set release version 0.8.0 (#2476)

* set release version 0.8.0

* make sure to include ring-flash-attn in docker image build

.github/workflows/base.yml

@@ -40,12 +40,24 @@ jobs:
             python_version: "3.11"
             pytorch: 2.6.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+          - cuda: "126"
+            cuda_version: 12.6.3
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.6.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
           - cuda: "128"
             cuda_version: 12.8.1
             cudnn_version: ""
             python_version: "3.11"
             pytorch: nightly
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: next
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -67,7 +79,7 @@ jobs:
         uses: docker/build-push-action@v4
         with:
           context: .
-          file: ${{ matrix.pytorch == 'nightly' && './docker/Dockerfile-base-nightly' || './docker/Dockerfile-base' }}
+          file: ${{ matrix.pytorch == 'nightly' && './docker/Dockerfile-base-nightly' || matrix.pytorch == 'next' && './docker/Dockerfile-base-next' || './docker/Dockerfile-base' }}
           push: ${{ github.event_name != 'pull_request' }}
           tags: ${{ steps.metadata.outputs.tags }}-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}${{ matrix.axolotl_extras != '' && '-' || '' }}${{ matrix.axolotl_extras }}
           labels: ${{ steps.metadata.outputs.labels }}

.github/workflows/main.yml

@@ -25,12 +25,12 @@ jobs:
             python_version: "3.11"
             pytorch: 2.5.1
             axolotl_extras: vllm
-            is_latest: true
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:
+            is_latest: true
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -87,12 +87,12 @@ jobs:
             python_version: "3.11"
             pytorch: 2.5.1
             axolotl_extras:
-            is_latest: true
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:
+            is_latest: true
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

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

@@ -42,8 +42,7 @@ jobs:
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.6.0
-            # awaiting vllm#12721
-            axolotl_extras:
+            axolotl_extras: vllm
             num_gpus: 2
             nightly_build: "true"
     runs-on: [self-hosted, modal]

.github/workflows/tests-nightly.yml

@@ -33,6 +33,15 @@ jobs:
       - name: Check out repository code
         uses: actions/checkout@v4
 
+      - name: Restore HF cache
+        id: hf-cache-restore
+        uses: actions/cache/restore@v4
+        with:
+          path: |
+            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/models--*
+          key: ${{ runner.os }}-hf-hub-cache-v2
+
       - name: Setup Python
         uses: actions/setup-python@v5
         with:
@@ -46,7 +55,7 @@ jobs:
 
       - name: Install PyTorch
         run: |
-          pip3 install torch==${{ matrix.pytorch_version }} --index-url https://download.pytorch.org/whl/cpu
+          pip3 install torch==${{ matrix.pytorch_version }}
 
       - name: Update requirements.txt
         run: |
@@ -58,8 +67,7 @@ jobs:
 
       - name: Install dependencies
         run: |
-          pip3 install --upgrade pip
-          pip3 install --upgrade packaging==23.2
+          pip3 show torch
           pip3 install --no-build-isolation -U -e .
           python scripts/unsloth_install.py | sh
           python scripts/cutcrossentropy_install.py | sh
@@ -73,10 +81,15 @@ jobs:
         run: |
           axolotl --help
 
+      - name: Pre-Download dataset fixture
+        run: |
+          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+
       - name: Run tests
         run: |
-          pytest -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ tests/
-          pytest tests/patched/
+          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
+          pytest -v tests/patched/
+          pytest -v tests/cli/
 
       - name: cleanup pip cache
         run: |

.github/workflows/tests.yml

@@ -96,6 +96,10 @@ jobs:
         run: |
           axolotl --help
 
+      - name: Pre-Download dataset fixture
+        run: |
+          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+
       - name: Run tests
         run: |
           pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
@@ -256,7 +260,7 @@ jobs:
             python_version: "3.11"
             pytorch: 2.6.0
             num_gpus: 1
-            axolotl_extras:
+            axolotl_extras: vllm
     steps:
       - name: Checkout
         uses: actions/checkout@v4

_quarto.yml

@@ -40,6 +40,7 @@ quartodoc:
         - cli.preprocess
         - cli.sweeps
         - cli.utils
+        - cli.vllm_serve
         - cli.cloud.base
         - cli.cloud.modal_
     - title: Trainers

cicd/multigpu.sh

@@ -2,4 +2,5 @@
 set -e
 
 # only run one test at a time so as not to OOM the GPU
-pytest -v -n2 /workspace/axolotl/tests/e2e/multigpu/
+pytest -v -n2 /workspace/axolotl/tests/e2e/multigpu/ --ignore=/workspace/axolotl/tests/e2e/multigpu/solo/
+pytest -v -n1 /workspace/axolotl/tests/e2e/multigpu/solo/

docker/Dockerfile

@@ -20,9 +20,9 @@ WORKDIR /workspace/axolotl
 
 # If AXOLOTL_EXTRAS is set, append it in brackets
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install --no-build-isolation -e .[deepspeed,flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
-        pip install --no-build-isolation -e .[deepspeed,flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
     fi
 
 RUN python scripts/unsloth_install.py | sh

docker/Dockerfile-base-next

@@ -0,0 +1,38 @@
+ARG CUDA_VERSION="12.8.1"
+ARG CUDNN_VERSION="8"
+ARG UBUNTU_VERSION="22.04"
+ARG MAX_JOBS=4
+
+FROM nvidia/cuda:$CUDA_VERSION-cudnn$CUDNN_VERSION-devel-ubuntu$UBUNTU_VERSION AS base-builder
+
+ENV PATH="/root/miniconda3/bin:${PATH}"
+
+ARG PYTHON_VERSION="3.11"
+ARG PYTORCH_VERSION="next"
+ARG CUDA="128"
+ARG TORCH_CUDA_ARCH_LIST="7.0 7.5 8.0 8.6 9.0+PTX"
+
+ENV PYTHON_VERSION=$PYTHON_VERSION
+ENV TORCH_CUDA_ARCH_LIST=$TORCH_CUDA_ARCH_LIST
+
+RUN apt-get update \
+    && apt-get install -y wget git build-essential ninja-build git-lfs libaio-dev pkg-config && rm -rf /var/lib/apt/lists/* \
+    && wget \
+    https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh \
+    && mkdir /root/.conda \
+    && bash Miniconda3-latest-Linux-x86_64.sh -b \
+    && rm -f Miniconda3-latest-Linux-x86_64.sh \
+    && conda create -n "py${PYTHON_VERSION}" python="${PYTHON_VERSION}"
+
+ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
+
+WORKDIR /workspace
+
+RUN python3 -m pip install --upgrade pip && pip3 install packaging && \
+    python3 -m pip install --no-cache-dir -U torch==2.7.0 --extra-index-url https://download.pytorch.org/whl/test/cu$CUDA && \
+    python3 -m pip install --no-cache-dir "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main" && \
+    python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"
+
+RUN git lfs install --skip-repo && \
+    pip3 install awscli && \
+    pip3 install -U --no-cache-dir pydantic==2.10.6

docs/cli.qmd

@@ -170,7 +170,7 @@ axolotl merge-sharded-fsdp-weights config.yml
 
 ### evaluate
 
-Evaluates a model's performance using metrics specified in the config.
+Evaluates a model's performance (loss etc) on the train and eval datasets.
 
 ```bash
 # Basic evaluation
@@ -197,6 +197,8 @@ lm_eval_batch_size: # Batch size for evaluation
 output_dir: # Directory to save evaluation results
 ```
 
+See [LM Eval Harness](https://github.com/EleutherAI/lm-evaluation-harness) for more details.
+
 ## Legacy CLI Usage
 
 While the new Click-based CLI is preferred, Axolotl still supports the legacy module-based CLI:
@@ -235,7 +237,7 @@ Create a cloud config YAML with your Modal settings:
 ```yaml
 # cloud_config.yml
 provider: modal
-gpu: a100  # Supported: l40s, a100-40gb, a100-80gb, a10g, h100, t4, l4
+gpu: a100       # Supported: l40s, a100-40gb, a100-80gb, a10g, h100, t4, l4
 gpu_count: 1    # Number of GPUs to use
 timeout: 86400  # Maximum runtime in seconds (24 hours)
 branch: main    # Git branch to use (optional)
@@ -248,7 +250,7 @@ volumes:        # Persistent storage volumes
   - name: axolotl-artifacts
     mount: /workspace/artifacts
 
-env:            # Environment variables
+secrets:        # Secrets to inject
   - WANDB_API_KEY
   - HF_TOKEN
 ```
@@ -274,15 +276,27 @@ axolotl lm-eval config.yml --cloud cloud_config.yml
 ### Cloud Configuration Options
 
 ```yaml
-provider: # compute provider, currently only `modal` is supported
-gpu: # GPU type to use
-gpu_count: # Number of GPUs (default: 1)
-memory: # RAM in GB (default: 128)
-timeout: # Maximum runtime in seconds
+provider:    # compute provider, currently only `modal` is supported
+gpu:         # GPU type to use
+gpu_count:   # Number of GPUs (default: 1)
+memory:      # RAM in GB (default: 128)
+timeout:     # Maximum runtime in seconds
 timeout_preprocess: # Preprocessing timeout
-branch: # Git branch to use
-docker_tag: # Custom Docker image tag
-volumes: # List of persistent storage volumes
-env: # Environment variables to pass
-secrets: # Secrets to inject
+branch:      # Git branch to use
+docker_tag:  # Custom Docker image tag
+volumes:     # List of persistent storage volumes
+
+# Environment variables to pass. Can be specified in two ways:
+# 1. As a string: Will load the value from the host computer's environment variables
+# 2. As a key-value pair: Will use the specified value directly
+# Example:
+# env:
+#   - CUSTOM_VAR  # Loads from host's $CUSTOM_VAR
+#   - {CUSTOM_VAR: "value"}  # Uses "value" directly
+env:
+
+# Secrets to inject. Same input format as `env` but for sensitive data.
+secrets:
+  # - HF_TOKEN
+  # - WANDB_API_KEY
 ```

docs/config.qmd

@@ -238,10 +238,10 @@ simpo_gamma: 0.5  # Target reward margin for the SimPO loss
 # grpo
 trl:
   use_vllm: # Optional[bool]. Whether to use VLLM for RL training.
-  vllm_device: # Optional[str]. Device to use for VLLM.
-  vllm_gpu_memory_utilization: # Optional[float]. GPU memory utilization for VLLM.
-  vllm_max_model_len: # Optional[int]. Maximum length of the model for VLLM.
-  vllm_dtype: # Optional[str]. Data type for VLLM.
+  vllm_server_host: # Optional[str]. Host of the vLLM server to connect to.
+  vllm_server_port: # Optional[int]. Port of the vLLM server to connect to.
+  vllm_server_timeout: # Optional[int]. Total timeout (in seconds) to wait for the vLLM server to respond.
+  vllm_guided_decoding_regex: # Optional[str]. Regex for vLLM guided decoding.
 
   beta: # Optional[float]. Beta parameter for the RL training. Same as `rl_beta`. Use
   max_completion_length: # Optional[int]. Maximum length of the completion for RL training.
@@ -320,9 +320,13 @@ total_num_tokens:
 sample_packing_group_size: 100000
 # The number of samples which can be packed into one sequence. Increase if using a large sequence_len with many short samples.
 sample_packing_bin_size: 200
+sample_pack_sequentially: # Optional[bool]. Whether to pack samples sequentially.
+
 # whether to concatenate samples during pretraining
 pretraining_sample_concatenation:
 
+curriculum_sampling: # Optional[bool]. Whether to use sequential sampling for curriculum learning
+
 # Use batch flattening for speedups when not using sample_packing
 batch_flattening:
 
@@ -354,7 +358,27 @@ lora_target_modules:
 #  - down_proj
 #  - up_proj
 lora_target_linear: # If true, will target all linear modules
-peft_layers_to_transform: # The layer indices to transform, otherwise, apply to all layers
+
+# List[int] | int. # The layer indices to transform, otherwise, apply to all layers
+# https://huggingface.co/docs/peft/v0.15.0/en/package_reference/lora#peft.LoraConfig.layers_to_transform
+peft_layers_to_transform:
+
+# Optional[bool]. Whether to use DoRA.
+# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#weight-decomposed-low-rank-adaptation-dora
+peft_use_dora:
+
+# Optional[bool]. Whether to use RSLoRA.
+# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#rank-stabilized-lora
+peft_use_rslora:
+
+# Optional[list[tuple[int, int]]]. List of layer indices to replicate.
+# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#memory-efficient-layer-replication-with-lora
+peft_layer_replication:
+
+# bool | Literal["gaussian", "eva", "olora", "pissa", "pissa_niter_[number of iters]", "corda", "loftq"]
+# How to initialize LoRA weights. Default to True which is MS original implementation.
+# https://huggingface.co/docs/peft/v0.15.0/en/developer_guides/lora#initialization
+peft_init_lora_weights:
 
 # If you added new tokens to the tokenizer, you may need to save some LoRA modules because they need to know the new tokens.
 # For LLaMA and Mistral, you need to save `embed_tokens` and `lm_head`. It may vary for other models.
@@ -486,7 +510,8 @@ train_on_inputs: false
 # Note that training loss may have an oscillating pattern with this enabled.
 group_by_length: false
 
-# Whether to use gradient checkpointing https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
+# Whether to use gradient checkpointing. Available options are: true, false, "offload".
+# https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
 gradient_checkpointing: false
 # additional kwargs to pass to the trainer for gradient checkpointing
 # gradient_checkpointing_kwargs:
@@ -587,26 +612,31 @@ max_grad_norm:
 # currently only supported on Llama and Mistral
 neftune_noise_alpha:
 
-# Whether to bettertransformers
+# Optional[bool]. Whether to bettertransformers
 flash_optimum:
-# Whether to use xformers attention patch https://github.com/facebookresearch/xformers:
+
+# Note: Only one of the following attention patches can be used at a time.
+# For example, if you set `xformers_attention` to `true`, do not set `flash_attention` to `true`.
+
+# Optional[bool]. Whether to use xformers attention patch https://github.com/facebookresearch/xformers:
 xformers_attention:
-# Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:
+# Optional[bool]. Whether to use flash attention patch https://github.com/Dao-AILab/flash-attention:
 flash_attention:
-flash_attn_cross_entropy:  # Whether to use flash-attention cross entropy implementation - advanced use only
-flash_attn_rms_norm:  # Whether to use flash-attention rms norm implementation - advanced use only
-flash_attn_fuse_qkv: # Whether to fuse QKV into a single operation
-flash_attn_fuse_mlp: # Whether to fuse part of the MLP into a single operation
-# Whether to use scaled-dot-product attention
+flash_attn_cross_entropy:  # Optional[bool]. Whether to use flash-attention cross entropy implementation - advanced use only
+flash_attn_rms_norm:  # Optional[bool]. Whether to use flash-attention rms norm implementation - advanced use only
+flash_attn_fuse_qkv: # Optional[bool]. Whether to fuse QKV into a single operation
+flash_attn_fuse_mlp: # Optional[bool]. Whether to fuse part of the MLP into a single operation
+# Optional[bool]. Whether to use scaled-dot-product attention
 # https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
 sdp_attention:
-# Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf
+# Optional[bool]. Shifted-sparse attention (only llama) - https://arxiv.org/pdf/2309.12307.pdf
 s2_attention:
+
 # Optional[bool]. Whether to use low_cpu_mem_usage
 low_cpu_mem_usage:
-# Resume from a specific checkpoint dir
+# Optional[str]. Resume from a specific checkpoint dir
 resume_from_checkpoint:
-# If resume_from_checkpoint isn't set and you simply want it to start where it left off.
+# Optional[bool]. If resume_from_checkpoint isn't set and you simply want it to start where it left off.
 # Be careful with this being turned on between different models.
 auto_resume_from_checkpoints: false
 

docs/faq.qmd

@@ -35,12 +35,22 @@ description: Frequently asked questions
 
 **Q: How to call Axolotl via custom python scripts?**
 
-> A: Yes, since Axolotl is just Python, please see `src/axolotl/cli/main.py` on how each command is called.
+> A: Since Axolotl is just Python, please see `src/axolotl/cli/main.py` on how each command is called.
 
 **Q: How to know the value to use for `fsdp_transformer_layer_cls_to_wrap`?**
 
 > A: This is the class name of the transformer layer to wrap with FSDP. For example, for `LlamaForCausalLM`, the value is `LlamaDecoderLayer`. To find this for a specific model, check the model's `PreTrainedModel` definition and look for `_no_split_modules` variable in the `modeling_<model_name>.py` file within `transformers` library.
 
+**Q: ValueError: Asking to pad but the tokenizer does not have a padding token. Please select a token to use as pad_token**
+
+> A: This is because the tokenizer does not have a padding token. Please add a padding token to the tokenizer via:
+
+> ```yaml
+> special_tokens:
+>   # str. If you're not sure, set to same as `eos_token`.
+>   pad_token: "..."
+> ```
+
 ### Chat templates
 
 **Q: `jinja2.exceptions.UndefinedError: 'dict object' has no attribute 'content' / 'role' / ____`**

docs/lora_optims.qmd

@@ -17,6 +17,7 @@ We currently support several common model architectures, including (but not limi
 - `qwen2`
 - `gemma`
 - `gemma2`
+- `gemma3`
 
 <details>
 

docs/rlhf.qmd

@@ -502,9 +502,48 @@ The input format is a simple JSON input with customizable fields based on the ab
 Check out our [GRPO cookbook](https://github.com/axolotl-ai-cloud/axolotl-cookbook/tree/main/grpo#training-an-r1-style-large-language-model-using-grpo).
 :::
 
+If you have multiple GPUs available, we reccomend using `vLLM` with the `GRPOTrainer` to significantly speedup trajectory generation during training.
+First, launch a `vLLM` server using `trl vllm-serve` - you may use a config file or CLI overrides to configure your vLLM server. In this example, we're
+using 4 GPUs - 2 for training, and 2 for vLLM:
+
+::: {.callout-important}
+Make sure you've installed the correct version of vLLM by including it as an extra when installing axolotl, e.g. `pip install axolotl[vllm]`.
+:::
+
+```yaml
+base_model: Qwen/Qwen2.5-1.5B-Instruct
+
+vllm:
+    host: 0.0.0.0
+    port: 8000
+    tensor_parallel_size: 2
+    gpu_memory_utilization: 0.85
+    dtype: auto
+    # max_model_len: # you may find it useful to set the vLLM model context length if you know this beforehand
+
+rl: grpo
+trl:
+    use_vllm: true
+    vllm_server_host: 0.0.0.0
+    vllm_server_port: 8000
+    vllm_server_timeout: 300
+```
+
+```bash
+CUDA_VISIBLE_DEVICES=2,3 axolotl vllm_serve grpo.yaml
+```
+
+Your `vLLM` instance will now attempt to spin up, and it's time to kick off training utilizing our remaining two GPUs. In another terminal, execute:
+
+```bash
+CUDA_VISIBLE_DEVICES=0,1 axolotl train grpo.yaml --num-processes 2
+```
+
+#### Reward functions
+
 GRPO uses custom reward functions and transformations. Please have them ready locally.
 
-For ex, to load OpenAI's GSM8K and use a random reward for completions:
+For example, to load OpenAI's GSM8K and use a random reward for completions:
 
 ```python
 # rewards.py
@@ -530,8 +569,6 @@ trl:
     beta: 0.001
     max_completion_length: 256
     use_vllm: True
-    vllm_device: auto
-    vllm_gpu_memory_utilization: 0.15
     num_generations: 4
     reward_funcs: ["rewards.rand_reward_func"]    # format: '{file_name}.{fn_name}'
     reward_weights: [1.0]

examples/gemma3/gemma-3-1b-qlora.yml

@@ -5,6 +5,9 @@ tokenizer_type: AutoTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
+# gemma3 doesn't seem to play nice with ddp
+ddp_find_unused_parameters: true
+
 load_in_8bit: false
 load_in_4bit: true
 strict: false
@@ -54,6 +57,8 @@ fp16:
 tf32: true
 
 gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
 early_stopping_patience:
 resume_from_checkpoint:
 local_rank:

examples/gemma3/gemma-3-4b-qlora.yml

@@ -0,0 +1,66 @@
+base_model: google/gemma-3-4b-it
+strict: false
+
+# gemma3 doesn't seem to play nice with ddp
+ddp_find_unused_parameters: true
+
+chat_template: gemma3
+datasets:
+  - path: cgato/SlimOrcaDedupCleaned
+    type: chat_template
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+adapter: lora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+pad_to_sequence_len: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: 'language_model.model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: true
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+local_rank:
+logging_steps: 1
+flash_attention: true
+eager_attention:
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:

examples/gemma3/gemma-3-4b-vision-qlora.yml

@@ -7,6 +7,9 @@ skip_prepare_dataset: true
 remove_unused_columns: false
 sample_packing: false
 
+# gemma3 doesn't seem to play nice with ddp
+ddp_find_unused_parameters: true
+
 chat_template: gemma3
 datasets:
   - path: HuggingFaceH4/llava-instruct-mix-vsft
@@ -17,7 +20,7 @@ dataset_prepared_path: last_run_prepared
 val_set_size: 0.01
 output_dir: ./outputs/out
 
-adapter: lora
+adapter: qlora
 lora_model_dir:
 
 sequence_len: 2048
@@ -48,6 +51,8 @@ fp16:
 tf32: true
 
 gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
 local_rank:
 logging_steps: 1
 flash_attention: true

examples/llama-3/lora-1b-sample-packing-sequentially.yml

@@ -0,0 +1,80 @@
+base_model: meta-llama/Llama-3.2-1B
+# optionally might have model_type or tokenizer_type
+model_type: LlamaForCausalLM
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+  - path: mhenrichsen/alpaca_2k_test
+    type: alpaca
+dataset_prepared_path:
+val_set_size: 0.0
+output_dir: ./outputs/lora-out
+
+test_value: true
+
+sequence_len: 4096
+sample_packing: true
+sample_packing_sequentially: true
+curriculum_sampling: true
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_fan_in_fan_out:
+lora_modules_to_save:
+  - embed_tokens
+  - lm_head
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+train_on_inputs: false
+group_by_length: false
+bf16: auto
+fp16:
+tf32: false
+
+gradient_checkpointing: true
+early_stopping_patience:
+resume_from_checkpoint:
+local_rank:
+logging_steps: 1
+xformers_attention:
+flash_attention: true
+s2_attention:
+
+warmup_steps: 10
+evals_per_epoch: 4
+eval_table_size:
+eval_max_new_tokens: 128
+saves_per_epoch: 1
+debug:
+deepspeed:
+weight_decay: 0.0
+fsdp:
+fsdp_config:
+special_tokens:
+  pad_token: <|end_of_text|>

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==0.45.3
+bitsandbytes==0.45.4
 triton>=3.0.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
@@ -12,12 +12,12 @@ liger-kernel==0.5.5
 packaging==23.2
 
 peft==0.15.0
-transformers==4.50.0
+transformers==4.50.3
 tokenizers>=0.21.1
 accelerate==1.5.2
 datasets==3.5.0
-deepspeed==0.16.4
-trl==0.15.1
+deepspeed==0.15.4
+trl==0.16.0
 
 optimum==1.16.2
 hf_transfer

setup.py

@@ -10,7 +10,7 @@ from pathlib import Path
 from setuptools import find_packages, setup
 
 
-def parse_requirements():
+def parse_requirements(extras_require_map):
     _install_requires = []
     _dependency_links = []
     with open("./requirements.txt", encoding="utf-8") as requirements_file:
@@ -67,6 +67,7 @@ def parse_requirements():
             if (major, minor) >= (2, 6):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 _install_requires.append("xformers==0.0.29.post2")
+                extras_require_map["vllm"] = ["vllm==0.8.1"]
             elif (major, minor) >= (2, 5):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 if patch == 0:
@@ -86,7 +87,7 @@ def parse_requirements():
 
     except PackageNotFoundError:
         pass
-    return _install_requires, _dependency_links
+    return _install_requires, _dependency_links, extras_require_map
 
 
 def get_package_version():
@@ -103,7 +104,50 @@ def get_package_version():
     return version_
 
 
-install_requires, dependency_links = parse_requirements()
+extras_require = {
+    "flash-attn": ["flash-attn==2.7.4.post1"],
+    "ring-flash-attn": [
+        "flash-attn==2.7.4.post1",
+        "ring-flash-attn>=0.1.4",
+        "yunchang==0.6.0",
+    ],
+    "deepspeed": [
+        "deepspeed==0.15.4",
+        "deepspeed-kernels",
+    ],
+    "mamba-ssm": [
+        "mamba-ssm==1.2.0.post1",
+        "causal_conv1d",
+    ],
+    "auto-gptq": [
+        "auto-gptq==0.5.1",
+    ],
+    "mlflow": [
+        "mlflow",
+    ],
+    "galore": [
+        "galore_torch",
+    ],
+    "apollo": [
+        "apollo-torch",
+    ],
+    "optimizers": [
+        "galore_torch",
+        "apollo-torch",
+        "lomo-optim==0.1.1",
+        "torch-optimi==0.2.1",
+    ],
+    "ray": [
+        "ray[train]",
+    ],
+    "vllm": [
+        "vllm==0.7.2",
+    ],
+}
+
+install_requires, dependency_links, extras_require_build = parse_requirements(
+    extras_require
+)
 
 setup(
     version=get_package_version(),
@@ -116,40 +160,5 @@ setup(
             "axolotl=axolotl.cli.main:main",
         ],
     },
-    extras_require={
-        "flash-attn": ["flash-attn==2.7.4.post1"],
-        "ring-flash-attn": ["ring-flash-attn>=0.1.4", "yunchang==0.6.0"],
-        "deepspeed": [
-            "deepspeed==0.16.4",
-            "deepspeed-kernels",
-        ],
-        "mamba-ssm": [
-            "mamba-ssm==1.2.0.post1",
-            "causal_conv1d",
-        ],
-        "auto-gptq": [
-            "auto-gptq==0.5.1",
-        ],
-        "mlflow": [
-            "mlflow",
-        ],
-        "galore": [
-            "galore_torch",
-        ],
-        "apollo": [
-            "apollo-torch",
-        ],
-        "optimizers": [
-            "galore_torch",
-            "apollo-torch",
-            "lomo-optim==0.1.1",
-            "torch-optimi==0.2.1",
-        ],
-        "ray": [
-            "ray[train]",
-        ],
-        "vllm": [
-            "vllm==0.7.2",
-        ],
-    },
+    extras_require=extras_require_build,
 )

src/axolotl/__init__.py

@@ -4,4 +4,4 @@ import pkgutil
 
 __path__ = pkgutil.extend_path(__path__, __name__)  # Make this a namespace package
 
-__version__ = "0.8.0.dev0"
+__version__ = "0.8.0"

src/axolotl/cli/args.py

@@ -35,6 +35,55 @@ class TrainerCliArgs:
     num_processes: Optional[int] = field(default=None)
 
 
+@dataclass
+class VllmServeCliArgs:
+    """Dataclass with CLI arguments for `axolotl vllm-serve` command."""
+
+    tensor_parallel_size: int = field(
+        default=1,
+        metadata={"help": "Number of tensor parallel workers to use."},
+    )
+    host: str = field(
+        default="0.0.0.0",  # nosec B104
+        metadata={"help": "Host address to run the server on."},
+    )
+    port: int = field(
+        default=8000,
+        metadata={"help": "Port to run the server on."},
+    )
+    gpu_memory_utilization: Optional[float] = field(
+        default=None,
+        metadata={
+            "help": "Ratio (between 0 and 1) of GPU memory to reserve for the model weights, activations, and KV "
+            "cache on the device dedicated to generation powered by vLLM. Higher values will increase the KV cache "
+            "size and thus improve the model's throughput. However, if the value is too high, it may cause "
+            "out-of-memory (OOM) errors during initialization."
+        },
+    )
+    dtype: Optional[str] = field(
+        default=None,
+        metadata={
+            "help": "Data type to use for vLLM generation. If set to 'auto', the data type will be automatically "
+            "determined based on the model configuration. Find the supported values in the vLLM documentation."
+        },
+    )
+    max_model_len: Optional[int] = field(
+        default=None,
+        metadata={
+            "help": "If set, the `max_model_len` to use for vLLM. This can be useful when running with reduced "
+            "`vllm_gpu_memory_utilization`, leading to a reduced KV cache size. If not set, vLLM will use the model "
+            "context size, which might be much larger than the KV cache, leading to inefficiencies."
+        },
+    )
+    enable_prefix_caching: Optional[bool] = field(
+        default=None,
+        metadata={
+            "help": "Whether to enable prefix caching in vLLM. If set to `True`, ensure that the model and the "
+            "hardware support this feature."
+        },
+    )
+
+
 @dataclass
 class EvaluateCliArgs:
     """Dataclass with CLI arguments for `axolotl evaluate` command."""

src/axolotl/cli/inference.py

@@ -256,7 +256,7 @@ def do_cli(
     """
     # pylint: disable=duplicate-code
     print_axolotl_text_art()
-    parsed_cfg = load_cfg(config, inference=True, **kwargs)
+    parsed_cfg = load_cfg(config, inference=True, rl=None, **kwargs)
     parsed_cfg.sample_packing = False
     parser = transformers.HfArgumentParser(InferenceCliArgs)
     parsed_cli_args, _ = parser.parse_args_into_dataclasses(

src/axolotl/cli/main.py

@@ -14,7 +14,12 @@ import yaml
 from dotenv import load_dotenv
 
 import axolotl
-from axolotl.cli.args import EvaluateCliArgs, PreprocessCliArgs, TrainerCliArgs
+from axolotl.cli.args import (
+    EvaluateCliArgs,
+    PreprocessCliArgs,
+    TrainerCliArgs,
+    VllmServeCliArgs,
+)
 from axolotl.cli.sweeps import generate_sweep_configs
 from axolotl.cli.utils import (
     add_options_from_config,
@@ -23,6 +28,7 @@ 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
@@ -316,6 +322,14 @@ def fetch(directory: str, dest: Optional[str]) -> None:
     fetch_from_github(f"{directory}/", dest)
 
 
+@cli.command()
+@click.argument("config", type=click.Path(exists=True, path_type=str))
+@add_options_from_dataclass(VllmServeCliArgs)
+@filter_none_kwargs
+def vllm_serve(config: str, **cli_args: VllmServeCliArgs):
+    do_vllm_serve(config, cli_args)
+
+
 cli.add_command(lm_eval)
 
 

src/axolotl/cli/merge_lora.py

@@ -74,8 +74,10 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs) -> None:
         load_in_8bit=False,
         load_in_4bit=False,
         flash_attention=False,
+        sequence_parallel_degree=None,
         deepspeed=None,
         fsdp=None,
+        fsdp_config=None,
         **kwargs,
     )
 
@@ -86,13 +88,6 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs) -> None:
             f"Target directory for merge: `{parsed_cfg.lora_model_dir}` does not exist."
         )
 
-    parsed_cfg.load_in_4bit = False
-    parsed_cfg.load_in_8bit = False
-    parsed_cfg.flash_attention = False
-    parsed_cfg.deepspeed = None
-    parsed_cfg.fsdp = None
-    parsed_cfg.fsdp_config = None
-
     do_merge_lora(cfg=parsed_cfg)
 
 

src/axolotl/cli/vllm_serve.py

@@ -0,0 +1,55 @@
+"""
+CLI to start the vllm server for online RL
+"""
+
+from pathlib import Path
+from typing import Union
+
+from trl.scripts.vllm_serve import ScriptArguments
+from trl.scripts.vllm_serve import main as vllm_serve_main
+
+from axolotl.cli.config import load_cfg
+
+
+def do_vllm_serve(
+    config: Union[Path, str],
+    cli_args: dict,
+):
+    """
+    Starts the VLLM server for serving LLM models used for online RL
+
+    Args
+        :param cfg: Parsed doct of the YAML config
+        :param cli_args: dict of additional command-line arguments of type VllmServeCliArgs
+
+    Returns:
+        process_id: the process id of the started VLLM server
+    """
+    cfg = load_cfg(config)
+    model = cfg.base_model
+
+    tensor_parallel_size = (
+        cli_args.get("tensor_parallel_size") or cfg.vllm.tensor_parallel_size
+    )
+    host = cli_args.get("host") or cfg.vllm.host
+    port = cli_args.get("port") or cfg.vllm.port
+    gpu_memory_utilization = (
+        cli_args.get("gpu_memory_utilization") or cfg.vllm.gpu_memory_utilization
+    )
+    dtype = cli_args.get("dtype") or cfg.vllm.dtype
+    max_model_len = cli_args.get("max_model_len") or cfg.vllm.max_model_len
+    enable_prefix_caching = (
+        cli_args.get("enable_prefix_caching") or cfg.vllm.enable_prefix_caching
+    )
+
+    vllm_script_args = ScriptArguments(
+        model,
+        tensor_parallel_size=tensor_parallel_size,
+        host=host,
+        port=port,
+        gpu_memory_utilization=gpu_memory_utilization,
+        dtype=dtype,
+        max_model_len=max_model_len,
+        enable_prefix_caching=enable_prefix_caching,
+    )
+    vllm_serve_main(vllm_script_args)

src/axolotl/core/trainer_builder.py

@@ -524,9 +524,15 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             and self.cfg.eval_steps
             and self.cfg.save_steps % self.cfg.eval_steps == 0
         ) or False
+
+        # handle ddp
+        ddp_find_unused_parameters = None
+        if self.cfg.ddp:
+            ddp_find_unused_parameters = bool(self.cfg.ddp_find_unused_parameters)
         training_arguments_kwargs["ddp_find_unused_parameters"] = (
-            False if self.cfg.ddp else None
+            ddp_find_unused_parameters
         )
+
         training_arguments_kwargs["group_by_length"] = self.cfg.group_by_length
         training_arguments_kwargs["curriculum_sampling"] = self.cfg.curriculum_sampling
         report_to = []

src/axolotl/core/trainers/base.py

@@ -8,12 +8,11 @@ import logging
 import os
 from collections import defaultdict
 from functools import wraps
-from typing import Any, Literal
+from typing import Literal
 
 import datasets
 import torch
 from datasets import Dataset
-from torch import nn
 from torch.utils.data import (
     BatchSampler,
     DataLoader,
@@ -28,6 +27,7 @@ from typing_extensions import override
 
 from axolotl.core.trainers.mixins import (
     OptimizerMixin,
+    RngLoaderMixin,
     SchedulerMixin,
     SequenceParallelMixin,
 )
@@ -40,7 +40,9 @@ from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
 LOG = logging.getLogger(__name__)
 
 
-class AxolotlTrainer(SchedulerMixin, OptimizerMixin, SequenceParallelMixin, Trainer):
+class AxolotlTrainer(
+    SchedulerMixin, OptimizerMixin, RngLoaderMixin, SequenceParallelMixin, Trainer
+):
     """Extend the base Trainer for axolotl helpers"""
 
     args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
@@ -112,6 +114,7 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, SequenceParallelMixin, Trai
             packing_efficiency_estimate=self.args.sample_packing_efficiency,
             batch_max_len=batch_max_len,
             batch_size=batch_size,
+            sequential=self.args.sample_packing_sequentially,
             drop_last=True,
         )
 
@@ -589,27 +592,3 @@ class AxolotlTrainer(SchedulerMixin, OptimizerMixin, SequenceParallelMixin, Trai
         output_dir = os.path.join(run_dir, checkpoint_folder)
         os.makedirs(output_dir, exist_ok=True)
         return super()._save_checkpoint(model, trial, **kwargs)
-
-    def training_step(
-        self,
-        model: nn.Module,
-        inputs: dict[str, torch.Tensor | Any],
-        num_items_in_batch: int | None = None,
-    ) -> torch.Tensor:
-        """
-        Perform a training step on a batch of inputs. Overrides the
-        `transformers.trainer.Trainer` method to handle sequence parallelism if
-        enabled.
-
-        Args:
-            model: Model to perform training step for.
-            inputs: Dictionary mapping.
-        """
-        # Set up sequence parallelism for this step if enabled
-        if self.args.sequence_parallel_degree > 1:
-            self._update_ring_flash_attn_params(inputs)
-
-        # Proceed with normal training step
-        loss = super().training_step(model, inputs, num_items_in_batch)
-
-        return loss

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

@@ -13,7 +13,7 @@ from transformers import Trainer
 from transformers.utils import is_sagemaker_mp_enabled
 from trl import DPOTrainer
 
-from axolotl.core.trainers.mixins import SchedulerMixin
+from axolotl.core.trainers.mixins import RngLoaderMixin, SchedulerMixin
 from axolotl.core.trainers.utils import (
     sanitize_kwargs_for_ds_tagging,
     sanitize_kwargs_for_tagging,
@@ -23,7 +23,7 @@ if is_sagemaker_mp_enabled():
     import smdistributed.modelparallel.torch as smp
 
 
-class AxolotlDPOTrainer(SchedulerMixin, DPOTrainer):
+class AxolotlDPOTrainer(RngLoaderMixin, SchedulerMixin, DPOTrainer):
     """
     Extend the base DPOTrainer for axolotl helpers
     """

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

@@ -40,18 +40,15 @@ class GRPOStrategy:
 
         if trl.use_vllm:
             grpo_args_kwargs["use_vllm"] = trl.use_vllm
-            grpo_args_kwargs["vllm_device"] = (
-                trl.vllm_device if trl.vllm_device else "auto"
-            )
-
-            if trl.vllm_gpu_memory_utilization:
-                grpo_args_kwargs["vllm_gpu_memory_utilization"] = (
-                    trl.vllm_gpu_memory_utilization
+            grpo_args_kwargs["vllm_server_host"] = trl.vllm_server_host
+            grpo_args_kwargs["vllm_server_port"] = trl.vllm_server_port
+            if trl.vllm_server_timeout:
+                grpo_args_kwargs["vllm_server_timeout"] = trl.vllm_server_timeout
+            if trl.vllm_guided_decoding_regex:
+                grpo_args_kwargs["vllm_guided_decoding_regex"] = (
+                    trl.vllm_guided_decoding_regex
                 )
 
-            if trl.vllm_max_model_len:
-                grpo_args_kwargs["vllm_max_model_len"] = trl.vllm_max_model_len
-
         if trl.num_generations:
             grpo_args_kwargs["num_generations"] = trl.num_generations
 
@@ -70,6 +67,25 @@ class GRPOStrategy:
         if trl.reward_weights:
             grpo_args_kwargs["reward_weights"] = trl.reward_weights
 
+        if trl.scale_rewards is not None:
+            grpo_args_kwargs["scale_rewards"] = trl.scale_rewards
+
+        if trl.temperature is not None:
+            grpo_args_kwargs["temperature"] = trl.temperature
+        if trl.top_p is not None:
+            grpo_args_kwargs["top_p"] = trl.top_p
+        if trl.top_k is not None:
+            grpo_args_kwargs["top_k"] = trl.top_k
+        if trl.min_p is not None:
+            grpo_args_kwargs["min_p"] = trl.min_p
+        if trl.repetition_penalty is not None:
+            grpo_args_kwargs["repetition_penalty"] = trl.repetition_penalty
+
+        if trl.num_iterations is not None:
+            grpo_args_kwargs["num_iterations"] = trl.num_iterations
+        if trl.epsilon is not None:
+            grpo_args_kwargs["epsilon"] = trl.epsilon
+
         return grpo_args_kwargs
 
     @classmethod

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

@@ -2,108 +2,68 @@
 Axolotl GRPO trainer
 """
 
-from accelerate.utils import is_peft_model
-from accelerate.utils.other import is_compiled_module
-from transformers import PreTrainedModel
-from trl import GRPOConfig, GRPOTrainer
-from trl.models import unwrap_model_for_generation
+from contextlib import nullcontext
 
-from axolotl.core.trainers.base import SchedulerMixin
+from accelerate.utils import is_deepspeed_available, is_peft_model
+from trl import GRPOTrainer
+from trl.extras.profiling import profiling_decorator
+
+from axolotl.core.trainers.mixins import RngLoaderMixin, SchedulerMixin
+
+if is_deepspeed_available():
+    import deepspeed
 
 
-# mypy: ignore-errors
-class AxolotlGRPOTrainer(SchedulerMixin, GRPOTrainer):
+class AxolotlGRPOTrainer(RngLoaderMixin, SchedulerMixin, GRPOTrainer):
     """
     Extend the base GRPOTrainer for axolotl helpers
     """
 
     _tag_names = ["trl", "grpo", "axolotl"]
 
-    def __init__(self, *args, **kwargs):
-        super().__init__(*args, **kwargs)
-
-        # pylint: disable=access-member-before-definition
-        # Enable gradient checkpointing if requested
-        if kwargs["args"].gradient_checkpointing:
-            # Ensure use_cache is disabled
-            if hasattr(self.model, "config"):
-                self.model.config.use_cache = False
-
-            # Enable gradient checkpointing on the base model for PEFT
-            if is_peft_model(self.model) and hasattr(
-                self.model.base_model, "gradient_checkpointing_enable"
-            ):
-                self.model.base_model.gradient_checkpointing_enable()
-            # Enable gradient checkpointing for non-PEFT models
-            elif hasattr(self.model, "gradient_checkpointing_enable"):
-                self.model.gradient_checkpointing_enable()
-            self.model = self._enable_gradient_checkpointing(self.model, kwargs["args"])
-        # pylint: enable=access-member-before-definition
-
-    def _enable_gradient_checkpointing(
-        self, model: PreTrainedModel, args: GRPOConfig
-    ) -> PreTrainedModel:
-        """Enables gradient checkpointing for the model."""
-        # pylint: disable=unused-argument,redefined-builtin
-        gradient_checkpointing_kwargs = args.gradient_checkpointing_kwargs or {}
-        use_reentrant = (
-            "use_reentrant" not in gradient_checkpointing_kwargs
-            or gradient_checkpointing_kwargs["use_reentrant"]
+    @profiling_decorator
+    def _move_model_to_vllm(self):
+        # For DeepSpeed ZeRO-3, we need to gather all parameters before operations
+        deepspeed_plugin = self.accelerator.state.deepspeed_plugin
+        zero_stage_3 = deepspeed_plugin is not None and deepspeed_plugin.zero_stage == 3
+        gather_if_zero3 = (
+            deepspeed.zero.GatheredParameters if zero_stage_3 else nullcontext
         )
 
-        if use_reentrant:
-            if hasattr(model, "enable_input_require_grads"):
-                model.enable_input_require_grads()
-            else:
+        if is_peft_model(self.model):
+            # With PEFT and DeepSpeed ZeRO Stage 3, we must gather the full model at once before merging, as merging
+            # adapters in a sharded manner is not supported.
+            with gather_if_zero3(list(self.model.parameters())):
+                self.model.merge_adapter()
 
-                def make_inputs_require_grad(module, input, output):
-                    output.requires_grad_(True)
+                # Update vLLM weights while parameters are gathered
+                for name, param in self.model.named_parameters():
+                    # When using PEFT, we need to recover the original parameter name and discard some parameters
+                    name = (
+                        name.removeprefix("base_model.model.")
+                        .removeprefix("base_model.model.")
+                        .replace(".base_layer", "")
+                    )
+                    if self.model.prefix in name:
+                        continue
+                    # When module to save, remove its prefix and discard the original module
+                    if "original_module" in name:
+                        continue
+                    name = name.replace("modules_to_save.default.", "")
 
-                model.get_input_embeddings().register_forward_hook(
-                    make_inputs_require_grad
-                )
+                    if self.accelerator.is_main_process:
+                        self.vllm_client.update_named_param(name, param.data)
 
-        return model
-        # pylint: enable=unused-argument,redefined-builtin
+                # Unmerge adapters while parameters are still gathered
+                self.model.unmerge_adapter()
+                # Parameters will automatically be repartitioned when exiting the context
+        else:
+            # For non-PEFT models, simply gather and update each parameter individually.
+            for name, param in self.model.named_parameters():
+                with gather_if_zero3([param]):
+                    if self.accelerator.is_main_process:
+                        self.vllm_client.update_named_param(name, param.data)
 
-    def _move_model_to_vllm(self):
-        with unwrap_model_for_generation(
-            self.model,
-            self.accelerator,
-            gather_deepspeed3_params=self.args.ds3_gather_for_generation,
-        ) as unwrapped_model:
-            if is_compiled_module(unwrapped_model):
-                unwrapped_model = (
-                    unwrapped_model._orig_mod  # pylint: disable=protected-access
-                )
-            if is_peft_model(unwrapped_model):
-                unwrapped_model.merge_adapter()
-                state_dict = unwrapped_model.state_dict()
-                # Remove base_model and base_layer prefixes
-                state_dict = {
-                    k.removeprefix("base_model.model.")
-                    .removeprefix("base_model.model.")
-                    .replace(".base_layer", ""): v
-                    for k, v in state_dict.items()
-                }
-                # Remove values with adapter prefix (example: "_lora")
-                state_dict = {
-                    k: v
-                    for k, v in state_dict.items()
-                    if unwrapped_model.prefix not in k
-                }
-                # When module to save, remove its prefix and discard the original module
-                state_dict = {
-                    k.replace("modules_to_save.default.", ""): v
-                    for k, v in state_dict.items()
-                    if "original_module" not in k
-                }
-            else:
-                state_dict = unwrapped_model.state_dict()
-            if self.accelerator.is_main_process:
-                llm_model = (
-                    self.llm.llm_engine.model_executor.driver_worker.model_runner.model
-                )
-                llm_model.load_weights(state_dict.items())
-            if is_peft_model(unwrapped_model):
-                unwrapped_model.unmerge_adapter()
+        # Reset cache on main process
+        if self.accelerator.is_main_process:
+            self.vllm_client.reset_prefix_cache()

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

@@ -4,5 +4,6 @@
 # flake8: noqa
 
 from .optimizer import OptimizerMixin
+from .rng_state_loader import RngLoaderMixin
 from .scheduler import SchedulerMixin
 from .sequence_parallel import SequenceParallelMixin

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

@@ -0,0 +1,67 @@
+"""
+Temporary fix/override for bug in resume from checkpoint
+
+See https://github.com/huggingface/transformers/pull/37162
+
+TODO: Remove when upstream added PR to release
+"""
+
+import logging
+import os
+import random
+
+import numpy as np
+import torch
+from transformers import Trainer, is_torch_npu_available
+from transformers.trainer import safe_globals
+from transformers.trainer_pt_utils import set_rng_state_for_device
+from transformers.training_args import ParallelMode
+
+LOG = logging.getLogger(__name__)
+
+
+class RngLoaderMixin(Trainer):
+    """
+    mixin for method override to load RNG states from a checkpoint
+    """
+
+    def _load_rng_state(self, checkpoint):
+        # Load RNG states from `checkpoint`
+        if checkpoint is None:
+            return
+
+        if self.args.world_size > 1:
+            process_index = self.args.process_index
+            rng_file = os.path.join(checkpoint, f"rng_state_{process_index}.pth")
+            if not os.path.isfile(rng_file):
+                LOG.info(
+                    f"Didn't find an RNG file for process {process_index}, if you are resuming a training that "
+                    "wasn't launched in a distributed fashion, reproducibility is not guaranteed."
+                )
+                return
+        else:
+            rng_file = os.path.join(checkpoint, "rng_state.pth")
+            if not os.path.isfile(rng_file):
+                LOG.info(
+                    "Didn't find an RNG file, if you are resuming a training that was launched in a distributed "
+                    "fashion, reproducibility is not guaranteed."
+                )
+                return
+
+        # Use safe_globals to ensure numpy RNG states can be deserialized safely under PyTorch 2.6+,
+        # which requires allowlisted classes when loading with weights_only=True.
+        with safe_globals():
+            checkpoint_rng_state = torch.load(rng_file)  # nosec B614
+        random.setstate(checkpoint_rng_state["python"])
+        np.random.set_state(checkpoint_rng_state["numpy"])
+        torch.random.set_rng_state(checkpoint_rng_state["cpu"])
+
+        is_distributed = self.args.parallel_mode == ParallelMode.DISTRIBUTED
+        if torch.cuda.is_available():
+            set_rng_state_for_device(
+                "CUDA", torch.cuda, checkpoint_rng_state, is_distributed
+            )
+        if is_torch_npu_available():
+            set_rng_state_for_device(
+                "NPU", torch.npu, checkpoint_rng_state, is_distributed
+            )

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

@@ -7,6 +7,7 @@ import torch
 import torch.distributed as dist
 import torch.nn.functional as F
 from datasets import Dataset
+from torch import nn
 from torch.utils.data import DistributedSampler, Sampler
 
 from axolotl.monkeypatch.attention.ring_attn import get_ring_attn_group
@@ -129,3 +130,53 @@ class SequenceParallelMixin:
         )
 
         update_ring_flash_attn_params(cu_seqlens, self.ring_attn_group)
+
+    def training_step(
+        self,
+        model: nn.Module,
+        inputs: dict[str, torch.Tensor | Any],
+        num_items_in_batch: int | None = None,
+    ) -> torch.Tensor:
+        """
+        Perform a training step on a batch of inputs. Overrides the
+        `transformers.trainer.Trainer` method to handle sequence parallelism if
+        enabled.
+
+        Args:
+            model: Model to perform training step for.
+            inputs: Dictionary mapping.
+        """
+        # Set up sequence parallelism for this step if enabled
+        if self.args.sequence_parallel_degree > 1:
+            self._update_ring_flash_attn_params(inputs)
+
+        # Proceed with normal training step
+        return super().training_step(model, inputs, num_items_in_batch)  # type: ignore
+
+    def prediction_step(
+        self,
+        model: nn.Module,
+        inputs: dict[str, torch.Tensor | Any],
+        prediction_loss_only: bool,
+        ignore_keys: list[str] | None = None,
+    ) -> tuple[torch.Tensor | None, torch.Tensor | None, torch.Tensor | None]:
+        """
+        Perform a prediction step on a batch of inputs. Overrides the
+        `transformers.trainer.Trainer` method to handle sequence parallelism if
+        enabled.
+
+        Args:
+            model: Model to perform prediction step for.
+            inputs: Dictionary mapping of inputs.
+            prediction_loss_only: Whether to return only the loss.
+            ignore_keys: Keys to ignore in the inputs.
+
+        Returns:
+            Tuple of (loss, logits, labels).
+        """
+        # Set up sequence parallelism for this prediction step if enabled
+        if self.args.sequence_parallel_degree > 1:
+            self._update_ring_flash_attn_params(inputs)
+
+        # Proceed with normal prediction step
+        return super().prediction_step(model, inputs, prediction_loss_only, ignore_keys)  # type: ignore

src/axolotl/core/trainers/trl.py

@@ -13,6 +13,7 @@ from trl import (
     RewardTrainer,
 )
 
+from axolotl.core.trainers.mixins import RngLoaderMixin
 from axolotl.core.trainers.mixins.scheduler import SchedulerMixin
 
 
@@ -74,7 +75,7 @@ class TRLPPOTrainer(PPOTrainer):
             )
 
 
-class AxolotlORPOTrainer(SchedulerMixin, ORPOTrainer):
+class AxolotlORPOTrainer(RngLoaderMixin, SchedulerMixin, ORPOTrainer):
     """
     Extend the base ORPOTrainer for axolotl helpers
     """
@@ -154,7 +155,7 @@ class AxolotlORPOTrainer(SchedulerMixin, ORPOTrainer):
         return loss, metrics
 
 
-class AxolotlKTOTrainer(SchedulerMixin, KTOTrainer):
+class AxolotlKTOTrainer(RngLoaderMixin, SchedulerMixin, KTOTrainer):
     """
     Extend the base KTOTrainer for axolotl helpers
     """
@@ -162,7 +163,7 @@ class AxolotlKTOTrainer(SchedulerMixin, KTOTrainer):
     tag_names = ["axolotl", "kto"]
 
 
-class AxolotlCPOTrainer(SchedulerMixin, CPOTrainer):
+class AxolotlCPOTrainer(RngLoaderMixin, SchedulerMixin, CPOTrainer):
     """
     Extend the base CPOTrainer for axolotl helpers
     """
@@ -244,7 +245,7 @@ class AxolotlCPOTrainer(SchedulerMixin, CPOTrainer):
         return loss, metrics
 
 
-class AxolotlRewardTrainer(SchedulerMixin, RewardTrainer):
+class AxolotlRewardTrainer(RngLoaderMixin, SchedulerMixin, RewardTrainer):
     """
     Extend the base RewardTrainer for axolotl helpers
     """
@@ -252,7 +253,7 @@ class AxolotlRewardTrainer(SchedulerMixin, RewardTrainer):
     tag_names = ["axolotl", "reward"]
 
 
-class AxolotlPRMTrainer(SchedulerMixin, PRMTrainer):
+class AxolotlPRMTrainer(RngLoaderMixin, SchedulerMixin, PRMTrainer):
     """
     Extend the base trl.PRMTrainer for axolotl helpers
     """

src/axolotl/core/training_args.py

@@ -34,6 +34,12 @@ class AxolotlTrainingMixins:
         default=False,
         metadata={"help": "Use sample packing for efficient training."},
     )
+    sample_packing_sequentially: bool = field(
+        default=False,
+        metadata={
+            "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."
+        },
+    )
     multipack_real_batches: bool = field(
         default=False,
         metadata={"help": "Use real batches for efficient training."},

src/axolotl/monkeypatch/gemma3.py

@@ -0,0 +1,238 @@
+"""Monkeypatch for gemma3 conditional generation forward to fix loss exploding"""
+
+# pylint: disable=duplicate-code
+
+from typing import Optional, Tuple, Union
+
+import torch
+from transformers.cache_utils import Cache
+from transformers.models.gemma3.modeling_gemma3 import (
+    _CONFIG_FOR_DOC,
+    GEMMA3_INPUTS_DOCSTRING,
+    Gemma3CausalLMOutputWithPast,
+    logger,
+)
+from transformers.utils import (
+    add_start_docstrings_to_model_forward,
+    is_torchdynamo_compiling,
+    replace_return_docstrings,
+)
+from transformers.utils.deprecation import deprecate_kwarg
+
+
+@deprecate_kwarg("num_logits_to_keep", version="4.50", new_name="logits_to_keep")
+@add_start_docstrings_to_model_forward(GEMMA3_INPUTS_DOCSTRING)
+@replace_return_docstrings(
+    output_type=Gemma3CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC
+)
+def new_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    pixel_values: torch.FloatTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Union[list[torch.FloatTensor], Cache]] = None,
+    token_type_ids: Optional[torch.LongTensor] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    inputs_embeds: Optional[torch.FloatTensor] = None,
+    labels: Optional[torch.LongTensor] = None,
+    use_cache: Optional[bool] = None,
+    output_attentions: Optional[bool] = None,
+    output_hidden_states: Optional[bool] = None,
+    return_dict: Optional[bool] = None,
+    logits_to_keep: Union[int, torch.Tensor] = 0,
+    **lm_kwargs,
+) -> Union[Tuple, Gemma3CausalLMOutputWithPast]:
+    r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+            config.text_config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.text_config.vocab_size]`.
+
+        logits_to_keep (`int` or `torch.Tensor`, *optional*):
+            If an `int`, compute logits for the last `logits_to_keep` tokens. If `0`, calculate logits for all
+            `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that
+            token can save memory, which becomes pretty significant for long sequences or large vocabulary size.
+            If a `torch.Tensor`, must be 1D corresponding to the indices to keep in the sequence length dimension.
+            This is useful when using packed tensor format (single dimension for batch and sequence length).
+
+    Returns:
+
+    Example:
+
+    ```python
+    >>> from PIL import Image
+    >>> import requests
+    >>> from transformers import AutoProcessor, Gemma3ForConditionalGeneration
+
+    >>> model = Gemma3ForConditionalGeneration.from_pretrained("google/Gemma3-test-224px-hf")
+    >>> processor = AutoProcessor.from_pretrained("google/Gemma3-test-224px-hf")
+
+    >>> prompt = "answer en Where is the cow standing?"
+    >>> url = "https://huggingface.co/gv-hf/Gemma3-test-224px-hf/resolve/main/cow_beach_1.png"
+    >>> image = Image.open(requests.get(url, stream=True).raw)
+
+    >>> inputs = processor(images=image, text=prompt,  return_tensors="pt")
+
+    >>> # Generate
+    >>> generate_ids = model.generate(**inputs, max_length=30)
+    >>> processor.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+    "answer en Where is the cow standing?\nbeach"
+    ```"""
+
+    if (input_ids is None) ^ (inputs_embeds is not None):
+        raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+    output_attentions = (
+        output_attentions
+        if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states
+        if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    is_training = token_type_ids is not None and labels is not None
+
+    # Replace image id with PAD if the image token is OOV, to avoid index-errors
+    if input_ids is not None and self.config.image_token_index >= self.vocab_size:
+        special_image_mask = input_ids == self.config.image_token_index
+        llm_input_ids = input_ids.clone()
+        llm_input_ids[special_image_mask] = 0
+    else:
+        llm_input_ids = input_ids
+
+    if inputs_embeds is None:
+        inputs_embeds = self.get_input_embeddings()(llm_input_ids)
+
+    if cache_position is None:
+        past_seen_tokens = (
+            past_key_values.get_seq_length() if past_key_values is not None else 0
+        )
+        cache_position = torch.arange(
+            past_seen_tokens,
+            past_seen_tokens + inputs_embeds.shape[1],
+            device=inputs_embeds.device,
+        )
+
+    # Merge text and images
+    if pixel_values is not None:
+        image_features = self.get_image_features(pixel_values)
+
+        if input_ids is None:
+            special_image_mask = inputs_embeds == self.get_input_embeddings()(
+                torch.tensor(
+                    self.config.image_token_index,
+                    dtype=torch.long,
+                    device=inputs_embeds.device,
+                )
+            )
+        else:
+            special_image_mask = (input_ids == self.config.image_token_index).unsqueeze(
+                -1
+            )
+            special_image_mask = special_image_mask.expand_as(inputs_embeds).to(
+                inputs_embeds.device
+            )
+
+        if (
+            not is_torchdynamo_compiling()
+            and inputs_embeds[special_image_mask].numel() != image_features.numel()
+        ):
+            image_tokens_in_text = (special_image_mask).sum(dim=1).sum(dim=0)[0]
+            raise ValueError(
+                f"Number of images does not match number of special image tokens in the input text. "
+                f"Got {image_tokens_in_text} image tokens in the text but {image_features.shape[0] * image_features.shape[1]} "
+                "tokens from image embeddings."
+            )
+        image_features = image_features.to(inputs_embeds.device, inputs_embeds.dtype)
+        inputs_embeds = inputs_embeds.masked_scatter(special_image_mask, image_features)
+
+    # mask out pad-token-ids in labels for BC
+    if labels is not None and self.pad_token_id in labels:
+        logger.warning_once(
+            "`labels` contains `pad_token_id` which will be masked with `config.ignore_index`. "
+            "You have to mask out `pad_token_id` when preparing `labels`, this behavior will be removed in v.4.46.",
+        )
+        labels = torch.where(
+            input_ids == self.pad_token_id, self.config.ignore_index, labels
+        )
+
+    causal_mask = self._update_causal_mask(  # pylint: disable=protected-access
+        attention_mask,
+        token_type_ids,
+        past_key_values,
+        cache_position,
+        inputs_embeds,
+        is_training,
+    )
+    outputs = self.language_model(
+        attention_mask=causal_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        return_dict=return_dict,
+        cache_position=cache_position,
+        logits_to_keep=logits_to_keep,
+        **lm_kwargs,
+    )
+
+    logits = outputs[0]
+    loss = None
+    if labels is not None:
+        if attention_mask is not None:
+            # Get the shifted attention mask
+            shift_attention_mask = attention_mask[:, -logits.shape[1] + 1 :].to(
+                logits.device
+            )  # +1 for shift
+
+            # Filter logits and labels based on attention mask
+            valid_indices = shift_attention_mask != 0
+            filtered_logits = logits[..., :-1, :][valid_indices]
+            filtered_labels = labels[..., 1:][valid_indices.to(labels.device)]
+
+            # TODO: do we need to handle num_items_in_batch given we filter the logits and labels?
+
+            loss = self.loss_function(
+                logits=filtered_logits,
+                labels=None,  # we pass shift_labels
+                shift_labels=filtered_labels,
+                vocab_size=self.config.text_config.vocab_size,
+                **lm_kwargs,
+            )
+        else:
+            # Standard case without filtering
+            loss = self.loss_function(
+                logits=logits,
+                labels=labels,
+                vocab_size=self.config.text_config.vocab_size,
+                **lm_kwargs,
+            )
+    if not return_dict:
+        output = (logits,) + outputs[1:]
+        return (loss,) + output if loss is not None else output
+
+    return Gemma3CausalLMOutputWithPast(
+        loss=loss,
+        logits=logits,
+        past_key_values=outputs.past_key_values,
+        hidden_states=outputs.hidden_states,
+        attentions=outputs.attentions,
+        image_hidden_states=image_features if pixel_values is not None else None,
+    )
+
+
+def patch_gemma3conditionalgeneration_forward():
+    from transformers.models.gemma3.modeling_gemma3 import (
+        Gemma3ForConditionalGeneration,
+    )
+
+    Gemma3ForConditionalGeneration.forward = new_forward

src/axolotl/monkeypatch/lora_kernels.py

@@ -252,12 +252,38 @@ def apply_lora_kernel_patches(
     LOG.setLevel(logging.INFO)
 
     # Choose activation based on model type
-    activation = model.config.hidden_act
+    activation = None
+    text_config = (
+        model.config.get_text_config()
+        if hasattr(model.config, "get_text_config")
+        else model.config
+    )
+    if hasattr(text_config, "hidden_act"):
+        activation = text_config.hidden_act
+    elif hasattr(text_config, "hidden_activation"):
+        activation = text_config.hidden_activation
+
+    # map activation to supported activation
+    if "gelu" in activation:
+        # gemma3 uses gelu_pytorch_tanh
+        activation = "gelu"
+
     if activation not in SUPPORTED_ACTIVATIONS:
         raise NotImplementedError(f"Activation {activation} is not supported")
 
+    layers = []
+    # check for multimodal models first
+    if hasattr(model, "language_model"):
+        layers = model.language_model.model.layers
+    elif hasattr(model, "model"):
+        layers = model.model.model.layers
+    else:
+        raise NotImplementedError(
+            f"Model type {model.config.model_type} is not supported yet. Please create an Issue."
+        )
+
     # Patch each layer
-    for layer in model.model.model.layers:
+    for layer in layers:
         # Add QKV, O fallback implementations to start
         # These will be overwritten later (if some conditions apply)
         layer.self_attn.apply_qkv = types.MethodType(

src/axolotl/monkeypatch/multipack.py

@@ -22,6 +22,7 @@ SUPPORTED_MULTIPACK_MODEL_TYPES = [
     "phi3",
     "gemma",
     "gemma2",
+    "gemma3",
     "gemma3_text",
     "cohere",
     "cohere2",

src/axolotl/train.py

@@ -509,6 +509,7 @@ def train(
     # Save the trained model and cleanup
     save_trained_model(cfg, trainer, model, safe_serialization)
     create_model_card(cfg, trainer)
-    cleanup_distributed()
+    if not cfg.use_ray:
+        cleanup_distributed()
 
     return model, tokenizer, trainer

src/axolotl/utils/collators/batching.py

@@ -112,6 +112,7 @@ class DataCollatorForSeq2Seq:
             self.local_world_size = dist.get_world_size(group=sp_group)
 
     def __call__(self, features, return_tensors=None):
+        has_attn_mask = "attention_mask" in features[0].keys()
         labels = None
         if return_tensors is None:
             return_tensors = self.return_tensors
@@ -164,6 +165,8 @@ class DataCollatorForSeq2Seq:
             pad_to_multiple_of=self.pad_to_multiple_of,
             return_tensors=return_tensors,
         )
+        if not has_attn_mask:
+            del features["attention_mask"]
 
         # prepare decoder_input_ids
         if (

src/axolotl/utils/data/shared.py

@@ -238,7 +238,8 @@ def load_dataset_w_config(
             trust_remote_code=config_dataset.trust_remote_code,
             **load_ds_kwargs,
         )
-    else:
+    elif config_dataset.data_files:
+        fp: str | list[str] | None = None
         if isinstance(config_dataset.data_files, str):
             fp = hf_hub_download(
                 repo_id=config_dataset.path,

src/axolotl/utils/models.py

@@ -535,6 +535,15 @@ class ModelLoader:
         self.auto_model_loader = AutoModelForCausalLM  # pylint: disable=invalid-name
 
     def apply_patches(self) -> None:
+        # patch gemma3 conditional generation forward before loading plugins
+        # as it could be overridden by plugins
+        if self.cfg.model_config_type == "gemma3":
+            from axolotl.monkeypatch.gemma3 import (
+                patch_gemma3conditionalgeneration_forward,
+            )
+
+            patch_gemma3conditionalgeneration_forward()
+
         # load any patches from plugins
         from axolotl.integrations.base import PluginManager
 

src/axolotl/utils/samplers/multipack.py

@@ -8,11 +8,13 @@ from typing import Any, Iterable, List, Union
 
 import numba
 import numpy as np
-from torch.utils.data import BatchSampler, Sampler
+from torch.utils.data import BatchSampler, Sampler, SequentialSampler
 
 from axolotl.utils.distributed import reduce_and_broadcast
 
-LOG = logging.getLogger("axolotl.utils.samplers.multipack")
+LOG = logging.getLogger(__name__)
+
+LOG.setLevel(logging.INFO)
 
 
 @numba.njit
@@ -103,6 +105,55 @@ def allocate(
     return result, s, len(result) * c * n
 
 
+@numba.njit
+def allocate_sequentially(lengths: np.ndarray, rank: int, c: int, n: int):
+    """
+    Sequential allocator that preserves example order
+
+    Parameters:
+    - lengths: The lengths of all examples
+    - rank: The current rank (for distributed training)
+    - c: The capacity of each bin (maximum sequence length)
+    - n: Number of ranks
+
+    Returns:
+    - result: List of batches for the current rank
+    - total_used: Number of actual example tokens
+    - total_slots: Maximum theoretical number of example tokens (number of bins * bin capacity)
+    """
+    result = []
+    total_used = 0
+
+    # First, do sequential packing into bins
+    all_bins = []
+    current_bin = [0 for i in range(0)]  # numba hint
+    remaining_capacity = c
+
+    for idx, size in enumerate(lengths):
+        if size <= remaining_capacity:
+            # Example fits in current bin
+            current_bin.append(idx)
+            remaining_capacity -= size
+            total_used += size
+        else:
+            # Example doesn't fit, start a new bin
+            if current_bin:  # Add non-empty bin to all_bins
+                all_bins.append(current_bin)
+            current_bin = [idx]
+            remaining_capacity = c - size
+            total_used += size
+
+    # Add the last bin if not empty
+    if current_bin:
+        all_bins.append(current_bin)
+
+    # Assign bins to ranks - each rank gets every n-th bin
+    for bin_idx in range(rank, len(all_bins), n):
+        result.append(all_bins[bin_idx])
+
+    return result, total_used, len(all_bins) * c
+
+
 class MultipackBatchSampler(BatchSampler):
     """Batch sampler class for multipack"""
 
@@ -115,6 +166,7 @@ class MultipackBatchSampler(BatchSampler):
         packing_efficiency_estimate: float = 1.0,
         drop_last: bool = False,
         num_count_samples: int = 16,
+        sequential: bool = False,
         **kwargs,
     ):
         super().__init__(sampler, batch_size, drop_last)
@@ -122,6 +174,7 @@ class MultipackBatchSampler(BatchSampler):
         self.batch_max_len = batch_max_len
         self.lengths: np.ndarray = lengths
         self.packing_efficiency_estimate = packing_efficiency_estimate or 1.0
+        self.sequential = sequential
 
         assert isinstance(self.lengths, np.ndarray)
 
@@ -136,6 +189,11 @@ class MultipackBatchSampler(BatchSampler):
         # the minimum packed dataset length across all ranks determined by a gather/broadcast
         self.len_across_ranks = None
 
+        if self.sequential and not isinstance(sampler, SequentialSampler):
+            LOG.warn(
+                "using sequential sample packing with non-sequential sampler, did you want to also enable curriculum_sampling?"
+            )
+
     def set_epoch(self, epoch: int):
         self.epoch = epoch
 
@@ -145,13 +203,21 @@ class MultipackBatchSampler(BatchSampler):
         lengths = self.lengths[indices]
         lengths_cumsum = np.cumsum(lengths)
 
-        batches, total_used, total_slots = allocate(
-            lengths=lengths,
-            lengths_cumsum=lengths_cumsum,
-            rank=0,
-            c=self.batch_max_len,
-            n=1,
-        )
+        if self.sequential:
+            batches, total_used, total_slots = allocate_sequentially(
+                lengths=lengths,
+                rank=0,
+                c=self.batch_max_len,
+                n=1,
+            )
+        else:
+            batches, total_used, total_slots = allocate(
+                lengths=lengths,
+                lengths_cumsum=lengths_cumsum,
+                rank=0,
+                c=self.batch_max_len,
+                n=1,
+            )
 
         batches = [
             [

src/axolotl/utils/schemas/config.py

@@ -46,6 +46,7 @@ from axolotl.utils.schemas.multimodal import MultiModalConfig
 from axolotl.utils.schemas.peft import LoraConfig, ReLoRAConfig
 from axolotl.utils.schemas.training import HyperparametersConfig
 from axolotl.utils.schemas.trl import TRLConfig
+from axolotl.utils.schemas.vllm import VllmConfig
 
 LOG = logging.getLogger(__name__)
 
@@ -86,6 +87,9 @@ class AxolotlInputConfig(
     trl: TRLConfig | None = Field(
         default_factory=lambda: TRLConfig(),  # pylint: disable=unnecessary-lambda
     )
+    vllm: VllmConfig | None = Field(
+        default_factory=lambda: VllmConfig(),  # pylint: disable=unnecessary-lambda
+    )
     reward_model: bool | None = None
     process_reward_model: bool | None = None
     num_labels: int | None = None
@@ -188,6 +192,7 @@ class AxolotlInputConfig(
     sample_packing: bool | None = None
     sample_packing_group_size: int | None = 100_000
     sample_packing_bin_size: int | None = 200
+    sample_packing_sequentially: bool | None = None
     eval_sample_packing: bool | None = None
     pad_to_sequence_len: bool | None = None
     curriculum_sampling: bool | None = None
@@ -1130,6 +1135,17 @@ class AxolotlInputConfig(
 
         return value
 
+    @model_validator(mode="before")
+    @classmethod
+    def check_muon_deepspeed_fsdp(cls, data):
+        if data.get("optimizer") == "muon" and (
+            data.get("deepspeed") or data.get("fsdp") or data.get("fsdp_config")
+        ):
+            raise ValueError(
+                "Muon optimizer is currently incompatible with DeepSpeed and FSDP"
+            )
+        return data
+
 
 class AxolotlConfigWCapabilities(AxolotlInputConfig):
     """wrapper to valdiate gpu capabilities with the configured options"""
@@ -1265,3 +1281,14 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
             if data["beta"] != data["trl"]["beta"]:
                 raise ValueError("beta and trl.beta must match or one must be removed")
         return data
+
+    @model_validator(mode="after")
+    def check_min_torch_version(self):
+        if self.env_capabilities and self.env_capabilities.torch_version:
+            torch_version = self.env_capabilities.torch_version
+            if version.parse(torch_version) < version.parse("2.5.1"):
+                LOG.warning(
+                    f"torch=={torch_version} may not be supported in future versions. Please consider upgrading to torch>=2.5.1."
+                )
+
+        return self

src/axolotl/utils/schemas/trl.py

@@ -20,27 +20,30 @@ class TRLConfig(BaseModel):
     )
 
     # GRPO specific args
-    # Ref: https://github.com/huggingface/trl/blob/e3244d2d096ff1e2e248c931d06d39e165e20623/trl/trainer/grpo_config.py#L22
-    use_vllm: bool | None = Field(
+    # Ref: https://github.com/huggingface/trl/blob/26d86757a7c7e24e397ea44f57ecce6031dfac01/trl/trainer/grpo_config.py#L23
+    use_vllm: bool = Field(
         default=False,
         json_schema_extra={"description": "Whether to use VLLM for RL training"},
     )
-    vllm_device: str | None = Field(
-        default="auto",
-        json_schema_extra={"description": "Device to use for VLLM"},
+    vllm_server_host: str | None = Field(
+        default="0.0.0.0",  # nosec B104
+        json_schema_extra={"description": "Host of the vLLM server to connect to"},
     )
-    vllm_gpu_memory_utilization: float | None = Field(
-        default=0.9,
-        json_schema_extra={"description": "GPU memory utilization for VLLM"},
+    vllm_server_port: int | None = Field(
+        default=8000,
+        json_schema_extra={"description": "Port of the vLLM server to connect to"},
     )
-    vllm_dtype: str | None = Field(
-        default="auto",
-        json_schema_extra={"description": "Data type for VLLM"},
-    )
-    vllm_max_model_len: int | None = Field(
+    vllm_server_timeout: int | None = Field(
         default=None,
         json_schema_extra={
-            "description": "Maximum length of the model context for VLLM"
+            "description": "Total timeout duration in seconds to wait for the vLLM server to be up. If the server is not up "
+            "after the timeout, a `ConnectionError` is raised."
+        },
+    )
+    vllm_guided_decoding_regex: str | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Regex for vLLM guided decoding. If `None` (default), guided decoding is disabled."
         },
     )
 
@@ -85,3 +88,48 @@ class TRLConfig(BaseModel):
             "description": "Sync steps for the reference model. Requires `sync_ref_model=True`."
         },
     )
+    scale_rewards: bool = Field(
+        default=True,
+        json_schema_extra={
+            "description": "Whether to scale the rewards for GRPO by dividing them by their standard deviation."
+        },
+    )
+
+    temperature: float | None = Field(
+        default=None,
+        json_schema_extra={"description": "Sampling temperature for the GRPO policy."},
+    )
+    top_p: float | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Top-p sampling probability for the generation policy."
+        },
+    )
+    top_k: int | None = Field(
+        default=None,
+        json_schema_extra={"description": "Top-k sampling for the generation policy."},
+    )
+    min_p: float | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Minimum probability for the generation policy."
+        },
+    )
+    repetition_penalty: float | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Float that penalizes new tokens based on whether they appear in the prompt and the generated text so far."
+        },
+    )
+    num_iterations: int | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Number of iterations per batch (denoted as μ in the algorithm) for GRPO."
+        },
+    )
+    epsilon: float | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Epsilon value for clipping in the GRPO algorithm."
+        },
+    )

src/axolotl/utils/schemas/vllm.py

@@ -0,0 +1,38 @@
+"""
+Pydantic models for VLLM configuration, used primarily for RL training with TRL + grpo
+"""
+
+from pydantic import BaseModel, Field
+
+
+class VllmConfig(BaseModel):
+    """
+    Configuration for VLLM server
+    """
+
+    device: str | None = Field(
+        default="auto",
+        json_schema_extra={"description": "Device to use for VLLM"},
+    )
+    tensor_parallel_size: int | None = Field(
+        default=None,
+        json_schema_extra={"description": "Tensor parallel size for VLLM"},
+    )
+    gpu_memory_utilization: float | None = Field(
+        default=0.9,
+        json_schema_extra={"description": "GPU memory utilization for VLLM"},
+    )
+    dtype: str | None = Field(
+        default="auto",
+        json_schema_extra={"description": "Data type for VLLM"},
+    )
+    max_model_len: int | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Maximum length of the model context for VLLM"
+        },
+    )
+    enable_prefix_caching: bool | None = Field(
+        default=None,
+        json_schema_extra={"description": "Enable prefix caching for VLLM"},
+    )

src/axolotl/utils/trainer.py

@@ -13,7 +13,7 @@ import torch
 import torch.cuda
 from accelerate.logging import get_logger
 from datasets import IterableDataset, disable_caching, enable_caching
-from torch.utils.data import DataLoader, RandomSampler
+from torch.utils.data import DataLoader, RandomSampler, SequentialSampler
 from transformers.utils import is_torch_bf16_gpu_available
 
 from axolotl.core.trainer_builder import HFCausalTrainerBuilder, HFRLTrainerBuilder
@@ -235,7 +235,7 @@ def drop_long_seq(sample, sequence_len=2048, min_sequence_len=2):
 
 
 def process_datasets_for_packing(cfg, train_dataset, eval_dataset):
-    if cfg.model_config_type == "mamba":
+    if cfg.model_config_type in ["mamba", "gemma3"]:
         LOG.info("dropping attention_mask column")
         train_dataset = train_dataset.remove_columns("attention_mask")
         if eval_dataset:
@@ -456,13 +456,18 @@ def calculate_total_num_steps(cfg, train_dataset, update=True):
             else:
                 sampler_batch_size = cfg.micro_batch_size
                 batch_max_len = cfg.sequence_len
+            if cfg.curriculum_sampling:
+                sampler = SequentialSampler(train_dataset)
+            else:
+                sampler = RandomSampler(train_dataset)
             sampler = MultipackBatchSampler(
-                sampler=RandomSampler(train_dataset),
+                sampler=sampler,
                 lengths=get_dataset_lengths(train_dataset),
                 batch_size=sampler_batch_size,
                 batch_max_len=batch_max_len,
                 group_size=cfg.sample_packing_group_size,
                 bin_size=cfg.sample_packing_bin_size,
+                sequential=cfg.sample_packing_sequentially,
                 drop_last=True,
             )
 

tests/conftest.py

@@ -8,11 +8,13 @@ import shutil
 import sys
 import tempfile
 import time
+from pathlib import Path
 
+import datasets
 import pytest
 import requests
-from datasets import load_dataset
 from huggingface_hub import snapshot_download
+from tokenizers import AddedToken
 from transformers import AutoTokenizer
 
 from tests.hf_offline_utils import disable_hf_offline, enable_hf_offline
@@ -48,6 +50,14 @@ def snapshot_download_w_retry(*args, **kwargs):
     return snapshot_download(*args, **kwargs)
 
 
+@pytest.fixture(scope="session", autouse=True)
+def download_ds_fixture_bundle():
+    ds_dir = snapshot_download_w_retry(
+        "axolotl-ai-internal/axolotl-oss-dataset-fixtures", repo_type="dataset"
+    )
+    return Path(ds_dir)
+
+
 @pytest.fixture(scope="session", autouse=True)
 def download_smollm2_135m_model():
     # download the model
@@ -101,42 +111,50 @@ def download_argilla_distilabel_capybara_dpo_7k_binarized_dataset():
 
 
 @pytest.fixture(scope="session", autouse=True)
-def download_argilla_ultrafeedback_binarized_preferences_cleaned_dataset():
+def download_argilla_distilabel_intel_orca_dpo_dataset():
     # download the dataset
     snapshot_download_w_retry(
-        "argilla/ultrafeedback-binarized-preferences-cleaned", repo_type="dataset"
+        "argilla/distilabel-intel-orca-dpo-pairs", repo_type="dataset"
     )
 
 
-@pytest.fixture(scope="session", autouse=True)
-def download_fozzie_alpaca_dpo_dataset():
-    # download the dataset
-    snapshot_download_w_retry(
-        "fozziethebeat/alpaca_messages_2k_dpo_test", repo_type="dataset"
-    )
-    snapshot_download_w_retry(
-        "fozziethebeat/alpaca_messages_2k_dpo_test",
-        repo_type="dataset",
-        revision="ea82cff",
-    )
+# @pytest.fixture(scope="session", autouse=True)
+# def download_argilla_ultrafeedback_binarized_preferences_cleaned_dataset():
+#     # download the dataset
+#     snapshot_download_w_retry(
+#         "argilla/ultrafeedback-binarized-preferences-cleaned", repo_type="dataset"
+#     )
 
 
-@pytest.fixture(scope="session")
-@disable_hf_offline
-def dataset_fozzie_alpaca_dpo_dataset(
-    download_fozzie_alpaca_dpo_dataset,
-):  # pylint: disable=unused-argument,redefined-outer-name
-    return load_dataset("fozziethebeat/alpaca_messages_2k_dpo_test", split="train")
+# @pytest.fixture(scope="session", autouse=True)
+# def download_fozzie_alpaca_dpo_dataset():
+#     # download the dataset
+#     snapshot_download_w_retry(
+#         "fozziethebeat/alpaca_messages_2k_dpo_test", repo_type="dataset"
+#     )
+#     snapshot_download_w_retry(
+#         "fozziethebeat/alpaca_messages_2k_dpo_test",
+#         repo_type="dataset",
+#         revision="ea82cff",
+#     )
 
 
-@pytest.fixture(scope="session")
-@disable_hf_offline
-def dataset_fozzie_alpaca_dpo_dataset_rev_ea82cff(
-    download_fozzie_alpaca_dpo_dataset,
-):  # pylint: disable=unused-argument,redefined-outer-name
-    return load_dataset(
-        "fozziethebeat/alpaca_messages_2k_dpo_test", split="train", revision="ea82cff"
-    )
+# @pytest.fixture(scope="session")
+# @disable_hf_offline
+# def dataset_fozzie_alpaca_dpo_dataset(
+#     download_fozzie_alpaca_dpo_dataset,
+# ):  # pylint: disable=unused-argument,redefined-outer-name
+#     return load_dataset("fozziethebeat/alpaca_messages_2k_dpo_test", split="train")
+#
+#
+# @pytest.fixture(scope="session")
+# @disable_hf_offline
+# def dataset_fozzie_alpaca_dpo_dataset_rev_ea82cff(
+#     download_fozzie_alpaca_dpo_dataset,
+# ):  # pylint: disable=unused-argument,redefined-outer-name
+#     return load_dataset(
+#         "fozziethebeat/alpaca_messages_2k_dpo_test", split="train", revision="ea82cff"
+#     )
 
 
 @pytest.fixture(scope="session", autouse=True)
@@ -263,7 +281,7 @@ def download_mlx_mistral_7b_model_fixture():
     )
 
 
-@pytest.fixture(scope="session", autouse=True)
+@pytest.fixture
 def download_llama2_model_fixture():
     # download the tokenizer only
     snapshot_download_w_retry(
@@ -273,7 +291,7 @@ def download_llama2_model_fixture():
     )
 
 
-@pytest.fixture(scope="session", autouse=True)
+@pytest.fixture
 @enable_hf_offline
 def tokenizer_huggyllama(
     download_huggyllama_model_fixture,
@@ -284,6 +302,57 @@ def tokenizer_huggyllama(
     return tokenizer
 
 
+@pytest.fixture
+@enable_hf_offline
+def tokenizer_huggyllama_w_special_tokens(
+    tokenizer_huggyllama,
+):  # pylint: disable=redefined-outer-name
+    tokenizer_huggyllama.add_special_tokens(
+        {
+            "bos_token": "<s>",
+            "eos_token": "</s>",
+            "unk_token": "<unk>",
+        }
+    )
+
+    return tokenizer_huggyllama
+
+
+@pytest.fixture
+@enable_hf_offline
+def tokenizer_llama2_7b(
+    download_llama2_model_fixture,
+):  # pylint: disable=unused-argument,redefined-outer-name
+    tokenizer = AutoTokenizer.from_pretrained("NousResearch/Llama-2-7b-hf")
+
+    return tokenizer
+
+
+@pytest.fixture
+@enable_hf_offline
+def tokenizer_mistral_7b_instruct(
+    download_mlx_mistral_7b_model_fixture,
+):  # pylint: disable=unused-argument,redefined-outer-name
+    return AutoTokenizer.from_pretrained("casperhansen/mistral-7b-instruct-v0.1-awq")
+
+
+@pytest.fixture
+def tokenizer_mistral_7b_instruct_chatml(tokenizer_mistral_7b_instruct):
+    tokenizer_mistral_7b_instruct.add_special_tokens(
+        {
+            "eos_token": AddedToken(
+                "<|im_end|>", rstrip=False, lstrip=False, normalized=False
+            )
+        }
+    )
+    tokenizer_mistral_7b_instruct.add_tokens(
+        [
+            AddedToken("<|im_start|>", rstrip=False, lstrip=False, normalized=False),
+        ]
+    )
+    return tokenizer_mistral_7b_instruct
+
+
 @pytest.fixture
 def temp_dir():
     # Create a temporary directory
@@ -349,6 +418,60 @@ def cleanup_monkeypatches():
                 globals().pop(module_global, None)
 
 
+@pytest.fixture
+def dataset_winglian_tiny_shakespeare(
+    download_ds_fixture_bundle: Path,
+):  # pylint: disable=redefined-outer-name
+    ds_path = download_ds_fixture_bundle / "winglian__tiny-shakespeare"
+    return datasets.load_from_disk(ds_path)
+
+
+@pytest.fixture
+def dataset_tatsu_lab_alpaca(
+    download_ds_fixture_bundle: Path,
+):  # pylint: disable=redefined-outer-name
+    ds_path = download_ds_fixture_bundle / "tatsu-lab__alpaca"
+    return datasets.load_from_disk(ds_path)["train"]
+
+
+@pytest.fixture
+def dataset_mhenrichsen_alpaca_2k_test(
+    download_ds_fixture_bundle: Path,
+):  # pylint: disable=redefined-outer-name
+    ds_path = download_ds_fixture_bundle / "mhenrichsen__alpaca_2k_test"
+    return datasets.load_from_disk(ds_path)["train"]
+
+
+@pytest.fixture
+def dataset_argilla_ultrafeedback_binarized_preferences_cleaned(
+    download_ds_fixture_bundle: Path,
+):  # pylint: disable=redefined-outer-name
+    ds_path = (
+        download_ds_fixture_bundle
+        / "argilla__ultrafeedback-binarized-preferences-cleaned"
+    )
+    return datasets.load_from_disk(ds_path)["train"]
+
+
+@pytest.fixture
+def dataset_fozziethebeat_alpaca_messages_2k_dpo_test(
+    download_ds_fixture_bundle: Path,
+):  # pylint: disable=redefined-outer-name
+    ds_path = download_ds_fixture_bundle / "fozziethebeat__alpaca_messages_2k_dpo_test"
+    return datasets.load_from_disk(ds_path)["train"]
+
+
+@pytest.fixture
+def dataset_fozziethebeat_alpaca_messages_2k_dpo_test_rev_ea82cff(
+    download_ds_fixture_bundle: Path,
+):  # pylint: disable=redefined-outer-name
+    ds_path = (
+        download_ds_fixture_bundle
+        / "fozziethebeat__alpaca_messages_2k_dpo_test__rev_ea82cff"
+    )
+    return datasets.load_from_disk(ds_path)["train"]
+
+
 # # pylint: disable=redefined-outer-name,unused-argument
 # def test_load_fixtures(
 #     download_smollm2_135m_model,

tests/e2e/multigpu/solo/test_grpo.py

@@ -0,0 +1,294 @@
+"""
+GRPO test suite
+"""
+
+import os
+import random
+import subprocess  # nosec B404
+import sys
+import time
+from pathlib import Path
+
+import pytest
+import requests
+import yaml
+from accelerate.test_utils import execute_subprocess_async
+from transformers.testing_utils import get_torch_dist_unique_port
+
+from axolotl.utils.dict import DictDefault
+
+from tests.e2e.utils import require_vllm
+
+
+def start_vllm(
+    model: str, env: dict | None = None, wait: int | None = None, quiet=False, **kwargs
+) -> int:
+    """
+    helper function to start the VLLM server in the background, mostly for testing purposes
+    """
+    cmd = [sys.executable, "-m", "trl.scripts.vllm_serve", "--model", model]
+
+    if tensor_parallel_size := kwargs.get("tensor_parallel_size"):
+        cmd.extend(["--tensor-parallel-size", str(tensor_parallel_size)])
+    if host := kwargs.get("host"):
+        cmd.extend(["--host", host])
+    if port := kwargs.get("port"):
+        cmd.extend(["--port", str(port)])
+    if gpu_memory_utilization := kwargs.get("gpu_memory_utilization"):
+        cmd.extend(["--gpu-memory-utilization", str(gpu_memory_utilization)])
+    if dtype := kwargs.get("dtype"):
+        cmd.extend(["--dtype", dtype])
+    if max_model_len := kwargs.get("max_model_len"):
+        cmd.extend(["--max-model-len", str(max_model_len)])
+    if kwargs.get("enable_prefix_caching"):
+        cmd.extend(["--enable-prefix-caching", "True"])
+
+    # print out the command to be executed
+    print(" ".join(cmd))
+
+    # start `trl vllm-serve` command in the background and capture the process id
+    process = subprocess.Popen(  # pylint: disable=consider-using-with
+        cmd,
+        env=env,
+        stdout=subprocess.DEVNULL if quiet else subprocess.PIPE,
+        stderr=subprocess.DEVNULL if quiet else subprocess.PIPE,
+    )  # nosec B603
+
+    # print out the process id so the user can easily kill it later
+    print(f"VLLM server process started (PID: {process.pid})")
+
+    # wait until the http server is ready, even if it 404s, but timeout after 60 seconds
+    started = False
+    if wait and host and port:
+        for _ in range(int(wait)):
+            try:
+                response = requests.get(f"http://{host}:{port}", timeout=1)
+                if int(response.status_code) in [200, 404]:
+                    started = True
+                    break
+            except requests.exceptions.RequestException:
+                pass
+
+            # also check if the process.pid is still running
+            if not process.poll() is None:
+                break
+
+            time.sleep(1)
+
+    if wait and not started:
+        print(
+            f"VLLM server process did not start within {wait} seconds. Please check your server logs."
+        )
+        process.kill()
+        raise RuntimeError(f"VLLM server process did not start within {wait} seconds.")
+
+    # return the process id
+    return process.pid
+
+
+class TestGRPO:
+    """
+    Test case for GRPO training using multilpe GPUs
+    """
+
+    def _utils_write_yaml_and_rewards(self, cfg, temp_dir, suffix=""):
+        # 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))
+        with open(f"rewards_{suffix}.py", "w", encoding="utf-8") as fout:
+            fout.write(
+                """import random
+def rand_reward_func(completions, **kwargs) -> list[float]:
+    return [random.uniform(0, 1) for _ in completions]
+
+def oai_gsm8k_transform(cfg, *args, **kwargs):
+    def transform_fn(example, tokenizer=None):
+        label = example["answer"].split("####")[-1].strip().replace(",", "")
+        return {
+            "prompt": [{"role": "user", "content": example["question"]},],
+            "answer": label,
+        }
+    return transform_fn, {"remove_columns": ["question"]}
+"""
+            )
+
+    @pytest.mark.parametrize(
+        "num_gpus",
+        [1, 2],
+    )
+    @require_vllm
+    def test_llama_dora(self, temp_dir, num_gpus):
+        rnd_reward_suffix = str(random.randint(1000, 9999))
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "chat_template": "llama3",
+                "rl": "grpo",
+                "trl": {
+                    "beta": 0.001,
+                    "max_completion_length": 256,
+                    "use_vllm": True,
+                    "num_generations": 4,
+                    "reward_funcs": [f"rewards_{rnd_reward_suffix}.rand_reward_func"],
+                },
+                "vllm": {
+                    "max_model_len": 800,
+                    "enable_prefix_caching": True,
+                },
+                "datasets": [
+                    {
+                        "path": "openai/gsm8k",
+                        "name": "main",
+                        "type": f"rewards_{rnd_reward_suffix}.oai_gsm8k_transform",
+                    },
+                ],
+                "adapter": "lora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "peft_use_dora": True,
+                "flash_attention": True,
+                "sequence_len": 1024,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "max_steps": 3,
+                "num_epochs": 1,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 2,
+                "warmup_steps": 10,
+                "val_set_size": 0.0,
+                "output_dir": temp_dir,
+                "learning_rate": 0.0001,
+                "optimizer": "adamw_torch_fused",
+                "lr_scheduler": "cosine",
+                "save_safetensors": True,
+                "bf16": "auto",
+                "use_tensorboard": True,
+            }
+        )
+
+        self._utils_write_yaml_and_rewards(cfg, temp_dir, suffix=rnd_reward_suffix)
+
+        current_env = os.environ.copy()
+        env = {
+            "NCCL_P2P_LEVEL": "LOC",
+            **current_env,
+            "CUDA_VISIBLE_DEVICES": "1",
+        }
+        vllm_process_id = start_vllm(
+            cfg.base_model,
+            env=env,
+            quiet=True,
+            wait=120,
+            gpu_memory_utilization=0.15,
+            max_model_len=cfg.vllm.max_model_len,
+            enable_prefix_caching=cfg.vllm.enable_prefix_caching,
+            host="0.0.0.0",
+            port=8000,
+        )
+
+        try:
+            execute_subprocess_async(
+                [
+                    "axolotl",
+                    "train",
+                    str(Path(temp_dir) / "config.yaml"),
+                    "--num-processes",
+                    str(num_gpus),
+                    "--main-process-port",
+                    f"{get_torch_dist_unique_port()}",
+                ],
+                env={"NCCL_P2P_LEVEL": "LOC", "NCCL_DEBUG": "INFO", **current_env},
+            )
+        finally:
+            os.kill(vllm_process_id, 9)
+
+    @pytest.mark.parametrize(
+        "num_gpus",
+        [1, 2],
+    )
+    @require_vllm
+    def test_llama_fft(self, temp_dir, num_gpus):
+        rnd_reward_suffix = str(random.randint(1000, 9999))
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "chat_template": "llama3",
+                "rl": "grpo",
+                "trl": {
+                    "beta": 0.001,
+                    "max_completion_length": 256,
+                    "use_vllm": True,
+                    "num_generations": 4,
+                    "reward_funcs": [f"rewards_{rnd_reward_suffix}.rand_reward_func"],
+                },
+                "vllm": {
+                    "max_model_len": 800,
+                    "enable_prefix_caching": True,
+                },
+                "datasets": [
+                    {
+                        "path": "openai/gsm8k",
+                        "name": "main",
+                        "type": f"rewards_{rnd_reward_suffix}.oai_gsm8k_transform",
+                    },
+                ],
+                "flash_attention": True,
+                "sequence_len": 1024,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "max_steps": 3,
+                "num_epochs": 1,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 2,
+                "warmup_steps": 10,
+                "val_set_size": 0.0,
+                "output_dir": temp_dir,
+                "learning_rate": 0.0001,
+                "optimizer": "adamw_torch_fused",
+                "lr_scheduler": "cosine",
+                "save_safetensors": True,
+                "bf16": "auto",
+                "use_tensorboard": True,
+            }
+        )
+
+        self._utils_write_yaml_and_rewards(cfg, temp_dir, suffix=rnd_reward_suffix)
+
+        current_env = os.environ.copy()
+        env = {
+            "NCCL_P2P_LEVEL": "LOC",  # nccl can be brittle, assume P2P isn't reliable
+            **current_env,
+            "CUDA_VISIBLE_DEVICES": "1",
+        }
+        vllm_process_id = start_vllm(
+            cfg.base_model,
+            env=env,
+            quiet=True,
+            wait=120,
+            gpu_memory_utilization=0.15,
+            max_model_len=cfg.vllm.max_model_len,
+            enable_prefix_caching=cfg.vllm.enable_prefix_caching,
+            host="0.0.0.0",
+            port=8000,
+        )
+
+        try:
+            execute_subprocess_async(
+                [
+                    "axolotl",
+                    "train",
+                    str(Path(temp_dir) / "config.yaml"),
+                    "--num-processes",
+                    str(num_gpus),
+                    "--main-process-port",
+                    f"{get_torch_dist_unique_port()}",
+                ],
+                env={"NCCL_P2P_LEVEL": "LOC", "NCCL_DEBUG": "INFO", **current_env},
+            )
+        finally:
+            os.kill(vllm_process_id, 9)

tests/e2e/multigpu/test_eval.py

@@ -52,9 +52,9 @@ class TestMultiGPUEval:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 5,
+                "max_steps": 2,
                 "micro_batch_size": 2,
-                "gradient_accumulation_steps": 4,
+                "gradient_accumulation_steps": 2,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_8bit",
@@ -121,9 +121,9 @@ class TestMultiGPUEval:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 5,
+                "max_steps": 2,
                 "micro_batch_size": 2,
-                "gradient_accumulation_steps": 4,
+                "gradient_accumulation_steps": 2,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_8bit",

tests/e2e/multigpu/test_gemma3.py

@@ -0,0 +1,100 @@
+"""
+E2E tests for multigpu lora tinyllama
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import pytest
+import yaml
+from accelerate.test_utils import execute_subprocess_async
+from huggingface_hub import snapshot_download
+from transformers.testing_utils import get_torch_dist_unique_port
+
+from axolotl.utils.dict import DictDefault
+
+from tests.e2e.utils import check_tensorboard
+
+LOG = logging.getLogger("axolotl.tests.e2e.multigpu")
+os.environ["WANDB_DISABLED"] = "true"
+
+AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_model():
+    # download the model
+    snapshot_download("axolotl-mirrors/gemma-3-4b-pt", repo_type="model")
+
+
+class TestMultiGPUGemma3:
+    """
+    Test case for Gemma3 models using LoRA
+    """
+
+    def test_lora_ddp_packed(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "axolotl-mirrors/gemma-3-4b-pt",
+                "sequence_len": 2048,
+                "ddp_find_unused_parameters": True,
+                "sample_packing": True,
+                "eval_sample_packing": False,
+                "pad_to_sequence_len": True,
+                "adapter": "lora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.0,
+                "chat_template": "gemma3",
+                "datasets": [
+                    {
+                        "path": "mlabonne/FineTome-100k",
+                        "type": "chat_template",
+                        "split": "train[:10%]",
+                        "field_messages": "conversations",
+                        "message_field_role": "from",
+                        "message_field_content": "value",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 2,
+                "micro_batch_size": 4,
+                "gradient_checkpointing": True,
+                "gradient_checkpointing_kwargs": {
+                    "use_reentrant": False,
+                },
+                "gradient_accumulation_steps": 2,
+                "output_dir": temp_dir,
+                "learning_rate": 0.0001,
+                "optimizer": "adamw_8bit",
+                "lr_scheduler": "cosine",
+                "flash_attention": True,
+                "use_tensorboard": True,
+                "bf16": True,
+            }
+        )
+
+        # 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(
+            [
+                "axolotl",
+                "train",
+                str(Path(temp_dir) / "config.yaml"),
+                "--num-processes",
+                "2",
+                "--main-process-port",
+                f"{get_torch_dist_unique_port()}",
+            ]
+        )
+
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 1.8, "Train Loss is too high"
+        )

tests/e2e/multigpu/test_grpo.py

@@ -1,175 +0,0 @@
-"""
-GRPO test suite
-"""
-
-import random
-from pathlib import Path
-
-import pytest
-import yaml
-from accelerate.test_utils import execute_subprocess_async
-from transformers.testing_utils import get_torch_dist_unique_port
-
-from axolotl.utils.dict import DictDefault
-
-from tests.e2e.utils import require_vllm
-
-
-class TestGRPO:
-    """
-    Test case for GRPO training using multilpe GPUs
-    """
-
-    def _utils_write_yaml_and_rewards(self, cfg, temp_dir, suffix=""):
-        # 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))
-        with open(f"rewards_{suffix}.py", "w", encoding="utf-8") as fout:
-            fout.write(
-                """import random
-def rand_reward_func(completions, **kwargs) -> list[float]:
-    return [random.uniform(0, 1) for _ in completions]
-
-def oai_gsm8k_transform(cfg, *args, **kwargs):
-    def transform_fn(example, tokenizer=None):
-        label = example["answer"].split("####")[-1].strip().replace(",", "")
-        return {
-            "prompt": [{"role": "user", "content": example["question"]},],
-            "answer": label,
-        }
-    return transform_fn, {"remove_columns": ["question"]}
-"""
-            )
-
-    @pytest.mark.parametrize(
-        "num_gpus",
-        [1, 2],
-    )
-    @require_vllm
-    def test_llama_dora(self, temp_dir, num_gpus):
-        rnd_reward_suffix = str(random.randint(1000, 9999))
-        cfg = DictDefault(
-            {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
-                "chat_template": "llama3",
-                "rl": "grpo",
-                "trl": {
-                    "beta": 0.001,
-                    "max_completion_length": 256,
-                    "use_vllm": True,
-                    "vllm_device": "auto" if num_gpus == 1 else "cuda:1",
-                    "vllm_gpu_memory_utilization": 0.15,
-                    "num_generations": 4,
-                    "reward_funcs": [f"rewards_{rnd_reward_suffix}.rand_reward_func"],
-                },
-                "datasets": [
-                    {
-                        "path": "openai/gsm8k",
-                        "name": "main",
-                        "type": f"rewards_{rnd_reward_suffix}.oai_gsm8k_transform",
-                    },
-                ],
-                "adapter": "lora",
-                "lora_r": 8,
-                "lora_alpha": 16,
-                "lora_dropout": 0.05,
-                "lora_target_linear": True,
-                "peft_use_dora": True,
-                "flash_attention": True,
-                "sequence_len": 1024,
-                "special_tokens": {
-                    "pad_token": "<|endoftext|>",
-                },
-                "max_steps": 5,
-                "num_epochs": 1,
-                "micro_batch_size": 4,
-                "gradient_accumulation_steps": 2,
-                "warmup_steps": 10,
-                "val_set_size": 0.0,
-                "output_dir": temp_dir,
-                "learning_rate": 0.0001,
-                "optimizer": "adamw_torch_fused",
-                "lr_scheduler": "cosine",
-                "save_safetensors": True,
-                "bf16": "auto",
-                "use_tensorboard": True,
-            }
-        )
-
-        self._utils_write_yaml_and_rewards(cfg, temp_dir, suffix=rnd_reward_suffix)
-
-        execute_subprocess_async(
-            [
-                "axolotl",
-                "train",
-                str(Path(temp_dir) / "config.yaml"),
-                "--num-processes",
-                str(num_gpus),
-                "--main-process-port",
-                f"{get_torch_dist_unique_port()}",
-            ]
-        )
-
-    @pytest.mark.parametrize(
-        "num_gpus",
-        [1, 2],
-    )
-    @require_vllm
-    def test_llama_fft(self, temp_dir, num_gpus):
-        rnd_reward_suffix = str(random.randint(1000, 9999))
-        cfg = DictDefault(
-            {
-                "base_model": "HuggingFaceTB/SmolLM2-135M",
-                "chat_template": "llama3",
-                "rl": "grpo",
-                "trl": {
-                    "beta": 0.001,
-                    "max_completion_length": 256,
-                    "use_vllm": True,
-                    "vllm_device": "auto" if num_gpus == 1 else "cuda:1",
-                    "vllm_gpu_memory_utilization": 0.15,
-                    "num_generations": 4,
-                    "reward_funcs": [f"rewards_{rnd_reward_suffix}.rand_reward_func"],
-                },
-                "datasets": [
-                    {
-                        "path": "openai/gsm8k",
-                        "name": "main",
-                        "type": f"rewards_{rnd_reward_suffix}.oai_gsm8k_transform",
-                    },
-                ],
-                "flash_attention": True,
-                "sequence_len": 1024,
-                "special_tokens": {
-                    "pad_token": "<|endoftext|>",
-                },
-                "max_steps": 5,
-                "num_epochs": 1,
-                "micro_batch_size": 4,
-                "gradient_accumulation_steps": 2,
-                "warmup_steps": 10,
-                "val_set_size": 0.0,
-                "output_dir": temp_dir,
-                "learning_rate": 0.0001,
-                "optimizer": "adamw_torch_fused",
-                "lr_scheduler": "cosine",
-                "save_safetensors": True,
-                "bf16": "auto",
-                "use_tensorboard": True,
-            }
-        )
-
-        self._utils_write_yaml_and_rewards(cfg, temp_dir, suffix=rnd_reward_suffix)
-
-        execute_subprocess_async(
-            [
-                "axolotl",
-                "train",
-                str(Path(temp_dir) / "config.yaml"),
-                "--num-processes",
-                str(num_gpus),
-                "--main-process-port",
-                f"{get_torch_dist_unique_port()}",
-            ]
-        )

tests/e2e/multigpu/test_llama.py

@@ -58,6 +58,7 @@ class TestMultiGPULlama:
                 "max_steps": 2,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
+                "gradient_checkpointing": True,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_8bit",
@@ -121,6 +122,7 @@ class TestMultiGPULlama:
                 "max_steps": 2,
                 "micro_batch_size": 1,
                 "gradient_accumulation_steps": gradient_accumulation_steps,
+                "gradient_checkpointing": True,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_8bit",
@@ -193,6 +195,7 @@ class TestMultiGPULlama:
                 "max_steps": 2,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 4,
+                "gradient_checkpointing": True,
                 "output_dir": temp_dir,
                 "warmup_steps": 0,
                 "learning_rate": 0.00001,
@@ -270,6 +273,7 @@ class TestMultiGPULlama:
                 "max_steps": 2,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 4,
+                "gradient_checkpointing": True,
                 "output_dir": temp_dir,
                 "warmup_steps": 0,
                 "learning_rate": 0.00001,
@@ -330,6 +334,7 @@ class TestMultiGPULlama:
                 "max_steps": 2,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": gradient_accumulation_steps,
+                "gradient_checkpointing": True,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
@@ -399,7 +404,8 @@ class TestMultiGPULlama:
                 "num_epochs": 1,
                 "max_steps": 2,
                 "micro_batch_size": 4,
-                "gradient_accumulation_steps": 4,
+                "gradient_accumulation_steps": 2,
+                "gradient_checkpointing": True,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
@@ -478,7 +484,8 @@ class TestMultiGPULlama:
                 "num_epochs": 1,
                 "max_steps": 2,
                 "micro_batch_size": 4,
-                "gradient_accumulation_steps": 4,
+                "gradient_accumulation_steps": 2,
+                "gradient_checkpointing": True,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",
@@ -778,9 +785,10 @@ class TestMultiGPULlama:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 5,
+                "max_steps": 2,
                 "micro_batch_size": 1,
                 "gradient_accumulation_steps": 1,
+                "gradient_checkpointing": True,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_torch_fused",

tests/e2e/multigpu/test_qwen2.py

@@ -46,7 +46,7 @@ class TestMultiGPUQwen2:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 5,
+                "max_steps": 2,
                 "warmup_steps": 20,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 2,

tests/e2e/multigpu/test_ray.py

@@ -50,7 +50,7 @@ class TestMultiGPURay:
                 "num_epochs": 1,
                 "max_steps": 2,
                 "micro_batch_size": 4,
-                "gradient_accumulation_steps": 4,
+                "gradient_accumulation_steps": 2,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_8bit",

tests/test_datasets.py

@@ -324,7 +324,7 @@ class TestDatasetPreparation:
 
     @enable_hf_offline
     def test_load_hub_with_revision_with_dpo(
-        self, dataset_fozzie_alpaca_dpo_dataset_rev_ea82cff
+        self, dataset_fozziethebeat_alpaca_messages_2k_dpo_test_rev_ea82cff
     ):
         """Verify that processing dpo data from the hub works with a specific revision"""
 
@@ -339,12 +339,10 @@ class TestDatasetPreparation:
         )
 
         # pylint: disable=duplicate-code
-        with patch(
-            "axolotl.utils.data.shared.load_dataset_w_config"
-        ) as mock_load_dataset:
+        with patch("axolotl.utils.data.rl.load_dataset_w_config") as mock_load_dataset:
             # Set up the mock to return different values on successive calls
             mock_load_dataset.return_value = (
-                dataset_fozzie_alpaca_dpo_dataset_rev_ea82cff
+                dataset_fozziethebeat_alpaca_messages_2k_dpo_test_rev_ea82cff
             )
 
             train_dataset, _ = load_prepare_preference_datasets(cfg)
@@ -354,7 +352,9 @@ class TestDatasetPreparation:
 
     @enable_hf_offline
     @pytest.mark.skip("datasets bug with local datasets when offline")
-    def test_load_local_hub_with_revision(self, tokenizer):
+    def test_load_local_hub_with_revision(
+        self, dataset_fozziethebeat_alpaca_messages_2k_dpo_test_rev_ea82cff, tokenizer
+    ):
         """Verify that a local copy of a hub dataset can be loaded with a specific revision"""
         with tempfile.TemporaryDirectory() as tmp_dir:
             tmp_ds_path = Path(tmp_dir) / "mhenrichsen/alpaca_2k_test"
@@ -386,13 +386,23 @@ class TestDatasetPreparation:
                 }
             )
 
-            dataset, _ = load_tokenized_prepared_datasets(tokenizer, cfg, prepared_path)
+            with patch(
+                "axolotl.utils.data.shared.load_dataset_w_config"
+            ) as mock_load_dataset:
+                # Set up the mock to return different values on successive calls
+                mock_load_dataset.return_value = (
+                    dataset_fozziethebeat_alpaca_messages_2k_dpo_test_rev_ea82cff
+                )
 
-            assert len(dataset) == 2000
-            assert "input_ids" in dataset.features
-            assert "attention_mask" in dataset.features
-            assert "labels" in dataset.features
-            shutil.rmtree(tmp_ds_path)
+                dataset, _ = load_tokenized_prepared_datasets(
+                    tokenizer, cfg, prepared_path
+                )
+
+                assert len(dataset) == 2000
+                assert "input_ids" in dataset.features
+                assert "attention_mask" in dataset.features
+                assert "labels" in dataset.features
+                shutil.rmtree(tmp_ds_path)
 
     @enable_hf_offline
     def test_loading_local_dataset_folder(self, tokenizer):

tests/test_exact_deduplication.py

@@ -238,21 +238,22 @@ class TestDeduplicateRLDataset:
 
     @enable_hf_offline
     def test_load_with_deduplication(
-        self, cfg, dataset_fozzie_alpaca_dpo_dataset_rev_ea82cff, tokenizer_huggyllama
+        self,
+        cfg,
+        dataset_fozziethebeat_alpaca_messages_2k_dpo_test_rev_ea82cff,
+        tokenizer_huggyllama,
     ):
         """Verify that loading with deduplication removes duplicates."""
 
         # pylint: disable=duplicate-code
         with (
-            patch(
-                "axolotl.utils.data.shared.load_dataset_w_config"
-            ) as mock_load_dataset,
+            patch("axolotl.utils.data.rl.load_dataset_w_config") as mock_load_dataset,
             patch("axolotl.utils.models.load_tokenizer") as mock_load_tokenizer,
         ):
             # Set up the mock to return different values on successive calls
             mock_load_dataset.side_effect = [
-                dataset_fozzie_alpaca_dpo_dataset_rev_ea82cff,
-                dataset_fozzie_alpaca_dpo_dataset_rev_ea82cff,
+                dataset_fozziethebeat_alpaca_messages_2k_dpo_test_rev_ea82cff,
+                dataset_fozziethebeat_alpaca_messages_2k_dpo_test_rev_ea82cff,
             ]
             mock_load_tokenizer.return_value = tokenizer_huggyllama
 
@@ -263,19 +264,20 @@ class TestDeduplicateRLDataset:
 
     @enable_hf_offline
     def test_load_without_deduplication(
-        self, cfg, dataset_fozzie_alpaca_dpo_dataset_rev_ea82cff, tokenizer_huggyllama
+        self,
+        cfg,
+        dataset_fozziethebeat_alpaca_messages_2k_dpo_test_rev_ea82cff,
+        tokenizer_huggyllama,
     ):
         # pylint: disable=duplicate-code
         with (
-            patch(
-                "axolotl.utils.data.shared.load_dataset_w_config"
-            ) as mock_load_dataset,
+            patch("axolotl.utils.data.rl.load_dataset_w_config") as mock_load_dataset,
             patch("axolotl.utils.models.load_tokenizer") as mock_load_tokenizer,
         ):
             # Set up the mock to return different values on successive calls
             mock_load_dataset.side_effect = [
-                dataset_fozzie_alpaca_dpo_dataset_rev_ea82cff,
-                dataset_fozzie_alpaca_dpo_dataset_rev_ea82cff,
+                dataset_fozziethebeat_alpaca_messages_2k_dpo_test_rev_ea82cff,
+                dataset_fozziethebeat_alpaca_messages_2k_dpo_test_rev_ea82cff,
             ]
             mock_load_tokenizer.return_value = tokenizer_huggyllama
 

tests/test_packed_batch_sampler.py

@@ -1,7 +1,7 @@
 """Module for testing streaming dataset sequence packing"""
 
 import pytest
-from datasets import concatenate_datasets, load_dataset
+from datasets import concatenate_datasets
 from torch.utils.data import DataLoader, RandomSampler
 from transformers import AutoTokenizer
 
@@ -27,7 +27,6 @@ class TestBatchedSamplerPacking:
     Test class for packing streaming dataset sequences
     """
 
-    @pytest.mark.skip(reason="TODO: fix hf offline mode for CI rate limits")
     @pytest.mark.parametrize(
         "batch_size, num_workers",
         [
@@ -38,14 +37,20 @@ class TestBatchedSamplerPacking:
         ],
     )
     @pytest.mark.parametrize("max_seq_length", [4096, 512])
+    @pytest.mark.parametrize("sequential", [True, False])
     @enable_hf_offline
-    def test_packing(self, batch_size, num_workers, tokenizer, max_seq_length):
+    def test_packing(
+        self,
+        dataset_winglian_tiny_shakespeare,
+        batch_size,
+        num_workers,
+        tokenizer,
+        max_seq_length,
+        sequential,
+    ):
         import axolotl.monkeypatch.data.batch_dataset_fetcher  # pylint: disable=unused-import  # noqa: F401
 
-        dataset = load_dataset(
-            "winglian/tiny-shakespeare",
-            split="train",
-        )
+        dataset = dataset_winglian_tiny_shakespeare["train"]
 
         cfg = DictDefault(
             {
@@ -55,7 +60,7 @@ class TestBatchedSamplerPacking:
         )
         ds_cfg = DictDefault(
             {
-                "field": "Text",
+                "field": "text",
             }
         )
         completion_strategy = load(tokenizer, cfg, ds_cfg)
@@ -75,6 +80,7 @@ class TestBatchedSamplerPacking:
             batch_max_len=max_seq_length,
             group_size=100000,
             bin_size=200,
+            sequential=sequential,
         )
 
         loader = DataLoader(

tests/test_prompt_tokenizers.py

@@ -2,13 +2,8 @@
 
 import json
 import logging
-import unittest
 from pathlib import Path
 
-import pytest
-from datasets import load_dataset
-from transformers import AddedToken, AutoTokenizer, LlamaTokenizer
-
 from axolotl.prompt_strategies.alpaca_chat import NoSystemPrompter
 from axolotl.prompt_strategies.alpaca_w_system import (
     InstructionWSystemPromptTokenizingStrategy,
@@ -61,24 +56,13 @@ test_data = {
 }
 
 
-class TestPromptTokenizationStrategies(unittest.TestCase):
+class TestPromptTokenizationStrategies:
     """
     Test class for prompt tokenization strategies.
     """
 
     @enable_hf_offline
-    def setUp(self) -> None:
-        # pylint: disable=duplicate-code
-        self.tokenizer = AutoTokenizer.from_pretrained("huggyllama/llama-7b")
-        self.tokenizer.add_special_tokens(
-            {
-                "bos_token": "<s>",
-                "eos_token": "</s>",
-                "unk_token": "<unk>",
-            }
-        )
-
-    def test_no_sys_prompt(self):
+    def test_no_sys_prompt(self, tokenizer_huggyllama_w_special_tokens):
         """
         tests the interface between the user and assistant parts
         """
@@ -86,7 +70,7 @@ class TestPromptTokenizationStrategies(unittest.TestCase):
         # pylint: disable=duplicate-code
         strat = AlpacaPromptTokenizingStrategy(
             prompter,
-            self.tokenizer,
+            tokenizer_huggyllama_w_special_tokens,
             False,
             2048,
         )
@@ -99,7 +83,8 @@ class TestPromptTokenizationStrategies(unittest.TestCase):
         assert example["labels"][world_idx] == 3186
         assert example["labels"][world_idx - 1] == -100
 
-    def test_alpaca(self):
+    @enable_hf_offline
+    def test_alpaca(self, tokenizer_huggyllama_w_special_tokens):
         """
         tests the interface between the user and assistant parts
         """
@@ -107,7 +92,7 @@ class TestPromptTokenizationStrategies(unittest.TestCase):
         prompter = AlpacaPrompter()
         strat = AlpacaPromptTokenizingStrategy(
             prompter,
-            self.tokenizer,
+            tokenizer_huggyllama_w_special_tokens,
             False,
             2048,
         )
@@ -118,28 +103,17 @@ class TestPromptTokenizationStrategies(unittest.TestCase):
         assert example["labels"][world_idx - 1] == -100
 
 
-class InstructionWSystemPromptTokenizingStrategyTest(unittest.TestCase):
+class TestInstructionWSystemPromptTokenizingStrategy:
     """
     Test class for prompt tokenization strategies with sys prompt from the dataset
     """
 
     @enable_hf_offline
-    def setUp(self) -> None:
-        # pylint: disable=duplicate-code
-        self.tokenizer = AutoTokenizer.from_pretrained("huggyllama/llama-7b")
-        self.tokenizer.add_special_tokens(
-            {
-                "bos_token": "<s>",
-                "eos_token": "</s>",
-                "unk_token": "<unk>",
-            }
-        )
-
-    def test_system_alpaca(self):
+    def test_system_alpaca(self, tokenizer_huggyllama_w_special_tokens):
         prompter = SystemDataPrompter(PromptStyle.CHAT.value)
         strat = InstructionWSystemPromptTokenizingStrategy(
             prompter,
-            self.tokenizer,
+            tokenizer_huggyllama_w_special_tokens,
             False,
             2048,
         )
@@ -160,18 +134,13 @@ class InstructionWSystemPromptTokenizingStrategyTest(unittest.TestCase):
         assert example["input_ids"][8] == 11889  # USER
 
 
-class Llama2ChatTokenizationTest(unittest.TestCase):
+class Llama2ChatTokenizationTest:
     """
     Test class for prompt tokenization strategies with sys prompt from the dataset
     """
 
     @enable_hf_offline
-    def setUp(self) -> None:
-        # pylint: disable=duplicate-code
-        self.tokenizer = LlamaTokenizer.from_pretrained("NousResearch/Llama-2-7b-hf")
-        # woraround because official Meta repos are not open
-
-    def test_llama2_chat_integration(self):
+    def test_llama2_chat_integration(self, tokenizer_llama2_7b):
         with open(
             Path(__file__).parent / "fixtures/conversation.json", encoding="utf-8"
         ) as fin:
@@ -186,16 +155,18 @@ class Llama2ChatTokenizationTest(unittest.TestCase):
         prompter = Llama2ChatPrompter()
         strat = LLama2ChatTokenizingStrategy(
             prompter,
-            self.tokenizer,
+            tokenizer_llama2_7b,
             False,
             4096,
         )
         example = strat.tokenize_prompt(conversation)
         for fields in ["input_ids", "attention_mask", "labels"]:
-            self.assertEqual(len(example[fields]), len(tokenized_conversation[fields]))
-            self.assertEqual(example[fields], tokenized_conversation[fields])
+            # pytest assert equals
 
-    def compare_with_transformers_integration(self):
+            assert len(example[fields]) == len(tokenized_conversation[fields])
+            assert example[fields] == tokenized_conversation[fields]
+
+    def compare_with_transformers_integration(self, tokenizer_llama2_7b):
         # this needs transformers >= v4.31.0
         from transformers.models.llama.tokenization_llama import B_SYS, E_SYS
         from transformers.pipelines.conversational import Conversation
@@ -234,49 +205,27 @@ If a question does not make any sense, or is not factually coherent, explain why
             generated_responses=answers,
         )
         # pylint: disable=W0212
-        hf_tokens = self.tokenizer._build_conversation_input_ids(hf_conf)
+        hf_tokens = tokenizer_llama2_7b._build_conversation_input_ids(hf_conf)
 
-        self.assertEqual(
-            hf_tokens, tokenized_conversation["input_ids"][: len(hf_tokens)]
-        )
+        assert hf_tokens == tokenized_conversation["input_ids"][: len(hf_tokens)]
 
 
-class OrpoTokenizationTest(unittest.TestCase):
+class OrpoTokenizationTest:
     """test case for the ORPO tokenization"""
 
     @enable_hf_offline
-    def setUp(self) -> None:
-        # pylint: disable=duplicate-code
-        tokenizer = LlamaTokenizer.from_pretrained(
-            "casperhansen/mistral-7b-instruct-v0.1-awq"
-        )
-        tokenizer.add_special_tokens(
-            {
-                "eos_token": AddedToken(
-                    "<|im_end|>", rstrip=False, lstrip=False, normalized=False
-                )
-            }
-        )
-        tokenizer.add_tokens(
-            [
-                AddedToken(
-                    "<|im_start|>", rstrip=False, lstrip=False, normalized=False
-                ),
-            ]
-        )
-        self.tokenizer = tokenizer
-        self.dataset = load_dataset(
-            "argilla/ultrafeedback-binarized-preferences-cleaned", split="train"
-        ).select([0])
-
-    @pytest.mark.skip(reason="TODO: fix hf hub offline to work with HF rate limits")
-    def test_orpo_integration(self):
+    def test_orpo_integration(
+        self,
+        tokenizer_mistral_7b_instruct_chatml,
+        dataset_argilla_ultrafeedback_binarized_preferences_cleaned,
+    ):
+        ds = dataset_argilla_ultrafeedback_binarized_preferences_cleaned.select([0])
         strat = load(
-            self.tokenizer,
+            tokenizer_mistral_7b_instruct_chatml,
             DictDefault({"train_on_inputs": False}),
             DictDefault({"chat_template": "chatml"}),
         )
-        res = strat.tokenize_prompt(self.dataset[0])
+        res = strat.tokenize_prompt(ds[0])
         assert "rejected_input_ids" in res
         assert "rejected_labels" in res
         assert "input_ids" in res
@@ -295,7 +244,3 @@ class OrpoTokenizationTest(unittest.TestCase):
 
         assert res["prompt_attention_mask"][0] == 1
         assert res["prompt_attention_mask"][-1] == 0
-
-
-if __name__ == "__main__":
-    unittest.main()

tests/test_validation_dataset.py

@@ -321,3 +321,48 @@ class TestValidationCheckDatasetConfig(BaseValidation):
         )
 
         validate_config(cfg)
+
+
+class TestOptimizerValidation(BaseValidation):
+    """
+    Test muon optimizer validation
+    """
+
+    def test_muon_deepspeed(self, minimal_cfg):
+        cfg = DictDefault(
+            minimal_cfg
+            | {
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    }
+                ],
+                "optimizer": "muon",
+                "deepspeed": "deepspeed_configs/zero3.json",
+            }
+        )
+
+        with pytest.raises(ValueError, match=r".*is currently incompatible with*"):
+            validate_config(cfg)
+
+    def test_muon_fsdp(self, minimal_cfg):
+        cfg = DictDefault(
+            minimal_cfg
+            | {
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    }
+                ],
+                "optimizer": "muon",
+                "fsdp": ["full_shard"],
+                "fsdp_config": {
+                    "fsdp_auto_wrap_policy": "TRANSFORMER_BASED_WRAP",
+                },
+            }
+        )
+
+        with pytest.raises(ValueError, match=r".*is currently incompatible with*"):
+            validate_config(cfg)