move import of llmcompressor to reset session inside test

Signed-off-by: Rahul Tuli <rtuli@redhat.com>

.github/workflows/base.yml

@@ -22,12 +22,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: "124"
-            cuda_version: 12.4.1
-            cudnn_version: ""
-            python_version: "3.11"
-            pytorch: 2.4.1
-            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
           - cuda: "124"
             cuda_version: 12.4.1
             cudnn_version: ""

.github/workflows/main.yml

@@ -15,11 +15,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -67,6 +62,7 @@ jobs:
             CUDA=${{ matrix.cuda }}
             PYTORCH_VERSION=${{ matrix.pytorch }}
             AXOLOTL_ARGS=${{ matrix.axolotl_args }}
+            AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}
           file: ./docker/Dockerfile
           push: ${{ github.event_name != 'pull_request' }}
           tags: |
@@ -82,11 +78,6 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"

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

@@ -9,6 +9,7 @@ on:
       - 'pyproject.toml'
       - '.github/workflows/multi-gpu-e2e.yml'
       - 'src/axolotl/core/trainers/mixins/sequence_parallel.py'
+      - 'src/axolotl/utils/distributed.py'
   workflow_dispatch:
   schedule:
     - cron: '0 0 * * 1,4'  # Runs at 00:00 UTC every monday & thursday
@@ -32,13 +33,6 @@ jobs:
             axolotl_extras: vllm
             num_gpus: 2
             nightly_build: "true"
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            axolotl_extras:  # no vllm support for 2.4.1
-            num_gpus: 2
-            nightly_build: "true"
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"

.github/workflows/nightlies.yml

@@ -12,11 +12,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -70,11 +65,6 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            axolotl_extras:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"

.github/workflows/tests-nightly.yml

@@ -26,7 +26,7 @@ jobs:
       max-parallel: 2
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.4.1", "2.5.1", "2.6.0"]
+        pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
     timeout-minutes: 20
 
     steps:
@@ -106,13 +106,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.4.1
-            num_gpus: 1
-            axolotl_extras:
-            nightly_build: "true"
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"

.github/workflows/tests.yml

@@ -27,6 +27,9 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
 
+env:
+  TRANSFORMERS_IS_CI: "yes"
+
 jobs:
   pre-commit:
     name: pre-commit
@@ -49,7 +52,7 @@ jobs:
       max-parallel: 2
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.4.1", "2.5.1", "2.6.0", "2.7.0"]
+        pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
     timeout-minutes: 20
 
     steps:
@@ -135,7 +138,7 @@ jobs:
       max-parallel: 1
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.4.1", "2.5.1", "2.6.0"]
+        pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
     timeout-minutes: 20
 
     steps:
@@ -258,6 +261,12 @@ 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"

.runpod/test-input.json

@@ -12,22 +12,22 @@
       "base_model": "HuggingFaceTB/SmolLM2-135M",
       "model_type": "AutoModelForCausalLM",
       "tokenizer_type": "AutoTokenizer",
-      "load_in_8bit": true,
-      "load_in_4bit": false,
+      "load_in_4bit": true,
       "strict": false,
       "datasets": [
         {
           "path": "mhenrichsen/alpaca_2k_test",
-          "type": "alpaca"
+          "type": "alpaca",
+          "split": "train[:10%]"
         }
       ],
-      "val_set_size": 0.05,
+      "val_set_size": 0.02,
       "output_dir": "./outputs/lora-out",
       "sequence_len": 4096,
       "sample_packing": true,
       "eval_sample_packing": false,
       "pad_to_sequence_len": true,
-      "adapter": "lora",
+      "adapter": "qlora",
       "lora_r": 32,
       "lora_alpha": 64,
       "lora_dropout": 0.05,
@@ -36,8 +36,8 @@
         "embed_tokens",
         "lm_head"
       ],
-      "gradient_accumulation_steps": 4,
-      "micro_batch_size": 2,
+      "gradient_accumulation_steps": 2,
+      "micro_batch_size": 1,
       "num_epochs": 1,
       "optimizer": "adamw_torch_fused",
       "lr_scheduler": "cosine",
@@ -56,7 +56,8 @@
       "weight_decay": 0.0,
       "special_tokens": {
         "pad_token": "<|endoftext|>"
-      }
+      },
+      "max_steps": 20
     },
     "timeout": 100000
   },

docs/config.qmd

@@ -184,6 +184,10 @@ datasets:
     # adding a system turn with empty content.
     drop_system_message:
 
+    # Optional[bool]. Whether to split the assistant turn based on a reasoning trace inside delimited tags
+    # defaults to False
+    split_thinking:
+
     # IMPORTANT: The following fields determine which parts of the conversation to train on.
     # Priority order: message_field_training > message_field_training_detail > train_on_inputs or role in roles_to_train
     # See examples at `docs/dataset-formats/conversation.qmd`

docs/custom_integrations.qmd

@@ -49,7 +49,8 @@ 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/multimodal.qmd

@@ -164,7 +164,7 @@ Here is an example of a multi-modal dataset:
         {
             "role": "user",
             "content": [
-                {"type": "image", "image": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
+                {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
                 {"type": "text", "text": "Describe this image in detail."}
             ]
         },

examples/llama-3/sparse-finetuning.yaml

@@ -0,0 +1,77 @@
+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

@@ -67,13 +67,13 @@ def parse_requirements(extras_require_map):
             if (major, minor) >= (2, 7):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 # _install_requires.append("xformers==0.0.29.post3")  # xformers seems to be hard pinned to 2.6.0
-                extras_require_map["vllm"] = ["vllm==0.8.4"]
+                extras_require_map["vllm"] = ["vllm==0.8.5"]
             elif (major, minor) >= (2, 6):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 _install_requires.append(
                     "xformers==0.0.29.post2"
                 )  # vllm needs post2 w torch 2.6
-                extras_require_map["vllm"] = ["vllm==0.8.4"]
+                extras_require_map["vllm"] = ["vllm==0.8.5"]
             elif (major, minor) >= (2, 5):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 if patch == 0:
@@ -149,6 +149,9 @@ 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__ = "0.9.0"
+__version__ = "0.10.0.dev0"

src/axolotl/cli/config.py

@@ -152,6 +152,12 @@ def prepare_plugins(cfg: DictDefault):
             plugin_manager.register(plugin_name)
 
 
+def plugin_set_cfg(cfg: DictDefault):
+    if cfg.get("plugins"):
+        plugin_manager = PluginManager.get_instance()
+        plugin_manager.cfg = cfg
+
+
 def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs) -> DictDefault:
     """
     Loads the `axolotl` configuration stored at `config`, validates it, and performs
@@ -213,5 +219,6 @@ def load_cfg(config: Union[str, Path] = Path("examples/"), **kwargs) -> DictDefa
     setup_wandb_env_vars(cfg)
     setup_mlflow_env_vars(cfg)
     setup_comet_env_vars(cfg)
+    plugin_set_cfg(cfg)
 
     return cfg

src/axolotl/cli/evaluate.py

@@ -1,6 +1,7 @@
 """CLI to run evaluation on a model."""
 
 import logging
+import os
 from pathlib import Path
 from typing import Union
 
@@ -14,6 +15,7 @@ from axolotl.cli.checks import check_accelerate_default_config, check_user_token
 from axolotl.cli.config import load_cfg
 from axolotl.common.datasets import load_datasets, load_preference_datasets
 from axolotl.evaluate import evaluate
+from axolotl.utils import set_pytorch_cuda_alloc_conf
 from axolotl.utils.dict import DictDefault
 
 LOG = logging.getLogger(__name__)
@@ -29,10 +31,14 @@ def do_evaluate(cfg: DictDefault, cli_args: TrainerCliArgs) -> None:
         cfg: Dictionary mapping `axolotl` config keys to values.
         cli_args: CLI arguments.
     """
+    # Enable expandable segments for cuda allocation to improve VRAM usage
+    set_pytorch_cuda_alloc_conf()
+
     # pylint: disable=duplicate-code
     print_axolotl_text_art()
     check_accelerate_default_config()
-    check_user_token()
+    if int(os.getenv("LOCAL_RANK", "0")) == 0:
+        check_user_token()
 
     if cfg.rl:
         dataset_meta = load_preference_datasets(cfg=cfg, cli_args=cli_args)

src/axolotl/cli/main.py

@@ -28,7 +28,6 @@ from axolotl.cli.utils import (
     fetch_from_github,
     filter_none_kwargs,
 )
-from axolotl.cli.vllm_serve import do_vllm_serve
 from axolotl.integrations.lm_eval.cli import lm_eval
 from axolotl.utils import set_pytorch_cuda_alloc_conf
 from axolotl.utils.schemas.config import AxolotlInputConfig
@@ -327,6 +326,8 @@ def fetch(directory: str, dest: Optional[str]) -> None:
 @add_options_from_dataclass(VllmServeCliArgs)
 @filter_none_kwargs
 def vllm_serve(config: str, **cli_args: VllmServeCliArgs):
+    from axolotl.cli.vllm_serve import do_vllm_serve
+
     do_vllm_serve(config, cli_args)
 
 

src/axolotl/common/architectures.py

@@ -11,5 +11,6 @@ MOE_ARCH_BLOCK = {
     ],
     "mixtral": "MixtralSparseMoeBlock",
     "qwen2_moe": "Qwen2MoeSparseMoeBlock",
+    "qwen3_moe": "Qwen3MoeSparseMoeBlock",
     "deepseek_v2": "DeepseekV2MoE",
 }

src/axolotl/core/trainer_builder.py

@@ -60,6 +60,7 @@ from axolotl.core.training_args import (
 from axolotl.integrations.base import PluginManager
 from axolotl.monkeypatch.multipack import SUPPORTED_MULTIPACK_MODEL_TYPES
 from axolotl.monkeypatch.relora import ReLoRACallback
+from axolotl.monkeypatch.trainer.lr import patch_trainer_get_lr
 from axolotl.processing_strategies import get_processing_strategy
 from axolotl.utils import is_comet_available, is_mlflow_available
 from axolotl.utils.callbacks import (
@@ -114,6 +115,8 @@ class TrainerBuilderBase(abc.ABC):
         if hasattr(model, "add_model_tags"):
             model.add_model_tags(["axolotl"])
 
+        patch_trainer_get_lr()
+
     @property
     def model_ref(self):
         return self._model_ref

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

@@ -70,6 +70,13 @@ class GRPOStrategy:
         if trl.scale_rewards is not None:
             grpo_args_kwargs["scale_rewards"] = trl.scale_rewards
 
+        if trl.loss_type is not None:
+            grpo_args_kwargs["loss_type"] = trl.loss_type
+        if trl.mask_truncated_completions is not None:
+            grpo_args_kwargs["mask_truncated_completions"] = (
+                trl.mask_truncated_completions
+            )
+
         if trl.temperature is not None:
             grpo_args_kwargs["temperature"] = trl.temperature
         if trl.top_p is not None:
@@ -85,6 +92,11 @@ class GRPOStrategy:
             grpo_args_kwargs["num_iterations"] = trl.num_iterations
         if trl.epsilon is not None:
             grpo_args_kwargs["epsilon"] = trl.epsilon
+        if trl.epsilon_high is not None:
+            grpo_args_kwargs["epsilon_high"] = trl.epsilon_high
+
+        if trl.use_liger_loss is not None:
+            grpo_args_kwargs["use_liger_loss"] = trl.use_liger_loss
 
         return grpo_args_kwargs
 

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

@@ -3,9 +3,10 @@
 import logging
 
 import torch
-from torch.optim.lr_scheduler import OneCycleLR
+from torch.optim.lr_scheduler import LRScheduler, OneCycleLR
 from transformers.trainer import Trainer
 
+from axolotl.integrations.base import PluginManager
 from axolotl.utils.schedulers import (
     RexLR,
     get_cosine_schedule_with_min_lr,
@@ -25,9 +26,9 @@ class SchedulerMixin(Trainer):
 
     def create_scheduler(
         self, num_training_steps: int, optimizer: torch.optim.Optimizer = None
-    ):
+    ) -> LRScheduler:
         """
-        Setup the scheduler. The optimizer of the trainer must have been set up either before this method is called or
+        Set up the scheduler. The optimizer of the trainer must have been set up either before this method is called or
         passed as an argument.
 
         Args:
@@ -47,7 +48,16 @@ class SchedulerMixin(Trainer):
         # fmt: off
         if self.lr_scheduler is None:  # type: ignore  # pylint: disable=access-member-before-definition
             # fmt: on
-            if self.args.alternate_lr_scheduler_type == "one_cycle":
+            plugin_manager = PluginManager.get_instance()
+            lr_scheduler: LRScheduler | None = plugin_manager.create_lr_scheduler(
+                trainer=self,
+                optimizer=optimizer,
+                num_training_steps=num_training_steps
+            )
+            if lr_scheduler is not None:
+                LOG.info(f"Using plugin-created lr_scheduler: {lr_scheduler}")
+                self.lr_scheduler = lr_scheduler
+            elif self.args.alternate_lr_scheduler_type == "one_cycle":
                 num_warmup_steps = self.args.get_warmup_steps(num_training_steps)
                 pct_start = num_warmup_steps / num_training_steps
                 extra_lr_kwargs = {}
@@ -110,4 +120,4 @@ class SchedulerMixin(Trainer):
             if use_cosine_min_lr:
                 LOG.warning("axolotl's cosine scheduler with min lr not used (e.g., because of deepspeed).")
 
-        return self.lr_scheduler
+        return self.lr_scheduler  # type: ignore

src/axolotl/core/trainers/relora.py

@@ -1,6 +1,7 @@
 """Module for ReLoRA trainer"""
 
 import torch
+from torch.optim.lr_scheduler import LRScheduler
 
 from axolotl.core.trainers.base import AxolotlTrainer
 from axolotl.monkeypatch.relora import ReLoRAScheduler
@@ -19,9 +20,11 @@ class ReLoRATrainer(AxolotlTrainer):
         self,
         num_training_steps: int,
         optimizer: torch.optim.Optimizer | None = None,
-    ):
+    ) -> LRScheduler:
         optimizer = self.optimizer if optimizer is None else optimizer
-        lr_scheduler = super().create_scheduler(num_training_steps, optimizer)
+        lr_scheduler: LRScheduler = super().create_scheduler(
+            num_training_steps, optimizer
+        )
 
         if self.args.relora_steps:
             warmup_steps = (
@@ -30,7 +33,7 @@ class ReLoRATrainer(AxolotlTrainer):
             anneal_steps = (
                 self.args.relora_anneal_steps if self.args.relora_anneal_steps else 1
             )
-            self.lr_scheduler = ReLoRAScheduler(
+            self.lr_scheduler = ReLoRAScheduler(  # type: ignore
                 optimizer,
                 lr_scheduler,
                 self.args.relora_steps,
@@ -38,6 +41,6 @@ class ReLoRATrainer(AxolotlTrainer):
                 warmup_steps,
             )
         else:
-            self.lr_scheduler = lr_scheduler
+            self.lr_scheduler = lr_scheduler  # type: ignore
 
-        return self.lr_scheduler
+        return self.lr_scheduler  # type: ignore

src/axolotl/evaluate.py

@@ -12,11 +12,12 @@ from datasets import Dataset
 from transformers.trainer import Trainer
 
 from axolotl.logging_config import configure_logging
-from axolotl.train import TrainDatasetMeta
-from axolotl.utils import set_pytorch_cuda_alloc_conf
+from axolotl.train import (
+    TrainDatasetMeta,
+    setup_model_and_tokenizer,
+)
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import cleanup_distributed
-from axolotl.utils.models import load_model, load_processor, load_tokenizer
 from axolotl.utils.trainer import setup_trainer
 
 project_root = os.path.abspath(os.path.join(os.path.dirname(__file__), ".."))
@@ -24,7 +25,7 @@ src_dir = os.path.join(project_root, "src")
 sys.path.insert(0, src_dir)
 
 configure_logging()
-LOG = get_logger("axolotl.evaluate")
+LOG = get_logger(__name__)
 
 
 def evaluate_dataset(
@@ -75,37 +76,22 @@ def evaluate(*, cfg: DictDefault, dataset_meta: TrainDatasetMeta) -> Dict[str, f
     Returns:
         Dictionary mapping metric names to their values.
     """
-    # pylint: disable=duplicate-code
-    # Enable expandable segments for cuda allocation to improve VRAM usage
-    set_pytorch_cuda_alloc_conf()
-
-    # Load tokenizer
-    LOG.debug(
-        f"loading tokenizer... {cfg.tokenizer_config or cfg.base_model_config}",
-        main_process_only=True,
-    )
-    tokenizer = load_tokenizer(cfg)
-
-    # Load processor for multimodal models if needed
-    processor = None
-    if cfg.is_multimodal:
-        processor = load_processor(cfg, tokenizer)
+    # Load tokenizer, processor and model
+    LOG.debug("loading model for evaluation...")
+    model, tokenizer, _, processor = setup_model_and_tokenizer(cfg)
 
     # Get datasets
+    # pylint: disable=duplicate-code
     train_dataset = dataset_meta.train_dataset
     eval_dataset = dataset_meta.eval_dataset
     total_num_steps = dataset_meta.total_num_steps
 
-    # Load model
-    LOG.debug("loading model for evaluation...")
-    model, _ = load_model(cfg, tokenizer, processor=processor)
-
     # Set up trainer
     trainer = setup_trainer(
-        cfg,
+        cfg=cfg,
         train_dataset=train_dataset,
         eval_dataset=eval_dataset,
-        model=(model, None, None),  # No need for model_ref or peft_config
+        model=model,
         tokenizer=tokenizer,
         processor=processor,
         total_num_steps=total_num_steps,

src/axolotl/integrations/base.py

@@ -24,6 +24,7 @@ import logging
 from typing import OrderedDict
 
 import torch
+from torch.optim.lr_scheduler import LRScheduler
 
 
 class BasePlugin:
@@ -41,7 +42,7 @@ class BasePlugin:
     post_lora_load(cfg, model): Performs actions after LoRA weights are loaded.
     post_model_load(cfg, model): Performs actions after the model is loaded, inclusive of any adapters.
     create_optimizer(cfg, trainer): Creates and returns an optimizer for training.
-    create_lr_scheduler(cfg, trainer, optimizer): Creates and returns a learning rate scheduler.
+    create_lr_scheduler(cfg, trainer, optimizer, num_training_steps): Creates and returns a learning rate scheduler.
     add_callbacks_pre_trainer(cfg, model): Adds callbacks to the trainer before training.
     add_callbacks_post_trainer(cfg, trainer): Adds callbacks to the trainer after training.
     """
@@ -146,8 +147,8 @@ class BasePlugin:
         """
 
     def create_lr_scheduler(
-        self, cfg, trainer, optimizer
-    ):  # pylint: disable=unused-argument
+        self, cfg, trainer, optimizer, num_training_steps
+    ) -> LRScheduler | None:  # pylint: disable=unused-argument
         """
         Creates and returns a learning rate scheduler.
 
@@ -155,9 +156,10 @@ class BasePlugin:
         cfg (dict): The configuration for the plugin.
         trainer (object): The trainer object for training.
         optimizer (object): The optimizer for training.
+        num_training_steps (int): Total number of training steps
 
         Returns:
-        object: The created learning rate scheduler.
+        object (LRScheduler): The created learning rate scheduler.
         """
 
     def add_callbacks_pre_trainer(self, cfg, model):  # pylint: disable=unused-argument
@@ -270,6 +272,7 @@ class PluginManager:
     plugins: OrderedDict[str, BasePlugin] = collections.OrderedDict()
 
     _instance = None
+    _cfg = None
 
     def __new__(cls):
         """
@@ -277,7 +280,9 @@ class PluginManager:
         """
         if cls._instance is None:
             cls._instance = super(PluginManager, cls).__new__(cls)
-            cls._instance.plugins = collections.OrderedDict()
+            cls._instance.plugins: OrderedDict[str, BasePlugin] = (
+                collections.OrderedDict()
+            )
         return cls._instance
 
     @staticmethod
@@ -290,6 +295,14 @@ class PluginManager:
             PluginManager()
         return PluginManager._instance  # type: ignore
 
+    @property
+    def cfg(self):
+        return self._cfg
+
+    @cfg.setter
+    def cfg(self, cfg):
+        self._cfg = cfg
+
     def register(self, plugin_name: str):
         """
         Registers a new plugin by its name.
@@ -409,29 +422,29 @@ class PluginManager:
                 return trainer_cls
         return None
 
-    def create_optimizer(self, cfg, trainer):
+    def create_optimizer(self, trainer):
         """
         Calls the create_optimizer method of all registered plugins and returns the first non-None optimizer.
 
         Parameters:
-        cfg (dict): The configuration for the plugins.
         trainer (object): The trainer object for training.
 
         Returns:
         object: The created optimizer, or None if none was found.
         """
         for plugin in self.plugins.values():
-            optimizer = plugin.create_optimizer(cfg, trainer)
+            optimizer = plugin.create_optimizer(self.cfg, trainer)
             if optimizer is not None:
                 return optimizer
         return None
 
-    def create_lr_scheduler(self, cfg, trainer, optimizer):
+    def create_lr_scheduler(
+        self, trainer, optimizer, num_training_steps
+    ) -> LRScheduler | None:
         """
         Calls the create_lr_scheduler method of all registered plugins and returns the first non-None scheduler.
 
         Parameters:
-        cfg (dict): The configuration for the plugins.
         trainer (object): The trainer object for training.
         optimizer (object): The optimizer for training.
 
@@ -439,7 +452,12 @@ class PluginManager:
         object: The created learning rate scheduler, or None if none was found.
         """
         for plugin in self.plugins.values():
-            scheduler = plugin.create_lr_scheduler(cfg, trainer, optimizer)
+            scheduler: LRScheduler | None = plugin.create_lr_scheduler(
+                self.cfg,
+                trainer=trainer,
+                optimizer=optimizer,
+                num_training_steps=num_training_steps,
+            )
             if scheduler is not None:
                 return scheduler
         return None

src/axolotl/integrations/cut_cross_entropy/__init__.py

@@ -25,7 +25,7 @@ import torch
 
 from axolotl.integrations.base import BasePlugin
 from axolotl.utils import get_pytorch_version
-from axolotl.utils.distributed import zero_only
+from axolotl.utils.distributed import is_main_process
 
 from .args import CutCrossEntropyArgs  # pylint: disable=unused-import. # noqa: F401
 
@@ -76,7 +76,7 @@ class CutCrossEntropyPlugin(BasePlugin):
                 cce_patch,
             )
 
-            with zero_only():
+            if is_main_process(use_environ=True):
                 LOG.info(
                     f"Applying Cut Cross Entropy to model type: {cfg.model_config_type}"
                 )

src/axolotl/integrations/kd/chat_template.py

@@ -37,6 +37,7 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
         train_on_eos=None,
         train_on_eot=None,
         eot_tokens=None,
+        split_thinking: bool | None = False,
         logprobs_field="logprobs",
         gen_temperature=1.0,
         kd_temperature=1.0,
@@ -54,6 +55,7 @@ class ChatTemplateStrategyWithKD(ChatTemplateStrategy):
             train_on_eos=train_on_eos,
             train_on_eot=train_on_eot,
             eot_tokens=eot_tokens,
+            split_thinking=split_thinking,
         )
 
     @property

src/axolotl/integrations/liger/__init__.py

@@ -23,8 +23,8 @@ import logging
 import sys
 
 from axolotl.integrations.base import BasePlugin
+from axolotl.utils.distributed import is_main_process
 
-from ...utils.distributed import zero_only
 from .args import LigerArgs  # pylint: disable=unused-import. # noqa: F401
 from .utils import patch_with_compile_disable
 
@@ -85,7 +85,7 @@ class LigerPlugin(BasePlugin):
                 kwargs["geglu"] = cfg.liger_glu_activation
             elif "swiglu" in liger_fn_sig.parameters:
                 kwargs["swiglu"] = cfg.liger_glu_activation
-            with zero_only():
+            if is_main_process(use_environ=True):
                 LOG.info(
                     f"Applying LIGER to {cfg.model_config_type} with kwargs: {kwargs}"
                 )

src/axolotl/integrations/llm_compressor/README.md

@@ -0,0 +1,108 @@
+# 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

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

src/axolotl/integrations/llm_compressor/args.py

@@ -0,0 +1,40 @@
+"""
+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

@@ -0,0 +1,171 @@
+"""
+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

@@ -0,0 +1,40 @@
+"""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/lora_kernels.py

@@ -23,22 +23,42 @@ from axolotl.utils.dict import DictDefault
 
 LOG = get_logger(__name__)
 
-ORIGINAL_QKV_CODE = """
+QKV_PATCHES = [
+    (
+        """
     query_states = self.q_proj(hidden_states).view(hidden_shape).transpose(1, 2)
     key_states = self.k_proj(hidden_states).view(hidden_shape).transpose(1, 2)
     value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
 """.lstrip(
-    "\n"
-)
-
-PATCHED_QKV_CODE = """
+            "\n"
+        ),
+        """
     query_states, key_states, value_states = self.apply_qkv(hidden_states)
     query_states = query_states.view(hidden_shape).transpose(1, 2)
     key_states = key_states.view(hidden_shape).transpose(1, 2)
     value_states = value_states.view(hidden_shape).transpose(1, 2)
 """.lstrip(
-    "\n"
-)
+            "\n"
+        ),
+    ),
+    (
+        """
+    query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+    key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
+    value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2)
+""".lstrip(
+            "\n"
+        ),
+        """
+    query_states, key_states, value_states = self.apply_qkv(hidden_states)
+    query_states = self.q_norm(query_states.view(hidden_shape)).transpose(1, 2)
+    key_states = self.k_norm(key_states.view(hidden_shape)).transpose(1, 2)
+    value_states = value_states.view(hidden_shape).transpose(1, 2)
+""".lstrip(
+            "\n"
+        ),
+    ),
+]
 
 ORIGINAL_O_CODE = """
     attn_output = self.o_proj(attn_output)
@@ -128,10 +148,11 @@ def get_attention_cls_from_config(cfg: DictDefault) -> Type[nn.Module]:
     try:
         # Dynamically import the module and attention class
         module_path = f"transformers.models.{model_type}.modeling_{model_type}"
-        module = __import__(
-            module_path, fromlist=[f"{model_type.capitalize()}Attention"]
+        model_cls_prefix = "".join(
+            [part.capitalize() for part in model_type.split("_")]
         )
-        attention_cls = getattr(module, f"{model_type.capitalize()}Attention")
+        module = __import__(module_path, fromlist=[f"{model_cls_prefix}Attention"])
+        attention_cls = getattr(module, f"{model_cls_prefix}Attention")
 
         return attention_cls
     except (ImportError, AttributeError) as e:
@@ -168,10 +189,18 @@ def patch_self_attn_lora(cfg: DictDefault):
     attention_cls._original_forward = self_attn_forward
     self_attn_forward, _ = detab_code(self_attn_forward)
 
-    assert ORIGINAL_QKV_CODE in self_attn_forward, "Original QKV code not found"
+    assert any(
+        qkv_options[0] in self_attn_forward for qkv_options in QKV_PATCHES
+    ), "Original QKV code not found"
     assert ORIGINAL_O_CODE in self_attn_forward, "Original O code not found"
 
-    self_attn_forward = self_attn_forward.replace(ORIGINAL_QKV_CODE, PATCHED_QKV_CODE)
+    for qkv_orig, qkv_patched in QKV_PATCHES:
+        if qkv_orig in self_attn_forward:
+            self_attn_forward = self_attn_forward.replace(
+                qkv_orig,
+                qkv_patched,
+            )
+            break
     self_attn_forward = self_attn_forward.replace(ORIGINAL_O_CODE, PATCHED_O_CODE)
     self_attn_forward = self_attn_forward.replace(
         "def forward(",

src/axolotl/monkeypatch/trainer/lr.py

@@ -0,0 +1,42 @@
+"""
+monkeypatch for Trainer _get_learning_rate method
+"""
+
+import logging
+
+import torch
+
+LOG = logging.getLogger(__name__)
+
+
+# TODO remove this patch once https://github.com/huggingface/transformers/pull/37881 is included in a release
+def _get_learning_rate(self):
+    if self.is_deepspeed_enabled:
+        # with deepspeed's fp16 and dynamic loss scale enabled the optimizer/scheduler steps may
+        # not run for the first few dozen steps while loss scale is too large, and thus during
+        # that time `get_last_lr` will fail if called during that warm up stage, so work around it:
+        try:
+            last_lr = self.lr_scheduler.get_last_lr()[0]
+        except AssertionError as e:
+            if "need to call step" in str(e):
+                LOG.warning(
+                    "tried to get lr value before scheduler/optimizer started stepping, returning lr=0"
+                )
+                last_lr = 0
+            else:
+                raise
+    else:
+        if isinstance(self.lr_scheduler, torch.optim.lr_scheduler.ReduceLROnPlateau):
+            last_lr = self.optimizer.param_groups[0]["lr"]
+        else:
+            last_lr = self.lr_scheduler.get_last_lr()[0]
+
+    if torch.is_tensor(last_lr):
+        last_lr = last_lr.item()
+    return last_lr
+
+
+def patch_trainer_get_lr():
+    from transformers.trainer import Trainer
+
+    Trainer._get_learning_rate = _get_learning_rate  # pylint: disable=protected-access

src/axolotl/prompt_strategies/chat_template.py

@@ -4,7 +4,7 @@ HF Chat Templates prompt strategy
 
 import logging
 from collections import defaultdict
-from typing import Any, Dict, List, Optional, Set, Union
+from typing import Any, Dict, List, Set, Union
 
 from pydantic import BaseModel
 from transformers import ProcessorMixin
@@ -29,12 +29,12 @@ class ChatTemplatePrompter(Prompter):
         chat_template: str,
         processor=None,
         max_length=2048,
-        message_property_mappings: Optional[Dict[str, str]] = None,
-        message_field_training: Optional[str] = None,
-        message_field_training_detail: Optional[str] = None,
+        message_property_mappings: Dict[str, str] | None = None,
+        message_field_training: str | None = None,
+        message_field_training_detail: str | None = None,
         field_messages: str = "messages",
         field_system: str = "system",
-        roles: Optional[Dict[str, List[str]]] = None,
+        roles: Dict[str, List[str]] | None = None,
         drop_system_message: bool = False,
     ):
         # check if message_property_mappings is None or empty dict
@@ -42,6 +42,7 @@ class ChatTemplatePrompter(Prompter):
             message_property_mappings = {
                 "role": "role",
                 "content": "content",
+                "reasoning_content": "reasoning_content",
             }
 
         if roles:
@@ -65,7 +66,7 @@ class ChatTemplatePrompter(Prompter):
         self.field_messages = field_messages
         self.field_system = field_system
         self.tokenizer = tokenizer
-        self.processor: Optional[ProcessorMixin] = processor
+        self.processor: ProcessorMixin | None = processor
         self.chat_template = chat_template
         self.max_length = max_length
         self.drop_system_message = drop_system_message
@@ -224,11 +225,11 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
         tokenizer,
         train_on_inputs: bool,
         sequence_len: int,
-        roles_to_train: Optional[List[str]] = None,
-        train_on_eos: Optional[str] = None,
-        train_on_eot: Optional[str] = None,
-        eot_tokens: Optional[List[str]] = None,
-        split_thinking: Optional[bool] = False,
+        roles_to_train: list[str] | None = None,
+        train_on_eos: str | None = None,
+        train_on_eot: str | None = None,
+        eot_tokens: list[str] | None = None,
+        split_thinking: bool | None = False,
     ):
         super().__init__(prompter, tokenizer, train_on_inputs, sequence_len)
         self.prompter: ChatTemplatePrompter = prompter
@@ -661,16 +662,46 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
         # if the role is assistant that we want to use reasoning_content
         if self.split_thinking and transformed_message["role"] == "assistant":
             content = transformed_message["content"]
-            pairs = [("<think>", "</think>"), ("<reasoning>", "</reasoning>")]
-            for pair in pairs:
-                if pair[0] in content and pair[1] in content:
-                    start_idx = content.find(pair[0])
-                    end_idx = content.find(pair[1])
-                    thinking_content = content[start_idx + len(pair[0]) : end_idx]
+            thinking_pairs = [
+                ("<think>", "</think>"),
+                ("<reasoning>", "</reasoning>"),
+                ("<|begin_of_thought|>", "<|end_of_thought|>"),
+            ]
+            content_pairs = [("<|begin_of_solution|>", "<|end_of_solution|>")]
+            for tpair in thinking_pairs:
+                # check if the thinking pair is in the content
+                if tpair[0] in content and tpair[1] in content:
+                    # find the start and end index of the thinking pair
+                    t_start_idx = content.find(tpair[0])
+                    t_end_idx = content.find(tpair[1])
+
+                    # get the thinking content
+                    thinking_content = content[t_start_idx + len(tpair[0]) : t_end_idx]
                     transformed_message["reasoning_content"] = thinking_content.strip()
-                    transformed_message["content"] = content[
-                        end_idx + len(pair[1]) :
-                    ].lstrip()
+
+                    # take remainder of the content
+                    # strip whitespace from beginning of the remainder (thinking tokens)
+                    remainder = content[t_end_idx + len(tpair[1]) :].lstrip()
+
+                    # check if the content pair is in the remainder
+                    cpair_found = False
+                    for cpair in content_pairs:
+                        if cpair[0] in remainder and cpair[1] in remainder:
+                            # find the start and end index of the content pair
+                            c_start_idx = remainder.find(cpair[0])
+                            c_end_idx = remainder.find(cpair[1])
+
+                            # get the content content
+                            content_content = remainder[
+                                c_start_idx + len(cpair[0]) : c_end_idx
+                            ]
+                            transformed_message["content"] = content_content.strip()
+                            cpair_found = True
+                            break
+
+                    # else, the content is the remainder
+                    if not cpair_found:
+                        transformed_message["content"] = remainder
                     break
 
         # Determine which keys in the original message were not mapped
@@ -714,7 +745,7 @@ class StrategyLoader:
         self,
         tokenizer,
         cfg,
-        ds_cfg: Optional[Union[Dict[str, Any], DatasetConfig]] = None,
+        ds_cfg: Union[Dict[str, Any], DatasetConfig] | None = None,
         processor=None,
     ):
         if ds_cfg is None:

src/axolotl/train.py

@@ -296,8 +296,23 @@ 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/callbacks/__init__.py

@@ -3,6 +3,7 @@
 from __future__ import annotations
 
 import gc
+import json
 import logging
 import os
 import traceback
@@ -808,11 +809,44 @@ class SaveAxolotlConfigtoWandBCallback(TrainerCallback):
                     artifact.add_file(temp_file.name)
                     wandb.log_artifact(artifact)
                     wandb.save(temp_file.name)
-                LOG.info(
-                    "The Axolotl config has been saved to the WandB run under files."
-                )
+                    LOG.info(
+                        "The Axolotl config has been saved to the WandB run under files."
+                    )
             except (FileNotFoundError, ConnectionError) as err:
                 LOG.warning(f"Error while saving Axolotl config to WandB: {err}")
+
+            if args.deepspeed:
+                try:
+                    # sync config to top level in run, cannot delete file right away because wandb schedules it to be synced even w/policy = 'now', so let OS delete it later.
+                    with NamedTemporaryFile(
+                        mode="w",
+                        delete=False,
+                        suffix=".json",
+                        prefix="deepspeed_config_",
+                    ) as temp_file:
+                        skip_upload = False
+                        if isinstance(args.deepspeed, dict):
+                            json.dump(args.deepspeed, temp_file, indent=4)
+                        elif isinstance(args.deepspeed, str) and os.path.exists(
+                            args.deepspeed
+                        ):
+                            copyfile(args.deepspeed, temp_file.name)
+                        else:
+                            skip_upload = True
+                        if not skip_upload:
+                            artifact = wandb.Artifact(
+                                f"deepspeed-config-{wandb.run.id}",
+                                type="deepspeed-config",
+                            )
+                            artifact.add_file(temp_file.name)
+                            wandb.log_artifact(artifact)
+                            wandb.save(temp_file.name)
+                            LOG.info(
+                                "The DeepSpeed config has been saved to the WandB run under files."
+                            )
+                except (FileNotFoundError, ConnectionError) as err:
+                    LOG.warning(f"Error while saving DeepSpeed config to WandB: {err}")
+
         return control
 
 

src/axolotl/utils/distributed.py

@@ -69,17 +69,27 @@ def barrier():
         dist.barrier()
 
 
-def is_main_process():
+def is_main_process(use_environ=False):
     """
     Check if the current process is the main process. If not in distributed mode,
     always return `True`.
+
+    Args:
+    - use_environ (bool, optional): Use environment variable to determine main process.
+
+    Returns:
+    - bool: `True` if the current process is the main process, `False` otherwise.
     """
+    if use_environ:
+        return os.environ.get("LOCAL_RANK", "0") == "0"
     if not is_distributed():
         return True
     return dist.get_rank() == 0
 
 
-def is_local_main_process():
+def is_local_main_process(use_environ=False):
+    if use_environ:
+        return os.environ.get("LOCAL_RANK", "0") == "0"
     return PartialState().is_local_main_process
 
 
@@ -99,17 +109,6 @@ def cleanup_distributed():
         torch.distributed.destroy_process_group()
 
 
-@contextmanager
-def zero_only():
-    """
-    Context manager that only runs the enclosed block on the main rank.
-    """
-    if is_main_process():
-        yield
-    else:
-        yield None
-
-
 @contextmanager
 def zero_first(is_main):
     """

src/axolotl/utils/models.py

@@ -68,7 +68,7 @@ from axolotl.utils.distributed import (
     get_device_count,
     get_device_type,
     is_local_main_process,
-    zero_only,
+    is_main_process,
 )
 from axolotl.utils.gradient_checkpointing import hf_grad_checkpoint_offload_wrapper
 from axolotl.utils.lora_embeddings import get_linear_embedding_layers
@@ -141,6 +141,22 @@ 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
@@ -437,7 +453,7 @@ def load_tokenizer(cfg):
             {"additional_special_tokens": additional_special_tokens}
         )
 
-    with zero_only():
+    if is_main_process(use_environ=True):
         LOG.debug(f"EOS: {tokenizer.eos_token_id} / {tokenizer.eos_token}")
         LOG.debug(f"BOS: {tokenizer.bos_token_id} / {tokenizer.bos_token}")
         LOG.debug(f"PAD: {tokenizer.pad_token_id} / {tokenizer.pad_token}")

src/axolotl/utils/schemas/config.py

@@ -1150,6 +1150,18 @@ class AxolotlInputConfig(
 
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def check_grpo_peft_liger(cls, data):
+        if (
+            data.get("rl") == "grpo"
+            and data.get("trl", {})
+            and data.get("trl").get("use_liger_loss")
+            and data.get("adapter")
+        ):
+            raise ValueError("PEFT + GRPO + Liger is not yet supported")
+        return data
+
     @model_validator(mode="after")
     def check_sequence_parallel_degree(self):
         if not self.sequence_parallel_degree:
@@ -1315,6 +1327,57 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
                     )
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def check_auto_enable_lora_kernels(cls, data):
+        # Only proceed if using LoRA or QLoRA adapter
+        if data.get("rl"):
+            # RL trainers not tested so don't enable kernels by default
+            return data
+        if data.get("adapter") in ["lora", "qlora"]:
+            # Skip if already set, using unsloth optimizations, or using 8-bit
+            unsloth_fields = ["unsloth_lora_mlp", "unsloth_lora_qkv", "unsloth_lora_o"]
+            kernel_fields = ["lora_mlp_kernel", "lora_qkv_kernel", "lora_o_kernel"]
+            if (
+                any(data.get(k) is not None for k in kernel_fields)
+                or any(data.get(k) for k in unsloth_fields)
+                or data.get("adapter") == "lora"
+                and data.get("load_in_8bit")
+            ):
+                return data
+
+            # Check multi-GPU compatibility
+            capabilities = data.get("capabilities")
+            is_multi_gpu = capabilities and capabilities.get("n_gpu", 0) > 1
+            is_fsdp = data.get("fsdp") is not None
+            is_fsdp2 = (
+                data.get("fsdp_config") is not None
+                and str(data.get("fsdp_config").get("fsdp_version")) == "2"
+            )
+
+            if (
+                not is_multi_gpu
+                or (is_multi_gpu and not is_fsdp)
+                or (is_multi_gpu and is_fsdp2)
+            ):
+                # Auto-enable kernels if not explicitly set by user
+                if data.get("lora_mlp_kernel") is None:
+                    data["lora_mlp_kernel"] = True
+
+                if data.get("lora_qkv_kernel") is None:
+                    data["lora_qkv_kernel"] = True
+
+                if data.get("lora_o_kernel") is None:
+                    data["lora_o_kernel"] = True
+
+                LOG.warning(
+                    "Auto-enabling LoRA kernel optimizations for faster training. "
+                    + "Please explicitly set `lora_*_kernel` config values to `false` to disable. "
+                    + "See https://docs.axolotl.ai/docs/lora_optims.html for more info."
+                )
+
+        return data
+
     @model_validator(mode="before")
     @classmethod
     def check_adopt_torch_version(cls, data):

src/axolotl/utils/schemas/trl.py

@@ -133,3 +133,25 @@ class TRLConfig(BaseModel):
             "description": "Epsilon value for clipping in the GRPO algorithm."
         },
     )
+    epsilon_high: float | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Upper-bound epsilon value for clipping in the GRPO algorithm."
+        },
+    )
+    use_liger_loss: bool | None = Field(
+        default=None,
+        json_schema_extra={"description": "Whether to use Liger loss for GRPO."},
+    )
+    loss_type: str | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Specifies the loss formulation to use. Supported values are `grpo`, `bnpo`, and `dr_grpo`."
+        },
+    )
+    mask_truncated_completions: bool = Field(
+        default=False,
+        json_schema_extra={
+            "description": "When enabled, truncated completions are excluded from the loss calculation."
+        },
+    )

tests/e2e/integrations/test_hooks.py

@@ -72,7 +72,7 @@ class LogHooksPlugin(BasePlugin):
             f.write("get_trainer_cls\n")
 
     def create_lr_scheduler(
-        self, cfg, trainer, optimizer
+        self, cfg, trainer, optimizer, num_training_steps
     ):  # pylint: disable=unused-argument
         with open(
             self.base_dir.joinpath("plugin_hooks.log"), "a", encoding="utf-8"
@@ -172,7 +172,7 @@ class TestPluginHooks:
             assert "post_model_load" in file_contents
             # assert "create_optimizer" in file_contents  # not implemented yet
             assert "get_trainer_cls" in file_contents
-            # assert "create_lr_scheduler" in file_contents  # not implemented yet
+            assert "create_lr_scheduler" in file_contents
             assert "add_callbacks_pre_trainer" in file_contents
             assert "add_callbacks_post_trainer" in file_contents
             assert "post_train" in file_contents

tests/e2e/integrations/test_llm_compressor.py

@@ -0,0 +1,111 @@
+"""
+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/patched/lora_kernels/test_lora_kernel_patching.py

@@ -2,14 +2,19 @@
 
 # pylint: disable=redefined-outer-name
 
+from pathlib import Path
+
 import pytest
 import torch
+import yaml
 from accelerate.state import PartialState
 from peft import PeftModelForCausalLM, get_peft_config
 from transformers import AutoModelForCausalLM, LlamaForCausalLM
 from transformers.models.llama.configuration_llama import LlamaConfig
 from transformers.models.llama.modeling_llama import LlamaAttention
+from transformers.models.qwen3_moe.modeling_qwen3_moe import Qwen3MoeAttention
 
+from axolotl.cli.config import load_cfg
 from axolotl.kernels.lora import (
     apply_lora_mlp_geglu,
     apply_lora_mlp_swiglu,
@@ -66,29 +71,36 @@ def small_llama_model():
     return LlamaForCausalLM(LlamaConfig(**config))
 
 
-def test_attention_patching_integration():
+@pytest.mark.parametrize(
+    "model_name,attention_cls",
+    [
+        ("HuggingFaceTB/SmolLM2-135M", LlamaAttention),
+        ("Qwen/Qwen3-30B-A3B", Qwen3MoeAttention),
+    ],
+)
+def test_attention_patching_integration(model_name, attention_cls):
     """Test attention patching in integration context."""
-    cfg = {"base_model": "TinyLlama/TinyLlama-1.1B-Chat-v1.0"}
+    cfg = {"base_model": model_name}
 
     # Store the original implementation
-    original_forward = getattr(LlamaAttention, "forward")
+    original_forward = getattr(attention_cls, "forward")
 
     # Apply patch
     patch_self_attn_lora(cfg)
 
     # Get the new forward method
-    patched_forward = LlamaAttention.forward
+    patched_forward = attention_cls.forward
 
     # Check the forward method was replaced
     assert original_forward is not patched_forward
     assert patched_forward.__name__ == "axolotl_attn_forward"
 
     # Check original implementation was stored
-    assert hasattr(LlamaAttention, "_original_forward")
+    assert hasattr(attention_cls, "_original_forward")
 
     # Clean up
-    setattr(LlamaAttention, "forward", original_forward)
-    delattr(LlamaAttention, "_original_forward")
+    setattr(attention_cls, "forward", original_forward)
+    delattr(attention_cls, "_original_forward")
 
 
 def test_swiglu_mlp_integration(small_llama_model):
@@ -413,3 +425,42 @@ def test_kernel_training_integration():
     # Verify correct activation function
     layer = model.model.model.layers[0]
     assert layer.mlp.forward.__func__ is apply_lora_mlp_swiglu
+
+
+def test_kernel_training_integration_auto_enable(temp_dir):
+    """Test model loading with auto-enabled kernel patches."""
+    # Create minimal config without explicitly setting kernel options
+    cfg = DictDefault(
+        {
+            "base_model": "HuggingFaceTB/SmolLM2-135M",
+            "tokenizer_config": "HuggingFaceTB/SmolLM2-135M",
+            "learning_rate": 0.000001,
+            "datasets": [
+                {
+                    "path": "mhenrichsen/alpaca_2k_test",
+                    "type": "alpaca",
+                }
+            ],
+            "micro_batch_size": 1,
+            "gradient_accumulation_steps": 1,
+            "adapter": "lora",
+            "lora_r": 8,
+            "lora_alpha": 16,
+            "lora_dropout": 0.0,
+            "lora_target_linear": True,
+            "sequence_len": 1024,
+        }
+    )
+
+    # Write cfg to yaml file
+    path = Path(temp_dir) / "config.yaml"
+    with open(path, "w", encoding="utf-8") as fout:
+        fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
+
+    # Load config
+    cfg = load_cfg(str(path))
+
+    # Verify kernel options were auto-enabled in the config
+    assert cfg.lora_mlp_kernel is True
+    assert cfg.lora_qkv_kernel is True
+    assert cfg.lora_o_kernel is True

tests/e2e/test_evaluate.py

@@ -0,0 +1,65 @@
+"""E2E smoke test for evaluate CLI command"""
+
+import os
+from pathlib import Path
+
+import yaml
+from accelerate.test_utils import execute_subprocess_async
+from transformers.testing_utils import get_torch_dist_unique_port
+
+from axolotl.utils.dict import DictDefault
+
+os.environ["WANDB_DISABLED"] = "true"
+
+
+class TestE2eEvaluate:
+    """Test cases for evaluate CLI"""
+
+    def test_evaluate(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
+                "sequence_len": 1024,
+                "val_set_size": 0.02,
+                "special_tokens": {
+                    "unk_token": "<unk>",
+                    "bos_token": "<s>",
+                    "eos_token": "</s>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 8,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch_fused",
+                "lr_scheduler": "cosine",
+                "max_steps": 20,
+            }
+        )
+
+        # write cfg to yaml file
+        Path(temp_dir).mkdir(parents=True, exist_ok=True)
+        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
+            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
+
+        execute_subprocess_async(
+            [
+                "accelerate",
+                "launch",
+                "--num-processes",
+                "2",
+                "--main_process_port",
+                f"{get_torch_dist_unique_port()}",
+                "-m",
+                "axolotl.cli.evaluate",
+                str(Path(temp_dir) / "config.yaml"),
+            ]
+        )

tests/e2e/utils.py

@@ -105,7 +105,25 @@ def require_vllm(test_case):
             return False
 
     return unittest.skipUnless(
-        is_vllm_installed(), "test requires a vllm to be installed"
+        is_vllm_installed(), "test requires 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)
 
 

tests/prompt_strategies/test_chat_templates_thinking.py

@@ -34,7 +34,31 @@ def messages_w_reasoning_fixture():
                         "content": "<think>lorem</think>\nwelcome",
                     },
                 ]
-            }
+            },
+            {
+                "messages": [
+                    {
+                        "role": "user",
+                        "content": "hello",
+                    },
+                    {
+                        "role": "assistant",
+                        "content": "<|begin_of_thought|>lorem<|end_of_thought|>\n<|begin_of_solution|>welcome\n<|end_of_solution|>",
+                    },
+                ]
+            },
+            {
+                "messages": [
+                    {
+                        "role": "user",
+                        "content": "hello",
+                    },
+                    {
+                        "role": "assistant",
+                        "content": "<reasoning>lorem</reasoning>\nwelcome",
+                    },
+                ]
+            },
         ]
     )
 
@@ -83,36 +107,37 @@ class TestSplitThinking:
                 }
             ),
         )
-        transformed_prompt = strategy.get_conversation_thread(messages_w_reasoning[0])
-        assert transformed_prompt[0]["role"] == "user"
-        assert transformed_prompt[1]["role"] == "assistant"
-        assert transformed_prompt[1]["reasoning_content"] == "lorem"
-        assert transformed_prompt[1]["content"] == "welcome"
+        for conversation in messages_w_reasoning:
+            transformed_prompt = strategy.get_conversation_thread(conversation)
+            assert transformed_prompt[0]["role"] == "user"
+            assert transformed_prompt[1]["role"] == "assistant"
+            assert transformed_prompt[1]["reasoning_content"] == "lorem"
+            assert transformed_prompt[1]["content"] == "welcome"
 
-        res = strategy.tokenize_prompt(messages_w_reasoning[0])
-        input_ids = res["input_ids"]
-        # fmt: off
-        expected_input_ids = [
-            151644,  # im_start
-            872,  # user
-            198,  # \n
-            14990,  # hello
-            151645,  # im_end
-            198,  # \n
-            151644,  # im_start
-            77091,  # assistant
-            198,  # \n
-            151667,  # think
-            198,  # \n
-            385, 1826,  # lorem
-            198,  # \n
-            151668,  # /think
-            271,  # \n
-            34084,  # welcome
-            151645,  # im_end
-            198,  # \n
-        ]
-        # fmt: on
-        assert (
-            input_ids == expected_input_ids
-        ), f"Input IDs mismatch: {input_ids} != {expected_input_ids}"
+            res = strategy.tokenize_prompt(conversation)
+            input_ids = res["input_ids"]
+            # fmt: off
+            expected_input_ids = [
+                151644,  # im_start
+                872,  # user
+                198,  # \n
+                14990,  # hello
+                151645,  # im_end
+                198,  # \n
+                151644,  # im_start
+                77091,  # assistant
+                198,  # \n
+                151667,  # think
+                198,  # \n
+                385, 1826,  # lorem
+                198,  # \n
+                151668,  # /think
+                271,  # \n
+                34084,  # welcome
+                151645,  # im_end
+                198,  # \n
+            ]
+            # fmt: on
+            assert (
+                input_ids == expected_input_ids
+            ), f"Input IDs mismatch: {input_ids} != {expected_input_ids}"