tag for v0.12.0 release (#3041)

* feat(doc): add links to new features on README

* fix merge error

* remove blurb about older FSDP2 integration

* update blog link

* chore: update cce commit

* feat: update model support into readme

* Update README.md

Co-authored-by: salman <salman.mohammadi@outlook.com>

* chore: lint num spaces

---------

Co-authored-by: Wing Lian <wing@axolotl.ai>
Co-authored-by: salman <salman.mohammadi@outlook.com>

.axolotl-complete.bash

@@ -0,0 +1,41 @@
+#!/bin/bash
+
+_axolotl_completions() {
+    local cur prev
+    COMPREPLY=()
+    cur="${COMP_WORDS[COMP_CWORD]}"
+    prev="${COMP_WORDS[COMP_CWORD-1]}"
+
+    # If we're completing the first argument (the command)
+    if [[ $COMP_CWORD -eq 1 ]]; then
+        mapfile -t COMPREPLY < <(compgen -W "delinearize-llama4 fetch lm-eval merge-sharded-fsdp-weights quantize vllm-serve evaluate inference merge-lora preprocess train" -- "$cur")
+        return 0
+    fi
+
+    # Commands that should complete with directories and YAML files
+    local -a yaml_commands=("merge-sharded-fsdp-weights" "quantize" "vllm-serve" "evaluate" "inference" "merge-lora" "preprocess" "train")
+
+    # Check if previous word is in our list
+    if [[ " ${yaml_commands[*]} " =~ (^|[[:space:]])$prev($|[[:space:]]) ]]; then
+        # Use filename completion which handles directories properly
+        compopt -o filenames
+        mapfile -t COMPREPLY < <(compgen -f -- "$cur")
+
+        # Filter to only include directories and YAML files
+        local -a filtered=()
+        for item in "${COMPREPLY[@]}"; do
+            if [[ -d "$item" ]] || [[ "$item" == *.yaml ]] || [[ "$item" == *.yml ]]; then
+                filtered+=("$item")
+            fi
+        done
+        COMPREPLY=("${filtered[@]}")
+
+        return 0
+    fi
+
+    # Default: no completion
+    return 0
+}
+
+# Remove the -o nospace option - let filenames handle it
+complete -F _axolotl_completions axolotl

.coderabbit.yaml

@@ -0,0 +1,16 @@
+# yaml-language-server: $schema=https://coderabbit.ai/integrations/schema.v2.json
+language: "en-US"
+early_access: false
+reviews:
+  profile: "chill"
+  request_changes_workflow: false
+  high_level_summary: true
+  review_status: true
+  collapse_walkthrough: true
+  poem: false
+  sequence_diagrams: false
+  auto_review:
+    enabled: true
+    drafts: false
+chat:
+  auto_reply: true

.github/workflows/base.yml

@@ -17,7 +17,7 @@ on:
 
 jobs:
   build-base:
-    if: github.repository_owner == 'axolotl-ai-cloud'
+    if: ${{ github.repository_owner == 'axolotl-ai-cloud' && (github.event_name != 'pull_request' || !github.event.pull_request.draft) }}
     timeout-minutes: 480
     # this job needs to be run on self-hosted GPU runners...
     runs-on: ubuntu-latest-m
@@ -54,7 +54,7 @@ jobs:
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-base"
           - cuda: "128"
-            cuda_version: 12.6.3
+            cuda_version: 12.8.1
             cudnn_version: ""
             python_version: "3.11"
             pytorch: 2.7.1
@@ -64,9 +64,16 @@ jobs:
             cuda_version: 12.8.1
             cudnn_version: ""
             python_version: "3.11"
-            pytorch: nightly
+            pytorch: 2.8.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
-            dockerfile: "Dockerfile-base-nightly"
+            dockerfile: "Dockerfile-base"
+#          - cuda: "128"
+#            cuda_version: 12.8.1
+#            cudnn_version: ""
+#            python_version: "3.11"
+#            pytorch: nightly
+#            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+#            dockerfile: "Dockerfile-base-nightly"
 #          # "next" is for release candidates of pytorch
 #          - cuda: "128"
 #            cuda_version: 12.8.1
@@ -108,7 +115,7 @@ jobs:
             PYTORCH_VERSION=${{ matrix.pytorch }}
             TORCH_CUDA_ARCH_LIST=${{ matrix.torch_cuda_arch_list }}
   build-base-uv:
-    if: github.repository_owner == 'axolotl-ai-cloud'
+    if: ${{ github.repository_owner == 'axolotl-ai-cloud' && (github.event_name != 'pull_request' || !github.event.pull_request.draft) }}
     timeout-minutes: 480
     runs-on: ubuntu-latest-m
     strategy:
@@ -122,6 +129,13 @@ jobs:
             pytorch: 2.6.0
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-uv-base"
+          - cuda: "126"
+            cuda_version: 12.6.3
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.7.1
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
           - cuda: "128"
             cuda_version: 12.8.1
             cudnn_version: ""
@@ -129,6 +143,13 @@ jobs:
             pytorch: 2.7.1
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-uv-base"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.8.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
     steps:
       - name: Checkout
         uses: actions/checkout@v4

.github/workflows/lint.yml

@@ -3,6 +3,7 @@ on:
   # check on PRs, and manual triggers
   merge_group:
   pull_request:
+      types: [opened, synchronize, reopened, ready_for_review]
       paths:
        - '**.py'
        - 'requirements.txt'
@@ -16,6 +17,7 @@ jobs:
   pre-commit:
     name: pre-commit
     runs-on: ubuntu-latest
+    if: ${{ !github.event.pull_request.draft }}
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v5

.github/workflows/main.yml

@@ -24,12 +24,13 @@ jobs:
             cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.7.0
-            axolotl_extras: vllm
+            axolotl_extras:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.7.1
-            axolotl_extras:
+            axolotl_extras: vllm
+            is_latest: true
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
@@ -87,7 +88,6 @@ jobs:
             python_version: "3.11"
             pytorch: 2.6.0
             axolotl_extras:
-            is_latest: true
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
@@ -98,6 +98,7 @@ jobs:
             python_version: "3.11"
             pytorch: 2.7.1
             axolotl_extras:
+            is_latest: true
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"

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

@@ -21,7 +21,7 @@ concurrency:
 
 jobs:
   test-axolotl-multigpu:
-    if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' }}
+    if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' && (github.event_name != 'pull_request' || !github.event.pull_request.draft) }}
     strategy:
       fail-fast: false
       matrix:
@@ -37,14 +37,14 @@ jobs:
             cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.7.0
-            axolotl_extras: vllm
+            axolotl_extras:
             num_gpus: 2
             nightly_build: "true"
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.7.1
-            axolotl_extras:
+            axolotl_extras: vllm
             num_gpus: 2
             nightly_build: "true"
     runs-on: [self-hosted, modal]

.github/workflows/preview-docs.yml

@@ -2,7 +2,7 @@ name: Preview
 on:
   workflow_dispatch:
   pull_request:
-    types: [opened, synchronize, reopened]
+    types: [opened, synchronize, reopened, ready_for_review]
 
     # Run the workflow only when one of these files changes
     paths:
@@ -25,6 +25,7 @@ permissions:
 jobs:
   preview:
     runs-on: ubuntu-latest
+    if: ${{ !github.event.pull_request.draft }}
     steps:
       - name: Check out repository
         uses: actions/checkout@v4
@@ -52,6 +53,7 @@ jobs:
 
       - name: Netlify Publish
         uses: nwtgck/actions-netlify@v3.0
+        if: ${{ github.event.pull_request.head.repo.full_name == github.repository }}
         id: netlify
         with:
           publish-dir: './_site'

.github/workflows/tests-nightly.yml

@@ -52,7 +52,7 @@ jobs:
 
       - name: Install PyTorch
         run: |
-          pip3 install torch==${{ matrix.pytorch_version }}
+          pip3 install torch==${{ matrix.pytorch_version }} torchvision
 
       - name: Update requirements.txt
         run: |
@@ -106,6 +106,13 @@ jobs:
             num_gpus: 1
             axolotl_extras:
             nightly_build: "true"
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            num_gpus: 1
+            axolotl_extras:
+            nightly_build: "true"
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -130,3 +137,45 @@ jobs:
       - name: Run tests job on Modal
         run: |
           modal run cicd.e2e_tests
+  docker-e2e-multigpu-tests:
+    if: github.repository_owner == 'axolotl-ai-cloud'
+    # this job needs to be run on self-hosted GPU runners...
+    runs-on: [self-hosted, modal]
+    timeout-minutes: 120
+    needs: [pre-commit, pytest, docker-e2e-tests]
+
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.7.1
+            num_gpus: 2
+            axolotl_extras:
+            nightly_build: "true"
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+      - name: Install Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.11"
+      - name: Install Modal
+        run: |
+          python -m pip install --upgrade pip
+          pip install modal==1.0.2 jinja2
+      - name: Update env vars
+        run: |
+          echo "BASE_TAG=main-base-py${{ matrix.python_version }}-cu${{ matrix.cuda }}-${{ matrix.pytorch }}" >> $GITHUB_ENV
+          echo "PYTORCH_VERSION=${{ matrix.pytorch}}" >> $GITHUB_ENV
+          echo "AXOLOTL_ARGS=${{ matrix.axolotl_args}}" >> $GITHUB_ENV
+          echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
+          echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
+          echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+          echo "NIGHTLY_BUILD=${{ matrix.nightly_build }}" >> $GITHUB_ENV
+          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
+      - name: Run tests job on Modal
+        run: |
+          modal run cicd.multigpu

.github/workflows/tests.yml

@@ -13,6 +13,7 @@ on:
       - 'cicd/cicd.sh'
       - 'cicd/Dockerfile.jinja'
   pull_request:
+      types: [opened, synchronize, reopened, ready_for_review]
       paths:
        - '**.py'
        - 'requirements.txt'
@@ -34,6 +35,7 @@ jobs:
   pre-commit:
     name: pre-commit
     runs-on: ubuntu-latest
+    if: ${{ !github.event.pull_request.draft }}
     steps:
       - uses: actions/checkout@v4
       - uses: actions/setup-python@v5
@@ -47,6 +49,7 @@ jobs:
   pytest:
     name: PyTest
     runs-on: ubuntu-latest
+    if: ${{ !github.event.pull_request.draft }}
 #    needs: [preload-cache]
     strategy:
       fail-fast: false
@@ -78,7 +81,7 @@ jobs:
 
       - name: Install PyTorch
         run: |
-          pip3 install torch==${{ matrix.pytorch_version }}
+          pip3 install torch==${{ matrix.pytorch_version }} torchvision
 
       - name: Install dependencies
         run: |
@@ -121,6 +124,7 @@ jobs:
   pytest-sdist:
     name: PyTest from Source Dist
     runs-on: ubuntu-latest
+    if: ${{ !github.event.pull_request.draft }}
     strategy:
       fail-fast: false
       matrix:
@@ -151,7 +155,7 @@ jobs:
 
       - name: Install PyTorch
         run: |
-          pip3 install torch==${{ matrix.pytorch_version }}
+          pip3 install torch==${{ matrix.pytorch_version }} torchvision
 
       - name: Install dependencies
         run: |
@@ -185,7 +189,7 @@ jobs:
 
   docker-e2e-tests-1st:
     # Run this job first as a gate for running the remainder of the test matrix
-    if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' }}
+    if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' && !github.event.pull_request.draft }}
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
@@ -235,7 +239,7 @@ jobs:
           modal run cicd.e2e_tests
 
   docker-e2e-tests:
-    if: github.repository_owner == 'axolotl-ai-cloud'
+    if: ${{ github.repository_owner == 'axolotl-ai-cloud' && !github.event.pull_request.draft }}
     # this job needs to be run on self-hosted GPU runners...
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
@@ -289,6 +293,7 @@ jobs:
     runs-on: [self-hosted, modal]
     timeout-minutes: 90
     needs: [docker-e2e-tests]
+    if: ${{ !github.event.pull_request.draft }}
 
     strategy:
       fail-fast: false

.pre-commit-config.yaml

@@ -27,7 +27,7 @@ repos:
     hooks:
     - id: pylint
 -   repo: https://github.com/pre-commit/mirrors-mypy
-    rev: v1.16.1
+    rev: v1.17.1
     hooks:
     - id: mypy
       additional_dependencies:

.runpod/README.md

@@ -119,14 +119,15 @@ datasets:
 
 ## Dataset Processing
 
-| Option                        | Default                    | Description                       |
-| ----------------------------- | -------------------------- | --------------------------------- |
-| `dataset_prepared_path`       | `"data/last_run_prepared"` | Path for prepared dataset         |
-| `push_dataset_to_hub`         | `""`                       | Push dataset to HF hub            |
-| `dataset_processes`           | `4`                        | Number of preprocessing processes |
-| `dataset_keep_in_memory`      | `false`                    | Keep dataset in memory            |
-| `shuffle_merged_datasets`     | `true`                     | Shuffle merged datasets           |
-| `dataset_exact_deduplication` | `true`                     | Deduplicate datasets              |
+| Option                            | Default                    | Description                         |
+| --------------------------------- | -------------------------- | ----------------------------------- |
+| `dataset_prepared_path`           | `"data/last_run_prepared"` | Path for prepared dataset           |
+| `push_dataset_to_hub`             | `""`                       | Push dataset to HF hub              |
+| `dataset_processes`               | `4`                        | Number of preprocessing processes   |
+| `dataset_keep_in_memory`          | `false`                    | Keep dataset in memory              |
+| `shuffle_merged_datasets`         | `true`                     | Shuffle merged datasets             |
+| `shuffle_before_merging_datasets` | `false`                    | Shuffle each dataset before merging |
+| `dataset_exact_deduplication`     | `true`                     | Deduplicate datasets                |
 
 ## LoRA Configuration
 
@@ -184,7 +185,6 @@ datasets:
 | `flash_attention`          | `false` | Use flash attention           |
 | `flash_attn_cross_entropy` | `false` | Flash attention cross entropy |
 | `flash_attn_rms_norm`      | `false` | Flash attention RMS norm      |
-| `flash_attn_fuse_qkv`      | `false` | Fuse QKV operations           |
 | `flash_attn_fuse_mlp`      | `false` | Fuse MLP operations           |
 | `sdp_attention`            | `false` | Use scaled dot product        |
 | `s2_attention`             | `false` | Use shifted sparse attention  |

.runpod/src/config/config.yaml

@@ -296,7 +296,6 @@
 # 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
 # # https://pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html
@@ -541,7 +540,6 @@ xformers_attention: ${XFORMERS_ATTENTION}
 flash_attention: ${FLASH_ATTENTION}
 flash_attn_cross_entropy: ${FLASH_ATTN_CROSS_ENTROPY}
 flash_attn_rms_norm: ${FLASH_ATTN_RMS_NORM}
-flash_attn_fuse_qkv: ${FLASH_ATTN_FUSE_QKV}
 flash_attn_fuse_mlp: ${FLASH_ATTN_FUSE_MLP}
 sdp_attention: ${SDP_ATTENTION}
 s2_attention: ${S2_ATTENTION}

README.md

@@ -25,15 +25,28 @@
 
 ## 🎉 Latest Updates
 
-- 2025/06: Magistral with mistral-common tokenizer support has been added to Axolotl. See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/magistral) to start training your own Magistral models with Axolotl!
+- 2025/07:
+  - ND Parallelism support has been added into Axolotl. Compose Context Parallelism (CP), Tensor Parallelism (TP), and Fully Sharded Data Parallelism (FSDP) within a single node and across multiple nodes. Check out the [blog post](https://huggingface.co/blog/accelerate-nd-parallel) for more info.
+  - Axolotl adds more models: [GPT-OSS](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/gpt-oss), [Gemma 3n](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/gemma3n), [Liquid Foundation Model 2 (LFM2)](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/lfm2), and [Arcee Foundation Models (AFM)](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/afm).
+  - FP8 finetuning with fp8 gather op is now possible in Axolotl via `torchao`. Get started [here](https://docs.axolotl.ai/docs/mixed_precision.html#sec-fp8)!
+  - [Voxtral](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/voxtral), [Magistral 1.1](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/magistral), and [Devstral](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/devstral) with mistral-common tokenizer support has been integrated in Axolotl!
+  - TiledMLP support for single-GPU to multi-GPU training with DDP, DeepSpeed and FSDP support has been added to support Arctic Long Sequence Training. (ALST). See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/alst) for using ALST with Axolotl!
 - 2025/05: Quantization Aware Training (QAT) support has been added to Axolotl. Explore the [docs](https://docs.axolotl.ai/docs/qat.html) to learn more!
-- 2025/04: Llama 4 support has been added in Axolotl. See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/llama-4) to start training your own Llama 4 models with Axolotl's linearized version!
 - 2025/03: Axolotl has implemented Sequence Parallelism (SP) support. Read the [blog](https://huggingface.co/blog/axolotl-ai-co/long-context-with-sequence-parallelism-in-axolotl) and [docs](https://docs.axolotl.ai/docs/sequence_parallelism.html) to learn how to scale your context length when fine-tuning.
+
+<details>
+
+<summary>Expand older updates</summary>
+
+- 2025/06: Magistral with mistral-common tokenizer support has been added to Axolotl. See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/magistral) to start training your own Magistral models with Axolotl!
+- 2025/04: Llama 4 support has been added in Axolotl. See [examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/llama-4) to start training your own Llama 4 models with Axolotl's linearized version!
 - 2025/03: (Beta) Fine-tuning Multimodal models is now supported in Axolotl. Check out the [docs](https://docs.axolotl.ai/docs/multimodal.html) to fine-tune your own!
 - 2025/02: Axolotl has added LoRA optimizations to reduce memory usage and improve training speed for LoRA and QLoRA in single GPU and multi-GPU training (DDP and DeepSpeed). Jump into the [docs](https://docs.axolotl.ai/docs/lora_optims.html) to give it a try.
 - 2025/02: Axolotl has added GRPO support. Dive into our [blog](https://huggingface.co/blog/axolotl-ai-co/training-llms-w-interpreter-feedback-wasm) and [GRPO example](https://github.com/axolotl-ai-cloud/grpo_code) and have some fun!
 - 2025/01: Axolotl has added Reward Modelling / Process Reward Modelling fine-tuning support. See [docs](https://docs.axolotl.ai/docs/reward_modelling.html).
 
+</details>
+
 ## ✨ Overview
 
 Axolotl is a tool designed to streamline post-training for various AI models.
@@ -79,6 +92,20 @@ docker run --gpus '"all"' --rm -it axolotlai/axolotl:main-latest
 
 Other installation approaches are described [here](https://docs.axolotl.ai/docs/installation.html).
 
+#### Cloud Providers
+
+<details>
+
+- [RunPod](https://runpod.io/gsc?template=v2ickqhz9s&ref=6i7fkpdz)
+- [Vast.ai](https://cloud.vast.ai?ref_id=62897&template_id=bdd4a49fa8bce926defc99471864cace&utm_source=github&utm_medium=developer_community&utm_campaign=template_launch_axolotl&utm_content=readme)
+- [PRIME Intellect](https://app.primeintellect.ai/dashboard/create-cluster?image=axolotl&location=Cheapest&security=Cheapest&show_spot=true)
+- [Modal](https://www.modal.com?utm_source=github&utm_medium=github&utm_campaign=axolotl)
+- [Novita](https://novita.ai/gpus-console?templateId=311)
+- [JarvisLabs.ai](https://jarvislabs.ai/templates/axolotl)
+- [Latitude.sh](https://latitude.sh/blueprint/989e0e79-3bf6-41ea-a46b-1f246e309d5c)
+
+</details>
+
 ### Your First Fine-tune
 
 ```bash
@@ -120,12 +147,6 @@ Contributions are welcome! Please see our [Contributing Guide](https://github.co
 
 ## ❤️ Sponsors
 
-Thank you to our sponsors who help make Axolotl possible:
-
-- [Modal](https://www.modal.com?utm_source=github&utm_medium=github&utm_campaign=axolotl) - Modal lets you run
-jobs in the cloud, by just writing a few lines of Python. Customers use Modal to deploy Gen AI models at large scale,
-fine-tune large language models, run protein folding simulations, and much more.
-
 Interested in sponsoring? Contact us at [wing@axolotl.ai](mailto:wing@axolotl.ai)
 
 ## 📜 License

_quarto.yml

@@ -35,25 +35,30 @@ quartodoc:
         - cli.train
         - cli.evaluate
         - cli.args
+        - cli.art
         - cli.checks
         - cli.config
+        - cli.delinearize_llama4
         - cli.inference
         - cli.merge_lora
         - cli.merge_sharded_fsdp_weights
         - cli.preprocess
-        - cli.sweeps
-        - cli.utils
+        - cli.quantize
         - cli.vllm_serve
         - cli.cloud.base
         - cli.cloud.modal_
-        - cli.quantize
+        - cli.utils
+        - cli.utils.args
+        - cli.utils.fetch
+        - cli.utils.load
+        - cli.utils.sweeps
+        - cli.utils.train
     - title: Trainers
       desc: Training implementations
       contents:
         - core.trainers.base
         - core.trainers.trl
         - core.trainers.mamba
-        - core.trainers.relora
         - core.trainers.dpo.trainer
         - core.trainers.grpo.trainer
         - core.trainers.grpo.sampler
@@ -268,6 +273,8 @@ website:
             - docs/batch_vs_grad.qmd
             - docs/dataset_preprocessing.qmd
             - docs/multipack.qmd
+            - docs/mixed_precision.qmd
+            - docs/optimizers.qmd
 
         - section: "Advanced Features"
           contents:
@@ -277,6 +284,7 @@ website:
             - docs/custom_integrations.qmd
             - docs/sequence_parallelism.qmd
             - docs/gradient_checkpointing.qmd
+            - docs/nd_parallelism.qmd
 
         - section: "Troubleshooting"
           contents:

cicd/Dockerfile-uv.jinja

@@ -11,7 +11,7 @@ ENV NIGHTLY_BUILD="{{ NIGHTLY_BUILD }}"
 ENV HF_HOME="{{ HF_HOME }}"
 
 RUN apt-get update && \
-    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev
+    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm
 
 WORKDIR /workspace
 

cicd/Dockerfile.jinja

@@ -12,7 +12,7 @@ ENV HF_HOME="{{ HF_HOME }}"
 ENV AXOLOTL_DATASET_PROCESSES="8"
 
 RUN apt-get update && \
-    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev
+    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm
 
 WORKDIR /workspace
 

cicd/multigpu.sh

@@ -2,7 +2,7 @@
 set -e
 
 # Only run two tests at a time to avoid OOM on GPU (with coverage collection)
-pytest -v -n2 \
+pytest -v --durations=10 -n2 \
   --ignore=/workspace/axolotl/tests/e2e/multigpu/solo/ \
   --ignore=/workspace/axolotl/tests/e2e/multigpu/patched/ \
   /workspace/axolotl/tests/e2e/multigpu/ \
@@ -19,5 +19,7 @@ pytest -v  --durations=10 -n1 /workspace/axolotl/tests/e2e/multigpu/patched/ \
   --cov-append \
   --cov-report=xml:multigpu-coverage.xml
 
-# Upload coverage to Codecov
-codecov upload-process -t "${CODECOV_TOKEN}" -f multigpu-coverage.xml -F multigpu,docker-tests,pytorch-${PYTORCH_VERSION} || true
+# Upload coverage to Codecov if CODECOV_TOKEN is available
+if [ -n "$CODECOV_TOKEN" ]; then
+  codecov upload-process -t "${CODECOV_TOKEN}" -f multigpu-coverage.xml -F multigpu,docker-tests,pytorch-${PYTORCH_VERSION} || true
+fi

cicd/single_gpu.py

@@ -65,6 +65,9 @@ GPU_CONFIG = f"L40S:{N_GPUS}"
 def run_cmd(cmd: str, run_folder: str):
     import subprocess  # nosec
 
+    sp_env = os.environ.copy()
+    sp_env["AXOLOTL_DATASET_PROCESSES"] = "8"
+
     # Propagate errors from subprocess.
-    if exit_code := subprocess.call(cmd.split(), cwd=run_folder):  # nosec
+    if exit_code := subprocess.call(cmd.split(), cwd=run_folder, env=sp_env):  # nosec
         exit(exit_code)  # pylint: disable=consider-using-sys-exit

codecov.yml

@@ -22,6 +22,7 @@ coverage:
         only_pulls: true
         flags: null
         paths: null
+        informational: true
     patch:
       default:
         # basic

deepspeed_configs/zero3.json

@@ -7,9 +7,9 @@
     "reduce_bucket_size": "auto",
     "stage3_prefetch_bucket_size": "auto",
     "stage3_param_persistence_threshold": "auto",
-    "stage3_max_live_parameters": 0,
-    "stage3_max_reuse_distance": 0,
-    "stage3_gather_16bit_weights_on_model_save": true
+    "max_live_parameters": 0,
+    "max_reuse_distance": 0,
+    "gather_16bit_weights_on_model_save": true
   },
   "bf16": {
     "enabled": "auto"

deepspeed_configs/zero3_bf16.json

@@ -7,9 +7,9 @@
     "reduce_bucket_size": "auto",
     "stage3_prefetch_bucket_size": "auto",
     "stage3_param_persistence_threshold": "auto",
-    "stage3_max_live_parameters": 0,
-    "stage3_max_reuse_distance": 0,
-    "stage3_gather_16bit_weights_on_model_save": true
+    "max_live_parameters": 0,
+    "max_reuse_distance": 0,
+    "gather_16bit_weights_on_model_save": true
   },
   "bf16": {
     "enabled": true

deepspeed_configs/zero3_bf16_cpuoffload_all.json

@@ -17,9 +17,9 @@
     "reduce_bucket_size": "auto",
     "stage3_prefetch_bucket_size": "auto",
     "stage3_param_persistence_threshold": "auto",
-    "stage3_max_live_parameters": 0,
-    "stage3_max_reuse_distance": 0,
-    "stage3_gather_16bit_weights_on_model_save": true
+    "max_live_parameters": 0,
+    "max_reuse_distance": 0,
+    "gather_16bit_weights_on_model_save": true
   },
   "bf16": {
     "enabled": true

deepspeed_configs/zero3_bf16_cpuoffload_params.json

@@ -13,9 +13,9 @@
     "reduce_bucket_size": "auto",
     "stage3_prefetch_bucket_size": "auto",
     "stage3_param_persistence_threshold": "auto",
-    "stage3_max_live_parameters": 0,
-    "stage3_max_reuse_distance": 0,
-    "stage3_gather_16bit_weights_on_model_save": true
+    "max_live_parameters": 0,
+    "max_reuse_distance": 0,
+    "gather_16bit_weights_on_model_save": true
   },
   "bf16": {
     "enabled": true

docker/Dockerfile

@@ -10,7 +10,9 @@ ARG PYTORCH_VERSION="2.1.2"
 ENV PYTORCH_VERSION=$PYTORCH_VERSION
 
 RUN apt-get update && \
-    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev rsync s3fs
+    apt-get install -y --allow-change-held-packages vim curl nano libnccl2 libnccl-dev rsync s3fs && \
+    rm -rf /var/cache/apt/archives && \
+    rm -rf /var/lib/apt/lists/*
 
 WORKDIR /workspace
 
@@ -23,17 +25,17 @@ RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
         pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray] $AXOLOTL_ARGS; \
-    fi
+    fi && \
+    python scripts/unsloth_install.py | sh && \
+    python scripts/cutcrossentropy_install.py | sh && \
+    pip install pytest && \
+    pip cache purge
 
-RUN python scripts/unsloth_install.py | sh
-RUN python scripts/cutcrossentropy_install.py | sh
-
-# So we can test the Docker image
-RUN pip install pytest
-
-# fix so that git fetch/pull from remote works
+# fix so that git fetch/pull from remote works with shallow clone
 RUN git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \
-    git config --get remote.origin.fetch
+    git config --get remote.origin.fetch && \
+    git config --global credential.helper store
 
-# helper for huggingface-login cli
-RUN git config --global credential.helper store
+COPY .axolotl-complete.bash /root/.axolotl-complete.bash
+RUN chmod +x /root/.axolotl-complete.bash && \
+    echo 'source /root/.axolotl-complete.bash' >> ~/.bashrc

docker/Dockerfile-base

@@ -16,12 +16,19 @@ 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/* \
+    && apt-get install -y --no-install-recommends \
+        wget git build-essential ninja-build git-lfs libaio-dev pkg-config \
+        ibverbs-providers ibverbs-utils infiniband-diags  \
+        librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm \
+    && rm -rf /var/cache/apt/archives \
+    && 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 tos accept --override-channels --channel https://repo.anaconda.com/pkgs/main \
+    && conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r \
     && conda create -n "py${PYTHON_VERSION}" python="${PYTHON_VERSION}"
 
 ENV PATH="/root/miniconda3/envs/py${PYTHON_VERSION}/bin:${PATH}"
@@ -31,12 +38,14 @@ WORKDIR /workspace
 RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel && \
     python3 -m pip install --no-cache-dir -U torch==${PYTORCH_VERSION}+cu${CUDA} torchvision --extra-index-url https://download.pytorch.org/whl/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"
+    python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" && \
+    python3 -m pip cache purge
 
 RUN git lfs install --skip-repo && \
     pip3 install awscli && \
     # The base image ships with `pydantic==1.8.2` which is not working
-    pip3 install -U --no-cache-dir pydantic==1.10.10
+    pip3 install -U --no-cache-dir pydantic==1.10.10 && \
+    pip3 cache purge
 
 RUN if [ "$PYTORCH_VERSION" = "2.6.0" ] && [ "$CUDA" = "124" ] ; then \
         FLASH_ATTENTION_FORCE_BUILD="TRUE" pip3 install --no-build-isolation flash-attn==2.8.0.post2; \

docker/Dockerfile-base-nightly

@@ -22,18 +22,22 @@ RUN apt-get update \
     && mkdir /root/.conda \
     && bash Miniconda3-latest-Linux-x86_64.sh -b \
     && rm -f Miniconda3-latest-Linux-x86_64.sh \
+    && conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/main \
+    && conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r \
     && 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 && \
+RUN python3 -m pip install --upgrade pip && pip3 install -U packaging==23.2 setuptools==75.8.0 wheel && \
     python3 -m pip install --no-cache-dir -U torch --extra-index-url https://download.pytorch.org/whl/nightly/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"
+    python3 -m pip install --no-cache-dir "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main" && \
+    python3 -m pip cache purge
 
 RUN git lfs install --skip-repo && \
     pip3 install awscli && \
     # The base image ships with `pydantic==1.8.2` which is not working
-    pip3 install -U --no-cache-dir pydantic==1.10.10
+    pip3 install -U --no-cache-dir pydantic==1.10.10 && \
+    pip3 cache purge

docker/Dockerfile-cloud

@@ -14,7 +14,10 @@ COPY scripts/motd /etc/motd
 
 RUN pip install jupyterlab notebook ipywidgets && \
     jupyter lab clean
-RUN apt install --yes --no-install-recommends openssh-server tmux iproute2 nvtop && \
+RUN apt update && \
+    apt install --yes --no-install-recommends openssh-server tmux iproute2 nvtop && \
+    rm -rf /var/cache/apt/archives && \
+    rm -rf /var/lib/apt/lists/* && \
     mkdir -p ~/.ssh && \
     chmod 700 ~/.ssh && \
     printf "\n[[ -z \"\$TMUX\"  ]] && { tmux attach-session -t ssh_tmux || tmux new-session -s ssh_tmux; exit; }\n" >> ~/.bashrc && \

docker/Dockerfile-cloud-no-tmux

@@ -9,13 +9,15 @@ ENV HF_HUB_ENABLE_HF_TRANSFER="1"
 EXPOSE 8888
 EXPOSE 22
 
-COPY scripts/cloud-entrypoint-term.sh /root/cloud-entrypoint.sh
+COPY scripts/cloud-entrypoint.sh /root/cloud-entrypoint.sh
 COPY scripts/motd /etc/motd
 
 RUN pip install jupyterlab notebook ipywidgets && \
     jupyter lab clean
-RUN apt install --yes --no-install-recommends openssh-server tmux sudo && \
-    pip3 install -U --no-cache-dir grpcio ray[default]==2.9.3 && \
+RUN apt update && \
+    apt install --yes --no-install-recommends openssh-server tmux iproute2 nvtop ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm && \
+    rm -rf /var/cache/apt/archives && \
+    rm -rf /var/lib/apt/lists/* && \
     mkdir -p ~/.ssh && \
     chmod 700 ~/.ssh && \
     printf "[ ! -z \"\$TERM\" -a -r /etc/motd ] && cat /etc/motd\n" >> ~/.bashrc && \

docs/cli.qmd

@@ -23,6 +23,20 @@ axolotl <command> [config.yml] [options]
 
 The config file can be local or a URL to a raw YAML file.
 
+### Launcher Arguments
+
+For commands that support multi-GPU (`train`, `evaluate`, ...), you can pass launcher-specific arguments using the `--` separator:
+
+```bash
+# Pass torchrun arguments
+axolotl train config.yml --launcher torchrun -- --nproc_per_node=2 --nnodes=1
+
+# Pass accelerate arguments
+axolotl train config.yml --launcher accelerate -- --config_file=accelerate_config.yml --num_processes=4
+```
+
+Arguments after `--` are passed directly to the launcher (torchrun, accelerate launch, etc.).
+
 ## Command Reference
 
 ### fetch
@@ -80,7 +94,11 @@ axolotl train config.yml \
     --num-epochs 3
 
 # Training without accelerate
-axolotl train config.yml --no-accelerate
+axolotl train config.yml --launcher python
+
+# Pass launcher-specific arguments using -- separator
+axolotl train config.yml --launcher torchrun -- --nproc_per_node=2 --nnodes=1
+axolotl train config.yml --launcher accelerate -- --config_file=accelerate_config.yml
 
 # Resume training from checkpoint
 axolotl train config.yml --resume-from-checkpoint path/to/checkpoint
@@ -175,6 +193,9 @@ Evaluates a model's performance (loss etc) on the train and eval datasets.
 ```bash
 # Basic evaluation
 axolotl evaluate config.yml
+
+# Evaluation with launcher arguments
+axolotl evaluate config.yml --launcher torchrun -- --nproc_per_node=2
 ```
 
 ### lm-eval
@@ -287,9 +308,6 @@ axolotl preprocess config.yml --cloud cloud_config.yml
 # Train on cloud
 axolotl train config.yml --cloud cloud_config.yml
 
-# Train without accelerate on cloud
-axolotl train config.yml --cloud cloud_config.yml --no-accelerate
-
 # Run lm-eval on cloud
 axolotl lm-eval config.yml --cloud cloud_config.yml
 ```

docs/dataset-formats/conversation.qmd

@@ -212,10 +212,11 @@ Instead of passing `tools` via the system prompt, an alternative method would be
 Tools need to follow [JSON schema](https://json-schema.org/learn/getting-started-step-by-step).
 :::
 
+Example config for Llama4:
 ```yaml
 chat_template: llama4
 datasets:
-  - path: ...
+  - path: Nanobit/text-tools-2k-test
     type: chat_template
     # field_tools: tools # default is `tools`
 ```

docs/faq.qmd

@@ -136,3 +136,7 @@ description: Frequently asked questions
 >   dynamic: false
 >   mode: max-autotune-no-cudagraphs
 > ```
+
+**Q: `ValueError("Backward pass should have cleared tracker of all tensors")`
+
+> A: This may happen due to edge cases in using the modern OffloadActivations context manager for CUDA streams. If you encounter this error, you may have success using the naive implementation with `offload_activations: legacy` in your YAML.

docs/installation.qmd

@@ -124,10 +124,13 @@ For providers supporting Docker:
 
 - Use `axolotlai/axolotl-cloud:main-latest`
 - Available on:
-  - [Latitude.sh](https://latitude.sh/blueprint/989e0e79-3bf6-41ea-a46b-1f246e309d5c)
-  - [JarvisLabs.ai](https://jarvislabs.ai/templates/axolotl)
-  - [RunPod](https://runpod.io/gsc?template=v2ickqhz9s&ref=6i7fkpdz)
-  - [Novita](https://novita.ai/gpus-console?templateId=311)
+    - [RunPod](https://runpod.io/gsc?template=v2ickqhz9s&ref=6i7fkpdz)
+    - [Vast.ai](https://cloud.vast.ai?ref_id=62897&template_id=bdd4a49fa8bce926defc99471864cace&utm_source=axolotl&utm_medium=partner&utm_campaign=template_launch_july2025&utm_content=docs_link)
+    - [PRIME Intellect](https://app.primeintellect.ai/dashboard/create-cluster?image=axolotl&location=Cheapest&security=Cheapest&show_spot=true)
+    - [Modal](https://www.modal.com?utm_source=github&utm_medium=github&utm_campaign=axolotl)
+    - [Novita](https://novita.ai/gpus-console?templateId=311)
+    - [JarvisLabs.ai](https://jarvislabs.ai/templates/axolotl)
+    - [Latitude.sh](https://latitude.sh/blueprint/989e0e79-3bf6-41ea-a46b-1f246e309d5c)
 
 ### Google Colab {#sec-colab}
 

docs/mixed_precision.qmd

@@ -0,0 +1,149 @@
+---
+title: "Mixed Precision Training"
+format:
+  html:
+    toc: true
+    toc-depth: 3
+    number-sections: true
+    code-tools: true
+execute:
+  enabled: false
+---
+
+Mixed precision training uses lower precision data types to reduce memory usage and increase training speed while maintaining model quality. Axolotl supports several mixed precision formats:
+
+- **FP16** - Half precision 16-bit (Pascal generation+)
+- **BF16** - Brain Float 16-bit (Ampere generation+)
+- **FP8** - 8-bit floating point (Hopper generation+)
+
+## FP16 Mixed Precision {#sec-fp16}
+
+### Overview {#sec-fp16-overview}
+
+FP16 is the traditional half-precision format, supported on older GPUs but can be less numerically stable than BF16.
+
+### Configuration {#sec-fp16-config}
+
+```{.yaml}
+fp16: true
+```
+
+### FP16 Considerations {#sec-fp16-considerations}
+
+- May require gradient scaling to prevent underflow
+- Less numerically stable than BF16
+- Can cause training instability with some model architectures
+- Consider using BF16 if your hardware supports it
+
+## BF16 Mixed Precision {#sec-bf16}
+
+### Overview {#sec-bf16-overview}
+
+BF16 (Brain Float 16) offers better numerical stability than FP16 and is the recommended mixed precision format for modern GPUs. It provides the same dynamic range as FP32 while using half the memory.
+
+### Configuration {#sec-bf16-config}
+
+```{.yaml}
+# Automatic BF16 detection (recommended)
+bf16: auto
+
+# Or explicitly enable
+bf16: true
+
+# For evaluation with BF16
+bf16: full  # Equivalent to bf16_full_eval in the HF trainer
+```
+
+## FP8 Mixed Precision {#sec-fp8}
+
+::: {.callout-note}
+FP8 support is experimental and requires compatible hardware (H100, H200) and recent PyTorch versions with TorchAO.
+:::
+
+### What is FP8? {#sec-fp8-overview}
+
+FP8 (8-bit floating point) can provide significant time savings compared to FP16/BF16 while maintaining training stability. Axolotl's implementation uses PyTorch's TorchAO library with "tensorwise" scaling strategy.
+
+### Requirements {#sec-fp8-software}
+
+- Hopper+ GPUs (H100/H200)
+- PyTorch 2.7+ (+ compatible TorchAO version)
+- CUDA 12.4+
+
+### Configuration {#sec-fp8-config}
+
+Add to your YAML config:
+
+```{.yaml}
+# Enable FP8 mixed precision
+fp8: true
+
+# Optional: Enable FP8 for FSDP all-gather operations
+fp8_enable_fsdp_float8_all_gather: true
+
+# Enable torch.compile (almost always necessary for FP8 speedups)
+torch_compile: true
+```
+
+::: {.callout-important}
+**torch.compile is critical for FP8 performance**
+
+FP8 training requires `torch_compile: true` to see meaningful speedups. Without compilation, FP8 may actually be slower and use more memory than FP16/BF16.
+:::
+
+### Advanced FP8 Configs {#sec-fp8-advanced}
+
+For [FSDP](multi-gpu.qmd#sec-fsdp) (Fully Sharded Data Parallel) training:
+
+```{.yaml}
+fp8: true
+fp8_enable_fsdp_float8_all_gather: true
+
+torch_compile: true
+
+# FSDP configuration
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: true
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  reshard_after_forward: true
+```
+
+## Best Practices {#sec-best-practices}
+
+### Choosing Precision Format {#sec-choosing-format}
+
+- **Start with automatic detection**: `bf16: auto`
+- **For Hopper+ (H100/H200)**: Try FP8 + torch.compile for maximum speed
+- **For Ampere (A100/RTX 30/40)**: Use BF16
+- **For older Pascal/Turing GPUs**: Use FP16 with caution
+- **For very old or unsupported GPUs**: Use FP32
+
+### Validation and Testing {#sec-validation}
+
+Always validate your mixed precision setup:
+
+- **Start with a small dataset** to verify stability
+- **Monitor loss curves** for irregularities
+- **Compare with FP32 baseline** when possible
+- **Test evaluation metrics** match expectations
+
+### FP8 Particulars {#sec-fp8-details}
+
+- Use cases
+  - Single GPU training
+  - Multi GPU training with FSDP2 or Deepspeed
+- Speedups
+  - Please refer to the [TorchAO FP8 training benchmarks](https://github.com/pytorch/ao/tree/main/torchao/float8#rowwise-scaling) for expected matmul speedups for different (M, K, N) settings
+  - Concrete number for LLaMA 3 8B training can be found [here](https://github.com/pytorch/ao/tree/main/torchao/float8#training-benchmarks)
+- Known issues:
+  - FP8 + DDP + `torch.compile` (causes [error](https://gist.github.com/djsaunde/0c1664c32e44a64d31b5e01b4aafe5c4))
+  - FP8 + FSDP2 + `torch.compile` + FSDP2 activation checkpointing tends to be _slower_ than the BF16 equivalent training
+  - Flash Attention 2 does not play nicely with `torch.compile`
+
+See `examples/llama-3/3b-fp8-fsdp2.yaml` for an optimized example config. Enabling FP8 mixed precision + FP8 all-gather training results in ~10% faster iterations per second vs. BF16 for a relatively small (3B param) model
+
+For more information on multi-GPU training, see our [Multi-GPU guide](multi-gpu.qmd).

docs/multi-gpu.qmd

@@ -98,8 +98,8 @@ fsdp_cpu_ram_efficient_loading | cpu_ram_efficient_loading
 fsdp_state_dict_type | state_dict_type
 fsdp_use_orig_params | **REMOVED**
 
-
-For example, if you were using the following FSDP1 config:
+For more details, please see the migration guide in the [torchtitan repo](https://github.com/pytorch/torchtitan/blob/main/docs/fsdp.md). In Axolotl,
+if you were using the following FSDP1 config:
 
 ```{.yaml}
 fsdp_version: 1

docs/multi-node.qmd

@@ -69,11 +69,19 @@ export NCCL_BUFFSIZE=2097152
 
 Run the following on each node:
 
+### Option 1: New Axolotl CLI with launcher args (Recommended)
+
+```bash
+axolotl train config.yaml --launcher torchrun -- --nnodes $num_nodes --nproc_per_node $gpu_per_node --rdzv_id $rdzv_id --rdzv_backend c10d --rdzv_endpoint "$head_node_ip:$head_node_port"
+```
+
+### Option 2: Direct torchrun (Legacy)
+
 ```bash
 torchrun --nnodes $num_nodes --nproc_per_node $gpu_per_node --rdzv_id $rdzv_id --rdzv_backend c10d --rdzv_endpoint "$head_node_ip:$head_node_port" -m axolotl.cli.train config.yaml
 ```
 
-Please make sure to substitute the placeholder variables.
+Please make sure to substitute the placeholder variables:
 
 - `num_nodes`: Number of nodes (containing GPUs)
 - `gpu_per_node`: Number of gpus per node
@@ -81,8 +89,6 @@ Please make sure to substitute the placeholder variables.
 - `head_node_port`: Port of the head node (make sure other machines can connect to this. Default 29400)
 - `rdzv_id`: A unique job ID that is used by the job across nodes.
 
-::: {.callout-note}
-You need to call `axolotl.cli.train` instead of `axolotl train` as the latter calls accelerate under the hood
-:::
+The new CLI approach (Option 1) is recommended as it provides consistent argument handling and works seamlessly with other Axolotl CLI features.
 
 More info on the available configs can be found on the Pytorch docs [here](https://pytorch.org/docs/stable/elastic/run.html)

docs/multimodal.qmd

@@ -14,6 +14,7 @@ format:
 - [Llava-1.5](#sec-llava-15)
 - [Mistral-Small-3.1](#sec-mistral-small-31)
 - [Gemma-3](#sec-gemma-3)
+- [Gemma-3n](#sec-gemma-3n)
 - [Qwen2-VL](#sec-qwen2-vl)
 - [Qwen2.5-VL](#sec-qwen25-vl)
 
@@ -110,6 +111,22 @@ base_model: google/gemma-3-4b-it
 chat_template: gemma3
 ```
 
+### Gemma-3n {#sec-gemma-3n}
+
+::: {.callout-warning}
+The model's initial loss and grad norm will be very high. We suspect this to be due to the Conv in the vision layers.
+:::
+
+::: {.callout-tip}
+Please make sure to install `timm` via `pip3 install timm==1.0.17`
+:::
+
+```yaml
+base_model: google/gemma-3n-E2B-it
+
+chat_template: gemma3n
+```
+
 ### Qwen2-VL {#sec-qwen2-vl}
 
 ```yaml
@@ -132,7 +149,9 @@ For multi-modal datasets, we adopt an extended `chat_template` format similar to
 
 - A message is a list of `role` and `content`.
 - `role` can be `system`, `user`, `assistant`, etc.
-- `content` is a list of `type` and (`text` or `image` or `path` or `url` or `base64`).
+- `content` is a list of `type` and (`text`, `image`, `path`, `url`, `base64`, or `audio`).
+
+### Image
 
 ::: {.callout-note}
 For backwards compatibility:
@@ -141,15 +160,29 @@ For backwards compatibility:
 - If `content` is a string, it will be converted to a list with `type` as `text`.
 :::
 
-::: {.callout-tip}
 For image loading, you can use the following keys within `content` alongside `"type": "image"`:
 
 - `"path": "/path/to/image.jpg"`
 - `"url": "https://example.com/image.jpg"`
 - `"base64": "..."`
 - `"image": PIL.Image`
+
+### Audio
+
+For audio loading, you can use the following keys within `content` alongside `"type": "audio"`:
+
+- `"path": "/path/to/audio.mp3"`
+- `"url": "https://example.com/audio.mp3"`
+- `"audio": np.ndarray`
+
+::: {.callout-tip}
+
+You may need to install `librosa` via `pip3 install librosa==0.11.0`.
+
 :::
 
+### Example
+
 Here is an example of a multi-modal dataset:
 ```json
 [
@@ -178,3 +211,9 @@ Here is an example of a multi-modal dataset:
   }
 ]
 ```
+
+## FAQ
+
+1. `PIL.UnidentifiedImageError: cannot identify image file ...`
+
+`PIL` could not retrieve the file at `url` using `requests`. Please check for typo. One alternative reason is that the request is blocked by the server.

docs/nd_parallelism.qmd

@@ -0,0 +1,108 @@
+---
+title: "N-D Parallelism (Beta)"
+---
+
+Axolotl enables training models at scale by composing different parallelism techniques. This is essential when:
+
+- A model's weights are too large to fit on a single GPU's memory.
+- A model's activations, especially with very long contexts, are too large for a single GPU.
+- You want to accelerate training by using multiple GPUs or nodes.
+
+or combinations of the above!
+
+## Core Concepts
+
+Parallelism strategies can be combined. The key is understanding how each one divides the workload. PyTorch's `DeviceMesh` is the modern way to manage these combinations, creating a logical grid of your GPUs and assigning different parallel strategies to different dimensions of the grid.
+
+### Data Parallelism {#sec-dp}
+
+Data Parallelism focuses on splitting the global data batch across GPUs.
+
+- Distributed Data Parallel (DDP): The classic approach. The full model is replicated on every GPU. Each GPU processes a different slice of the data batch. Gradients are then averaged across all GPUs after the backward pass to keep the models synchronized. This can substantially improve data throughput compared to single-device training, but requires that each GPU is able to hold the entire model, its gradients, and optimizer states.
+
+- [Fully Sharded Data Parallel (FSDP)](multi-gpu.qmd#fully-sharded-data-parallel-(fsdp)): A highly memory-efficient form of data parallelism (inspired by DeepSpeed's ZeRO). Instead of replicating the model, FSDP shards the model's *parameters, gradients, and optimizer states* across the GPUs in the data-parallel group. During computation, each GPU receives the specific parameters it needs via an `all_gather` operation just before they are used, and they can be discarded immediately after (`reshard-after-forward`).
+    - FSDP maps to ZeRO stages:
+        - ZeRO-2 (`reshard_after_forward=False`): Shards gradients and optimizer states. Model weights are replicated on each GPU.
+        - ZeRO-3 (`reshard_after_forward=True`): Shards gradients, optimizer states, AND model parameters. This provides the most memory savings at the cost of more communication (re-gathering parameters for both forward and backward passes).
+
+### [Experimental] Tensor Parallelism (TP) {#sec-tp}
+
+Also known as "horizontal model parallelism," as described in the [Megatron-LM paper](https://arxiv.org/pdf/1909.08053.pdf). Instead of splitting the batch, TP splits the model's layers themselves across GPUs.
+
+- How it works: For a linear layer `Y = XA`, the weight matrix `A` is split column-wise (`A = [A_1, A_2]`). The computation becomes `Y_1 = XA_1` and `Y_2 = XA_2`, which can happen in parallel on different GPUs. The final output `Y` is simply the concatenation of `Y_1` and `Y_2`. Check [this comment](https://github.com/huggingface/transformers/issues/10321#issuecomment-783543530) for more detailed info.
+- Requirement: TP involves frequent, small communications within a forward/backward pass. It requires a very fast interconnect between GPUs (e.g., NVLink) and is typically not recommended across different nodes.
+
+### Context Parallelism (CP) {#sec-cp}
+
+Context Parallelism, also called [Sequence Parallelism](sequence_parallelism.qmd), addresses the memory bottleneck from long sequences. The input sequence itself is split along the sequence length dimension and distributed across GPUs.
+
+- How it works: If you have a sequence of 8192 tokens and a `context_parallel_size` of 4, each GPU will only handle a chunk of 2048 tokens.
+- The Challenge: Attention is not local; every token needs to "attend to" every other token. Splitting the sequence breaks this.
+- The Solution (`ring-flash-attention`): An efficient communication protocol is used. To compute attention for its local sequence chunk, each GPU passes its Key-Value (KV) cache to its neighbor in a "ring." After `N-1` steps, every GPU has seen the KV-cache from all other GPUs, allowing it to compute the correct attention values for its chunk. This is implemented using the highly optimized `flash-attention` kernel at each step.
+
+### Hybrid Sharding Data Parallel (HSDP) {#sec-hsdp}
+
+HSDP is a 2D strategy that intelligently combines FSDP and DDP, typically for multi-node training.
+
+- Intra-Node (within a machine): Use FSDP. This is efficient because GPUs on the same node have fast interconnects (NVLink), making the `all_gather` operations for sharded parameters fast.
+- Inter-Node (across machines): Use DDP. The gradient synchronization between nodes is less frequent than FSDP's parameter gathering, making it a better fit for the slower node-to-node network (e.g., Ethernet/Infiniband).
+- Example: With 2 nodes of 8 GPUs each (16 total), you could have `dp_shard_size=8` (FSDP within each node) and `dp_replicate_size=2` (DDP across the two nodes).
+
+## Usage
+
+```yaml
+# FSDP config. See https://docs.axolotl.ai/docs/multi-gpu.html#sec-fsdp
+fsdp_version: 2
+fsdp_config:
+  # ...
+
+# The number of GPUs to shard the model parameters across (FSDP dimension).
+dp_shard_size: 4
+
+# The number of times to replicate the sharded model (DDP dimension).
+dp_replicate_size: 2
+
+# Number of GPUs for Tensor Parallelism.
+tensor_parallel_size: 1  # (default is 1, no TP)
+
+# Number of GPUs for Context/Sequence Parallelism.
+context_parallel_size: 1 # (default is 1, no CP)
+```
+
+Note: We recommend FSDP. DeepSpeed is only compatible with `tensor_parallel_size`.
+
+## Examples
+
+::: {.callout-tip}
+See our example configs [here](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/distributed-parallel).
+:::
+
+1.  HSDP on 2 nodes with 4 GPUs each (8 GPUs total):
+    - You want FSDP within each node and DDP across nodes.
+    - Set `dp_shard_size: 4` and `dp_replicate_size: 2`.
+
+2.  FSDP + TP on a single 8-GPU node:
+    - You want to split the model across 4 GPUs using FSDP, and further split each layer across 2 GPUs with TP.
+    - Set `dp_shard_size: 4` and `tensor_parallel_size: 2`.
+
+3.  FSDP + CP on a single 8-GPU node for long context:
+    - You want to shard the model across all 8 GPUs and also split the sequence length across all 8 GPUs.
+    - Set `dp_shard_size: 8` and `context_parallel_size: 8`. Note: this means the data parallel group and context parallel group are the same. A more common setup might be to shard across a smaller group.
+
+## Support Matrix
+
+This matrix describes how different parallelism methods can be combined in Axolotl.
+
+| Combination | `dp_replicate_size` | `dp_shard_size` | `tp_size` | `cp_size` | Status & Notes |
+| --- | :---: | :---: |:---:|:---:|---|
+| **FSDP** (ZeRO-3) | 1 | >1 | 1 | 1 | ✅ Fully supported. Shards model across all GPUs. |
+| **HSDP** | >1 | >1 | 1 | 1 | ✅ Fully supported. FSDP intra-node, DDP inter-node. |
+| **FSDP + TP** | 1 | >1 | >1 | 1 | ✅ **2D Parallelism**. Shards the model across a `dp_shard` group, and TP-splits layers within the `tp` group. |
+| **HSDP + TP** | >1 | >1 | >1 | 1 | ✅ **3D Parallelism**. A powerful but complex combination. |
+| **FSDP + CP** | 1 | >1 | 1 | >1 | ✅ **2D Parallelism**. Combines FSDP with context parallelism. |
+| **FSDP + TP + CP**| 1 | >1 | >1| >1| ✅ **3D Parallelism**. Another advanced combination. |
+| DDP + TP/CP | >1 | 1 | >1 | >1 | ❌ **Not Supported**. The `ParallelismConfig` explicitly prevents this, as composing pure DDP with TP or CP is currently not supported. You should use FSDP + TP/CP instead (`dp_shard_size > 1`). |
+| Just TP / CP | 1 | 1 | >1 | >1 | ✅ Supported. Useful for inference or when the model fits on one GPU but context is too long. |
+
+- `tp_size` refers to `tensor_parallel_size`
+- `cp_size` refers to `context_parallel_size`

docs/optimizers.qmd

@@ -0,0 +1,129 @@
+---
+title: Optimizers
+description: Configuring optimizers
+---
+
+## Overview
+
+Axolotl supports all optimizers supported by [transformers OptimizerNames](https://github.com/huggingface/transformers/blob/51f94ea06d19a6308c61bbb4dc97c40aabd12bad/src/transformers/training_args.py#L142-L187)
+
+Here is a list of optimizers supported by transformers as of `v4.54.0`:
+
+- `adamw_torch`
+- `adamw_torch_fused`
+- `adamw_torch_xla`
+- `adamw_torch_npu_fused`
+- `adamw_apex_fused`
+- `adafactor`
+- `adamw_anyprecision`
+- `adamw_torch_4bit`
+- `adamw_torch_8bit`
+- `ademamix`
+- `sgd`
+- `adagrad`
+- `adamw_bnb_8bit`
+- `adamw_8bit`  # alias for adamw_bnb_8bit
+- `ademamix_8bit`
+- `lion_8bit`
+- `lion_32bit`
+- `paged_adamw_32bit`
+- `paged_adamw_8bit`
+- `paged_ademamix_32bit`
+- `paged_ademamix_8bit`
+- `paged_lion_32bit`
+- `paged_lion_8bit`
+- `rmsprop`
+- `rmsprop_bnb`
+- `rmsprop_bnb_8bit`
+- `rmsprop_bnb_32bit`
+- `galore_adamw`
+- `galore_adamw_8bit`
+- `galore_adafactor`
+- `galore_adamw_layerwise`
+- `galore_adamw_8bit_layerwise`
+- `galore_adafactor_layerwise`
+- `lomo`
+- `adalomo`
+- `grokadamw`
+- `schedule_free_radam`
+- `schedule_free_adamw`
+- `schedule_free_sgd`
+- `apollo_adamw`
+- `apollo_adamw_layerwise`
+- `stable_adamw`
+
+
+## Custom Optimizers
+
+Enable custom optimizers by passing a string to the `optimizer` argument. Each optimizer will receive beta and epsilon args, however, some may accept additional args which are detailed below.
+
+### optimi_adamw
+
+```yaml
+optimizer: optimi_adamw
+```
+
+### ao_adamw_4bit
+
+Deprecated: Please use `adamw_torch_4bit`.
+
+### ao_adamw_8bit
+
+Deprecated: Please use `adamw_torch_8bit`.
+
+### ao_adamw_fp8
+
+
+```yaml
+optimizer: ao_adamw_fp8
+```
+
+### adopt_adamw
+
+GitHub: [https://github.com/iShohei220/adopt](https://github.com/iShohei220/adopt)
+Paper: [https://arxiv.org/abs/2411.02853](https://arxiv.org/abs/2411.02853)
+
+```yaml
+optimizer: adopt_adamw
+```
+
+### came_pytorch
+
+GitHub: [https://github.com/yangluo7/CAME/tree/master](https://github.com/yangluo7/CAME/tree/master)
+Paper: [https://arxiv.org/abs/2307.02047](https://arxiv.org/abs/2307.02047)
+
+```yaml
+optimizer: came_pytorch
+
+# optional args (defaults below)
+adam_beta1: 0.9
+adam_beta2: 0.999
+adam_beta3: 0.9999
+adam_epsilon: 1e-30
+adam_epsilon2: 1e-16
+```
+
+### muon
+
+Blog: [https://kellerjordan.github.io/posts/muon/](https://kellerjordan.github.io/posts/muon/)
+Paper: [https://arxiv.org/abs/2502.16982v1](https://arxiv.org/abs/2502.16982v1)
+
+```yaml
+optimizer: muon
+```
+
+### dion
+
+Microsoft's Dion (DIstributed OrthoNormalization) optimizer is a scalable and communication-efficient
+orthonormalizing optimizer that uses low-rank approximations to reduce gradient communication.
+
+GitHub: [https://github.com/microsoft/dion](https://github.com/microsoft/dion)
+Paper: [https://arxiv.org/pdf/2504.05295](https://arxiv.org/pdf/2504.05295)
+Note: Implementation written for PyTorch 2.7+ for DTensor
+
+```yaml
+optimizer: dion
+dion_lr: 0.01
+dion_momentum: 0.95
+lr: 0.00001  # learning rate for embeddings and parameters that fallback to AdamW
+```

docs/sequence_parallelism.qmd

@@ -22,7 +22,7 @@ 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
+context_parallel_size: 4  # Split sequences across 4 GPUs
 # Optional; strides across the key dimension. Larger values use more memory but should make training faster.
 heads_k_stride: 1
 # Optional; one of "varlen_llama3" or "batch_ring". Defaults to
@@ -30,7 +30,7 @@ heads_k_stride: 1
 ring_attn_func:
 ```
 
-The `sequence_parallel_degree` should be a divisor of the total number of GPUs. For example:
+The `context_parallel_size` should be a divisor of the total number of GPUs. For example:
 
 - With 8 GPUs, valid values would be 2, 4, or 8
 - With 4 GPUs, valid values would be 2 or 4
@@ -66,7 +66,7 @@ sequence_len: 8192
 
 ...
 
-sequence_parallel_degree: 4  # Split each sequence into 4 parts, one per GPU
+context_parallel_size: 4  # Split each sequence into 4 parts, one per GPU
 # Optional; strides across the key dimension. Larger values use more memory but should make training faster.
 heads_k_stride: 1
 # Optional; one of "varlen_llama3" or "batch_ring". Defaults to
@@ -89,12 +89,12 @@ Sequence parallelism is compatible with Axolotl's sample packing functionality.
 
 ## Effect on Batch Size
 
-When using sequence parallelism, your effective global batch size is **divided** by the `sequence_parallel_degree`. This happens because:
+When using sequence parallelism, your effective global batch size is **divided** by the `context_parallel_size`. This happens because:
 
-- Each group of `sequence_parallel_degree` GPUs works on the same batch (just different parts of each sequence)
+- Each group of `context_parallel_size` GPUs works on the same batch (just different parts of each sequence)
 - The number of batches processed per step decreases
 
 For example:
 - With 8 GPUs and no sequence parallelism: 8 different batches processed per step
-- With 8 GPUs and `sequence_parallel_degree=4`: Only 2 different batches processed per step (each split across 4 GPUs)
+- With 8 GPUs and `context_parallel_size=4`: Only 2 different batches processed per step (each split across 4 GPUs)
 - If your per-GPU `micro_batch_size` is 2, the global batch size decreases from 16 to 4

examples/alst/README.md

@@ -0,0 +1,9 @@
+# Arctic Long Sequence Training (ALST)
+
+Artic Long Sequence Training (ALST) is a technique for training long context models using a variety of optimization
+techniques. It is a combination of:
+- TiledMLP: Leverage tiling over the sequence dimension on MLP layers to reduce memory usage
+- Tiled Loss: Using optimized loss functions like Liger-Kernel or Cut Cross Entropy to reduce memory usage
+- Activation Offloading: Offload activations to CPU RAM to reduce memory usage
+
+For more information, you can check out the ALST paper [here](https://www.arxiv.org/abs/2506.13996).

examples/alst/llama3-8b-deepspeed-alst.yaml

@@ -0,0 +1,53 @@
+base_model: meta-llama/Llama-3.1-8B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+datasets:
+  - path: togethercomputer/Long-Data-Collections
+    type: completion
+    field: text
+    data_files:
+      - pretrain/rp_sub.jsonl.zst
+  - path: princeton-nlp/TextbookChapters
+    type: completion
+    field: chapter
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 500_000
+min_sample_len: 200_000
+sample_packing: true
+
+tiled_mlp: true
+context_parallel_size: 8
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_8bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+activation_offloading: legacy
+
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 100
+saves_per_epoch: 1
+evals_per_epoch: 2
+weight_decay: 0.0
+special_tokens:
+  pad_token: <|end_of_text|>
+
+deepspeed: deepspeed_configs/zero3_bf16_cpuoffload_all.json
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/alst/llama3-8b-fsdp2-alst.yaml

@@ -0,0 +1,59 @@
+base_model: meta-llama/Llama-3.1-8B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+datasets:
+  - path: togethercomputer/Long-Data-Collections
+    type: completion
+    field: text
+    data_files:
+      - pretrain/rp_sub.jsonl.zst
+  - path: princeton-nlp/TextbookChapters
+    type: completion
+    field: chapter
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 500_000
+min_sample_len: 200_000
+sample_packing: true
+
+tiled_mlp: true
+context_parallel_size: 8
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_8bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+activation_offloading: legacy
+
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 100
+saves_per_epoch: 1
+evals_per_epoch: 2
+weight_decay: 0.0
+special_tokens:
+  pad_token: <|end_of_text|>
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false  # offloading is currently not compatible with SP + torchao optimizer
+  state_dict_type: SHARDED_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  reshard_after_forward: true
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/arcee/README.md

@@ -0,0 +1,53 @@
+# Finetune ArceeAI's AFM with Axolotl
+
+[Arcee Foundation Models (AFM)](https://huggingface.co/collections/arcee-ai/afm-45b-68823397c351603014963473) are a family of 4.5B parameter open weight models trained by Arcee.ai.
+
+This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+Thanks to the team at Arcee.ai for using Axolotl in supervised fine-tuning the AFM model.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as AFM is only on nightly or use our latest [Docker images](https://docs.axolotl.ai/docs/docker.html).
+
+    Here is an example of how to install from main for pip:
+
+```bash
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
+git clone https://github.com/axolotl-ai-cloud/axolotl.git
+cd axolotl
+
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation -e '.[flash-attn]'
+```
+
+2. Run the finetuning example:
+
+```bash
+axolotl train examples/arcee/afm-4.5b-qlora.yaml
+```
+
+This config uses about 7.8GiB VRAM.
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### TIPS
+
+- For inference, the official Arcee.ai team recommends `top_p: 0.95`, `temperature: 0.5`, `top_k: 50`, and `repeat_penalty: 1.1`.
+- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+
+## Optimization Guides
+
+- [Multi-GPU Training](https://docs.axolotl.ai/docs/multi-gpu.html)
+- [Multi-Node Training](https://docs.axolotl.ai/docs/multi-node.html)
+- [LoRA Optimizations](https://docs.axolotl.ai/docs/lora_optims.html)
+
+## Related Resources
+
+- [AFM Blog](https://docs.arcee.ai/arcee-foundation-models/introduction-to-arcee-foundation-models)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl Website](https://axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/arcee/afm-4.5b-qlora.yaml

@@ -0,0 +1,64 @@
+base_model: arcee-ai/AFM-4.5B
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - 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: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/archived/cerebras/btlm-ft.yml

@@ -66,7 +66,7 @@ flash_optimum:
 gptq_groupsize:
 gptq_model_v1:
 
-warmup_steps: 32
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 save_total_limit:

examples/archived/cerebras/qlora.yml

@@ -43,7 +43,7 @@ xformers_attention: true
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.1

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

@@ -17,7 +17,7 @@ output_dir: ./outputs/lora-out
 
 sequence_len: 4096
 sample_packing: true
-pad_to_sequence_len: true
+
 
 adapter: lora
 lora_model_dir:
@@ -47,7 +47,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

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

@@ -20,7 +20,7 @@ lora_model_dir:
 
 sequence_len: 4096
 sample_packing: true
-pad_to_sequence_len: true
+
 
 lora_r: 32
 lora_alpha: 16
@@ -48,7 +48,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

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

@@ -17,7 +17,7 @@ output_dir: ./outputs/lora-out
 
 sequence_len: 4096
 sample_packing: true
-pad_to_sequence_len: true
+
 
 adapter: lora
 lora_model_dir:
@@ -47,7 +47,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

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

@@ -20,7 +20,7 @@ lora_model_dir:
 
 sequence_len: 4096
 sample_packing: true
-pad_to_sequence_len: true
+
 
 lora_r: 32
 lora_alpha: 16
@@ -48,7 +48,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

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

@@ -17,7 +17,7 @@ output_dir: ./outputs/lora-out
 
 sequence_len: 4096
 sample_packing: true
-pad_to_sequence_len: true
+
 
 adapter: lora
 lora_model_dir:
@@ -47,7 +47,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

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

@@ -20,7 +20,7 @@ lora_model_dir:
 
 sequence_len: 4096
 sample_packing: true
-pad_to_sequence_len: true
+
 
 lora_r: 32
 lora_alpha: 16
@@ -48,7 +48,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/dbrx/16bit-lora.yaml

@@ -54,7 +54,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch:
 saves_per_epoch: 1
 

examples/archived/dbrx/8bit-lora.yaml

@@ -57,7 +57,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch:
 saves_per_epoch: 1
 

examples/archived/dbrx/fft-ds-zero3.yaml

@@ -41,7 +41,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch:
 saves_per_epoch: 1
 

examples/archived/deepcoder/deepcoder-14B-preview-lora.yml

@@ -21,7 +21,7 @@ output_dir: ./outputs/lora-out
 sequence_len: 4096
 sample_packing: true
 eval_sample_packing: false
-pad_to_sequence_len: true
+
 
 adapter: lora
 lora_model_dir:
@@ -51,7 +51,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 1
 saves_per_epoch: 1
 weight_decay: 0.0

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

@@ -47,7 +47,7 @@ xformers_attention: true
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 40
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

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

@@ -77,7 +77,7 @@ xformers_attention: true
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.000001

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

@@ -44,7 +44,7 @@ xformers_attention: true
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 40
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/gemma/qlora.yml

@@ -25,7 +25,7 @@ lora_target_linear: true
 sequence_len: 4096
 sample_packing: true
 eval_sample_packing: false
-pad_to_sequence_len: true
+
 
 wandb_project:
 wandb_entity:

examples/archived/gptj/qlora.yml

@@ -40,7 +40,7 @@ xformers_attention: true
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.1

examples/archived/jeopardy-bot/config.yml

@@ -41,7 +41,7 @@ xformers_attention: true
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 20
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.1

examples/archived/mpt-7b/config.yml

@@ -42,7 +42,7 @@ logging_steps: 5
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 20
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0001

examples/archived/openllama-3b/config.yml

@@ -42,7 +42,7 @@ logging_steps: 1
 flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 20
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.1

examples/archived/openllama-3b/lora.yml

@@ -50,7 +50,7 @@ logging_steps: 1
 flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 20
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.1

examples/archived/openllama-3b/qlora.yml

@@ -43,7 +43,7 @@ logging_steps: 1
 flash_attention: true
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 20
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.1

examples/archived/qwen/lora.yml

@@ -49,7 +49,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention:
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/qwen/qlora.yml

@@ -49,7 +49,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention:
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/qwen/qwen2-moe-lora.yaml

@@ -45,7 +45,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/qwen/qwen2-moe-qlora.yaml

@@ -48,7 +48,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/redpajama/config-3b.yml

@@ -43,7 +43,7 @@ logging_steps: 5
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 20
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0001

examples/archived/replit-3b/config-lora.yml

@@ -41,7 +41,7 @@ logging_steps: 1
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 20
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0

examples/archived/stablelm-2/1.6b/fft.yml

@@ -16,7 +16,7 @@ output_dir: ./outputs/out
 
 sequence_len: 4096
 sample_packing: true
-pad_to_sequence_len: true
+
 
 adapter:
 lora_model_dir:
@@ -47,10 +47,9 @@ logging_steps: 1
 flash_attention: true
 flash_attn_cross_entropy: false
 flash_attn_rms_norm: true
-flash_attn_fuse_qkv: false
 flash_attn_fuse_mlp: true
 
-warmup_steps: 100
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 

examples/archived/stablelm-2/1.6b/lora.yml

@@ -19,7 +19,7 @@ output_dir: ./outputs/lora-out
 
 sequence_len: 4096
 sample_packing: true
-pad_to_sequence_len: true
+
 
 adapter: lora
 lora_model_dir:
@@ -51,7 +51,7 @@ flash_attention: true
 flash_attn_cross_entropy: false
 flash_attn_rms_norm: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/starcoder2/qlora.yml

@@ -19,7 +19,7 @@ lora_model_dir:
 
 sequence_len: 8192
 sample_packing: true
-pad_to_sequence_len: true
+
 
 lora_r: 32
 lora_alpha: 16
@@ -48,7 +48,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 20
+warmup_ratio: 0.1
 evals_per_epoch: 4
 eval_steps:
 saves_per_epoch: 4

examples/archived/tiny-llama/lora-mps.yml

@@ -17,7 +17,7 @@ output_dir: ./outputs/lora-out
 
 sequence_len: 4096
 sample_packing: true
-pad_to_sequence_len: true
+
 eval_sample_packing: false
 
 adapter: lora
@@ -49,7 +49,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: false
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 0
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/tiny-llama/lora.yml

@@ -17,7 +17,7 @@ output_dir: ./outputs/lora-out
 sequence_len: 4096
 sample_packing: true
 eval_sample_packing: false
-pad_to_sequence_len: true
+
 
 adapter: lora
 lora_model_dir:
@@ -47,7 +47,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/tiny-llama/pretrain.yml

@@ -38,7 +38,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch:
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/tiny-llama/qlora.yml

@@ -21,7 +21,7 @@ lora_model_dir:
 sequence_len: 4096
 sample_packing: true
 eval_sample_packing: false
-pad_to_sequence_len: true
+
 
 lora_r: 32
 lora_alpha: 16
@@ -49,7 +49,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/xgen-7b/xgen-7b-8k-qlora.yml

@@ -75,7 +75,7 @@ xformers_attention: true
 flash_attention:
 gptq_groupsize:
 gptq_model_v1:
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 4
 saves_per_epoch: 1
 weight_decay: 0.0

examples/archived/yi-34B-chat/qlora.yml

@@ -20,7 +20,7 @@ special_tokens:
 datasets:
   - path: mhenrichsen/alpaca_2k_test
     type: alpaca
-warmup_steps: 10
+warmup_ratio: 0.1
 
 # Iterations
 num_epochs: 1

examples/cloud/modal.yaml

@@ -26,5 +26,3 @@ timeout: 86400
 # Preprocess specific configurations
 memory_preprocess: 32
 timeout_preprocess: 14400
-
-# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

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

@@ -27,7 +27,7 @@ lora_target_linear: true
 sequence_len: 2048
 sample_packing: true
 eval_sample_packing: false
-pad_to_sequence_len: true
+
 
 wandb_project:
 wandb_entity:

examples/colab-notebooks/colab-axolotl-example.ipynb

@@ -40,7 +40,7 @@
         "%%capture\n",
         "# This step can take ~5-10 minutes to install dependencies\n",
         "!pip install --no-build-isolation axolotl[flash-attn]>=0.9.1\n",
-        "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@50cef19\""
+        "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@0ee9ee8\""
       ]
     },
     {

examples/deepcogito/cogito-v1-preview-llama-3B-lora.yml

@@ -21,7 +21,7 @@ output_dir: ./outputs/lora-out
 sequence_len: 4096
 sample_packing: true
 eval_sample_packing: false
-pad_to_sequence_len: true
+
 
 adapter: lora
 lora_model_dir:
@@ -51,7 +51,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 1
 saves_per_epoch: 1
 weight_decay: 0.0

examples/deepcogito/cogito-v1-preview-qwen-14B-lora.yml

@@ -21,7 +21,7 @@ output_dir: ./outputs/lora-out
 sequence_len: 4096
 sample_packing: true
 eval_sample_packing: false
-pad_to_sequence_len: true
+
 
 adapter: lora
 lora_model_dir:
@@ -51,7 +51,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 10
+warmup_ratio: 0.1
 evals_per_epoch: 1
 saves_per_epoch: 1
 weight_decay: 0.0

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

@@ -12,7 +12,7 @@ output_dir: ./outputs/out
 
 sequence_len: 2048
 sample_packing: true
-pad_to_sequence_len: true
+
 
 wandb_project:
 wandb_entity:
@@ -37,7 +37,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 100
+warmup_ratio: 0.1
 evals_per_epoch: 2
 saves_per_epoch: 1
 weight_decay: 0.0

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

@@ -30,7 +30,7 @@ output_dir: ./outputs/out
 
 sequence_len: 4096
 sample_packing: true
-pad_to_sequence_len: true
+
 
 wandb_project:
 wandb_entity:
@@ -61,7 +61,7 @@ resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
 
-warmup_steps: 100
+warmup_ratio: 0.1
 evals_per_epoch: 2
 saves_per_epoch: 1
 weight_decay: 0.0

examples/devstral/README.md

@@ -10,17 +10,14 @@ Thanks to the team at MistralAI for giving us early access to prepare for this r
 
 ## Getting started
 
-1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as Devstral is only on nightly or use our latest [Docker images](https://docs.axolotl.ai/docs/docker.html).
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
 
-    Here is an example of how to install from main for pip:
+    Here is an example of how to install from pip:
 
 ```bash
-# Ensure you have Pytorch installed (Pytorch 2.6.0+)
-git clone https://github.com/axolotl-ai-cloud/axolotl.git
-cd axolotl
-
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
 pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
-pip3 install --no-build-isolation -e '.[flash-attn]'
+pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. Run the finetuning example:

examples/devstral/devstral-small-qlora.yml

@@ -25,7 +25,7 @@ lora_model_dir:
 
 sequence_len: 2048
 sample_packing: true
-pad_to_sequence_len: true
+
 
 lora_r: 32
 lora_alpha: 16

examples/distributed-parallel/README.md

@@ -0,0 +1,52 @@
+# ND Parallelism Examples
+
+This directory contains example configurations for training models using ND Parallelism in Axolotl. These examples demonstrate how to compose different parallelism strategies (FSDP, TP, CP, HSDP) for efficient multi-GPU training.
+
+## Quick Start
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
+
+2. Run the command below:
+
+```bash
+# Train Qwen3 8B with FSDP + TP + CP on a single 8-GPU node
+axolotl train examples/distributed-parallel/qwen3-8b-fsdp-tp-cp.yaml
+
+# Train Llama 3.1 8B with HSDP + TP on 2 nodes (16 GPUs total)
+axolotl train examples/distributed-parallel/llama-3_1-8b-hsdp-tp.yaml
+```
+
+## Example Configurations
+
+### Single Node (8 GPUs)
+
+**Qwen3 8B with FSDP + TP + CP** ([qwen3-8b-fsdp-tp-cp.yaml](./qwen3-8b-fsdp-tp-cp.yaml))
+- Uses all 3 parallelism dimensions on a single node
+- Ideal for: when model weights, activations, and/or context are too large to fit on single GPU
+
+```yaml
+dp_shard_size: 2         # FSDP across 2 GPUs
+tensor_parallel_size: 2  # TP across 2 GPUs
+context_parallel_size: 2 # CP across 2 GPUs
+# Total: 2 × 2 × 2 = 8 GPUs
+```
+
+### Multi-Node
+
+**Llama 3.1 8B with HSDP + TP** ([llama-3_1-8b-hsdp-tp.yaml](./llama-3_1-8b-hsdp-tp.yaml))
+- FSDP & TP within nodes, DDP across nodes to minimize inter-node communication
+- Ideal for: Scaling to multiple nodes while maintaining training efficiency
+
+```yaml
+dp_shard_size: 4        # FSDP within each 4-GPU group
+tensor_parallel_size: 2 # TP within each node
+dp_replicate_size: 2    # DDP across 2 groups
+# Total: (4 × 2) × 2 = 16 GPUs (2 nodes)
+```
+
+## Learn More
+
+- [ND Parallelism Documentation](https://docs.axolotl.ai/docs/nd_parallelism.html)
+- [Blog: Accelerate ND-Parallel Guide](https://huggingface.co/blog/accelerate-nd-parallel)
+- [Multi-GPU Training Guide](https://docs.axolotl.ai/docs/multi-gpu.html)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/distributed-parallel/llama-3_1-8b-hsdp-tp.yaml

@@ -0,0 +1,47 @@
+base_model: meta-llama/Llama-3.1-8B
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+dp_shard_size: 4
+dp_replicate_size: 2
+tensor_parallel_size: 2
+# context_parallel_size: 2
+
+dataset_prepared_path: last_run_prepared
+
+special_tokens:
+  pad_token: <|end_of_text|>
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  state_dict_type: FULL_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: LlamaDecoderLayer
+  reshard_after_forward: true
+
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+
+output_dir: ./outputs/ndp-out/
+
+sequence_len: 2048
+sample_packing: true
+flash_attention: true
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 2
+optimizer: adamw_torch_fused
+lr_scheduler: constant_with_warmup
+learning_rate: 2e-6
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1

examples/distributed-parallel/qwen3-8b-fsdp-tp-cp.yaml

@@ -0,0 +1,46 @@
+base_model: Qwen/Qwen3-8B
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+dp_shard_size: 2
+# dp_replicate_size: 1
+context_parallel_size: 2
+tensor_parallel_size: 2
+
+dataset_prepared_path: last_run_prepared
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  state_dict_type: FULL_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen3DecoderLayer
+  reshard_after_forward: true
+
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+
+output_dir: ./outputs/ndp-out/
+
+sequence_len: 8192
+sample_packing: true
+flash_attention: true
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1  # must be 1 when using context parallel
+num_epochs: 2
+optimizer: adamw_torch_fused
+lr_scheduler: constant_with_warmup
+learning_rate: 2e-6
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+
+special_tokens:

examples/falcon-h1/falcon-h1-1b-deep-qlora.yaml

@@ -38,7 +38,7 @@ lora_target_modules:
 sequence_len: 2048
 sample_packing: false
 eval_sample_packing: false
-pad_to_sequence_len: true
+
 
 wandb_project:
 wandb_entity:

examples/falcon-h1/falcon-h1-1b-qlora.yaml

@@ -38,7 +38,7 @@ lora_target_modules:
 sequence_len: 2048
 sample_packing: false
 eval_sample_packing: false
-pad_to_sequence_len: true
+
 
 wandb_project:
 wandb_entity:

examples/falcon-h1/falcon-h1-34b-qlora.yaml

@@ -38,7 +38,7 @@ lora_target_modules:
 sequence_len: 2048
 sample_packing: false
 eval_sample_packing: false
-pad_to_sequence_len: true
+
 
 wandb_project:
 wandb_entity:

examples/falcon-h1/falcon-h1-3b-qlora.yaml

@@ -38,7 +38,7 @@ lora_target_modules:
 sequence_len: 2048
 sample_packing: false
 eval_sample_packing: false
-pad_to_sequence_len: true
+
 
 wandb_project:
 wandb_entity:

examples/falcon-h1/falcon-h1-500m-qlora.yaml

@@ -38,7 +38,7 @@ lora_target_modules:
 sequence_len: 2048
 sample_packing: false
 eval_sample_packing: false
-pad_to_sequence_len: true
+
 
 wandb_project:
 wandb_entity:

examples/falcon-h1/falcon-h1-7b-qlora.yaml

@@ -38,7 +38,7 @@ lora_target_modules:
 sequence_len: 2048
 sample_packing: false
 eval_sample_packing: false
-pad_to_sequence_len: true
+
 
 wandb_project:
 wandb_entity:

examples/gemma2/qlora.yml

@@ -31,7 +31,7 @@ lora_target_linear: true
 sequence_len: 2048
 sample_packing: true
 eval_sample_packing: false
-pad_to_sequence_len: true
+
 
 wandb_project:
 wandb_entity: