configurable weight scale normalization for MoE expert drift

* Fix DO_NOT_TRACK not being correctly handled

* add unit tests and lint

---------

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

.github/CONTRIBUTING.md

@@ -68,7 +68,7 @@ You can skip certain CI checks by including specific keywords in your commit mes
 
 ### Code Style
 
-axolotl uses [{codestyle}]({URLofCodestyle}) as its code style guide. Please ensure that your code follows these guidelines.
+axolotl uses [Ruff](https://docs.astral.sh/ruff/) as its code style guide. Please ensure that your code follows these guidelines.
 
 Use the pre-commit linter to ensure that your code is formatted consistently.
 ```bash
@@ -83,6 +83,6 @@ Write clear and concise commit messages that briefly describe the changes made i
 
 - [GitHub Help](https://help.github.com/)
 - [GitHub Pull Request Documentation](https://docs.github.com/en/github/collaborating-with-issues-and-pull-requests)
-- [{codestyle}]({URLofCodestyle})
+- [Ruff](https://docs.astral.sh/ruff/)
 
 Thank you once again for your interest in contributing to axolotl. We look forward to collaborating with you and creating an even better project together!

.github/workflows/base.yml

@@ -15,6 +15,9 @@ on:
       - '.github/workflows/base.yml'
   workflow_dispatch:
 
+permissions:
+  contents: read
+
 jobs:
   build-base:
     if: ${{ github.repository_owner == 'axolotl-ai-cloud' && (github.event_name != 'pull_request' || !github.event.pull_request.draft) }}
@@ -27,14 +30,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - 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-base"
-            platforms: "linux/amd64"
           - cuda: "128"
             cuda_version: 12.8.1
             cudnn_version: ""
@@ -124,7 +119,7 @@ jobs:
           images: |
             axolotlai/axolotl-base
       - name: Login to Docker Hub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         if: ${{ github.event_name != 'pull_request' && env.HAS_DOCKERHUB_CREDS == 'true' }}
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
@@ -132,7 +127,7 @@ jobs:
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
       - name: Build
-        uses: docker/build-push-action@v4
+        uses: docker/build-push-action@v5
         with:
           context: .
           file: ./docker/${{ matrix.dockerfile }}
@@ -161,14 +156,14 @@ jobs:
             cuda_version: 12.8.1
             cudnn_version: ""
             python_version: "3.11"
-            pytorch: 2.8.0
+            pytorch: 2.9.1
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-uv-base"
-            platforms: "linux/amd64"
+            platforms: "linux/amd64,linux/arm64"
           - cuda: "128"
             cuda_version: 12.8.1
             cudnn_version: ""
-            python_version: "3.11"
+            python_version: "3.12"
             pytorch: 2.9.1
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-uv-base"
@@ -239,7 +234,7 @@ jobs:
           images: |
             axolotlai/axolotl-base-uv
       - name: Login to Docker Hub
-        uses: docker/login-action@v2
+        uses: docker/login-action@v3
         if: ${{ github.event_name != 'pull_request' && env.HAS_DOCKERHUB_CREDS == 'true' }}
         with:
           username: ${{ secrets.DOCKERHUB_USERNAME }}
@@ -247,7 +242,7 @@ jobs:
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
       - name: Build
-        uses: docker/build-push-action@v4
+        uses: docker/build-push-action@v5
         with:
           context: .
           file: ./docker/${{ matrix.dockerfile }}

.github/workflows/lint.yml

@@ -13,6 +13,9 @@ on:
        - ".pre-commit-config.yaml"
   workflow_dispatch:
 
+permissions:
+  contents: read
+
 jobs:
   pre-commit:
     name: pre-commit

.github/workflows/main.yml

@@ -8,6 +8,9 @@ on:
       - "v*"
   workflow_dispatch:
 
+permissions:
+  contents: read
+
 jobs:
   build-axolotl:
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
@@ -15,12 +18,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.8.0
-            axolotl_extras:
-            platforms: "linux/amd64"
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
@@ -110,6 +107,12 @@ jobs:
             pytorch: 2.9.1
             axolotl_extras:
             platforms: "linux/amd64,linux/arm64"
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.12"
+            pytorch: 2.9.1
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
             is_latest: true
           - cuda: 128
             cuda_version: 12.8.1
@@ -174,14 +177,9 @@ jobs:
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
     strategy:
+      fail-fast: false
       matrix:
         include:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.8.0
-            axolotl_extras:
-            platforms: "linux/amd64"
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
@@ -259,6 +257,7 @@ jobs:
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
     strategy:
+      fail-fast: false
       matrix:
         include:
           - cuda: 128
@@ -266,6 +265,12 @@ jobs:
             python_version: "3.11"
             pytorch: 2.9.1
             axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.12"
+            pytorch: 2.9.1
+            axolotl_extras:
             is_latest: true
             platforms: "linux/amd64,linux/arm64"
           - cuda: 128
@@ -326,6 +331,7 @@ jobs:
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...
     strategy:
+      fail-fast: false
       matrix:
         include:
           - cuda: 128

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

@@ -20,6 +20,9 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
 
+permissions:
+  contents: read
+
 env:
   MODAL_IMAGE_BUILDER_VERSION: "2025.06"
 
@@ -30,12 +33,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.8.0
-            axolotl_extras: fbgemm-gpu
-            num_gpus: 2
 #          - cuda: 129
 #            cuda_version: 12.9.1
 #            python_version: "3.12"
@@ -78,8 +75,9 @@ jobs:
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
           echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
+        env:
+          CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
         run: |
           modal run -m cicd.multigpu

.github/workflows/nightlies.yml

@@ -5,6 +5,9 @@ on:
   schedule:
     - cron: '0 0 * * *'  # Runs at 00:00 UTC every day
 
+permissions:
+  contents: read
+
 jobs:
   build-axolotl:
     if: ${{ ! contains(github.event.commits[0].message, '[skip docker]') && github.repository_owner == 'axolotl-ai-cloud' }}
@@ -12,11 +15,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.8.0
-            axolotl_extras:
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
@@ -64,11 +62,6 @@ jobs:
     strategy:
       matrix:
         include:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.8.0
-            axolotl_extras:
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"

.github/workflows/precommit-autoupdate.yml

@@ -5,6 +5,8 @@ on:
     - cron: '0 0 1 * *'  # Run monthly
   workflow_dispatch:  # Manual kickoff
 
+permissions: {}
+
 jobs:
   auto-update:
     runs-on: ubuntu-latest

.github/workflows/pypi.yml

@@ -3,9 +3,11 @@ name: publish pypi
 on:
   push:
     tags:
-      - 'v*'
+      - "v*"
   workflow_dispatch:
 
+permissions: {}
+
 jobs:
   setup_release:
     name: Create Release
@@ -28,7 +30,8 @@ jobs:
       name: pypi
       url: https://pypi.org/p/axolotl
     permissions:
-      id-token: write  # IMPORTANT: this permission is mandatory for trusted publishing
+      contents: read
+      id-token: write # IMPORTANT: this permission is mandatory for trusted publishing
     steps:
       - name: Check out repository code
         uses: actions/checkout@v4
@@ -46,7 +49,7 @@ jobs:
 
       - name: Extract tag name
         id: tag
-        run: echo ::set-output name=TAG_NAME::$(echo $GITHUB_REF | cut -d / -f 3)
+        run: echo "TAG_NAME=$(echo $GITHUB_REF | cut -d / -f 3)" >> "$GITHUB_OUTPUT"
 
       - name: Update version in VERSION file
         run: |

.github/workflows/tests-nightly.yml

@@ -8,6 +8,9 @@ on:
     paths:
       - '.github/workflows/tests-nightly.yml'
 
+permissions:
+  contents: read
+
 jobs:
   pre-commit:
     name: pre-commit
@@ -41,7 +44,7 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.12"]  # TODO include py3.14 once https://github.com/mistralai/mistral-common/pull/194 is merged
-        pytorch_version: ["2.8.0", "2.9.1", "2.10.0"]
+        pytorch_version: ["2.9.1", "2.10.0"]
     timeout-minutes: 20
 
     steps:
@@ -156,8 +159,9 @@ jobs:
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
           echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
           echo "NIGHTLY_BUILD=${{ matrix.nightly_build }}" >> $GITHUB_ENV
-          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
+        env:
+          CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
         run: |
           modal run cicd.e2e_tests
   docker-e2e-multigpu-tests:
@@ -198,7 +202,8 @@ jobs:
           echo "CUDA=${{ 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
+        env:
+          CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
         run: |
           modal run cicd.multigpu

.github/workflows/tests.yml

@@ -28,6 +28,9 @@ concurrency:
   group: ${{ github.workflow }}-${{ github.ref }}
   cancel-in-progress: ${{ github.ref != 'refs/heads/main' }}
 
+permissions:
+  contents: read
+
 env:
   TRANSFORMERS_IS_CI: "yes"
 
@@ -65,13 +68,11 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python_version: ["3.12"]  # TODO include py3.14 once https://github.com/mistralai/mistral-common/pull/194 is merged
-        pytorch_version: ["2.8.0", "2.9.1", "2.10.0"]
-#        exclude:
-#          - python_version: "3.14"
-#            pytorch_version: "2.8.0"
-#          - python_version: "3.14"
-#            pytorch_version: "2.9.1"
+        python_version: ["3.12", "3.14"]
+        pytorch_version: ["2.9.1", "2.10.0"]
+        exclude:
+          - python_version: "3.14"
+            pytorch_version: "2.9.1"
     timeout-minutes: 20
 
     steps:
@@ -161,13 +162,11 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python_version: ["3.12"]  # TODO include py3.14 once https://github.com/mistralai/mistral-common/pull/194 is merged
-        pytorch_version: ["2.8.0", "2.9.1", "2.10.0"]
-#        exclude:
-#          - python_version: "3.14"
-#            pytorch_version: "2.8.0"
-#          - python_version: "3.14"
-#            pytorch_version: "2.9.1"
+        python_version: ["3.12", "3.14"]
+        pytorch_version: ["2.9.1", "2.10.0"]
+        exclude:
+          - python_version: "3.14"
+            pytorch_version: "2.9.1"
     timeout-minutes: 30
 
     steps:
@@ -303,9 +302,10 @@ jobs:
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
           echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
+        env:
+          CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
         run: |
           modal run cicd.e2e_tests
 
@@ -325,13 +325,6 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 128
-            cuda_version: 12.8.1
-            python_version: "3.11"
-            pytorch: 2.8.0
-            num_gpus: 1
-            gpu_type: "B200"
-            axolotl_extras: fbgemm-gpu
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
@@ -371,9 +364,10 @@ jobs:
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
           echo "GPU_TYPE=${{ matrix.gpu_type || 'L40S'}}" >> $GITHUB_ENV
-          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
           echo "E2E_DOCKERFILE=${{ matrix.dockerfile || 'Dockerfile.jinja'}}" >> $GITHUB_ENV
       - name: Run tests job on Modal
+        env:
+          CODECOV_TOKEN: ${{ secrets.CODECOV_TOKEN }}
         run: |
           modal run cicd.e2e_tests
 
@@ -413,7 +407,6 @@ jobs:
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |
           modal run cicd.cleanup

.pre-commit-config.yaml

@@ -11,7 +11,7 @@ repos:
     -   id: no-commit-to-branch
         args: ['--branch', 'main']
 -   repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.15.4
+    rev: v0.15.8
     hooks:
     -   id: ruff
         args: [--fix]

AGENTS.md

@@ -0,0 +1,94 @@
+# Axolotl
+
+Fine-tuning framework for LLMs. Config-driven: every training run is defined by a single YAML file.
+
+## Tech Stack
+
+Python, PyTorch, HuggingFace Transformers, TRL, PEFT (LoRA/QLoRA), DeepSpeed, FSDP, vLLM (for GRPO generation).
+
+## Commands
+
+```bash
+axolotl train config.yaml              # Train (single or multi-GPU, auto-detected)
+axolotl preprocess config.yaml         # Tokenize dataset and validate config
+axolotl preprocess config.yaml --debug # Inspect tokenized samples and label masking
+axolotl inference config.yaml          # Interactive inference
+axolotl merge-lora config.yaml         # Merge LoRA adapter into base model
+axolotl vllm-serve config.yaml         # Start vLLM server for GRPO/EBFT training
+axolotl fetch examples                 # Download example configs
+```
+
+## Training Methods
+
+| Method | Config Key | When to Use |
+|--------|-----------|-------------|
+| SFT | *(default)* | Input-output pairs, instruction tuning |
+| DPO/IPO | `rl: dpo` / `rl: ipo` | Paired preference data (chosen vs rejected) |
+| KTO | `rl: kto` | Unpaired binary preference labels |
+| ORPO | `rl: orpo` | Single-stage alignment, no ref model |
+| GRPO | `rl: grpo` | RL with verifiable reward functions (math, code) |
+| EBFT | `rl: ebft` | Feature-matching rewards from internal representations |
+
+Agent-specific references:
+- [docs/agents/sft.md](docs/agents/sft.md) — supervised fine-tuning
+- [docs/agents/preference_tuning.md](docs/agents/preference_tuning.md) — DPO, IPO, KTO, ORPO, SimPO
+- [docs/agents/grpo.md](docs/agents/grpo.md) — GRPO online RL with reward functions
+- [docs/agents/reward_modelling.md](docs/agents/reward_modelling.md) — outcome and process reward models
+- [docs/agents/pretraining.md](docs/agents/pretraining.md) — continual pretraining
+
+## Config Pattern
+
+All training is config-driven. A YAML file specifies model, adapter, dataset(s), and hyperparameters:
+
+```yaml
+base_model: meta-llama/Llama-3.1-8B-Instruct
+adapter: lora                    # or qlora, or omit for full fine-tune
+datasets:
+  - path: my_dataset
+    type: chat_template          # prompt strategy (see docs/dataset-formats/)
+output_dir: ./outputs/lora-out
+```
+
+Config schema: `src/axolotl/utils/schemas/config.py` (AxolotlInputConfig).
+
+## Project Structure
+
+```
+src/axolotl/
+  cli/                           # CLI entry points (train, preprocess, inference, merge_lora, vllm_serve)
+  core/
+    builders/                    # TrainerBuilder classes (causal.py for SFT, rl.py for RLHF)
+    trainers/                    # Trainer classes, mixins (optimizer, scheduler, packing)
+      dpo/                       # DPO trainer and config
+      grpo/                      # GRPO trainer and sampler
+  loaders/                       # Model, tokenizer, adapter, processor loading
+  prompt_strategies/             # Dataset format handlers (chat_template, alpaca, dpo/, kto/, orpo/)
+  utils/schemas/                 # Pydantic config schemas (config, model, training, peft, trl, fsdp)
+  integrations/                  # Plugins (liger, cut_cross_entropy, swanlab, nemo_gym)
+  monkeypatch/                   # Runtime patches for HF transformers
+
+examples/                        # Example YAML configs by model (llama-3/, qwen2/, mistral/, ebft/)
+deepspeed_configs/               # DeepSpeed JSON configs (zero2, zero3)
+docs/                            # Quarto documentation site
+```
+
+## Code Conventions
+
+- Config-driven: features are toggled via YAML, not code changes
+- Prompt strategies: `src/axolotl/prompt_strategies/` — each `type:` value maps to a function
+- Plugin system: `plugins:` list in config loads integration modules
+- Trainer mixins: `core/trainers/mixins/` for composable trainer behaviors
+- Schemas: all config validation via Pydantic in `utils/schemas/`
+
+## Key Documentation
+
+- [Getting Started](docs/getting-started.qmd) — quickstart tutorial
+- [Choosing a Method](docs/choosing_method.qmd) — SFT vs DPO vs GRPO decision guide
+- [Config Reference](docs/config-reference.qmd) — all config options
+- [Dataset Formats](docs/dataset-formats/) — chat_template, alpaca, input_output, completion
+- [RLHF](docs/rlhf.qmd) — DPO, KTO, ORPO, GRPO, EBFT configs and dataset formats
+- [GRPO Deep Dive](docs/grpo.qmd) — async training, custom rewards, scaling
+- [vLLM Serving](docs/vllm_serving.qmd) — vLLM setup for GRPO/EBFT
+- [Multi-GPU](docs/multi-gpu.qmd) — FSDP and DeepSpeed
+- [Training Stability](docs/training_stability.qmd) — debugging loss, NaN, OOM
+- [Debugging](docs/debugging.qmd) — VSCode setup, Docker debugging

README.md

@@ -30,7 +30,7 @@
 ## 🎉 Latest Updates
 
 - 2026/03:
-  - New model support has been added in Axolotl for [Qwen3.5, Qwen3.5 MoE](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/qwen3.5), [GLM-4.7-Flash](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm47-flash), [GLM-4.6V](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm46v), and [GLM-4.5-Air](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm45).
+  - New model support has been added in Axolotl for [Mistral Small 4](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/mistral4), [Qwen3.5, Qwen3.5 MoE](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/qwen3.5), [GLM-4.7-Flash](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm47-flash), [GLM-4.6V](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm46v), and [GLM-4.5-Air](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/glm45).
   - [MoE expert quantization](https://docs.axolotl.ai/docs/expert_quantization.html) support (via `quantize_moe_experts: true`) greatly reduces VRAM when training MoE models (FSDP2 compat).
 - 2026/02:
   - [ScatterMoE LoRA](https://github.com/axolotl-ai-cloud/axolotl/pull/3410) support. LoRA fine-tuning directly on MoE expert weights using custom Triton kernels.
@@ -75,7 +75,7 @@ Features:
 - **Multimodal Training**: Fine-tune vision-language models (VLMs) including LLaMA-Vision, Qwen2-VL, Pixtral, LLaVA, SmolVLM2, GLM-4.6V, InternVL 3.5, Gemma 3n, and audio models like Voxtral with image, video, and audio support.
 - **Training Methods**: Full fine-tuning, LoRA, QLoRA, GPTQ, QAT, Preference Tuning (DPO, IPO, KTO, ORPO), RL (GRPO, GDPO), and Reward Modelling (RM) / Process Reward Modelling (PRM).
 - **Easy Configuration**: Re-use a single YAML configuration file across the full fine-tuning pipeline: dataset preprocessing, training, evaluation, quantization, and inference.
-- **Performance Optimizations**: [Multipacking](https://docs.axolotl.ai/docs/multipack.html), [Flash Attention](https://github.com/Dao-AILab/flash-attention), [Xformers](https://github.com/facebookresearch/xformers), [Flex Attention](https://pytorch.org/blog/flexattention/), [SageAttention](https://github.com/thu-ml/SageAttention), [Liger Kernel](https://github.com/linkedin/Liger-Kernel), [Cut Cross Entropy](https://github.com/apple/ml-cross-entropy/tree/main), [ScatterMoE](https://docs.axolotl.ai/docs/custom_integrations.html#kernels-integration), [Sequence Parallelism (SP)](https://docs.axolotl.ai/docs/sequence_parallelism.html), [LoRA optimizations](https://docs.axolotl.ai/docs/lora_optims.html), [Multi-GPU training (FSDP1, FSDP2, DeepSpeed)](https://docs.axolotl.ai/docs/multi-gpu.html), [Multi-node training (Torchrun, Ray)](https://docs.axolotl.ai/docs/multi-node.html), and many more!
+- **Performance Optimizations**: [Multipacking](https://docs.axolotl.ai/docs/multipack.html), [Flash Attention 2/3/4](https://docs.axolotl.ai/docs/attention.html#flash-attention), [Xformers](https://docs.axolotl.ai/docs/attention.html#xformers), [Flex Attention](https://docs.axolotl.ai/docs/attention.html#flex-attention), [SageAttention](https://docs.axolotl.ai/docs/attention.html#sageattention), [Liger Kernel](https://docs.axolotl.ai/docs/custom_integrations.html#liger-kernels), [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy), [ScatterMoE](https://docs.axolotl.ai/docs/custom_integrations.html#kernels-integration), [Sequence Parallelism (SP)](https://docs.axolotl.ai/docs/sequence_parallelism.html), [LoRA optimizations](https://docs.axolotl.ai/docs/lora_optims.html), [Multi-GPU training (FSDP1, FSDP2, DeepSpeed)](https://docs.axolotl.ai/docs/multi-gpu.html), [Multi-node training (Torchrun, Ray)](https://docs.axolotl.ai/docs/multi-node.html), and many more!
 - **Flexible Dataset Handling**: Load from local, HuggingFace, and cloud (S3, Azure, GCP, OCI) datasets.
 - **Cloud Ready**: We ship [Docker images](https://hub.docker.com/u/axolotlai) and also [PyPI packages](https://pypi.org/project/axolotl/) for use on cloud platforms and local hardware.
 
@@ -87,7 +87,7 @@ Features:
 
 - NVIDIA GPU (Ampere or newer for `bf16` and Flash Attention) or AMD GPU
 - Python 3.11
-- PyTorch ≥2.8.0
+- PyTorch ≥2.9.1
 
 ### Google Colab
 

_quarto.yml

@@ -128,11 +128,9 @@ quartodoc:
         - monkeypatch.mistral_attn_hijack_flash
         - monkeypatch.multipack
         - monkeypatch.relora
-        - monkeypatch.llama_expand_mask
         - monkeypatch.lora_kernels
         - monkeypatch.utils
         - monkeypatch.btlm_attn_hijack_flash
-        - monkeypatch.llama_patch_multipack
         - monkeypatch.stablelm_attn_hijack_flash
         - monkeypatch.trainer_fsdp_optim
         - monkeypatch.transformers_fa_utils
@@ -240,6 +238,7 @@ website:
         - section: "Getting Started"
           contents:
             - docs/getting-started.qmd
+            - docs/choosing_method.qmd
             - docs/installation.qmd
             - docs/inference.qmd
             - section: "Model Guides"
@@ -304,6 +303,9 @@ website:
           contents:
             - docs/multimodal.qmd
             - docs/rlhf.qmd
+            - docs/grpo.qmd
+            - docs/ebft.qmd
+            - docs/vllm_serving.qmd
             - docs/reward_modelling.qmd
             - docs/lr_groups.qmd
             - docs/lora_optims.qmd
@@ -336,6 +338,7 @@ website:
         - section: "Troubleshooting"
           contents:
             - docs/faq.qmd
+            - docs/training_stability.qmd
             - docs/debugging.qmd
             - docs/nccl.qmd
 

benchmarks/bench_entropy.py

@@ -0,0 +1,208 @@
+"""Benchmark for entropy_from_logits Triton kernel vs original chunked implementation.
+
+Usage: CUDA_VISIBLE_DEVICES=0 python benchmarks/bench_entropy.py
+"""
+
+import gc
+import statistics
+
+import torch
+import torch.nn.functional as F
+
+from axolotl.monkeypatch.trainer.utils import entropy_from_logits
+
+V = 151936  # Qwen vocab
+WARMUP = 5
+BENCH_ITERS = 20
+MEM_ITERS = 10
+
+
+def entropy_from_logits_original(logits: torch.Tensor, chunk_size: int = 128):
+    """Original chunked implementation (reference)."""
+    original_shape = logits.shape[:-1]
+    num_classes = logits.shape[-1]
+    flat_logits = logits.reshape(-1, num_classes)
+    entropies = []
+    for chunk in flat_logits.split(chunk_size, dim=0):
+        logps = F.log_softmax(chunk, dim=-1)
+        chunk_entropy = -(torch.exp(logps) * logps).sum(-1)
+        entropies.append(chunk_entropy)
+    return torch.cat(entropies, dim=0).reshape(original_shape)
+
+
+def _clean_gpu():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_peak_memory_stats()
+    torch.cuda.reset_accumulated_memory_stats()
+    torch.cuda.synchronize()
+
+
+def profile_time(fn, logits, n_iters=BENCH_ITERS):
+    for _ in range(WARMUP):
+        out = fn(logits, chunk_size=128)
+        del out
+    torch.cuda.synchronize()
+
+    times = []
+    for _ in range(n_iters):
+        s = torch.cuda.Event(enable_timing=True)
+        e = torch.cuda.Event(enable_timing=True)
+        s.record()
+        out = fn(logits, chunk_size=128)
+        e.record()
+        torch.cuda.synchronize()
+        times.append(s.elapsed_time(e))
+        del out
+    return times
+
+
+def profile_memory(fn, logits, n_iters=MEM_ITERS):
+    for _ in range(WARMUP):
+        out = fn(logits, chunk_size=128)
+        del out
+    torch.cuda.synchronize()
+
+    peaks = []
+    for _ in range(n_iters):
+        _clean_gpu()
+        base = torch.cuda.max_memory_allocated()
+        out = fn(logits, chunk_size=128)
+        torch.cuda.synchronize()
+        peaks.append(torch.cuda.max_memory_allocated() - base)
+        del out
+    return [p / 1e6 for p in peaks]
+
+
+def fmt(values, unit=""):
+    mean = statistics.mean(values)
+    std = statistics.stdev(values) if len(values) > 1 else 0.0
+    return f"{mean:8.2f} ± {std:5.2f} {unit}  [min={min(values):.2f}, max={max(values):.2f}]"
+
+
+def benchmark_contiguous():
+    print("=" * 60)
+    print(
+        f"CONTIGUOUS BENCHMARK  (warmup={WARMUP}, time={BENCH_ITERS}, mem={MEM_ITERS})"
+    )
+    print("=" * 60)
+
+    configs = [
+        (1, 2048),
+        (1, 8192),
+        (1, 16384),
+        (4, 4096),
+        (8, 2048),
+        (16, 2048),
+        (16, 4096),
+    ]
+
+    for B, L in configs:
+        mem_gb = B * L * V * 2 / 1e9
+        if mem_gb > 28:
+            print(f"\n  skip B={B}, L={L} ({mem_gb:.1f} GB)")
+            continue
+
+        N = B * L
+        print(f"\n{'─' * 60}")
+        print(f"B={B:2d}, L={L:5d}  ({N:6d} rows, logits {mem_gb:.2f} GB)")
+        print(f"{'─' * 60}")
+
+        torch.manual_seed(42)
+        logits = torch.randn(B, L, V, device="cuda", dtype=torch.bfloat16)
+
+        t_orig = profile_time(entropy_from_logits_original, logits)
+        t_triton = profile_time(entropy_from_logits, logits)
+        orig_mean = statistics.mean(t_orig)
+        triton_mean = statistics.mean(t_triton)
+
+        print("  TIME (ms):")
+        print(f"    original: {fmt(t_orig, 'ms')}")
+        print(f"    triton:   {fmt(t_triton, 'ms')}")
+        print(f"    speedup:  {orig_mean / triton_mean:.2f}x")
+
+        m_orig = profile_memory(entropy_from_logits_original, logits)
+        m_triton = profile_memory(entropy_from_logits, logits)
+        orig_peak = statistics.mean(m_orig)
+        triton_peak = statistics.mean(m_triton)
+
+        print("  MEMORY (peak overhead):")
+        print(f"    original: {fmt(m_orig, 'MB')}")
+        print(f"    triton:   {fmt(m_triton, 'MB')}")
+        print(f"    saved:    {orig_peak - triton_peak:.1f} MB")
+
+        del logits
+        _clean_gpu()
+
+
+def benchmark_noncontiguous():
+    print("\n" + "=" * 60)
+    print(
+        f"NON-CONTIGUOUS BENCHMARK  (warmup={WARMUP}, time={BENCH_ITERS}, mem={MEM_ITERS})"
+    )
+    print("=" * 60)
+
+    configs = [
+        (4, 2048, "transpose"),
+        (4, 8192, "transpose"),
+        (8, 2048, "transpose"),
+        (4, 4096, "slice_batch"),
+    ]
+
+    for B, L, method in configs:
+        torch.manual_seed(42)
+
+        if method == "transpose":
+            raw = torch.randn(L, B, V, device="cuda", dtype=torch.bfloat16)
+            logits_nc = raw.transpose(0, 1)
+            raw_gb = L * B * V * 2 / 1e9
+        elif method == "slice_batch":
+            raw = torch.randn(B * 2, L, V, device="cuda", dtype=torch.bfloat16)
+            logits_nc = raw[::2]
+            raw_gb = B * 2 * L * V * 2 / 1e9
+        else:
+            continue
+
+        if raw_gb > 28:
+            print(f"\n  skip B={B}, L={L}, {method} ({raw_gb:.1f} GB)")
+            del raw, logits_nc
+            torch.cuda.empty_cache()
+            continue
+
+        N = B * L
+        print(f"\n{'─' * 60}")
+        print(f"B={B}, L={L}  {method}  ({N} rows, raw {raw_gb:.2f} GB)")
+        print(f"{'─' * 60}")
+
+        def original_with_copy(logits, chunk_size=128):
+            return entropy_from_logits_original(
+                logits.contiguous(), chunk_size=chunk_size
+            )
+
+        t_orig = profile_time(original_with_copy, logits_nc)
+        t_triton = profile_time(entropy_from_logits, logits_nc)
+        orig_mean = statistics.mean(t_orig)
+        triton_mean = statistics.mean(t_triton)
+
+        print("  TIME (ms):")
+        print(f"    orig+copy:     {fmt(t_orig, 'ms')}")
+        print(f"    triton-strided:{fmt(t_triton, 'ms')}")
+        print(f"    speedup:       {orig_mean / triton_mean:.2f}x")
+
+        m_orig = profile_memory(original_with_copy, logits_nc)
+        m_triton = profile_memory(entropy_from_logits, logits_nc)
+        orig_peak = statistics.mean(m_orig)
+        triton_peak = statistics.mean(m_triton)
+
+        print("  MEMORY (peak overhead):")
+        print(f"    orig+copy:     {fmt(m_orig, 'MB')}")
+        print(f"    triton-strided:{fmt(m_triton, 'MB')}")
+        print(f"    saved:         {orig_peak - triton_peak:.1f} MB")
+
+        del raw, logits_nc
+        _clean_gpu()
+
+
+if __name__ == "__main__":
+    benchmark_contiguous()
+    benchmark_noncontiguous()

benchmarks/bench_scattermoe_lora.py

@@ -0,0 +1,284 @@
+"""Benchmark for ScatterMoE LoRA Triton kernels.
+
+Measures forward, backward dX, and backward dA/dB kernels at common MoE
+model shapes. Reports per-kernel timings, LoRA overhead vs base scatter2scatter,
+and full fwd+bwd autograd throughput.
+
+Usage:
+  CUDA_VISIBLE_DEVICES=0 python benchmarks/bench_scattermoe_lora.py
+  CUDA_VISIBLE_DEVICES=0 python benchmarks/bench_scattermoe_lora.py --ranks 16 64
+  CUDA_VISIBLE_DEVICES=0 python benchmarks/bench_scattermoe_lora.py --models Qwen/Qwen3.5-35B-A3B
+"""
+
+import argparse
+import gc
+import time
+from functools import partial
+
+import torch
+
+from axolotl.integrations.kernels.libs.scattermoe_lora.kernels import (
+    lora_ops,
+    ops as base_ops,
+)
+from axolotl.integrations.kernels.libs.scattermoe_lora.parallel_experts import (
+    flatten_sort_count,
+)
+from axolotl.integrations.kernels.libs.scattermoe_lora.parallel_linear_lora import (
+    ScatterMoELoRA,
+)
+
+DEVICE = "cuda"
+DTYPE = torch.bfloat16
+WARMUP = 5
+ITERS = 20
+
+# ─── Model configs ──────────────────────────────────────────────────────────
+
+BUILTIN_CONFIGS = {
+    "Qwen3.5-35B-A3B": (256, 2048, 512, 8),  # E, H, I, k
+    "Qwen3-30B-A3B": (128, 2048, 768, 8),
+    "OLMoE-1B-7B": (64, 2048, 1024, 8),
+    "Mixtral-8x7B": (8, 4096, 14336, 2),
+}
+
+
+def _resolve_config(spec):
+    """Resolve a model spec to (E, H, I, k). Accepts builtin names or HF IDs."""
+    key = spec.lower().replace("/", "-")
+    for name, cfg in BUILTIN_CONFIGS.items():
+        if key in name.lower() or name.lower() in key:
+            return name, cfg
+
+    from transformers import AutoConfig
+
+    hf_cfg = AutoConfig.from_pretrained(spec, trust_remote_code=True)
+    if callable(getattr(hf_cfg, "get_text_config", None)):
+        tc = hf_cfg.get_text_config()
+        if hasattr(tc, "model_type") and tc.model_type != hf_cfg.model_type:
+            hf_cfg = tc
+    hidden = hf_cfg.hidden_size
+    inter = getattr(hf_cfg, "moe_intermediate_size", None) or hf_cfg.intermediate_size
+    experts = (
+        getattr(hf_cfg, "num_experts", None)
+        or getattr(hf_cfg, "num_local_experts", None)
+        or getattr(hf_cfg, "n_routed_experts", None)
+    )
+    top_k = (
+        getattr(hf_cfg, "num_experts_per_tok", None)
+        or getattr(hf_cfg, "num_experts_per_token", None)
+        or 2
+    )
+    name = spec.split("/")[-1]
+    return name, (experts, hidden, inter, top_k)
+
+
+# ─── Benchmark helpers ──────────────────────────────────────────────────────
+
+
+def _clean():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.synchronize()
+
+
+def _bench(fn, warmup=WARMUP, iters=ITERS):
+    for _ in range(warmup):
+        fn()
+    torch.cuda.synchronize()
+    times = []
+    for _ in range(iters):
+        torch.cuda.synchronize()
+        t0 = time.perf_counter()
+        fn()
+        torch.cuda.synchronize()
+        times.append((time.perf_counter() - t0) * 1000)
+    times.sort()
+    return times[len(times) // 2]
+
+
+def _setup(num_experts, K, N, T, top_k, R):
+    torch.manual_seed(42)
+    x = torch.randn(T, K, device=DEVICE, dtype=DTYPE)
+    W = torch.randn(num_experts, K, N, device=DEVICE, dtype=DTYPE) * 0.02
+    lora_A = torch.randn(R * num_experts, K, device=DEVICE, dtype=DTYPE) * 0.01
+    lora_B = torch.randn(N, R * num_experts, device=DEVICE, dtype=DTYPE) * 0.01
+    logits = torch.randn(T, num_experts, device=DEVICE)
+    _, top_idx = torch.topk(torch.softmax(logits, dim=-1), top_k, dim=-1)
+    sei, ssi, eo = flatten_sort_count(top_idx, num_experts)
+    gx = base_ops.group(x, ssi, fan_out=top_k)
+    dy = torch.randn(gx.size(0), N, device=DEVICE, dtype=DTYPE)
+    return x, W, lora_A, lora_B, sei, ssi, eo, gx, dy
+
+
+# ─── Kernel wrappers (avoid B023 loop-variable capture) ──────────────────────
+
+
+def _call_fwd(x, W, sei, ssi, top_k, lA, lB):
+    return lora_ops.scatter2scatter_lora(
+        X=x,
+        W=W,
+        sorted_expert_idxs=sei,
+        sorted_scattered_idxs=ssi,
+        k=top_k,
+        lora_A=lA,
+        lora_B=lB,
+        scaling=2.0,
+    )
+
+
+def _call_base(x, W, sei, ssi, top_k):
+    return base_ops.scatter2scatter(
+        X=x,
+        W=W,
+        sorted_expert_idxs=sei,
+        sorted_scattered_idxs=ssi,
+        k=top_k,
+    )
+
+
+def _call_dx(dy, W, sei, ssi, lA, lB):
+    return lora_ops.scatter2scatter_lora_dX(
+        DY=dy,
+        W=W,
+        sorted_expert_idxs=sei,
+        sorted_scattered_idxs=ssi,
+        k=1,
+        lora_A=lA,
+        lora_B=lB,
+        scaling=2.0,
+        dy_grouped=True,
+        dx_grouped=False,
+    )
+
+
+def _call_bwd(dy, gx, lA, lB, eo, num_experts):
+    return lora_ops.group_bwd_lora(
+        DY=dy,
+        X=gx,
+        lora_A=lA,
+        lora_B=lB,
+        expert_offsets=eo,
+        E=num_experts,
+        scaling=2.0,
+    )
+
+
+# ─── Main ────────────────────────────────────────────────────────────────────
+
+
+def main():
+    parser = argparse.ArgumentParser(description="ScatterMoE LoRA kernel benchmark")
+    parser.add_argument(
+        "--models",
+        "-m",
+        nargs="+",
+        help="Model names or HF IDs (default: all builtins)",
+    )
+    parser.add_argument("--ranks", "-r", nargs="+", type=int, default=[16, 32, 64])
+    parser.add_argument("--seq-len", "-T", type=int, default=2048)
+    args = parser.parse_args()
+
+    T = args.seq_len
+    print(f"GPU: {torch.cuda.get_device_name()}")
+    print(f"T={T}, ranks={args.ranks}\n")
+
+    if args.models:
+        configs = [_resolve_config(m) for m in args.models]
+    else:
+        configs = list(BUILTIN_CONFIGS.items())
+
+    for model_name, (num_experts, hidden, inter, top_k) in configs:
+        print(f"{'=' * 70}")
+        print(f"  {model_name}: E={num_experts}, H={hidden}, I={inter}, k={top_k}")
+        print(f"{'=' * 70}")
+
+        for R in args.ranks:
+            for proj, K, N in [("gate_up", hidden, 2 * inter), ("down", inter, hidden)]:
+                _clean()
+                x, W, lA, lB, sei, ssi, eo, gx, dy = _setup(
+                    num_experts, K, N, T, top_k, R
+                )
+
+                # Forward with LoRA (auto-dispatched: fused or split)
+                dispatch = (
+                    "split"
+                    if (
+                        num_experts <= lora_ops._SPLIT_LORA_FWD_MAX_EXPERTS
+                        and K * N >= lora_ops._SPLIT_LORA_FWD_THRESHOLD
+                    )
+                    else "fused"
+                )
+                t_fwd = _bench(partial(_call_fwd, x, W, sei, ssi, top_k, lA, lB))
+                t_base = _bench(partial(_call_base, x, W, sei, ssi, top_k))
+                t_dx = _bench(partial(_call_dx, dy, W, sei, ssi, lA, lB))
+                t_bwd = _bench(partial(_call_bwd, dy, gx, lA, lB, eo, num_experts))
+
+                total = t_fwd + t_dx + t_bwd
+                overhead = t_fwd / t_base - 1 if t_base > 0 else 0
+
+                print(
+                    f"  R={R:>2} {proj:<8}  "
+                    f"fwd={t_fwd:>6.2f}ms [{dispatch}]  "
+                    f"base={t_base:>6.2f}ms "
+                    f"(+{overhead * 100:.0f}%)  "
+                    f"dx={t_dx:>6.2f}ms  bwd={t_bwd:>6.2f}ms  "
+                    f"total={total:>6.2f}ms"
+                )
+
+                # Full autograd fwd+bwd with memory measurement
+                x_ag = x.clone().requires_grad_(True)
+                lA_ag = lA.clone().requires_grad_(True)
+                lB_ag = lB.clone().requires_grad_(True)
+
+                def _run_autograd(
+                    _x=x_ag,
+                    _W=W,
+                    _k=top_k,
+                    _sei=sei,
+                    _ssi=ssi,
+                    _eo=eo,
+                    _lA=lA_ag,
+                    _lB=lB_ag,
+                ):
+                    out = ScatterMoELoRA.apply(
+                        _x,
+                        _W,
+                        _k,
+                        _sei,
+                        _ssi,
+                        _eo,
+                        _lA,
+                        _lB,
+                        2.0,
+                        None,
+                        None,
+                        False,
+                        False,
+                        True,
+                        False,
+                    )
+                    out.sum().backward()
+                    _x.grad = None
+                    _lA.grad = None
+                    _lB.grad = None
+
+                t_full = _bench(_run_autograd)
+
+                _clean()
+                torch.cuda.reset_peak_memory_stats()
+                mem_before = torch.cuda.memory_allocated()
+                _run_autograd()
+                torch.cuda.synchronize()
+                mem_peak = torch.cuda.max_memory_allocated() - mem_before
+
+                print(
+                    f"         full_fwd_bwd={t_full:>6.2f}ms  "
+                    f"peak_delta={mem_peak / 1e6:>6.1f}MB"
+                )
+
+        print()
+
+
+if __name__ == "__main__":
+    main()

benchmarks/bench_selective_logsoftmax.py

@@ -0,0 +1,191 @@
+"""Benchmark for selective_log_softmax Triton kernel vs original implementation.
+
+Usage: CUDA_VISIBLE_DEVICES=0 python benchmarks/bench_selective_logsoftmax.py
+"""
+
+import gc
+import statistics
+
+import torch
+
+from axolotl.monkeypatch.trainer.utils import (
+    selective_log_softmax,
+    selective_log_softmax_original,
+)
+
+V = 151936  # Qwen vocab
+WARMUP = 5
+BENCH_ITERS = 20
+MEM_ITERS = 10
+
+
+def _clean_gpu():
+    gc.collect()
+    torch.cuda.empty_cache()
+    torch.cuda.reset_peak_memory_stats()
+    torch.cuda.reset_accumulated_memory_stats()
+    torch.cuda.synchronize()
+
+
+def profile_time(fn, args, n_iters=BENCH_ITERS):
+    for _ in range(WARMUP):
+        fn(*args)
+    torch.cuda.synchronize()
+
+    times = []
+    for _ in range(n_iters):
+        s = torch.cuda.Event(enable_timing=True)
+        e = torch.cuda.Event(enable_timing=True)
+        s.record()
+        fn(*args)
+        e.record()
+        torch.cuda.synchronize()
+        times.append(s.elapsed_time(e))
+    return times
+
+
+def profile_memory(fn, args, n_iters=MEM_ITERS):
+    for _ in range(WARMUP):
+        out = fn(*args)
+        del out
+    torch.cuda.synchronize()
+
+    peaks = []
+    for _ in range(n_iters):
+        _clean_gpu()
+        base = torch.cuda.max_memory_allocated()
+        out = fn(*args)
+        torch.cuda.synchronize()
+        peaks.append(torch.cuda.max_memory_allocated() - base)
+        del out
+    return [p / 1e6 for p in peaks]
+
+
+def fmt(values, unit=""):
+    mean = statistics.mean(values)
+    std = statistics.stdev(values) if len(values) > 1 else 0.0
+    return f"{mean:8.2f} ± {std:5.2f} {unit}  [min={min(values):.2f}, max={max(values):.2f}]"
+
+
+def benchmark_forward():
+    print("=" * 60)
+    print(f"FORWARD BENCHMARK  (warmup={WARMUP}, time={BENCH_ITERS}, mem={MEM_ITERS})")
+    print("=" * 60)
+
+    configs = [
+        (1, 2048),
+        (1, 8192),
+        (4, 4096),
+        (8, 2048),
+        (16, 2048),
+        (16, 4096),
+    ]
+
+    for B, L in configs:
+        mem_gb = B * L * V * 2 / 1e9
+        if mem_gb > 28:
+            print(f"\n  skip B={B}, L={L} ({mem_gb:.1f} GB)")
+            continue
+
+        N = B * L
+        print(f"\n{'─' * 60}")
+        print(f"B={B:2d}, L={L:5d}  ({N:6d} rows, logits {mem_gb:.2f} GB)")
+        print(f"{'─' * 60}")
+
+        torch.manual_seed(42)
+        logits = torch.randn(B, L, V, device="cuda", dtype=torch.bfloat16)
+        index = torch.randint(0, V, (B, L), device="cuda")
+
+        t_orig = profile_time(selective_log_softmax_original, (logits, index))
+        t_triton = profile_time(selective_log_softmax, (logits, index))
+        orig_mean = statistics.mean(t_orig)
+        triton_mean = statistics.mean(t_triton)
+
+        print("  TIME (ms):")
+        print(f"    original: {fmt(t_orig, 'ms')}")
+        print(f"    triton:   {fmt(t_triton, 'ms')}")
+        print(f"    speedup:  {orig_mean / triton_mean:.2f}x")
+
+        m_orig = profile_memory(selective_log_softmax_original, (logits, index))
+        m_triton = profile_memory(selective_log_softmax, (logits, index))
+        orig_peak = statistics.mean(m_orig)
+        triton_peak = statistics.mean(m_triton)
+
+        print("  MEMORY (peak overhead):")
+        print(f"    original: {fmt(m_orig, 'MB')}")
+        print(f"    triton:   {fmt(m_triton, 'MB')}")
+        print(f"    saved:    {orig_peak - triton_peak:.1f} MB")
+
+        del logits, index
+        _clean_gpu()
+
+
+def benchmark_backward():
+    print("\n" + "=" * 60)
+    print(f"FWD+BWD BENCHMARK  (warmup={WARMUP}, time={BENCH_ITERS}, mem={MEM_ITERS})")
+    print("=" * 60)
+
+    configs = [
+        (1, 2048),
+        (1, 8192),
+        (4, 4096),
+        (8, 2048),
+        (16, 2048),
+        (16, 4096),
+    ]
+
+    def fwd_bwd_original(logits, index):
+        logits.grad = None
+        out = selective_log_softmax_original(logits, index)
+        out.sum().backward()
+
+    def fwd_bwd_triton(logits, index):
+        logits.grad = None
+        out = selective_log_softmax(logits, index)
+        out.sum().backward()
+
+    for B, L in configs:
+        mem_gb = B * L * V * 2 / 1e9
+        if mem_gb > 20:
+            print(f"\n  skip B={B}, L={L} ({mem_gb:.1f} GB, need room for grads)")
+            continue
+
+        N = B * L
+        print(f"\n{'─' * 60}")
+        print(f"B={B:2d}, L={L:5d}  ({N:6d} rows, logits {mem_gb:.2f} GB)")
+        print(f"{'─' * 60}")
+
+        torch.manual_seed(42)
+        logits_orig = torch.randn(
+            B, L, V, device="cuda", dtype=torch.bfloat16, requires_grad=True
+        )
+        logits_tri = logits_orig.detach().clone().requires_grad_(True)
+        index = torch.randint(0, V, (B, L), device="cuda")
+
+        t_orig = profile_time(fwd_bwd_original, (logits_orig, index))
+        t_triton = profile_time(fwd_bwd_triton, (logits_tri, index))
+        orig_mean = statistics.mean(t_orig)
+        triton_mean = statistics.mean(t_triton)
+
+        print("  FWD+BWD TIME (ms):")
+        print(f"    original: {fmt(t_orig, 'ms')}")
+        print(f"    triton:   {fmt(t_triton, 'ms')}")
+        print(f"    speedup:  {orig_mean / triton_mean:.2f}x")
+
+        m_orig = profile_memory(fwd_bwd_original, (logits_orig, index))
+        m_triton = profile_memory(fwd_bwd_triton, (logits_tri, index))
+        orig_peak = statistics.mean(m_orig)
+        triton_peak = statistics.mean(m_triton)
+
+        print("  FWD+BWD MEMORY (peak overhead):")
+        print(f"    original: {fmt(m_orig, 'MB')}")
+        print(f"    triton:   {fmt(m_triton, 'MB')}")
+        print(f"    saved:    {orig_peak - triton_peak:.1f} MB")
+
+        del logits_orig, logits_tri, index
+        _clean_gpu()
+
+
+if __name__ == "__main__":
+    benchmark_forward()
+    benchmark_backward()

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 ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm
+    apt-get install -y --allow-change-held-packages vim curl nano zstd 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_NUM_PROC="8"
 
 RUN apt-get update && \
-    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
+    apt-get install -y --allow-change-held-packages vim curl nano zstd libnccl2 libnccl-dev ibverbs-providers ibverbs-utils infiniband-diags librdmacm-dev librdmacm1 rdmacm-utils slurm-wlm
 
 WORKDIR /workspace
 

cicd/cicd.sh

@@ -3,11 +3,23 @@ set -e
 
 python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
 
-# curl -L https://axolotl-ci.b-cdn.net/hf-cache.tar.zst | tar -xpf - -C "${HF_HOME}/hub/"  --use-compress-program unzstd --strip-components=1
-hf download "NousResearch/Meta-Llama-3-8B"
-hf download "NousResearch/Meta-Llama-3-8B-Instruct"
-hf download "microsoft/Phi-4-reasoning"
-hf download "microsoft/Phi-3.5-mini-instruct"
+set -o pipefail
+for i in 1 2 3; do
+  if curl --silent --show-error --fail -L \
+    https://axolotl-ci.b-cdn.net/hf-cache.tar.zst \
+    | tar -xpf - -C "${HF_HOME}/hub/" --use-compress-program unzstd --strip-components=1; then
+    echo "HF cache extracted successfully"
+    break
+  fi
+  echo "Attempt $i failed, cleaning up and retrying in 15s..."
+  rm -rf "${HF_HOME}/hub/"*
+  sleep 15
+done
+# hf download "NousResearch/Meta-Llama-3-8B"
+# hf download "NousResearch/Meta-Llama-3-8B-Instruct"
+# hf download "microsoft/Phi-4-reasoning"
+# hf download "microsoft/Phi-3.5-mini-instruct"
+# hf download "microsoft/Phi-3-medium-128k-instruct"
 
 # Run unit tests with initial coverage report
 pytest -v --durations=10 -n8 \

cicd/single_gpu.py

@@ -68,10 +68,6 @@ def run_cmd(cmd: str, run_folder: str):
     sp_env["AXOLOTL_DATASET_NUM_PROC"] = "8"
 
     # Propagate errors from subprocess.
-    try:
-        exit_code = subprocess.call(cmd.split(), cwd=run_folder, env=sp_env)  # nosec
-        if exit_code:
-            print(f"Command '{cmd}' failed with exit code {exit_code}")
-            return exit_code
-    except Exception as e:  # pylint: disable=broad-except
-        print(f"Command '{cmd}' failed with exception {e}")
+    exit_code = subprocess.call(cmd.split(), cwd=run_folder, env=sp_env)  # nosec
+    if exit_code:
+        raise RuntimeError(f"Command '{cmd}' failed with exit code {exit_code}")

codecov.yml

@@ -37,6 +37,7 @@ coverage:
         only_pulls: false
         flags: null
         paths: null
+        informational: true
 
 parsers:
   gcov:

docker/Dockerfile-cloud-uv

@@ -22,6 +22,7 @@ RUN apt update && \
     chmod 700 ~/.ssh && \
     printf "\n[[ -z \"\$TMUX\"  ]] && { tmux attach-session -t ssh_tmux || tmux new-session -s ssh_tmux; exit; }\n" >> ~/.bashrc && \
     printf "[ ! -z \"\$TERM\" -a -r /etc/motd ] && cat /etc/motd\n" >> ~/.bashrc && \
+    printf "source /workspace/axolotl-venv/bin/activate\n" >> ~/.bashrc && \
     chmod +x /workspace/axolotl/scripts/cloud-entrypoint.sh && \
     chmod +x /root/cloud-entrypoint.sh && \
     echo 'set-option -g history-limit 5000' >> ~/.tmux.conf

docker/Dockerfile-uv-base

@@ -36,22 +36,22 @@ RUN uv pip install packaging setuptools wheel psutil \
     && uv pip install awscli pydantic
 
 RUN if [ "$TARGETARCH" = "amd64" ]; then \
-        uv pip install --no-build-isolation "causal_conv1d @ git+https://github.com/Dao-AILab/causal-conv1d.git@main"; \
-        uv pip install "mamba_ssm @ git+https://github.com/state-spaces/mamba.git@main"; \
+        MAMBA_SKIP_CUDA_BUILD=TRUE CAUSAL_CONV1D_SKIP_CUDA_BUILD=TRUE uv pip install --no-build-isolation mamba_ssm causal_conv1d; \
     fi
 
 # Map Python version (e.g., 3.12 -> cp312)
 RUN PYTHON_CP="cp$(echo $PYTHON_VERSION | tr -d '.')" && \
     # Map PyTorch version (e.g., 2.9.1 -> torch2.9, 2.10.0 -> torch2.10)
     TORCH_TAG="torch$(echo $PYTORCH_VERSION | grep -oP '^\d+\.\d+')" && \
+    LINUX_TAG="manylinux_" && \
     # Map architecture
     case "$TARGETARCH" in \
-        amd64) ARCH_TAG="x86_64" ;; \
-        arm64) ARCH_TAG="aarch64" ;; \
+        amd64) ARCH_TAG="2_24_x86_64.manylinux_2_28_x86_64" ;; \
+        arm64) ARCH_TAG="2_34_aarch64" ;; \
         *) echo "Unsupported architecture: $TARGETARCH"; exit 1 ;; \
     esac && \
     WHL_VERSION="v0.7.16" && \
-    WHL_FILE="flash_attn-2.8.3+cu${CUDA}${TORCH_TAG}-${PYTHON_CP}-${PYTHON_CP}-linux_${ARCH_TAG}.whl" && \
+    WHL_FILE="flash_attn-2.8.3+cu${CUDA}${TORCH_TAG}-${PYTHON_CP}-${PYTHON_CP}-${LINUX_TAG}${ARCH_TAG}.whl" && \
     wget -nv "https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/${WHL_VERSION}/${WHL_FILE}" && \
     uv pip install --no-cache-dir "${WHL_FILE}" && \
     rm "${WHL_FILE}"

docs/agents/grpo.md

@@ -0,0 +1,71 @@
+# GRPO — Agent Reference
+
+Online RL with verifiable reward functions. For full config reference, async features, and scaling, see [grpo.qmd](../grpo.qmd). For vLLM setup, see [vllm_serving.qmd](../vllm_serving.qmd).
+
+## Architecture
+
+```
+Terminal 1 (GPU 0)                    Terminal 2 (GPU 1)
+┌──────────────────────┐              ┌──────────────────────────────────┐
+│  vLLM Server         │   HTTP       │  Trainer                         │
+│  Serves base model   │◄────────────►│  1. Send prompts to vLLM         │
+│  + LoRA adapter      │  /generate   │  2. Score completions (rewards)  │
+│                      │  /set_lora   │  3. Compute advantages           │
+│  Punica kernels for  │              │  4. PPO-clip gradient update     │
+│  LoRA inference      │              │  5. Sync LoRA weights to vLLM    │
+└──────────────────────┘              └──────────────────────────────────┘
+```
+
+## Components Required
+
+1. A YAML config with `rl: grpo`
+2. A reward module (Python file with reward functions)
+3. A running vLLM server (`axolotl vllm-serve config.yaml`)
+
+## Reward Function Signature
+
+```python
+def my_reward(completions, **kwargs) -> list[float]:
+    # completions[i][0]["content"] = text of i-th completion
+    # **kwargs contains dataset columns not removed by transform
+    return [score_for_each_completion]
+```
+
+Multiple rewards: `reward_funcs: [r1, r2]` with `reward_weights: [1.0, 0.5]`.
+
+## Key Async Features
+
+| Feature | Config | Purpose |
+|---------|--------|---------|
+| Async prefetch | `async_prefetch: true` | Overlap generation with training |
+| LoRA sync | `vllm_lora_sync: true` | Fast adapter sync via filesystem |
+| Streaming scoring | `streaming_partial_batch: true` | Score one group at a time |
+| Zero-adv skip | `skip_zero_advantage_batches: true` | Skip batches with no learning signal |
+| Replay buffer | `replay_buffer_size: 100` | Cache high-signal groups |
+| IS correction | `vllm_importance_sampling_correction: true` | Fix off-policy distribution shift |
+
+## Health Checks
+
+- `rewards/*/mean` > 0.15 within 20 steps (else: test reward function standalone)
+- `reward_std` > 0 on most steps (else: no learning signal)
+- `entropy` 0.05-0.5 (< 0.01 = mode collapse)
+- `grad_norm` 0.001-1.0 (> 10 = unstable, 0.0 = zero-advantage skip)
+
+See [training_stability.qmd](../training_stability.qmd) for detailed diagnostics.
+
+## File Map
+
+```
+src/axolotl/
+  cli/train.py                     # Entry point
+  cli/vllm_serve.py                # Entry point for vLLM server
+  core/trainers/grpo/
+    trainer.py                     # AxolotlGRPOTrainer
+    sampler.py                     # Sampling utilities
+  core/builders/rl.py              # HFRLTrainerBuilder — routes rl type → trainer
+  scripts/vllm_serve_lora.py       # vLLM serve script with LoRA sync support
+  utils/schemas/trl.py             # TRL config schema (all trl: options)
+
+docs/grpo.qmd                     # Full user docs: async, rewards, scaling, config reference
+docs/vllm_serving.qmd             # vLLM server modes, LoRA sync, weight sync
+```

docs/agents/preference_tuning.md

@@ -0,0 +1,121 @@
+# Preference Learning (RLHF) — Agent Reference
+
+Reference for DPO, IPO, KTO, ORPO, and SimPO. For config templates and dataset format examples, see [rlhf.qmd](../rlhf.qmd). For GRPO, see [grpo.qmd](../grpo.qmd). For EBFT, see [ebft.qmd](../ebft.qmd).
+
+## Method Overview
+
+| Method | Data Requirement | Key Idea | Best For |
+|--------|-----------------|----------|----------|
+| **DPO** | Paired (chosen + rejected) | Implicit reward via preference pairs | General alignment, most common |
+| **IPO** | Paired (chosen + rejected) | DPO with different loss (avoids overfitting) | When DPO overfits |
+| **KTO** | Unpaired (completion + binary label) | Kahneman-Tversky loss, no pairs needed | When you only have thumbs-up/down |
+| **ORPO** | Paired (chosen + rejected) | Combined SFT + preference, no ref model | Single-stage alignment, saves VRAM |
+| **SimPO** | Paired (chosen + rejected) | Length-normalized, no ref model | Simple setup, length-robust |
+
+Default: start with DPO. All methods require `sample_packing: false`.
+
+## Architecture
+
+```
+┌──────────────┐   ┌───────────────┐   ┌───────────────┐
+│ Policy Model │   │ Reference     │   │ Preference    │
+│ (trainable)  │   │ Model (frozen)│   │ Dataset       │
+└──────┬───────┘   └──────┬────────┘   └──────┬────────┘
+       └──────────┬───────┘                    │
+                  v                            │
+       Forward pass on chosen + rejected <─────┘
+                  │
+       Preference Loss (DPO/IPO/KTO/...)
+                  │
+       Backprop + Update
+
+Exception: ORPO and SimPO do NOT use a reference model (~50% less VRAM).
+```
+
+No vLLM server needed (unlike GRPO). Offline RL with pre-collected preference data.
+
+## Method Selection
+
+1. Paired preference data (chosen + rejected)?
+   - Default → `rl: dpo`
+   - Overfitting → `rl: ipo`
+   - VRAM-limited → `rl: orpo` (no ref model)
+   - Length-sensitive → `rl: simpo` (no ref model)
+2. Only binary labels (good/bad)? → `rl: kto`
+3. Single-stage training (no separate SFT)? → `rl: orpo`
+
+| | DPO | IPO | KTO | ORPO | SimPO |
+|---|---|---|---|---|---|
+| **Reference model** | Yes | Yes | Yes | No | No |
+| **VRAM overhead** | ~2x model | ~2x model | ~2x model | ~1x model | ~1x model |
+| **TRL trainer class** | DPOTrainer | DPOTrainer | KTOTrainer | ORPOTrainer | CPOTrainer |
+
+## Prompt Strategy Resolution
+
+The `type` field resolves to a Python function:
+
+```
+type: "chatml.intel"
+  → axolotl.prompt_strategies.dpo.chatml.intel(cfg, **kwargs)
+  → returns transform_fn(sample) → {"prompt", "chosen", "rejected"}
+
+type: "chat_template.default"
+  → axolotl.prompt_strategies.dpo.chat_template.default(cfg, dataset_idx, **kwargs)
+
+type: {"field_prompt": "prompt", ...}   (dict)
+  → axolotl.prompt_strategies.dpo.user_defined.default(...)
+```
+
+Module base: `axolotl.prompt_strategies.{rl_method}` — replace `dpo` with `kto` or `orpo`.
+
+## Healthy Training Indicators
+
+| Metric | Healthy Range | Problem |
+|--------|--------------|---------|
+| `train/loss` | Decreasing, 0.3-0.7 | Flat or increasing = broken data or too high LR |
+| `rewards/chosen` | Increasing | Flat = model not learning preferences |
+| `rewards/rejected` | Decreasing | Increasing = model prefers wrong responses |
+| `rewards/margins` | Positive and increasing | Negative = prefers rejected over chosen |
+| `rewards/accuracies` | > 0.5, toward 0.7+ | < 0.5 = worse than random |
+| `logps/rejected` | Decreasing | Increasing = reward hacking |
+| `grad_norm` | 0.01 - 10.0 | > 100 = exploding gradients |
+
+Method-specific: DPO/IPO watch `rewards/margins`; KTO loss is noisier; ORPO monitor SFT + odds ratio components; SimPO check length-normalized reward separation.
+
+## Known Issues
+
+| Issue | Fix |
+|-------|-----|
+| Sample packing crash | Set `sample_packing: false` (required for all preference methods) |
+| KTO `KeyError: 'label'` | Ensure dataset has boolean `label` column |
+| ORPO/KTO `KeyError` during tokenization | Add `remove_unused_columns: false` |
+| ORPO template not applied | ORPO requires explicit `chat_template` setting |
+| OOM with ref model (DPO/IPO/KTO) | Use LoRA/QLoRA, or switch to ORPO/SimPO (no ref model) |
+| IPO + label_smoothing | Do not set `dpo_label_smoothing` when `rl: ipo` |
+
+Full troubleshooting: [training_stability.qmd](../training_stability.qmd)
+
+## File Map
+
+```
+src/axolotl/
+  core/trainers/dpo/              # DPO trainer, args, strategy
+  core/builders/rl.py             # HFRLTrainerBuilder — routes rl type → trainer class
+  core/training_args.py           # AxolotlKTOConfig, AxolotlORPOConfig, AxolotlCPOConfig
+  prompt_strategies/
+    dpo/                          # DPO/IPO/SimPO dataset strategies
+      chat_template.py            # chat_template.default, chat_template.argilla_chat
+      chatml.py                   # chatml.default/intel/icr/argilla_chat/prompt_pairs/ultra
+      llama3.py                   # llama3 variants (same subtypes as chatml)
+      user_defined.py             # Custom field mapping
+      passthrough.py              # No transform
+    kto/                          # KTO dataset strategies (chatml, llama3, user_defined)
+    orpo/                         # ORPO dataset strategies (chat_template.argilla)
+  utils/schemas/enums.py          # RLType enum (dpo, ipo, kto, orpo, simpo, grpo, gdpo, ebft)
+  utils/schemas/config.py         # All rl/dpo/kto/orpo/simpo config fields
+
+docs/rlhf.qmd                    # Full user docs: all dataset formats, config templates
+docs/choosing_method.qmd          # SFT vs DPO vs GRPO decision guide
+examples/qwen2/dpo.yaml           # DPO example
+examples/llama-3/qlora-1b-kto.yaml  # KTO example
+```

docs/agents/pretraining.md

@@ -0,0 +1,75 @@
+# Pretraining / Continual Pretraining — Agent Reference
+
+Train on raw text with no input masking. Two approaches depending on dataset size.
+
+## When to Use
+
+- Continual pretraining on domain-specific corpora
+- Adapting a base model to a new language or domain before fine-tuning
+- Pretraining-style data where the entire text is the training signal
+
+## Choosing an Approach
+
+| | Non-streaming (`type: completion`) | Streaming (`pretraining_dataset`) |
+|---|---|---|
+| **Dataset size** | Fits in memory | Too large to fit in memory |
+| **Tokenization** | Pre-tokenized before training | On-demand during training |
+| **Config key** | `datasets:` | `pretraining_dataset:` |
+| **Long text handling** | Splits texts exceeding `sequence_len` | Concatenates into fixed-length sequences |
+| **Benefit** | Can preprocess on CPU, transfer to GPU | Start training immediately, no preprocessing |
+
+## Non-Streaming: `type: completion`
+
+For smaller datasets that fit in memory. Pre-tokenizes the entire dataset.
+
+```yaml
+datasets:
+  - path: my_corpus
+    type: completion
+    # field: text              # Column name (default: "text")
+```
+
+## Streaming: `pretraining_dataset`
+
+For large corpora. Streams data on-demand without loading everything into memory.
+
+```yaml
+pretraining_dataset:
+  - path: HuggingFaceFW/fineweb-edu
+    type: pretrain
+    text_column: text
+    split: train
+
+max_steps: 1000                          # Required — axolotl can't infer dataset size
+streaming_multipack_buffer_size: 10000   # Buffer for sample packing
+pretrain_multipack_attn: true            # Prevent cross-attention between packed samples
+```
+
+`max_steps` is required for streaming — one step = `sequence_len * micro_batch_size * gradient_accumulation_steps * num_gpus` tokens.
+
+Full streaming docs: [streaming.qmd](../streaming.qmd)
+
+## Dataset Format
+
+```json
+{"text": "The complete document text goes here."}
+```
+
+## Key Settings
+
+- `sample_packing: true` + `pad_to_sequence_len: true` — pack documents into fixed-length sequences
+- `flash_attention: true` — required for sample packing
+- No adapter — typically full fine-tune for pretraining
+- `train_on_inputs: true` — default for completion (all tokens trained on)
+
+## File Map
+
+```
+src/axolotl/
+  prompt_strategies/completion.py    # Non-streaming: completion prompt strategy (no masking)
+  utils/data/sft.py                  # Non-streaming: dataset loading and processing
+  utils/data/streaming.py            # Streaming: encode_streaming(), wrap_streaming_dataset()
+  utils/schemas/config.py            # Config fields: pretraining_dataset, pretrain_multipack_attn, etc.
+
+examples/streaming/pretrain.yaml     # Full streaming pretraining example config
+```

docs/agents/reward_modelling.md

@@ -0,0 +1,48 @@
+# Reward Modelling — Agent Reference
+
+Train models to score responses for use as reward signals in RL. For full docs, see [reward_modelling.qmd](../reward_modelling.qmd).
+
+## Types
+
+### Outcome Reward Models (ORM)
+
+Train a classifier to predict preference over entire interactions. Uses `AutoModelForSequenceClassification`.
+
+```yaml
+base_model: google/gemma-2-2b
+model_type: AutoModelForSequenceClassification
+num_labels: 1
+reward_model: true
+chat_template: gemma
+datasets:
+  - path: argilla/distilabel-intel-orca-dpo-pairs
+    type: bradley_terry.chat_template
+```
+
+Dataset format: `{"system": "...", "input": "...", "chosen": "...", "rejected": "..."}`
+
+### Process Reward Models (PRM)
+
+Train a token classifier to score each reasoning step. Uses `AutoModelForTokenClassification`.
+
+```yaml
+base_model: Qwen/Qwen2.5-3B
+model_type: AutoModelForTokenClassification
+num_labels: 2
+process_reward_model: true
+datasets:
+  - path: trl-lib/math_shepherd
+    type: stepwise_supervised
+```
+
+Dataset format: see [stepwise_supervised.qmd](../dataset-formats/stepwise_supervised.qmd).
+
+## File Map
+
+```
+src/axolotl/
+  core/builders/causal.py                    # Handles reward_model flag in trainer builder
+  prompt_strategies/bradley_terry/           # Bradley-Terry prompt strategies
+  prompt_strategies/stepwise_supervised.py   # PRM dataset strategy
+  utils/schemas/config.py                    # reward_model, process_reward_model config fields
+```

docs/agents/sft.md

@@ -0,0 +1,115 @@
+# SFT — Agent Reference
+
+Supervised fine-tuning pipeline reference. For config templates and dataset format examples, see [getting-started.qmd](../getting-started.qmd) and [dataset-formats/](../dataset-formats/).
+
+## Architecture
+
+```
+YAML Config → axolotl train config.yaml
+
+  1. Load base model (+ quantization if QLoRA/8-bit)
+  2. Apply adapter layers (LoRA/QLoRA) if configured
+  3. Load + tokenize dataset(s)
+     - Apply prompt template (chat_template / alpaca / custom)
+     - Mask inputs (train_on_inputs: false)
+     - Pack samples into sequences (sample_packing: true)
+  4. Training loop (HuggingFace Trainer)
+     - forward → loss → backward → optimizer step → lr scheduler step
+  5. Save model / adapter weights + tokenizer
+
+Multi-GPU: FSDP or DeepSpeed shards model across GPUs automatically.
+```
+
+## Components Required
+
+1. A YAML config — model, dataset(s), adapter settings, hyperparameters
+2. A dataset — HuggingFace Hub, local JSONL/JSON/Parquet, or S3/GCS path
+3. (Optional) A custom prompt strategy — for non-standard dataset formats
+
+No external server processes needed (unlike GRPO which requires vLLM).
+
+## Dataset Format Decision Tree
+
+```
+Is your data in chat/message format?
+  ├─ YES: OpenAI message format (role/content)?
+  │   ├─ YES ──────────────────────> type: chat_template  (recommended)
+  │   └─ NO (custom field names) ──> type: chat_template + message_property_mappings
+  └─ NO: Instruction/response pairs?
+      ├─ YES ──> type: alpaca       (instruction, input, output)
+      └─ NO: Raw text?
+          ├─ YES with segments ─────> type: input_output  (template-free masking)
+          └─ YES continuous ────────> type: completion     (pretraining-style)
+```
+
+Full format specs: [dataset-formats/](../dataset-formats/)
+
+## Model Size to Adapter Choice
+
+| Model Size | LoRA | QLoRA (4-bit) | Full Fine-Tune | VRAM (approx) |
+|-----------|------|---------------|----------------|---------------|
+| 1-3B | Preferred | Low-budget option | Single GPU OK | 8-16 GB (LoRA) |
+| 7-8B | Preferred | Good balance | Needs multi-GPU | 16-24 GB (LoRA) |
+| 13-14B | Preferred | Good balance | Multi-GPU required | 24-40 GB (LoRA) |
+| 30-70B | LoRA or QLoRA | Preferred for single GPU | Multi-node | 40-80 GB (QLoRA) |
+
+## Hyperparameter Ranges
+
+| Parameter | LoRA | QLoRA | Full FT |
+|-----------|------|-------|---------|
+| `learning_rate` | 1e-4 to 3e-4 | 1e-4 to 3e-4 | 1e-5 to 5e-5 |
+| `lora_r` | 16-64 | 16-64 | N/A |
+| `lora_alpha` | 1-2x `lora_r` | 1-2x `lora_r` | N/A |
+| `micro_batch_size` | 2-8 | 2-4 | 1-2 |
+| `gradient_accumulation_steps` | 2-8 | 4-16 | 4-16 |
+| `num_epochs` | 1-3 | 1-3 | 1-3 |
+| `optimizer` | `adamw_8bit` | `adamw_bnb_8bit` | `adamw_torch_fused` |
+
+Effective batch = micro_batch * grad_accum * num_gpus. Lower LR for larger models.
+
+## Healthy Training Indicators
+
+| Metric | Healthy | Problem |
+|--------|---------|---------|
+| `train_loss` | Decreasing, starting ~2-4 for chat models | Flat or increasing from step 1 — data or LR issue |
+| `eval_loss` | Decreasing, tracks train_loss | Increasing while train_loss decreases — overfitting |
+| `grad_norm` | 0.1-10, relatively stable | Spikes >100 — instability. 0.0 — frozen weights |
+| `learning_rate` | Follows scheduler curve | Flat or NaN — config issue |
+
+Watch for: loss never decreasing (check `train_on_inputs`, dataset, LR), loss goes to 0 quickly (overfitting), eval_loss diverging (reduce epochs, add regularization). See [training_stability.qmd](../training_stability.qmd).
+
+## Known Issues
+
+| Issue | Fix |
+|-------|-----|
+| OOM during training | Reduce `micro_batch_size`, enable `gradient_checkpointing`, reduce `sequence_len` |
+| `sample_packing` + SDPA + bf16 = 0.0 loss | Use `flash_attention: true` or disable `sample_packing` |
+| Missing chat template error | Set `chat_template: chatml` explicitly |
+| Label masking wrong | Run `axolotl preprocess config.yaml --debug` and inspect labels |
+| Loss NaN | Use `bf16: auto`, lower LR, check data for empty samples |
+| Tokenizer pad token / infinite loss | Set `special_tokens: pad_token: "<\|end_of_text\|>"` |
+| FSDP save hangs | Use `fsdp_state_dict_type: FULL_STATE_DICT` |
+| DeepSpeed CheckpointError | Set `use_reentrant: true` in `gradient_checkpointing_kwargs` |
+
+Full troubleshooting: [training_stability.qmd](../training_stability.qmd), [debugging.qmd](../debugging.qmd)
+
+## File Map
+
+```
+src/axolotl/
+  cli/train.py                     # Entry point for `axolotl train`
+  cli/preprocess.py                # Entry point for `axolotl preprocess`
+  core/builders/causal.py          # HFCausalTrainerBuilder — wires config → SFT trainer
+  core/trainers/base.py            # AxolotlTrainer — base trainer class
+  core/trainers/mixins/            # Packing, optimizer, scheduler, checkpoints
+  prompt_strategies/               # Format handlers: chat_template, alpaca, completion, input_output
+  utils/schemas/config.py          # AxolotlInputConfig — main config schema
+  utils/schemas/datasets.py        # SFTDataset, DatasetConfig
+  utils/schemas/peft.py            # LoraConfig — LoRA parameters
+  integrations/liger/              # Liger kernel plugin
+
+examples/llama-3/                  # LoRA, QLoRA, full FT example configs
+docs/getting-started.qmd           # Quickstart with config templates
+docs/optimizations.qmd             # Flash attention, gradient checkpointing, sample packing
+docs/multi-gpu.qmd                 # FSDP and DeepSpeed setup
+```

docs/attention.qmd

@@ -13,9 +13,10 @@ sdp_attention: true
 
 For more details: [PyTorch docs](https://docs.pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html)
 
-## Flash Attention 2
+## Flash Attention
 
-Uses efficient kernels to compute attention.
+Axolotl supports Flash Attention 2, 3, and 4. The best available version is used automatically
+based on your installed packages and GPU.
 
 ```yaml
 flash_attention: true
@@ -23,11 +24,9 @@ flash_attention: true
 
 For more details: [Flash Attention](https://github.com/Dao-AILab/flash-attention/)
 
-### Nvidia
+### Flash Attention 2
 
-Requirements: Ampere, Ada, or Hopper GPUs
-
-Note: For Turing GPUs or lower, please use other attention methods.
+Requirements: Ampere, Ada, or Hopper GPUs (Turing or lower not supported)
 
 ```bash
 pip install flash-attn --no-build-isolation
@@ -35,11 +34,12 @@ pip install flash-attn --no-build-isolation
 
 ::: {.callout-tip}
 
-If you get `undefined symbol` while training, ensure you installed PyTorch prior to Axolotl. Alternatively, try reinstall or downgrade a version.
+If you get `undefined symbol` while training, ensure you installed PyTorch prior to Axolotl.
+Alternatively, try reinstall or downgrade a version.
 
 :::
 
-#### Flash Attention 3
+### Flash Attention 3
 
 Requirements: Hopper only and CUDA 12.8 (recommended)
 
@@ -50,6 +50,44 @@ cd flash-attention/hopper
 python setup.py install
 ```
 
+### Flash Attention 4
+
+Requirements: Hopper or Blackwell GPUs
+
+```bash
+pip install flash-attn-4
+```
+
+Or from source:
+
+```bash
+git clone https://github.com/Dao-AILab/flash-attention.git
+cd flash-attention/flash_attn/cute
+
+pip install -e .
+
+# FA2's flash_attn package includes a cute/ stub that shadows FA4.
+# Remove it so Python can find the real FA4 module:
+rm -r $(python -c "import flash_attn; print(flash_attn.__path__[0])")/cute
+```
+
+::: {.callout-note}
+
+**Hopper (SM90) users**: The backward kernel is not yet included in the pip package. To use FA4
+for training on Hopper, install from source using the instructions above.
+
+:::
+
+::: {.callout-warning}
+
+FA4 only supports head dimensions up to 128 (`d ≤ 128`). The DeepSeek shape `(192, 128)` is
+also supported but only on Blackwell. Axolotl automatically detects incompatible head dimensions
+and falls back to FA2/3.
+
+:::
+
+For more details: [flash-attention/flash_attn/cute](https://github.com/Dao-AILab/flash-attention/tree/main/flash_attn/cute)
+
 ### AMD
 
 Requirements: ROCm 6.0 and above.

docs/choosing_method.qmd

@@ -0,0 +1,206 @@
+---
+title: "Which Fine-Tuning Method Should I Use?"
+description: "A decision guide for choosing the right fine-tuning method, adapter, and hardware configuration in Axolotl."
+format:
+  html:
+    toc: true
+    toc-depth: 3
+    number-sections: true
+execute:
+  enabled: false
+---
+
+## Overview {#sec-overview}
+
+Axolotl supports four broad categories of fine-tuning, each suited to different data types, objectives, and resource constraints.
+
+| Method | What It Does | Data You Need |
+|--------|-------------|---------------|
+| **Supervised Fine-Tuning (SFT)** | Teaches the model to produce specific outputs given inputs | Input-output pairs (instructions, conversations, completions) |
+| **Preference Learning (DPO/KTO/ORPO)** | Steers the model toward preferred outputs and away from dispreferred ones | Chosen/rejected response pairs (DPO, ORPO) or binary labels (KTO) |
+| **Reinforcement Learning (GRPO)** | Optimizes the model against a reward signal through online generation | A reward function (code or model-based) and a prompt dataset |
+| **Reward Modeling** | Trains a model to score responses, for use as a reward signal in RL | Preference pairs ranked by quality |
+
+Each method is configured through a YAML file with `rl: <method>` (or omitted for SFT). All methods support LoRA, QLoRA, and full fine-tuning unless otherwise noted.
+
+## Decision Tree {#sec-decision-tree}
+
+Use the following flowchart to choose your method. Start at the top and follow the path that matches your situation.
+
+```
+Do you have a reward function (code-based or model-based)?
+├── YES
+│   └── Use GRPO (rl: grpo)
+│       The model generates its own completions and learns from reward scores.
+│       Best for: math, code, reasoning, tasks with verifiable answers.
+│       See: rlhf.qmd#grpo
+│
+└── NO
+    │
+    Do you have preference pairs (chosen vs. rejected responses)?
+    ├── YES
+    │   │
+    │   Are they paired (same prompt, one chosen, one rejected)?
+    │   ├── YES → Use DPO (rl: dpo)
+    │   │         Direct optimization without a separate reward model.
+    │   │         See: rlhf.qmd#dpo
+    │   │
+    │   └── NO (only binary good/bad labels)
+    │       └── Use KTO (rl: kto)
+    │           Works with unpaired preference data.
+    │           See: rlhf.qmd#kto
+    │
+    └── NO
+        │
+        Do you have input-output examples?
+        ├── YES → Use SFT
+        │         The simplest and most common method.
+        │         See: getting-started.qmd
+        │
+        └── NO
+            └── You need to create training data first.
+                Consider generating preference pairs with an LLM judge,
+                or writing a reward function for GRPO.
+```
+
+::: {.callout-tip}
+**When in doubt, start with SFT.** It is the most straightforward method and works well for most tasks. You can always move to preference learning or RL later to further refine behavior.
+:::
+
+### Method Comparison at a Glance
+
+| Criterion | SFT | DPO | KTO | GRPO |
+|-----------|-----|-----|-----|------|
+| Data complexity | Low (input-output pairs) | Medium (preference pairs) | Medium (binary labels) | Low (prompts + reward code) |
+| Compute cost | Low | Medium | Medium | High (requires vLLM server) |
+| Learning signal | Supervised | Contrastive | Contrastive | Online reward |
+| Online generation | No | No | No | Yes |
+| Reward model needed | No | No | No | No (uses reward functions) |
+| Best for | Task adaptation, instruction following | Safety, style alignment | Unpaired preference data | Reasoning, math, code |
+
+::: {.callout-note}
+**ORPO** is an alternative to DPO that combines SFT and preference optimization in a single training stage, removing the need for a separate SFT step. Configure with `rl: orpo`. See [rlhf.qmd](rlhf.qmd) for details.
+:::
+
+## Adapter Selection {#sec-adapter-selection}
+
+Once you have chosen a method, decide how to apply the parameter updates. The three main options trade off VRAM usage against model quality.
+
+### QLoRA
+
+- **How it works**: The base model is loaded in 4-bit (NF4) quantization. Small low-rank adapter matrices are trained in higher precision on top.
+- **VRAM savings**: Roughly 4x reduction in model memory compared to full fine-tuning.
+- **Quality**: Slight degradation due to quantization noise, but often negligible for task-specific fine-tuning.
+- **When to use**: When your GPU cannot fit the model in full precision, or when you want fast experimentation.
+
+```yaml
+adapter: qlora
+load_in_4bit: true
+lora_r: 32
+lora_alpha: 64
+lora_target_linear: true
+```
+
+### LoRA
+
+- **How it works**: The base model is loaded at full precision (or 8-bit). Low-rank adapter matrices are trained alongside.
+- **VRAM savings**: Roughly 2-3x reduction compared to full fine-tuning (model weights are frozen, only adapters + optimizer states for adapters are stored).
+- **Quality**: Very close to full fine-tuning for most tasks, especially with higher rank values.
+- **When to use**: When you have enough VRAM for the base model but not for full optimizer states.
+
+```yaml
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_target_linear: true
+```
+
+::: {.callout-tip}
+For GRPO training, LoRA is strongly recommended. The vLLM server needs to sync weights from the trainer, and LoRA sync (`trl.vllm_lora_sync: true`) is far more efficient than syncing full merged weights. See [vLLM Serving](vllm_serving.qmd) for details.
+:::
+
+### Full Fine-Tuning
+
+- **How it works**: All model parameters are updated during training. No adapters.
+- **VRAM savings**: None. Requires memory for model weights, gradients, and optimizer states (roughly 4x model size in bf16 with AdamW).
+- **Quality**: Highest potential quality, especially for large distribution shifts.
+- **When to use**: When you have ample GPU memory or multi-GPU setups, and need maximum performance. Also required for pre-training.
+
+```yaml
+# No adapter or load_in_* lines needed
+micro_batch_size: 1
+gradient_accumulation_steps: 16
+```
+
+### Quick Comparison
+
+| | QLoRA | LoRA | Full |
+|---|---|---|---|
+| Trainable params | ~0.1-1% | ~0.1-1% | 100% |
+| Model memory | ~25% of full | ~50-100% of full | 100% |
+| Optimizer memory | Tiny (adapters only) | Tiny (adapters only) | 2x model size (AdamW) |
+| Training speed | Slower (dequantization overhead) | Baseline | Faster per-step (no adapter overhead) |
+| Inference | Merge or serve with adapter | Merge or serve with adapter | Direct |
+| Multi-GPU required? | Rarely | For 13B+ models | For 7B+ models |
+
+## Hardware Mapping {#sec-hardware-mapping}
+
+The tables below provide approximate GPU memory requirements. Actual usage depends on context length, batch size, and optimizer choice.
+
+### SFT / Preference Learning
+
+| Model Size | QLoRA (4-bit) | LoRA (bf16) | Full (bf16 + AdamW) |
+|------------|--------------|-------------|---------------------|
+| 1-3B | 6-8 GB | 8-12 GB | 24-32 GB |
+| 7-8B | 10-14 GB | 16-24 GB | 60-80 GB |
+| 13-14B | 16-20 GB | 28-40 GB | 120+ GB |
+| 30-34B | 24-32 GB | 64-80 GB | 2-4x 80 GB |
+| 70-72B | 40-48 GB | 2x 80 GB | 4-8x 80 GB |
+
+::: {.callout-important}
+These estimates assume a short context length (512-2048 tokens) and micro_batch_size of 1-2. Longer sequences and larger batches increase memory significantly due to activations. Use [gradient checkpointing](gradient_checkpointing.qmd) to reduce activation memory at the cost of ~30% slower training.
+:::
+
+### GRPO (RL Training)
+
+GRPO requires additional GPU(s) for the vLLM generation server. Plan for at least two GPUs: one for training, one for vLLM.
+
+| Model Size | Training GPU (LoRA, bf16) | vLLM GPU | Total GPUs |
+|------------|--------------------------|----------|------------|
+| 0.5-3B | 1x 24 GB | 1x 24 GB | 2x 24 GB |
+| 7-8B | 1x 80 GB | 1x 80 GB | 2x 80 GB |
+| 13-14B | 1-2x 80 GB | 1-2x 80 GB | 2-4x 80 GB |
+| 30-72B | 2-4x 80 GB (FSDP/DeepSpeed) | 2-4x 80 GB (tensor parallel) | 4-8x 80 GB |
+
+::: {.callout-tip}
+For single-GPU GRPO, use `vllm_mode: colocate` with `vllm_enable_sleep_mode: true`. The vLLM engine shares the GPU and offloads VRAM when not generating. This works for smaller models (up to ~3B on a 24 GB GPU) but is slower than the two-GPU server mode.
+:::
+
+### Multi-GPU Threshold
+
+You need multi-GPU training when:
+
+- **Full fine-tuning** of models 7B+ (use FSDP or DeepSpeed ZeRO)
+- **LoRA** of models 30B+ (or 13B+ with long contexts)
+- **GRPO** almost always (separate vLLM server), unless using colocate mode
+
+See [Multi-GPU Training](multi-gpu.qmd) for FSDP and DeepSpeed configuration.
+
+## Quick Links {#sec-quick-links}
+
+| Method | Config Key | Documentation | Example Config |
+|--------|-----------|---------------|----------------|
+| SFT | *(default, no `rl:` key)* | [Getting Started](getting-started.qmd) | `examples/llama-3/lora-1b.yml` |
+| DPO | `rl: dpo` | [RLHF - DPO](rlhf.qmd#dpo) | See rlhf.qmd |
+| KTO | `rl: kto` | [RLHF - KTO](rlhf.qmd#kto) | See rlhf.qmd |
+| ORPO | `rl: orpo` | [RLHF - ORPO](rlhf.qmd#orpo) | See rlhf.qmd |
+| GRPO | `rl: grpo` | [RLHF - GRPO](rlhf.qmd#grpo), [vLLM Serving](vllm_serving.qmd) | See rlhf.qmd |
+| Reward Modeling | `rl: reward_trainer` | [Reward Modelling](reward_modelling.qmd) | See reward_modelling.qmd |
+
+### Related Guides
+
+- [Configuration Reference](config-reference.qmd) -- Full list of all config options
+- [Dataset Formats](dataset-formats) -- How to structure your training data
+- [Optimizations](optimizations.qmd) -- Flash attention, gradient checkpointing, mixed precision
+- [Multi-GPU Training](multi-gpu.qmd) -- FSDP and DeepSpeed setup
+- [vLLM Serving](vllm_serving.qmd) -- Setting up vLLM for GRPO training

docs/dataset-formats/index.qmd

@@ -22,89 +22,46 @@ For `pretraining_dataset:` specifically, please refer to the [Pre-training secti
 
 ## Pre-training
 
-When aiming to train on large corpora of text datasets, pre-training is your go-to choice. Due to the size of these datasets, downloading the entire-datasets before beginning training would be prohibitively time-consuming. Axolotl supports [streaming](https://huggingface.co/docs/datasets/en/stream) to only load batches into memory at a time.
-
-A sample format for a pre-training dataset is as follows:
+Pre-training trains on raw text corpora with no input masking. The dataset format is simple:
 
 ```json
 {"text": "first row"}
 {"text": "second row"}
-...
 ```
 
-It is typically recommended to save your dataset as `.jsonl` due to its flexibility and simplicity.
+Axolotl supports two approaches:
 
-Axolotl supports loading from a Hugging Face hub repo or from local files.
+### Streaming (large datasets)
 
-### Pre-training from Hugging Face hub datasets
-
-As an example, to train using a Hugging Face dataset `hf_org/name`, you can pass the following config:
-
-```yaml
-pretraining_dataset: hf_org/name
-```
-
-### Pre-training from local dataset files
-
-Given a few corpus files: `A.jsonl`, `B.jsonl`, and `C.jsonl`, your config will look like the below:
+For large corpora that don't fit in memory, use `pretraining_dataset` with [streaming](../streaming.qmd). Data is tokenized on-demand during training.
 
 ```yaml
 pretraining_dataset:
-  - path: json
-    data_files:
-      - A.jsonl
-      - B.jsonl
-      - C.jsonl
-```
-
-While we recommend `.jsonl`, you can also use the other formats (`csv`, `parquet`, `arrow`, `SQL`, `Webdataset`) that are supported by [`Dataset.load_dataset`](https://huggingface.co/docs/datasets/loading#local-and-remote-files)
-
-### Pre-training without streaming
-
-In the case that the dataset is small and can be loaded entirely into memory, another approach to running pre-training is to use the `completion` format. This would mean that the entire dataset is pre-tokenized instead of on-demand in streaming.
-
-One benefit of this is that the tokenization can be performed separately on a CPU-only machine, and then transferred to a GPU machine for training to save costs.
-
-From Hugging Face:
-
-```yaml
-datasets:
-  - path: hf_org/name
-    type: completion
-```
-
-From local files:
-
-```yaml
-datasets:
-  - path: A.jsonl
-    type: completion
-
-  - path: B.jsonl
-    type: completion
+  - path: HuggingFaceFW/fineweb-edu
+    type: pretrain
+    text_column: text
+    split: train
 ```
 
 ::: {.callout-important}
-For `completion` only, Axolotl would split texts if it exceeds the context length into multiple smaller prompts. If you are interested in having this for `pretraining_dataset` too, please let us know or help make a PR!
+Streaming requires `max_steps` in your config — Axolotl cannot infer the dataset size. One step = `sequence_len * micro_batch_size * gradient_accumulation_steps * num_gpus` tokens.
 :::
 
-### Pre-training dataset configuration tips
+See [Streaming Datasets](../streaming.qmd) for full configuration details.
 
-#### Setting max_steps
+### Non-streaming (smaller datasets)
 
-When using streaming for large datasets, Axolotl does not know in advance how large the dataset is and does not know when to stop.
+For datasets that fit in memory, use `type: completion` under `datasets:`. The entire dataset is pre-tokenized before training, which can be done on a CPU-only machine.
 
-Therefore, it is necessary to set `max_steps: int` in your config for pre-training to run, so that Axolotl knows when to stop training.
+```yaml
+datasets:
+  - path: my_corpus
+    type: completion
+```
 
-One step is equal to `sequence_len * micro_batch_size * gradient_accumulation_steps * total_num_gpus` tokens.
-
-#### Group_by_length
-
-It is recommended to leave this off if downloading from Hugging Face hub as it would download the entire dataset which can be very large.
-
-### Reference
-
-Please see docs [here](pretraining.qmd).
+::: {.callout-note}
+With `completion`, texts exceeding `sequence_len` are split into multiple samples automatically.
+:::
 
 ## Supervised fine-tuning (SFT)
 

docs/dataset-formats/pretraining.qmd

@@ -4,29 +4,9 @@ description: Data format for a pre-training completion task.
 order: 1
 ---
 
-For pretraining, there is no prompt template or roles.  The only required field is `text`:
-
-```{.json filename="data.jsonl"}
-{"text": "first row"}
-{"text": "second row"}
-...
-```
-
-:::{.callout-note}
-
-### Streaming is recommended for large datasets
-
-Axolotl usually loads the entire dataset into memory. This will be challenging for large datasets. Use the following config to enable streaming:
-
-```{.yaml filename="config.yaml"}
-pretraining_dataset:
-  - name:
-    path:
-    split:
-    text_column: # column in dataset with the data, usually `text`
-    type: pretrain
-    trust_remote_code:
-    skip: # number of rows of data to skip over from the beginning
-```
+::: {.callout-note}
+Pre-training documentation has been consolidated:
 
+- **Streaming pretraining** (large datasets): See [Streaming Datasets](../streaming.qmd#pretraining-with-streaming)
+- **Non-streaming pretraining** (`type: completion`): See [Dataset Formats](index.qmd#pre-training)
 :::

docs/debugging.qmd

@@ -6,6 +6,10 @@ description: How to debug Axolotl
 
 This document provides some tips and tricks for debugging Axolotl.  It also provides an example configuration for debugging with VSCode.  A good debugging setup is essential to understanding how Axolotl code works behind the scenes.
 
+::: {.callout-tip}
+For training-specific debugging (loss spikes, NaN gradients, OOM errors, RL training stability), see [Training Stability & Debugging](training_stability.qmd).
+:::
+
 ## Table of Contents
 
 - [General Tips](#general-tips)
@@ -85,7 +89,7 @@ If you developing on a remote host, you can easily use VSCode to debug remotely.
 
 The easiest way to get started is to modify the [.vscode/launch.json](../.vscode/launch.json) file in this project.  This is just an example configuration, so you may need to modify or copy it to suit your needs.
 
-For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 accelerate launch -m axolotl.cli.train dev_chat_template.yml`, you would use the below configuration[^1].  Note that we add additional flags that override the axolotl config and incorporate the tips above (see the comments). We also set the working directory to `devtools` and set the `env` variable `HF_HOME` to a temporary folder that is later partially deleted.  This is because we want to delete the HF dataset cache before each run in order to ensure that the data preprocessing code is run from scratch.
+For example, to mimic the command `cd devtools && CUDA_VISIBLE_DEVICES=0 axolotl train dev_chat_template.yml`, you would use the below configuration[^1].  Note that we add additional flags that override the axolotl config and incorporate the tips above (see the comments). We also set the working directory to `devtools` and set the `env` variable `HF_HOME` to a temporary folder that is later partially deleted.  This is because we want to delete the HF dataset cache before each run in order to ensure that the data preprocessing code is run from scratch.
 
 ```json
 // .vscode/launch.json
@@ -242,6 +246,6 @@ style="border-radius: 10px; display: block; margin: auto;" width="560" height="3
 </div>
 <br>
 
-[^1]: The config actually mimics the command `CUDA_VISIBLE_DEVICES=0 python -m accelerate.commands.launch -m axolotl.cli.train devtools/chat_template.yml`, but this is the same thing.
+[^1]: The VSCode config uses `accelerate.commands.launch` as the Python module entry point, which is what `axolotl train` invokes under the hood.
 
 [^2]: Many of the below flags are recommended best practices by Nvidia when using nvidia-container-toolkit.  You can read more about these flags [here](https://docs.nvidia.com/deeplearning/frameworks/user-guide/index.html).

docs/ebft.qmd

@@ -0,0 +1,556 @@
+---
+title: "EBFT Training"
+description: "Energy-Based Fine-Tuning uses feature-matching rewards from internal representations to train language models without external reward functions."
+order: 9
+back-to-top-navigation: true
+toc: true
+toc-expand: 2
+toc-depth: 4
+---
+
+## Overview
+
+Energy-Based Fine-Tuning (EBFT) is a training method that optimizes language models by matching the **internal feature representations** of generated text to those of ground-truth completions. Instead of relying on external reward models or hand-crafted reward functions, EBFT extracts hidden states from intermediate layers of a frozen copy of the model and uses cosine similarity between generated and reference features as the reward signal.
+
+Paper: ["Matching Features, Not Tokens: Energy-Based Fine-Tuning of Language Models"](https://arxiv.org/abs/2603.12248) (Jelassi et al., 2026)
+
+### How EBFT Differs from Other RL Methods
+
+| Method | Reward Signal | Requires | Best For |
+|--------|--------------|----------|----------|
+| **GRPO** | External reward function(s) | Custom reward code or reward model | Tasks with verifiable answers (math, code) |
+| **DPO** | Preference pairs (chosen vs rejected) | Paired preference data | Alignment with human preferences |
+| **EBFT** | Feature similarity to ground truth | Ground-truth completions | Any task with reference outputs |
+
+EBFT's key advantage is that it needs only ground-truth completions -- no reward engineering, no preference annotation, and no reward model training. The model's own internal representations serve as the reward signal. This makes it particularly effective for:
+
+- Code generation (match features of known-good solutions)
+- Instruction following with reference outputs
+- Continual pretraining on unstructured text (strided mode)
+- Multi-turn dialogue with reference conversations
+
+### Reward Formulation
+
+The EBFT reward for each generated completion is:
+
+```
+reward = alignment_coef * cosine_similarity(gen_features, gt_features)
+       - diversity_coef * mean_pairwise_similarity(gen_features)
+```
+
+- **Alignment**: How closely the generated output's internal representations match the ground truth. Higher is better.
+- **Diversity**: Penalizes generated samples that are too similar to each other (prevents mode collapse). Lower is better.
+- **CFM loss** (Cross-Feature Matching): Tracks `||mean(gen_features) - gt_features||^2` as a diagnostic. This is the quantity that EBFT ultimately minimizes.
+
+## Modes
+
+EBFT supports three operational modes, each suited to different use cases.
+
+### Structured Mode (Sync)
+
+Uses vLLM on a separate GPU for generation, with sequential generate-score-train steps. This is the simplest mode and recommended for getting started.
+
+```
+GPU 0: vLLM Server (generates completions, receives weight syncs)
+GPU 1: Trainer (feature extraction, reward computation, GRPO training)
+```
+
+**When to use**: Standard instruction-following or QA datasets where you have prompt/completion pairs. Requires 2 GPUs.
+
+### Structured Mode (Async)
+
+Same architecture as sync, but overlaps generation of the next batch with training on the current batch. Faster throughput at the cost of slightly stale weights during generation.
+
+**When to use**: Same data as sync mode, but when you want faster training and can tolerate weight staleness (controlled by `vllm_sync_interval`).
+
+### Strided Mode
+
+Runs entirely on a single GPU with no vLLM dependency. Places anchor points throughout a document and generates short rollouts at each anchor using block-parallel attention patterns.
+
+```
+Single GPU: Base model + LoRA adapter
+  - Strided block-parallel generation (flex_attention)
+  - Feature extraction via disable_adapter()
+  - No vLLM needed
+```
+
+**When to use**: Unstructured text data (raw code, prose, documents) where there is no natural prompt/completion split. Also works with structured data that includes prompt boundaries. Requires only 1 GPU.
+
+## Quick Start
+
+### Structured Mode
+
+This minimal example fine-tunes Qwen2-0.5B on code data using EBFT with vLLM generation.
+
+**Step 1**: Create a config file `ebft_quickstart.yaml`:
+
+```yaml
+base_model: Qwen/Qwen2-0.5B-Instruct
+
+rl: ebft
+
+ebft:
+  feature_layers: [0.25, 0.5, 0.75]
+  embed_method: last_token
+  alignment_coef: 1.0
+  diversity_coef: 1.0
+
+trl:
+  num_generations: 4
+  max_completion_length: 256
+  temperature: 0.7
+  use_vllm: true
+  vllm_server_host: 0.0.0.0
+  vllm_server_port: 8000
+  vllm_lora_sync: true
+  vllm_sync_interval: 3
+  use_data_producer: true
+  async_prefetch: false
+  scale_rewards: true
+  loss_type: grpo
+
+vllm:
+  gpu_memory_utilization: 0.5
+  max_model_len: 1024
+
+datasets:
+  - path: nvidia/OpenCodeInstruct
+    type: ebft_opencode.transform
+    split: train[:500]
+
+# Standard training settings (see getting-started.qmd for details)
+adapter: lora
+lora_r: 16
+lora_alpha: 32
+lora_target_linear: true
+sequence_len: 1024
+micro_batch_size: 2
+gradient_accumulation_steps: 4
+max_steps: 20
+learning_rate: 5.0e-6
+bf16: auto
+flash_attention: true
+gradient_checkpointing: true
+output_dir: ./outputs/ebft-quickstart
+```
+
+**Step 2**: Start vLLM on GPU 0:
+
+```bash
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve ebft_quickstart.yaml
+```
+
+**Step 3**: Wait approximately 30 seconds for vLLM to initialize, then start training on GPU 1:
+
+```bash
+CUDA_VISIBLE_DEVICES=1 axolotl train ebft_quickstart.yaml
+```
+
+::: {.callout-important}
+The `micro_batch_size` must be divisible by `num_generations`. For example, with `num_generations: 4`, valid values are 4, 8, 12, etc.
+:::
+
+### Dataset Format
+
+Structured mode datasets must produce two fields after the transform:
+
+- `prompt`: Either a string or a list of chat messages (`[{"role": "user", "content": "..."}]`)
+- `ground_truth`: A string containing the reference completion
+
+Example raw dataset row:
+
+```json
+{
+  "input": "Write a function to compute fibonacci numbers.",
+  "output": "def fibonacci(n):\n    if n <= 1:\n        return n\n    return fibonacci(n-1) + fibonacci(n-2)"
+}
+```
+
+The `ebft_opencode.transform` converts this to the required `{prompt, ground_truth}` format automatically.
+
+## Feature Extraction
+
+EBFT extracts hidden states from intermediate transformer layers and pools them into per-sequence embeddings. These embeddings are compared between generated and ground-truth completions to compute rewards.
+
+### Feature Layers
+
+The `feature_layers` parameter specifies which layers to extract, as fractions of total model depth:
+
+```yaml
+ebft:
+  feature_layers: [0.25, 0.5, 0.75]  # Quarter, middle, three-quarter depth
+```
+
+For a 32-layer model, this extracts layers 8, 16, and 24. The hidden states from all selected layers are concatenated along the feature dimension, producing embeddings of size `num_layers * hidden_dim`.
+
+::: {.callout-tip}
+Using multiple layers captures both low-level syntactic features (early layers) and high-level semantic features (later layers). The default `[0.25, 0.5, 0.75]` works well across model sizes.
+:::
+
+### Embed Methods
+
+The `embed_method` controls how per-token hidden states are pooled into a single vector per sequence:
+
+| Method | Description | Output Shape | Notes |
+|--------|-------------|-------------|-------|
+| `last_token` | Hidden state at the last non-padding token | `(B, D)` | Default. Good for autoregressive models where the last token summarizes the sequence. |
+| `mean_pooling` | Mean of all non-padding token states | `(B, D)` | Considers the entire sequence equally. |
+| `completion_mean` | Mean over completion tokens only (excludes prompt) | `(B, D)` | Focuses reward signal on generated content. Requires prompt length information. |
+| `concat` | Concatenation of states at 25%, 50%, 75% positions | `(B, 3*D)` | Captures positional structure. Higher dimensional. |
+
+```yaml
+ebft:
+  embed_method: completion_mean  # Focus on completion features
+```
+
+### SVD Whitening
+
+Whitening decorrelates the feature dimensions so that no single direction dominates the feature-matching loss. This is computed via SVD on the generated embeddings, with the same transform applied to the ground-truth embeddings.
+
+```yaml
+ebft:
+  use_whitening: true
+```
+
+When whitening is enabled, the reward computation applies a whitening matrix `W = U @ diag(1/S) @ U^T` derived from the SVD of generated embeddings. This ensures all feature dimensions contribute equally to the alignment reward.
+
+::: {.callout-note}
+Singular values scale with `sqrt(batch_size)`, so reward magnitudes are batch-size dependent. This is acceptable because the number of samples per prompt (`n_samples_per_prompt` or `num_generations`) is fixed during training.
+:::
+
+### Alignment and Diversity Coefficients
+
+The two reward components are weighted by coefficients:
+
+```yaml
+ebft:
+  alignment_coef: 1.0   # Weight for cosine similarity with ground truth
+  diversity_coef: 1.0   # Weight for pairwise similarity penalty
+```
+
+Both values are scaled by 2 internally (per paper equation 7). The final reward per sample is:
+
+```
+reward_j = 2 * alignment_coef * cos(gen_j, gt)
+         - 2 * diversity_coef * (1/(n-1)) * sum_{j' != j} dot(gen_j, gen_j')
+```
+
+Setting `diversity_coef: 0.0` disables the diversity penalty entirely, which may be appropriate when `num_generations` is small (e.g., 2).
+
+## Strided Mode
+
+Strided mode is designed for training on unstructured text data where there is no natural prompt/completion boundary. Instead of generating full completions with vLLM, it places **anchor points** at regular intervals throughout each document and generates short rollouts at each anchor using block-parallel attention.
+
+### How Block-Parallel Generation Works
+
+Given a document of length `S` tokens:
+
+1. **Anchor placement**: Starting at position `anchor_offset`, place anchors every `stride` tokens. Each anchor defines a block.
+2. **Context window**: Each block sees `context_length` tokens of preceding context from the original document.
+3. **Generation**: At each anchor, generate `generate_max_len` tokens autoregressively, conditioned only on the context window.
+4. **Parallelism**: All blocks are processed in a single forward pass using a specialized attention mask that prevents information leakage between blocks.
+
+```
+Document:   [tok0, tok1, ..., tok_S]
+                    |         |         |
+                 anchor_0   anchor_1  anchor_2
+                    |         |         |
+             [ctx][gen]  [ctx][gen]  [ctx][gen]
+```
+
+The attention mask ensures:
+
+- Prompt tokens use standard causal attention
+- Each generated block attends to its own context window and its own preceding generated tokens
+- Blocks do not attend to each other's generated tokens
+
+When `flex_attention` is available (PyTorch >= 2.5), the mask is compiled into efficient fused kernels. Otherwise, a dense 4D attention mask is used as a fallback.
+
+### Strided Mode Configuration
+
+```yaml
+base_model: meta-llama/Llama-3.2-1B
+rl: ebft
+
+ebft:
+  mode: strided
+  stride: 8                    # Tokens between anchor points
+  context_length: 8            # Context window per block
+  generate_max_len: 8          # Tokens to generate per block
+  n_samples_per_prompt: 4      # Independent rollouts per document
+  temperature: 0.6
+  feature_layers: [0.25, 0.5, 0.75]
+  embed_method: last_token
+  use_whitening: true
+  alignment_coef: 1.0
+  diversity_coef: 1.0
+  rl_coef: 1.0                # RL policy gradient loss weight
+  ce_coef: 0.03               # Cross-entropy loss on GT tokens
+  advantage_estimator: rloo    # rloo, group_norm, or reinforce
+  min_completion_prefix: 8     # Skip anchors in prompt region
+
+datasets:
+  - path: nvidia/OpenCodeInstruct
+    type: ebft_strided_structured.transform
+    split: train[:1%]
+
+sequence_len: 2048
+micro_batch_size: 1
+gradient_accumulation_steps: 2
+
+adapter: lora
+lora_r: 16
+lora_alpha: 32
+lora_target_linear: true
+
+bf16: auto
+flex_attention: true
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: true          # Required with flex_attention
+```
+
+Run with a single command (no vLLM needed):
+
+```bash
+CUDA_VISIBLE_DEVICES=0 axolotl train config.yaml
+```
+
+### Advantage Estimators
+
+Strided mode supports three advantage estimation methods:
+
+| Estimator | Formula | Requirements |
+|-----------|---------|-------------|
+| `rloo` | Leave-one-out baseline: `reward_j - mean(rewards_{-j})` | `n_samples_per_prompt >= 2` |
+| `group_norm` | Group normalization: `(reward_j - mean) / std` | `n_samples_per_prompt >= 2` |
+| `reinforce` | Raw reward as advantage (no baseline) | Works with `n_samples_per_prompt = 1` |
+
+::: {.callout-warning}
+When `n_samples_per_prompt: 1`, the trainer automatically falls back to `reinforce` and disables the diversity penalty (which requires multiple samples).
+:::
+
+### Strided Mode Constraints
+
+- **`flex_attention: true`** is strongly recommended. Without it, dense 4D masks consume significantly more memory.
+- **`torch_compile: true`** must NOT be set. `flex_attention` compiles its own kernels internally; adding `torch_compile` causes conflicts and OOM.
+- **Gradient checkpointing** must use `use_reentrant: true`. Non-reentrant checkpointing causes `CheckpointError` with `flex_attention` block masks.
+- **`activation_offloading`** is incompatible with `flex_attention`.
+
+### Cross-Entropy Loss
+
+Strided mode supports an optional cross-entropy loss term on ground-truth tokens. This acts as a regularizer to prevent the model from drifting too far from the original distribution:
+
+```yaml
+ebft:
+  ce_coef: 0.03    # Small CE coefficient
+  rl_coef: 1.0     # RL loss coefficient
+```
+
+The total loss is `rl_coef * rl_loss + ce_coef * ce_loss`. For structured mode, `ce_coef` is typically `0.0` since vLLM generation provides sufficient learning signal.
+
+## Dataset Formats
+
+EBFT provides several built-in dataset transforms in `src/axolotl/prompt_strategies/ebft/`.
+
+### Built-In Transforms
+
+| Transform | Input Format | Output Fields | Use Case |
+|-----------|-------------|---------------|----------|
+| `ebft_opencode.transform` | `{input, output}` | `{prompt, ground_truth}` | OpenCodeInstruct, structured QA |
+| `ebft_strided_structured.transform` | `{input, output}` | `{input_ids, labels, prompt_length}` | Strided mode with structured data |
+| `ebft_strided_chat.transform` | `{messages: [...]}` | `{input_ids, labels, prompt_length}` | Strided mode with chat data |
+| `ebft_chat_multiturn.transform` | `{messages: [...]}` | `{prompt, ground_truth, remaining_turns}` | Multi-turn: first-turn target |
+| `ebft_chat_multiturn.transform_last_turn` | `{messages: [...]}` | `{prompt, ground_truth}` | Multi-turn: last-turn target |
+| `ebft_chat_multiturn.transform_all_turns` | `{messages: [...]}` | `{prompt[], ground_truth[]}` | Multi-turn: one example per turn |
+| `ebft_reasoning.transform` | `{messages: [...]}` (with `<think>`) | `{prompt, ground_truth}` | Reasoning/thinking datasets |
+
+### Structured Mode Datasets
+
+For structured (sync/async) mode, the transform must produce `prompt` and `ground_truth` fields:
+
+```yaml
+datasets:
+  - path: nvidia/OpenCodeInstruct
+    type: ebft_opencode.transform
+    split: train[:500]
+```
+
+### Multi-Turn Datasets
+
+Multi-turn transforms extract conversation data for sequential rollout. The `transform` variant targets the first assistant turn, while `transform_last_turn` targets the final turn:
+
+```yaml
+datasets:
+  - path: your/multiturn-dataset
+    type: ebft_chat_multiturn.transform
+```
+
+When `remaining_turns` is present in the dataset output, the trainer performs sequential rollouts: it generates the first assistant turn with vLLM, then continues generating subsequent turns by building up the conversation history.
+
+### Strided Mode Datasets
+
+Strided transforms tokenize the full document and produce `input_ids`, `labels`, and `prompt_length`:
+
+```yaml
+datasets:
+  - path: nvidia/OpenCodeInstruct
+    type: ebft_strided_structured.transform
+    split: train[:1%]
+```
+
+### Custom Transforms
+
+To use your own dataset format, write a transform function:
+
+```python
+def transform(cfg, **kwargs):
+    def transform_fn(example, tokenizer=None):
+        return {
+            "prompt": [{"role": "user", "content": example["question"]}],
+            "ground_truth": example["answer"],
+        }
+    return transform_fn, {"remove_columns": "__all__"}
+```
+
+The `"__all__"` sentinel removes all original dataset columns after the mapping step. Reference this transform in your config:
+
+```yaml
+datasets:
+  - path: your/dataset
+    type: your_module.transform
+```
+
+## Configuration Reference
+
+### Common Parameters (All Modes)
+
+These parameters are set under the `ebft:` key in the YAML config.
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `mode` | `"structured"` or `"strided"` | `"structured"` | EBFT operating mode |
+| `feature_layers` | `list[float]` | `[0.25, 0.5, 0.75]` | Fractional layer depths for feature extraction |
+| `embed_method` | `string` | `"last_token"` | Pooling method: `last_token`, `mean_pooling`, `completion_mean`, or `concat` |
+| `use_whitening` | `bool` | `false` | Apply SVD whitening to feature embeddings before reward computation |
+| `alignment_coef` | `float` | `1.0` | Weight for alignment reward (cosine similarity with ground truth) |
+| `diversity_coef` | `float` | `1.0` | Weight for diversity penalty (pairwise dot product between samples) |
+| `ce_coef` | `float` | `0.0` | Cross-entropy loss coefficient on ground-truth tokens |
+| `adaptive_max_tokens` | `bool` | `true` | Dynamically set vLLM `max_tokens` based on ground-truth length (structured mode) |
+| `gt_length_multiplier` | `float` | `1.5` | Multiplier for ground-truth token count when computing adaptive max tokens (min 0.1) |
+
+### Strided Mode Parameters
+
+These additional parameters apply only when `mode: strided`.
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `stride` | `int` | `8` | Number of tokens between anchor points (must be >= 1) |
+| `context_length` | `int` | `8` | Context window size for each generated block (must be >= 1) |
+| `generate_max_len` | `int` | `8` | Number of tokens to generate per block (must be >= 1) |
+| `n_samples_per_prompt` | `int` | `4` | Number of independent rollouts per document (must be >= 1) |
+| `temperature` | `float` | `0.6` | Sampling temperature for strided generation |
+| `top_p` | `float` | `1.0` | Top-p nucleus sampling threshold |
+| `rl_coef` | `float` | `1.0` | RL policy gradient loss coefficient |
+| `advantage_estimator` | `string` | `"rloo"` | Advantage estimation method: `rloo`, `group_norm`, or `reinforce` |
+| `min_completion_prefix` | `int` | `0` | Minimum tokens into the completion span before placing anchors |
+
+### Structured Mode TRL Parameters
+
+These are set under the `trl:` key and control the GRPO training loop.
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `num_generations` | `int` | -- | Number of completions generated per prompt |
+| `max_completion_length` | `int` | -- | Maximum tokens per generated completion |
+| `temperature` | `float` | `0.7` | Sampling temperature for vLLM generation |
+| `use_vllm` | `bool` | -- | Enable vLLM generation backend |
+| `vllm_lora_sync` | `bool` | `false` | Sync LoRA adapters via filesystem (recommended) |
+| `vllm_sync_interval` | `int` | `1` | Steps between weight syncs to vLLM |
+| `use_data_producer` | `bool` | -- | Required for sync mode with LoRA sync |
+| `async_prefetch` | `bool` | `false` | Enable async generation (overlaps with training) |
+| `streaming_partial_batch` | `bool` | `false` | Score groups incrementally (async mode) |
+| `skip_zero_advantage_batches` | `bool` | `false` | Skip micro-batches where all advantages are zero |
+| `scale_rewards` | `bool` | -- | Normalize rewards within each prompt group |
+| `loss_type` | `string` | `"grpo"` | Loss type for policy optimization |
+| `epsilon` | `float` | `0.2` | Clipping parameter for importance sampling |
+
+### Stop Tokens
+
+vLLM needs explicit stop token IDs for generation. Common configurations:
+
+```yaml
+trl:
+  generation_kwargs:
+    stop_token_ids: [151645, 151643]   # Qwen: <|im_end|>, <|endoftext|>
+```
+
+### Multi-Turn Chat Settings
+
+For multi-turn conversations with Qwen3.5, disable thinking mode to prevent `<think>` tags in completions:
+
+```yaml
+trl:
+  chat_template_kwargs:
+    enable_thinking: false
+```
+
+## Monitoring
+
+### Key Metrics
+
+EBFT logs several custom metrics to wandb and the training console. Here is what to watch for:
+
+| Metric | Healthy Range | Interpretation |
+|--------|--------------|----------------|
+| `ebft/alignment` | 0.3 -- 0.9, trending upward | Cosine similarity between generated and ground-truth features. Higher means the model is learning to produce representations that match the reference. |
+| `ebft/diversity` | 0.01 -- 0.1 | Mean pairwise similarity between different generations for the same prompt. Values above 1.0 indicate mode collapse. |
+| `ebft/cfm_loss` | Below 10, trending downward | Cross-Feature Matching loss. This is the core quantity being minimized. Consistently above 100 indicates instability. |
+| `ebft/reward` | Trending upward (may start negative) | Combined reward signal. If stuck at -1.0, the diversity penalty is dominating alignment. |
+| `grad_norm` | 0.1 -- 3.0 | Gradient magnitude. Values of 0.0 indicate zero-advantage skip (normal). Values above 10 suggest instability. |
+| `entropy` | 0.05 -- 0.5 | Policy entropy. Values below 0.01 suggest mode collapse. |
+| `IS ratio min` | Above 0.1 | Importance sampling ratio minimum. Near-zero values mean the policy is too far off-policy; increase `vllm_sync_interval`. |
+
+### Console Log Example
+
+During training, you will see periodic EBFT reward logs:
+
+```
+ebft reward | align +0.412 ^ | divers +0.023 v | cfm 4.231 v | reward +0.389 ^
+```
+
+The arrows indicate the desired direction: alignment and reward should trend upward, while diversity and CFM loss should trend downward.
+
+### Troubleshooting
+
+| Symptom | Likely Cause | Fix |
+|---------|-------------|-----|
+| `alignment` stays below 0.1 | Feature layers not capturing useful information | Try different `feature_layers` or `embed_method` |
+| `diversity` exceeds 1.0 | Mode collapse -- generations are too similar | Increase `diversity_coef` or `temperature` |
+| `reward` stuck at -1.0 | Diversity penalty dominates alignment | Reduce `diversity_coef` or increase `alignment_coef` |
+| `grad_norm` consistently 0.0 | All micro-batches have zero advantage | Increase `num_generations` or check data quality |
+| `CheckpointError` in strided mode | Incompatible gradient checkpointing settings | Set `use_reentrant: true` in `gradient_checkpointing_kwargs` |
+| OOM during training | Logits tensor too large | Reduce `sequence_len` or `micro_batch_size`; strided mode uses chunked lm_head to mitigate this |
+| vLLM 500 errors | `truncate_prompt_tokens` not supported | Ensure you are using `axolotl vllm-serve` (not `trl vllm-serve`) |
+
+### Feature Network Memory
+
+In PEFT (LoRA) mode, the feature network shares base weights with the actor model by using the `disable_adapter()` context manager. This saves an entire model copy in VRAM (approximately 1--16 GB depending on model size). For non-PEFT training, a separate frozen deepcopy is created.
+
+::: {.callout-note}
+The `disable_adapter()` approach relies on an invariant: `merge_adapter()` is never called on the base weights. All weight sync paths (LoRA sync, HTTP, NCCL) compute merged weights as new tensors or save the adapter to the filesystem, leaving base weights unmodified.
+:::
+
+## Examples
+
+Complete example configurations are available in `examples/ebft/`:
+
+| Config | Model | Mode | Description |
+|--------|-------|------|-------------|
+| `llama-1b-ebft-strided-structured.yaml` | Llama 3.2 1B | Strided | Single-GPU strided training on code data |
+| `qwen3-4b-ebft-structured.yaml` | Qwen3 4B | Structured (sync) | Two-GPU structured training |
+| `qwen3-4b-ebft-structured-async.yaml` | Qwen3 4B | Structured (async) | Two-GPU async training with prefetch |
+| `qwen3-8b-ebft-structured.yaml` | Qwen3 8B | Structured (sync) | Two-GPU structured training for larger model |
+| `qwen35-4b-ebft-structured.yaml` | Qwen3.5 4B | Structured (sync) | Two-GPU with Qwen3.5 |
+| `qwen35-4b-ebft-structured-async.yaml` | Qwen3.5 4B | Structured (async) | Two-GPU async with Qwen3.5 |
+| `qwen35-9b-ebft-structured.yaml` | Qwen3.5 9B | Structured (sync) | Two-GPU structured for 9B model |

docs/getting-started.qmd

@@ -170,17 +170,26 @@ More details can be found in [Merging LoRA weights](inference.qmd#sec-merging).
 
 ## Next Steps {#sec-next-steps}
 
-Now that you have the basics, you might want to:
+Now that you have the basics, explore these guides based on what you want to do:
 
-- Try different model architectures
-- Experiment with hyperparameters
-- Use more advanced training methods
-- Scale up to larger models
+**Choose your path:**
 
-Check our other guides for details on these topics:
+- [Choosing a Fine-Tuning Method](choosing_method.qmd) — SFT vs LoRA vs QLoRA vs GRPO vs DPO, with hardware recommendations
 
-- [Configuration Guide](config-reference.qmd) - Full configuration options
-- [Dataset Loading](dataset_loading.qmd) - Loading datasets from various sources
-- [Dataset Formats](dataset-formats) - Working with different data formats
-- [Multi-GPU Training](multi-gpu.qmd)
-- [Multi-Node Training](multi-node.qmd)
+**Core guides:**
+
+- [Dataset Loading](dataset_loading.qmd) — Loading datasets from various sources
+- [Dataset Formats](dataset-formats) — Working with different data formats
+- [Optimizations](optimizations.qmd) — Flash attention, gradient checkpointing, sample packing
+- [Training Stability & Debugging](training_stability.qmd) — Monitoring metrics, fixing NaN, OOM debugging
+
+**Advanced training methods:**
+
+- [RLHF / Preference Learning](rlhf.qmd) — DPO, KTO, GRPO, EBFT
+- [GRPO Training](grpo.qmd) — RL with custom rewards and vLLM generation
+- [vLLM Serving](vllm_serving.qmd) — Setting up vLLM for GRPO
+
+**Scaling up:**
+
+- [Multi-GPU Training](multi-gpu.qmd) — DeepSpeed, FSDP, DDP
+- [Multi-Node Training](multi-node.qmd) — Distributed training across machines

docs/gradient_checkpointing.qmd

@@ -1,5 +1,5 @@
 ---
-title: Gradient Checkpointing and Activation Offloading
+title: Gradient Checkpointing, Activation Offloading, and Layer Offloading
 ---
 
 Gradient checkpointing and activation offloading are techniques used to optimize the performance of deep learning
@@ -27,3 +27,33 @@ The `activation_offloading: legacy` naively offloads activations to CPU and with
 
 For resource constrained environments with limited CPU memory, `activation_offloading: disk` offloads
 activations to disk instead of CPU RAM so that much larger context lengths can be trained with minimal memory.
+
+### Enabling Layer Offloading
+
+```yaml
+layer_offloading: true
+```
+
+Layer offloading reduces GPU memory usage by moving frozen (non-trainable) decoder layer parameters to CPU
+and streaming them back to GPU one layer at a time during the forward and backward passes. This is
+particularly useful for LoRA/QLoRA training where most of the model's parameters are frozen — only the
+trainable adapter weights stay on GPU permanently.
+
+During training, forward and backward hooks on each decoder layer handle the transfer automatically:
+
+- **Forward pass:** Before a layer executes, its frozen params are loaded to GPU. The next layer is
+  prefetched asynchronously on a separate CUDA stream for overlap.
+- **Backward pass:** Same pattern in reverse — the current layer's frozen params are loaded and the
+  previous layer is prefetched.
+
+After each layer finishes, its frozen params are offloaded back to CPU pinned memory.
+
+This approach trades some CPU-GPU transfer overhead for significant GPU memory savings — the freed memory
+is roughly equal to the size of all frozen parameters across all decoder layers, minus one layer's worth
+that is kept on GPU at any given time.
+
+**Requirements:**
+
+- CUDA GPU (CPU-only training is not supported for this feature)
+- Works with any HuggingFace model architecture that uses decoder layers (Llama, Mistral, Qwen, etc.)
+- Best combined with LoRA/QLoRA where most parameters are frozen

docs/grpo.qmd

@@ -0,0 +1,611 @@
+---
+title: "GRPO Training"
+description: "Group Relative Policy Optimization — a reinforcement learning method for training language models with verifiable reward functions."
+order: 8
+---
+
+## Overview
+
+Group Relative Policy Optimization (GRPO) is a reinforcement learning method that improves language models by generating multiple completions per prompt, scoring them with reward functions, and using the relative ranking within each group to compute advantage estimates. Unlike DPO, which requires pre-collected preference pairs, GRPO generates its own training data online and can work with any programmatic reward signal (math correctness, format compliance, code execution results, etc.).
+
+Use GRPO when you have a task with a verifiable reward signal and want the model to discover solution strategies on its own. Use DPO when you already have human preference data. Use SFT when you have gold-standard completions to imitate directly.
+
+Axolotl's GRPO implementation builds on TRL and adds async generation, streaming scoring, importance sampling correction, replay buffers, and multi-GPU scaling via FSDP and DeepSpeed.
+
+
+## Architecture
+
+GRPO training uses a two-process architecture: a vLLM server for fast generation and a trainer process for scoring and gradient updates.
+
+```
+Terminal 1 (GPU 0)                    Terminal 2 (GPU 1)
+┌──────────────────────┐              ┌──────────────────────────────────┐
+│  vLLM Server         │              │  Trainer                         │
+│                      │   HTTP       │                                  │
+│  Serves base model   │◄────────────►│  Background thread:              │
+│  + LoRA adapter      │  /generate   │    Send prompts to vLLM          │
+│                      │  /set_lora   │    Pad & collate completions     │
+│  Punica kernels for  │              │                                  │
+│  LoRA inference      │              │  Main thread:                    │
+│                      │              │    Score completions (rewards)   │
+└──────────────────────┘              │    Compute policy log-probs      │
+                                      │    Calculate advantages          │
+                                      │    PPO-clip gradient update      │
+                                      │    Sync LoRA weights to vLLM     │
+                                      └──────────────────────────────────┘
+```
+
+**Data flow for each training step:**
+
+1. The background thread sends prompts to vLLM, which generates `num_generations` completions per prompt.
+2. The main thread scores completions using your reward functions.
+3. Advantages are computed within each prompt group (group-relative normalization).
+4. Policy log-probabilities are computed by running a forward pass on the training model.
+5. The PPO-clip loss is computed and gradients are applied.
+6. Periodically, LoRA adapter weights are synced back to vLLM so future generations reflect the updated policy.
+
+With async prefetch enabled, step 1 for the *next* batch runs concurrently with steps 2-6 for the *current* batch.
+
+
+## Quick Start
+
+A GRPO training run requires three components: a YAML config, a reward module (Python file), and a running vLLM server.
+
+### 1. Write a reward module
+
+Create a file called `rewards.py` in your working directory:
+
+```python
+# rewards.py
+import re
+
+
+def accuracy_reward(completions, answer, **kwargs) -> list[float]:
+    """Check if the completion contains the correct numerical answer."""
+    rewards = []
+    for completion, correct in zip(completions, answer):
+        text = completion[0]["content"]
+        # Extract the last number from the completion
+        numbers = re.findall(r"-?\d+(?:\.\d+)?", text)
+        predicted = numbers[-1] if numbers else ""
+        rewards.append(1.0 if predicted == str(correct) else 0.0)
+    return rewards
+
+
+def format_reward(completions, **kwargs) -> list[float]:
+    """Reward completions that use a structured thinking format."""
+    rewards = []
+    for completion in completions:
+        text = completion[0]["content"]
+        has_think = "<think>" in text and "</think>" in text
+        has_answer = "<answer>" in text and "</answer>" in text
+        rewards.append(1.0 if has_think and has_answer else 0.0)
+    return rewards
+
+
+def prompt_transform(cfg, *args, **kwargs):
+    """Convert GSM8K dataset rows into chat prompts."""
+    def transform_fn(example, tokenizer=None):
+        label = example["answer"].split("####")[-1].strip().replace(",", "")
+        return {
+            "prompt": [
+                {"role": "system", "content": "Solve the math problem. Show your reasoning in <think> tags and your final numerical answer in <answer> tags."},
+                {"role": "user", "content": example["question"]},
+            ],
+            "answer": label,
+        }
+    return transform_fn, {"remove_columns": ["question"]}
+```
+
+### 2. Write the config
+
+Create `config.yaml`:
+
+```yaml
+base_model: Qwen/Qwen2.5-1.5B-Instruct
+
+rl: grpo
+chat_template: tokenizer_default
+
+vllm:
+  host: 0.0.0.0
+  port: 8000
+  gpu_memory_utilization: 0.85
+  dtype: auto
+  max_model_len: 2048
+
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_target_linear: true
+
+trl:
+  use_vllm: true
+  use_data_producer: true
+  vllm_server_host: 0.0.0.0
+  vllm_server_port: 8000
+  vllm_server_timeout: 300
+  vllm_lora_sync: true
+  num_generations: 8
+  max_completion_length: 512
+  temperature: 0.7
+  reward_funcs:
+    - rewards.accuracy_reward
+    - rewards.format_reward
+  reward_weights:
+    - 1.0
+    - 0.5
+
+datasets:
+  - path: openai/gsm8k
+    name: main
+    type: rewards.prompt_transform
+    split: train
+
+skip_prepare_dataset: true
+val_set_size: 0.0
+sequence_len: 512
+micro_batch_size: 2
+gradient_accumulation_steps: 4
+max_steps: 200
+learning_rate: 5.0e-6
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+warmup_steps: 10
+
+bf16: true
+flash_attention: true
+gradient_checkpointing: true
+
+special_tokens:
+  pad_token: "<|endoftext|>"
+
+output_dir: ./grpo-output
+logging_steps: 1
+```
+
+### 3. Start vLLM and train
+
+```bash
+# Terminal 1: Start vLLM server on GPU 0
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve config.yaml
+
+# Wait 30-90 seconds for model loading and CUDA graph capture
+
+# Terminal 2: Train on GPU 1
+CUDA_VISIBLE_DEVICES=1 axolotl train config.yaml
+```
+
+:::{.callout-tip}
+Use `tmux` or separate terminal sessions to manage the two processes. The vLLM server must remain running for the entire training duration.
+:::
+
+
+## Custom Reward Functions
+
+### Function signature
+
+TRL calls reward functions with this signature:
+
+```python
+def my_reward(completions, **kwargs) -> list[float]:
+```
+
+- `completions` is a list of single-element lists, where each element is a dict `{"role": "assistant", "content": "..."}`. So `completions[i][0]["content"]` gives you the text of the i-th completion.
+- `**kwargs` contains all dataset columns that were *not* removed by the dataset transform. This is how you pass ground truth answers, metadata, or any other information to your reward function.
+- Return a `list[float]` with the same length as `completions`. You may return `None` for individual elements to exclude them from aggregation.
+
+### Example: accuracy reward with answer extraction
+
+```python
+def accuracy_reward(completions, answer, **kwargs) -> list[float]:
+    rewards = []
+    for completion, correct_answer in zip(completions, answer):
+        text = completion[0]["content"]
+        # Extract answer from <answer>...</answer> tags
+        match = re.search(r"<answer>(.*?)</answer>", text, re.DOTALL)
+        predicted = match.group(1).strip() if match else ""
+        rewards.append(1.0 if predicted == str(correct_answer) else 0.0)
+    return rewards
+```
+
+### Example: length penalty
+
+```python
+def length_penalty(completions, **kwargs) -> list[float]:
+    """Penalize very short or very long completions."""
+    rewards = []
+    for completion in completions:
+        length = len(completion[0]["content"])
+        if length < 50:
+            rewards.append(-0.5)
+        elif length > 2000:
+            rewards.append(-0.2)
+        else:
+            rewards.append(0.0)
+    return rewards
+```
+
+### Multiple rewards and weighting
+
+You can combine multiple reward functions with different weights:
+
+```yaml
+trl:
+  reward_funcs:
+    - rewards.accuracy_reward
+    - rewards.format_reward
+    - rewards.length_penalty
+  reward_weights:
+    - 1.0    # accuracy is most important
+    - 0.5    # format compliance
+    - 0.1    # mild length preference
+```
+
+Rewards are combined by the `multi_objective_aggregation` strategy:
+
+- `sum_then_normalize` (default): weights and sums all rewards first, then normalizes across the group.
+- `normalize_then_sum` (GDPO): normalizes each reward independently, then sums. This prevents one reward from dominating and is recommended when using multiple reward functions with different scales.
+
+```yaml
+trl:
+  multi_objective_aggregation: normalize_then_sum
+```
+
+### Dataset transforms
+
+The dataset transform converts raw HuggingFace dataset rows into chat-format prompts:
+
+```python
+def prompt_transform(cfg, *args, **kwargs):
+    def map_fn(example, tokenizer=None):
+        return {
+            "prompt": [
+                {"role": "system", "content": "You are a helpful assistant."},
+                {"role": "user", "content": example["question"]},
+            ],
+            # Keep 'answer' column for the reward function
+            "answer": example["answer"],
+        }
+    # Remove columns consumed by the transform; keep columns needed by rewards
+    return map_fn, {"remove_columns": ["question"]}
+```
+
+The transform returns a tuple of `(map_function, kwargs_dict)`. The `remove_columns` in the kwargs dict removes columns that are no longer needed. Columns that your reward functions reference via `**kwargs` (like `answer`) must *not* be removed.
+
+:::{.callout-warning}
+The reward module must be importable from the directory where you run `axolotl train`. If your reward file is `rewards.py`, the import path is `rewards.accuracy_reward`. If it is inside a package `my_rewards/scoring.py`, use `my_rewards.scoring.accuracy_reward`.
+:::
+
+### Reward models (neural network rewards)
+
+Instead of a Python function, you can pass a HuggingFace model path as a reward function. TRL will load it as a reward model and use its scalar output as the reward:
+
+```yaml
+trl:
+  reward_funcs:
+    - OpenAssistant/reward-model-deberta-v3-large-v2
+    - rewards.format_reward
+  reward_weights:
+    - 1.0
+    - 0.3
+```
+
+### Using math_verify
+
+The `math_verify` library provides robust mathematical answer verification but uses `signal.alarm()` internally, which only works in the main thread. If you use `math_verify` in a reward function, set `reward_num_workers` to use subprocess workers:
+
+```yaml
+trl:
+  reward_num_workers: 4
+```
+
+Each worker runs in its own subprocess with its own main thread, so `signal.alarm()` works correctly.
+
+
+## vLLM Setup
+
+GRPO requires a running vLLM server for generation. For a complete guide on server modes, LoRA sync, weight synchronization, and restart procedures, see [vLLM Serving](vllm_serving.qmd).
+
+The minimal setup:
+
+```yaml
+vllm:
+  host: 0.0.0.0
+  port: 8000
+  gpu_memory_utilization: 0.85
+
+trl:
+  use_vllm: true
+  vllm_lora_sync: true         # Recommended with LoRA — faster sync, no NCCL contention
+  vllm_sync_interval: 5        # Sync weights every 5 steps
+```
+
+```bash
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve config.yaml   # GPU 0: vLLM
+CUDA_VISIBLE_DEVICES=1 axolotl train config.yaml         # GPU 1: training
+```
+
+:::{.callout-warning}
+vLLM must be restarted between experiments — stale weight syncs corrupt server state. See [Restart Requirements](vllm_serving.qmd#sec-restart).
+:::
+
+
+## Async Training Features
+
+Async GRPO overlaps generation and training to reduce wall-clock time. While the model trains on the current batch, the next batch is already being generated by vLLM.
+
+### Enabling async prefetch
+
+```yaml
+trl:
+  use_data_producer: true
+  async_prefetch: true
+  prefetch_depth: 1
+  vllm_sync_interval: 2
+```
+
+- `use_data_producer: true` enables the data producer protocol (required for all async features).
+- `async_prefetch: true` runs generation in a background thread.
+- `prefetch_depth` controls how many batches to prefetch ahead (1 is usually sufficient).
+- `vllm_sync_interval` controls how often LoRA weights are synced to vLLM (every N optimizer steps). Lower values mean fresher generations but more sync overhead.
+
+:::{.callout-tip}
+Because the background thread generates with slightly stale model weights, async mode benefits from importance sampling correction (see next section). Enable `vllm_importance_sampling_correction: true` when using `async_prefetch: true`.
+:::
+
+### Streaming partial batch
+
+Instead of scoring the entire batch at once, streaming mode scores one prompt group at a time. This reduces peak memory during scoring and enables finer-grained zero-advantage skipping.
+
+```yaml
+trl:
+  streaming_partial_batch: true
+  streaming_min_groups: 1
+```
+
+`streaming_min_groups` controls the minimum number of prompt groups scored per chunk. Setting it to 1 gives maximum granularity.
+
+### Zero-advantage batch skipping
+
+When all advantages in a micro-batch are zero (every completion in the group got the same reward), there is no learning signal. This feature skips the forward/backward pass entirely for such micro-batches.
+
+```yaml
+trl:
+  skip_zero_advantage_batches: true   # default
+```
+
+This is enabled by default and logged as `skipped_zero_adv_batches` in training metrics. It is a safety net, not a major optimization -- it only saves significant time when the model cannot solve any prompts in the batch.
+
+### Replay buffer
+
+The replay buffer caches rollout groups that had learning signal (non-zero reward variance) and replaces zero-signal groups in later batches. This improves data utilization when many prompts yield no reward variance.
+
+```yaml
+trl:
+  replay_buffer_size: 100
+  replay_recompute_logps: true
+```
+
+:::{.callout-warning}
+When `replay_recompute_logps: false`, replayed data uses stale log-probabilities which creates an IS mismatch. Keep the default `true` unless you have a specific reason to disable it.
+:::
+
+### Deferred re-rolling
+
+Prompts where the model gets zero reward for all generations are buffered and re-injected into later batches, when the model may have improved enough to produce useful completions.
+
+```yaml
+trl:
+  reroll_start_fraction: 0.5   # Start re-rolling after 50% of training
+  reroll_max_groups: 1          # Max groups to replace per batch
+```
+
+Set `reroll_start_fraction: 1.0` to disable. This is most useful for tasks where the model starts weak but steadily improves.
+
+### Parallel reward workers
+
+Reward functions that use `signal.alarm()` (like `math_verify`) only work in the main thread. Parallel reward workers run each function in its own subprocess:
+
+```yaml
+trl:
+  reward_num_workers: 4
+```
+
+Work is sharded across workers by prompt group. For simple reward functions, a single worker is usually sufficient -- the overhead of IPC can exceed the computation time.
+
+
+## Importance Sampling and Off-Policy Correction
+
+When using async prefetch, completions are generated from a slightly older policy. IS correction adjusts the gradient to account for this mismatch.
+
+```yaml
+trl:
+  vllm_importance_sampling_correction: true
+  importance_sampling_level: token     # 'token' recommended (especially with Liger kernel)
+  off_policy_mask_threshold: 0.5       # KL threshold — masks sequences that are too off-policy
+```
+
+Use `token` level IS. Sequence-level has numerical issues with Liger's chunked computation. The `off_policy_mask_threshold` (OPSM) is a safety net that drops sequences where KL divergence exceeds the threshold — 0.5 is a reasonable starting point.
+
+For detailed coverage of IS modes (`token_mask`, `token_truncate`, etc.), capping, and bias-corrected KL, see [vLLM Serving — IS Correction](vllm_serving.qmd#sec-weight-sync).
+
+
+## Scaling
+
+### FP8 training
+
+FP8 quantization halves model VRAM usage with minimal impact on training quality. It does not significantly speed up computation for small models but allows larger models to fit in memory.
+
+```yaml
+fp8: true
+torch_compile: true
+```
+
+:::{.callout-warning}
+FP8 requires patching for zero-padding edge cases. The `act_quant_kernel` can produce NaN when input is all zeros (padding positions). If you see NaN in grad norms, check whether your padding token embedding is non-zero.
+:::
+
+### FSDP (Fully Sharded Data Parallel)
+
+FSDP distributes model parameters across multiple GPUs for training while vLLM runs on a separate GPU:
+
+```yaml
+fsdp:
+  - full_shard
+  - auto_wrap
+fsdp_config:
+  fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+```
+
+Launch with:
+
+```bash
+# GPU 0: vLLM
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve config.yaml
+
+# GPUs 0,1: Training (FSDP will use both visible GPUs)
+CUDA_VISIBLE_DEVICES=0,1 axolotl train config.yaml
+```
+
+:::{.callout-warning}
+`async_prefetch: true` can deadlock with FSDP because background threads perform unsynchronized FSDP collectives across ranks. With multi-GPU FSDP, only rank 0 generates in the background thread and results are broadcast to all ranks. If you still see hangs, set `async_prefetch: false`.
+:::
+
+### DeepSpeed ZeRO-3
+
+```yaml
+deepspeed: deepspeed_configs/zero3_bf16.json
+gradient_checkpointing_kwargs:
+  use_reentrant: true   # Required -- non-reentrant causes CheckpointError with ZeRO-3
+```
+
+:::{.callout-note}
+DeepSpeed ZeRO-3 requires `use_reentrant: true` for gradient checkpointing. This is the opposite of the FSDP recommendation. Non-reentrant checkpointing causes tensor metadata mismatches during recomputation with ZeRO-3's parameter partitioning.
+:::
+
+### Multi-GPU considerations
+
+| Concern | Recommendation |
+|---------|---------------|
+| vLLM GPU allocation | Dedicate one or more GPUs to vLLM; do not share with trainer GPUs |
+| Weight sync contention | Use `vllm_lora_sync: true` to avoid NCCL contention between training and vLLM |
+| FSDP + async | Use `async_prefetch: false` or rely on rank-0-only background generation |
+| DeepSpeed + gradient checkpoint | Must use `use_reentrant: true` |
+| OOM during scoring | Reduce `micro_batch_size` or `num_generations`. The logits tensor scales with `batch_size * vocab_size` |
+
+
+## Monitoring and Debugging
+
+For detailed metric ranges, failure diagnosis, and OOM debugging, see [Training Stability & Debugging](training_stability.qmd).
+
+Quick health checks during GRPO training:
+
+- `rewards/*/mean` should be > 0.15 within 20 steps — if it stays at 0, test your reward function standalone
+- `reward_std` should be > 0 on most steps — all-zero means no learning signal
+- `entropy` in 0.05-0.5 — below 0.01 suggests mode collapse
+- `grad_norm` in 0.001-1.0 — > 10 is unstable, 0.0 is expected when zero-advantage skip fires
+
+:::{.callout-tip}
+Pipe training output to a log file: `axolotl train config.yaml 2>&1 | tee /tmp/training.log`
+:::
+
+
+## Configuration Reference
+
+All GRPO-specific options live under the `trl:` key in your config. Standard training options (`learning_rate`, `micro_batch_size`, etc.) are set at the top level as usual.
+
+### Core GRPO
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `use_vllm` | bool | `false` | Enable vLLM for generation |
+| `vllm_mode` | `"server"` or `"colocate"` | `null` | vLLM deployment mode |
+| `vllm_server_host` | str | `"0.0.0.0"` | vLLM server hostname |
+| `vllm_server_port` | int | `8000` | vLLM server port |
+| `vllm_server_timeout` | int | `null` | Timeout (seconds) for vLLM responses |
+| `num_generations` | int | `null` | Completions generated per prompt |
+| `generation_batch_size` | int | `null` | Number of unique prompts per generation step |
+| `max_completion_length` | int | `null` | Maximum tokens per completion |
+| `beta` | float | `null` | KL penalty coefficient |
+| `num_iterations` | int | `null` | Iterations per batch (mu in the GRPO paper) |
+| `epsilon` | float | `null` | PPO clipping lower bound |
+| `epsilon_high` | float | `null` | PPO clipping upper bound |
+| `loss_type` | str | `null` | Loss formulation: `grpo`, `bnpo`, or `dr_grpo` |
+| `scale_rewards` | bool | `true` | Normalize rewards by standard deviation |
+| `mask_truncated_completions` | bool | `false` | Exclude truncated completions from loss |
+
+### Reward functions
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `reward_funcs` | list[str] | `null` | Import paths to reward functions or HF model IDs |
+| `reward_weights` | list[float] | `null` | Relative weights for each reward function |
+| `multi_objective_aggregation` | str | `null` | `"sum_then_normalize"` (GRPO) or `"normalize_then_sum"` (GDPO) |
+| `rollout_func` | str | `null` | Import path to custom rollout function for OpenEnv-style tasks |
+
+### Generation parameters
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `temperature` | float | `null` | Sampling temperature |
+| `top_p` | float | `null` | Nucleus sampling probability |
+| `top_k` | int | `null` | Top-k sampling |
+| `min_p` | float | `null` | Minimum probability threshold |
+| `repetition_penalty` | float | `null` | Penalty for repeated tokens |
+| `generation_kwargs` | dict | `null` | Additional vLLM SamplingParams (e.g., `stop_token_ids`) |
+| `chat_template_kwargs` | dict | `null` | Chat template kwargs (e.g., `{enable_thinking: false}`) |
+| `vllm_guided_decoding_regex` | str | `null` | Regex constraint for guided decoding |
+
+### Async pipeline
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `use_data_producer` | bool | `false` | Enable data producer protocol (required for async features) |
+| `async_prefetch` | bool | `false` | Generate next batch in background thread |
+| `prefetch_depth` | int | `null` | Number of batches to prefetch ahead |
+| `vllm_sync_interval` | int | `null` | Sync LoRA weights to vLLM every N steps |
+| `vllm_lora_sync` | bool | `false` | Use filesystem LoRA sync instead of NCCL merge |
+| `streaming_partial_batch` | bool | `null` | Score prompt groups incrementally |
+| `streaming_min_groups` | int | `null` | Minimum groups per streaming chunk |
+| `skip_zero_advantage_batches` | bool | `true` | Skip micro-batches with zero learning signal |
+| `reward_num_workers` | int | `1` | Subprocess workers for reward computation |
+| `vllm_enable_sleep_mode` | bool | `null` | Offload vLLM weights when idle (colocate mode) |
+
+### Importance sampling
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `vllm_importance_sampling_correction` | bool | `null` | Enable IS correction for async distribution shift |
+| `importance_sampling_level` | `"token"` or `"sequence"` | `null` | Granularity of IS ratios. Use `token` with Liger |
+| `vllm_importance_sampling_mode` | str | `null` | `token_mask`, `token_truncate`, `sequence_mask`, or `sequence_truncate` |
+| `vllm_importance_sampling_cap` | float | `null` | Cap C for IS ratio clipping/masking |
+| `off_policy_mask_threshold` | float | `null` | KL threshold for off-policy sequence masking (OPSM) |
+| `use_bias_correction_kl` | bool | `null` | Apply IS correction to KL divergence term |
+
+### Replay and re-roll
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `replay_buffer_size` | int | `0` | Max cached high-signal groups. 0 = disabled |
+| `replay_recompute_logps` | bool | `true` | Recompute log-probs for replayed data with current model |
+| `reroll_start_fraction` | float | `1.0` | Start re-rolling failed prompts after this fraction of training. 1.0 = disabled |
+| `reroll_max_groups` | int | `1` | Max prompt groups to replace with re-rolls per batch |
+
+### Reference model
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `sync_ref_model` | bool | `false` | Periodically sync reference model with training model |
+| `ref_model_mixup_alpha` | float | `0.9` | EMA coefficient for reference model sync |
+| `ref_model_sync_steps` | int | `64` | Sync reference model every N steps |
+
+### Logging
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `log_completions` | bool | `false` | Log sample completions to W&B |
+| `num_completions_to_print` | int | `null` | Number of completions to print per step |
+| `use_liger_loss` | bool | `null` | Use Liger fused kernel for GRPO loss (reduces VRAM) |

docs/multimodal.qmd

@@ -13,12 +13,14 @@ format:
 - [Pixtral](#sec-pixtral)
 - [Llava-1.5](#sec-llava-15)
 - [Mistral-Small-3.1](#sec-mistral-small-31)
+- [Mistral-Small-4](#sec-mistral-small-4)
 - [Magistral-Small-2509](#sec-magistral-small-2509)
 - [Voxtral](#sec-voxtral)
 - [Gemma-3](#sec-gemma-3)
 - [Gemma-3n](#sec-gemma-3n)
 - [Qwen2-VL](#sec-qwen2-vl)
 - [Qwen2.5-VL](#sec-qwen25-vl)
+- [Qwen3.5](#sec-qwen3-5)
 - [GLM-4.6V](#sec-glm-4-6v)
 - [SmolVLM2](#sec-smolvlm2)
 - [LFM2-VL](#sec-lfm2-vl)
@@ -108,6 +110,12 @@ Please make sure to install vision lib via `pip install 'mistral-common[opencv]=
 base_model: mistralai/Mistral-Small-3.1-24B-Instruct-2503
 ```
 
+### Mistral-Small-4 {#sec-mistral-small-4}
+
+```yaml
+base_model: mistralai/Mistral-Small-4-119B-2603
+```
+
 ### Magistral-Small-2509 {#sec-magistral-small-2509}
 
 ::: {.callout-tip}
@@ -184,6 +192,14 @@ base_model: Qwen/Qwen3-VL-4B-Instruct
 chat_template: qwen2_vl  # same as qwen2-vl
 ```
 
+### Qwen3.5 {#sec-qwen3-5}
+
+```yaml
+base_model: Qwen/Qwen3.5-9B
+
+chat_template: qwen3_5
+```
+
 ### GLM-4.6V {#sec-glm-4-6v}
 
 Both GLM-4.6V (106B MoE) and GLM-4.6V-Flash (9B) are supported.

docs/optimizations.qmd

@@ -54,6 +54,13 @@ These techniques save VRAM by changing how activations are handled.
 - Activation Offloading: moves activations to CPU RAM or disk, trading I/O overhead for VRAM.
 - Learn more: [Gradient Checkpointing and Offloading Docs](gradient_checkpointing.qmd)
 
+### Layer Offloading
+
+Offloads frozen (non-trainable) decoder layer parameters to CPU and streams them back to GPU one layer at a time during forward/backward passes using CUDA stream prefetching. Especially effective for LoRA/QLoRA where most parameters are frozen.
+
+- **Config:** `layer_offloading: true`
+- **Learn more:** [Layer Offloading Docs](gradient_checkpointing.qmd#enabling-layer-offloading)
+
 ### Cut Cross Entropy (CCE)
 
 Reduces VRAM usage by using an optimized cross-entropy loss calculation.

docs/rlhf.qmd

@@ -16,8 +16,12 @@ feedback. Various methods include, but not limited to:
 - [Identity Preference Optimization (IPO)](#ipo)
 - [Kahneman-Tversky Optimization (KTO)](#kto)
 - [Odds Ratio Preference Optimization (ORPO)](#orpo)
-- [Group Relative Policy Optimization (GRPO)](#grpo)
+- [Group Relative Policy Optimization (GRPO)](#grpo) — see also the [GRPO deep dive](grpo.qmd) for async features, custom rewards, and scaling
 - [Group Reward-Decoupled Policy Optimization (GDPO)](#gdpo)
+- [Energy-Based Fine-Tuning (EBFT)](#ebft) — see also the [EBFT guide](ebft.qmd) for detailed mode comparisons and configuration
+- [NeMo Gym Integration](#nemo-gym-integration)
+
+For help choosing between these methods, see [Choosing a Fine-Tuning Method](choosing_method.qmd).
 
 
 ## RLHF using Axolotl
@@ -513,7 +517,7 @@ The input format is a simple JSON input with customizable fields based on the ab
 ### GRPO
 
 ::: {.callout-tip}
-Check out our [GRPO cookbook](https://github.com/axolotl-ai-cloud/grpo_code).
+Check out our [GRPO cookbook](https://github.com/axolotl-ai-cloud/grpo_code). For a comprehensive guide covering async training, custom rewards, importance sampling, and scaling, see the [GRPO deep dive](grpo.qmd).
 :::
 
 In the latest GRPO implementation, `vLLM` is used to significantly speedup trajectory generation during training. In this example, we're using 4 GPUs - 2 for training, and 2 for vLLM:
@@ -721,6 +725,213 @@ trl:
 
 For more information, see [GRPO docs](https://huggingface.co/docs/trl/v0.17.0/en/grpo_trainer#loss-types).
 
+#### Async GRPO
+
+Async GRPO overlaps vLLM generation with training by producing rollouts in a background thread. While the model trains on the current batch, the next batch is already being generated. This can significantly reduce wall-clock time per step.
+
+```yaml
+trl:
+  use_data_producer: true     # Enable data producer protocol
+  use_vllm: true
+  async_prefetch: true         # Generate rollouts in background thread
+  prefetch_depth: 1            # Number of rollouts to prefetch
+  vllm_sync_interval: 2        # Sync weights to vLLM every N steps
+```
+
+::: {.callout-note}
+Because the background thread generates completions with slightly stale model weights, async GRPO uses importance sampling correction to account for the distribution shift. This is controlled by `vllm_importance_sampling_correction: true` (default when async is enabled).
+:::
+
+##### vLLM LoRA Sync
+
+By default, weight sync to vLLM merges the LoRA adapter into the base model and broadcasts all parameters via NCCL. LoRA sync is a faster alternative that saves only the adapter weights to the filesystem and has vLLM load them natively using Punica kernels.
+
+```yaml
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_target_linear: true
+
+trl:
+  vllm_lora_sync: true         # Enable native LoRA sync
+```
+
+When `vllm_lora_sync: true` is set, axolotl automatically selects the LoRA-aware vLLM serve module. Start vLLM as usual:
+
+```bash
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve config.yaml
+```
+
+Then start training on a separate GPU:
+
+```bash
+CUDA_VISIBLE_DEVICES=1 axolotl train config.yaml
+```
+
+::: {.callout-tip}
+LoRA sync is especially beneficial with multi-GPU training (FSDP/DeepSpeed), where NCCL merge-sync can cause GPU contention with vLLM generation.
+:::
+
+##### Streaming Partial Batch
+
+Instead of scoring the entire batch at once, streaming mode scores one prompt group at a time. This enables finer-grained zero-advantage skipping and reduces peak memory usage during scoring.
+
+```yaml
+trl:
+  streaming_partial_batch: true
+```
+
+##### Importance Sampling Correction
+
+When using async prefetch, completions are generated from a slightly older version of the model. Importance sampling (IS) correction adjusts the policy gradient to account for this distribution shift.
+
+```yaml
+trl:
+  vllm_importance_sampling_correction: true   # Enable IS correction
+  importance_sampling_level: token             # 'token' or 'sequence'
+  off_policy_mask_threshold: 0.5              # Mask sequences with IS ratio below this
+```
+
+- `importance_sampling_level: token` applies per-token IS ratios (recommended with Liger kernel)
+- `importance_sampling_level: sequence` applies per-sequence IS ratios
+- `off_policy_mask_threshold` masks out sequences where the IS ratio indicates they are too far off-policy
+
+##### Replay Buffer
+
+The replay buffer caches rollout groups that had learning signal (non-zero reward variance) and uses them to replace zero-signal groups in later batches.
+
+```yaml
+trl:
+  replay_buffer_size: 100       # Max cached groups (0 = disabled)
+  replay_recompute_logps: true  # Recompute log-probs for replayed data (recommended)
+```
+
+::: {.callout-note}
+When `replay_recompute_logps: true` (default), old log-probabilities are recomputed using the current model weights. This fixes the IS mismatch that would otherwise occur when replaying stale data.
+:::
+
+##### Deferred Re-rolling
+
+Failed prompts (where the model produces zero reward for all generations) are buffered and re-injected into later batches when the model may be better equipped to solve them.
+
+```yaml
+trl:
+  reroll_start_fraction: 0.5    # Start re-rolling after 50% of training
+  reroll_max_groups: 1          # Max groups to replace per batch
+```
+
+##### Zero-Advantage Batch Skipping
+
+When all advantages in a micro-batch are zero (no learning signal), the forward/backward pass is skipped entirely. This is enabled by default and logged as `skipped_zero_adv_batches=1`.
+
+```yaml
+trl:
+  skip_zero_advantage_batches: true   # default
+```
+
+##### Parallel Reward Workers
+
+Reward functions that use `signal.alarm()` (e.g., `math_verify`) must run in the main thread. Parallel reward workers use subprocesses to work around this limitation while enabling concurrent reward computation.
+
+```yaml
+trl:
+  reward_num_workers: 4         # Number of subprocess workers (1 = no parallelism)
+```
+
+##### Full Async GRPO Example
+
+```yaml
+base_model: Qwen/Qwen2.5-1.5B-Instruct
+
+vllm:
+    host: 0.0.0.0
+    port: 8000
+    gpu_memory_utilization: 0.35
+    dtype: auto
+
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_target_linear: true
+
+rl: grpo
+trl:
+  use_data_producer: true
+  use_vllm: true
+  async_prefetch: true
+  prefetch_depth: 1
+  vllm_sync_interval: 2
+  vllm_lora_sync: true
+  streaming_partial_batch: true
+  vllm_importance_sampling_correction: true
+  off_policy_mask_threshold: 0.5
+  importance_sampling_level: token
+  num_generations: 8
+  max_completion_length: 512
+  reward_funcs:
+    - rewards.accuracy_reward
+  reroll_start_fraction: 0.5
+  replay_buffer_size: 100
+  reward_num_workers: 4
+  skip_zero_advantage_batches: true
+
+datasets:
+  - path: AI-MO/NuminaMath-TIR
+    type: rewards.prompt_transform
+    split: train
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+max_steps: 500
+learning_rate: 1e-5
+bf16: true
+gradient_checkpointing: true
+```
+
+```bash
+# Terminal 1: Start vLLM on GPU 0
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve config.yaml
+
+# Terminal 2: Train on GPU 1
+CUDA_VISIBLE_DEVICES=1 axolotl train config.yaml
+```
+
+##### Multi-GPU Async GRPO
+
+Async GRPO supports FSDP and DeepSpeed ZeRO-3 for multi-GPU training. vLLM runs on one GPU while training is distributed across the remaining GPUs.
+
+**FSDP:**
+
+```yaml
+fsdp:
+  - full_shard
+  - auto_wrap
+fsdp_config:
+  fsdp_transformer_layer_cls_to_wrap: Qwen2DecoderLayer
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+```
+
+**DeepSpeed ZeRO-3:**
+
+```yaml
+deepspeed: deepspeed_configs/zero3_bf16.json
+gradient_checkpointing_kwargs:
+  use_reentrant: true   # Required for ZeRO-3
+```
+
+```bash
+# Terminal 1: Start vLLM on GPU 0
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve config.yaml
+
+# Terminal 2: Train on GPUs 0,1
+CUDA_VISIBLE_DEVICES=0,1 axolotl train config.yaml
+```
+
+::: {.callout-important}
+With multi-GPU async prefetch, only rank 0 generates completions in the background thread. Results are broadcast to all ranks on the main thread. This avoids FSDP/DeepSpeed collective deadlocks from unsynchronized background threads.
+:::
+
 ### GDPO
 
 GDPO (Group Reward-Decoupled Policy Optimization) extends GRPO for multi-reward training. It addresses the **reward advantage collapse** problem by normalizing each reward function independently before combining them.
@@ -830,6 +1041,306 @@ simpo_gamma: 0.5  # default in CPOTrainer
 
 This method uses the same dataset format as [DPO](#dpo).
 
+### EBFT {#ebft}
+
+::: {.callout-tip}
+For a detailed guide on EBFT modes, feature extraction, and configuration, see the [EBFT guide](ebft.qmd).
+:::
+
+EBFT (Energy-Based Fine-Tuning) fine-tunes language models by optimizing a **feature-matching loss** rather than relying on external reward functions. A frozen copy of the model extracts embeddings from both generated and ground-truth completions, and the generator is updated via REINFORCE to match the ground-truth feature moments.
+
+Paper: ["Matching Features, Not Tokens: Energy-Based Fine-Tuning of Language Models"](https://arxiv.org/abs/2603.12248) (Jelassi et al., 2026)
+
+**Key advantages:**
+
+- No reward model or verifier required — works on any (prompt, completion) data
+- Applicable to non-verifiable tasks (code, translation, creative writing)
+- Operates on model rollouts (not teacher forcing), reducing distribution shift
+
+EBFT supports two modes:
+
+- **Structured mode**: For QA/instruction data with prompt + completion pairs. Uses vLLM for generation (like GRPO).
+- **Strided mode**: For unstructured text without prompt/completion splits. Uses strided block-parallel generation with flex_attention — no vLLM needed.
+
+#### Structured Mode
+
+```yaml
+base_model: Qwen/Qwen3-4B
+
+rl: ebft
+
+ebft:
+  feature_layers: [0.25, 0.5, 0.75]    # Extract features at 25%, 50%, 75% depth
+  embed_method: last_token
+  use_whitening: false
+  alignment_coef: 1.0                    # Cosine similarity reward weight
+  diversity_coef: 1.0                    # Pairwise dot product penalty
+  ce_coef: 0.0                          # Cross-entropy on GT tokens (0 = off)
+
+trl:
+  num_generations: 4
+  max_completion_length: 256
+  temperature: 0.7
+  use_vllm: true
+  vllm_server_host: 0.0.0.0
+  vllm_server_port: 8000
+  vllm_lora_sync: true                   # LoRA adapter sync (recommended)
+  vllm_sync_interval: 3
+  use_data_producer: true
+  async_prefetch: true                   # Set false for sync mode
+  scale_rewards: true
+  loss_type: grpo
+  epsilon: 0.2
+
+vllm:
+  gpu_memory_utilization: 0.5
+  max_model_len: 2048
+
+datasets:
+  - path: nvidia/OpenCodeInstruct
+    type: ebft_opencode.transform
+    split: train[:500]
+
+adapter: lora
+lora_r: 16
+lora_alpha: 32
+lora_target_linear: true
+```
+
+```bash
+# Terminal 1: Start vLLM
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve config.yaml
+
+# Terminal 2: Train
+CUDA_VISIBLE_DEVICES=1 axolotl train config.yaml
+```
+
+#### Strided Mode
+
+For unstructured text (raw code, prose). No vLLM needed — runs on a single GPU.
+
+```yaml
+base_model: meta-llama/Llama-3.2-1B
+
+rl: ebft
+
+ebft:
+  mode: strided
+  stride: 8
+  context_length: 8
+  generate_max_len: 8
+  n_samples_per_prompt: 4
+  temperature: 0.6
+  feature_layers: [0.25, 0.5, 0.75]
+  embed_method: last_token
+  use_whitening: true
+  alignment_coef: 1.0
+  diversity_coef: 1.0
+  rl_coef: 1.0
+  ce_coef: 0.03
+  advantage_estimator: rloo
+
+datasets:
+  - path: nvidia/OpenCodeInstruct
+    type: ebft_strided_structured.transform
+    split: train[:1%]
+
+flash_attention: false
+flex_attention: true     # Strided mode uses flex_attention
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: true    # Required for flex_attention
+```
+
+```bash
+CUDA_VISIBLE_DEVICES=0 axolotl train config.yaml
+```
+
+::: {.callout-tip}
+See `examples/ebft/` for complete example configs covering Llama 1B/3B/8B and Qwen3 4B/8B models in both modes.
+:::
+
+#### EBFT Configuration Reference
+
+| Parameter | Default | Description |
+|-----------|---------|-------------|
+| `ebft.feature_layers` | `[0.25, 0.5, 0.75]` | Layer depths for feature extraction (fractional) |
+| `ebft.embed_method` | `last_token` | Feature pooling: `last_token`, `mean_pooling`, `concat` |
+| `ebft.use_whitening` | `false` | SVD whitening of feature dimensions |
+| `ebft.alignment_coef` | `1.0` | Cosine similarity reward weight |
+| `ebft.diversity_coef` | `1.0` | Pairwise dot product penalty weight |
+| `ebft.ce_coef` | `0.0` | Cross-entropy loss on ground-truth tokens |
+| `ebft.mode` | `structured` | `structured` (vLLM) or `strided` (no vLLM) |
+| `ebft.stride` | — | Tokens between anchor points (strided mode) |
+| `ebft.context_length` | — | Context window per block (strided mode) |
+| `ebft.generate_max_len` | — | Tokens to generate per block (strided mode) |
+| `ebft.n_samples_per_prompt` | — | Rollouts per document (strided mode) |
+| `ebft.advantage_estimator` | `grpo` | `grpo` or `rloo` (strided mode) |
+
+### NeMo Gym Integration
+
+[NeMo Gym](https://github.com/NVIDIA-NeMo/Gym) provides 50+ verified RL environments (math, coding, tool-use, reasoning) with deterministic reward signals. The axolotl integration supports both **single-turn** (call `/verify` after generation) and **multi-turn** (agent-based tool execution via `/run`).
+
+#### Single-Turn (Simplest)
+
+For environments that only need answer verification (math, coding challenges). No agent server needed — the reward function calls `/verify` directly on the resource server.
+
+```yaml
+base_model: Qwen/Qwen2.5-0.5B-Instruct
+
+rl: grpo
+chat_template: tokenizer_default
+
+trl:
+  use_vllm: false                          # Colocate mode (single GPU)
+  num_generations: 4
+  max_completion_length: 128
+  temperature: 0.9
+  reward_funcs:
+    - axolotl.integrations.nemo_gym.rewards.reward_nemo_gym_verify
+
+plugins:
+  - axolotl.integrations.nemo_gym.NemoGymPlugin
+
+nemo_gym_enabled: true
+nemo_gym_dir: ~/Gym
+nemo_gym_auto_start: false
+nemo_gym_head_port: 11000
+nemo_gym_datasets:
+  - path: resources_servers/reasoning_gym/data/train_basic_arithmetic.jsonl
+    server_name: reasoning_gym
+
+datasets:
+  - path: ~/Gym/resources_servers/reasoning_gym/data/train_basic_arithmetic.jsonl
+    type: chat_template
+    field_messages: responses_create_params.input
+    message_field_content: content
+    message_field_role: role
+```
+
+```bash
+# Terminal 1: Start NeMo Gym resource server
+cd ~/Gym && .venv/bin/ng_run \
+    "+config_paths=[resources_servers/reasoning_gym/configs/resources_only.yaml]" \
+    "+skip_venv_if_present=true"
+
+# Terminal 2: Train
+CUDA_VISIBLE_DEVICES=0 axolotl train config.yaml
+```
+
+::: {.callout-note}
+`nemo_gym_datasets.path` is relative to `nemo_gym_dir`. Don't use absolute paths or they will be double-joined.
+:::
+
+#### Multi-Turn with Async GRPO (Recommended)
+
+For environments with tool-use (weather, search, databases). An agent server orchestrates multi-turn interactions: generate → parse tool calls → execute tools → feed results back → repeat until done.
+
+```yaml
+base_model: Qwen/Qwen3-0.6B
+
+rl: grpo
+chat_template: tokenizer_default
+
+adapter: lora
+lora_r: 16
+lora_alpha: 32
+lora_target_modules: [q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj]
+
+trl:
+  use_vllm: true
+  vllm_mode: server
+  vllm_server_host: localhost
+  vllm_server_port: 8000
+  vllm_lora_sync: true
+  vllm_sync_interval: 5
+  use_data_producer: true
+  async_prefetch: true                     # 3x speedup
+  num_generations: 4
+  max_completion_length: 512
+  temperature: 0.8
+  reward_funcs:
+    - axolotl.integrations.nemo_gym.rewards.reward_env
+
+plugins:
+  - axolotl.integrations.nemo_gym.NemoGymPlugin
+
+nemo_gym_enabled: true
+nemo_gym_auto_start: false
+nemo_gym_head_port: 11000
+nemo_gym_multi_turn: true
+nemo_gym_verify_timeout: 120
+nemo_gym_datasets:
+  - path: resources_servers/example_single_tool_call/data/weather_tool_calling.jsonl
+    server_name: example_single_tool_call
+
+datasets:
+  - path: ~/Gym/resources_servers/example_single_tool_call/data/weather_tool_calling.jsonl
+    type: chat_template
+    field_messages: responses_create_params.input
+    message_field_content: content
+    message_field_role: role
+
+vllm:
+  gpu_memory_utilization: 0.85
+  max_model_len: 2048
+```
+
+Multi-turn requires three services running:
+
+```bash
+# Terminal 1: vLLM with LoRA + tool calling
+VLLM_ALLOW_RUNTIME_LORA_UPDATING=1 CUDA_VISIBLE_DEVICES=0 \
+    python -m vllm.entrypoints.openai.api_server \
+    --model Qwen/Qwen3-0.6B --max-model-len 2048 \
+    --gpu-memory-utilization 0.85 \
+    --enable-lora --max-lora-rank 64 \
+    --enable-auto-tool-choice --tool-call-parser hermes
+
+# Terminal 2: NeMo Gym servers (resource + model proxy + agent)
+cd ~/Gym && .venv/bin/ng_run \
+    "+config_paths=[configs/axolotl_tool_calling.yaml]" \
+    "+skip_venv_if_present=true"
+
+# Terminal 3: Training
+CUDA_VISIBLE_DEVICES=1 axolotl train config.yaml
+```
+
+::: {.callout-important}
+Multi-turn requires a NeMo Gym agent config YAML that defines three components: a resource server (tools + `/verify`), a model server proxy (forwards to your vLLM), and an agent server (orchestrates `/run`). See the [NeMo Gym README](https://github.com/NVIDIA-NeMo/Gym) for agent config format.
+:::
+
+#### NeMo Gym Prerequisites
+
+```bash
+# Clone and set up NeMo Gym
+git clone https://github.com/NVIDIA-NeMo/Gym.git ~/Gym
+cd ~/Gym
+uv venv --python 3.12 && source .venv/bin/activate && uv sync
+
+# Fix pycosat build (GCC 13+)
+CFLAGS="" uv pip install pycosat --python .venv/bin/python --no-build-isolation
+```
+
+#### NeMo Gym Configuration Reference
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `nemo_gym_enabled` | bool | — | Enable the NeMo Gym integration |
+| `nemo_gym_dir` | str | `~/Gym` | Path to NeMo Gym repo |
+| `nemo_gym_auto_start` | bool | `true` | Auto-start resource servers |
+| `nemo_gym_head_port` | int | `11000` | Head server port |
+| `nemo_gym_multi_turn` | bool | `false` | Enable multi-turn via agent `/run` |
+| `nemo_gym_verify_timeout` | int | `30` | Per-request timeout (seconds) |
+| `nemo_gym_datasets` | list | required | Dataset configs with `path` and `server_name` |
+
+#### Reward Functions
+
+| Function | Mode | Description |
+|----------|------|-------------|
+| `axolotl.integrations.nemo_gym.rewards.reward_nemo_gym_verify` | Single-turn | Calls `/verify`, returns binary reward |
+| `axolotl.integrations.nemo_gym.rewards.reward_env` | Multi-turn | Passthrough reward from agent `/run` |
+
 ### Using local dataset files
 
 ```yaml

docs/training_stability.qmd

@@ -0,0 +1,443 @@
+---
+title: "Training Stability & Debugging"
+order: 15
+description: "Guide to monitoring, debugging, and stabilizing training runs in axolotl"
+---
+
+This guide covers practical techniques for monitoring training health, diagnosing instability, and resolving common failures in both supervised fine-tuning (SFT) and reinforcement learning (GRPO/EBFT) workflows.
+
+## Monitoring Training
+
+### Key Metrics for SFT
+
+Every SFT run should be monitored through at least these four metrics:
+
+| Metric | What It Tells You | Healthy Range |
+|--------|-------------------|---------------|
+| `train/loss` | How well the model fits training data | Decreasing; typically 0.5--2.0 for chat fine-tuning |
+| `eval/loss` | Generalization performance | Tracks train loss with small gap; divergence signals overfitting |
+| `grad_norm` | Gradient magnitude | 0.1--10.0; spikes above 100 indicate instability |
+| `learning_rate` | Current LR from scheduler | Should follow expected schedule (warmup then decay) |
+
+::: {.callout-tip}
+## Set Up Logging Early
+Enable W&B or TensorBoard from the start. Debugging a failed run without metrics is guesswork.
+
+```yaml
+wandb_project: my-project
+wandb_run_id:   # optional, for resuming
+logging_steps: 1
+```
+:::
+
+### Key Metrics for RL (GRPO)
+
+GRPO training logs a richer set of metrics. These are the critical ones:
+
+| Metric | Healthy Range | Red Flag |
+|--------|---------------|----------|
+| `rewards/<name>/mean` | > 0.15 within 20 steps | Stays at 0 -- reward function is broken or task is too hard |
+| `reward_std` | > 0 on most steps | Always 0 -- no learning signal (all completions get the same reward) |
+| `frac_reward_zero_std` | < 0.8 | 1.0 on every step -- zero-advantage skip fires constantly, no gradient updates |
+| `grad_norm` | 0.001--1.0 | 0.0 is acceptable occasionally (zero-adv skip); > 10.0 is unstable |
+| `entropy` | 0.05--0.5 | < 0.01 suggests mode collapse; > 1.0 suggests the model is not converging |
+| `kl` | 0.0--0.5 | > 2.0 suggests policy has diverged too far from reference |
+| `sampling/sampling_logp_difference/mean` | < 0.1 | > 1.0 means policy has diverged far from vLLM server weights |
+| `sampling/importance_sampling_ratio/min` | > 0.1 | Near 0 indicates stale off-policy data; increase `vllm_sync_interval` |
+| `clip_ratio/region_mean` | < 0.1 | > 0.3 means PPO clipping is too aggressive |
+| `completions/mean_length` | Task-dependent | Monotonically increasing to max length suggests reward hacking |
+| `completions/clipped_ratio` | < 0.3 | > 0.8 means most completions hit `max_completion_length` -- increase it |
+
+::: {.callout-note}
+## EBFT-Specific Metrics
+For EBFT training, also monitor `ebft/alignment` (should trend upward, healthy 0.3--0.9), `ebft/diversity` (healthy 0.01--0.1; > 1.0 indicates mode collapse), and `ebft/cfm_loss` (should trend downward, < 10).
+:::
+
+## SFT Stability
+
+### Loss Plateau
+
+**Symptom**: Loss stops decreasing early in training, well above expected values.
+
+**Causes and fixes**:
+
+- **Learning rate too low**: Increase by 2--5x. Typical ranges: full fine-tune 1e-5 to 5e-5, LoRA 1e-4 to 3e-4.
+- **Insufficient warmup**: Set `warmup_steps` to 5--10% of total steps. Too-aggressive learning at the start can push the model into a flat region.
+- **Data quality**: Check that labels are correctly masked. Use `axolotl preprocess` and inspect tokenized samples to confirm only the target tokens are trainable.
+- **Weight decay too high**: Default 0.01 is usually fine. Values above 0.1 can suppress learning in LoRA.
+
+### Loss Spikes
+
+**Symptom**: Loss suddenly jumps by 2--10x then (possibly) recovers.
+
+**Causes and fixes**:
+
+- **Bad data samples**: A single malformed or extremely long example can cause a spike. Enable `sample_packing: false` temporarily and check if spikes correlate with specific batches.
+- **Learning rate too high**: Reduce by 2--5x, or increase warmup.
+- **Gradient accumulation mismatch**: Effective batch size = `micro_batch_size * gradient_accumulation_steps * num_gpus`. Very large effective batch sizes amplify gradient noise.
+- **Mixed precision issues**: With `bf16: true`, some operations can lose precision. If spikes are severe, try `fp32` for diagnosis.
+
+### Overfitting
+
+**Symptom**: Train loss keeps decreasing but eval loss starts increasing.
+
+**Fixes**:
+
+- Increase `val_set_size` (e.g., 0.05) and monitor `eval/loss`.
+- Reduce `num_epochs` or `max_steps`.
+- Increase `weight_decay` (try 0.01--0.1).
+- Use a smaller LoRA rank (`lora_r`). Typical values: 8--32.
+- Increase dropout: `lora_dropout: 0.05`.
+
+## RL/GRPO Stability
+
+### Reward Never Increases
+
+If `rewards/*/mean` stays at 0 for more than 20 steps:
+
+1. **Test reward function standalone**: Run it outside training with known inputs to verify it returns nonzero values.
+   ```bash
+   cd experiments && python -c "import my_rewards; print(my_rewards.accuracy_reward(...))"
+   ```
+2. **Check dataset columns**: The reward function receives `**kwargs` containing dataset columns. Verify the columns it needs (e.g., `answer`) are not removed by the dataset transform.
+3. **Check completion content**: Enable `log_completions: true` in the `trl:` config and inspect logged completions in W&B. If completions are empty or incoherent, the model may be too weak for the task.
+4. **Verify vLLM is serving the right model**: Hit the vLLM health endpoint and confirm the model name matches your config.
+
+### Entropy Collapse (Mode Collapse)
+
+**Symptom**: `entropy` drops below 0.01; all completions become nearly identical.
+
+**Fixes**:
+
+- Increase `temperature` in generation kwargs (try 0.8--1.0).
+- Reduce learning rate.
+- Add a KL penalty term (`beta` parameter in GRPO config).
+- Check that `num_generations` is sufficient (16+ gives better advantage estimates).
+
+### IS Ratio Divergence
+
+**Symptom**: `sampling/importance_sampling_ratio/min` drops near 0, or `sampling/sampling_logp_difference/mean` exceeds 1.0.
+
+This means the policy has diverged significantly from the weights used by vLLM for generation. The importance sampling correction becomes unreliable.
+
+**Fixes**:
+
+- Decrease `vllm_sync_interval` (sync weights more often).
+- Enable `off_policy_mask_threshold` (e.g., 0.5) to mask stale off-policy samples.
+- Use `importance_sampling_level: token` for finer-grained correction.
+
+### Gradient Norm Instability
+
+**Symptom**: `grad_norm` oscillates wildly or exceeds 10.0 regularly.
+
+**Fixes**:
+
+- Enable gradient clipping: `max_grad_norm: 1.0` (default in most configs).
+- Reduce learning rate.
+- Increase `gradient_accumulation_steps` to smooth out noisy batches.
+- Check for NaN issues (see next section).
+
+## MoE Weight Scale Drift
+
+**Symptom**: Model works on short prompts but loses coherence on long conversations — repeating itself, "philosophizing", or generating broken code. Particularly affects MoE models with recurrent/SSM components (e.g. DeltaNet linear attention).
+
+**Root cause**: In MoE models trained with AdamW, rarely-activated experts accumulate smaller second-moment estimates. This gives them a disproportionately large effective learning rate, causing their weights to drift to higher variance than the group norm. In recurrent components like `conv1d` in DeltaNet layers, this amplifies short-range context and washes out long-range state.
+
+**Detection**: Use `normalize_weight_scales` with `dry_run: true` to scan for anomalies without modifying weights:
+
+```yaml
+normalize_weight_scales:
+  - name_pattern: 'linear_attn\.conv1d\.weight'
+    threshold: 1.3
+    dry_run: true
+```
+
+This logs any tensors matching the pattern whose standard deviation exceeds 1.3x the group median. Example output:
+
+```
+normalize_weight_scales [DRY RUN]: pattern 'linear_attn\.conv1d\.weight' —
+  3/30 tensors outside 1.3x threshold (median std=0.062733):
+    layers.36.linear_attn.conv1d.weight: std=0.101870 (1.62x median)
+    layers.37.linear_attn.conv1d.weight: std=0.102362 (1.63x median)
+    layers.38.linear_attn.conv1d.weight: std=0.089227 (1.42x median)
+```
+
+Each rule accepts:
+
+- `name_pattern`: regex matched against parameter names. All matching tensors form a group.
+- `threshold`: flag tensors whose std deviates from the group median by more than this factor (default: 1.5).
+- `dry_run`: when `true`, log anomalies without modifying weights (default: `false`).
+
+Multiple rules can target different tensor patterns:
+
+```yaml
+normalize_weight_scales:
+  - name_pattern: 'linear_attn\.conv1d\.weight'
+    threshold: 1.3
+  - name_pattern: 'experts\.gate_up_proj'
+    threshold: 1.5
+    dry_run: true  # just check these, don't fix
+```
+
+The transform runs after model loading but before adapter injection, so it modifies the base model weights directly.
+
+## NaN and Inf Handling
+
+### Common Causes
+
+| Cause | Where It Manifests | Detection |
+|-------|-------------------|-----------|
+| FP8 zero-scale division | Forward pass logits | `grad_norm: nan`, loss becomes NaN immediately |
+| Gradient explosion | Backward pass | `grad_norm` spikes to inf, then loss goes NaN |
+| Bad data (empty sequences) | Logprob computation | NaN in specific batches only |
+| Numerical overflow in log-softmax | Loss computation | Large negative logprobs cause exp() overflow |
+
+### FP8-Specific NaN Issues
+
+FP8 quantization (`fp8: true`) can produce NaN when the activation quantization kernel divides by `max(abs(x)) / 448`. If the input tensor is all zeros (e.g., padding positions), the scale becomes 0, causing division by zero.
+
+**Fixes applied in axolotl**:
+
+- The `act_quant_kernel` has a zero-guard: `s = tl.where(s == 0, 1.0, s)`.
+- A safety net `nan_to_num(logits, nan=0.0)` is applied in `_get_per_token_logps_and_entropies`.
+- Embedding padding is zero-padded for FP8 compatibility.
+
+::: {.callout-important}
+## After Modifying Triton Kernels
+If you patch any Triton JIT kernel (e.g., the FP8 quantization kernels in transformers), you must clear the Triton cache for changes to take effect:
+
+```bash
+rm -rf ~/.triton/cache
+```
+:::
+
+### General NaN Debugging Steps
+
+1. **Enable anomaly detection** (slow, but pinpoints the source):
+   ```python
+   torch.autograd.set_detect_anomaly(True)
+   ```
+2. **Check grad_norm**: If it goes to NaN, the backward pass is the problem. If loss is NaN but grad_norm was fine on the previous step, the forward pass is the problem.
+3. **Reduce to single GPU, single batch**: Eliminate distributed training variables.
+4. **Inspect data**: Print the batch that triggers NaN. Look for empty sequences, extreme token IDs, or unexpected padding patterns.
+
+## OOM Debugging
+
+Out-of-memory errors are the most common training failure. Use this systematic approach, from least to most disruptive:
+
+### Step 1: Reduce Batch Size
+
+The single highest-impact change. VRAM scales roughly linearly with batch size.
+
+```yaml
+micro_batch_size: 1              # Start here
+gradient_accumulation_steps: 16  # Increase to maintain effective batch size
+```
+
+For GRPO specifically, the logits tensor for policy logprob computation can be very large. `batch_size * num_generations * seq_len * vocab_size` in bf16. For example, with `num_generations: 16` and `micro_batch_size: 8`, the logits tensor alone is:
+
+```
+8 * 16 * 2048 * 151936 * 2 bytes = ~75 GB  (way too large)
+```
+
+Reduce `micro_batch_size` to 2--4 for GRPO.
+
+### Step 2: Enable Gradient Checkpointing
+
+Trades compute for memory by recomputing activations during the backward pass instead of storing them.
+
+```yaml
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false     # Recommended default
+```
+
+::: {.callout-warning}
+## Reentrant Checkpointing Exceptions
+Some configurations require `use_reentrant: true`:
+
+- DeepSpeed ZeRO-3 (non-reentrant causes `CheckpointError`)
+- EBFT strided mode with flex_attention
+:::
+
+### Step 3: Use Quantization
+
+Load the base model in reduced precision:
+
+```yaml
+# 4-bit QLoRA
+adapter: qlora
+load_in_4bit: true
+
+# 8-bit
+load_in_8bit: true
+
+# FP8 (saves ~50% model VRAM, same compute speed as bf16)
+fp8: true
+```
+
+### Step 4: Reduce Sequence Length
+
+```yaml
+sequence_len: 1024     # Down from 2048 or 4096
+```
+
+For GRPO, also reduce `max_completion_length`. Memory scales quadratically with sequence length when using standard attention.
+
+### Step 5: Use Flash Attention
+
+Reduces attention memory from O(n^2) to O(n):
+
+```yaml
+flash_attention: true
+```
+
+### Step 6: Offload with DeepSpeed
+
+For extreme cases, offload optimizer states or parameters to CPU:
+
+```yaml
+deepspeed: deepspeed_configs/zero3_bf16.json
+```
+
+### Diagnosing the Specific Culprit
+
+Use the `profiler_steps` config option to capture GPU memory snapshots:
+
+```yaml
+profiler_steps: [1, 2]
+```
+
+This generates PyTorch profiler traces you can inspect to see exactly which tensor allocation caused the OOM.
+
+## Common Errors
+
+| Error Message | Likely Cause | Fix |
+|---------------|-------------|-----|
+| `exitcode: -9` | System RAM exhaustion | Reduce dataset size, `dataset_num_proc`, or number of data workers |
+| `exitcode: -7` (DeepSpeed) | DeepSpeed version issue | `pip install -U deepspeed` |
+| `CUDA out of memory` | GPU VRAM exhaustion | Follow OOM debugging steps above |
+| `RuntimeError: NCCL communicator was aborted` | GPU communication failure | See [NCCL docs](nccl.qmd); check `NCCL_DEBUG=INFO` output |
+| `ValueError: Asking to pad but the tokenizer does not have a padding token` | Missing pad token | Add `special_tokens: { pad_token: "<\|endoftext\|>" }` to config |
+| `'DummyOptim' object has no attribute 'step'` | DeepSpeed on single GPU | Remove `deepspeed:` section from config |
+| `unable to load strategy X` then `None is not callable` | Reward module not importable | Run `cd experiments && python -c "import my_rewards"` to check |
+| `generation_batch_size not divisible by num_generations` | micro_batch_size too small | Set `micro_batch_size >= num_generations` and make it divisible |
+| `'weight' must be 2-D` | FSDP1 flattened parameters | Use `fsdp_version: 2` or skip `unwrap_model` when FSDP is enabled |
+| `CheckpointError` (tensor count mismatch) | Non-reentrant checkpointing + ZeRO-3 or flex_attention | Set `use_reentrant: true` in `gradient_checkpointing_kwargs` |
+| `BFloat16` TypeError during weight sync | NumPy does not support bf16 | Fixed in axolotl's `weight_serde.py` (auto bf16 to fp16 conversion) |
+| `Content end boundary is before start boundary` | Chat template parsing issue | Check `eos_token` matches template; file a GitHub issue if persistent |
+| `CAS service error` during data processing | HuggingFace XET issue | Set `export HF_HUB_DISABLE_XET=1` |
+| Training hangs (multi-GPU) | FSDP + async prefetch deadlock | Set `async_prefetch: false` with FSDP |
+
+## Profiling
+
+### PyTorch Profiler
+
+Axolotl supports PyTorch profiler integration via the config:
+
+```yaml
+profiler_steps: [1, 2, 3]
+```
+
+This captures profiler traces for the specified steps. View them in TensorBoard:
+
+```bash
+tensorboard --logdir output_dir/runs
+```
+
+Or open the `.json` trace file in `chrome://tracing`.
+
+### CUDA Memory Snapshots
+
+For detailed memory analysis, use PyTorch's memory snapshot API. Add this to your training script or use it interactively:
+
+```python
+import torch
+
+# Enable memory history tracking
+torch.cuda.memory._record_memory_history()
+
+# ... run your training step ...
+
+# Save snapshot
+torch.cuda.memory._dump_snapshot("memory_snapshot.pickle")
+```
+
+Visualize with PyTorch's memory visualizer:
+
+```bash
+python -m torch.cuda.memory._viz memory_snapshot.pickle
+```
+
+### Quick GPU Memory Check
+
+During training, monitor GPU utilization in a separate terminal:
+
+```bash
+watch -n 1 nvidia-smi
+```
+
+For programmatic access within axolotl, the logged metrics `memory/max_alloc` and `memory/max_reserved` come from `torch.cuda.max_memory_allocated()` and `torch.cuda.max_memory_reserved()`. Note these report PyTorch's view of memory, which may differ from `nvidia-smi` (see [FAQ](faq.qmd)).
+
+## W&B and Logging
+
+### Enabling Logging
+
+```yaml
+wandb_project: my-project
+wandb_entity: my-team          # optional
+wandb_run_id: run-123          # optional, for resuming
+wandb_name: experiment-name    # optional
+logging_steps: 1               # log every step (recommended for RL)
+```
+
+### Debug Logging
+
+For detailed axolotl-internal debug output:
+
+```bash
+AXOLOTL_LOG_LEVEL=DEBUG axolotl train config.yaml 2>&1 | tee /tmp/training.log
+```
+
+::: {.callout-tip}
+## Always Log to a File
+Pipe training output to a log file so you can inspect it after the run:
+
+```bash
+axolotl train config.yaml 2>&1 | tee /tmp/my_run.log
+```
+:::
+
+### What Axolotl Logs
+
+**SFT metrics** (logged every `logging_steps`):
+
+- `train/loss`, `eval/loss` -- training and validation loss
+- `train/grad_norm` -- gradient L2 norm (before clipping)
+- `train/learning_rate` -- current learning rate
+- `memory/max_alloc`, `memory/max_reserved` -- peak GPU memory
+
+**GRPO/RL metrics** (logged every step):
+
+- `rewards/<name>/mean`, `rewards/<name>/std` -- per-reward-function statistics
+- `reward`, `reward_std` -- aggregated reward across all reward functions
+- `frac_reward_zero_std` -- fraction of prompt groups where all completions got the same reward
+- `completions/mean_length`, `completions/min_length`, `completions/max_length` -- completion token lengths
+- `completions/clipped_ratio` -- fraction of completions that hit the max length
+- `completions/mean_terminated_length`, `completions/min_terminated_length`, `completions/max_terminated_length` -- lengths of naturally terminated completions
+- `kl` -- KL divergence between policy and reference
+- `entropy` -- policy entropy (measure of output diversity)
+- `clip_ratio/region_mean`, `clip_ratio/low_mean`, `clip_ratio/high_mean` -- PPO clipping statistics
+- `sampling/sampling_logp_difference/mean`, `sampling/sampling_logp_difference/max` -- log-probability difference between policy and sampling distribution
+- `sampling/importance_sampling_ratio/min`, `sampling/importance_sampling_ratio/mean`, `sampling/importance_sampling_ratio/max` -- IS ratio statistics for off-policy correction
+- `num_tokens` -- total tokens processed
+
+### Reading W&B Charts
+
+For a healthy GRPO run, expect to see:
+
+1. **`reward/mean`**: Gradual upward trend. May start near 0 and reach 0.3--0.8 depending on task difficulty. Not monotonic -- fluctuations are normal.
+2. **`entropy`**: Gradual decrease from initial values (often 0.3--0.6) as the model becomes more confident. Should not collapse to near-zero.
+3. **`grad_norm`**: Mostly in the 0.001--1.0 range. Occasional 0.0 values are fine (zero-advantage skip). Persistent values above 10.0 need investigation.
+4. **`kl`**: Starts near 0 and grows slowly. If it shoots up rapidly, the policy is diverging from the reference.
+5. **`completions/mean_length`**: Should reflect the task's natural answer length. If it steadily increases to `max_completion_length`, the model may be reward-hacking by generating longer outputs.

docs/vllm_serving.qmd

@@ -0,0 +1,318 @@
+---
+title: "vLLM Serving for GRPO Training"
+description: "How to configure and run vLLM as a generation backend for GRPO reinforcement learning in Axolotl."
+format:
+  html:
+    toc: true
+    toc-depth: 3
+    number-sections: true
+execute:
+  enabled: false
+---
+
+## Overview {#sec-overview}
+
+GRPO (Group Relative Policy Optimization) trains a language model by generating completions, scoring them with reward functions, and updating the policy to favor higher-reward outputs. The generation step is the bottleneck: producing thousands of tokens per training step with the policy model is slow using standard HuggingFace generation.
+
+Axolotl uses [vLLM](https://github.com/vllm-project/vllm) as a high-throughput generation backend. vLLM runs as a separate process (either on a dedicated GPU or colocated on the training GPU) and serves completions via an HTTP API. The trainer sends prompts to vLLM, receives completions, scores them, and performs gradient updates.
+
+```
+┌──────────────────────┐       HTTP        ┌──────────────────────┐
+│   Trainer (GPU 1)    │ ───────────────── │   vLLM Server (GPU 0)│
+│                      │  prompts/compls   │                      │
+│  - Policy model      │ ◄──────────────── │  - Same base model   │
+│  - Reward scoring    │                   │  - Fast generation   │
+│  - Gradient updates  │  weight sync      │  - LoRA adapter      │
+│  - LoRA adapter      │ ─────────────────►│    (periodically     │
+│                      │  (every N steps)  │     updated)         │
+└──────────────────────┘                   └──────────────────────┘
+```
+
+::: {.callout-important}
+vLLM must serve the **same base model** specified in your training config. If the models do not match, weight synchronization will silently produce incorrect results.
+:::
+
+## Server Mode {#sec-server-mode}
+
+Server mode runs vLLM as an external process on dedicated GPU(s). This is the recommended configuration for most setups.
+
+### Starting the Server
+
+Use the `axolotl vllm-serve` command with your training config:
+
+```bash
+# Terminal 1: Start vLLM on GPU 0
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve grpo_config.yaml
+```
+
+```bash
+# Terminal 2: Start training on GPU 1
+CUDA_VISIBLE_DEVICES=1 axolotl train grpo_config.yaml
+```
+
+The server reads vLLM settings from the `vllm:` section of your config and starts an HTTP server (default: `http://0.0.0.0:8000`).
+
+::: {.callout-tip}
+Use `tmux` or `screen` to manage the vLLM server process. Typical startup time is 30-90 seconds depending on model size and whether CUDA graphs are captured.
+:::
+
+### Minimal Server Config
+
+```yaml
+base_model: Qwen/Qwen2.5-1.5B-Instruct
+
+vllm:
+  host: 0.0.0.0
+  port: 8000
+  gpu_memory_utilization: 0.85
+  dtype: auto
+  max_model_len: 4096
+
+rl: grpo
+trl:
+  use_vllm: true
+  vllm_server_host: 0.0.0.0
+  vllm_server_port: 8000
+  vllm_server_timeout: 300
+```
+
+### Multi-GPU vLLM
+
+For larger models, use tensor parallelism across multiple GPUs:
+
+```yaml
+vllm:
+  tensor_parallel_size: 2
+  gpu_memory_utilization: 0.85
+```
+
+```bash
+# vLLM on GPUs 2,3; training on GPUs 0,1
+CUDA_VISIBLE_DEVICES=2,3 axolotl vllm-serve grpo_config.yaml
+CUDA_VISIBLE_DEVICES=0,1 axolotl train grpo_config.yaml --num-processes 2
+```
+
+::: {.callout-note}
+Due to how TRL maps vLLM device indices, the vLLM instance should use the **last** N GPUs (highest device indices), while training uses the first N.
+:::
+
+## Colocate Mode {#sec-colocate-mode}
+
+Colocate mode runs vLLM on the same GPU as the trainer. This is useful when you only have a single GPU.
+
+```yaml
+trl:
+  use_vllm: true
+  vllm_mode: colocate
+  vllm_enable_sleep_mode: true
+```
+
+With `vllm_enable_sleep_mode: true`, vLLM offloads its VRAM allocation when not actively generating, freeing memory for training. When the trainer needs new completions, vLLM wakes up and reclaims VRAM.
+
+::: {.callout-warning}
+Colocate mode is significantly slower than server mode because generation and training cannot overlap. The GPU alternates between the two workloads. This mode is practical only for smaller models (up to ~3B on a 24 GB GPU).
+:::
+
+**When to use colocate mode:**
+
+- You have exactly one GPU
+- The model fits in memory with both vLLM and training active (with sleep mode), or is small enough to time-share
+- You accept the performance tradeoff for simpler setup (no separate vLLM process to manage)
+
+**When to use server mode:**
+
+- You have two or more GPUs
+- You want maximum throughput (generation overlaps with training via async prefetch)
+- You are running larger models (7B+)
+
+## LoRA Sync {#sec-lora-sync}
+
+LoRA sync is the recommended weight synchronization method when training with LoRA adapters. Instead of merging adapter weights into the base model and broadcasting the full merged weights over NCCL, it saves only the LoRA adapter files to the filesystem and tells vLLM to load them natively.
+
+### How It Works
+
+1. The trainer calls `model.save_pretrained()` to write the LoRA adapter weights to a temporary directory
+2. The trainer sends an HTTP POST to `/set_lora_adapter/` on the vLLM server
+3. vLLM loads the adapter using its native LoRA support (Punica kernels)
+4. Generation uses the updated adapter on the next request
+
+### Benefits
+
+- **Smaller sync payload**: Transfers ~40 MB of LoRA weights instead of ~1.4 GB+ of merged model weights (for a typical 0.5-3B model)
+- **No NCCL communicator**: Eliminates the need for a cross-GPU NCCL communication channel, removing GPU contention between vLLM generation and weight sync
+- **Faster sync**: ~200 ms per sync vs. 350 ms to 5+ seconds for NCCL merge sync
+- **Simpler multi-GPU**: No need to set up NCCL groups between trainer and vLLM processes
+
+### Configuration
+
+```yaml
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_target_linear: true
+
+trl:
+  vllm_lora_sync: true    # Enables LoRA sync mode
+  vllm_sync_interval: 5   # Sync every 5 training steps
+```
+
+Setting `vllm_lora_sync: true` automatically selects the LoRA-aware vLLM serve script (`axolotl.scripts.vllm_serve_lora`). You do not need to set `vllm.serve_module` manually.
+
+::: {.callout-important}
+LoRA sync requires that you are training with a LoRA adapter (`adapter: lora` or `adapter: qlora`). It is not applicable to full fine-tuning.
+:::
+
+## Weight Synchronization {#sec-weight-sync}
+
+During GRPO training, the policy model on the trainer is continuously updated via gradient steps. The vLLM server, however, still holds the old weights. Periodically, the trainer must push updated weights to vLLM so that future generations reflect the improved policy.
+
+### Sync Interval
+
+The `vllm_sync_interval` parameter controls how often weights are synced:
+
+```yaml
+trl:
+  vllm_sync_interval: 5   # Sync every 5 optimizer steps
+```
+
+**Tradeoffs:**
+
+- **Lower interval** (e.g., 1-3): Fresher generations, better on-policy data, but more sync overhead per step
+- **Higher interval** (e.g., 5-10): Less overhead, but generations become increasingly off-policy between syncs
+- **Recommended**: 3-5 for most setups. Axolotl includes importance sampling correction (`vllm_importance_sampling_correction: true`) to handle mild distribution mismatch from stale vLLM weights.
+
+### Sync Methods
+
+| Method | Config | Payload | Mechanism | Typical Time |
+|--------|--------|---------|-----------|-------------|
+| **LoRA sync** | `vllm_lora_sync: true` | LoRA adapter only (~40 MB) | Filesystem + HTTP | ~200 ms |
+| **NCCL merge sync** | Default (no lora_sync) | Full merged weights (~1.4 GB+) | HTTP trigger + NCCL broadcast | 350 ms - 5 s |
+
+::: {.callout-tip}
+If you are training with LoRA (which is recommended for GRPO), always enable `vllm_lora_sync: true`. The performance difference is substantial, especially as training progresses and NCCL contention increases.
+:::
+
+### Importance Sampling Correction
+
+When vLLM weights are stale (between syncs), the generated data is slightly off-policy. Axolotl can correct for this:
+
+```yaml
+trl:
+  vllm_importance_sampling_correction: true
+  importance_sampling_level: token          # 'token' or 'sequence'
+  off_policy_mask_threshold: 0.5            # KL threshold for masking stale sequences
+```
+
+- **Token-level IS** is recommended when using Liger kernel (sequence-level has numerical issues with chunked computation)
+- **Off-policy sequence masking (OPSM)** drops sequences that have diverged too far from the current policy, providing a safety net against stale data
+
+## Restart Requirements {#sec-restart}
+
+::: {.callout-warning}
+**vLLM must be restarted between training runs.** Weight syncs from a previous run leave the server in a corrupted state. If you start a new training run against a stale vLLM server, the model may fail to learn.
+:::
+
+### When to Restart
+
+- Before every new training experiment
+- After a training run crashes or is interrupted
+- If you change the base model in your config
+
+### How to Restart
+
+Killing vLLM reliably requires terminating both the main process and its background EngineCore subprocess:
+
+```bash
+# Kill all vLLM-related processes
+pkill -9 -f "vllm|EngineCore"
+
+# Verify GPU memory is freed
+nvidia-smi
+
+# Restart the server
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve grpo_config.yaml
+```
+
+::: {.callout-tip}
+A single `kill` often does not fully stop vLLM. Always use `kill -9` and verify with `nvidia-smi` that GPU memory has been released before restarting.
+:::
+
+### Health Check
+
+The vLLM server exposes a health endpoint. Wait for it to return 200 before starting training:
+
+```bash
+# For the LoRA serve script (trailing slash required)
+curl http://localhost:8000/health/
+
+# For the default TRL serve script
+curl http://localhost:8000/health
+```
+
+## Configuration Reference {#sec-config-reference}
+
+### vLLM Server Options (`vllm:` section)
+
+These control the vLLM server process started by `axolotl vllm-serve`.
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `host` | str | `0.0.0.0` | Host address for the vLLM server |
+| `port` | int | `8000` | Port for the vLLM server |
+| `device` | str | `auto` | Device to use for vLLM |
+| `tensor_parallel_size` | int | `None` | Number of GPUs for tensor parallelism |
+| `data_parallel_size` | int | `None` | Number of data parallel replicas |
+| `gpu_memory_utilization` | float | `0.9` | Fraction of GPU memory for vLLM (0.0-1.0) |
+| `dtype` | str | `auto` | Data type (`auto`, `float16`, `bfloat16`) |
+| `max_model_len` | int | `None` | Maximum model context length. Set explicitly if the default is too large for your GPU |
+| `enable_prefix_caching` | bool | `None` | Enable prefix caching for repeated prompt prefixes |
+| `enable_reasoning` | bool | `None` | Enable reasoning mode for models with thinking tokens |
+| `reasoning_parser` | str | `None` | Parser for reasoning output |
+| `enforce_eager` | bool | `None` | Disable CUDA graph capture (required for some architectures like Qwen3.5 hybrid attention) |
+| `serve_module` | str | `None` | Python module for vLLM serve script. Auto-set when `vllm_lora_sync: true` |
+| `worker_extension_cls` | str | `None` | vLLM worker extension class for weight sync |
+
+### Trainer vLLM Options (`trl:` section)
+
+These control how the trainer interacts with vLLM.
+
+| Option | Type | Default | Description |
+|--------|------|---------|-------------|
+| `use_vllm` | bool | `false` | Enable vLLM for generation |
+| `vllm_mode` | str | `None` | `server` (external process) or `colocate` (same GPU) |
+| `vllm_server_host` | str | `0.0.0.0` | Host of the vLLM server to connect to |
+| `vllm_server_port` | int | `8000` | Port of the vLLM server to connect to |
+| `vllm_server_timeout` | int | `None` | Timeout in seconds for vLLM requests |
+| `vllm_lora_sync` | bool | `false` | Sync LoRA adapters via filesystem instead of NCCL merge |
+| `vllm_sync_interval` | int | `None` | Sync weights every N optimizer steps |
+| `vllm_enable_sleep_mode` | bool | `None` | Offload vLLM VRAM when idle (colocate mode) |
+| `vllm_guided_decoding_regex` | str | `None` | Regex constraint for guided decoding |
+
+For async pipeline and off-policy correction options, see the [GRPO Configuration Reference](grpo.qmd#configuration-reference).
+
+## Complete Example {#sec-complete-example}
+
+For a full working GRPO config including vLLM, LoRA sync, async generation, rewards, and dataset setup, see the [GRPO Quick Start](grpo.qmd#quick-start). That config includes all the vLLM settings covered in this guide.
+
+```bash
+# Terminal 1: Start vLLM
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve grpo_config.yaml
+
+# Wait for health check to pass
+curl http://localhost:8000/health/
+
+# Terminal 2: Start training
+CUDA_VISIBLE_DEVICES=1 axolotl train grpo_config.yaml
+```
+
+## Troubleshooting {#sec-troubleshooting}
+
+| Problem | Likely Cause | Solution |
+|---------|-------------|----------|
+| Training hangs waiting for vLLM | Server not started or wrong port | Check `curl http://localhost:8000/health/` and verify `vllm_server_host`/`vllm_server_port` match |
+| OOM on vLLM GPU | `gpu_memory_utilization` too high or `max_model_len` too large | Reduce `gpu_memory_utilization` to 0.7 or set `max_model_len` explicitly |
+| OOM on training GPU | Batch too large for policy logprobs | Reduce `micro_batch_size` or `num_generations` |
+| Accuracy stays at zero | Stale vLLM from previous run | Restart vLLM: `pkill -9 -f "vllm\|EngineCore"`, verify with `nvidia-smi`, restart |
+| `ResponseValidationError` from vLLM | Missing logprobs in response | Ensure you are using the correct serve module (auto-selected with `vllm_lora_sync: true`) |
+| Weight sync takes 5+ seconds | NCCL contention with vLLM generation | Switch to `vllm_lora_sync: true` to eliminate NCCL |
+| `async_prefetch` deadlocks with FSDP | Background threads run unsynchronized FSDP collectives | Set `async_prefetch: false` when using FSDP or DeepSpeed multi-GPU |

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@e8ad129\""
+    "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@63b15e6\""
    ]
   },
   {

examples/ebft/README.md

@@ -0,0 +1,211 @@
+# Energy-Based Fine-Tuning (EBFT)
+
+EBFT is an integration of ["Matching Features, Not Tokens: Energy-Based Fine-Tuning of Language Models"](https://arxiv.org/abs/2603.12248) (Jelassi et al., 2026) into axolotl.
+
+## Overview
+
+EBFT fine-tunes language models by optimizing a **feature-matching loss** rather than relying on external reward functions or verifiers. A frozen copy of the model (the "feature network") extracts embeddings from both generated and ground-truth completions, and the generator is updated via REINFORCE to match the ground-truth feature moments.
+
+**Key advantages over SFT:**
+- Operates on model rollouts (not teacher forcing), reducing distribution shift
+- Provides dense sequence-level supervision without a task-specific verifier
+- Improves both downstream accuracy and validation cross-entropy simultaneously
+
+**Key advantages over RLVR:**
+- No reward model or verifier required — works on any (prompt, completion) data
+- Applicable to non-verifiable tasks (e.g., raw code, translation, creative writing)
+- Maintains distributional calibration (low feature-matching loss)
+
+## Two Modes
+
+EBFT supports two modes depending on your data format:
+
+### Structured Mode (`mode: structured`, default)
+For **QA/instruction data** with prompt + completion pairs (e.g., OpenCodeInstruct, ALMA translation).
+- Extends GRPOTrainer — uses vLLM for fast rollout generation
+- RLOO advantages and clipped policy gradient from GRPO
+- Feature-matching rewards replace external reward functions
+
+### Strided Mode (`mode: strided`)
+For **unstructured text** without prompt/completion splits (e.g., raw code, prose, SwallowCode).
+- Uses **strided block-parallel generation** — multiple short rollouts at different anchor points within a document
+- No vLLM needed — generation uses custom strided attention masks
+- Uses **torch flex_attention** with compiled block masks for efficient fused attention kernels (~2x faster than eager attention)
+- Compatible with gradient checkpointing via automatic dtype normalization
+- This is the core EBFT algorithm from the paper (Section F)
+
+### Common to both modes:
+- **Frozen feature network** — deep copy of the model at initialization (frozen, eval mode)
+- **Feature extraction** — hidden states at configurable layer depths (default: 25%, 50%, 75%), L2-normalized per layer before concatenation
+- **Feature-matching rewards** — cosine similarity (alignment) minus pairwise dot-product (diversity), scaled by 2 per paper equation (7)
+- **SVD whitening** — decorrelates feature dimensions; the paper shows removing it causes the largest degradation
+- **CFM loss tracking** — conditional feature-matching loss (paper eq 2) logged as `ebft/cfm_loss`
+- **FSDP2 compatible** — feature network stays outside FSDP wrapping (frozen, inference-only)
+
+## Quick Start
+
+### Structured Mode (QA data + vLLM)
+
+```bash
+# 1. Start vLLM server (LoRA serve module auto-selected when vllm_lora_sync: true)
+CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve examples/ebft/qwen3-4b-ebft-structured-async.yaml
+
+# 2. Train on a separate GPU
+CUDA_VISIBLE_DEVICES=1 axolotl train examples/ebft/qwen3-4b-ebft-structured-async.yaml
+```
+
+### Strided Mode (unstructured text)
+
+```bash
+# No vLLM needed — strided generation is built-in
+axolotl train examples/ebft/llama-3b-ebft-strided-fft.yaml
+```
+
+## Configuration
+
+### Common EBFT Settings
+
+```yaml
+rl: ebft
+
+ebft:
+  # Feature network: which layers to extract hidden states from
+  # Values are fractions of total depth (0.0 = embedding, 1.0 = final layer)
+  feature_layers: [0.25, 0.5, 0.75]
+
+  # How to pool per-token hidden states into sequence embeddings
+  # Options: "last_token" (recommended), "mean_pooling", "concat"
+  embed_method: last_token
+
+  # SVD whitening — strongly recommended (paper shows largest degradation without it)
+  use_whitening: true
+
+  # Reward = alignment_coef * alignment - diversity_coef * diversity
+  # Per paper Variant (i) (eq 49): alignment uses cosine similarity (normalized),
+  # diversity uses raw dot product — both are bounded after whitening.
+  alignment_coef: 1.0
+  diversity_coef: 1.0
+
+  # Cross-entropy loss on ground-truth tokens (mixed objective, paper Section 2.1)
+  # 0.0 = pure feature matching; 0.03 = recommended balance; 0.1 = CE-dominated
+  ce_coef: 0.0
+```
+
+### Strided Mode Settings
+
+```yaml
+ebft:
+  mode: strided
+  stride: 8              # tokens between anchor points (paper default: 8)
+  context_length: 8      # context window per block (paper default: 8)
+  generate_max_len: 8    # tokens generated per block (paper default: 8)
+  n_samples_per_prompt: 4  # independent rollouts per document (>= 2 for RLOO)
+  temperature: 0.6
+  rl_coef: 1.0           # RL loss weight
+  advantage_estimator: rloo  # rloo (recommended), group_norm, or reinforce
+```
+
+### Structured Mode Settings (via TRL)
+
+```yaml
+trl:
+  num_generations: 4           # samples per prompt
+  max_completion_length: 256   # max tokens to generate
+  temperature: 1.0
+  use_vllm: true
+  scale_rewards: true
+  loss_type: grpo
+  epsilon: 0.2
+```
+
+### Dataset Format
+
+**Structured mode** — QA data with prompt + ground-truth completion:
+```yaml
+datasets:
+  - path: nvidia/OpenCodeInstruct
+    type: ebft_opencode.transform
+```
+Transform returns: `{"prompt": ..., "ground_truth": ...}`
+
+**Strided mode** — raw text tokenized to fixed length:
+```yaml
+datasets:
+  - path: sjelassi/swallow_code_20m
+    type: ebft_pretrain.transform
+```
+Transform returns: `{"input_ids": ..., "attention_mask": ..., "labels": ...}`
+
+## How It Works
+
+### Structured Mode
+1. **Generate**: For each prompt, generate `num_generations` completions via vLLM
+2. **Extract features**: Forward both generated and ground-truth sequences through the frozen feature network
+3. **Compute rewards**: `2 * alignment - 2 * diversity` (paper eq 7)
+4. **RLOO advantages**: subtract leave-one-out group mean
+5. **Policy gradient**: clipped PPO-style loss
+
+### Strided Mode
+1. **Anchor selection**: Pick `num_blocks = (seq_len - gen_len - ctx_len) / stride + 1` anchor points across the document
+2. **Block-parallel generation**: At each anchor, generate `gen_len` tokens using a custom strided attention mask via `flex_attention` compiled block masks
+3. **Feature extraction**: Forward the full sequence (prompt + generated) through the frozen feature network **with the strided attention mask** — this is critical for correct feature representations
+4. **Per-block rewards**:
+   - **Alignment** = `2 * cosine_similarity(gen_block_emb, gt_block_emb)` — normalized, bounded in [-2, 2]
+   - **Diversity** = `2 * mean_pairwise_dot_product(gen_block_embs)` — raw dot product on whitened vectors
+   - **Reward** = `alignment_coef * alignment - diversity_coef * diversity`
+5. **RLOO advantages**: leave-one-out baseline across `n_samples_per_prompt` rollouts per block
+6. **Policy gradient**: REINFORCE loss on generated tokens, weighted by per-block advantages
+
+### Tracked Metrics
+
+| Metric | Description |
+|--------|-------------|
+| `ebft/alignment` | Mean cosine similarity between generated and GT features (higher = better) |
+| `ebft/diversity` | Mean pairwise similarity between samples (lower = more diverse) |
+| `ebft/mean_reward` | alignment - diversity (should trend upward) |
+| `ebft/cfm_loss` | Conditional feature-matching loss ‖E[φ(ŷ)] - φ(y)‖² (paper eq 2, lower = better) |
+| `ebft/rl_loss` | REINFORCE policy gradient loss |
+| `ebft/ce_loss` | Cross-entropy loss on ground-truth tokens (when `ce_coef > 0`) |
+| `ebft/advantages_std` | RLOO advantage standard deviation (should be non-zero) |
+
+## Tips and Recommendations
+
+### Reward coefficients
+- **`use_whitening: true`**: Strongly recommended. The paper's ablation (Figure 7) shows removing whitening causes the largest performance degradation. Safe to use with `diversity_coef > 0`.
+- **`diversity_coef`**: Default 1.0. Per the paper's Variant (i) (eq 49), alignment uses cosine similarity while diversity uses raw dot product. After whitening, both are bounded and on compatible scales.
+- **`n_samples_per_prompt`**: Must be >= 2 for diversity and RLOO. 4 is the paper's default.
+- **`ce_coef`**: The paper ablates `γ ∈ {0, 0.03, 0.1}`. `0.03` balances CE and RL signals; `0.1` causes CE to dominate the gradient. `0.0` gives pure feature matching.
+
+### Feature extraction
+- **`feature_layers: [0.25, 0.5, 0.75]`**: Extracts and concatenates hidden states from 25%, 50%, 75% depth. Each layer is L2-normalized independently before concatenation. The paper shows this works better than mean pooling or single-layer extraction.
+- **`embed_method: last_token`**: Uses the last token's hidden state per block. The paper shows this outperforms mean pooling (Figure 7).
+
+### Performance
+- **`torch_compile: true`**: Recommended for strided mode. Provides additional speedup via graph compilation.
+- **flex_attention**: Strided mode automatically uses `flex_attention` with compiled block masks when available (~2x faster than eager attention). Works with gradient checkpointing via automatic dtype normalization. Falls back to eager attention with dense 4D masks if flex_attention is unavailable.
+
+### Memory
+- EBFT requires a frozen copy of the model (the feature network), roughly doubling model memory.
+- **LoRA** is recommended to reduce trainable parameter memory. The feature network is always a frozen copy of the base model (without LoRA adapters).
+- With 2 GPUs visible, the trainer automatically places the feature network on the second GPU.
+- **FSDP2** is supported — the feature network stays outside FSDP wrapping since it's frozen and inference-only. With `cpu_ram_efficient_loading`, the feature network is loaded separately from pretrained weights.
+
+## Example Configs
+
+| Config | Mode | Model | Description |
+|--------|------|-------|-------------|
+| `llama-1b-ebft-opencode.yaml` | Structured | Llama-3.2-1B | QA coding with vLLM |
+| `llama-1b-ebft-opencode-novllm.yaml` | Structured | Llama-3.2-1B | QA coding without vLLM |
+| `llama-3b-ebft-strided-fft.yaml` | Strided | Llama-3.2-3B | Unstructured code with LoRA |
+| `llama-1b-ebft-strided.yaml` | Strided | Llama-3.2-1B | Quick validation |
+
+## Citation
+
+```bibtex
+@article{jelassi2026matching,
+  title={Matching Features, Not Tokens: Energy-Based Fine-Tuning of Language Models},
+  author={Jelassi, Samy and Kwun, Mujin and Zhao, Rosie and Li, Yuanzhi and Fusi, Nicolo and Du, Yilun and Kakade, Sham M. and Domingo-Enrich, Carles},
+  journal={arXiv preprint arXiv:2603.12248},
+  year={2026}
+}
+```

examples/ebft/ebft_opencode.py

@@ -0,0 +1,28 @@
+"""
+Dataset transform for nvidia/OpenCodeInstruct with EBFT.
+
+Maps the dataset's `input` (prompt) and `output` (code solution) fields
+to the format expected by the EBFT trainer.
+"""
+
+
+def transform(cfg, *args, **kwargs):
+    def transform_fn(example, tokenizer=None):
+        return {
+            "prompt": [
+                {"role": "user", "content": example["input"]},
+            ],
+            "ground_truth": example["output"],
+        }
+
+    return transform_fn, {
+        "remove_columns": [
+            "id",
+            "domain",
+            "generation_algorithm",
+            "llm_judgement",
+            "unit_tests",
+            "tests_execution_status",
+            "average_test_score",
+        ]
+    }

examples/ebft/ebft_pretrain.py

@@ -0,0 +1,31 @@
+"""
+Dataset transform for unstructured text data with strided EBFT.
+
+Tokenizes raw text into fixed-length input_ids for the strided trainer.
+Sequences are padded to sequence_len for uniform batching.
+"""
+
+
+def transform(cfg, *args, **kwargs):
+    seq_len = cfg.sequence_len
+
+    def transform_fn(example, tokenizer=None):
+        text = example.get("question", example.get("text", ""))
+        if tokenizer is None:
+            return {"prompt": text}
+
+        encoded = tokenizer(
+            text,
+            truncation=True,
+            max_length=seq_len,
+            padding="max_length",
+            add_special_tokens=True,
+            return_tensors=None,
+        )
+        return {
+            "input_ids": encoded["input_ids"],
+            "attention_mask": encoded["attention_mask"],
+            "labels": list(encoded["input_ids"]),
+        }
+
+    return transform_fn, {"remove_columns": ["question", "answer"]}

examples/ebft/ebft_strided_structured.py

@@ -0,0 +1,80 @@
+"""
+Dataset transform for structured (prompt, completion) data with strided EBFT.
+
+Tokenizes prompt and completion separately, concatenates into a single
+input_ids sequence, and marks prompt tokens with labels=-100 so the
+strided trainer knows where to place anchors (completion span only).
+
+Works with datasets that have chat-style fields (e.g., nvidia/OpenCodeInstruct).
+"""
+
+
+def transform(cfg, *args, **kwargs):
+    seq_len = cfg.sequence_len
+
+    def transform_fn(example, tokenizer=None):
+        # Extract prompt and completion from the example
+        prompt_text = example.get(
+            "input", example.get("prompt", example.get("question", ""))
+        )
+        completion_text = example.get(
+            "output", example.get("completion", example.get("answer", ""))
+        )
+
+        if tokenizer is None:
+            return {"prompt": prompt_text}
+
+        pad_id = tokenizer.pad_token_id or tokenizer.eos_token_id
+
+        # Tokenize prompt and completion separately
+        prompt_enc = tokenizer(
+            prompt_text,
+            truncation=False,
+            add_special_tokens=True,
+            return_tensors=None,
+        )
+        completion_enc = tokenizer(
+            completion_text,
+            truncation=False,
+            add_special_tokens=False,
+            return_tensors=None,
+        )
+
+        prompt_ids = prompt_enc["input_ids"]
+        completion_ids = completion_enc["input_ids"]
+
+        # Truncate to fit within seq_len (prioritize keeping prompt + some completion)
+        total_len = len(prompt_ids) + len(completion_ids)
+        if total_len > seq_len:
+            # Truncate completion first, then prompt if needed
+            max_completion = seq_len - len(prompt_ids)
+            if max_completion < 1:
+                # Prompt alone exceeds seq_len — truncate prompt, keep at least 1 completion token
+                prompt_ids = prompt_ids[: seq_len - 1]
+                completion_ids = completion_ids[:1]
+            else:
+                completion_ids = completion_ids[:max_completion]
+
+        input_ids = prompt_ids + completion_ids
+        prompt_length = len(prompt_ids)
+
+        # Labels: -100 for prompt tokens, input_ids for completion tokens
+        labels = [-100] * prompt_length + completion_ids
+
+        # Pad to seq_len
+        pad_len = seq_len - len(input_ids)
+        attention_mask = [1] * len(input_ids) + [0] * pad_len
+        labels = labels + [-100] * pad_len
+        input_ids = input_ids + [pad_id] * pad_len
+
+        return {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "labels": labels,
+            "prompt_length": prompt_length,
+        }
+
+    # Signal to remove all original columns (filtered to existing ones at map time)
+    return transform_fn, {
+        "remove_columns": "__all__",
+    }

examples/ebft/llama-1b-ebft-opencode-novllm.yaml

@@ -0,0 +1,64 @@
+# EBFT validation config — no vLLM, uses HF generate for simplicity
+# Run: CUDA_VISIBLE_DEVICES=0 axolotl train examples/ebft/llama-1b-ebft-opencode-novllm.yaml
+
+base_model: meta-llama/Llama-3.2-1B
+chat_template: llama3
+rl: ebft
+
+ebft:
+  feature_layers: [0.25, 0.5, 0.75]
+  embed_method: last_token
+  use_whitening: false
+  alignment_coef: 1.0
+  diversity_coef: 1.0
+  ce_coef: 0.0
+
+trl:
+  num_generations: 4
+  max_completion_length: 128
+  temperature: 1.0
+  use_vllm: false
+  scale_rewards: true
+  loss_type: grpo
+  epsilon: 0.2
+
+datasets:
+  - path: nvidia/OpenCodeInstruct
+    type: ebft_opencode.transform
+    split: train[:1%]
+
+sequence_len: 512
+micro_batch_size: 2
+gradient_accumulation_steps: 2
+num_epochs: 1
+max_steps: 10
+
+learning_rate: 1.0e-5
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+warmup_steps: 2
+weight_decay: 0.01
+
+adapter: lora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_linear: true
+
+bf16: auto
+flash_attention: true
+gradient_checkpointing: true
+
+special_tokens:
+  pad_token: "<|end_of_text|>"
+
+val_set_size: 0.0
+output_dir: ./outputs/ebft-validation
+
+wandb_project: ebft
+wandb_run_id:
+wandb_watch:
+wandb_log_model:
+
+logging_steps: 1
+save_steps: 100

examples/ebft/llama-1b-ebft-opencode.yaml

@@ -0,0 +1,81 @@
+# EBFT: Energy-Based Fine-Tuning with Llama-3.2-1B on OpenCodeInstruct
+#
+# Paper: "Matching Features, Not Tokens" (Jelassi et al., 2026)
+# https://arxiv.org/abs/2603.12248
+#
+# Prerequisites:
+#   1. Start vLLM server on a separate GPU:
+#      CUDA_VISIBLE_DEVICES=1 python -m trl.scripts.vllm_serve \
+#          --model meta-llama/Llama-3.2-1B \
+#          --host 0.0.0.0 --port 8000 \
+#          --gpu-memory-utilization 0.4 --dtype bfloat16
+#
+#   2. Run training:
+#      CUDA_VISIBLE_DEVICES=0 axolotl train examples/ebft/llama-1b-ebft-opencode.yaml
+
+base_model: meta-llama/Llama-3.2-1B
+chat_template: llama3
+
+# --- Training method ---
+rl: ebft
+
+# --- EBFT configuration ---
+ebft:
+  feature_layers: [0.25, 0.5, 0.75]  # extract hidden states at 25%, 50%, 75% depth
+  embed_method: last_token            # pool to sequence embedding via last token
+  use_whitening: false                # SVD whitening (disable for speed in small runs)
+  alignment_coef: 1.0                 # cosine similarity with ground-truth features
+  diversity_coef: 1.0                 # pairwise similarity penalty
+  ce_coef: 0.0                        # cross-entropy on ground-truth (0 = pure feature matching)
+
+# --- Generation settings (via TRL/GRPO infrastructure) ---
+trl:
+  num_generations: 4                  # samples per prompt for RLOO
+  max_completion_length: 256          # max generated tokens
+  temperature: 1.0
+  use_vllm: true
+  scale_rewards: true
+  loss_type: grpo
+  epsilon: 0.2
+
+# --- Dataset ---
+datasets:
+  - path: nvidia/OpenCodeInstruct
+    type: ebft_opencode.transform
+    split: train[:1%]        # first 1% for validation runs
+
+# --- Training hyperparameters ---
+sequence_len: 1024
+micro_batch_size: 2
+gradient_accumulation_steps: 4
+num_epochs: 1
+max_steps: 50
+
+learning_rate: 1.0e-5
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+warmup_steps: 5
+weight_decay: 0.01
+
+# --- LoRA (recommended to reduce memory with frozen feature network) ---
+adapter: lora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_linear: true
+
+# --- Hardware ---
+bf16: auto
+flash_attention: true
+gradient_checkpointing: true
+
+special_tokens:
+  pad_token: "<|end_of_text|>"
+
+val_set_size: 0.0
+output_dir: ./outputs/ebft-llama-1b-opencode
+
+# --- Logging ---
+use_tensorboard: true
+logging_steps: 1
+save_steps: 25

examples/ebft/llama-1b-ebft-strided-structured.yaml

@@ -0,0 +1,65 @@
+# EBFT Strided Structured Mode: For structured (prompt, completion) data
+# Uses strided block-parallel generation on completion spans — no vLLM needed.
+#
+# Run: CUDA_VISIBLE_DEVICES=0 axolotl train examples/ebft/llama-1b-ebft-strided-structured.yaml
+
+base_model: meta-llama/Llama-3.2-1B
+rl: ebft
+
+ebft:
+  mode: strided                       # strided block-parallel generation
+  stride: 8                           # tokens between anchor points
+  context_length: 8                   # context window per block
+  generate_max_len: 8                 # tokens to generate per block
+  n_samples_per_prompt: 4             # rollouts per document
+  temperature: 0.6
+  top_p: 1.0
+  feature_layers: [0.25, 0.5, 0.75]
+  embed_method: last_token
+  use_whitening: true
+  alignment_coef: 1.0
+  diversity_coef: 1.0
+  rl_coef: 1.0
+  ce_coef: 0.03                       # small CE weight for structured data
+  advantage_estimator: rloo
+  min_completion_prefix: 8            # skip anchors too close to prompt boundary
+
+datasets:
+  - path: nvidia/OpenCodeInstruct
+    type: ebft_strided_structured.transform
+    split: train[:1%]
+
+sequence_len: 2048
+micro_batch_size: 1
+gradient_accumulation_steps: 2
+num_epochs: 1
+# max_steps: 10
+
+learning_rate: 1.0e-6
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+warmup_steps: 5
+
+adapter: lora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_linear: true
+
+bf16: auto
+flash_attention: false  # strided EBFT overrides to flex_attention (or eager fallback) at runtime
+flex_attention: true    # fused flex_attention kernel compiles itself; don't set torch_compile: true
+                        # (full-model compile conflicts with gradient checkpointing + flex_attention)
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: true   # required for flex_attention (non-reentrant causes CheckpointError)
+
+special_tokens:
+  pad_token: "<|end_of_text|>"
+
+val_set_size: 0.0
+output_dir: ./outputs/ebft-strided-structured
+
+wandb_project: ebft
+logging_steps: 1
+save_steps: 100

examples/ebft/llama-1b-ebft-strided.yaml

@@ -0,0 +1,60 @@
+# EBFT Strided Mode: For unstructured text data (raw code, prose)
+# Uses strided block-parallel generation — no vLLM needed.
+#
+# Run: CUDA_VISIBLE_DEVICES=0 axolotl train examples/ebft/llama-1b-ebft-strided.yaml
+
+base_model: meta-llama/Llama-3.2-1B
+rl: ebft
+
+ebft:
+  mode: strided                       # strided block-parallel generation
+  stride: 8                           # tokens between anchor points
+  context_length: 8                   # context window per block
+  generate_max_len: 8                 # tokens to generate per block
+  n_samples_per_prompt: 4             # rollouts per document
+  temperature: 0.6
+  top_p: 1.0
+  feature_layers: [0.25, 0.5, 0.75]
+  embed_method: last_token
+  use_whitening: true
+  alignment_coef: 1.0
+  diversity_coef: 1.0
+  rl_coef: 1.0
+  ce_coef: 0.0
+  advantage_estimator: rloo
+
+datasets:
+  - path: sjelassi/swallow_code_20m
+    type: ebft_pretrain.transform
+    split: train[:100]
+
+sequence_len: 256
+micro_batch_size: 1
+gradient_accumulation_steps: 2
+num_epochs: 1
+max_steps: 5
+
+learning_rate: 1.0e-6
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+warmup_steps: 2
+
+adapter: lora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.05
+lora_target_linear: true
+
+bf16: auto
+flash_attention: false  # strided EBFT overrides to flex_attention (or eager fallback) at runtime
+gradient_checkpointing: true
+
+special_tokens:
+  pad_token: "<|end_of_text|>"
+
+val_set_size: 0.0
+output_dir: ./outputs/ebft-strided-validation
+
+wandb_project: ebft
+logging_steps: 1
+save_steps: 100

examples/ebft/llama-3b-ebft-strided-fft.yaml

@@ -0,0 +1,69 @@
+# EBFT Strided: LoRA Llama-3.2-3B on SwallowCode, 100 steps
+# Actor on GPU 0, frozen feature network on GPU 1
+#
+# Run: CUDA_VISIBLE_DEVICES=0,1 python -m axolotl.cli.train examples/ebft/llama-3b-ebft-strided-fft.yaml
+
+base_model: meta-llama/Llama-3.2-3B
+rl: ebft
+
+ebft:
+  mode: strided
+  stride: 8
+  context_length: 8
+  generate_max_len: 8
+  n_samples_per_prompt: 4
+  temperature: 0.6
+  top_p: 1.0
+  feature_layers: [0.25, 0.5, 0.75]
+  embed_method: last_token
+  use_whitening: true
+  alignment_coef: 1.0
+  diversity_coef: 1.0
+  rl_coef: 1.0
+  ce_coef: 0.0   # paper recommends 0.03 for mixed objective; 0.1 causes CE to dominate
+  advantage_estimator: rloo
+
+datasets:
+  - path: sjelassi/swallow_code_20m
+    type: ebft_pretrain.transform
+    split: train[:5000]
+
+sequence_len: 1024
+micro_batch_size: 1
+gradient_accumulation_steps: 4
+num_epochs: 1
+max_steps: 100
+
+learning_rate: 1.0e-5
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+warmup_steps: 10
+weight_decay: 0.01
+
+adapter: lora
+lora_r: 32
+lora_alpha: 64
+lora_dropout: 0.05
+lora_target_linear: true
+
+bf16: auto
+torch_dtype: bfloat16
+flash_attention: false
+gradient_checkpointing: true
+torch_compile: true
+gradient_checkpointing_kwargs:
+  use_reentrant: true
+ddp: false
+device_map:
+  "": 0
+
+special_tokens:
+  pad_token: "<|end_of_text|>"
+
+val_set_size: 0.0
+output_dir: ./outputs/ebft-llama3b-strided
+
+wandb_project: ebft
+wandb_name: llama3b-strided-lora-100steps
+logging_steps: 1
+save_steps: 50

examples/ebft/llama-8b-ebft-strided-fft.yaml

@@ -0,0 +1,58 @@
+# EBFT Strided: Full-parameter Llama-3.1-8B on SwallowCode, 100 steps
+# Feature network is CPU-offloaded to fit in single 32GB GPU
+#
+# Run: CUDA_VISIBLE_DEVICES=0 python -m axolotl.cli.train examples/ebft/llama-8b-ebft-strided-fft.yaml
+
+base_model: meta-llama/Llama-3.1-8B
+rl: ebft
+
+ebft:
+  mode: strided
+  stride: 8
+  context_length: 8
+  generate_max_len: 8
+  n_samples_per_prompt: 4
+  temperature: 0.6
+  top_p: 1.0
+  feature_layers: [0.25, 0.5, 0.75]
+  embed_method: last_token
+  use_whitening: true
+  alignment_coef: 1.0
+  diversity_coef: 1.0
+  rl_coef: 1.0
+  ce_coef: 0.0
+  advantage_estimator: rloo
+
+datasets:
+  - path: sjelassi/swallow_code_20m
+    type: ebft_pretrain.transform
+    split: train[:5000]
+
+sequence_len: 1024
+micro_batch_size: 1
+gradient_accumulation_steps: 4
+num_epochs: 1
+max_steps: 100
+
+learning_rate: 1.0e-6
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+warmup_steps: 10
+weight_decay: 0.01
+
+bf16: auto
+flash_attention: false  # strided EBFT uses flex_attention at runtime
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+
+special_tokens:
+  pad_token: "<|end_of_text|>"
+
+val_set_size: 0.0
+output_dir: ./outputs/ebft-llama8b-strided
+
+wandb_project: ebft
+wandb_name: llama8b-strided-fft-100steps
+logging_steps: 1
+save_steps: 50

examples/ebft/qwen35-4b-ebft-structured-async.yaml

@@ -0,0 +1,86 @@
+# EBFT Structured Mode: Qwen3.5-4B (hybrid linear attention)
+#
+# Qwen3.5 uses hybrid attention: linear attention (conv1d) on 3/4 of layers,
+# full attention every 4th layer. This tests EBFT compatibility.
+#
+# Prerequisites:
+#   1. Start vLLM on GPU 0:
+#      CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve examples/ebft/qwen35-4b-ebft-structured-async.yaml
+#
+#   2. Run training on GPU 1:
+#      CUDA_VISIBLE_DEVICES=1 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
+#          axolotl train examples/ebft/qwen35-4b-ebft-structured-async.yaml
+
+base_model: Qwen/Qwen3.5-4B
+
+rl: ebft
+
+ebft:
+  feature_layers: [0.25, 0.5, 0.75]
+  embed_method: last_token
+  use_whitening: false
+  alignment_coef: 1.0
+  diversity_coef: 1.0
+  ce_coef: 0.0
+
+trl:
+  num_generations: 4
+  max_completion_length: 256
+  temperature: 0.7
+  use_vllm: true
+  vllm_server_host: 0.0.0.0
+  vllm_server_port: 8000
+  scale_rewards: true
+  loss_type: grpo
+  epsilon: 0.2
+  generation_kwargs:
+    stop_token_ids: [248044, 248046]  # <|endoftext|>, <|im_end|>
+  chat_template_kwargs:
+    enable_thinking: false
+  async_prefetch: true
+  vllm_server_timeout: 300
+
+vllm:
+  gpu_memory_utilization: 0.5
+  max_model_len: 2048
+  serve_module: axolotl.scripts.vllm_serve_lora
+  enforce_eager: true
+
+datasets:
+  - path: nvidia/OpenCodeInstruct
+    type: ebft_opencode.transform
+    split: train[:500]
+
+sequence_len: 1024
+micro_batch_size: 1
+gradient_accumulation_steps: 4
+num_epochs: 1
+max_steps: 10
+
+learning_rate: 5.0e-6
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+warmup_steps: 3
+weight_decay: 0.01
+
+adapter: lora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.0
+# Target full-attention q/k/v/o on layers 3,7,11,15,19,23,27,31 + MLP on all layers
+# Avoids linear_attn modules (in_proj_qkv, in_proj_z, etc.) which break vLLM LoRA
+lora_target_modules: ".*\\.layers\\.(3|7|11|15|19|23|27|31)\\.self_attn\\.(q|k|v|o)_proj|.*\\.mlp\\.(gate|up|down)_proj"
+
+bf16: auto
+flash_attention: true
+gradient_checkpointing: true
+
+special_tokens:
+  pad_token: "<|endoftext|>"
+
+val_set_size: 0.0
+output_dir: ./outputs/ebft-qwen35-4b-structured-async
+
+wandb_project: ebft
+logging_steps: 1
+save_steps: 50

examples/ebft/qwen35-4b-ebft-structured.yaml

@@ -0,0 +1,77 @@
+# EBFT Structured Mode: Qwen3.5-4B (hybrid linear attention)
+#
+# Qwen3.5 uses hybrid attention: linear attention (conv1d) on 3/4 of layers,
+# full attention every 4th layer. This tests EBFT compatibility.
+#
+# Prerequisites:
+#   1. Start vLLM on GPU 0:
+#      CUDA_VISIBLE_DEVICES=0 trl vllm-serve --model Qwen/Qwen3.5-4B \
+#          --gpu-memory-utilization 0.5 --max-model-len 2048 --enforce-eager
+#
+#   2. Run training on GPU 1:
+#      CUDA_VISIBLE_DEVICES=1 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
+#          axolotl train examples/ebft/qwen35-4b-ebft-structured.yaml
+
+base_model: Qwen/Qwen3.5-4B
+
+rl: ebft
+
+ebft:
+  feature_layers: [0.25, 0.5, 0.75]
+  embed_method: last_token
+  use_whitening: false
+  alignment_coef: 1.0
+  diversity_coef: 1.0
+  ce_coef: 0.0
+
+trl:
+  num_generations: 4
+  max_completion_length: 256
+  temperature: 0.7
+  use_vllm: true
+  vllm_server_host: 0.0.0.0
+  vllm_server_port: 8000
+  scale_rewards: true
+  loss_type: grpo
+  epsilon: 0.2
+  generation_kwargs:
+    stop_token_ids: [248044, 248046]  # <|endoftext|>, <|im_end|>
+  chat_template_kwargs:
+    enable_thinking: false  # disable Qwen3.5 thinking mode for shorter completions
+
+datasets:
+  - path: nvidia/OpenCodeInstruct
+    type: ebft_opencode.transform
+    split: train[:500]
+
+sequence_len: 1024
+micro_batch_size: 1
+gradient_accumulation_steps: 4
+num_epochs: 1
+max_steps: 10
+
+learning_rate: 5.0e-6
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+warmup_steps: 3
+weight_decay: 0.01
+
+adapter: lora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.0
+lora_target_modules: ".*\\.layers\\.(3|7|11|15|19|23|27|31)\\.self_attn\\.(q|k|v|o)_proj|.*\\.mlp\\.(gate|up|down)_proj"
+
+bf16: auto
+flash_attention: true
+gradient_checkpointing: true
+
+special_tokens:
+  pad_token: "<|endoftext|>"
+
+val_set_size: 0.0
+output_dir: ./outputs/ebft-qwen35-4b-structured
+
+wandb_project: ebft
+logging_steps: 1
+save_steps: 50

examples/ebft/qwen35-9b-ebft-structured.yaml

@@ -0,0 +1,82 @@
+# EBFT Structured Mode: Qwen3.5-9B (hybrid linear attention)
+#
+# Prerequisites:
+#   1. Start vLLM on GPU 0:
+#      CUDA_VISIBLE_DEVICES=0 axolotl vllm-serve examples/ebft/qwen35-9b-ebft-structured.yaml
+#
+#   2. Run training on GPU 1:
+#      CUDA_VISIBLE_DEVICES=1 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
+#          axolotl train examples/ebft/qwen35-9b-ebft-structured.yaml
+
+base_model: Qwen/Qwen3.5-9B
+
+rl: ebft
+
+ebft:
+  feature_layers: [0.25, 0.5, 0.75]
+  embed_method: last_token
+  use_whitening: false
+  alignment_coef: 1.0
+  diversity_coef: 1.0
+  ce_coef: 0.0
+
+trl:
+  num_generations: 4
+  max_completion_length: 256
+  temperature: 0.7
+  use_vllm: true
+  vllm_server_host: 0.0.0.0
+  vllm_server_port: 8000
+  scale_rewards: true
+  loss_type: grpo
+  epsilon: 0.2
+  generation_kwargs:
+    stop_token_ids: [248044, 248046]  # <|endoftext|>, <|im_end|>
+  chat_template_kwargs:
+    enable_thinking: false
+  vllm_server_timeout: 300
+
+vllm:
+  gpu_memory_utilization: 0.7
+  max_model_len: 2048
+  serve_module: axolotl.scripts.vllm_serve_lora
+  enforce_eager: true
+
+datasets:
+  - path: nvidia/OpenCodeInstruct
+    type: ebft_opencode.transform
+    split: train[:500]
+
+sequence_len: 1024
+micro_batch_size: 1
+gradient_accumulation_steps: 4
+num_epochs: 1
+max_steps: 10
+
+learning_rate: 3.0e-6
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+warmup_steps: 3
+weight_decay: 0.01
+
+adapter: lora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.0
+# Target full-attention q/k/v/o on layers 3,7,11,15,19,23,27,31 + MLP on all layers
+# Avoids linear_attn modules (in_proj_qkv, in_proj_z, etc.) which break vLLM LoRA
+lora_target_modules: ".*\\.layers\\.(3|7|11|15|19|23|27|31)\\.self_attn\\.(q|k|v|o)_proj|.*\\.mlp\\.(gate|up|down)_proj"
+
+bf16: auto
+flash_attention: true
+gradient_checkpointing: true
+
+special_tokens:
+  pad_token: "<|endoftext|>"
+
+val_set_size: 0.0
+output_dir: ./outputs/ebft-qwen35-9b-structured
+
+wandb_project: ebft
+logging_steps: 1
+save_steps: 50

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

@@ -1,8 +1,5 @@
 base_model: google/gemma-3-1b-it
 
-model_type: Gemma3ForCausalLM
-cls_model_config: Gemma3TextConfig
-
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
@@ -27,6 +24,11 @@ datasets:
 val_set_size: 0.0
 output_dir: ./outputs/out
 
+# Freeze vision tower
+unfrozen_parameters:
+  - ^model\.language_model\..*
+  - ^lm_head\..*
+
 adapter: qlora
 lora_r: 32
 lora_alpha: 16

examples/gemma3/gemma-3-270m-qlora.yml

@@ -1,8 +1,5 @@
 base_model: google/gemma-3-270m-it
 
-model_type: Gemma3ForCausalLM
-cls_model_config: Gemma3TextConfig
-
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
@@ -27,6 +24,11 @@ datasets:
 val_set_size: 0.0
 output_dir: ./outputs/out
 
+# Freeze vision tower
+unfrozen_parameters:
+  - ^model\.language_model\..*
+  - ^lm_head\..*
+
 adapter: qlora
 lora_r: 32
 lora_alpha: 16

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

@@ -1,9 +1,5 @@
 base_model: google/gemma-3-4b-it
 
-# Need to set else transformers tries to load vision too
-model_type: Gemma3ForCausalLM
-cls_model_config: Gemma3TextConfig
-
 load_in_4bit: true
 
 # gemma3 doesn't seem to play nice with ddp
@@ -24,6 +20,11 @@ dataset_prepared_path: last_run_prepared
 val_set_size: 0.01
 output_dir: ./outputs/out
 
+# Freeze vision tower
+unfrozen_parameters:
+  - ^model\.language_model\..*
+  - ^lm_head\..*
+
 adapter: qlora
 lora_model_dir:
 

examples/gemma4/26b-a4b-moe-qlora.yaml

@@ -0,0 +1,104 @@
+# Gemma 4 26B-A4B MoE QLoRA with ScatterMoE kernels
+#
+# Validated: 50 steps on FineTome-100k, loss 7.4 -> 2.4, single RTX 5090 (32GB)
+#
+# Key notes:
+# - Flash Attention 2 is NOT supported (global_head_dim=512 > FA2 max of 256).
+#   Use sdp_attention instead.
+# - Gemma 4 is multimodal (text+vision+audio). For text-only SFT, restrict
+#   LoRA to the text backbone via lora_target_linear_modules regex.
+# - MoE experts use `experts_implementation: scattermoe` — Gemma 4 embeds MoE
+#   directly in the decoder layer (no SparseMoeBlock), so we register ScatterMoE
+#   via the transformers ExpertsInterface.
+# - Expert LoRA targets are `experts.gate_up_proj` / `experts.down_proj`
+#   (no `mlp.` prefix, unlike Qwen/Mixtral).
+# - micro_batch_size: 1 fits 2048 seq_len on 32GB GPU with SDP attention.
+#   Use micro_batch_size: 4 with 1024 seq_len, or on 48GB+ GPUs.
+
+base_model: google/gemma-4-26B-A4B
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+  - axolotl.integrations.kernels.KernelsPlugin
+  - axolotl.integrations.liger.LigerPlugin
+use_kernels: true
+use_scattermoe: true
+experts_implementation: scattermoe
+torch_compile: false
+liger_layer_norm: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+liger_rms_norm_gated: true
+strict: false
+
+chat_template: gemma4
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:10%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+val_set_size: 0.05
+output_dir: ./outputs/gemma4-26b-a4b-qlora
+
+sequence_len: 2048
+sample_packing: true
+
+load_in_4bit: true
+quantize_moe_experts: true
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0
+
+# Restrict LoRA to text backbone only (skip vision/audio encoders).
+# lora_target_modules is intentionally empty — all module targeting is done
+# via regex in lora_target_linear_modules below.
+lora_target_modules: []
+lora_target_linear_modules:
+  - language_model\.model\.layers\.\d+\.self_attn\.(q|k|v|o)_proj
+
+# MoE expert LoRA (3D Parameter tensors, not nn.Linear)
+lora_target_parameters:
+  - experts.gate_up_proj
+  - experts.down_proj
+
+lora_mlp_kernel: false
+lora_qkv_kernel: false
+lora_o_kernel: false
+
+bnb_config_kwargs:
+  bnb_4bit_use_double_quant: true
+
+wandb_project: gemma4-qlora
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+activation_offloading: true
+logging_steps: 1
+
+# FA2 not supported — Gemma4 global_head_dim=512 exceeds FA2 max of 256
+flash_attention: false
+sdp_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/glm45/README.md

@@ -58,6 +58,14 @@ datasets:
 - **LoRA kernels**: Incompatible with this model. Must be explicitly disabled (`lora_*_kernel: false`).
 - Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
 
+### GGUF / llama.cpp loading error (missing tensors)
+
+If you see `missing tensor 'blk.X.attn_norm.weight'` when loading a GLM-4 / GLM4-MoE model in llama.cpp, this is likely
+caused by `num_nextn_predict_layers` being set to `1` in `config.json` while the MTP weights were not exported (possible
+after PEFT/QLoRA training).
+
+**Fix:** Set `"num_nextn_predict_layers": 0` in your `config.json` before converting to GGUF.
+
 ## Optimization Guides
 
 Please check the [Optimizations doc](https://docs.axolotl.ai/docs/optimizations.html).

examples/mistral4/README.md

@@ -0,0 +1,82 @@
+# Finetune Mistral Small 4 with Axolotl
+
+Mistral Small 4 is a 119B parameter (6.5B active) multimodal MoE model from MistralAI that unifies instruct, reasoning, and coding capabilities into a single model. It is available on HuggingFace at [Mistral-Small-4-119B-2603](https://huggingface.co/mistralai/Mistral-Small-4-119B-2603).
+
+Thanks to the team at MistralAI for giving us early access to prepare for this release.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
+
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage
+
+3. Install transformers from main
+
+  ```bash
+  pip install git+https://github.com/huggingface/transformers.git
+  ```
+
+4. Run one of the example configs:
+
+  ```bash
+  # text-only
+  axolotl train examples/mistral4/qlora-text.yml  # no experts ~69 GiB, experts ~93 GiB
+  axolotl train examples/mistral4/fft-text.yml
+
+  # text + vision
+  # run: wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
+  axolotl train examples/mistral4/qlora-vision.yml  # no experts ~68 GiB
+  axolotl train examples/mistral4/fft-vision.yml
+  ```
+
+Note: FFT configs provided as reference. Please adjust hyperparameters as needed.
+
+## Reasoning Effort
+
+The chat template supports a `reasoning_effort` variable to control the model's reasoning depth:
+
+- `"none"` — instruct mode (default)
+- `"high"` — reasoning mode with explicit thinking steps
+
+Pass it via `chat_template_kwargs` under your dataset config:
+
+```yaml
+datasets:
+  - path: your/dataset
+    type: chat_template
+    chat_template_kwargs:
+      reasoning_effort: high
+```
+
+## Thinking Support
+
+The chat template supports a `thinking` content type in assistant messages for training on reasoning traces (rendered as `[THINK]...[/THINK]` blocks).
+
+To use thinking datasets, add the `thinking` mapping via `message_property_mappings`:
+
+```yaml
+datasets:
+  - path: your/thinking-dataset
+    type: chat_template
+    message_property_mappings:
+      role: role
+      content: content
+      thinking: thinking
+    chat_template_kwargs:
+      reasoning_effort: high
+```
+
+See the [Magistral thinking guide](../magistral/think/README.md) for dataset format details.
+
+## Tips
+
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The text dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+- The vision model requires multi-modal dataset format as documented [here](https://docs.axolotl.ai/docs/multimodal.html#dataset-format).
+
+## Related Resources
+
+- [MistralAI Mistral Small 4 Blog](https://mistral.ai/news/mistral-small-4)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/mistral4/fft-text.yml

@@ -0,0 +1,58 @@
+base_model: axolotl-ai-co/Mistral-Small-4-119B-2603-BF16
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+  - axolotl.integrations.kernels.KernelsPlugin
+use_kernels: true
+use_sonicmoe: true
+
+# only train language model layers, freeze vision tower
+unfrozen_parameters:
+  - model.language_model.*
+  - lm_head
+  - embed_tokens
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+sequence_len: 2048
+sample_packing: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: false
+  state_dict_type: FULL_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Mistral4DecoderLayer
+  reshard_after_forward: true
+  activation_checkpointing: true

examples/mistral4/fft-vision.yml

@@ -0,0 +1,57 @@
+base_model: axolotl-ai-co/Mistral-Small-4-119B-2603-BF16
+processor_type: AutoProcessor
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+  - axolotl.integrations.kernels.KernelsPlugin
+use_kernels: true
+use_sonicmoe: true
+
+# vision requirements
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+datasets:
+  - path: Nanobit/text-vision-2k-test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+sequence_len: 2048
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+
+fsdp_version: 2
+fsdp_config:
+  offload_params: false
+  cpu_ram_efficient_loading: false
+  state_dict_type: FULL_STATE_DICT
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Mistral4DecoderLayer
+  reshard_after_forward: true
+  activation_checkpointing: true

examples/mistral4/qlora-text.yml

@@ -0,0 +1,58 @@
+base_model: axolotl-ai-co/Mistral-Small-4-119B-2603-BF16
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_4bit: true
+quantize_moe_experts: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+adapter: qlora
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+# uncomment to train on expert layers
+# lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
+# lora_mlp_kernel: false
+# lora_qkv_kernel: false
+# lora_o_kernel: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: true
+tf32: true
+
+gradient_checkpointing: true
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0

examples/mistral4/qlora-vision.yml

@@ -0,0 +1,63 @@
+base_model: axolotl-ai-co/Mistral-Small-4-119B-2603-BF16
+processor_type: AutoProcessor
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_4bit: true
+quantize_moe_experts: true
+
+# vision chat template requirements
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+datasets:
+  - path: Nanobit/text-vision-2k-test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+adapter: qlora
+
+sequence_len: 2048
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|cross_attn|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+# uncomment to train on expert layers
+# lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
+# lora_mlp_kernel: false
+# lora_qkv_kernel: false
+# lora_o_kernel: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: true
+tf32: true
+
+gradient_checkpointing: true
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0

examples/nemotron-h/120b-a12b-qlora.yaml

@@ -0,0 +1,74 @@
+base_model: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
+
+# LoRA kernel patches are incompatible with this architecture — see README.
+lora_mlp_kernel: false
+lora_qkv_kernel: false
+lora_o_kernel: false
+
+chat_template: tokenizer_default
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+val_set_size: 0.0
+output_dir: ./outputs/out
+dataset_prepared_path: last_run_prepared
+
+sequence_len: 4096
+sample_packing: true
+
+use_cut_cross_entropy: true
+
+load_in_4bit: true
+quantize_moe_experts: true
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.0
+lora_target_modules:
+  # Attention projection layers (present in ~12 attention layers out of 88)
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+  # To also train MoE expert weights, add them via lora_target_parameters
+  # (they are 3D nn.Parameter tensors, not nn.Linear — no gate_proj):
+  #   lora_target_parameters:
+  #     - up_proj
+  #     - down_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 2
+saves_per_epoch: 1
+weight_decay: 0.0
+
+special_tokens:

examples/nemotron-h/README.md

@@ -0,0 +1,48 @@
+# Nemotron-H (nvidia/NVIDIA-Nemotron-3-*)
+
+Hybrid Mamba2 / Attention / MoE architecture (`model_type: nemotron_h`).
+
+| Model | Total params | Active params | Layers |
+|---|---|---|---|
+| NVIDIA-Nemotron-3-Super-120B-A12B-BF16 | 120B | ~12B | 88 |
+| NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 | 30B | ~3B | — |
+
+## Requirements
+
+```bash
+pip install mamba-ssm causal-conv1d   # fast Mamba2 CUDA kernels
+```
+
+## Architecture notes
+
+- Three block types per layer: **Mamba2** (selective SSM), **Attention** (sparse), **MoE** (mixture-of-experts).
+- Only ~12 out of 88 blocks are attention layers (120B variant).
+- MLP activation is `relu2` via `mlp_hidden_act` (not the usual `hidden_act`).
+
+## LoRA kernel patches
+
+All three LoRA Triton kernel patches must be disabled:
+
+```yaml
+lora_qkv_kernel: false   # attention lives in NemotronHBlock.mixer, not layer.self_attn
+lora_o_kernel: false     # same reason
+lora_mlp_kernel: false   # relu2 (mlp_hidden_act) is not supported by lora_mlp_kernel
+```
+
+## MoE expert weights
+
+NemotronH experts store `up_proj` and `down_proj` as 3D `nn.Parameter` tensors
+(shape `[num_experts, out_dim, in_dim]`), **not** `nn.Linear` modules — there is no
+`gate_proj`. To fine-tune them alongside attention, use `lora_target_parameters`
+instead of `lora_target_modules`:
+
+```yaml
+lora_target_parameters:
+  - up_proj
+  - down_proj
+```
+
+## Limitations
+
+- **MoE Triton kernels**: `lora_mlp_kernel` is not supported for NemotronH's MoE expert layers. The expert weights are 3D `nn.Parameter` tensors (not `nn.Linear`), which the Triton kernel does not support. Keep `lora_mlp_kernel: false`.
+- **Gradient checkpointing**: Only supported when `sample_packing: true`. Without sample packing the upstream model marks `supports_gradient_checkpointing = False`.

examples/nemotron-h/nano-30b-a3b-qlora.yaml

@@ -0,0 +1,74 @@
+# See examples/nemotron-h/README.md for architecture notes and requirements.
+base_model: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
+
+# LoRA kernel patches are incompatible with this architecture — see README.
+lora_mlp_kernel: false
+lora_qkv_kernel: false
+lora_o_kernel: false
+
+chat_template: tokenizer_default
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+val_set_size: 0.0
+output_dir: ./outputs/out
+dataset_prepared_path: last_run_prepared
+
+sequence_len: 4096
+sample_packing: true
+
+use_cut_cross_entropy: true
+
+load_in_4bit: true
+quantize_moe_experts: true
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0.0
+lora_target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+  # To also train MoE expert weights, add them via lora_target_parameters
+  # (they are 3D nn.Parameter tensors, not nn.Linear — no gate_proj):
+  #   lora_target_parameters:
+  #     - up_proj
+  #     - down_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+
+special_tokens:

examples/nemotron/nemotron-mini-4b-qlora.yaml

@@ -0,0 +1,57 @@
+base_model: nvidia/Nemotron-Mini-4B-Instruct
+
+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/nemotron-mini-4b-qlora
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 4096
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+  - up_proj
+  - down_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
+
+special_tokens:

examples/qwen3.5/122b-a10b-moe-qlora-fsdp.yaml

@@ -0,0 +1,85 @@
+base_model: Qwen/Qwen3.5-122B-A10B
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+strict: false
+
+chat_template: qwen3_5
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+val_set_size: 0.0
+output_dir: ./outputs/out
+dataset_prepared_path: last_run_prepared
+
+sequence_len: 2048
+sample_packing: true
+
+load_in_4bit: true
+quantize_moe_experts: true
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0
+lora_target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+  # Add gate_up_proj and down_proj to also target shared experts (nn.Linear):
+  # - gate_up_proj
+  # - down_proj
+
+# Target routed experts (3D nn.Parameter tensors, not nn.Linear — use lora_target_parameters):
+# lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+lora_mlp_kernel: false
+lora_qkv_kernel: false
+lora_o_kernel: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+
+fsdp_config:
+  fsdp_version: 2
+  offload_params: true
+  cpu_ram_efficient_loading: false
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen3_5MoeDecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true

examples/qwen3.5/122b-a10b-moe-qlora.yaml

@@ -31,8 +31,12 @@ lora_target_modules:
   - k_proj
   - v_proj
   - o_proj
+  # Add gate_up_proj and down_proj to also target shared experts (nn.Linear):
+  # - gate_up_proj
+  # - down_proj
 
-#lora_target_parameters:
+# Target routed experts (3D nn.Parameter tensors, not nn.Linear — use lora_target_parameters):
+# lora_target_parameters:
 #   - mlp.experts.gate_up_proj
 #   - mlp.experts.down_proj
 
@@ -52,7 +56,6 @@ learning_rate: 0.0002
 bf16: auto
 tf32: true
 
-
 lora_mlp_kernel: false
 lora_qkv_kernel: false
 lora_o_kernel: false

examples/qwen3.5/27b-fft.yaml

@@ -0,0 +1,59 @@
+base_model: Qwen/Qwen3.5-27B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# Full fine-tune (FFT) of the text-only path of Qwen3.5-27B.
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+strict: false
+
+chat_template: qwen3_5
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+val_set_size: 0.0
+output_dir: ./outputs/out
+dataset_prepared_path: last_run_prepared
+
+sequence_len: 2048
+sample_packing: true
+
+# Freeze vision encoder
+unfrozen_parameters:
+  - model\.language_model\..*
+  - lm_head\..*
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/qwen3.5/27b-qlora-fsdp.yaml

@@ -0,0 +1,81 @@
+base_model: Qwen/Qwen3.5-27B
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+strict: false
+
+chat_template: qwen3_5
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+val_set_size: 0.0
+output_dir: ./outputs/out
+dataset_prepared_path: last_run_prepared
+
+sequence_len: 2048
+sample_packing: true
+
+load_in_4bit: true
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+lora_target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+  - down_proj
+  - up_proj
+  # Uncomment below to also target the linear attention projections.
+  # These use separate in_proj_qkv / in_proj_z / out_proj (Qwen3.5-specific).
+  # - linear_attn.in_proj_qkv
+  # - linear_attn.in_proj_z
+  # - linear_attn.out_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+
+fsdp_config:
+  fsdp_version: 2
+  offload_params: false
+  cpu_ram_efficient_loading: false
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen3_5DecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true

examples/qwen3.5/27b-qlora.yaml

@@ -1,9 +1,7 @@
 base_model: Qwen/Qwen3.5-27B
+
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-# Note: Qwen3.5 is an early-fusion VLM (image+text). This config fine-tunes
-# the text-only path. For multimodal (image+text) fine-tuning, add image
-# columns to your dataset following axolotl's multimodal dataset format.
 
 plugins:
   - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

examples/qwen3.5/35b-a3b-moe-qlora-fsdp.yaml

@@ -0,0 +1,85 @@
+base_model: Qwen/Qwen3.5-35B-A3B
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+strict: false
+
+chat_template: qwen3_5
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+val_set_size: 0.0
+output_dir: ./outputs/out
+dataset_prepared_path: last_run_prepared
+
+sequence_len: 2048
+sample_packing: true
+
+load_in_4bit: true
+quantize_moe_experts: true
+adapter: qlora
+lora_r: 16
+lora_alpha: 32
+lora_dropout: 0
+lora_target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+  # Add gate_up_proj and down_proj to also target shared experts (nn.Linear):
+  # - gate_up_proj
+  # - down_proj
+
+# Target routed experts (3D nn.Parameter tensors, not nn.Linear — use lora_target_parameters):
+# lora_target_parameters:
+#   - mlp.experts.gate_up_proj
+#   - mlp.experts.down_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+lora_mlp_kernel: false
+lora_qkv_kernel: false
+lora_o_kernel: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 4
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+
+fsdp_config:
+  fsdp_version: 2
+  offload_params: true
+  cpu_ram_efficient_loading: false
+  auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  transformer_layer_cls_to_wrap: Qwen3_5MoeDecoderLayer
+  state_dict_type: FULL_STATE_DICT
+  sharding_strategy: FULL_SHARD
+  reshard_after_forward: true
+  activation_checkpointing: true

examples/qwen3.5/35b-a3b-moe-qlora.yaml

@@ -1,8 +1,18 @@
-base_model: Qwen/Qwen3.5-35B-A3B
+base_model: Qwen/Qwen3.5-35B-A3B-Base
 
 plugins:
   - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
-strict: false
+  - axolotl.integrations.kernels.KernelsPlugin
+  - axolotl.integrations.liger.LigerPlugin
+use_kernels: true
+use_scattermoe: true
+liger_layer_norm: true
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+liger_rms_norm_gated: true
+
+torch_compile: false
 
 chat_template: qwen3_5
 datasets:
@@ -13,6 +23,7 @@ datasets:
     message_property_mappings:
       role: from
       content: value
+
 val_set_size: 0.0
 output_dir: ./outputs/out
 dataset_prepared_path: last_run_prepared
@@ -31,11 +42,19 @@ lora_target_modules:
   - k_proj
   - v_proj
   - o_proj
+  # Add gate_up_proj and down_proj to also target shared experts (nn.Linear):
+  # - gate_up_proj
+  # - down_proj
 
-#lora_target_parameters:
+# Target routed experts (3D nn.Parameter tensors, not nn.Linear — use lora_target_parameters):
+# lora_target_parameters:
 #   - mlp.experts.gate_up_proj
 #   - mlp.experts.down_proj
 
+lora_qkv_kernel: true
+lora_o_kernel: true
+lora_mlp_kernel: false
+
 wandb_project:
 wandb_entity:
 wandb_watch:
@@ -43,22 +62,17 @@ wandb_name:
 wandb_log_model:
 
 gradient_accumulation_steps: 2
-micro_batch_size: 1
+micro_batch_size: 4
 num_epochs: 1
-optimizer: adamw_torch_4bit
+optimizer: adamw_torch_8bit
 lr_scheduler: cosine
 learning_rate: 0.0002
 
 bf16: auto
 tf32: true
 
-lora_mlp_kernel: false
-lora_qkv_kernel: false
-lora_o_kernel: false
-
 gradient_checkpointing: true
-gradient_checkpointing_kwargs:
-  use_reentrant: false
+activation_offloading: true
 resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true

examples/qwen3.5/9b-fft-vision.yaml

@@ -0,0 +1,49 @@
+base_model: Qwen/Qwen3.5-9B
+processor_type: AutoProcessor
+
+# Required for multimodal training
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+chat_template: qwen3_5
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 4096
+pad_to_sequence_len: false
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/qwen3.5/9b-lora-vision.yaml

@@ -1,10 +1,6 @@
-base_model: Qwen/Qwen3.5-7B
+base_model: Qwen/Qwen3.5-9B
 processor_type: AutoProcessor
 
-# Qwen3.5-7B and above are early-fusion VLMs (Qwen3_5ForConditionalGeneration).
-# Vision and text tokens are processed together by the same transformer layers.
-# Note: Qwen3.5-2B is a text-only model — the smallest VLM is Qwen3.5-7B.
-
 # These 3 lines are required for vision/multimodal training
 skip_prepare_dataset: true
 remove_unused_columns: false
@@ -30,8 +26,6 @@ lora_r: 32
 lora_alpha: 16
 lora_dropout: 0.05
 # Targets the language model attention and MLP layers.
-# Qwen3.5 is early-fusion: all layers (including those seeing vision tokens) share
-# the same transformer stack, so standard attention targets work for both modalities.
 lora_target_modules:
   - q_proj
   - k_proj

examples/qwen3.5/README.md

@@ -1,15 +1,6 @@
 # Finetune Qwen3.5 with Axolotl
 
-[Qwen3.5](https://huggingface.co/collections/Qwen/qwen35-68452f3bc6e4b7cfb4e1c803) is a hybrid architecture model series combining Gated DeltaNet linear attention with standard Transformer attention. Models from 7B onwards are early-fusion vision-language models (`Qwen3_5ForConditionalGeneration`), meaning vision and text tokens are processed through the same transformer stack. The 2B variant is text-only.
-
-Available configs:
-
-| Config | Model | Type |
-|---|---|---|
-| `27b-qlora.yaml` | Qwen3.5-27B | Dense VLM, text-only path |
-| `35b-a3b-moe-qlora.yaml` | Qwen3.5-35B-A3B | MoE, text-only path |
-| `122b-a10b-moe-qlora.yaml` | Qwen3.5-122B-A10B | MoE, text-only path |
-| `7b-lora-vision.yaml` | Qwen3.5-7B | Vision+text (multimodal) |
+[Qwen3.5](https://huggingface.co/collections/Qwen/qwen35) is a hybrid architecture model series combining Gated DeltaNet linear attention with standard Transformer attention. All Qwen3.5 models are early-fusion vision-language models: dense variants use `Qwen3_5ForConditionalGeneration` and MoE variants use `Qwen3_5MoeForConditionalGeneration`.
 
 ## Getting started
 
@@ -18,35 +9,78 @@ Available configs:
 2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
 
 3. Install FLA for sample packing support with the Gated DeltaNet linear attention layers:
-```bash
-pip3 uninstall -y causal-conv1d && pip3 install flash-linear-attention==0.4.1
+  ```bash
+  pip3 uninstall -y causal-conv1d && pip3 install flash-linear-attention==0.4.1
+  ```
+  > FLA is required when `sample_packing: true`. Without it, training raises a `RuntimeError` on packed sequences. Vision configs use `sample_packing: false` so FLA is optional there.
+
+4. Pick any config from the table below and run:
+
+    ```bash
+    axolotl train examples/qwen3.5/<config>.yaml
+    ```
+
+Available configs:
+
+| Config | Model | Type | Peak VRAM |
+|---|---|---|---|
+| `9b-lora-vision.yaml` | Qwen3.5-9B | Vision+text LoRA, single GPU | — |
+| `9b-fft-vision.yaml` | Qwen3.5-9B | Vision+text FFT, single GPU | ~61 GiB |
+| `27b-qlora.yaml` | Qwen3.5-27B | Dense, text-only QLoRA | ~47 GiB |
+| `27b-fft.yaml` | Qwen3.5-27B | Dense, text-only FFT (vision frozen) | ~53 GiB |
+| `27b-qlora-fsdp.yaml` | Qwen3.5-27B | Dense, text-only QLoRA + FSDP2 | — |
+| `35b-a3b-moe-qlora.yaml` | Qwen3.5-35B-A3B | MoE, text-only QLoRA | — |
+| `35b-a3b-moe-qlora-fsdp.yaml` | Qwen3.5-35B-A3B | MoE, text-only QLoRA + FSDP2 | — |
+| `122b-a10b-moe-qlora.yaml` | Qwen3.5-122B-A10B | MoE, text-only QLoRA | — |
+| `122b-a10b-moe-qlora-fsdp.yaml` | Qwen3.5-122B-A10B | MoE, text-only QLoRA + FSDP2 | — |
+
+### Gated DeltaNet Linear Attention
+
+Qwen3.5 interleaves standard attention with Gated DeltaNet linear attention layers. To apply LoRA to them, add to `lora_target_modules`:
+
+```yaml
+lora_target_modules:
+  # ... standard projections ...
+  - linear_attn.in_proj_qkv
+  - linear_attn.in_proj_z
+  - linear_attn.out_proj
 ```
-> FLA is required when `sample_packing: true`. Without it, training raises a `RuntimeError` on packed sequences. Vision configs use `sample_packing: false` so FLA is optional there.
 
-4. Run a finetuning example:
+### Routed Experts (MoE)
 
-```bash
-# Dense 27B text-only (QLoRA, ~47 GiB VRAM with sample packing)
-axolotl train examples/qwen3.5/27b-qlora.yaml
+To apply LoRA to routed expert parameters, add `lora_target_parameters`:
 
-# MoE 35B-A3B text-only (QLoRA)
-axolotl train examples/qwen3.5/35b-a3b-moe-qlora.yaml
-
-# MoE 122B-A10B text-only (QLoRA)
-axolotl train examples/qwen3.5/122b-a10b-moe-qlora.yaml
-
-# 7B vision+text (LoRA, multimodal dataset)
-axolotl train examples/qwen3.5/7b-lora-vision.yaml
+```yaml
+lora_target_parameters:
+  - mlp.experts.gate_up_proj
+  - mlp.experts.down_proj
+#  - mlp.gate.weight  # router
 ```
 
+### Shared Experts (MoE)
+
+Shared experts use `nn.Linear` (unlike routed experts which are 3D `nn.Parameter` tensors), so they can be targeted via `lora_target_modules`. To also train shared expert projections alongside attention, uncomment `gate_up_proj` and `down_proj` in `lora_target_modules`:
+
+```yaml
+lora_target_modules:
+  - q_proj
+  - k_proj
+  - v_proj
+  - o_proj
+  # Add gate_up_proj and down_proj to also target shared experts (nn.Linear):
+  # - gate_up_proj
+  # - down_proj
+```
+
+Use `lora_target_parameters` (see [Routed Experts](#routed-experts-moe) above) to target routed experts separately.
+
 ### TIPS
 
-- For inference, you can experiment with `temperature: 0.7`, `top_p: 0.8`, `top_k: 20`, and `min_p: 0`.
-- You can run a full finetuning by removing `adapter: qlora` and `load_in_4bit: true`. See [Multi-GPU](#optimization-guides) below.
+- For inference hyp, please see the respective model card details.
+- You can run a full finetuning of smaller configs by removing `adapter: qlora` and `load_in_4bit: true`. See [Multi-GPU](#optimization-guides) below.
 - Read more on loading 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).
-- For **multimodal** finetuning, set `processor_type: AutoProcessor`, `skip_prepare_dataset: true`, and `remove_unused_columns: false` as shown in `7b-lora-vision.yaml`.
-- The Gated DeltaNet linear attention layers (`linear_attn.*`) can optionally be added to `lora_target_modules` — they are commented out by default.
+- For **multimodal** finetuning, set `processor_type: AutoProcessor`, `skip_prepare_dataset: true`, and `remove_unused_columns: false` as shown in `9b-lora-vision.yaml`.
 
 ## Optimization Guides
 

pyproject.toml

@@ -61,5 +61,11 @@ skip-magic-trailing-comma = false
 line-ending = "auto"
 docstring-code-format = false
 
+[tool.pytest.ini_options]
+addopts = "-m 'not slow'"
+markers = [
+    "slow: marks tests as slow",
+]
+
 [tool.uv.extra-build-dependencies]
 axolotl = ["huggingface_hub"]

requirements.txt

@@ -12,7 +12,7 @@ packaging==26.0
 huggingface_hub>=1.1.7
 peft>=0.18.1
 tokenizers>=0.22.1
-transformers==5.3.0
+transformers==5.5.0
 accelerate==1.13.0
 datasets==4.5.0
 deepspeed>=0.18.6,<0.19.0
@@ -61,12 +61,12 @@ zstandard==0.22.0
 fastcore
 
 # lm eval harness
-lm_eval==0.4.7
+lm_eval==0.4.11
 langdetect==1.0.9
 immutabledict==4.2.0
 antlr4-python3-runtime==4.13.2
 
-torchao==0.16.0
+torchao==0.17.0
 openenv-core==0.1.0
 schedulefree==1.4.1
 
@@ -75,4 +75,4 @@ axolotl-contribs-mit==0.0.6
 # telemetry
 posthog==6.7.11
 
-mistral-common==1.8.8
+mistral-common==1.11.0

scripts/cutcrossentropy_install.py

@@ -29,5 +29,5 @@ UV_PREFIX = "uv " if USE_UV else ""
 
 print(
     UNINSTALL_PREFIX
-    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@e8ad129"'
+    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@63b15e6"'
 )

setup.py

@@ -81,16 +81,23 @@ def parse_requirements(extras_require_map):
                     f"https://download.pytorch.org/whl/{torch_cuda_version}"
                 )
 
-            if (major, minor) >= (2, 9):
+            if (major, minor) >= (2, 10):
+                extras_require_map.pop("fbgemm-gpu")
+                extras_require_map["fbgemm-gpu"] = [
+                    "fbgemm-gpu==1.5.0",
+                    "fbgemm-gpu-genai==1.5.0",
+                ]
+                if not install_xformers:
+                    _install_requires.pop(_install_requires.index(xformers_version))
+                extras_require_map["vllm"] = ["vllm>=0.17.1"]
+            elif (major, minor) >= (2, 9):
                 extras_require_map.pop("fbgemm-gpu")
                 extras_require_map["fbgemm-gpu"] = [
                     "fbgemm-gpu==1.4.0",
                     "fbgemm-gpu-genai==1.4.2",
                 ]
-                extras_require_map["vllm"] = ["vllm==0.11.1"]
                 if not install_xformers:
                     _install_requires.pop(_install_requires.index(xformers_version))
-                extras_require_map["vllm"] = ["vllm==0.13.0"]
                 if patch == 0:
                     extras_require_map["vllm"] = ["vllm==0.13.0"]
                 else:

src/axolotl/cli/checks.py

@@ -3,6 +3,8 @@
 import os
 from pathlib import Path
 
+import httpcore
+import httpx
 from accelerate.commands.config import config_args
 from huggingface_hub import HfApi
 from huggingface_hub.utils import LocalTokenNotFoundError
@@ -47,7 +49,7 @@ def check_user_token() -> bool:
             "Error verifying HuggingFace token. Remember to log in using `hf auth login` and get your access token from https://huggingface.co/settings/tokens if you want to use gated models or datasets."
         )
         return False
-    except HTTPError:
+    except (HTTPError, httpcore.ConnectError, httpx.ConnectError):
         LOG.warning(
             "Error accessing HuggingFace. This may be due to a network issue or rate limiting."
         )

src/axolotl/cli/cloud/modal_.py

@@ -90,9 +90,8 @@ class ModalCloud(Cloud):
         # grab the sha256 hash from docker hub for this image+tag
         # this ensures that we always get the latest image for this tag, even if it's already cached
         try:
-            manifest = subprocess.check_output(  # nosec B602
-                f"docker manifest inspect {docker_image}",
-                shell=True,
+            manifest = subprocess.check_output(  # nosec
+                ["docker", "manifest", "inspect", docker_image],
             ).decode("utf-8")
             sha256_hash = json.loads(manifest)["manifests"][0]["digest"]
         except subprocess.CalledProcessError:

src/axolotl/cli/config.py

@@ -11,7 +11,7 @@ from urllib.parse import urlparse
 import requests
 import torch
 import yaml
-from transformers.utils import is_torch_bf16_gpu_available
+from transformers.utils import is_torch_bf16_gpu_available, is_torch_tf32_available
 
 from axolotl.integrations.base import PluginManager
 from axolotl.telemetry.errors import send_errors
@@ -300,7 +300,7 @@ def load_cfg(
     try:
         device_props = torch.cuda.get_device_properties("cuda")
         gpu_version = "sm_" + str(device_props.major) + str(device_props.minor)
-    except:
+    except (RuntimeError, AssertionError):
         gpu_version = None
 
     prepare_plugins(cfg)
@@ -310,6 +310,7 @@ def load_cfg(
         capabilities={
             "bf16": is_torch_bf16_gpu_available(),
             "fp8": compute_supports_fp8(),
+            "tf32": is_torch_tf32_available(),
             "n_gpu": int(os.environ.get("WORLD_SIZE", 1)),
             "compute_capability": gpu_version,
         },

src/axolotl/cli/merge_lora.py

@@ -4,9 +4,11 @@ from pathlib import Path
 from typing import Union
 
 import fire
+import torch
 
 from axolotl.cli.config import load_cfg
 from axolotl.cli.utils import load_model_and_tokenizer
+from axolotl.cli.utils.lora_merge import merge_lora_sharded_efficient
 from axolotl.telemetry.errors import send_errors
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.logging import get_logger
@@ -17,16 +19,30 @@ LOG = get_logger(__name__)
 @send_errors
 def do_merge_lora(*, cfg: DictDefault) -> None:
     """
-    Calls `transformers`' `merge_and_unload` on the model given in the `axolotl` config
-    along with the LoRA adapters to combine them into a single base model.
+    Merges LoRA adapters with base model using either memory-efficient or legacy approach.
 
     Args:
         cfg: Dictionary mapping `axolotl` config keys to values.
     """
+    merge_method = str(getattr(cfg, "merge_method", "memory_efficient"))
+    if merge_method == "legacy":
+        LOG.debug("Using legacy LoRA merging method...")
+        _do_merge_lora_legacy(cfg=cfg)
+    else:
+        LOG.debug("Using memory-efficient LoRA merging method...")
+        _do_merge_lora_efficient(cfg=cfg)
+
+
+def _do_merge_lora_legacy(*, cfg: DictDefault) -> None:
+    """
+    Legacy LoRA merging using merge_and_unload.
+    Loads the full model into memory before merging.
+    """
+    LOG.debug("Using legacy LoRA merging method...")
     model, tokenizer, processor = load_model_and_tokenizer(cfg=cfg)
 
     LOG.info("Running merge of LoRA with base model...")
-    model = model.merge_and_unload(progressbar=True, safe_merge=True)
+    model = model.merge_and_unload(progressbar=True)
     try:
         model.to(dtype=cfg.torch_dtype)
     except ValueError as e:
@@ -52,6 +68,58 @@ def do_merge_lora(*, cfg: DictDefault) -> None:
             processor.save_pretrained(str(Path(cfg.output_dir) / "merged"))
 
 
+def _do_merge_lora_efficient(*, cfg: DictDefault) -> None:
+    """
+    Memory-efficient LoRA merging using shard-by-shard processing.
+    Does not load the full model into memory.
+
+    Supports standard LoRA, RSLoRA, and DoRA. Unsupported methods (AdaLoRA, VeRA)
+    will raise NotImplementedError — use legacy method for those.
+    """
+    LOG.debug("Using memory-efficient LoRA merging method...")
+
+    output_path = Path(cfg.output_dir) / "merged"
+    safe_tensors = getattr(cfg, "save_safetensors", True)
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+
+    # Detect NF4 quantization from config to simulate QLoRA training dynamics.
+    # Check both current and original (pre-override) config values since do_cli
+    # forces load_in_4bit=False for the legacy path.
+    simulate_nf4 = bool(
+        getattr(cfg, "load_in_4bit", False)
+        or getattr(cfg, "_original_load_in_4bit", False)
+        or getattr(cfg, "adapter", None) == "qlora"
+        or getattr(cfg, "_original_adapter", None) == "qlora"
+    )
+
+    bnb_config_kwargs = getattr(cfg, "bnb_config_kwargs", None) or {}
+    nf4_blocksize = bnb_config_kwargs.get("blocksize", None)
+    nf4_double_quant = bnb_config_kwargs.get(
+        "bnb_4bit_use_double_quant",
+        getattr(cfg, "bnb_4bit_use_double_quant", True),
+    )
+
+    # Detect MoE expert quantization
+    simulate_nf4_experts = bool(
+        getattr(cfg, "quantize_moe_experts", False)
+        or getattr(cfg, "_original_quantize_moe_experts", False)
+    )
+
+    merge_lora_sharded_efficient(
+        base_model_path=cfg.base_model,
+        lora_adapter_path=cfg.lora_model_dir,
+        output_path=output_path,
+        safe_tensors=safe_tensors,
+        device=device,
+        simulate_nf4=simulate_nf4,
+        simulate_nf4_experts=simulate_nf4_experts,
+        nf4_blocksize=nf4_blocksize,
+        nf4_double_quant=nf4_double_quant,
+    )
+
+    LOG.debug("Memory-efficient LoRA merge completed successfully!")
+
+
 def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs) -> None:
     """
     Parses `axolotl` config, CLI args, and calls `do_merge_lora`. Note that various
@@ -66,6 +134,12 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs) -> None:
         ValueError: If target directory for LoRA merged model does not exist.
     """
 
+    # Pre-load config to detect original quantization settings before overrides
+    raw_cfg = load_cfg(config, **kwargs)
+    original_load_in_4bit = getattr(raw_cfg, "load_in_4bit", False)
+    original_adapter = getattr(raw_cfg, "adapter", None)
+    original_quantize_moe_experts = getattr(raw_cfg, "quantize_moe_experts", False)
+
     parsed_cfg = load_cfg(
         config,
         merge_lora=True,
@@ -80,11 +154,16 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs) -> None:
         **kwargs,
     )
 
+    # Stash original quantization settings for NF4 simulation in efficient merge
+    parsed_cfg._original_load_in_4bit = original_load_in_4bit
+    parsed_cfg._original_adapter = original_adapter
+    parsed_cfg._original_quantize_moe_experts = original_quantize_moe_experts
+
     if not parsed_cfg.lora_model_dir and parsed_cfg.output_dir:
         parsed_cfg.lora_model_dir = parsed_cfg.output_dir
     if not Path(parsed_cfg.lora_model_dir).exists():
         raise ValueError(
-            f"Target directory for merge: `{parsed_cfg.lora_model_dir}` does not exist."
+            f"Target directory for LoRA adapter weights does not exist: `{parsed_cfg.lora_model_dir}`"
         )
 
     do_merge_lora(cfg=parsed_cfg)

src/axolotl/cli/merge_sharded_fsdp_weights.py

@@ -196,12 +196,10 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
     state.wait_for_everyone()
     LOG.info(
         f"FSDP SHARDED_STATE_DICT weights successfully merged to: {output_path}",
-        main_process_only=True,
     )
     LOG.info(
         "Merged weights are only the safetensors and doesn't include the model configuration "
         f"or tokenizer which may be found in {parsed_cfg.output_dir}.",
-        main_process_only=True,
     )
 
 

src/axolotl/cli/quantize.py

@@ -5,7 +5,7 @@ CLI to post-training quantize a model using torchao
 from pathlib import Path
 from typing import Union
 
-from transformers import AutoConfig, AutoModelForCausalLM, TorchAoConfig
+from transformers import AutoConfig, AutoModelForCausalLM
 
 from axolotl.cli.config import load_cfg
 from axolotl.loaders import load_processor, load_tokenizer
@@ -93,17 +93,22 @@ def do_quantize(
         weight_dtype, activation_dtype, group_size
     )
 
-    ao_config = TorchAoConfig(
-        quant_type=quantization_config,
-        include_input_output_embeddings=quantize_embedding,
-    )
-    model.config.quantization_config = ao_config
-
     LOG.info(f"Saving quantized model to: {str(Path(output_dir) / 'quantized')}.")
-    model.save_pretrained(
-        str(Path(output_dir) / "quantized"),
-        progressbar=True,
-    )
+    try:
+        model.save_pretrained(
+            str(Path(output_dir) / "quantized"),
+            progressbar=True,
+        )
+    except NotImplementedError:
+        LOG.warning(
+            "Model weight conversions do not support reverse_op, "
+            "retrying save with save_original_format=False"
+        )
+        model.save_pretrained(
+            str(Path(output_dir) / "quantized"),
+            progressbar=True,
+            save_original_format=False,
+        )
     tokenizer.save_pretrained(
         str(Path(output_dir) / "quantized"),
         progressbar=True,

src/axolotl/cli/train.py

@@ -84,8 +84,11 @@ def do_cli(config: Union[Path, str] = Path("examples/"), **kwargs):
                 storage_path=Path(parsed_cfg.output_dir).absolute().as_posix(),
             ),
         )
-        return trainer.fit()
-    return do_train(parsed_cfg, parsed_cli_args)
+
+        trainer.fit()
+        return
+
+    do_train(parsed_cfg, parsed_cli_args)
 
 
 def ray_train_func(kwargs: dict):

src/axolotl/cli/utils/lora_merge.py

@@ -0,0 +1,982 @@
+import gc
+import math
+import os
+import shutil
+from pathlib import Path
+from typing import Dict, Optional, Union
+
+import safetensors
+import safetensors.torch
+import torch
+from huggingface_hub import snapshot_download
+from peft import LoraConfig
+from tqdm import tqdm
+
+from axolotl.utils.logging import get_logger
+
+LOG = get_logger(__name__)
+
+
+def _simulate_nf4_roundtrip(
+    tensor: torch.Tensor,
+    blocksize: Optional[int] = None,
+    compress_statistics: bool = True,
+    device: Optional[Union[str, torch.device]] = None,
+) -> torch.Tensor:
+    """
+    Simulate NF4 quantization roundtrip to match QLoRA training dynamics.
+
+    During QLoRA training, base weights are quantized to NF4 and dequantized on-the-fly
+    for each forward pass. The LoRA adapters learn to compensate for the quantization
+    noise in the dequantized weights. To match this at merge time, we apply the same
+    quantize → dequantize roundtrip so the merged result reflects what the model saw
+    during training.
+
+    Args:
+        tensor: Base model weight tensor (fp16/bf16/fp32)
+        blocksize: NF4 quantization block size (default: bitsandbytes default)
+        compress_statistics: Whether to use double quantization
+        device: Device for quantization computation.  bitsandbytes requires a
+            CUDA device; defaults to "cuda" when available.
+
+    Returns:
+        Tensor after NF4 quantize → dequantize roundtrip, in original dtype
+    """
+    import bitsandbytes.functional as bnb_F
+
+    quant_device: torch.device
+    if device is None:
+        quant_device = torch.device("cuda")
+    elif isinstance(device, str):
+        quant_device = torch.device(device)
+    else:
+        quant_device = device
+
+    if quant_device.type == "cuda" and not torch.cuda.is_available():
+        raise RuntimeError(
+            "NF4 simulation requires CUDA but no GPU is available. "
+            "Either run on a machine with a GPU or disable NF4 simulation."
+        )
+
+    original_dtype = tensor.dtype
+    original_shape = tensor.shape
+
+    # bitsandbytes requires float32 input for quantization and contiguous+CUDA tensor
+    flat = tensor.reshape(-1).to(torch.float32).contiguous().to(quant_device)
+
+    quant_kwargs = {
+        "quant_type": "nf4",
+        "compress_statistics": compress_statistics,
+    }
+    if blocksize is not None:
+        quant_kwargs["blocksize"] = blocksize
+
+    quantized, quant_state = bnb_F.quantize_4bit(flat, **quant_kwargs)
+    dequantized = bnb_F.dequantize_4bit(quantized, quant_state, quant_type="nf4")
+
+    return dequantized.reshape(original_shape).to(original_dtype).cpu()
+
+
+def find_lora_weights(
+    lora_state: Dict[str, torch.Tensor],
+    key: str,
+    weight_renamings: Optional[Dict[str, str]] = None,
+) -> tuple[Optional[torch.Tensor], Optional[torch.Tensor]]:
+    """
+    Find corresponding LoRA A and B weights for a given key.
+
+    Also tries keys after applying weight renamings (from transformers v5
+    conversion mappings) in case the checkpoint key names differ from the
+    runtime model key names used by the LoRA adapter.
+    """
+    import re
+
+    clean_key = key[:-7] if key.endswith(".weight") else key
+
+    # Try the direct key first
+    a_key = f"base_model.model.{clean_key}.lora_A.weight"
+    b_key = f"base_model.model.{clean_key}.lora_B.weight"
+
+    lora_a = lora_state.get(a_key)
+    lora_b = lora_state.get(b_key)
+
+    if lora_a is not None and lora_b is not None:
+        return lora_a, lora_b
+
+    # Try renamed keys (checkpoint format → runtime format)
+    if weight_renamings:
+        for src_pattern, tgt_pattern in weight_renamings.items():
+            renamed_key = re.sub(src_pattern, tgt_pattern, clean_key)
+            if renamed_key != clean_key:
+                a_key = f"base_model.model.{renamed_key}.lora_A.weight"
+                b_key = f"base_model.model.{renamed_key}.lora_B.weight"
+                lora_a = lora_state.get(a_key)
+                lora_b = lora_state.get(b_key)
+                if lora_a is not None and lora_b is not None:
+                    return lora_a, lora_b
+
+    return None, None
+
+
+def _find_param_wrapper_lora(
+    lora_state: Dict[str, torch.Tensor],
+    key: str,
+    tensor_shape: Optional[tuple] = None,
+) -> tuple[Optional[torch.Tensor], Optional[torch.Tensor], Optional[str]]:
+    """
+    Find LoRA weights from a ParamWrapper (lora_target_parameters) that targets
+    a parent module containing this weight as a sub-parameter.
+
+    For example, base weight key 'model.layers.0.mlp.experts.down_proj' may have
+    LoRA at 'base_model.model.model.layers.0.mlp.experts.lora_A.weight' (targeting
+    the 'experts' module with 'down_proj' as the parameter_name).
+
+    When tensor_shape is provided, validates that the LoRA dimensions match the
+    target tensor (important when multiple ParamWrappers are nested and each
+    nesting level has different LoRA dimensions).
+
+    Returns (lora_A, lora_B, parameter_name) or (None, None, None).
+    """
+    clean_key = key[:-7] if key.endswith(".weight") else key
+    # Strip trailing parameter name to get the parent module path
+    # e.g., "model.layers.0.mlp.experts.down_proj" → parent="model.layers.0.mlp.experts", param="down_proj"
+    parts = clean_key.rsplit(".", 1)
+    if len(parts) != 2:
+        return None, None, None
+
+    parent_key, param_name = parts
+
+    # PEFT's ParamWrapper nesting: when multiple parameters are targeted on
+    # the same module, it nests wrappers. The outer wrapper's LoRA is at
+    # parent.lora_A/B and inner wrappers use parent.base_layer.lora_A/B,
+    # parent.base_layer.base_layer.lora_A/B, etc.
+    prefixes_to_try = [
+        f"base_model.model.{parent_key}",
+    ]
+    # Walk up .base_layer nesting levels (typically 1-2 deep)
+    for depth in range(1, 4):
+        bl = ".base_layer" * depth
+        prefixes_to_try.append(f"base_model.model.{parent_key}{bl}")
+
+    for prefix in prefixes_to_try:
+        a_key = f"{prefix}.lora_A.weight"
+        b_key = f"{prefix}.lora_B.weight"
+        lora_a = lora_state.get(a_key)
+        lora_b = lora_state.get(b_key)
+        if lora_a is None or lora_b is None:
+            continue
+
+        # When tensor_shape is given, verify dimensions match before returning.
+        # This prevents returning a mismatched LoRA from a different nesting level.
+        if tensor_shape is not None and len(tensor_shape) >= 3:
+            num_experts = tensor_shape[0]
+            if not (
+                lora_a.shape[0] == lora_b.shape[1]
+                and lora_a.shape[0] % num_experts == 0
+                and lora_a.shape[1] == tensor_shape[1]
+                and lora_b.shape[0] == tensor_shape[2]
+            ):
+                continue  # Dimensions don't match, try next nesting level
+
+        return lora_a, lora_b, param_name
+
+    return None, None, None
+
+
+def _build_peft_layer_and_get_delta(
+    lora_a: torch.Tensor,
+    lora_b: torch.Tensor,
+    lora_config_dict: Dict,
+    base_tensor: torch.Tensor,
+    adapter_name: str = "default",
+    is_param_wrapper: bool = False,
+    magnitude: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """
+    Use PEFT's own layer classes to compute the LoRA delta weight.
+
+    Instead of re-implementing the merge math for every LoRA variant, this
+    constructs a lightweight PEFT layer, loads the A/B weights, and calls
+    ``get_delta_weight`` (or ``merge`` for DoRA) which handles standard LoRA,
+    RSLoRA, DoRA, and ParamWrapper (expert-blocked) LoRA.
+
+    Returns the delta tensor (same shape as base_tensor).
+    """
+    import warnings
+
+    import torch.nn as nn
+
+    r_total = lora_a.shape[0]
+    in_features = lora_a.shape[1]
+    out_features = lora_b.shape[0]
+    lora_alpha = lora_config_dict.get("lora_alpha", lora_config_dict.get("r", 1))
+    use_rslora = bool(lora_config_dict.get("use_rslora", False))
+    use_dora = bool(lora_config_dict.get("use_dora", False)) and magnitude is not None
+
+    if is_param_wrapper:
+        from peft.tuners.lora.layer import ParamWrapper
+
+        num_experts = base_tensor.shape[0]
+        r = r_total // num_experts
+
+        class _FakeModule(nn.Module):
+            pass
+
+        fake = _FakeModule()
+        fake.register_parameter(
+            "weight", nn.Parameter(base_tensor.clone(), requires_grad=False)
+        )
+
+        with warnings.catch_warnings():
+            warnings.simplefilter("ignore", UserWarning)
+            layer = ParamWrapper(
+                fake,
+                adapter_name=adapter_name,
+                parameter_name="weight",
+                r=r,
+                lora_alpha=lora_alpha,
+                use_rslora=use_rslora,
+            )
+        layer.lora_A[adapter_name].weight.data = lora_a
+        layer.lora_B[adapter_name].weight.data = lora_b
+        return layer.get_delta_weight(adapter_name)
+    else:
+        from peft.tuners.lora.layer import Linear as LoraLinear
+
+        base_layer = nn.Linear(in_features, out_features, bias=False)
+        base_layer.weight.data = base_tensor.clone()
+
+        fan_in_fan_out = bool(
+            lora_config_dict.get("fan_in_fan_out", False)
+            or lora_config_dict.get("lora_fan_in_fan_out", False)
+        )
+
+        layer = LoraLinear(
+            base_layer,
+            adapter_name=adapter_name,
+            r=r_total,
+            lora_alpha=lora_alpha,
+            fan_in_fan_out=fan_in_fan_out,
+            use_rslora=use_rslora,
+            use_dora=use_dora,
+        )
+        layer.lora_A[adapter_name].weight.data = lora_a
+        layer.lora_B[adapter_name].weight.data = lora_b
+
+        if use_dora:
+            # DoRA merges magnitude normalization into the weight directly.
+            # Use PEFT's merge() which handles DoRA internally, then
+            # compute the delta as merged_weight - original_weight.
+            mag_layer = layer.lora_magnitude_vector[adapter_name]
+            mag_layer.weight = nn.Parameter(magnitude)
+            layer.merge(adapter_names=[adapter_name])
+            return base_layer.weight.data - base_tensor
+
+        return layer.get_delta_weight(adapter_name)
+
+
+def get_model_shards(model_path: Path) -> list[Path]:
+    """Find all model shards in the given path."""
+    shards: list[Path] = []
+
+    patterns = ["model*.safetensors", "pytorch_model*.bin"]
+
+    for pattern in patterns:
+        shards.extend(model_path.glob(pattern))
+        if shards:
+            break
+
+    return sorted(shards)
+
+
+def copy_non_model_files(
+    input_path: Path, output_path: Path, model_shards: list[Path]
+) -> None:
+    """
+    Copy all non-model files to the output directory.
+
+    Args:
+        input_path: Source directory
+        output_path: Destination directory
+        model_shards: List of model shard files to skip
+    """
+    LOG.info("Copying non-model files to output directory...")
+
+    shard_names = {shard.name for shard in model_shards}
+
+    for filepath in input_path.glob("*"):
+        if filepath.is_dir():
+            continue
+        if filepath.name in shard_names:
+            continue
+        if (
+            filepath.name.startswith("model") and filepath.suffix == ".safetensors"
+        ) or (filepath.name.startswith("pytorch_model") and filepath.suffix == ".bin"):
+            continue
+        if filepath.suffix == ".gguf":
+            continue
+        # Skip weight-map index files — they reference shard filenames that may
+        # change during the merge (e.g. .bin → .safetensors).  A correct index
+        # is regenerated after all shards have been written.
+        if filepath.name.endswith(".index.json"):
+            continue
+
+        LOG.debug(f"Copying {filepath.name} to output")
+        shutil.copy2(filepath, output_path)
+
+
+def _find_dora_magnitude(
+    lora_state: Dict[str, torch.Tensor],
+    key: str,
+    weight_renamings: Optional[Dict[str, str]] = None,
+) -> Optional[torch.Tensor]:
+    """
+    Find DoRA magnitude vector for a given key.
+    """
+    import re
+
+    clean_key = key[:-7] if key.endswith(".weight") else key
+    mag_key = f"base_model.model.{clean_key}.lora_magnitude_vector"
+    result = lora_state.get(mag_key)
+    if result is not None:
+        return result
+
+    if weight_renamings:
+        for src_pattern, tgt_pattern in weight_renamings.items():
+            renamed_key = re.sub(src_pattern, tgt_pattern, clean_key)
+            if renamed_key != clean_key:
+                mag_key = f"base_model.model.{renamed_key}.lora_magnitude_vector"
+                result = lora_state.get(mag_key)
+                if result is not None:
+                    return result
+
+    return None
+
+
+def _should_nf4_roundtrip(
+    key: str,
+    tensor: torch.Tensor,
+    simulate_nf4: bool,
+    simulate_nf4_experts: bool,
+) -> bool:
+    """Determine if a tensor should undergo NF4 quantization roundtrip."""
+    if tensor.ndim < 2:
+        return False
+    if simulate_nf4:
+        return True
+    if simulate_nf4_experts and tensor.ndim >= 3 and "expert" in key.lower():
+        return True
+    return False
+
+
+def _merge_tensor_with_lora(
+    tensor: torch.Tensor,
+    key: str,
+    lora_state: Dict[str, torch.Tensor],
+    scale: float,
+    lora_config_dict: Dict,
+    device: str,
+    simulate_nf4: bool = False,
+    simulate_nf4_experts: bool = False,
+    nf4_blocksize: Optional[int] = None,
+    nf4_double_quant: bool = True,
+    use_dora: bool = False,
+    weight_renamings: Optional[Dict[str, str]] = None,
+) -> tuple[torch.Tensor, bool]:
+    """
+    Helper function to merge a single tensor with its corresponding LoRA weights.
+
+    Args:
+        tensor: Base model tensor
+        key: Tensor key/name
+        lora_state: Dictionary containing LoRA weights
+        scale: LoRA scaling factor (alpha/r)
+        lora_config_dict: LoRA configuration dictionary
+        device: Device to perform computations on
+        simulate_nf4: Whether to simulate NF4 quantization roundtrip for all weights
+        simulate_nf4_experts: Whether to simulate NF4 roundtrip for MoE expert tensors only
+        nf4_blocksize: Block size for NF4 quantization
+        nf4_double_quant: Whether to use double quantization
+        use_dora: Whether to apply DoRA (Weight-Decomposed LoRA) merging
+        weight_renamings: Optional key renamings from transformers conversion mapping
+
+    Returns:
+        Tuple of (merged tensor, whether LoRA was applied)
+    """
+    lora_a, lora_b = find_lora_weights(lora_state, key, weight_renamings)
+
+    do_nf4 = _should_nf4_roundtrip(key, tensor, simulate_nf4, simulate_nf4_experts)
+
+    if lora_a is not None and lora_b is not None:
+        LOG.debug(f"Merging LoRA for {key}: {lora_a.shape}, {lora_b.shape}")
+
+        original_dtype = tensor.dtype
+
+        # Simulate NF4 quantization roundtrip to match QLoRA training dynamics
+        if do_nf4:
+            tensor = _simulate_nf4_roundtrip(
+                tensor,
+                blocksize=nf4_blocksize,
+                compress_statistics=nf4_double_quant,
+                device=device,
+            )
+
+        magnitude = (
+            _find_dora_magnitude(lora_state, key, weight_renamings)
+            if use_dora
+            else None
+        )
+        delta = _build_peft_layer_and_get_delta(
+            lora_a.to(device),
+            lora_b.to(device),
+            lora_config_dict,
+            tensor.to(device),
+            magnitude=magnitude.to(device) if magnitude is not None else None,
+        )
+        merged_tensor = (
+            (tensor.to(device).to(torch.float32) + delta.to(torch.float32))
+            .to(original_dtype)
+            .detach()
+            .cpu()
+        )
+        return merged_tensor, True
+    else:
+        # Try ParamWrapper LoRA (lora_target_parameters) — the LoRA targets a
+        # parent module and this weight is a sub-parameter of that module.
+        if tensor.ndim >= 3:
+            pw_a, pw_b, param_name = _find_param_wrapper_lora(
+                lora_state, key, tensor_shape=tuple(tensor.shape)
+            )
+            if pw_a is not None and pw_b is not None:
+                LOG.debug(
+                    f"Merging ParamWrapper LoRA for {key} "
+                    f"(param={param_name}): {pw_a.shape}, {pw_b.shape}"
+                )
+                if do_nf4:
+                    tensor = _simulate_nf4_roundtrip(
+                        tensor,
+                        blocksize=nf4_blocksize,
+                        compress_statistics=nf4_double_quant,
+                        device=device,
+                    )
+                original_dtype = tensor.dtype
+                delta = _build_peft_layer_and_get_delta(
+                    pw_a.to(device),
+                    pw_b.to(device),
+                    lora_config_dict,
+                    tensor.to(device),
+                    is_param_wrapper=True,
+                )
+                merged = (
+                    (tensor.to(device).to(torch.float32) + delta.to(torch.float32))
+                    .to(original_dtype)
+                    .detach()
+                    .cpu()
+                )
+                return merged, True
+
+        if do_nf4:
+            tensor = _simulate_nf4_roundtrip(
+                tensor,
+                blocksize=nf4_blocksize,
+                compress_statistics=nf4_double_quant,
+                device=device,
+            )
+        return tensor.detach().cpu(), False
+
+
+def _get_conversion_info(base_model_path: Path) -> tuple[Dict[str, str], list]:
+    """
+    Load the model's config.json and check if transformers has WeightRenaming
+    or WeightConverter mappings for this model type.
+
+    Returns:
+        - dict of {source_pattern: target_pattern} for simple renamings
+        - list of WeightConverter objects for fuse/unfuse operations
+    """
+    import json as _json
+
+    config_path = base_model_path / "config.json"
+    if not config_path.exists():
+        return {}, []
+
+    try:
+        with open(config_path) as f:
+            model_config = _json.load(f)
+    except (OSError, _json.JSONDecodeError):
+        return {}, []
+
+    model_type = model_config.get("model_type")
+    if not model_type:
+        return {}, []
+
+    try:
+        from transformers.conversion_mapping import get_checkpoint_conversion_mapping
+        from transformers.core_model_loading import WeightConverter, WeightRenaming
+    except ImportError:
+        return {}, []
+
+    conversions = get_checkpoint_conversion_mapping(model_type)
+    if not conversions:
+        return {}, []
+
+    renamings = {}
+    weight_converters = []
+    for conv in conversions:
+        if isinstance(conv, WeightRenaming):
+            # WeightRenaming stores patterns as lists internally
+            src_list = (
+                conv.source_patterns
+                if isinstance(conv.source_patterns, list)
+                else [conv.source_patterns]
+            )
+            tgt_list = (
+                conv.target_patterns
+                if isinstance(conv.target_patterns, list)
+                else [conv.target_patterns]
+            )
+            if len(src_list) == 1 and len(tgt_list) == 1:
+                renamings[src_list[0]] = tgt_list[0]
+        elif isinstance(conv, WeightConverter):
+            weight_converters.append(conv)
+
+    return renamings, weight_converters
+
+
+def _fuse_and_unfuse_with_merge(
+    shard_tensors: Dict[str, torch.Tensor],
+    weight_converters: list,
+    lora_state: Dict[str, torch.Tensor],
+    scale: float,
+    lora_config_dict: Dict,
+    device: str,
+    simulate_nf4: bool = False,
+    simulate_nf4_experts: bool = False,
+    nf4_blocksize: Optional[int] = None,
+    nf4_double_quant: bool = True,
+    use_dora: bool = False,
+    weight_renamings: Optional[Dict[str, str]] = None,
+) -> tuple[Dict[str, torch.Tensor], int, set]:
+    """
+    For tensors matching WeightConverter patterns (MoE expert weights):
+    1. Fuse checkpoint-format tensors into runtime-format (e.g., per-expert → fused 3D)
+    2. Apply NF4 roundtrip + LoRA merge on the fused tensor
+    3. Unfuse back to checkpoint format for saving
+
+    Returns:
+        - Updated tensor dict
+        - Count of merged LoRA targets
+        - Set of keys that were processed (fused/merged/unfused) and should be
+          skipped by the per-tensor merge pass to avoid double NF4 roundtrip
+    """
+    import re
+
+    from transformers.core_model_loading import Concatenate, MergeModulelist
+
+    result = dict(shard_tensors)  # Start with all tensors
+    merged_count = 0
+    processed_keys: set = set()  # Keys that were fuse/unfuse processed
+
+    for converter in weight_converters:
+        src_patterns = (
+            converter.source_patterns
+            if isinstance(converter.source_patterns, list)
+            else [converter.source_patterns]
+        )
+        tgt_patterns = (
+            converter.target_patterns
+            if isinstance(converter.target_patterns, list)
+            else [converter.target_patterns]
+        )
+
+        # Build regex for each source pattern
+        pattern_regexes = []
+        for pat in src_patterns:
+            regex_str = re.escape(pat).replace(r"\.\*\.", r"\.(\d+)\.")
+            regex_str = (
+                regex_str.rstrip(r"\$") if regex_str.endswith(r"\$") else regex_str
+            )
+            pattern_regexes.append(re.compile(r"(.*\.)?" + regex_str + "$"))
+
+        # Group matching keys by layer prefix and source pattern
+        # {layer_prefix: {pat_idx: {expert_idx: (key, tensor)}}}
+        layer_groups: Dict[str, Dict[int, Dict[int, tuple[str, torch.Tensor]]]] = {}
+
+        for key in list(result.keys()):
+            for pat_idx, pat_regex in enumerate(pattern_regexes):
+                match = pat_regex.match(key)
+                if match:
+                    prefix = match.group(1) or ""
+                    # Extract expert index from the matched portion
+                    remaining = key[len(prefix) :]
+                    expert_match = re.search(r"\.(\d+)\.", remaining)
+                    expert_idx = int(expert_match.group(1)) if expert_match else 0
+
+                    layer_groups.setdefault(prefix, {}).setdefault(pat_idx, {})[
+                        expert_idx
+                    ] = (key, result[key])
+                    break
+
+        # Process each layer group
+        for prefix, pat_groups in layer_groups.items():
+            # Check we have all source patterns for this layer
+            if not pat_groups:
+                continue
+
+            # Step 1: Fuse — MergeModulelist (stack experts) per source pattern
+            fused_per_pattern = {}
+            original_keys_per_pattern: Dict[int, list[str]] = {}
+            num_experts = None
+
+            for pat_idx in sorted(pat_groups.keys()):
+                expert_data = pat_groups[pat_idx]
+                sorted_indices = sorted(expert_data.keys())
+                if num_experts is None:
+                    num_experts = len(sorted_indices)
+
+                sorted_tensors = [expert_data[idx][1] for idx in sorted_indices]
+                original_keys_per_pattern[pat_idx] = [
+                    expert_data[idx][0] for idx in sorted_indices
+                ]
+                fused_per_pattern[src_patterns[pat_idx]] = torch.stack(
+                    sorted_tensors, dim=0
+                )
+
+            # Apply remaining operations (Concatenate)
+            fused_tensor = None
+            has_concat = False
+            concat_dim = 1  # default
+
+            for op in converter.operations:
+                if isinstance(op, MergeModulelist):
+                    pass  # Already handled
+                elif isinstance(op, Concatenate):
+                    has_concat = True
+                    concat_dim = op.dim
+                    tensors_to_cat = [
+                        fused_per_pattern[sp]
+                        for sp in src_patterns
+                        if sp in fused_per_pattern
+                    ]
+                    if len(tensors_to_cat) > 1:
+                        fused_tensor = torch.cat(tensors_to_cat, dim=concat_dim)
+                    elif tensors_to_cat:
+                        fused_tensor = tensors_to_cat[0]
+
+            if not has_concat and len(fused_per_pattern) == 1:
+                fused_tensor = next(iter(fused_per_pattern.values()))
+
+            if fused_tensor is None:
+                continue
+
+            # Step 2: Build the fused key name and merge LoRA
+            fused_key = prefix + tgt_patterns[0]
+
+            # Apply NF4 roundtrip on the fused tensor (matching training dynamics)
+            do_nf4 = _should_nf4_roundtrip(
+                fused_key, fused_tensor, simulate_nf4, simulate_nf4_experts
+            )
+            if do_nf4:
+                fused_tensor = _simulate_nf4_roundtrip(
+                    fused_tensor,
+                    blocksize=nf4_blocksize,
+                    compress_statistics=nf4_double_quant,
+                    device=device,
+                )
+
+            # Try to find and merge LoRA weights for the fused key
+            lora_a, lora_b = find_lora_weights(lora_state, fused_key, weight_renamings)
+            if lora_a is not None and lora_b is not None:
+                LOG.debug(
+                    f"Merging LoRA for fused key {fused_key}: {lora_a.shape}, {lora_b.shape}"
+                )
+                original_dtype = fused_tensor.dtype
+                magnitude = (
+                    _find_dora_magnitude(lora_state, fused_key, weight_renamings)
+                    if use_dora
+                    else None
+                )
+                delta = _build_peft_layer_and_get_delta(
+                    lora_a.to(device),
+                    lora_b.to(device),
+                    lora_config_dict,
+                    fused_tensor.to(device),
+                    magnitude=magnitude.to(device) if magnitude is not None else None,
+                )
+                fused_tensor = (
+                    (
+                        fused_tensor.to(device).to(torch.float32)
+                        + delta.to(torch.float32)
+                    )
+                    .to(original_dtype)
+                    .detach()
+                    .cpu()
+                )
+                merged_count += 1
+
+            # Step 3: Save in fused format (runtime format) so that the merged
+            # model can be loaded directly without needing WeightConverter
+            # fusion during from_pretrained (which can OOM for large MoE models).
+            # Remove the original per-expert keys and save the fused tensor
+            # under the runtime key name.
+            for pat_idx in sorted(original_keys_per_pattern.keys()):
+                for ok in original_keys_per_pattern[pat_idx]:
+                    result.pop(ok, None)
+                    processed_keys.add(ok)
+
+            result[fused_key] = fused_tensor.detach().cpu()
+            processed_keys.add(fused_key)
+
+    return result, merged_count, processed_keys
+
+
+def merge_lora_sharded_efficient(
+    base_model_path: Union[str, Path],
+    lora_adapter_path: Union[str, Path],
+    output_path: Union[str, Path],
+    device: str = "cpu",
+    safe_tensors: bool = True,
+    simulate_nf4: bool = False,
+    simulate_nf4_experts: bool = False,
+    nf4_blocksize: Optional[int] = None,
+    nf4_double_quant: bool = True,
+) -> None:
+    """
+    Memory-efficient LoRA merging that processes shards individually
+    without loading the full model into memory.
+
+    Args:
+        simulate_nf4: Apply NF4 roundtrip to ALL weight tensors (for QLoRA)
+        simulate_nf4_experts: Apply NF4 roundtrip only to MoE expert tensors
+            (for quantize_moe_experts). Expert tensors are identified by having
+            "expert" in the key name and ndim >= 3.
+    """
+    base_model_path = Path(base_model_path)
+    lora_adapter_path = Path(lora_adapter_path)
+    output_path = Path(output_path)
+
+    if "/" in str(base_model_path) and not base_model_path.exists():
+        base_model_path = Path(snapshot_download(str(base_model_path)))
+
+    # Check for weight conversion requirements (transformers v5)
+    weight_renamings, weight_converters = _get_conversion_info(base_model_path)
+    if weight_renamings:
+        LOG.debug(f"Found {len(weight_renamings)} weight renamings for this model type")
+    if weight_converters:
+        LOG.debug(
+            f"Found {len(weight_converters)} weight converters (fuse/unfuse) for this model type. "
+            f"Will fuse→merge→unfuse within each shard."
+        )
+
+    os.makedirs(output_path, exist_ok=True)
+
+    config_file = lora_adapter_path / "adapter_config.json"
+    if not config_file.exists():
+        raise FileNotFoundError(f"LoRA config not found: {config_file}")
+
+    lora_config_dict = LoraConfig.from_json_file(str(config_file))
+    if not lora_config_dict.get("r") or lora_config_dict["r"] <= 0:
+        raise ValueError("LoRA config 'r' must be > 0")
+
+    use_dora = bool(lora_config_dict.get("use_dora", False))
+
+    unsupported_methods = []
+
+    # Check for AdaLoRA (Adaptive LoRA)
+    if lora_config_dict.get("use_adalora", False):
+        unsupported_methods.append("AdaLoRA (Adaptive LoRA)")
+
+    # Check for VeRA (Vector-based Random Matrix Adaptation)
+    if lora_config_dict.get("use_vera", False):
+        unsupported_methods.append("VeRA (Vector-based Random Matrix Adaptation)")
+
+    # Check for other advanced LoRA variants by task_type
+    task_type = lora_config_dict.get("task_type", "")
+    if task_type and task_type not in [
+        "CAUSAL_LM",
+        "SEQ_2_SEQ_LM",
+        "TOKEN_CLS",
+        "SEQ_CLS",
+        "QUESTION_ANS",
+    ]:
+        unsupported_methods.append(f"Task type: {task_type}")
+
+    # Check for rank adaptation patterns (AdaLoRA indicators)
+    # Use .get() so empty dicts/None don't false-positive
+    if any(
+        lora_config_dict.get(key)
+        for key in ["rank_pattern", "alpha_pattern", "target_rank"]
+    ):
+        unsupported_methods.append("AdaLoRA (rank adaptation detected)")
+
+    # Check for advanced initialization methods
+    init_lora_weights = lora_config_dict.get("init_lora_weights", "")
+    if init_lora_weights and init_lora_weights not in [
+        "gaussian",
+        "loftq",
+        True,
+        False,
+    ]:
+        unsupported_methods.append(f"Advanced initialization: {init_lora_weights}")
+
+    if unsupported_methods:
+        methods_str = ", ".join(unsupported_methods)
+        raise NotImplementedError(
+            f"Memory-efficient LoRA merge only supports standard LoRA. "
+            f"Detected unsupported methods: {methods_str}. "
+            f"Please use the legacy merge method for advanced LoRA variants."
+        )
+
+    use_rslora = bool(lora_config_dict.get("use_rslora", False))
+    if use_rslora:
+        scale = float(lora_config_dict["lora_alpha"]) / math.sqrt(
+            float(lora_config_dict["r"])
+        )
+    else:
+        scale = float(lora_config_dict["lora_alpha"]) / float(lora_config_dict["r"])
+
+    LOG.debug(f"LoRA scale factor: {scale} (rslora={use_rslora})")
+
+    if simulate_nf4:
+        LOG.info(
+            "NF4 simulation enabled: base weights will undergo quantize→dequantize "
+            "roundtrip before LoRA merge to match QLoRA training dynamics"
+        )
+
+    lora_file = lora_adapter_path / "adapter_model.safetensors"
+    if not lora_file.exists():
+        lora_file = lora_adapter_path / "adapter_model.bin"
+        if not lora_file.exists():
+            raise FileNotFoundError(
+                f"LoRA adapter weights not found in {lora_adapter_path}"
+            )
+
+    LOG.debug(f"Loading LoRA weights from {lora_file}")
+
+    if lora_file.suffix == ".safetensors":
+        lora_state = safetensors.torch.load_file(lora_file)
+    else:
+        lora_state = torch.load(lora_file, map_location="cpu", weights_only=True)  # nosec B614
+    LOG.debug("Keeping LoRA weights on CPU; will move per-tensor during merge")
+
+    model_shards = get_model_shards(base_model_path)
+    if not model_shards:
+        raise FileNotFoundError(f"No model shards found in {base_model_path}")
+
+    LOG.debug(f"Found {len(model_shards)} model shards in {base_model_path}")
+    copy_non_model_files(base_model_path, output_path, model_shards)
+
+    merged_count = 0
+    total_tensors = 0
+    # Track weight_map for index regeneration: {tensor_key: shard_filename}
+    weight_map: Dict[str, str] = {}
+
+    for shard_path in tqdm(model_shards, desc="Merging shards"):
+        merged_tensors = {}
+        metadata = {}
+
+        # Load all tensors from the shard
+        if shard_path.suffix == ".safetensors":
+            with safetensors.safe_open(shard_path, framework="pt", device="cpu") as f:
+                if hasattr(f, "metadata") and f.metadata():
+                    metadata = f.metadata()
+                shard_tensors = {key: f.get_tensor(key) for key in f.keys()}
+        else:
+            shard_tensors = torch.load(  # nosec B614: loading trusted model weights
+                shard_path, map_location="cpu", weights_only=True
+            )
+
+        total_tensors += len(shard_tensors)
+
+        # Step 1: Handle fused weight conversions (MoE experts) if applicable
+        fused_keys: set = set()
+        if weight_converters:
+            shard_tensors, fused_merged, fused_keys = _fuse_and_unfuse_with_merge(
+                shard_tensors,
+                weight_converters,
+                lora_state,
+                scale,
+                lora_config_dict,
+                device,
+                simulate_nf4=simulate_nf4,
+                simulate_nf4_experts=simulate_nf4_experts,
+                nf4_blocksize=nf4_blocksize,
+                nf4_double_quant=nf4_double_quant,
+                use_dora=use_dora,
+                weight_renamings=weight_renamings,
+            )
+            merged_count += fused_merged
+
+        # Step 2: Merge remaining (non-fused) tensors with LoRA
+        # Skip keys already processed by fuse/unfuse to avoid double NF4 roundtrip
+        for key, tensor in shard_tensors.items():
+            if key in fused_keys:
+                merged_tensors[key] = tensor.detach().cpu()
+                continue
+            merged_tensor, was_merged = _merge_tensor_with_lora(
+                tensor,
+                key,
+                lora_state,
+                scale,
+                lora_config_dict,
+                device,
+                simulate_nf4=simulate_nf4,
+                simulate_nf4_experts=simulate_nf4_experts,
+                nf4_blocksize=nf4_blocksize,
+                nf4_double_quant=nf4_double_quant,
+                use_dora=use_dora,
+                weight_renamings=weight_renamings,
+            )
+            merged_tensors[key] = merged_tensor
+            if was_merged:
+                merged_count += 1
+
+        output_shard_path = output_path / shard_path.name
+        merged_tensors = {k: v.detach().cpu() for k, v in merged_tensors.items()}
+
+        if safe_tensors:
+            if not str(output_shard_path).endswith(".safetensors"):
+                output_shard_path = output_path / (shard_path.stem + ".safetensors")
+            safetensors.torch.save_file(
+                merged_tensors, output_shard_path, metadata=metadata
+            )
+        else:
+            if shard_path.suffix == ".safetensors":
+                safetensors.torch.save_file(
+                    merged_tensors, output_shard_path, metadata=metadata
+                )
+            else:
+                torch.save(merged_tensors, output_shard_path)
+
+        for tensor_key in merged_tensors:
+            weight_map[tensor_key] = output_shard_path.name
+
+        del merged_tensors, shard_tensors
+        if device != "cpu" and torch.cuda.is_available():
+            torch.cuda.empty_cache()
+        gc.collect()
+
+    # Regenerate weight-map index if the model was sharded
+    if len(model_shards) > 1 and weight_map:
+        import json as _json
+
+        index_name = (
+            "model.safetensors.index.json"
+            if safe_tensors
+            else "pytorch_model.bin.index.json"
+        )
+        index = {
+            "metadata": {"total_size": total_tensors},
+            "weight_map": weight_map,
+        }
+        with open(output_path / index_name, "w") as f:
+            _json.dump(index, f, indent=2)
+        LOG.debug(f"Wrote weight-map index: {index_name}")
+
+    if merged_count == 0:
+        LOG.warning(
+            "No LoRA weights were matched to base model tensors. "
+            "This may indicate a key name mismatch between the checkpoint format "
+            "and the LoRA adapter. Consider using merge_method: legacy."
+        )
+    LOG.info(f"Applied LoRA to {merged_count}/{total_tensors} tensors")

src/axolotl/cli/vllm_serve.py

@@ -38,7 +38,14 @@ def do_vllm_serve(
     cfg = load_cfg(config)
     model = cfg.base_model
 
-    serve_module = cli_args.get("serve_module", "trl.scripts.vllm_serve")
+    # Determine serve module: explicit CLI/config > default (axolotl's LoRA-aware serve).
+    # We default to axolotl's serve module instead of TRL's because TRL's sends
+    # truncate_prompt_tokens which is unsupported in vLLM 0.17+.
+    serve_module = cli_args.get("serve_module") or getattr(
+        cfg.vllm, "serve_module", None
+    )
+    if serve_module is None:
+        serve_module = "axolotl.scripts.vllm_serve_lora"
     vllm_serve_main = __import__(serve_module, fromlist=["main"]).main
     tensor_parallel_size = 1
     data_parallel_size = 1
@@ -68,7 +75,13 @@ def do_vllm_serve(
         cli_args.get("enable_reasoning") or cfg.vllm.enable_reasoning or False
     )
 
-    vllm_script_args = AxolotlScriptArguments(
+    cli_enforce_eager = cli_args.get("enforce_eager")
+    cfg_enforce_eager = getattr(cfg.vllm, "enforce_eager", None)
+    raw_enforce_eager = (
+        cfg_enforce_eager if cli_enforce_eager is None else cli_enforce_eager
+    )
+    enforce_eager = bool(raw_enforce_eager) if raw_enforce_eager is not None else False
+    base_kwargs = dict(
         model=model,
         tensor_parallel_size=tensor_parallel_size,
         data_parallel_size=data_parallel_size,
@@ -78,7 +91,28 @@ def do_vllm_serve(
         dtype=dtype,
         max_model_len=max_model_len,
         enable_prefix_caching=enable_prefix_caching,
-        reasoning_parser=reasoning_parser,
-        enable_reasoning=enable_reasoning,
+        enforce_eager=enforce_eager,
     )
+
+    # Use LoRAScriptArguments when serving with native LoRA support
+    if serve_module == "axolotl.scripts.vllm_serve_lora":
+        from axolotl.scripts.vllm_serve_lora import LoRAScriptArguments
+
+        lora_kwargs = {}
+        if hasattr(cfg, "lora_r") and cfg.lora_r:
+            lora_kwargs["max_lora_rank"] = cfg.lora_r
+        # Disable native LoRA in vLLM if not using vllm_lora_sync
+        # (merged weight sync via batch_update doesn't need vLLM LoRA mode)
+        if not getattr(cfg.trl, "vllm_lora_sync", False):
+            lora_kwargs["enable_lora"] = False
+        if getattr(cfg.vllm, "worker_extension_cls", None):
+            lora_kwargs["worker_extension_cls"] = cfg.vllm.worker_extension_cls
+        vllm_script_args = LoRAScriptArguments(**base_kwargs, **lora_kwargs)
+    else:
+        vllm_script_args = AxolotlScriptArguments(
+            **base_kwargs,
+            reasoning_parser=reasoning_parser,
+            enable_reasoning=enable_reasoning,
+        )
+
     vllm_serve_main(vllm_script_args)

src/axolotl/common/architectures.py

@@ -16,10 +16,12 @@ MOE_ARCH_BLOCK = {
     "qwen3_vl_moe": "Qwen3VLMoeTextSparseMoeBlock",
     "deepseek_v2": "DeepseekV2MoE",
     "deepseek_v3": "DeepseekV3MoE",
+    "mistral4": "Mistral4MoE",
     "gpt_oss": "GptOssDecoderLayer",
     "lfm2_moe": "Lfm2MoeSparseMoeBlock",
     "afmoe": "AfmoeMoE",
     "glm4_moe": "Glm4MoeDecoderLayer",
     "glm4_moe_lite": "Glm4MoeLiteDecoderLayer",
     "glm_moe_dsa": "GlmMoeDsaDecoderLayer",
+    "nemotron_h": "NemotronHMoE",
 }

src/axolotl/common/datasets.py

@@ -118,7 +118,7 @@ def load_preference_datasets(
     train_dataset, eval_dataset = prepare_preference_datasets(cfg, tokenizer)
 
     total_num_steps: int | None = None
-    if cfg.rl is not RLType.GRPO:
+    if cfg.rl not in {RLType.GRPO, RLType.EBFT}:
         total_num_steps = int(
             math.ceil(len(train_dataset) * cfg.num_epochs / cfg.batch_size)
         )

src/axolotl/convert.py

@@ -67,7 +67,7 @@ class JsonToJsonlConverter:
         self.json_parser = json_parser
         self.jsonl_serializer = jsonl_serializer
 
-    def convert(self, input_file_path, output_file_path):
+    def convert(self, input_file_path):
         content = self.file_reader.read(input_file_path)
         data = self.json_parser.parse(content)
         # data = [r for r in data if r["conversations"]]  # vicuna cleaned has rows with empty conversations

src/axolotl/core/builders/base.py

@@ -250,7 +250,7 @@ class TrainerBuilderBase(abc.ABC):
 
     def _configure_precision_settings(self, training_args_kwargs: dict):
         training_args_kwargs["fp16"] = (self.cfg.fp16 and not self.cfg.bf16) or False
-        training_args_kwargs["tf32"] = self.cfg.tf32
+        training_args_kwargs["tf32"] = True if self.cfg.tf32 is True else False
         if self.cfg.bf16 == "full":
             training_args_kwargs["bf16_full_eval"] = True
         else:
@@ -329,7 +329,7 @@ class TrainerBuilderBase(abc.ABC):
                 optimizer_cls = AdamW
                 optimizer_kwargs.update(adam_kwargs)
             elif self.cfg.optimizer == "ao_adamw_fp8":
-                from torchao.prototype.low_bit_optim import AdamWFp8
+                from torchao.optim.adam import AdamWFp8
 
                 optimizer_cls = AdamWFp8
                 optimizer_kwargs.update(adam_kwargs)
@@ -353,6 +353,30 @@ class TrainerBuilderBase(abc.ABC):
                 adam_kwargs["eps"] = (eps1, eps2)
 
                 optimizer_kwargs.update(adam_kwargs)
+            elif self.cfg.optimizer == "flash_adamw":
+                from flashoptim import FlashAdamW
+
+                optimizer_cls = FlashAdamW
+                optimizer_kwargs.update(adam_kwargs)
+            elif self.cfg.optimizer == "flash_adam":
+                from flashoptim import FlashAdam
+
+                optimizer_cls = FlashAdam
+                optimizer_kwargs.update(adam_kwargs)
+            elif self.cfg.optimizer == "flash_sgd":
+                from flashoptim import FlashSGD
+
+                optimizer_cls = FlashSGD
+            elif self.cfg.optimizer == "flash_sgdw":
+                from flashoptim import FlashSGDW
+
+                optimizer_cls = FlashSGDW
+            elif self.cfg.optimizer == "flash_lion":
+                from flashoptim import FlashLion
+
+                optimizer_cls = FlashLion
+                if "betas" in adam_kwargs:
+                    optimizer_kwargs["betas"] = adam_kwargs["betas"]
             else:
                 raise ValueError(
                     f"Unhandled optimizer: {self.cfg.optimizer}. Please raise an Issue."
@@ -484,6 +508,8 @@ class TrainerBuilderBase(abc.ABC):
             training_args_kwargs["accelerator_config"] = AcceleratorConfig()
 
     def _configure_gradient_checkpointing(self, training_args_kwargs: dict):
+        if self.cfg.layer_offloading:
+            training_args_kwargs["layer_offloading"] = True
         if self.cfg.activation_offloading is True:
             # don't use the HF gradient checkpointing, manually wrap
             training_args_kwargs["gradient_checkpointing"] = False

src/axolotl/core/builders/causal.py

@@ -421,6 +421,13 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             trainer_kwargs["dataset_tags"] = [
                 d["path"] for d in self.cfg.datasets if not Path(d["path"]).is_dir()
             ]
+        # TRL's RewardTrainer validates num_labels=1 on pre-loaded models; ensure the
+        # config reflects this regardless of how the model was instantiated.
+        if (
+            self.cfg.reward_model
+            and getattr(self.model.config, "num_labels", None) != 1
+        ):
+            self.model.config.num_labels = 1
         trainer = trainer_cls(
             model=self.model,
             train_dataset=self.train_dataset,

src/axolotl/core/builders/rl.py

@@ -54,8 +54,16 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         if self.cfg.rl in {RLType.GRPO, RLType.GDPO}:
             from axolotl.core.trainers.grpo import GRPOStrategy
 
+            async_grpo = bool(
+                self.cfg.trl
+                and (
+                    getattr(self.cfg.trl, "async_prefetch", False)
+                    or getattr(self.cfg.trl, "use_data_producer", False)
+                )
+            )
             trainer_cls = GRPOStrategy.get_trainer_class(
-                sequence_parallel=self.cfg.context_parallel_size > 1
+                sequence_parallel=self.cfg.context_parallel_size > 1,
+                async_grpo=async_grpo,
             )
             trainer_cls_args.extend(GRPOStrategy.set_trainer_args(self.cfg))
             trainer_kwargs.update(GRPOStrategy.set_trainer_kwargs(self.cfg))
@@ -70,6 +78,11 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             trainer_cls = AxolotlKTOTrainer
         elif self.cfg.rl is RLType.SIMPO:
             trainer_cls = AxolotlCPOTrainer
+        elif self.cfg.rl is RLType.EBFT:
+            from axolotl.core.trainers.ebft import EBFTStrategy
+
+            trainer_cls = EBFTStrategy.get_trainer_class(self.cfg)
+            trainer_kwargs.update(EBFTStrategy.set_trainer_kwargs(self.cfg))
         else:
             raise ValueError(f"Unsupported RL: {self.cfg.rl}")
 
@@ -114,9 +127,6 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             # trl does some odd mapping of alpha to beta to reuse the beta parameter ???
             training_args_kwargs["beta"] = self.cfg.orpo_alpha
 
-        if self.cfg.rpo_alpha is not None:
-            training_args_kwargs["rpo_alpha"] = self.cfg.rpo_alpha
-
         if self.cfg.use_wandb:
             training_args_kwargs["run_name"] = self.cfg.wandb_name
 
@@ -151,9 +161,34 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         elif self.cfg.rl in {RLType.GRPO, RLType.GDPO}:
             from axolotl.core.trainers.grpo import GRPOStrategy
 
-            training_args_cls = GRPOStrategy.get_training_args_class()
+            async_grpo = bool(
+                self.cfg.trl
+                and (
+                    getattr(self.cfg.trl, "async_prefetch", False)
+                    or getattr(self.cfg.trl, "use_data_producer", False)
+                )
+            )
+            training_args_cls = GRPOStrategy.get_training_args_class(
+                async_grpo=async_grpo
+            )
             training_args_kwargs.update(GRPOStrategy.set_training_args_kwargs(self.cfg))
             blocklist_args_kwargs = GRPOStrategy.get_blocklist_args_kwargs()
+            if not async_grpo:
+                # Filter out async/fast-async-only fields not in standard GRPOConfig.
+                # These are defined in FastAsyncGRPOConfig and only used by
+                # AxolotlAsyncGRPOConfig. Standard GRPOConfig rejects them.
+                import dataclasses
+
+                from trl import GRPOConfig as _BaseGRPOConfig
+
+                from axolotl.core.trainers.grpo.fast_async_trainer import (
+                    FastAsyncGRPOConfig,
+                )
+
+                async_only_fields = {
+                    f.name for f in dataclasses.fields(FastAsyncGRPOConfig)
+                } - {f.name for f in dataclasses.fields(_BaseGRPOConfig)}
+                blocklist_args_kwargs.extend(list(async_only_fields))
             if self.cfg.rl is RLType.GDPO:
                 training_args_kwargs.setdefault(
                     "multi_objective_aggregation", "normalize_then_sum"
@@ -162,6 +197,13 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
         elif self.cfg.rl in [RLType.DPO, RLType.IPO]:
             training_args_cls = AxolotlDPOConfig
             training_args_kwargs.update(DPOStrategy.set_training_args_kwargs(self.cfg))
+
+        elif self.cfg.rl is RLType.EBFT:
+            from axolotl.core.trainers.ebft import EBFTStrategy
+
+            training_args_cls = EBFTStrategy.get_training_args_class(self.cfg)
+            training_args_kwargs.update(EBFTStrategy.set_training_args_kwargs(self.cfg))
+            blocklist_args_kwargs = EBFTStrategy.get_blocklist_args_kwargs(self.cfg)
         else:
             raise ValueError(f"Unsupported RL: {self.cfg.rl}")
 
@@ -191,12 +233,12 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
 
         if self.eval_dataset:
             trainer_kwargs["eval_dataset"] = self.eval_dataset
-        if self.cfg.adapter and self.peft_config and self.cfg.rl is not RLType.GRPO:
+        if (
+            self.cfg.adapter
+            and self.peft_config
+            and self.cfg.rl not in (RLType.GRPO, RLType.ORPO, RLType.EBFT)
+        ):
             trainer_kwargs["peft_config"] = self.peft_config
-        if self.cfg.precompute_ref_log_probs is not None:
-            trainer_kwargs["precompute_ref_log_probs"] = (
-                self.cfg.precompute_ref_log_probs
-            )
 
         trainer_cls, trainer_cls_args = self._get_trainer_cls(trainer_kwargs)
 
@@ -217,13 +259,36 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
             trainer_kwargs, trainer_cls
         )
 
-        trainer = trainer_cls(
-            *trainer_cls_args,
-            args=training_args,
-            train_dataset=self.train_dataset,
-            callbacks=self.get_callbacks(),
-            **trainer_kwargs,
-        )
+        # Allow FP8-quantized models to be fine-tuned with LoRA adapters.
+        # transformers' validate_quantization_for_training blocks FP8 because
+        # hf_quantizer.is_trainable is False, but LoRA only trains the adapters
+        # (base weights stay frozen in FP8).
+        _orig_validate_quant = None
+        if (
+            self.cfg.adapter
+            and hasattr(self.model, "is_quantized")
+            and self.model.is_quantized
+        ):
+            import transformers.trainer as _trainer_module
+
+            _orig_validate_quant = _trainer_module.validate_quantization_for_training
+            _trainer_module.validate_quantization_for_training = lambda model: None
+
+        try:
+            trainer = trainer_cls(
+                *trainer_cls_args,
+                args=training_args,
+                train_dataset=self.train_dataset,
+                callbacks=self.get_callbacks(),
+                **trainer_kwargs,
+            )
+        finally:
+            if _orig_validate_quant is not None:
+                import transformers.trainer as _trainer_module
+
+                _trainer_module.validate_quantization_for_training = (
+                    _orig_validate_quant
+                )
         if self.cfg.fsdp_config or self.cfg.fsdp:
             ensure_dtype(trainer.model, dtype=self.cfg.torch_dtype)
             if self.cfg.rl in [RLType.DPO, RLType.IPO] and trainer.ref_model: