1] == seq_len // 2
-    #     assert result_rank1["input_ids"].shape[1] == seq_len // 2
+        assert result_rank1["input_ids"].shape[1] == seq_len // 2
 
-    #     # For a 2-rank system with 8 tokens, check specific zigzag pattern
+        # For a 2-rank system with 8 tokens, check specific zigzag pattern
-    #     # Rank 0 should get chunks [0, 1] and [6, 7]
+        # Rank 0 should get chunks [0, 1] and [6, 7]
-    #     # Rank 1 should get chunks [2, 3] and [4, 5]
+        # Rank 1 should get chunks [2, 3] and [4, 5]
-    #     if seq_len == 8:
+        if seq_len == 8:
-    #         # Create expected tensors for comparison
+            # Create expected tensors for comparison
-    #         rank0_expected = torch.cat(
+            rank0_expected = torch.cat(
-    #             [original_input_ids[:, :2], original_input_ids[:, 6:8]], dim=1
+                [original_input_ids[:, :2], original_input_ids[:, 6:8]], dim=1
-    #         )
+            )
 
-    #         rank1_expected = torch.cat(
+            rank1_expected = torch.cat(
-    #             [original_input_ids[:, 2:4], original_input_ids[:, 4:6]], dim=1
+                [original_input_ids[:, 2:4], original_input_ids[:, 4:6]], dim=1
-    #         )
+            )
 
-    #         assert torch.equal(result_rank0["input_ids"], rank0_expected)
+            assert torch.equal(result_rank0["input_ids"], rank0_expected)
-    #         assert torch.equal(result_rank1["input_ids"], rank1_expected)
+            assert torch.equal(result_rank1["input_ids"], rank1_expected)
 
     def test_partial_application(self, sequence_parallel_batch):
         """Test that we can create a partially applied version of the function."""
@@ -429,12 +390,11 @@ class TestApplySequenceParallelism:
             apply_sequence_parallelism,
             local_rank=0,
             local_world_size=2,
-            gradient_accumulation_steps=1,
             ring_attn_func=RingAttnFunc.BATCH_RING,
         )
 
         # Use the partially applied function
-        result, _, _ = rank0_ring_parallel(batch=batch)
+        result = rank0_ring_parallel(batch=batch)
 
         # Verify it works as expected
         assert result["input_ids"].shape[1] == original_input_ids.shape[1] // 2
@@ -452,15 +412,13 @@ class TestApplySequenceParallelism:
         original_input_ids = batch["input_ids"].clone()
 
         # This should run without error even though position_ids is missing
-        result, _, _ = apply_sequence_parallelism(
+        result = apply_sequence_parallelism(
             batch=batch,
             local_rank=0,
             local_world_size=2,
-            gradient_accumulation_steps=1,
             ring_attn_func=RingAttnFunc.BATCH_RING,
         )
 
         # Verification should pass
-        assert "position_ids" in result
+        assert "position_ids" not in result
-        assert result["input_ids"].shape[1] == result["position_ids"].shape[1]
         assert result["input_ids"].shape[1] == original_input_ids.shape[1] // 2

tests/e2e/utils.py

@@ -105,25 +105,7 @@ def require_vllm(test_case):
             return False
 
     return unittest.skipUnless(
-        is_vllm_installed(), "test requires vllm to be installed"
+        is_vllm_installed(), "test requires a vllm to be installed"
-    )(test_case)
-
-
-def require_llmcompressor(test_case):
-    """
-    Decorator marking a test that requires a llmcompressor to be installed
-    """
-
-    def is_llmcompressor_installed():
-        try:
-            import llmcompressor  # pylint: disable=unused-import  # noqa: F401
-
-            return True
-        except ImportError:
-            return False
-
-    return unittest.skipUnless(
-        is_llmcompressor_installed(), "test requires llmcompressor to be installed"
     )(test_case)