add preliminary fp8 support

* fsdp2 support

* use accelerate release 1.6.0

* allow 8bit optims with fsdp2

* liger + torch compile fix

* add fsdp2 e2e tests

* use transformers commit with fsdp2 support

* skip zero3 tests for this PR for now

* fix fsdp2 config for ci

* make sure both flex and flash attn work with fsdp2, skip fix untrained tokens

* okay, actually use fdsp2...

* more fixes to flex for fsdp2

* make sure to patch all the loaded models

* additional validation for fsdp2, bump dep versions

.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

@@ -24,6 +24,13 @@ jobs:
       fail-fast: false
       matrix:
         include:
+          - cuda: 124
+            cuda_version: 12.4.1
+            python_version: "3.11"
+            pytorch: 2.6.0
+            axolotl_extras: vllm
+            num_gpus: 2
+            nightly_build: "true"
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
@@ -38,14 +45,6 @@ jobs:
             axolotl_extras: vllm
             num_gpus: 2
             nightly_build: "true"
-          - cuda: 124
-            cuda_version: 12.4.1
-            python_version: "3.11"
-            pytorch: 2.6.0
-            # awaiting vllm#12721
-            axolotl_extras:
-            num_gpus: 2
-            nightly_build: "true"
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     steps:

.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/
@@ -207,7 +211,7 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.5.1
+            pytorch: 2.6.0
             num_gpus: 1
             axolotl_extras: vllm
     steps:
@@ -254,9 +258,9 @@ jobs:
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.5.1
             num_gpus: 1
-            axolotl_extras:
+            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
@@ -243,6 +244,7 @@ website:
             - docs/unsloth.qmd
             - docs/torchao.qmd
             - docs/custom_integrations.qmd
+            - docs/sequence_parallelism.qmd
 
         - section: "Troubleshooting"
           contents:

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  --durations=10 -n2 /workspace/axolotl/tests/e2e/multigpu/ --ignore=/workspace/axolotl/tests/e2e/multigpu/solo/
+pytest -v  --durations=10 -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
 
@@ -658,6 +688,9 @@ ddp_broadcast_buffers:
 # subsequences, or set to 4 to split into four equal-sized subsequences.
 # See https://axolotl-ai-cloud.github.io/axolotl/docs/sequence_parallelism.html for more details.
 sequence_parallel_degree:
+# Optional; strides across the key dimension. Larger values use more memory but should make training faster.
+# Must evenly divide the number of KV heads in your model.
+heads_k_stride: 1
 
 # Path to torch distx for optim 'adamw_anyprecision'
 torchdistx_path:

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/multi-gpu.qmd

@@ -18,6 +18,7 @@ Axolotl supports several methods for multi-GPU training:
 
 - DeepSpeed (recommended)
 - FSDP (Fully Sharded Data Parallel)
+- Sequence parallelism
 - FSDP + QLoRA
 
 ## DeepSpeed {#sec-deepspeed}
@@ -66,6 +67,28 @@ fsdp_config:
   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
 ```
 
+## Sequence parallelism {#sec-sequence-parallelism}
+
+We support sequence parallelism (SP) via the
+[ring-flash-attention](https://github.com/zhuzilin/ring-flash-attention) project. This
+allows one to split up sequences across GPUs, which is useful in the event that a
+single sequence causes OOM errors during model training.
+
+First, install `ring-flash-attn`, recommended via `pip install axolotl[ring-flash-attn]`,
+or from source with `pip install .[ring-flash-attn]`.
+
+Your Axolotl YAML config should contain the following lines:
+
+```{.yaml}
+sequence_parallel_degree: 4  # Split each sequence into 4 parts, one per GPU
+flash_attention: true  # Required with sequence parallelism
+
+# Optional; strides across the key dimension. Larger values use more memory but will make training faster.
+heads_k_stride: 1
+```
+
+See our [dedicated guide](sequence_parallelism.qmd) for more details.
+
 ### FSDP + QLoRA {#sec-fsdp-qlora}
 
 For combining FSDP with QLoRA, see our [dedicated guide](fsdp_qlora.qmd).

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]

docs/sequence_parallelism.qmd

@@ -25,6 +25,8 @@ To enable sequence parallelism, add the following to your configuration file:
 ```yaml
 # Set to a divisor (> 1) of the number of GPUs available
 sequence_parallel_degree: 4  # Split sequences across 4 GPUs
+# Optional; strides across the key dimension. Larger values use more memory but should make training faster.
+heads_k_stride: 1
 ```
 
 The `sequence_parallel_degree` should be a divisor of the total number of GPUs. For example:
@@ -58,11 +60,16 @@ To use sequence parallelism, you need:
 ## Example
 
 ```yaml
-# Example config with sequence parallelism
 base_model: meta-llama/Llama-3-8B-Instruct
 sequence_len: 8192
-sequence_parallel_degree: 2  # Split each sequence into 4 parts
+
+...
+
+sequence_parallel_degree: 4  # Split each sequence into 4 parts, one per GPU
 flash_attention: true  # Required with sequence parallelism
+# Optional; strides across the key dimension. Larger values use more memory but should make training faster.
+heads_k_stride: 1
+
 ...
 ```
 

examples/cerebras/btlm-ft.yml

@@ -8,9 +8,6 @@ tokenizer_type: GPT2Tokenizer
 trust_remote_code: true
 tokenizer_use_fast: true
 tokenizer_legacy: true
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 push_dataset_to_hub:
 hf_use_auth_token: true
@@ -34,7 +31,6 @@ lora_alpha:
 lora_dropout:
 lora_target_modules:
 lora_target_linear:
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -58,16 +54,12 @@ learning_rate: 0.000085
 train_on_inputs: true
 group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 
 gradient_checkpointing: false
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
 
-xformers_attention:
 flash_attention: true
 sdp_attention:
 flash_optimum:
@@ -80,8 +72,6 @@ evals_per_epoch: 4
 saves_per_epoch: 1
 save_total_limit:
 
-debug:
-deepspeed:
 weight_decay: 0.1
 special_tokens:
   pad_token: "<|endoftext|>"

examples/cerebras/qlora.yml

@@ -22,7 +22,6 @@ lora_target_modules:
   - c_attn
   - c_proj
 lora_target_linear:
-lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
@@ -36,15 +35,10 @@ optimizer: paged_adamw_8bit
 torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.0002
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
 xformers_attention: true
 flash_attention:
@@ -53,10 +47,6 @@ gptq_model_v1:
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|endoftext|>"

examples/code-llama/13b/lora.yml

@@ -26,7 +26,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -41,29 +40,18 @@ 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
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/code-llama/13b/qlora.yml

@@ -26,9 +26,7 @@ pad_to_sequence_len: true
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -43,28 +41,18 @@ optimizer: paged_adamw_32bit
 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
 
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/code-llama/34b/lora.yml

@@ -26,7 +26,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -41,29 +40,18 @@ 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
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/code-llama/34b/qlora.yml

@@ -26,9 +26,7 @@ pad_to_sequence_len: true
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -43,28 +41,18 @@ optimizer: paged_adamw_32bit
 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
 
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/code-llama/7b/lora.yml

@@ -26,7 +26,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -41,29 +40,18 @@ 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
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/code-llama/7b/qlora.yml

@@ -26,9 +26,7 @@ pad_to_sequence_len: true
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -43,28 +41,18 @@ optimizer: paged_adamw_32bit
 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
 
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/cohere/command-r-7b-qlora.yml

@@ -44,28 +44,16 @@ optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_ratio: 0.1
 evals_per_epoch:
-eval_table_size:
-eval_max_new_tokens: 128
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:

examples/dbrx/16bit-lora.yaml

@@ -3,9 +3,6 @@ base_model: LnL-AI/dbrx-base-converted-v2
 # hub_model_id: username/custom_model_name
 
 trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -48,26 +45,20 @@ optimizer: paged_adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: false  # don't use with fsdp_activation_checkpointing
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
 evals_per_epoch:
 saves_per_epoch: 1
-debug:
+
 weight_decay: 0.0
 fsdp:
   - full_shard

examples/dbrx/8bit-lora.yaml

@@ -48,26 +48,20 @@ optimizer: paged_adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: false  # don't use with fsdp_activation_checkpointing
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
 evals_per_epoch:
 saves_per_epoch: 1
-debug:
+
 weight_decay: 0.0
 fsdp:
   - full_shard

examples/dbrx/fft-ds-zero3.yaml

@@ -3,9 +3,6 @@ base_model: LnL-AI/dbrx-base-converted-v2
 # hub_model_id: username/custom_model_name
 
 trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -35,25 +32,19 @@ optimizer: paged_adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
 evals_per_epoch:
 saves_per_epoch: 1
-debug:
+
 weight_decay: 0.0
 deepspeed: deepspeed_configs/zero3_bf16.json

examples/deepseek-v2/fft-fsdp-16b.yaml

@@ -2,9 +2,6 @@ base_model: deepseek-ai/DeepSeek-V2-Lite
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -31,27 +28,19 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 2e-5
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 100
 evals_per_epoch: 2
-eval_table_size:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
 special_tokens:
 fsdp:

examples/deepseek-v2/qlora-fsdp-2_5.yaml

@@ -52,27 +52,19 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 2e-5
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 100
 evals_per_epoch: 2
-eval_table_size:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
 special_tokens:
 fsdp:

examples/falcon/config-7b-lora.yml

@@ -25,9 +25,7 @@ max_packed_sequence_len:
 lora_r: 16
 lora_alpha: 32
 lora_dropout: 0.0
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
@@ -41,15 +39,10 @@ optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.00003
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
 xformers_attention: true
 flash_attention:
@@ -58,11 +51,7 @@ gptq_model_v1:
 warmup_steps: 40
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|endoftext|>"
   bos_token: "<|endoftext|>"

examples/falcon/config-7b-qlora.yml

@@ -38,9 +38,7 @@ lora_alpha: 16
 # 0.05 for 33B and 65B models
 lora_dropout: 0.05
 # add LoRA modules on all linear layers of the base model
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -67,10 +65,7 @@ lr_scheduler: cosine
 # - 2e-4 for 7b & 13b
 # - 1e-4 for 33b & 64b
 learning_rate: 0.0002
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 gradient_checkpointing: true
 # stop training after this many evaluation losses have increased in a row
@@ -78,7 +73,6 @@ gradient_checkpointing: true
 early_stopping_patience: 3
 resume_from_checkpoint:
 auto_resume_from_checkpoints: true
-local_rank:
 logging_steps: 1
 xformers_attention: true
 flash_attention:
@@ -87,11 +81,7 @@ gptq_model_v1:
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.000001
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|endoftext|>"
   bos_token: "<|endoftext|>"

examples/falcon/config-7b.yml

@@ -7,9 +7,6 @@ tokenizer_type: AutoTokenizer
 
 # required by falcon custom model code: https://huggingface.co/tiiuae/falcon-7b/tree/main
 trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
 gptq: false
 strict: false
 push_dataset_to_hub:
@@ -25,9 +22,7 @@ max_packed_sequence_len:
 lora_r: 64
 lora_alpha: 32
 lora_dropout: 0.0
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
@@ -41,15 +36,10 @@ optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.00003
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
 xformers_attention: true
 flash_attention:
@@ -58,11 +48,7 @@ gptq_model_v1:
 warmup_steps: 40
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|endoftext|>"
   bos_token: "<|endoftext|>"

examples/gemma/qlora.yml

@@ -42,28 +42,16 @@ 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
 
 warmup_ratio: 0.1
 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:

examples/gemma2/qlora.yml

@@ -48,28 +48,16 @@ optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_ratio: 0.1
 evals_per_epoch:
-eval_table_size:
-eval_max_new_tokens: 128
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:

examples/gemma2/reward-model.yaml

@@ -5,9 +5,6 @@ num_labels: 1
 tokenizer_type: AutoTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 reward_model: true
@@ -38,8 +35,6 @@ optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: true
 fp16:
 tf32: true
@@ -47,21 +42,12 @@ tf32: true
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_ratio: 0.1
 evals_per_epoch:
-eval_table_size:
-eval_max_new_tokens: 128
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:

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
@@ -47,28 +50,18 @@ optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 
 gradient_checkpointing: true
-early_stopping_patience:
+gradient_checkpointing_kwargs:
+  use_reentrant: false
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_ratio: 0.1
 evals_per_epoch:
-eval_table_size:
-eval_max_new_tokens: 128
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:

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

@@ -0,0 +1,61 @@
+base_model: google/gemma-3-4b-it
+strict: false
+
+load_in_4bit: true
+
+# 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: qlora
+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
+
+bf16: true
+fp16:
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+logging_steps: 1
+flash_attention: true
+eager_attention:
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0

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

@@ -2,11 +2,16 @@ base_model: google/gemma-3-4b-it
 processor_type: AutoProcessor
 strict: false
 
+load_in_4bit: true
+
 # these 3 lines are needed for now to handle vision chat templates w images
 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 +22,7 @@ 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
@@ -41,14 +46,13 @@ 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
-local_rank:
+gradient_checkpointing_kwargs:
+  use_reentrant: false
 logging_steps: 1
 flash_attention: true
 eager_attention:
@@ -56,8 +60,4 @@ eager_attention:
 warmup_ratio: 0.1
 evals_per_epoch: 1
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:

examples/gptj/qlora.yml

@@ -18,9 +18,7 @@ max_packed_sequence_len:
 lora_r: 8
 lora_alpha: 32
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
@@ -34,15 +32,10 @@ optimizer: paged_adamw_8bit
 torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.0001
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
 xformers_attention: true
 flash_attention:
@@ -51,10 +44,6 @@ gptq_model_v1:
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|endoftext|>"

examples/jamba/qlora.yaml

@@ -40,26 +40,18 @@ optimizer: paged_adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.00001
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
 evals_per_epoch:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
 special_tokens:

examples/jamba/qlora_deepspeed.yaml

@@ -39,26 +39,20 @@ optimizer: paged_adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.00001
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
 evals_per_epoch:
 saves_per_epoch: 1
-debug:
+
 deepspeed: deepspeed_configs/zero2.json
 weight_decay: 0.0
 special_tokens:

examples/jamba/qlora_fsdp_large.yaml

@@ -39,8 +39,6 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 0.00001
 
-train_on_inputs: false
-group_by_length: false
 bf16: true
 tf32: true
 

examples/jeopardy-bot/config.yml

@@ -33,13 +33,9 @@ optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.00003
-train_on_inputs: false
-group_by_length: false
 bf16: auto
 tf32: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 5
 xformers_attention: true
 flash_attention:
@@ -48,11 +44,7 @@ gptq_model_v1:
 warmup_steps: 20
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
-fsdp:
-fsdp_config:
 tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/llama-2/fft_optimized.yml

@@ -4,9 +4,6 @@ model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -26,7 +23,6 @@ lora_r:
 lora_alpha:
 lora_dropout:
 lora_target_linear:
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -41,18 +37,12 @@ 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
 flash_attn_cross_entropy: false
 flash_attn_rms_norm: true
@@ -61,11 +51,8 @@ flash_attn_fuse_mlp: true
 
 warmup_steps: 100
 evals_per_epoch: 4
-eval_table_size:
 saves_per_epoch: 1
-debug:
+
 deepspeed: #deepspeed_configs/zero2.json # multi-gpu only
 weight_decay: 0.1
-fsdp:
-fsdp_config:
 special_tokens:

examples/llama-2/gptq-lora.yml

@@ -10,8 +10,6 @@ gptq_disable_exllama: true
 
 tokenizer_use_fast: true
 tokenizer_legacy: true
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 push_dataset_to_hub:
 hf_use_auth_token: true
@@ -33,7 +31,6 @@ lora_target_modules:
   - q_proj
   - v_proj
 lora_target_linear:
-lora_fan_in_fan_out:
 wandb_project:
 wandb_watch:
 wandb_name:
@@ -50,26 +47,19 @@ torchdistx_path:
 lr_scheduler: cosine
 lr_quadratic_warmup: true
 learning_rate: 0.000017
-train_on_inputs: false
-group_by_length: false
 bf16: false
 fp16: false
 float16: true
 tf32: true
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention:
 sdp_attention:
 flash_optimum:
 warmup_steps: 100
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
 special_tokens:
   bos_token: "<s>"

examples/llama-2/lisa.yml

@@ -4,9 +4,6 @@ model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -26,7 +23,6 @@ lora_r:
 lora_alpha:
 lora_dropout:
 lora_target_linear:
-lora_fan_in_fan_out:
 
 lisa_n_layers: 4
 lisa_step_interval: 20
@@ -45,18 +41,12 @@ optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 5e-5 # recommendation from lisa paper for 7b
 
-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
 flash_attn_cross_entropy: false
 flash_attn_rms_norm: true
@@ -65,13 +55,8 @@ flash_attn_fuse_mlp: true
 
 warmup_steps: 100
 evals_per_epoch: 4
-eval_table_size:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/llama-2/loftq.yml

@@ -4,9 +4,6 @@ model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -26,7 +23,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 peft:
   loftq_config:
     loftq_bits: 4
@@ -44,29 +40,16 @@ 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:

examples/llama-2/lora.yml

@@ -26,7 +26,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -41,29 +40,16 @@ 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:

examples/llama-2/qlora-fsdp.yml

@@ -26,9 +26,7 @@ pad_to_sequence_len: true
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -43,28 +41,19 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 0.00001
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
 evals_per_epoch: 4
-eval_table_size:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
 fsdp:
   - full_shard

examples/llama-2/qlora.yml

@@ -26,9 +26,7 @@ pad_to_sequence_len: true
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -43,27 +41,16 @@ optimizer: paged_adamw_32bit
 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
 
 warmup_steps: 10
 evals_per_epoch: 4
-eval_table_size:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:

examples/llama-2/relora.yml

@@ -24,9 +24,7 @@ pad_to_sequence_len: true
 lora_r: 8
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 relora_steps: 150
 relora_warmup_steps: 10
@@ -45,28 +43,18 @@ 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
 
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/llama-3-vision/lora-11b.yaml

@@ -45,14 +45,11 @@ 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
-local_rank:
 logging_steps: 1
 flash_attention: true
 eager_attention:
@@ -60,8 +57,4 @@ eager_attention:
 warmup_ratio: 0.1
 evals_per_epoch: 1
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:

examples/llama-3/fft-8b-liger-fsdp.yaml

@@ -42,27 +42,19 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 2e-5
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 100
 evals_per_epoch: 2
-eval_table_size:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
 fsdp:
   - full_shard

examples/llama-3/fft-8b.yaml

@@ -1,9 +1,6 @@
 base_model: NousResearch/Meta-Llama-3.1-8B
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -30,29 +27,19 @@ optimizer: paged_adamw_8bit
 lr_scheduler: cosine
 learning_rate: 2e-5
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 100
 evals_per_epoch: 2
-eval_table_size:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: <|end_of_text|>

examples/llama-3/instruct-dpo-lora-8b.yml

@@ -42,7 +42,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -57,28 +56,15 @@ 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:

examples/llama-3/instruct-lora-8b.yml

@@ -37,7 +37,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -52,30 +51,17 @@ 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|>

examples/llama-3/lora-1b-deduplicate-dpo.yml

@@ -58,7 +58,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -73,28 +72,15 @@ 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:

examples/llama-3/lora-1b-deduplicate-sft.yml

@@ -31,7 +31,6 @@ 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
@@ -49,30 +48,17 @@ 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|>

examples/llama-3/lora-1b-kernels.yml

@@ -1,9 +1,6 @@
 base_model: NousResearch/Llama-3.2-1B
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -24,7 +21,6 @@ lora_r: 16
 lora_alpha: 32
 # Currently, we don't support dropout with our custom Triton kernels
 # lora_dropout: 0.05
-lora_fan_in_fan_out:
 lora_target_modules:
   - gate_proj
   - down_proj
@@ -53,18 +49,12 @@ optimizer: adamw_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
 
 loss_watchdog_threshold: 5.0
@@ -73,10 +63,6 @@ loss_watchdog_patience: 3
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|end_of_text|>"

examples/llama-3/lora-1b-ray.yml

@@ -1,9 +1,6 @@
 base_model: NousResearch/Llama-3.2-1B
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -24,7 +21,6 @@ pad_to_sequence_len: true
 lora_r: 16
 lora_alpha: 32
 lora_dropout: 0.05
-lora_fan_in_fan_out:
 lora_target_modules:
   - gate_proj
   - down_proj
@@ -47,18 +43,12 @@ optimizer: adamw_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
 
 loss_watchdog_threshold: 5.0
@@ -67,11 +57,9 @@ loss_watchdog_patience: 3
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
+
 deepspeed: deepspeed_configs/zero3.json
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|end_of_text|>"
 

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

@@ -0,0 +1,66 @@
+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_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
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+  pad_token: <|end_of_text|>

examples/llama-3/lora-1b.yml

@@ -1,9 +1,6 @@
 base_model: NousResearch/Llama-3.2-1B
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -24,7 +21,6 @@ pad_to_sequence_len: true
 lora_r: 16
 lora_alpha: 32
 lora_dropout: 0.05
-lora_fan_in_fan_out:
 lora_target_modules:
   - gate_proj
   - down_proj
@@ -47,18 +43,12 @@ optimizer: adamw_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
 
 loss_watchdog_threshold: 5.0
@@ -67,10 +57,6 @@ loss_watchdog_patience: 3
 warmup_steps: 10
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|end_of_text|>"

examples/llama-3/lora-8b.yml

@@ -27,7 +27,6 @@ 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
@@ -45,30 +44,17 @@ 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|>

examples/llama-3/qlora-1b-kto.yaml

@@ -32,7 +32,6 @@ lora_r: 32
 lora_alpha: 64
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -47,31 +46,19 @@ optimizer: adamw_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 20
 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|>"

examples/llama-3/qlora-1b.yml

@@ -24,7 +24,6 @@ pad_to_sequence_len: true
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_fan_in_fan_out:
 lora_target_modules:
   - gate_proj
   - down_proj
@@ -47,18 +46,12 @@ 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
 
 loss_watchdog_threshold: 5.0
@@ -66,13 +59,7 @@ loss_watchdog_patience: 3
 
 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|>"

examples/llama-3/qlora-fsdp-405b.yaml

@@ -24,7 +24,6 @@ pad_to_sequence_len: true
 lora_r: 16
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
 
 gradient_accumulation_steps: 4
@@ -34,8 +33,6 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 0.00001
 
-train_on_inputs: false
-group_by_length: false
 bf16: true
 tf32: true
 

examples/llama-3/qlora-fsdp-70b.yaml

@@ -26,9 +26,7 @@ pad_to_sequence_len: true
 lora_r: 8
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -43,28 +41,19 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 0.00001
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: false
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 10
 evals_per_epoch: 4
-eval_table_size:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
 fsdp:
   - full_shard

examples/llama-3/qlora.yml

@@ -26,9 +26,7 @@ pad_to_sequence_len: true
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -43,28 +41,17 @@ optimizer: paged_adamw_32bit
 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
 
 warmup_steps: 10
 evals_per_epoch: 4
-eval_table_size:
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: "<|end_of_text|>"

examples/llama4/scout-lora.yaml

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

examples/llava/lora-7b.yaml

@@ -41,14 +41,11 @@ 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
-local_rank:
 logging_steps: 1
 flash_attention: true
 eager_attention:
@@ -56,8 +53,4 @@ eager_attention:
 warmup_ratio: 0.1
 evals_per_epoch: 1
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:

examples/mamba/config.yml

@@ -5,9 +5,6 @@ tokenizer_type: AutoTokenizer
 tokenizer_config: EleutherAI/gpt-neox-20b
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -38,27 +35,17 @@ train_on_inputs: false
 group_by_length: true
 
 bf16: auto
-fp16:
 tf32: true
 
 gradient_checkpointing: false
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_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:
 tokens:
 save_safetensors: False

examples/mistral/bigstral-ds-zero3.yaml

@@ -6,9 +6,6 @@ tokenizer_type: LlamaTokenizer
 # hub_model_id: username/custom_model_name
 
 trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 unfrozen_parameters:
@@ -40,27 +37,19 @@ optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0001
 
-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
 
 save_total_limit: 1
 save_steps:
-debug:
+
 deepspeed: deepspeed_configs/zero3_bf16_cpuoffload_params.json
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   eos_token: "<|im_end|>"
 tokens:

examples/mistral/config.yml

@@ -4,9 +4,6 @@ model_type: MistralForCausalLM
 tokenizer_type: LlamaTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -34,28 +31,16 @@ optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.000005
 
-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
 
 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:

examples/mistral/lora-mps.yml

@@ -4,9 +4,6 @@ model_type: MistralForCausalLM
 tokenizer_type: LlamaTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -28,7 +25,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 lora_target_modules:
   - gate_proj
   - down_proj
@@ -51,18 +47,13 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
 fp16: false
 tf32: true
 
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: false
 sdp_attention: true
 
@@ -71,12 +62,6 @@ loss_watchdog_patience: 3
 
 warmup_steps: 10
 evals_per_epoch: 4
-eval_table_size:
-eval_table_max_new_tokens: 128
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:

examples/mistral/lora.yml

@@ -27,7 +27,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 lora_target_modules:
   - gate_proj
   - down_proj
@@ -50,18 +49,12 @@ 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
 
 loss_watchdog_threshold: 5.0
@@ -69,12 +62,6 @@ loss_watchdog_patience: 3
 
 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:

examples/mistral/mistral-dpo-qlora.yml

@@ -40,7 +40,6 @@ lora_r: 8
 lora_alpha: 16
 lora_dropout: 0.2
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 lora_target_modules:
   - gate_proj
@@ -67,31 +66,18 @@ optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0001
 
-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: false
-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:
   bos_token: "<|im_start|>"
   eos_token: "<|im_end|>"

examples/mistral/mistral-qlora-fsdp.yml

@@ -32,7 +32,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -47,18 +46,12 @@ optimizer: paged_adamw_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
 
 loss_watchdog_threshold: 5.0
@@ -66,10 +59,8 @@ loss_watchdog_patience: 3
 
 warmup_steps: 10
 evals_per_epoch: 4
-eval_table_size:
-eval_max_new_tokens: 128
 saves_per_epoch: 1
-debug:
+
 weight_decay: 0.0
 fsdp:
   - full_shard

examples/mistral/mistral-qlora-orpo.yml

@@ -32,7 +32,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 lora_target_modules:
   - gate_proj
   - down_proj
@@ -55,18 +54,12 @@ 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
 
 loss_watchdog_threshold: 5.0
@@ -74,12 +67,6 @@ loss_watchdog_patience: 3
 
 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:

examples/mistral/mistral-small-3.1-24B-lora.yml

@@ -43,14 +43,11 @@ 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
-local_rank:
 logging_steps: 1
 flash_attention: false # PixtralVisionModel does not support Flash Attention 2.0 yet.
 eager_attention:
@@ -58,9 +55,5 @@ eager_attention:
 warmup_ratio: 0.1
 evals_per_epoch: 1
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:

examples/mistral/mixtral-8x22b-qlora-fsdp.yml

@@ -30,7 +30,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -45,18 +44,12 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 loss_watchdog_threshold: 5.0
@@ -64,10 +57,8 @@ loss_watchdog_patience: 3
 
 warmup_steps: 10
 evals_per_epoch: 4
-eval_table_size:
-eval_max_new_tokens: 128
 saves_per_epoch: 1
-debug:
+
 weight_decay: 0.0
 fsdp:
   - full_shard

examples/mistral/mixtral-qlora-fsdp.yml

@@ -32,7 +32,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -47,18 +46,12 @@ optimizer: adamw_torch_fused
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 loss_watchdog_threshold: 5.0
@@ -66,10 +59,8 @@ loss_watchdog_patience: 3
 
 warmup_steps: 10
 evals_per_epoch: 4
-eval_table_size:
-eval_max_new_tokens: 128
 saves_per_epoch: 1
-debug:
+
 weight_decay: 0.0
 fsdp:
   - full_shard

examples/mistral/mixtral.yml

@@ -41,7 +41,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 #lora_target_modules:
 #  - gate
 #  - q_proj
@@ -65,18 +64,12 @@ 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
 
 loss_watchdog_threshold: 5.0
@@ -84,12 +77,8 @@ loss_watchdog_patience: 3
 
 warmup_steps: 10
 evals_per_epoch: 4
-eval_table_size:
-eval_max_new_tokens: 128
 saves_per_epoch: 1
-debug:
+
 deepspeed: deepspeed_configs/zero2.json
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:

examples/mistral/mixtral_22.yml

@@ -6,9 +6,6 @@ tokenizer_type: LlamaTokenizer
 # hub_model_id: username/custom_model_name
 
 trust_remote_code: true
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 unfrozen_parameters:
@@ -38,27 +35,19 @@ optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0001
 
-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
 
 save_total_limit: 1
 save_steps:
-debug:
+
 deepspeed: deepspeed_configs/zero3_bf16_cpuoffload_all.json
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   eos_token: "<|im_end|>"
 tokens:

examples/mistral/qlora.yml

@@ -27,7 +27,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 lora_target_modules:
   - gate_proj
   - down_proj
@@ -50,18 +49,12 @@ 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
 
 loss_watchdog_threshold: 5.0
@@ -69,12 +62,6 @@ loss_watchdog_patience: 3
 
 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:

examples/mpt-7b/config.yml

@@ -35,26 +35,17 @@ optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.0000002
-train_on_inputs: false
-group_by_length: false
 bf16: auto
 tf32: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 5
-xformers_attention:
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0001
-fsdp:
-fsdp_config:
 tokens:
   pad_token: "<|padding|>"
   bos_token: "<|endoftext|>"

examples/openllama-3b/config.yml

@@ -4,9 +4,6 @@ model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 push_dataset_to_hub:
 datasets:
@@ -23,7 +20,6 @@ lora_alpha:
 lora_dropout:
 lora_target_modules:
 lora_target_linear:
-lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
@@ -37,29 +33,20 @@ optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.000003
-train_on_inputs: false
-group_by_length: false
 float16: true
 bf16: false
 fp16: false
 tf32: false
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/openllama-3b/lora.yml

@@ -29,7 +29,6 @@ lora_target_modules:
   - v_proj
   - k_proj
   - o_proj
-lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
@@ -43,29 +42,19 @@ optimizer: adamw_bnb_8bit
 torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.0002
-train_on_inputs: false
-group_by_length: false
 bf16: false
 fp16: true
 tf32: false
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 gptq_groupsize:
-s2_attention:
 gptq_model_v1:
 warmup_steps: 20
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/openllama-3b/qlora.yml

@@ -21,9 +21,7 @@ sample_packing: true
 lora_r: 8
 lora_alpha: 32
 lora_dropout: 0.05
-lora_target_modules:
 lora_target_linear: true
-lora_fan_in_fan_out:
 wandb_project:
 wandb_entity:
 wandb_watch:
@@ -37,28 +35,19 @@ optimizer: paged_adamw_32bit
 torchdistx_path:
 lr_scheduler: cosine
 learning_rate: 0.0002
-train_on_inputs: false
-group_by_length: false
 bf16: false
 fp16: true
 tf32: false
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
-fsdp:
-fsdp_config:
 special_tokens:
   bos_token: "<s>"
   eos_token: "</s>"

examples/phi/lora-3.5.yaml

@@ -37,7 +37,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -52,28 +51,16 @@ optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
-train_on_inputs: false
-group_by_length: false
 bfloat16: true
 bf16: true
 fp16:
 tf32: false
 
 gradient_checkpointing: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
-s2_attention:
 
 warmup_steps: 10
 evals_per_epoch: 4
-eval_table_size:
-eval_max_new_tokens: 128
 saves_per_epoch: 4
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:

examples/phi/phi-ft.yml

@@ -4,9 +4,6 @@ model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -27,7 +24,6 @@ lora_r:
 lora_alpha:
 lora_dropout:
 lora_target_linear:
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -45,30 +41,20 @@ max_grad_norm: 1.0
 lr_scheduler: cosine
 learning_rate: 0.000003
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: True
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 100
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
-fsdp:
-fsdp_config:
 resize_token_embeddings_to_32x: true
 special_tokens:
   pad_token: "<|endoftext|>"

examples/phi/phi-qlora.yml

@@ -27,7 +27,6 @@ lora_r: 64
 lora_alpha: 32
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -45,30 +44,20 @@ max_grad_norm: 1.0
 lr_scheduler: cosine
 learning_rate: 0.000003
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: True
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 100
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
-fsdp:
-fsdp_config:
 resize_token_embeddings_to_32x: true
 special_tokens:
   pad_token: "<|endoftext|>"

examples/phi/phi2-ft.yml

@@ -4,9 +4,6 @@ model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -27,7 +24,6 @@ lora_r:
 lora_alpha:
 lora_dropout:
 lora_target_linear:
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -45,30 +41,20 @@ max_grad_norm: 1.0
 lr_scheduler: cosine
 learning_rate: 0.000003
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: True
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 100
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
-fsdp:
-fsdp_config:
 resize_token_embeddings_to_32x: true
 special_tokens:
   pad_token: "<|endoftext|>"

examples/phi/phi3-ft-fsdp.yml

@@ -4,9 +4,6 @@ model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -28,7 +25,6 @@ lora_r:
 lora_alpha:
 lora_dropout:
 lora_target_linear:
-lora_fan_in_fan_out:
 
 wandb_project: phi3
 wandb_entity:
@@ -46,27 +42,19 @@ max_grad_norm: 1.0
 lr_scheduler: cosine
 learning_rate: 0.000003
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
-fp16:
 tf32: true
 
 gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_attention: true
 
 warmup_steps: 100
 evals_per_epoch: 4
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.1
 fsdp:
   - full_shard

examples/phi/phi3-ft.yml

@@ -7,9 +7,6 @@ tokenizer_type: AutoTokenizer
 # hub_model_id: username/custom_model_name
 
 chat_template: phi_3
-
-load_in_8bit: false
-load_in_4bit: false
 strict: false
 
 datasets:
@@ -30,7 +27,6 @@ lora_r: 64
 lora_alpha: 32
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 gradient_accumulation_steps: 1
 micro_batch_size: 2
@@ -42,8 +38,6 @@ max_grad_norm: 1.0
 lr_scheduler: cosine
 learning_rate: 5.0e-6
 
-train_on_inputs: false
-group_by_length: false
 bf16: auto
 
 gradient_checkpointing: true
@@ -55,9 +49,9 @@ flash_attention: true
 
 eval_steps: 1000
 save_steps: 5000
-eval_table_size: 2
 eval_batch_size: 2
 eval_sample_packing: false
+eval_table_size: 2
 eval_max_new_tokens: 32
 eval_causal_lm_metrics: ["perplexity"]
 do_causal_lm_eval: true

examples/pixtral/lora-12b.yml

@@ -41,14 +41,11 @@ 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
-local_rank:
 logging_steps: 1
 flash_attention: false # PixtralVisionModel does not support Flash Attention 2.0 yet
 eager_attention:
@@ -56,10 +53,6 @@ eager_attention:
 warmup_ratio: 0.1
 evals_per_epoch: 1
 saves_per_epoch: 1
-debug:
-deepspeed:
 weight_decay: 0.0
-fsdp:
-fsdp_config:
 special_tokens:
   pad_token: <pad>

examples/pythia-12b/config.yml

@@ -5,9 +5,6 @@ model_type: GPTNeoXForCausalLM
 tokenizer_type: AutoTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-
-load_in_8bit: false
-load_in_4bit: false
 gptq: false
 device_map: auto
 datasets:
@@ -22,7 +19,6 @@ max_packed_sequence_len: 2048
 lora_r: 64
 lora_alpha: 32
 lora_dropout: 0.0
-lora_target_modules:
 lora_target_linear: true
 lora_fan_in_fan_out: true  # pythia/GPTNeoX lora specific
 wandb_project:
@@ -37,16 +33,10 @@ num_epochs: 5
 learning_rate: 0.00003
 optimizer: adamw_bnb_8bit
 lr_scheduler: cosine
-train_on_inputs: false
-group_by_length: false
 bf16: false
 fp16: false
 float16: true
 tf32: true
 flash_optimum: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 gradient_checkpointing: true
-fsdp:
-fsdp_config:

examples/pythia/lora.yml

@@ -28,13 +28,9 @@ gradient_accumulation_steps: 1
 micro_batch_size: 4
 num_epochs: 4
 learning_rate: 0.00001
-train_on_inputs: false
-group_by_length: false
 bf16: auto
 tf32: true
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 weight_decay: 0.1
 evals_per_epoch: 4
 logging_steps: 1

examples/qwen/lora.yml

@@ -28,7 +28,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -43,28 +42,16 @@ 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: false
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_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:

examples/qwen/qlora.yml

@@ -28,7 +28,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 lora_target_linear: true
-lora_fan_in_fan_out:
 
 wandb_project:
 wandb_entity:
@@ -43,28 +42,16 @@ 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: false
-early_stopping_patience:
 resume_from_checkpoint:
-local_rank:
 logging_steps: 1
-xformers_attention:
 flash_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: