fix: missing key in enum

* feat: update handling for mistraltokenizer decode

* fix: update mistral common package version

* fix: to use correct release

* fix triton path

---------

Co-authored-by: Wing Lian <wing@axolotl.ai>

.github/workflows/main.yml

@@ -20,12 +20,11 @@ jobs:
             python_version: "3.11"
             pytorch: 2.5.1
             axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras: vllm
-            is_latest: true
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
@@ -88,8 +87,8 @@ jobs:
             python_version: "3.11"
             pytorch: 2.5.1
             axolotl_extras:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:
@@ -146,8 +145,8 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:

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

@@ -26,11 +26,11 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
-            axolotl_extras: vllm
+            axolotl_extras:
             num_gpus: 2
             nightly_build: "true"
           - cuda: 124

.github/workflows/tests.yml

@@ -195,12 +195,12 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
             num_gpus: 1
-            axolotl_extras: vllm
+            axolotl_extras:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
@@ -247,8 +247,8 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.6.0
             num_gpus: 1
@@ -311,7 +311,7 @@ jobs:
             python_version: "3.11"
             pytorch: 2.6.0
             num_gpus: 1
-            axolotl_extras: vllm
+            axolotl_extras:
     steps:
       - name: Checkout
         uses: actions/checkout@v4

.pre-commit-config.yaml

@@ -19,7 +19,7 @@ repos:
     hooks:
       - id: isort
 -   repo: https://github.com/PyCQA/flake8
-    rev: 7.2.0
+    rev: 7.3.0
     hooks:
     - id: flake8
 -   repo: https://github.com/pylint-dev/pylint
@@ -27,7 +27,7 @@ repos:
     hooks:
     - id: pylint
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.16.0
+    rev: v1.16.1
     hooks:
     - id: mypy
       additional_dependencies:
@@ -36,7 +36,7 @@ repos:
             'pydantic>=2.5.3',
         ]
 -   repo: https://github.com/PyCQA/bandit
-    rev: 1.8.3
+    rev: 1.8.5
     hooks:
     -   id: bandit
         args: [

README.md

@@ -43,7 +43,7 @@ Features:
 - **Multiple Model Support**: Train various models like LLaMA, Mistral, Mixtral, Pythia, and more. We are compatible with HuggingFace transformers causal language models.
 - **Training Methods**: Full fine-tuning, LoRA, QLoRA, GPTQ, QAT, Preference Tuning (DPO, IPO, KTO, ORPO), RL (GRPO), Multimodal, and Reward Modelling (RM) / Process Reward Modelling (PRM).
 - **Easy Configuration**: Re-use a single YAML file between dataset preprocess, training, evaluation, quantization, and inference.
-- **Performance Optimizations**: [Multipacking](https://docs.axolotl.ai/docs/multipack.html), [Flash Attention](https://github.com/Dao-AILab/flash-attention), [Xformers](https://github.com/facebookresearch/xformers), [Flex Attention](https://pytorch.org/blog/flexattention/), [Liger Kernel](https://github.com/linkedin/Liger-Kernel), [Cut Cross Entropy](https://github.com/apple/ml-cross-entropy/tree/main), Sequence Parallelism (SP), LoRA optimizations, Multi-GPU training (FSDP1, FSDP2, DeepSpeed), Multi-node training (Torchrun, Ray), and many more!
+- **Performance Optimizations**: [Multipacking](https://docs.axolotl.ai/docs/multipack.html), [Flash Attention](https://github.com/Dao-AILab/flash-attention), [Xformers](https://github.com/facebookresearch/xformers), [Flex Attention](https://pytorch.org/blog/flexattention/), [Liger Kernel](https://github.com/linkedin/Liger-Kernel), [Cut Cross Entropy](https://github.com/apple/ml-cross-entropy/tree/main), [Sequence Parallelism (SP)](https://docs.axolotl.ai/docs/sequence_parallelism.html), [LoRA optimizations](https://docs.axolotl.ai/docs/lora_optims.html), [Multi-GPU training (FSDP1, FSDP2, DeepSpeed)](https://docs.axolotl.ai/docs/multi-gpu.html), [Multi-node training (Torchrun, Ray)](https://docs.axolotl.ai/docs/multi-node.html), and many more!
 - **Flexible Dataset Handling**: Load from local, HuggingFace, and cloud (S3, Azure, GCP, OCI) datasets.
 - **Cloud Ready**: We ship [Docker images](https://hub.docker.com/u/axolotlai) and also [PyPI packages](https://pypi.org/project/axolotl/) for use on cloud platforms and local hardware.
 
@@ -59,6 +59,8 @@ Features:
 
 ### Installation
 
+#### Using pip
+
 ```bash
 pip3 install -U packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
@@ -68,6 +70,13 @@ axolotl fetch examples
 axolotl fetch deepspeed_configs  # OPTIONAL
 ```
 
+#### Using Docker
+
+Installing with Docker can be less error prone than installing in your own environment.
+```bash
+docker run --gpus '"all"' --rm -it axolotlai/axolotl:main-latest
+```
+
 Other installation approaches are described [here](https://docs.axolotl.ai/docs/installation.html).
 
 ### Your First Fine-tune

cicd/single_gpu.py

@@ -32,6 +32,8 @@ df_args = {
     "NIGHTLY_BUILD": os.environ.get("NIGHTLY_BUILD", ""),
     "CODECOV_TOKEN": os.environ.get("CODECOV_TOKEN", ""),
     "HF_HOME": "/workspace/data/huggingface-cache/hub",
+    "PYTHONUNBUFFERED": os.environ.get("PYTHONUNBUFFERED", "1"),
+    "DEEPSPEED_LOG_LEVEL": os.environ.get("DEEPSPEED_LOG_LEVEL", "WARNING"),
 }
 
 dockerfile_contents = df_template.render(**df_args)

docker/Dockerfile-base

@@ -37,7 +37,3 @@ RUN git lfs install --skip-repo && \
     pip3 install awscli && \
     # The base image ships with `pydantic==1.8.2` which is not working
     pip3 install -U --no-cache-dir pydantic==1.10.10
-
-RUN if [ "$PYTORCH_VERSION" = "2.7.1" ] ; then \
-        pip3 install flash-attn==2.7.4.post1; \
-    fi

docker/Dockerfile-uv-base

@@ -34,7 +34,3 @@ RUN uv pip install packaging setuptools wheel psutil \
     && uv pip install --no-build-isolation "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" \
     && uv pip install "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" \
     && uv pip install awscli pydantic
-
-RUN if [ "$PYTORCH_VERSION" = "2.7.1" ] ; then \
-        uv pip install --no-build-isolation flash-attn==2.7.4.post1; \
-    fi

docs/dataset-formats/conversation.qmd

@@ -9,7 +9,7 @@ order: 3
 Chat Template strategy uses a jinja2 template that converts a list of messages into a prompt. Support using tokenizer's template, a supported template, or custom jinja2.
 
 ```{.json filename="data.jsonl"}
-{"conversations": [{"role": "...", "content": "..."}]}
+{"messages": [{"role": "...", "content": "..."}, {"role": "...", "content": "..."}, ...]}
 ```
 
 See [configs](../config-reference.qmd) for full configs and supported templates.

docs/docker.qmd

@@ -9,7 +9,7 @@ format:
 This section describes the different Docker images that are released by AxolotlAI at [Docker Hub](https://hub.docker.com/u/axolotlai).
 
 ::: {.callout-important}
-For Blackwell GPUs, please use the tags with Pytorch 2.7.1 and CUDA 12.8.
+For Blackwell GPUs, please use the tags with PyTorch 2.7.1 and CUDA 12.8.
 :::
 
 ## Base
@@ -34,6 +34,7 @@ Tags examples:
 
 - `main-base-py3.11-cu128-2.7.1`
 - `main-base-py3.11-cu126-2.7.1`
+- `main-base-py3.11-cu126-2.6.0`
 - `main-base-py3.11-cu124-2.6.0`
 - `main-base-py3.11-cu124-2.5.1`
 
@@ -73,13 +74,15 @@ There may be some extra tags appended to the image, like `-vllm` which installs
 
 Tags examples:
 
-- `main-py3.11-cu126-2.7.0`
+- `main-py3.11-cu128-2.7.1`
+- `main-py3.11-cu126-2.7.1`
+- `main-py3.11-cu126-2.6.0`
 - `main-py3.11-cu124-2.6.0`
 - `main-py3.11-cu124-2.5.1`
 - `main-latest`
 - `main-20250303-py3.11-cu124-2.6.0`
 - `main-20250303-py3.11-cu124-2.5.1`
-- `0.9.2`
+- `0.10.1`
 
 ## Cloud
 

docs/multimodal.qmd

@@ -16,6 +16,7 @@ format:
 - [Gemma-3](#sec-gemma-3)
 - [Qwen2-VL](#sec-qwen2-vl)
 - [Qwen2.5-VL](#sec-qwen25-vl)
+- [Phi3-V](#sec-phi3-v)
 
 ## Usage
 
@@ -126,6 +127,15 @@ base_model: Qwen/Qwen2.5-VL-7B-Instruct
 chat_template: qwen2_vl  # same as qwen2-vl
 ```
 
+### Phi3-V {#sec-phi3-v}
+
+```yaml
+base_model: microsoft/Phi-3.5-vision-instruct
+
+trust_remote_code: true
+chat_template: phi_35_vl
+```
+
 ## Dataset Format
 
 For multi-modal datasets, we adopt an extended `chat_template` format similar to OpenAI's Message format.

requirements.txt

@@ -1,7 +1,7 @@
 --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
 
 # START section of dependencies that don't install on Darwin/MacOS
-bitsandbytes==0.45.4
+bitsandbytes==0.46.0
 triton>=3.0.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
@@ -15,7 +15,7 @@ huggingface_hub==0.32.2
 peft==0.15.2
 transformers==4.52.4
 tokenizers>=0.21.1
-accelerate==1.7.0
+accelerate==1.8.1
 datasets==3.6.0
 deepspeed>=0.17.0
 trl==0.18.2
@@ -68,4 +68,4 @@ schedulefree==1.4.1
 axolotl-contribs-lgpl==0.0.6
 axolotl-contribs-mit==0.0.3
 
-mistral-common==1.6.0
+mistral-common==1.6.3

setup.py

@@ -111,9 +111,9 @@ def get_package_version():
 
 
 extras_require = {
-    "flash-attn": ["flash-attn==2.7.4.post1"],
+    "flash-attn": ["flash-attn==2.8.0.post2"],
     "ring-flash-attn": [
-        "flash-attn==2.7.4.post1",
+        "flash-attn==2.8.0.post2",
         "ring-flash-attn>=0.1.4",
         "yunchang==0.6.0",
     ],

src/axolotl/common/datasets.py

@@ -75,13 +75,17 @@ def load_datasets(
 
         num_examples = cli_args.debug_num_examples if cli_args else 1
         text_only = cli_args.debug_text_only if cli_args else False
-        train_samples = sample_dataset(train_dataset, num_examples)
-        check_dataset_labels(
-            train_samples,
-            tokenizer,
-            num_examples=num_examples,
-            text_only=text_only,
-        )
+        try:
+            train_samples = sample_dataset(train_dataset, num_examples)
+            check_dataset_labels(
+                train_samples,
+                tokenizer,
+                num_examples=num_examples,
+                text_only=text_only,
+            )
+        except AttributeError:
+            # can't sample iterable datasets
+            pass
 
         LOG.info("printing prompters...")
         for prompter in prompters:

src/axolotl/core/builders/causal.py

@@ -253,6 +253,10 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         training_arguments_kwargs["eval_sample_packing"] = bool(
             self.cfg.eval_sample_packing
         )
+        if self.cfg.sample_packing_sequentially is not None:
+            training_arguments_kwargs["sample_packing_sequentially"] = (
+                self.cfg.sample_packing_sequentially
+            )
         if self.cfg.sample_packing_bin_size is not None:
             training_arguments_kwargs["sample_packing_bin_size"] = (
                 self.cfg.sample_packing_bin_size
@@ -413,7 +417,8 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 or self.cfg.micro_batch_size > 1
             ):
                 return DataCollatorForSeq2Seq(self.tokenizer, **kwargs)
-            return None
+            if not (self.cfg.sample_packing and self.cfg.pretrain_multipack_attn):
+                return None
 
         if self.cfg.model_config_type == "mamba":
             return MambaDataCollator(tokenizer=self.tokenizer)

src/axolotl/core/trainers/base.py

@@ -116,6 +116,7 @@ class AxolotlTrainer(
             sequential=self.args.sample_packing_sequentially,
             drop_last=True,
             num_processes=self.args.dataset_num_proc,
+            mp_start_method=self.args.sample_packing_mp_start_method or "fork",
         )
 
         len(sampler)

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

@@ -28,7 +28,7 @@ class DPOStrategy:
         training_args_kwargs["max_completion_length"] = None
         training_args_kwargs["max_length"] = cfg.sequence_len
         training_args_kwargs["max_prompt_length"] = cfg.sequence_len
-        training_args_kwargs["generate_during_eval"] = cfg.use_wandb
+        training_args_kwargs["generate_during_eval"] = cfg.dpo_generate_during_eval
         if cfg.dpo_use_weighting is not None:
             training_args_kwargs["use_weighting"] = cfg.dpo_use_weighting
         if cfg.dpo_padding_free is not None:

src/axolotl/core/training_args_base.py

@@ -38,6 +38,10 @@ class AxolotlTrainingMixins:
             "help": "Use next-fit sample packing that preserves the order of samples coming from the sampler. Use in combination with curriculum_sampling for fully sequential packing."
         },
     )
+    sample_packing_mp_start_method: str | None = field(
+        default=None,
+        metadata={"help": "The multiprocessing start method to use."},
+    )
     multipack_real_batches: bool = field(
         default=False,
         metadata={"help": "Use real batches for efficient training."},

src/axolotl/loaders/constants.py

@@ -1,6 +1,7 @@
 """Shared constants for axolotl.loaders module"""
 
 from transformers import (
+    AutoModelForCausalLM,
     Gemma3ForConditionalGeneration,
     Llama4ForConditionalGeneration,
     LlavaForConditionalGeneration,
@@ -18,4 +19,6 @@ MULTIMODAL_AUTO_MODEL_MAPPING = {
     "qwen2_5_vl": Qwen2_5_VLForConditionalGeneration,
     "mistral3": Mistral3ForConditionalGeneration,
     "gemma3": Gemma3ForConditionalGeneration,
+    # phi3_v modeling code is not available in transformers yet
+    "phi3_v": AutoModelForCausalLM,
 }

src/axolotl/loaders/model.py

@@ -776,6 +776,9 @@ class ModelLoader:
         dist_dtype: torch.dtype,
         before_kbit_train_or_finetune: bool,
     ):
+        dest = {"dtype": dist_dtype}
+        if self.cfg.lora_on_cpu:
+            dest["device"] = "cpu"
         for name, module in self.model.named_modules():
             if "norm" in name:
                 module.to(dist_dtype)
@@ -786,4 +789,4 @@ class ModelLoader:
                     # don't upcast lm_head for btlm
                     continue
             if any(m in name for m in embedding_modules) and hasattr(module, "weight"):
-                module.to(dist_dtype)
+                module.to(**dest)

src/axolotl/loaders/patch_manager.py

@@ -65,6 +65,7 @@ class PatchManager:
         self._apply_mistral_cross_entropy_patch()
         self._apply_self_attention_lora_patch()
         self._apply_gemma3_conditional_generation_forward_patch()
+        self._apply_sequence_parallel_patches()
 
     def apply_post_model_load_patches(self, model: PreTrainedModel):
         """Apply patches that require the model instance."""
@@ -231,6 +232,17 @@ class PatchManager:
 
             patch_gemma3_conditional_generation_forward()
 
+    def _apply_sequence_parallel_patches(self):
+        """Apply sequence parallelism patches."""
+        if self.cfg.sequence_parallel_degree and self.cfg.sequence_parallel_degree > 1:
+            from axolotl.monkeypatch.ring_attn.patch import (
+                patch_prepare_data_loader,
+                patch_prepare_device_mesh,
+            )
+
+            patch_prepare_data_loader()
+            patch_prepare_device_mesh(self.cfg.sequence_parallel_degree, self.cfg.fsdp)
+
     def _patch_attention(self):
         """Apply attention-specific patches based on model type."""
         if not (self.cfg.flash_attention and hasattr(self.model_config, "model_type")):

src/axolotl/monkeypatch/multipack.py

@@ -42,6 +42,10 @@ def patch_for_multipack(model_type, model_name=None, has_remote_code=False):
     if has_remote_code:
         patch_remote(model_name)
     elif hasattr(transformers, "modeling_flash_attention_utils"):
+        # sanity check in case upstream api changes on this
+        assert hasattr(
+            transformers.modeling_flash_attention_utils, "_get_unpad_data"
+        ), "transformers api changed for _get_unpad_data for flash attention"
         transformers.modeling_flash_attention_utils._get_unpad_data = (  # pylint: disable=protected-access
             get_unpad_data
         )

src/axolotl/monkeypatch/ring_attn/patch.py

@@ -152,7 +152,7 @@ def update_ring_attn_params(position_ids: torch.Tensor | None):
 def patch_prepare_data_loader():
     """Patch `accelerate.data_loader.prepare_data_loader` to respect the SP degree.
 
-    Raies:
+    Raises:
         RuntimeError: If source code to patch does not exist.
     """
     original_fn = accelerate.data_loader.prepare_data_loader
@@ -168,23 +168,34 @@ def patch_prepare_data_loader():
         ORIGINAL_PREPARE_DATALOADER_CODE, NEW_PREPARE_DATALOADER_CODE
     )
 
+    items_to_import = []
+    for item in dir(accelerate.data_loader):
+        if item in patched_source:
+            items_to_import.append(item)
+
     # Create a new function from the patched source
     namespace = {}
     exec(  # pylint: disable=exec-used  # nosec B102
-        patched_source, accelerate.data_loader.__dict__, namespace
+        f"from accelerate.data_loader import ({', '.join(items_to_import)})",
+        globals(),
+    )
+    exec(  # pylint: disable=exec-used  # nosec B102
+        patched_source, globals(), namespace
     )
-    patched_function = namespace["prepare_data_loader"]
 
-    accelerate.data_loader.prepare_data_loader = patched_function
+    patched_function = namespace["prepare_data_loader"]
+    original_fn.__code__ = patched_function.__code__
+
     LOG.info("Patched accelerate.data_loader.prepare_data_loader for SP support")
 
 
-def patch_prepare_device_mesh(sequence_parallel_degree: int):
+def patch_prepare_device_mesh(sequence_parallel_degree: int, fsdp: bool = False):
     """Patches the `Accelerator._prepare_device_mesh` method to create a device mesh
     that includes sequence parallelism with the specified degree.
 
     Args:
-        sequence_parallel_degree (int): The degree of sequence parallelism to use.
+        sequence_parallel_degree: The degree of sequence parallelism to use.
+        fsdp: Whether to use FSDP.
     """
 
     def _prepare_device_mesh(self):
@@ -207,12 +218,14 @@ def patch_prepare_device_mesh(sequence_parallel_degree: int):
         )
         device_ids = list(range(world_size))
 
-        # Note that we use "cp" instead of "sp" to match the PyTorch native "context
-        # parallelism" implementation naming
+        # NOTE: We use "cp" instead of "sp" to match the PyTorch native "context
+        # parallelism" implementation naming.
+        # NOTE: We have a simplified FSDP handling here; i.e., if FSDP is enabled, we
+        # only use "fsdp" and "cp" for the device mesh.
         return dist.DeviceMesh(
             "cuda",
             torch.tensor(device_ids).reshape(mesh_shape),
-            mesh_dim_names=("dp", "cp"),
+            mesh_dim_names=("dp", "cp") if not fsdp else ("fsdp", "cp"),
         )
 
     # Replace the original method with our new method

src/axolotl/processing_strategies.py

@@ -264,6 +264,23 @@ class Gemma3ProcessingStrategy(ProcessingStrategy):
         return labels
 
 
+class Phi35VLProcessingStrategy(ProcessingStrategy):
+    """Processing Strategy class for Phi-3.5-vision-instruct"""
+
+    def __init__(
+        self,
+        processor: ProcessorMixin,
+        chat_template: Optional[str] = None,
+        image_size: int | tuple[int, int] | None = None,
+        image_resize_algorithm: Resampling | None = None,
+    ):
+        super().__init__(processor, chat_template, image_size, image_resize_algorithm)
+        self.image_token = "<|image|>"  # nosec
+        self.image_token_id = processor.tokenizer.convert_tokens_to_ids(
+            self.image_token
+        )
+
+
 def get_processing_strategy(
     processor: ProcessorMixin,
     chat_template,
@@ -279,6 +296,10 @@ def get_processing_strategy(
         return Gemma3ProcessingStrategy(
             processor, chat_template, image_size, image_resize_algorithm
         )
+    if chat_template_type == "phi_35_vl":
+        return Phi35VLProcessingStrategy(
+            processor, chat_template, image_size, image_resize_algorithm
+        )
     if chat_template_type in [
         "llama3_2_vision",
         "llama4",

src/axolotl/prompt_strategies/chat_template.py

@@ -103,6 +103,7 @@ class ChatTemplatePrompter(Prompter):
         chat_template_kwargs = {
             "chat_template": self.chat_template,
             "add_generation_prompt": add_generation_prompt,
+            **self.chat_template_kwargs,
         }
 
         if tools:

src/axolotl/train.py

@@ -223,6 +223,9 @@ def execute_training(
             )
 
         LOG.info("Starting trainer...")
+        # TODO: disabling for now as not compatible with FSDP2 + torchao low bit optimizers
+        # if cfg.bf16:
+        #     torch.set_default_dtype(torch.bfloat16)
         trainer.train(resume_from_checkpoint=resume_from_checkpoint)
 
 

src/axolotl/utils/chat_templates.py

@@ -32,6 +32,7 @@ _CHAT_TEMPLATES = {
     "llava": "{% for message in messages %}{% if message['role'] != 'system' %}{{ message['role'].upper() + ': '}}{% endif %}{# Render all images first #}{% for content in message['content'] | selectattr('type', 'equalto', 'image') %}{{ '<image>\n' }}{% endfor %}{# Render all text next #}{% if message['role'] != 'assistant' %}{% for content in message['content'] | selectattr('type', 'equalto', 'text') %}{{ content['text'] + ' '}}{% endfor %}{% else %}{% for content in message['content'] | selectattr('type', 'equalto', 'text') %}{% generation %}{{ content['text'] + ' '}}{% endgeneration %}{% endfor %}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ 'ASSISTANT:' }}{% endif %}",
     "phi_3": "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'system') %}{{'<|system|>' + '\n' + message['content'] + '<|end|>' + '\n'}}{% elif (message['role'] == 'user') %}{{'<|user|>' + '\n' + message['content'] + '<|end|>' + '\n' + '<|assistant|>' + '\n'}}{% elif message['role'] == 'assistant' %}{{message['content'] + '<|end|>' + '\n'}}{% endif %}{% endfor %}",
     "phi_35": "{% for message in messages %}{% if message['role'] == 'system' and message['content'] %}{{'<|system|>\n' + message['content'] + '<|end|>\n'}}{% elif message['role'] == 'user' %}{{'<|user|>\n' + message['content'] + '<|end|>\n'}}{% elif message['role'] == 'assistant' %}{{'<|assistant|>\n' + message['content'] + '<|end|>\n'}}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|assistant|>\n' }}{% endif %}",
+    "phi_35_vl": "{% set image_count = namespace(value=0) %}{% for message in messages %}{{'<|' + message['role'] + '|>\n' }}{% if message['content'] is string %}{{ message['content'] }}{% else %}{% set message_images = [] %}{% set message_text = [] %}{% for chunk in message['content'] %}{% if chunk['type'] == 'image' or 'image' in chunk or 'image_url' in chunk %}{% set image_count.value = image_count.value + 1 %}{% set _ = message_images.append('<|image_' + image_count.value|string + '|>\n') %}{% elif chunk['type'] == 'text' %}{% set _ = message_text.append(chunk['text']) %}{% endif %}{% endfor %}{{ message_images | join('') }}{{ message_text | join('') }}{% endif %}{{ '<|end|>\n' }}{% endfor %}{% if add_generation_prompt and messages[-1]['role'] != 'assistant' %}{{- '<|assistant|>\n' -}}{% endif %}",
     "phi_4": "{% set system_message = 'You are Phi, a language model trained by Microsoft to help users. Your role as an assistant involves thoroughly exploring questions through a systematic thinking process before providing the final precise and accurate solutions. This requires engaging in a comprehensive cycle of analysis, summarizing, exploration, reassessment, reflection, backtracing, and iteration to develop well-considered thinking process. Please structure your response into two main sections: Thought and Solution using the specified format: <think> {Thought section} </think> {Solution section}. In the Thought section, detail your reasoning process in steps. Each step should include detailed considerations such as analysing questions, summarizing relevant findings, brainstorming new ideas, verifying the accuracy of the current steps, refining any errors, and revisiting previous steps. In the Solution section, based on various attempts, explorations, and reflections from the Thought section, systematically present the final solution that you deem correct. The Solution section should be logical, accurate, and concise and detail necessary steps needed to reach the conclusion. Now, try to solve the following question through the above guidelines:' -%}{%- if messages and messages[0]['role'] == 'system' -%}{%- set system_message = messages[0]['content'] -%}{%- set messages = messages[1:] -%}{%- endif -%}<|im_start|>system<|im_sep|>{{ system_message }}<|im_end|>{% for message in messages %}{% if (message['role'] == 'user') %}{{'<|im_start|>user<|im_sep|>' + message['content'] + '<|im_end|>'}}{% elif (message['role'] == 'assistant') %}{{'<|im_start|>assistant<|im_sep|>'}}{% generation %}{{message['content'] + '<|im_end|>'}}{% endgeneration %}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant<|im_sep|>' }}{% endif %}",
     "deepseek_v2": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{{ bos_token }}{% for message in messages %}{% if message['role'] == 'user' %}{{ '<｜User｜>' + message['content'] }}{% elif message['role'] == 'assistant' %}{{ '<｜Assistant｜>' + message['content'] + eos_token }}{% elif message['role'] == 'system' %}{{ message['content'] + '\n\n' }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<｜Assistant｜>' }}{% endif %}",
     "deepseek_v3": "{% if not add_generation_prompt is defined %}{% set add_generation_prompt = false %}{% endif %}{% set ns = namespace(is_first=false, is_tool=false, is_output_first=true, system_prompt='', is_first_sp=true) %}{%- for message in messages %}{%- if message['role'] == 'system' %}{%- if ns.is_first_sp %}{% set ns.system_prompt = ns.system_prompt + message['content'] %}{% set ns.is_first_sp = false %}{%- else %}{% set ns.system_prompt = ns.system_prompt + '\\n\\n' + message['content'] %}{%- endif %}{%- endif %}{%- endfor %}{{ bos_token }}{{ ns.system_prompt }}{%- for message in messages %}{%- if message['role'] == 'user' %}{%- set ns.is_tool = false -%}{{'<｜User｜>' + message['content']}}{%- endif %}{%- if message['role'] == 'assistant' and 'tool_calls' in message %}{%- set ns.is_tool = false -%}{%- for tool in message['tool_calls'] %}{%- if not ns.is_first %}{%- if message['content'] is none %}{{'<｜Assistant｜><｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- else %}{{'<｜Assistant｜>' + message['content'] + '<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- endif %}{%- set ns.is_first = true -%}{%- else %}{{'\\n' + '<｜tool▁call▁begin｜>' + tool['type'] + '<｜tool▁sep｜>' + tool['function']['name'] + '\\n' + '```json' + '\\n' + tool['function']['arguments'] + '\\n' + '```' + '<｜tool▁call▁end｜>'}}{%- endif %}{%- endfor %}{{'<｜tool▁calls▁end｜><｜end▁of▁sentence｜>'}}{%- endif %}{%- if message['role'] == 'assistant' and 'tool_calls' not in message %}{%- if ns.is_tool %}{{'<｜tool▁outputs▁end｜>' + message['content'] + '<｜end▁of▁sentence｜>'}}{%- set ns.is_tool = false -%}{%- else %}{% set content = message['content'] %}{% if '</think>' in content %}{% set content = content.split('</think>')[-1] %}{% endif %}{{'<｜Assistant｜>' + content + '<｜end▁of▁sentence｜>'}}{%- endif %}{%- endif %}{%- if message['role'] == 'tool' %}{%- set ns.is_tool = true -%}{%- if ns.is_output_first %}{{'<｜tool▁outputs▁begin｜><｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- set ns.is_output_first = false %}{%- else %}{{'<｜tool▁output▁begin｜>' + message['content'] + '<｜tool▁output▁end｜>'}}{%- endif %}{%- endif %}{%- endfor -%}{% if ns.is_tool %}{{'<｜tool▁outputs▁end｜>'}}{% endif %}{% if add_generation_prompt and not ns.is_tool %}{{'<｜Assistant｜>'}}{% endif %}",

src/axolotl/utils/ctx_managers/sequence_parallel.py

@@ -12,8 +12,6 @@ from transformers.utils import ModelOutput
 
 from axolotl.monkeypatch.ring_attn import (
     get_ring_attn_group,
-    patch_prepare_data_loader,
-    patch_prepare_device_mesh,
     register_ring_attn,
     update_ring_attn_params,
 )
@@ -238,12 +236,6 @@ class SequenceParallelContextManager:
             ring_attn_func=self.ring_attn_func,
         )
 
-        # Patches for accelerate functionality
-        patch_prepare_data_loader()
-        patch_prepare_device_mesh(
-            sequence_parallel_degree=self.sequence_parallel_degree
-        )
-
     def _register_model_hooks(self):
         # Forward pre-hook to apply sequence parallelism
         def sequence_parallel_pre_hook(_, args, kwargs):

src/axolotl/utils/data/pretraining.py

@@ -224,10 +224,10 @@ def wrap_pretraining_dataset(
     remove_columns = []
     if dataset.features is None:
         for first_row in dataset:
-            remove_columns = first_row.keys()
+            remove_columns = list(first_row.keys())
             break
     else:
-        remove_columns = dataset.features.keys()
+        remove_columns = list(dataset.features.keys())
 
     dataset = dataset.map(
         encode,
@@ -267,6 +267,7 @@ def encode_packed_pretraining(
         batch_size=1,
         batch_max_len=batch_size * max_seq_length,
         drop_last=True,
+        num_processes=1,
     )
 
     chunked_data = defaultdict(list)

src/axolotl/utils/data/sft.py

@@ -334,7 +334,10 @@ def _load_raw_datasets(
     dataset = merge_datasets(datasets, cfg)
 
     if not cfg.skip_prepare_dataset:
-        dataset = drop_long_seq_in_dataset(dataset, cfg)
+        if split == "test" and cfg.eval_sequence_len:
+            dataset = drop_long_seq_in_dataset(dataset, cfg.eval_sequence_len, cfg)
+        else:
+            dataset = drop_long_seq_in_dataset(dataset, cfg.sequence_len, cfg)
         if cfg.sample_packing:
             dataset, _ = process_datasets_for_packing(cfg, dataset, None)
 

src/axolotl/utils/data/shared.py

@@ -524,13 +524,24 @@ def merge_datasets(datasets: list[Dataset], cfg: DictDefault) -> Dataset:
         Merged dataset.
     """
     if len(datasets) == 1:
-        return datasets[0]
+        ds = datasets[0]
+
+        # Do not shuffle if curriculum sampling is enabled
+        if cfg.curriculum_sampling:
+            return ds
+
+        return ds.shuffle(seed=cfg.seed)
 
     LOG.info("Merging datasets...")
     merged_dataset = concatenate_datasets(datasets)
 
     if cfg.shuffle_merged_datasets:
         LOG.debug("Shuffling merged datasets...")
+        if cfg.curriculum_sampling:
+            LOG.warning(
+                "Shuffling merged datasets with curriculum sampling is not recommended. "
+                "This will randomize the order of samples."
+            )
         merged_dataset = merged_dataset.shuffle(seed=cfg.seed)
     else:
         LOG.debug("Not shuffling merged datasets.")

src/axolotl/utils/data/utils.py

@@ -148,11 +148,14 @@ def deduplicate_and_log_datasets(
     return dataset, other_dataset
 
 
-def drop_long_seq_in_dataset(dataset: Dataset, cfg: DictDefault) -> Dataset:
+def drop_long_seq_in_dataset(
+    dataset: Dataset, sequence_len: int, cfg: DictDefault
+) -> Dataset:
     """Remove sequences longer than configured maximum from dataset.
 
     Args:
         dataset: Dataset to filter.
+        sequence_len: Maximum length for sequences to keep
         cfg: Dictionary mapping `axolotl` config keys to values.
 
     Returns:
@@ -167,7 +170,7 @@ def drop_long_seq_in_dataset(dataset: Dataset, cfg: DictDefault) -> Dataset:
 
     drop_long = functools.partial(
         drop_long_seq,
-        sequence_len=cfg.sequence_len,
+        sequence_len=sequence_len,
         min_sequence_len=cfg.min_sample_len,
     )
 
@@ -187,7 +190,7 @@ def drop_long_seq_in_dataset(dataset: Dataset, cfg: DictDefault) -> Dataset:
 
     drop_long_kwargs = {}
     if filter_map_kwargs:
-        drop_long_kwargs["desc"] = "Dropping Long Sequences"
+        drop_long_kwargs["desc"] = f"Dropping Long Sequences (>{sequence_len})"
 
     dataset = dataset.filter(
         drop_long,

src/axolotl/utils/mistral_tokenizer.py

@@ -8,7 +8,7 @@ from typing import TYPE_CHECKING, Optional
 import numpy as np
 from huggingface_hub import hf_hub_download
 from mistral_common.tokens.tokenizers.mistral import MistralTokenizer
-from mistral_common.tokens.tokenizers.tekken import Tekkenizer
+from mistral_common.tokens.tokenizers.tekken import SpecialTokenPolicy, Tekkenizer
 from torch import Tensor
 from transformers.utils import PaddingStrategy
 
@@ -251,10 +251,13 @@ class HFMistralTokenizer:
             token_ids = [token_ids]
 
         if skip_special_tokens:
-            return self._mistral.instruct_tokenizer.tokenizer.decode(token_ids)
+            return self._mistral.instruct_tokenizer.tokenizer.decode(
+                token_ids, special_token_policy=SpecialTokenPolicy.IGNORE
+            )
 
-        # to_string returns a string with special tokens
-        return self._mistral.instruct_tokenizer.tokenizer.to_string(token_ids)
+        return self._mistral.instruct_tokenizer.tokenizer.decode(
+            token_ids, special_token_policy=SpecialTokenPolicy.KEEP
+        )
 
     def _create_mistral_chat_completion_request(
         self, conversation: list[dict], tools: list[dict] | None = None

src/axolotl/utils/samplers/multipack.py

@@ -127,7 +127,7 @@ def pack_parallel(
     bin_size: int,
     num_processes: int | None = None,
     safe_mode: bool = True,
-    mp_start_method: str | None = "spawn",
+    mp_start_method: str | None = "fork",
 ) -> list[list[int]]:
     """Pack sequences into bins using parallel processing.
 
@@ -260,12 +260,13 @@ class MultipackBatchSampler(BatchSampler):
         lengths: np.ndarray,  # Sequence lengths
         packing_efficiency_estimate: float = 1.0,  # Initial efficiency estimate
         drop_last: bool = True,  # Whether to drop final batches (might be incomplete)
-        num_count_samples: int = 8,  # Number of times to estimate batch count
+        num_count_samples: int = 4,  # Number of times to estimate batch count
         sequential: bool = False,  # Whether to use sequential packing
         group_size: int = 100_000,  # Size of groups for parallel packing
         bin_size: int = 200,  # The max number of samples that can be packed in a single bin
         num_processes: int | None = None,  # Number of processes for parallel packing
         safe_mode: bool = True,  # Conservative packing to prevent training instability
+        mp_start_method: str = "fork",
         **kwargs,  # pylint: disable=unused-argument
     ):
         super().__init__(sampler, batch_size, drop_last)
@@ -278,6 +279,7 @@ class MultipackBatchSampler(BatchSampler):
         self.bin_size = bin_size
         self.num_processes = num_processes
         self.safe_mode = safe_mode
+        self.mp_start_method = mp_start_method
 
         assert isinstance(self.lengths, np.ndarray)
 
@@ -333,13 +335,15 @@ class MultipackBatchSampler(BatchSampler):
             bins = [[indices[b_idx] for b_idx in bin_indices] for bin_indices in bins]
         else:
             # Use parallel packing
+            num_processes = self.num_processes or 1
             all_bins = pack_parallel(
                 lengths,
                 bin_capacity=self.batch_max_len,
                 group_size=self.group_size,
                 bin_size=self.bin_size,
-                num_processes=self.num_processes,
+                num_processes=min(4, num_processes) if num_processes else 4,
                 safe_mode=self.safe_mode,
+                mp_start_method=self.mp_start_method,
             )
 
             # Map bin indices back to original indices

src/axolotl/utils/schemas/config.py

@@ -146,6 +146,7 @@ class AxolotlInputConfig(
     dpo_label_smoothing: float | None = None
     dpo_norm_loss: bool | None = None
     dpo_padding_free: bool | None = None
+    dpo_generate_during_eval: bool | None = None
 
     datasets: (
         Annotated[
@@ -366,6 +367,12 @@ class AxolotlInputConfig(
             "description": "The maximum length of an input to train with, this should typically be less than 2048 as most models have a token/context limit of 2048"
         },
     )
+    eval_sequence_len: int | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "The maximum length of an input for evaluation. If not specified, defaults to sequence_len"
+        },
+    )
     min_sample_len: int | None = None
     max_prompt_len: int = Field(
         default=512,
@@ -393,6 +400,12 @@ class AxolotlInputConfig(
         default=None,
         json_schema_extra={"description": "Whether to pack samples sequentially"},
     )
+    sample_packing_mp_start_method: str | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "The multiprocessing start method to use for packing. Should be 'fork', 'spawn' or 'forkserver'"
+        },
+    )
     eval_sample_packing: bool | None = Field(
         default=None,
         json_schema_extra={
@@ -772,6 +785,12 @@ class AxolotlInputConfig(
             "description": "Custom jinja template for chat template. This will be only used if chat_template is set to `jinja` or `null` (in which case chat_template is automatically set to `jinja`). Default is null."
         },
     )
+    chat_template_kwargs: dict[str, Any] | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Additional kwargs to pass to the chat template. This is useful for customizing the chat template. For example, you can pass `thinking=False` to add a generation prompt to the chat template."
+        },
+    )
     eot_tokens: list[str] | None = Field(
         default=None,
         json_schema_extra={

src/axolotl/utils/schemas/enums.py

@@ -48,6 +48,8 @@ class ChatTemplate(str, Enum):
     llama4 = "llama4"
     phi_3 = "phi_3"
     phi_35 = "phi_35"
+    phi_35_vl = "phi_35_vl"
+    phi_4 = "phi_4"
     deepseek_v2 = "deepseek_v2"
     deepseek_v3 = "deepseek_v3"
     jamba = "jamba"

src/axolotl/utils/schemas/validation.py

@@ -462,6 +462,20 @@ class TrainingValidationMixin:
 
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def pretrain_with_tps(cls, data):
+        if data.get("pretraining_dataset") and data.get(
+            "include_tokens_per_second", False
+        ):
+            # combining these would raise `TypeError: cannot pickle 'dict_keys' object`
+            # due to trying to count the number of tokens total in the dataset
+            raise ValueError(
+                "pretraining_dataset and include_tokens_per_second cannot be used together."
+            )
+
+        return data
+
 
 class LoRAValidationMixin:
     """Validation methods related to LoRA/QLoRA configuration."""

src/axolotl/utils/trainer.py

@@ -381,6 +381,7 @@ def process_pretraining_datasets_for_packing(
     if not skip_position_ids:
         train_dataset = train_dataset.map(
             add_position_ids,
+            batched=True,
             desc="Add position_id column (Pretraining Sample Packing)",
         )
     if drop_attention_mask:
@@ -467,6 +468,7 @@ def calculate_total_num_steps(cfg, train_dataset, update=True):
                 sequential=cfg.sample_packing_sequentially,
                 drop_last=True,
                 num_processes=cfg.dataset_processes,
+                mp_start_method=cfg.sample_packing_mp_start_method or "fork",
             )
 
             data_loader = DataLoader(

tests/conftest.py

@@ -4,12 +4,14 @@ shared pytest fixtures
 
 import functools
 import importlib
+import logging
 import os
 import shutil
 import sys
 import tempfile
 import time
-from pathlib import Path
+from pathlib import Path, PosixPath
+from typing import Generator
 
 import datasets
 import pytest
@@ -24,6 +26,8 @@ from tests.hf_offline_utils import (
     hf_offline_context,
 )
 
+logging.getLogger("filelock").setLevel(logging.CRITICAL)
+
 
 def retry_on_request_exceptions(max_retries=3, delay=1):
     # pylint: disable=duplicate-code
@@ -411,7 +415,7 @@ def tokenizer_mistral_7b_instruct_chatml(tokenizer_mistral_7b_instruct):
 
 
 @pytest.fixture
-def temp_dir():
+def temp_dir() -> Generator[str, None, None]:
     # Create a temporary directory
     _temp_dir = tempfile.mkdtemp()
     yield _temp_dir
@@ -419,6 +423,11 @@ def temp_dir():
     shutil.rmtree(_temp_dir)
 
 
+@pytest.fixture(scope="function", autouse=True)
+def unique_triton_cache_dir(temp_dir: str | PosixPath) -> None:
+    os.environ["TRITON_CACHE_DIR"] = str(temp_dir) + "/.triton/cache"
+
+
 @pytest.fixture(scope="function", autouse=True)
 def cleanup_monkeypatches():
     from transformers import Trainer

tests/e2e/multigpu/patched/test_sp.py

@@ -54,6 +54,7 @@ class TestSequenceParallelism:
                 "micro_batch_size": micro_batch_size,
                 "gradient_accumulation_steps": 2,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_8bit",
                 "lr_scheduler": "cosine",

tests/e2e/multigpu/solo/test_flex.py

@@ -54,6 +54,7 @@ class TestPackedFlex:
                 "gradient_accumulation_steps": 2,
                 "gradient_checkpointing": True,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",

tests/e2e/multigpu/solo/test_grpo.py

@@ -309,6 +309,7 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
                 "warmup_steps": 10,
                 "val_set_size": 0.0,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.0001,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
@@ -400,6 +401,7 @@ def oai_gsm8k_transform(cfg, *args, **kwargs):
                 "warmup_steps": 10,
                 "val_set_size": 0.0,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.0001,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",

tests/e2e/multigpu/test_eval.py

@@ -38,12 +38,13 @@ class TestMultiGPUEval:
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
                 "lora_modules_to_save": ["embed_tokens", "lm_head"],
-                "val_set_size": 0.004,
+                "val_set_size": 0.05,
                 "special_tokens": {"pad_token": "<|endoftext|>"},
                 "datasets": [
                     {
                         "path": "teknium/GPT4-LLM-Cleaned",
                         "type": "alpaca",
+                        "split": "train[:5%]",
                     },
                 ],
                 "num_epochs": 1,
@@ -51,6 +52,7 @@ class TestMultiGPUEval:
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 2,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_8bit",
                 "lr_scheduler": "cosine",
@@ -107,12 +109,13 @@ class TestMultiGPUEval:
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
                 "lora_modules_to_save": ["embed_tokens", "lm_head"],
-                "val_set_size": 0.0004,
+                "val_set_size": 0.01,
                 "special_tokens": {"pad_token": "<|endoftext|>"},
                 "datasets": [
                     {
                         "path": "teknium/GPT4-LLM-Cleaned",
                         "type": "alpaca",
+                        "split": "train[:5%]",
                     },
                 ],
                 "num_epochs": 1,
@@ -120,6 +123,7 @@ class TestMultiGPUEval:
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 2,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_8bit",
                 "lr_scheduler": "cosine",

tests/e2e/multigpu/test_gemma3.py

@@ -64,6 +64,7 @@ class TestMultiGPUGemma3:
                 },
                 "gradient_accumulation_steps": 2,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.0001,
                 "optimizer": "adamw_8bit",
                 "lr_scheduler": "cosine",

tests/e2e/multigpu/test_llama.py

@@ -62,6 +62,7 @@ class TestMultiGPULlama:
                 "gradient_accumulation_steps": 2,
                 # "gradient_checkpointing": True,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_8bit",
                 "lr_scheduler": "cosine",
@@ -127,6 +128,7 @@ class TestMultiGPULlama:
                 "gradient_accumulation_steps": gradient_accumulation_steps,
                 # "gradient_checkpointing": True,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_8bit",
                 "lr_scheduler": "cosine",
@@ -200,6 +202,7 @@ class TestMultiGPULlama:
                 "gradient_accumulation_steps": 2,
                 # "gradient_checkpointing": True,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "warmup_steps": 0,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_8bit",
@@ -278,6 +281,7 @@ class TestMultiGPULlama:
                 "gradient_accumulation_steps": 2,
                 # "gradient_checkpointing": True,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "warmup_steps": 0,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_8bit",
@@ -340,6 +344,7 @@ class TestMultiGPULlama:
                 "gradient_accumulation_steps": gradient_accumulation_steps,
                 # "gradient_checkpointing": True,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
@@ -412,6 +417,7 @@ class TestMultiGPULlama:
                 "gradient_accumulation_steps": 2,
                 # "gradient_checkpointing": True,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
@@ -491,6 +497,7 @@ class TestMultiGPULlama:
                 "gradient_accumulation_steps": 2,
                 "gradient_checkpointing": True,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_8bit",
                 "lr_scheduler": "cosine",
@@ -573,6 +580,7 @@ class TestMultiGPULlama:
                 "gradient_accumulation_steps": 2,
                 # "gradient_checkpointing": True,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
@@ -669,6 +677,7 @@ class TestMultiGPULlama:
                 "micro_batch_size": 1,
                 "gradient_accumulation_steps": gradient_accumulation_steps,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
@@ -743,6 +752,7 @@ class TestMultiGPULlama:
                 "micro_batch_size": 1,
                 "gradient_accumulation_steps": gradient_accumulation_steps,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
@@ -817,6 +827,7 @@ class TestMultiGPULlama:
                 "micro_batch_size": 1,
                 "gradient_accumulation_steps": gradient_accumulation_steps,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",

tests/e2e/multigpu/test_qwen2.py

@@ -46,6 +46,7 @@ class TestMultiGPUQwen2:
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 2,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",

tests/e2e/multigpu/test_ray.py

@@ -48,6 +48,7 @@ class TestMultiGPURay:
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 2,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_8bit",
                 "lr_scheduler": "cosine",
@@ -107,6 +108,7 @@ class TestMultiGPURay:
                 "micro_batch_size": 1,
                 "gradient_accumulation_steps": gradient_accumulation_steps,
                 "output_dir": temp_dir,
+                "dataset_prepared_path": temp_dir + "/last_run_prepared",
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch",
                 "lr_scheduler": "cosine",

tests/e2e/patched/lora_kernels/test_lora_kernel_patching.py

@@ -396,7 +396,7 @@ def test_model_architecture(model_config):
 
 
 # pylint: disable=duplicate-code
-def test_kernel_training_integration():
+def test_kernel_training_integration(temp_dir):
     """Test model loading with kernel patches enabled."""
     from axolotl.cli.utils import load_model_and_tokenizer
 
@@ -426,6 +426,14 @@ def test_kernel_training_integration():
         }
     )
 
+    # 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))
+
     # Load model
     model, _, _ = load_model_and_tokenizer(cfg=cfg)
 
@@ -505,7 +513,7 @@ def test_kernel_training_integration_auto_enable(temp_dir):
     assert found_patched_attn
 
 
-def test_kernel_training_integration_dropout_non_zero():
+def test_kernel_training_integration_dropout_non_zero(temp_dir):
     """Test model loading with dropout non-zero should not patch."""
 
     from axolotl.cli.utils import load_model_and_tokenizer
@@ -533,6 +541,14 @@ def test_kernel_training_integration_dropout_non_zero():
         }
     )
 
+    # 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))
+
     # Get original attention class
     attention_cls = get_attention_cls_from_config(cfg)
 

tests/test_packed_batch_sampler.py

@@ -70,7 +70,7 @@ class TestBatchedSamplerPacking:
         )
         train_dataset = concatenate_datasets([dataset_wrapper])
 
-        train_dataset = drop_long_seq_in_dataset(train_dataset, cfg)
+        train_dataset = drop_long_seq_in_dataset(train_dataset, cfg.sequence_len, cfg)
 
         lengths = get_dataset_lengths(train_dataset)
         batch_sampler = MultipackBatchSampler(