hmmm

* feat: update thinking and chat_template notes

* fix: grammar

.coderabbit.yaml

@@ -12,6 +12,6 @@ reviews:
   auto_review:
     enabled: true
     drafts: false
-    auto_incremental_review: true
+    auto_incremental_review: false
 chat:
   auto_reply: true

.github/workflows/main.yml

@@ -36,6 +36,11 @@ jobs:
             python_version: "3.11"
             pytorch: 2.7.1
             axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.8.0
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -110,6 +115,11 @@ jobs:
             python_version: "3.11"
             pytorch: 2.7.1
             axolotl_extras:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.8.0
+            axolotl_extras:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout
@@ -169,6 +179,12 @@ jobs:
             pytorch: 2.7.1
             axolotl_extras: vllm
             is_latest: true
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.8.0
+            axolotl_extras:
+            is_latest:
     runs-on: axolotl-gpu-runner
     steps:
       - name: Checkout

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

@@ -33,13 +33,6 @@ jobs:
             axolotl_extras:
             num_gpus: 2
             nightly_build: "true"
-          - cuda: 126
-            cuda_version: 12.6.3
-            python_version: "3.11"
-            pytorch: 2.7.0
-            axolotl_extras:
-            num_gpus: 2
-            nightly_build: "true"
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
@@ -47,6 +40,13 @@ jobs:
             axolotl_extras: vllm
             num_gpus: 2
             nightly_build: "true"
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.8.0
+            axolotl_extras: fbgemm-gpu
+            num_gpus: 2
+            nightly_build: "true"
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     steps:

.github/workflows/tests.yml

@@ -55,7 +55,7 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.6.0", "2.7.0", "2.7.1"]
+        pytorch_version: ["2.6.0", "2.7.1", "2.8.0"]
     timeout-minutes: 20
 
     steps:
@@ -130,7 +130,7 @@ jobs:
       fail-fast: false
       matrix:
         python_version: ["3.11"]
-        pytorch_version: ["2.6.0", "2.7.0", "2.7.1"]
+        pytorch_version: ["2.6.0", "2.7.1", "2.8.0"]
     timeout-minutes: 20
 
     steps:
@@ -240,7 +240,7 @@ jobs:
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.7.1
             num_gpus: 1
             axolotl_extras:
             dockerfile: "Dockerfile-uv.jinja"
@@ -298,6 +298,13 @@ jobs:
             pytorch: 2.7.1
             num_gpus: 1
             axolotl_extras:
+          - 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
     steps:
       - name: Checkout
         uses: actions/checkout@v4
@@ -318,6 +325,7 @@ jobs:
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "MODAL_IMAGE_BUILDER_VERSION=2024.10" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
+          echo "GPU_TYPE=${{ 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
@@ -334,10 +342,10 @@ jobs:
       fail-fast: false
       matrix:
         include:
-          - cuda: 124
-            cuda_version: 12.4.1
+          - cuda: 126
+            cuda_version: 12.6.3
             python_version: "3.11"
-            pytorch: 2.6.0
+            pytorch: 2.7.1
             num_gpus: 1
             axolotl_extras:
     steps:

.pre-commit-config.yaml

@@ -11,10 +11,10 @@ repos:
     -   id: no-commit-to-branch
         args: ['--branch', 'main']
 -   repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.12.9
+    rev: v0.12.12
     hooks:
     -   id: ruff
-        args: [--fix]
+        args: [--fix, --select, I]
     -   id: ruff-format
 -   repo: https://github.com/pre-commit/mirrors-mypy
     rev: v1.17.1

CITATION.cff

@@ -1,6 +1,6 @@
 cff-version: 1.2.0
 type: software
-title: "Axolotl: Post-Training for AI Models"
+title: "Axolotl: Open Source LLM Post-Training"
 message: "If you use this software, please cite it as below."
 authors:
   - name: "Axolotl maintainers and contributors"

README.md

@@ -5,6 +5,9 @@
         <img alt="Axolotl" src="https://raw.githubusercontent.com/axolotl-ai-cloud/axolotl/887513285d98132142bf5db2a74eb5e0928787f1/image/axolotl_logo_digital_black.svg" width="400" height="104" style="max-width: 100%;">
     </picture>
 </p>
+  <p align="center">
+      <strong>A Free and Open Source LLM Fine-tuning Framework</strong><br>
+  </p>
 
 <p align="center">
     <img src="https://img.shields.io/github/license/axolotl-ai-cloud/axolotl.svg?color=blue" alt="GitHub License">
@@ -17,6 +20,7 @@
     <br/>
     <a href="https://discord.com/invite/HhrNrHJPRb"><img src="https://img.shields.io/badge/discord-7289da.svg?style=flat-square&logo=discord" alt="discord" style="height: 20px;"></a>
     <a href="https://twitter.com/axolotl_ai"><img src="https://img.shields.io/twitter/follow/axolotl_ai?style=social" alt="twitter" style="height: 20px;"></a>
+    <a href="https://colab.research.google.com/github/axolotl-ai-cloud/axolotl/blob/main/examples/colab-notebooks/colab-axolotl-example.ipynb"><img src="https://colab.research.google.com/assets/colab-badge.svg" alt="google-colab" style="height: 20px;"></a>
     <br/>
     <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/tests-nightly.yml/badge.svg" alt="tests-nightly">
     <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/multi-gpu-e2e.yml/badge.svg" alt="multigpu-semi-weekly tests">
@@ -49,20 +53,21 @@
 
 ## ✨ Overview
 
-Axolotl is a tool designed to streamline post-training for various AI models.
+Axolotl is a free and open-source tool designed to streamline post-training and fine-tuning for the latest large language models (LLMs).
 
 Features:
 
-- **Multiple Model Support**: Train various models like LLaMA, Mistral, Mixtral, Pythia, and more. We are compatible with HuggingFace transformers causal language models.
-- **Training Methods**: Full fine-tuning, LoRA, QLoRA, GPTQ, QAT, Preference Tuning (DPO, IPO, KTO, ORPO), RL (GRPO), Multimodal, and Reward Modelling (RM) / Process Reward Modelling (PRM).
-- **Easy Configuration**: Re-use a single YAML file between dataset preprocess, training, evaluation, quantization, and inference.
+- **Multiple Model Support**: Train various models like GPT-OSS, LLaMA, Mistral, Mixtral, Pythia, and many more models available on the Hugging Face Hub.
+- **Multimodal Training**: Fine-tune vision-language models (VLMs) including LLaMA-Vision, Qwen2-VL, Pixtral, LLaVA, SmolVLM2, 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), 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/), [Liger Kernel](https://github.com/linkedin/Liger-Kernel), [Cut Cross Entropy](https://github.com/apple/ml-cross-entropy/tree/main), [Sequence Parallelism (SP)](https://docs.axolotl.ai/docs/sequence_parallelism.html), [LoRA optimizations](https://docs.axolotl.ai/docs/lora_optims.html), [Multi-GPU training (FSDP1, FSDP2, DeepSpeed)](https://docs.axolotl.ai/docs/multi-gpu.html), [Multi-node training (Torchrun, Ray)](https://docs.axolotl.ai/docs/multi-node.html), and many more!
 - **Flexible Dataset Handling**: Load from local, HuggingFace, and cloud (S3, Azure, GCP, OCI) datasets.
 - **Cloud Ready**: We ship [Docker images](https://hub.docker.com/u/axolotlai) and also [PyPI packages](https://pypi.org/project/axolotl/) for use on cloud platforms and local hardware.
 
 
 
-## 🚀 Quick Start
+## 🚀 Quick Start - LLM Fine-tuning in Minutes
 
 **Requirements**:
 
@@ -70,6 +75,10 @@ Features:
 - Python 3.11
 - PyTorch ≥2.6.0
 
+### Google Colab
+
+[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/axolotl-ai-cloud/axolotl/blob/main/examples/colab-notebooks/colab-axolotl-example.ipynb#scrollTo=msOCO4NRmRLa)
+
 ### Installation
 
 #### Using pip
@@ -155,7 +164,7 @@ If you use Axolotl in your research or projects, please cite it as follows:
 
 ```bibtex
 @software{axolotl,
-  title = {Axolotl: Post-Training for AI Models},
+  title = {Axolotl: Open Source LLM Post-Training},
   author = {{Axolotl maintainers and contributors}},
   url = {https://github.com/axolotl-ai-cloud/axolotl},
   license = {Apache-2.0},

_quarto.yml

@@ -153,7 +153,7 @@ quartodoc:
         - utils.distributed
         - utils.dict
         - utils.optimizers.adopt
-        - utils.data.pretraining
+        - utils.data.streaming
         - utils.data.sft
         - utils.quantization
     - title: Schemas
@@ -272,6 +272,7 @@ website:
           contents:
             - docs/batch_vs_grad.qmd
             - docs/dataset_preprocessing.qmd
+            - docs/streaming.qmd
             - docs/multipack.qmd
             - docs/mixed_precision.qmd
             - docs/optimizers.qmd

cicd/single_gpu.py

@@ -57,7 +57,8 @@ VOLUME_CONFIG = {
 }
 
 N_GPUS = int(os.environ.get("N_GPUS", 1))
-GPU_CONFIG = f"L40S:{N_GPUS}"
+GPU_TYPE = os.environ.get("GPU_TYPE", "L40S")
+GPU_CONFIG = f"{GPU_TYPE}:{N_GPUS}"
 
 
 def run_cmd(cmd: str, run_folder: str):

codecov.yml

@@ -12,7 +12,7 @@ coverage:
       default:
         # basic
         target: auto
-        threshold: 0%
+        threshold: 1%
         base: auto
         # advanced
         branches: null
@@ -27,7 +27,7 @@ coverage:
       default:
         # basic
         target: auto
-        threshold: 0%
+        threshold: 1%
         base: auto
         # advanced
         branches: null

docs/installation.qmd

@@ -134,7 +134,7 @@ For providers supporting Docker:
 
 ### Google Colab {#sec-colab}
 
-Use our [example notebook](../examples/colab-notebooks/colab-axolotl-example.ipynb).
+[![](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/axolotl-ai-cloud/axolotl/blob/main/examples/colab-notebooks/colab-axolotl-example.ipynb#scrollTo=msOCO4NRmRLa)
 
 ## Platform-Specific Instructions {#sec-platform-specific}
 

docs/multi-gpu.qmd

@@ -63,15 +63,6 @@ Start from Stage 1 -> Stage 2 -> Stage 3.
 
 :::
 
-::: {.callout-tip}
-
-Using ZeRO Stage 3 with Single-GPU training
-
-ZeRO Stage 3 can be used for training on a single GPU by manually setting the environment variables:
-`WORLD_SIZE=1 LOCAL_RANK=0 MASTER_ADDR=0.0.0.0 MASTER_PORT=29500`
-
-:::
-
 ## Fully Sharded Data Parallel (FSDP) {#sec-fsdp}
 
 ::: {.callout-note}

docs/quantize.qmd

@@ -51,3 +51,11 @@ axolotl quantize qat.yml
 ```
 
 This ensures that an identical quantization configuration is used to quantize the model as was used to train it.
+
+
+::: {.callout-note}
+
+If you have configured pushing to hub with `hub_model_id`, your model hub name will have the quantization schema appended to it,
+e.g. `axolotl-ai-cloud/qat-nvfp4-llama3B` will become `axolotl-ai-cloud/qat-nvfp4-llama3B-nvfp4w`
+
+:::

docs/reward_modelling.qmd

@@ -11,6 +11,7 @@ We support the reward modelling techniques supported by `trl`.
 ### (Outcome) Reward Models
 
 Outcome reward models are trained using data which contains preference annotations for an entire interaction between the user and model (e.g. rather than per-turn or per-step).
+For improved training stability, you can use the `center_rewards_coefficient` parameter to encourage mean-zero reward outputs ([see TRL docs](https://huggingface.co/docs/trl/v0.10.1/en/reward_trainer#centering-rewards)).
 
 ```yaml
 base_model: google/gemma-2-2b

docs/streaming.qmd

@@ -0,0 +1,120 @@
+---
+title: Streaming Datasets
+description: How to use streaming mode for large-scale datasets and memory-efficient training
+order: 10
+---
+
+Streaming enables memory-efficient training with large datasets by loading data
+incrementally rather than loading the entire dataset into memory at once.
+
+Use streaming when:
+
+- Your dataset is too large to fit in memory (e.g. when you're doing pretraining with massive text corpora)
+- You want to start training immediately without preprocessing the entire dataset
+
+Streaming works with both remote and locally stored datasets!
+
+::: {.callout-note}
+Streaming currently only supports a single dataset. Multi-dataset support will be added soon.
+:::
+
+
+## Configuration
+
+### Basic Streaming
+
+Enable streaming mode by setting the `streaming` flag:
+
+```yaml
+streaming: true
+```
+
+### Pretraining with Streaming
+
+For pretraining tasks, streaming is automatically enabled when using `pretraining_dataset`:
+
+```yaml
+pretraining_dataset:
+  - path: HuggingFaceFW/fineweb-edu
+    type: pretrain
+    text_column: text
+    split: train
+
+# Optionally, enable sample packing
+streaming_multipack_buffer_size: 10000
+sample_packing: true
+```
+
+### SFT with Streaming
+
+For supervised fine-tuning with streaming:
+
+```yaml
+streaming: true
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+    split: train
+
+# Optionally, enable sample packing
+streaming_multipack_buffer_size: 10000
+sample_packing: true
+```
+
+## Configuration Options
+
+### `streaming_multipack_buffer_size`
+
+Controls the buffer size for multipack streaming (default: 10,000). This determines how
+many samples are buffered before packing. Larger buffers can improve packing efficiency
+but use more memory.
+
+### `shuffle_merged_datasets`
+
+When enabled, shuffles the streaming dataset using the buffer. This requires additional
+memory for the shuffle buffer.
+
+## Sample Packing with Streaming
+
+Sample packing is supported for streaming datasets. When enabled, multiple samples are
+packed into a single sequence to maximize GPU utilization:
+
+```yaml
+sample_packing: true
+streaming_multipack_buffer_size: 10000
+
+# For SFT: attention is automatically isolated between packed samples
+# For pretraining: control with pretrain_multipack_attn
+pretrain_multipack_attn: true  # prevent cross-attention between packed samples
+```
+
+For more information, see our [documentation](multipack.qmd) on multipacking.
+
+## Important Considerations
+
+### Memory Usage
+
+While streaming reduces memory usage compared to loading entire datasets, you still need
+to consider:
+
+- You can control the memory usage by adjusting `streaming_multipack_buffer_size`
+- Sample packing requires buffering multiple samples
+- Shuffling requires additional memory for the shuffle buffer
+
+### Performance
+
+- Streaming may have slightly higher latency compared to preprocessed datasets, as samples are processed on-the-fly
+- Network speed and disk read speed are important when streaming from remote sources or a local dataset, respectively
+- Consider using `axolotl preprocess` for smaller or more frequently used datasets
+
+### Evaluation Datasets
+
+Evaluation datasets are not streamed to ensure consistent evaluation metrics. They're
+loaded normally even when training uses streaming.
+
+## Examples
+
+See the `examples/streaming/` directory for complete configuration examples:
+
+- `pretrain.yaml`: Pretraining with streaming dataset
+- `sft.yaml`: Supervised fine-tuning with streaming

examples/cloud/baseten.yaml

@@ -0,0 +1,10 @@
+provider: baseten
+project_name:
+
+secrets:
+  - HF_TOKEN
+  - WANDB_API_KEY
+
+gpu: h100
+gpu_count: 8
+node_count: 1

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@0ee9ee8\""
+    "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@c6a32c5\""
    ]
   },
   {
@@ -176,8 +176,8 @@
     }
    ],
    "source": [
-    "from axolotl.utils.dict import DictDefault\n",
     "from axolotl.cli.config import load_cfg\n",
+    "from axolotl.utils.dict import DictDefault\n",
     "\n",
     "# Axolotl provides full control and transparency over model and training configuration\n",
     "config = DictDefault(\n",

examples/devstral/README.md

@@ -20,7 +20,13 @@ pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
-2. Run the finetuning example:
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage
+
+```bash
+python scripts/cutcrossentropy_install.py | sh
+```
+
+3. Run the finetuning example:
 
 ```bash
 axolotl train examples/devstral/devstral-small-qlora.yml

examples/gemma3/270m-qlora.yml

@@ -0,0 +1,68 @@
+base_model: google/gemma-3-270m-it
+# optionally might have model_type or tokenizer_type
+model_type: AutoModelForCausalLM
+tokenizer_type: AutoTokenizer
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# gemma3 doesn't seem to play nice with ddp
+ddp_find_unused_parameters: true
+
+load_in_8bit: false
+load_in_4bit: true
+
+# huggingface repo
+chat_template: gemma3
+eot_tokens:
+  - <end_of_turn>
+datasets:
+  - path: cgato/SlimOrcaDedupCleaned
+    type: chat_template
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+adapter: qlora
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+sequence_len: 2048
+sample_packing: true
+eval_sample_packing: 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:
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

examples/gpt-oss/README.md

@@ -106,6 +106,16 @@ See [Nanobit/text-tools-2k-test](https://huggingface.co/datasets/Nanobit/text-to
 
 Refer to [our docs](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#using-tool-use) for more info.
 
+### Thinking and chat_template masking conflict
+
+OpenAI’s Harmony template hides `thinking` in all non-final turns, which conflicts with Axolotl’s `chat_template` masking.
+
+If your dataset has `thinking` content mid-turn, there are two paths we recommend:
+
+- Train only on the last turn. This can be accomplished via chat_template's [train on last doc](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#training-on-last-message).
+
+- Adjust your dataset to only have `thinking` content in the last turn.
+
 ### TIPS
 
 - Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).

examples/hunyuan/README.md

@@ -0,0 +1,85 @@
+# Finetune HunYuan with Axolotl
+
+Tencent released a family of opensource models called HunYuan with varying parameter scales of 0.5B, 1.8B, 4B, and 7B scale for both Pre-trained and Instruct variants. The models can be found at [HuggingFace](https://huggingface.co/collections/tencent/hunyuan-dense-model-6890632cda26b19119c9c5e7). This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as HunYuan is only on nightly or use our latest [Docker images](https://docs.axolotl.ai/docs/docker.html).
+
+    Here is an example of how to install from main for pip:
+
+```bash
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
+git clone https://github.com/axolotl-ai-cloud/axolotl.git
+cd axolotl
+
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation -e '.[flash-attn]'
+
+# Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
+python scripts/cutcrossentropy_install.py | sh
+```
+
+2. Run the finetuning example:
+
+```bash
+axolotl train examples/hunyuan/hunyuan-v1-dense-qlora.yaml
+```
+
+This config uses about 4.7 GB VRAM.
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### Dataset
+
+HunYuan Instruct models can choose to enter a slow think or fast think pattern. For best performance on fine-tuning their Instruct models, your dataset should be adjusted to match their pattern.
+
+```python
+# fast think pattern
+messages = [
+    {"role": "system", "content": "You are a helpful assistant."},
+    {"role": "user", "content": "/no_think What color is the sun?" },
+    {"role": "assistant", "content": "<think>\n\n</think>\n<answer>\nThe sun is yellow.\n</answer>"}
+]
+
+# slow think pattern
+messages = [
+    {"role": "system", "content": "You are a helpful assistant."},
+    {"role": "user", "content": "/no_think What color is the sun?" },
+    {"role": "assistant", "content": "<think>\nThe user is asking about the color of the sun. I need to ...\n</think>\n<answer>\nThe sun is yellow.\n</answer>"}
+]
+```
+
+### TIPS
+
+- For inference, the official Tencent team recommends
+
+```json
+
+{
+  "do_sample": true,
+  "top_k": 20,
+  "top_p": 0.8,
+  "repetition_penalty": 1.05,
+  "temperature": 0.7
+}
+
+```
+
+- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+
+## Optimization Guides
+
+- [Multi-GPU Training](https://docs.axolotl.ai/docs/multi-gpu.html)
+- [Multi-Node Training](https://docs.axolotl.ai/docs/multi-node.html)
+- [LoRA Optimizations](https://docs.axolotl.ai/docs/lora_optims.html)
+
+## Related Resources
+
+- [Tencent HunYuan Blog](https://hunyuan.tencent.com/)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl Website](https://axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/hunyuan/hunyuan-v1-dense-qlora.yaml

@@ -0,0 +1,64 @@
+base_model: tencent/Hunyuan-0.5B-Instruct
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+lora_target_modules:
+  - gate_proj
+  - down_proj
+  - up_proj
+  - q_proj
+  - v_proj
+  - k_proj
+  - o_proj
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/llama-3/3b-qat-fsdp2-nvfp4.yaml

@@ -0,0 +1,64 @@
+base_model: meta-llama/Llama-3.2-3B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: false
+load_in_4bit: false
+strict: false
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+liger_layer_norm: true
+liger_fused_linear_cross_entropy: true
+
+datasets:
+  - path: yahma/alpaca-cleaned
+    type: alpaca
+    split: train[:95%]
+
+output_dir: ./outputs/qat_out/
+dataset_prepared_path: ./outputs/dataset_prepared
+
+sequence_len: 8192
+flash_attention: true
+
+qat:
+  activation_dtype: nvfp4
+  weight_dtype: nvfp4
+  group_size: 16 # only group_size of 16 is supported with nvfp4
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_checkpointing: true
+gradient_accumulation_steps: 1
+micro_batch_size: 64
+num_epochs: 1
+optimizer: adamw_torch_fused
+
+cosine_constant_lr_ratio: 0
+cosine_min_lr_ratio: 1.0
+learning_rate: 2e-5
+save_only_model: true
+bf16: true
+
+resume_from_checkpoint:
+logging_steps: 1
+
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+warmup_ratio: 0.1
+weight_decay: 0.0
+
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/llama-3/3b-qat-fsdp2.yaml

@@ -15,20 +15,18 @@ liger_glu_activation: true
 liger_layer_norm: true
 liger_fused_linear_cross_entropy: true
 
+
 datasets:
   - path: yahma/alpaca-cleaned
     type: alpaca
+    split: train[:95%]
 
 output_dir: ./outputs/qat_out/
+dataset_prepared_path: ./outputs/qat_out/dataset_prepared
 
-sample_packing: true
-
-sequence_len: 512
-
-flex_attention: true
-flex_attn_compile_kwargs:
-  dynamic: false
-  mode: max-autotune-no-cudagraphs
+sample_packing: false
+sequence_len: 8192
+flash_attention: true
 
 qat:
   activation_dtype: int8
@@ -67,7 +65,7 @@ fsdp:
 fsdp_config:
   fsdp_version: 2
   fsdp_offload_params: false
-  fsdp_cpu_ram_efficient_loading: true
+  fsdp_cpu_ram_efficient_loading: false
   fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
   fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
   fsdp_state_dict_type: FULL_STATE_DICT
@@ -76,6 +74,6 @@ fsdp_config:
   fsdp_activation_checkpointing: true
 
 special_tokens:
-  pad_token: <|end_of_text|>
+  pad_token: <|finetune_right_pad_id|>
 
 # save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/llama-3/diffusion/pretrain-1b.yaml

@@ -0,0 +1,56 @@
+base_model: meta-llama/Llama-3.2-1B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+pretraining_dataset:
+  - path: wikitext
+    name: wikitext-103-raw-v1
+    type: completion
+    field: text
+
+plugins:
+  - axolotl.integrations.diffusion.DiffusionPlugin
+
+diffusion:
+  noise_schedule: cosine
+  min_mask_ratio: 0.15
+  max_mask_ratio: 0.85
+  num_diffusion_steps: 128
+  eps: 5e-4
+  importance_weighting: true
+  mask_token_id: 128002
+  generate_samples: true
+  generation_interval: 250
+
+output_dir: ./outputs/model-out
+
+sequence_len: 512
+sample_packing: true
+
+gradient_accumulation_steps: 8
+micro_batch_size: 4
+max_steps: 10000
+warmup_ratio: 0.1
+
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 3e-4
+sdp_attention: true
+
+bf16: auto
+tf32: true
+
+logging_steps: 1
+save_strategy: steps
+save_steps: 1000
+
+special_tokens:
+  pad_token: "<|end_of_text|>"
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/llama-3/diffusion/sft-1b.yaml

@@ -0,0 +1,59 @@
+base_model: meta-llama/Llama-3.2-1B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+datasets:
+  - path: teknium/GPT4-LLM-Cleaned
+    type: alpaca
+val_set_size: 0.05
+
+plugins:
+  - axolotl.integrations.diffusion.DiffusionPlugin
+
+diffusion:
+  noise_schedule: cosine
+  min_mask_ratio: 0.1
+  max_mask_ratio: 0.9
+  num_diffusion_steps: 128
+  eps: 1e-3
+  importance_weighting: true
+  mask_token_id: 128002
+  generate_samples: true
+  generation_interval: 250
+
+output_dir: ./outputs/model-out
+
+sequence_len: 512
+sample_packing: true
+eval_sample_packing: true
+
+gradient_accumulation_steps: 4
+micro_batch_size: 4
+num_epochs: 1
+warmup_steps: 0.1
+
+optimizer: adamw_8bit
+lr_scheduler: cosine
+learning_rate: 1e-5
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+sdp_attention: true
+
+logging_steps: 1
+save_strategy: best
+eval_strategy: epoch
+
+special_tokens:
+  pad_token: "<|end_of_text|>"
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/magistral/README.md

@@ -18,7 +18,13 @@ pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
 pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
-2. Run the finetuning example:
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage
+
+```bash
+python scripts/cutcrossentropy_install.py | sh
+```
+
+3. Run the finetuning example:
 
 ```bash
 axolotl train examples/magistral/magistral-small-qlora.yaml

examples/qwen3/reward-model.yaml

@@ -0,0 +1,44 @@
+base_model: Skywork/Skywork-Reward-V2-Qwen3-8B
+model_type: AutoModelForSequenceClassification
+num_labels: 1
+
+reward_model: true
+center_rewards_coefficient: 0.01  # Incentivize mean-zero rewards for improved stability
+chat_template: qwen3
+datasets:
+  - path: argilla/distilabel-intel-orca-dpo-pairs
+    type: bradley_terry.chat_template
+
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 8192
+sample_packing: false
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+deepspeed: deepspeed_configs/zero1.json
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+eval_batch_size: 1
+num_epochs: 3
+optimizer: adamw_bnb_8bit
+lr_scheduler: linear
+learning_rate: 0.00002
+
+bf16: true
+tf32: true
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+warmup_ratio: 0.1
+logging_steps: 1
+weight_decay: 0.01

examples/seed-oss/README.md

@@ -0,0 +1,54 @@
+# Finetune ByteDance's Seed-OSS with Axolotl
+
+[Seed-OSS](https://huggingface.co/collections/ByteDance-Seed/seed-oss-68a609f4201e788db05b5dcd) are a series of 36B parameter open source models trained by ByteDance's Seed Team.
+
+This guide shows how to fine-tune it with Axolotl with multi-turn conversations and proper masking.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html). You need to install from main as Seed-OSS is only on nightly or use our latest [Docker images](https://docs.axolotl.ai/docs/docker.html).
+
+    Here is an example of how to install from main for pip:
+
+```bash
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
+git clone https://github.com/axolotl-ai-cloud/axolotl.git
+cd axolotl
+
+pip3 install packaging==23.2 setuptools==75.8.0 wheel ninja
+pip3 install --no-build-isolation -e '.[flash-attn]'
+
+# Install Cut Cross Entropy
+python scripts/cutcrossentropy_install.py | sh
+```
+
+2. Run the finetuning example:
+
+```bash
+axolotl train examples/seed-oss/seed-oss-36b-qlora.yaml
+```
+
+This config uses about 27.7 GiB VRAM.
+
+Let us know how it goes. Happy finetuning! 🚀
+
+### TIPS
+
+- For inference, the official Seed Team recommends `top_p=0.95` and `temperature=1.1`.
+- You can run a full finetuning by removing the `adapter: qlora` and `load_in_4bit: true` from the config.
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+
+## Optimization Guides
+
+- [Multi-GPU Training](https://docs.axolotl.ai/docs/multi-gpu.html)
+- [Multi-Node Training](https://docs.axolotl.ai/docs/multi-node.html)
+- [LoRA Optimizations](https://docs.axolotl.ai/docs/lora_optims.html)
+
+## Related Resources
+
+- [ByteDance Seed Website](https://seed.bytedance.com/)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl Website](https://axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/seed-oss/seed-oss-36b-qlora.yaml

@@ -0,0 +1,56 @@
+base_model: ByteDance-Seed/Seed-OSS-36B-Instruct
+
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_8bit: false
+load_in_4bit: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir: ./outputs/lora-out
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/streaming/README.md

@@ -0,0 +1,50 @@
+# Streaming Dataset Examples
+
+This directory contains example configurations for using Axolotl's streaming dataset
+functionality, which enables memory-efficient training with large datasets.
+
+## Examples
+
+Run the following examples with e.g. `axolotl train examples/streaming/sft.yaml`; no
+`axolotl preprocess` required!
+
+### Pretraining (`pretrain.yaml`)
+
+Demonstrates streaming configuration for pretraining tasks using the fineweb-edu dataset
+with SmolLM2-135M.
+
+- Uses `pretraining_dataset` configuration for automatic streaming
+- Multipack attention control to prevent cross-attention between packed sequences
+- Buffer size configuration for memory management
+
+### SFT (`sft.yaml`)
+
+Shows how to use streaming for supervised fine-tuning with the Alpaca dataset.
+
+- Explicit `streaming: true` flag for SFT datasets
+- Memory-efficient training on instruction datasets
+- Evaluation datasets are currently not streamed
+
+## Key Configuration Options
+
+### `streaming`
+- Enables streaming mode for standard datasets
+- Automatically enabled for `pretraining_dataset`
+
+### `streaming_multipack_buffer_size`
+- Controls buffer size for sample packing (default: 10,000)
+- Larger values improve packing efficiency but use more memory
+- Adjust based on available memory
+
+### `shuffle_merged_datasets`
+- Enables shuffling of streaming datasets
+- Requires additional memory for shuffle buffer
+
+### `sample_packing`
+- Packs multiple samples into single sequences
+- Minimize per-step padding tokens
+
+## Performance Tips
+
+- Download small / frequently-used datasets locally for better performance
+- Larger buffer sizes improve packing efficiency

examples/streaming/pretrain.yaml

@@ -0,0 +1,57 @@
+base_model: HuggingFaceTB/SmolLM2-135M
+
+# Streaming pretraining configuration
+pretraining_dataset:
+  - path: HuggingFaceFW/fineweb-edu
+    name: sample-10BT
+    type: pretrain
+    text_column: text
+    split: train
+
+# Streaming-specific settings
+streaming_multipack_buffer_size: 10000
+shuffle_merged_datasets: true
+
+# Training configuration
+max_steps: 1000
+output_dir: ./outputs/smollm2-135m-pretrain-streaming
+
+# Sequence and packing settings
+sequence_len: 1024
+sample_packing: true
+pretrain_multipack_attn: true  # Prevent cross-attention between packed sequences
+flash_attention: true
+
+# Batch size settings
+gradient_accumulation_steps: 8
+micro_batch_size: 1
+
+# Optimizer and scheduler
+optimizer: adamw_torch
+lr_scheduler: cosine
+learning_rate: 5e-4
+warmup_ratio: 0.1
+weight_decay: 0.01
+
+# Precision and performance
+bf16: auto
+tf32: true
+
+# Logging and checkpointing
+logging_steps: 10
+save_strategy: steps
+save_steps: 250
+save_total_limit: 3
+
+# Weights & Biases (optional)
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+# Special tokens
+special_tokens:
+  pad_token: "<|endoftext|>"
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/streaming/sft.yaml

@@ -0,0 +1,55 @@
+base_model: HuggingFaceTB/SmolLM2-135M
+
+# Dataset configuration
+datasets:
+  - path: tatsu-lab/alpaca
+    type: alpaca
+    split: train
+
+# Streaming-specific settings
+streaming: true
+streaming_multipack_buffer_size: 10000
+shuffle_merged_datasets: true
+
+# Training configuration
+max_steps: 1000
+output_dir: ./outputs/smollm2-135m-sft-streaming
+
+# Sequence and packing settings
+sequence_len: 1024
+sample_packing: true
+flash_attention: true
+
+# Batch size settings
+gradient_accumulation_steps: 4
+micro_batch_size: 1
+
+# Optimizer and scheduler
+optimizer: adamw_torch
+lr_scheduler: cosine
+learning_rate: 2e-4
+warmup_ratio: 0.1
+weight_decay: 0.0
+
+# Precision and performance
+bf16: auto
+tf32: true
+
+# Logging and checkpointing
+logging_steps: 10
+save_strategy: steps
+save_steps: 100
+save_total_limit: 3
+
+# Weights & Biases (optional)
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+# Special tokens
+special_tokens:
+  pad_token: "<|endoftext|>"
+
+# save_first_step: true  # uncomment this to validate checkpoint saving works with your config

examples/voxtral/README.md

@@ -22,6 +22,9 @@ pip3 install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 # audio
 pip3 install librosa==0.11.0
 pip3 install 'mistral_common[audio]==1.8.3'
+
+# Install CCE https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
+python scripts/cutcrossentropy_install.py | sh
 ```
 
 3. Run the finetuning example:

requirements.txt

@@ -2,8 +2,7 @@
 
 # START section of dependencies that don't install on Darwin/MacOS
 bitsandbytes==0.47.0
-# triton 3.4.0 is not compatible with CCE
-triton>=3.0.0,<3.4.0
+triton>=3.0.0
 mamba-ssm==1.2.0.post1
 xformers>=0.0.23.post1
 autoawq==0.2.7.post3
@@ -14,7 +13,7 @@ packaging==23.2
 
 huggingface_hub>=0.33.0
 peft>=0.17.0
-transformers==4.55.3
+transformers==4.56.1
 tokenizers>=0.21.1
 accelerate==1.10.0
 datasets==4.0.0
@@ -65,7 +64,7 @@ langdetect==1.0.9
 immutabledict==4.2.0
 antlr4-python3-runtime==4.13.2
 
-torchao==0.12.0
+torchao==0.13.0
 schedulefree==1.4.1
 
 axolotl-contribs-lgpl==0.0.6

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@0ee9ee8"'
+    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@c6a32c5"'
 )

setup.py

@@ -64,7 +64,9 @@ def parse_requirements(extras_require_map):
             else:
                 raise ValueError("Invalid version format")
 
-            if (major, minor) >= (2, 7):
+            if (major, minor) >= (2, 8):
+                pass
+            elif (major, minor) >= (2, 7):
                 _install_requires.pop(_install_requires.index(xformers_version))
                 if patch == 0:
                     _install_requires.append("xformers==0.0.30")
@@ -125,7 +127,7 @@ extras_require = {
         "yunchang==0.6.0",
     ],
     "deepspeed": [
-        "deepspeed==0.17.2",
+        "deepspeed==0.17.5",
         "deepspeed-kernels",
     ],
     "mamba-ssm": [
@@ -160,6 +162,7 @@ extras_require = {
     "llmcompressor": [
         "llmcompressor==0.5.1",
     ],
+    "fbgemm-gpu": ["fbgemm-gpu-genai>=1.2.0"],
 }
 install_requires, dependency_links, extras_require_build = parse_requirements(
     extras_require

src/axolotl/cli/args.py

@@ -14,9 +14,13 @@ class PreprocessCliArgs:
     prompter: Optional[str] = field(default=None)
     download: Optional[bool] = field(default=True)
     iterable: Optional[bool] = field(
-        default=None,
+        default=False,
         metadata={
-            "help": "Use IterableDataset for streaming processing of large datasets"
+            "help": (
+                "Deprecated in v0.13.0, will be removed in v0.14.0. For streaming "
+                "datasets, use 'axolotl train' and set 'streaming: true' in your YAML "
+                "config, or pass --streaming instead in the CLI."
+            )
         },
     )
 
@@ -111,6 +115,7 @@ class QuantizeCliArgs:
     quantize_embedding: Optional[bool] = field(default=None)
     group_size: Optional[int] = field(default=None)
     output_dir: Optional[str] = field(default=None)
+    hub_model_id: Optional[str] = field(default=None)
 
 
 @dataclass

src/axolotl/cli/cloud/__init__.py

@@ -7,6 +7,8 @@ from typing import Literal
 
 import yaml
 
+from axolotl.cli.cloud.base import Cloud
+from axolotl.cli.cloud.baseten import BasetenCloud
 from axolotl.cli.cloud.modal_ import ModalCloud
 from axolotl.utils.dict import DictDefault
 
@@ -38,8 +40,15 @@ def do_cli_train(
     cwd=None,
     **kwargs,
 ) -> None:
-    cloud_cfg = load_cloud_cfg(cloud_config)
-    cloud = ModalCloud(cloud_cfg)
+    cloud_cfg: DictDefault = load_cloud_cfg(cloud_config)
+    provider = cloud_cfg.provider or "modal"
+    cloud: Cloud | None
+    if provider == "modal":
+        cloud = ModalCloud(cloud_cfg)
+    elif provider == "baseten":
+        cloud = BasetenCloud(cloud_cfg.to_dict())
+    else:
+        raise ValueError(f"Unsupported cloud provider: {provider}")
     with open(config, "r", encoding="utf-8") as file:
         config_yaml = file.read()
     local_dirs = {}

src/axolotl/cli/cloud/baseten/__init__.py

@@ -0,0 +1,48 @@
+"""Baseten Cloud CLI"""
+
+import shutil
+import subprocess  # nosec B404
+import tempfile
+from os.path import dirname
+from typing import Literal
+
+import yaml
+
+from axolotl.cli.cloud.base import Cloud
+
+
+class BasetenCloud(Cloud):
+    """Baseten Cloud Axolotl CLI"""
+
+    def __init__(self, config: dict):
+        self.config = config
+
+    def preprocess(self, config_yaml: str, *args, **kwargs) -> None:
+        raise NotImplementedError(
+            "Separate preprocess function for Baseten is not "
+            "implemented and will happen during hte train step."
+        )
+
+    def train(
+        self,
+        config_yaml: str,
+        launcher: Literal["accelerate", "torchrun", "python"] = "accelerate",
+        launcher_args: list[str] | None = None,
+        local_dirs: dict[str, str] | None = None,  # pylint: disable=unused-argument
+        **kwargs,
+    ):
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            config = self.config.copy()
+            config["launcher"] = launcher
+            config["launcher_args"] = launcher_args
+            with open(tmp_dir + "/cloud.yaml", "w", encoding="utf-8") as cloud_fout:
+                yaml.dump(config, cloud_fout)
+            with open(tmp_dir + "/train.yaml", "w", encoding="utf-8") as config_fout:
+                config_fout.write(config_yaml)
+            shutil.copyfile(dirname(__file__) + "/template/run.sh", tmp_dir + "/run.sh")
+            shutil.copyfile(
+                dirname(__file__) + "/template/train_sft.py", tmp_dir + "/train_sft.py"
+            )
+            subprocess.run(  # nosec B603 B607
+                ["truss", "train", "push", "train_sft.py"], cwd=tmp_dir, check=False
+            )

src/axolotl/cli/cloud/baseten/template/run.sh

@@ -0,0 +1,9 @@
+#!/bin/bash
+set -eux
+
+export NCCL_SOCKET_IFNAME="^docker0,lo"
+export NCCL_IB_DISABLE=0
+export NCCL_TIMEOUT=1800000
+
+axolotl preprocess train.yaml
+axolotl train train.yaml --launcher ${AXOLOTL_LAUNCHER} ${AXOLOTL_LAUNCHER_ARGS}

src/axolotl/cli/cloud/baseten/template/train_sft.py

@@ -0,0 +1,71 @@
+"""
+Baseten Training Script for Axolotl
+"""
+
+# pylint: skip-file
+import yaml
+from truss.base import truss_config
+
+# Import necessary classes from the Baseten Training SDK
+from truss_train import definitions
+
+cloud_config = yaml.safe_load(open("cloud.yaml", "r"))
+gpu = cloud_config.get("gpu", "h100")
+gpu_count = int(cloud_config.get("gpu_count", 1))
+node_count = int(cloud_config.get("node_count", 1))
+project_name = cloud_config.get("project_name", "axolotl-project") or "axolotl-project"
+secrets = cloud_config.get("secrets", [])
+launcher = cloud_config.get("launcher", "accelerate")
+launcher_args = cloud_config.get("launcher_args", [])
+script_name = "run.sh"
+
+launcher_args_str = ""
+if launcher_args:
+    launcher_args_str = "-- " + " ".join(launcher_args)
+
+# 1. Define a base image for your training job
+# must use torch 2.7.0 for vllm
+BASE_IMAGE = "axolotlai/axolotl:main-py3.11-cu126-2.7.1"
+
+# 2. Define the Runtime Environment for the Training Job
+# This includes start commands and environment variables.a
+# Secrets from the baseten workspace like API keys are referenced using
+# `SecretReference`.
+
+env_vars = {
+    "AXOLOTL_LAUNCHER": launcher,
+    "AXOLOTL_LAUNCHER_ARGS": launcher_args_str,
+}
+for secret_name in secrets:
+    env_vars[secret_name] = definitions.SecretReference(name=secret_name)
+
+training_runtime = definitions.Runtime(
+    start_commands=[  # Example: list of commands to run your training script
+        f"/bin/sh -c 'chmod +x ./{script_name} && ./{script_name}'"
+    ],
+    environment_variables=env_vars,
+)
+
+# 3. Define the Compute Resources for the Training Job
+training_compute = definitions.Compute(
+    node_count=node_count,
+    accelerator=truss_config.AcceleratorSpec(
+        accelerator=truss_config.Accelerator.H100,
+        count=gpu_count,
+    ),
+)
+
+# 4. Define the Training Job
+# This brings together the image, compute, and runtime configurations.
+my_training_job = definitions.TrainingJob(
+    image=definitions.Image(base_image=BASE_IMAGE),
+    compute=training_compute,
+    runtime=training_runtime,
+)
+
+
+# This config will be pushed using the Truss CLI.
+# The association of the job to the project happens at the time of push.
+first_project_with_job = definitions.TrainingProject(
+    name=project_name, job=my_training_job
+)

src/axolotl/cli/inference.py

@@ -14,10 +14,14 @@ from transformers import GenerationConfig, TextIteratorStreamer, TextStreamer
 from axolotl.cli.args import InferenceCliArgs
 from axolotl.cli.config import load_cfg
 from axolotl.cli.utils import load_model_and_tokenizer
-from axolotl.utils.chat_templates import (
-    get_chat_template,
-    get_chat_template_from_config,
+from axolotl.cli.utils.diffusion import (
+    diffusion_inference,
+    launch_diffusion_gradio_ui,
+    render_html,
+    run_diffusion,
 )
+from axolotl.integrations.base import PluginManager
+from axolotl.utils.chat_templates import get_chat_template_from_config
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.logging import get_logger
 
@@ -32,6 +36,7 @@ def get_multi_line_input() -> str:
         Possibly multi-line, possibly empty stdin input as a string.
     """
     print("Give me an instruction (Ctrl + D to submit): ")
+    print("=" * 80)
 
     instruction = ""
     for line in sys.stdin:
@@ -46,9 +51,9 @@ def do_inference(
     cli_args: InferenceCliArgs,
 ):
     """
-    Runs inference on the command line in a loop. User input is accepted, a chat template
-    is (optionally) applied, and the model specified in the `axolotl` config is used to
-    generate completions according to a default generation config.
+    Runs inference on the command line in a loop. User input is accepted, a chat
+    template is (optionally) applied, and the model specified in the `axolotl` config is
+    used to generate completions according to a default generation config.
 
     Args:
         cfg: Dictionary mapping `axolotl` config keys to values.
@@ -64,17 +69,31 @@ def do_inference(
             importlib.import_module("axolotl.prompters"), prompter
         )
     elif cfg.chat_template:
-        chat_template_str = get_chat_template(cfg.chat_template, tokenizer=tokenizer)
-    elif cfg.datasets[0].type == "chat_template":
+        chat_template_str = get_chat_template_from_config(
+            cfg, ds_cfg=None, tokenizer=tokenizer
+        )
+    elif cfg.datasets and cfg.datasets[0].type == "chat_template":
         chat_template_str = get_chat_template_from_config(
             cfg=cfg, ds_cfg=cfg.datasets[0], tokenizer=tokenizer
         )
 
     model = model.to(cfg.device, dtype=cfg.torch_dtype)
 
+    # Detect diffusion mode
+    plugin_manager = PluginManager.get_instance()
+    is_diffusion = any(
+        plugin.__class__.__name__ == "DiffusionPlugin"
+        for plugin in plugin_manager.plugins.values()
+    )
+
+    if is_diffusion:
+        print("=" * 80)
+        print("Commands:")
+        print(":complete N -> completion mode with N tokens (default 64)")
+        print(":mask R     -> random masking with ratio R (0.0–1.0)")
+
     while True:
         print("=" * 80)
-        # support for multiline inputs
         instruction = get_multi_line_input()
         if not instruction:
             return
@@ -104,9 +123,19 @@ def do_inference(
         else:
             batch = tokenizer(prompt, return_tensors="pt", add_special_tokens=True)
 
-        print("=" * 40)
+        print("=" * 80)
         model.eval()
         with torch.no_grad():
+            if is_diffusion:
+                diffusion_inference(
+                    model=model,
+                    tokenizer=tokenizer,
+                    cfg=cfg,
+                    prompt=prompt,
+                    chat_template_str=chat_template_str,
+                )
+                continue
+
             generation_config = GenerationConfig(
                 repetition_penalty=1.1,
                 max_new_tokens=1024,
@@ -129,7 +158,7 @@ def do_inference(
                 generation_config=generation_config,
                 streamer=streamer,
             )
-        print("=" * 40)
+        print("=" * 80)
         print(tokenizer.decode(generated["sequences"].cpu().tolist()[0]))
 
 
@@ -159,10 +188,33 @@ def do_inference_gradio(
             importlib.import_module("axolotl.prompters"), prompter
         )
     elif cfg.chat_template:
-        chat_template_str = get_chat_template(cfg.chat_template, tokenizer=tokenizer)
+        chat_template_str = get_chat_template_from_config(
+            cfg, ds_cfg=None, tokenizer=tokenizer
+        )
+    elif cfg.datasets and cfg.datasets[0].type == "chat_template":
+        chat_template_str = get_chat_template_from_config(
+            cfg=cfg, ds_cfg=cfg.datasets[0], tokenizer=tokenizer
+        )
 
     model = model.to(cfg.device, dtype=cfg.torch_dtype)
 
+    # Detect diffusion mode
+    plugin_manager = PluginManager.get_instance()
+    is_diffusion = any(
+        plugin.__class__.__name__ == "DiffusionPlugin"
+        for plugin in plugin_manager.plugins.values()
+    )
+
+    if is_diffusion:
+        launch_diffusion_gradio_ui(
+            model=model,
+            tokenizer=tokenizer,
+            cfg=cfg,
+            prompter_module=prompter_module,
+            chat_template_str=chat_template_str,
+        )
+        return
+
     def generate(instruction):
         if not instruction:
             return

src/axolotl/cli/merge_lora.py

@@ -43,7 +43,10 @@ def do_merge_lora(*, cfg: DictDefault) -> None:
             safe_serialization=safe_serialization,
             progressbar=True,
         )
-        tokenizer.save_pretrained(str(Path(cfg.output_dir) / "merged"))
+        tokenizer.save_pretrained(
+            str(Path(cfg.output_dir) / "merged"),
+            save_jinja_files=cfg.tokenizer_save_jinja_files,
+        )
 
         if processor:
             processor.save_pretrained(str(Path(cfg.output_dir) / "merged"))

src/axolotl/cli/preprocess.py

@@ -35,10 +35,20 @@ def do_preprocess(cfg: DictDefault, cli_args: PreprocessCliArgs) -> None:
     check_accelerate_default_config()
     check_user_token()
 
+    if cli_args.iterable:
+        LOG.error(
+            "The --iterable CLI argument for 'axolotl preprocess' is no longer "
+            "supported. For training, set 'streaming: true' in your YAML config or "
+            "pass '--streaming' in your 'axolotl train' command for on-the-fly "
+            "preprocessing."
+        )
+        return
+
     for key in ["skip_prepare_dataset", "pretraining_dataset"]:
         if cfg.get(key):
             LOG.error(
-                f"You have set `{key}:`. `preprocess` is not needed. Run the `axolotl train` CLI directly instead."
+                f"You have set `{key}:`. `preprocess` is not needed. Run the 'axolotl "
+                "train' CLI directly instead."
             )
             return
 

src/axolotl/cli/quantize.py

@@ -5,12 +5,17 @@ CLI to post-training quantize a model using torchao
 from pathlib import Path
 from typing import Union
 
-from transformers import AutoModelForCausalLM
+from transformers import AutoConfig, AutoModelForCausalLM, TorchAoConfig
 
 from axolotl.cli.config import load_cfg
 from axolotl.loaders import load_tokenizer
 from axolotl.utils.logging import get_logger
-from axolotl.utils.quantization import TorchIntDType, quantize_model_for_ptq
+from axolotl.utils.quantization import (
+    TorchAOQuantDType,
+    get_quantization_config,
+    quantization_config_to_str,
+    quantize_model,
+)
 
 LOG = get_logger(__name__)
 
@@ -43,13 +48,13 @@ def do_quantize(
             "No quantization configuration found. Please specify either qat or quantization in your config file."
         )
 
-    model_path = cli_args.get("model_path") or cfg.output_dir
+    model_path = cli_args.get("base_model") or cfg.output_dir
     if weight_dtype := cli_args.get("weight_dtype"):
-        weight_dtype = TorchIntDType[weight_dtype]
+        weight_dtype = TorchAOQuantDType.from_string(weight_dtype)
     else:
         weight_dtype = quantize_cfg.weight_dtype
     if activation_dtype := cli_args.get("activation_dtype"):
-        activation_dtype = TorchIntDType[activation_dtype]
+        activation_dtype = TorchAOQuantDType.from_string(activation_dtype)
     else:
         activation_dtype = quantize_cfg.activation_dtype
     group_size = cli_args.get("group_size") or quantize_cfg.group_size
@@ -57,10 +62,15 @@ def do_quantize(
         cli_args.get("quantize_embedding") or quantize_cfg.quantize_embedding
     )
     output_dir = cli_args.get("output_dir") or cfg.output_dir
+    hub_model_id = cli_args.get("hub_model_id") or cfg.hub_model_id
 
-    LOG.info(f"Loading model from {model_path}...")
+    LOG.info(f"Loading model from {model_path}.")
     tokenizer = load_tokenizer(cfg)
-    model = AutoModelForCausalLM.from_pretrained(model_path, device_map="auto")
+    config = AutoConfig.from_pretrained(model_path)
+    torch_dtype = config.torch_dtype if hasattr(config, "torch_dtype") else None
+    model = AutoModelForCausalLM.from_pretrained(
+        model_path, device_map="auto", torch_dtype=torch_dtype
+    )
 
     LOG.info(
         f"Quantizing model with configuration: \n"
@@ -70,11 +80,21 @@ def do_quantize(
         f"\tquantize_embedding: {quantize_embedding}"
     )
 
-    quantize_model_for_ptq(
+    quantize_model(
         model, weight_dtype, group_size, activation_dtype, quantize_embedding
     )
 
-    LOG.info(f"Saving quantized model to: {str(Path(output_dir) / 'quantized')}...")
+    quantization_config = get_quantization_config(
+        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"),
         safe_serialization=False,
@@ -84,5 +104,16 @@ def do_quantize(
         str(Path(output_dir) / "quantized"),
         safe_serialization=False,
         progressbar=True,
+        save_jinja_files=cfg.tokenizer_save_jinja_files,
     )
-    LOG.info(f"Quantized model saved to: {str(Path(output_dir) / 'quantized')}...")
+
+    if hub_model_id:
+        hub_model_id = (
+            hub_model_id.rstrip("-")
+            + f"-{quantization_config_to_str[type(quantization_config)]}"
+        )
+        model.push_to_hub(hub_model_id, safe_serialization=False)
+        tokenizer.push_to_hub(hub_model_id)
+        LOG.info(f"Quantized model pushed to: {hub_model_id}.")
+
+    LOG.info(f"Quantized model saved to: {str(Path(output_dir) / 'quantized')}.")

src/axolotl/cli/utils/diffusion.py

@@ -0,0 +1,375 @@
+"""Helpers for diffusion-mode inference in CLI and Gradio."""
+
+from __future__ import annotations
+
+import gradio as gr
+import torch
+from colorama import Fore, Style
+
+from axolotl.integrations.diffusion import generate, resolve_mask_token_id
+from axolotl.utils.dict import DictDefault
+
+
+def diffusion_inference(
+    model,
+    tokenizer,
+    cfg,
+    prompt: str,
+    chat_template_str: str | None = None,
+):
+    """Diffusion inference helper method."""
+    mode = "random"
+    completion_tokens = 0
+    target_mask_ratio = None
+    mode, completion_tokens, target_mask_ratio, cleaned = _parse_commands(prompt)
+
+    if cleaned:
+        prompt = cleaned
+
+    info = run_diffusion(
+        model=model,
+        tokenizer=tokenizer,
+        cfg=cfg,
+        prompt=prompt,
+        chat_template_str=chat_template_str,
+        mode=mode,
+        target_mask_ratio=target_mask_ratio,
+        completion_tokens=completion_tokens,
+    )
+    masked_text = info["masked_text"]
+    mask_ratio = info["mask_ratio"]
+    generated_ids = info["generated_ids"]
+    masked_positions = info["masked_positions"]
+    orig_ids = info["orig_ids"]
+
+    # Display with masked preview and colored diff
+    if masked_text is not None and mask_ratio is not None:
+        print(f"Masked ({mask_ratio:.1%}):\n{masked_text}\n")
+    if generated_ids is not None:
+        # Compute per-token style
+        styles: list[str] = []
+        for i, tid in enumerate(generated_ids):
+            if i in masked_positions:
+                if i < len(orig_ids) and tid == orig_ids[i]:
+                    styles.append("green")  # correct fill
+                elif i < len(orig_ids):
+                    styles.append("red")  # incorrect fill
+                else:
+                    styles.append("normal")  # appended
+            else:
+                same = i < len(orig_ids) and tid == orig_ids[i]
+                styles.append("dim" if same else "normal")
+
+        # Group contiguous spans by style
+        styled_spans: list[tuple[str, int, int]] = []
+        if generated_ids:
+            current_style = styles[0]
+            start = 0
+            for i in range(1, len(generated_ids)):
+                s = styles[i]
+                if s != current_style:
+                    styled_spans.append((current_style, start, i))
+                    current_style, start = s, i
+            styled_spans.append((current_style, start, len(generated_ids)))
+
+        out_parts = []
+        for style_name, a, b in styled_spans:
+            chunk_text = tokenizer.decode(generated_ids[a:b], skip_special_tokens=False)
+            if style_name == "green":
+                out_parts.append(Fore.GREEN + chunk_text + Style.RESET_ALL)
+            elif style_name == "red":
+                out_parts.append(Fore.RED + chunk_text + Style.RESET_ALL)
+            else:
+                if style_name == "dim":
+                    out_parts.append(Style.DIM + chunk_text + Style.RESET_ALL)
+                else:
+                    out_parts.append(chunk_text)
+        print("Generated:\n" + "".join(out_parts))
+    else:
+        print("Generated:\n(no output)")
+
+
+def _parse_commands(text: str):
+    """
+    Parse leading diffusion commands.
+
+    Supported at start of input (can be chained):
+      :complete N  -> completion mode with N tokens (default 64)
+      :mask R      -> random masking with ratio R in [0, 1]
+    """
+    tokens = text.strip().split()
+    i = 0
+    mode = "random"
+    completion_tokens = 0
+    target_mask_ratio = None
+    consumed = 0
+    while i < len(tokens) and tokens[i].startswith(":"):
+        cmd = tokens[i]
+        i += 1
+        consumed = i
+        if cmd == ":complete":
+            mode = "completion"
+            if i < len(tokens):
+                try:
+                    completion_tokens = int(tokens[i])
+                    i += 1
+                    consumed = i
+                except Exception:
+                    completion_tokens = 64
+            else:
+                completion_tokens = 64
+        elif cmd == ":mask":
+            mode = "random"
+            if i < len(tokens):
+                try:
+                    target_mask_ratio = float(tokens[i])
+                    i += 1
+                    consumed = i
+                except Exception:
+                    target_mask_ratio = None
+        else:
+            i -= 1
+            consumed = i
+            break
+
+    cleaned = " ".join(tokens[consumed:])
+
+    return mode, completion_tokens, target_mask_ratio, cleaned
+
+
+def run_diffusion(
+    *,
+    model,
+    tokenizer,
+    cfg: DictDefault,
+    prompt: str,
+    chat_template_str: str | None,
+    mode: str = "random",
+    target_mask_ratio: float | None = None,
+    completion_tokens: int = 0,
+):
+    """Run a single diffusion generation and return a structured result dict."""
+    if chat_template_str:
+        batch = tokenizer.apply_chat_template(
+            [{"role": "user", "content": prompt}],
+            return_tensors="pt",
+            add_special_tokens=True,
+            add_generation_prompt=True,
+            chat_template=chat_template_str,
+            tokenize=True,
+            return_dict=True,
+        )
+    else:
+        batch = tokenizer(prompt, return_tensors="pt", add_special_tokens=True)
+
+    mask_token_id = resolve_mask_token_id(tokenizer, cfg, allow_add=False)
+
+    seq = batch["input_ids"].to(cfg.device)
+    gen_mode = "completion" if mode == "completion" else "random"
+    comp_tokens = int(completion_tokens) if gen_mode == "completion" else 0
+
+    result = generate(
+        model,
+        tokenizer,
+        original_sequence=seq[:1],
+        num_diffusion_steps=cfg.diffusion.num_diffusion_steps,
+        temperature=cfg.diffusion.generation_temperature,
+        mask_token_id=int(mask_token_id),
+        mode=gen_mode,  # type: ignore[arg-type]
+        completion_tokens=comp_tokens,
+        target_mask_ratio=target_mask_ratio,
+    )
+
+    masked_text = result.get("masked") if isinstance(result, dict) else None
+    mask_ratio = result.get("mask_ratio") if isinstance(result, dict) else None
+    generated_ids = result.get("generated_ids") if isinstance(result, dict) else None
+    masked_positions = (
+        set(result.get("masked_positions") or []) if isinstance(result, dict) else set()
+    )
+    orig_ids = seq[0].detach().cpu().tolist()
+
+    return {
+        "masked_text": masked_text,
+        "mask_ratio": mask_ratio,
+        "generated_ids": generated_ids,
+        "masked_positions": masked_positions,
+        "orig_ids": orig_ids,
+    }
+
+
+def render_html(
+    *,
+    generated_ids: list[int] | None,
+    orig_ids: list[int],
+    masked_positions: set[int],
+    tokenizer,
+) -> str:
+    """Render HTML visualizing diffusion outputs."""
+    if not generated_ids:
+        return "<pre>Generated:\n(no output)</pre>"
+
+    def _style_for(i: int, tid: int) -> str:
+        if i in masked_positions:
+            if i < len(orig_ids) and tid == orig_ids[i]:
+                return "green"
+            if i < len(orig_ids):
+                return "red"
+            return "normal"
+        same = i < len(orig_ids) and tid == orig_ids[i]
+        return "dim" if same else "normal"
+
+    # Group contiguous spans by style to reduce HTML size
+    spans: list[tuple[str, int, int]] = []
+    if generated_ids:
+        cur = _style_for(0, generated_ids[0])
+        start = 0
+        for i in range(1, len(generated_ids)):
+            s = _style_for(i, generated_ids[i])
+            if s != cur:
+                spans.append((cur, start, i))
+                cur, start = s, i
+        spans.append((cur, start, len(generated_ids)))
+
+    html_parts = []
+    for style_name, a, b in spans:
+        txt = tokenizer.decode(generated_ids[a:b], skip_special_tokens=False)
+        if style_name == "green":
+            html_parts.append(f'<span style="color:#2e7d32">{txt}</span>')
+        elif style_name == "red":
+            html_parts.append(f'<span style="color:#c62828">{txt}</span>')
+        elif style_name == "dim":
+            html_parts.append(f'<span style="opacity:0.6">{txt}</span>')
+        else:
+            html_parts.append(txt)
+
+    legend = (
+        '<div style="font-size:0.9em;margin-bottom:4px">'
+        '<span style="color:#2e7d32">correct</span>, '
+        '<span style="color:#c62828">incorrect</span>, '
+        '<span style="opacity:0.6">unchanged</span>'
+        "</div>"
+    )
+
+    return (
+        legend
+        + '<pre style="white-space:pre-wrap">Generated:\n'
+        + "".join(html_parts)
+        + "</pre>"
+    )
+
+
+def launch_diffusion_gradio_ui(
+    *,
+    model,
+    tokenizer,
+    cfg: DictDefault,
+    prompter_module=None,
+    chat_template_str: str | None = None,
+):
+    """Build and launch a simple Gradio UI for diffusion inference."""
+    with gr.Blocks(
+        title=cfg.get("gradio_title", "Axolotl Diffusion Interface")
+    ) as demo:
+        gr.Markdown(
+            """
+            ## Axolotl Diffusion Inference
+            - Mode "Random" masks tokens at a target ratio and fills them.
+            - Mode "Completion" appends N masked tokens at the end and fills them.
+            """
+        )
+
+        with gr.Row():
+            mode = gr.Radio(
+                choices=["random", "completion"],
+                value="random",
+                label="Mode",
+            )
+            mask_ratio = gr.Slider(
+                minimum=0.0,
+                maximum=1.0,
+                step=0.05,
+                value=0.4,
+                label="Mask ratio (random mode)",
+                interactive=True,
+            )
+            completion_tokens = gr.Number(
+                value=64,
+                precision=0,
+                label="Completion tokens (completion mode)",
+                interactive=True,
+                visible=False,
+            )
+
+        instruction = gr.Textbox(label="Instruction", lines=6)
+        run_btn = gr.Button("Generate")
+
+        masked_preview = gr.Textbox(label="Masked preview", lines=6)
+        html_out = gr.HTML(label="Generated")
+
+        def _toggle_controls(selected_mode: str):
+            return (
+                gr.update(visible=(selected_mode == "random")),
+                gr.update(visible=(selected_mode == "completion")),
+            )
+
+        mode.change(
+            _toggle_controls,
+            inputs=[mode],
+            outputs=[mask_ratio, completion_tokens],
+        )
+
+        def _gen(instruction_text: str, selected_mode: str, mratio: float, ctoks: int):
+            if not instruction_text:
+                return "", "<pre>Generated:\n(no output)</pre>"
+
+            if prompter_module:
+                prompt: str = next(
+                    prompter_module().build_prompt(
+                        instruction=instruction_text.strip("\n")
+                    )
+                )
+            else:
+                prompt = instruction_text.strip()
+
+            info = run_diffusion(
+                model=model,
+                tokenizer=tokenizer,
+                cfg=cfg,
+                prompt=prompt,
+                chat_template_str=chat_template_str,
+                mode=selected_mode,
+                target_mask_ratio=mratio if selected_mode == "random" else None,
+                completion_tokens=int(ctoks) if selected_mode == "completion" else 0,
+            )
+
+            masked_text = info.get("masked_text")
+            mask_ratio_val = info.get("mask_ratio")
+            generated_ids = info.get("generated_ids")
+            masked_positions = info.get("masked_positions") or set()
+            orig_ids = info.get("orig_ids") or []
+
+            preview = (
+                f"Masked ({mask_ratio_val:.1%}):\n{masked_text}"
+                if masked_text is not None and mask_ratio_val is not None
+                else ""
+            )
+            html = render_html(
+                generated_ids=generated_ids,
+                orig_ids=orig_ids,
+                masked_positions=masked_positions,
+                tokenizer=tokenizer,
+            )
+            return preview, html
+
+        run_btn.click(
+            _gen,
+            inputs=[instruction, mode, mask_ratio, completion_tokens],
+            outputs=[masked_preview, html_out],
+        )
+
+        demo.queue().launch(
+            show_api=False,
+            share=cfg.get("gradio_share", True),
+            server_name=cfg.get("gradio_server_name", "127.0.0.1"),
+            server_port=cfg.get("gradio_server_port", None),
+        )

src/axolotl/common/datasets.py

@@ -55,13 +55,11 @@ def load_datasets(
     """
     tokenizer = load_tokenizer(cfg)
     processor = load_processor(cfg, tokenizer=tokenizer) if cfg.processor_type else None
-    preprocess_iterable = getattr(cli_args, "iterable", False)
 
     train_dataset, eval_dataset, total_num_steps, prompters = prepare_datasets(
         cfg,
         tokenizer,
         processor=processor,
-        preprocess_iterable=preprocess_iterable,
     )
 
     if (

src/axolotl/core/builders/base.py

@@ -24,9 +24,7 @@ from pathlib import Path
 from typing import Any
 
 import torch
-from transformers import (
-    TrainerCallback,
-)
+from transformers import TrainerCallback
 from transformers.trainer_pt_utils import AcceleratorConfig
 
 from axolotl.integrations.base import PluginManager
@@ -512,6 +510,7 @@ class TrainerBuilderBase(abc.ABC):
                 self.cfg.eval_batch_size
             )
 
+        training_args_kwargs["include_tkps"] = self.cfg.include_tkps
         training_args_kwargs["max_steps"] = self.cfg.max_steps or total_num_steps or -1
         training_args_kwargs["num_train_epochs"] = self.cfg.num_epochs
 

src/axolotl/core/builders/causal.py

@@ -10,6 +10,7 @@ import transformers
 from transformers import (
     DataCollatorWithFlattening,
     EarlyStoppingCallback,
+    Trainer,
 )
 from trl.trainer.utils import RewardDataCollatorWithPadding
 
@@ -35,6 +36,7 @@ from axolotl.utils.callbacks import (
 )
 from axolotl.utils.callbacks.lisa import lisa_callback_factory
 from axolotl.utils.callbacks.qat import QATCallback
+from axolotl.utils.callbacks.tokens_per_second import TokensPerSecondCallback
 from axolotl.utils.chat_templates import get_chat_template_from_config
 from axolotl.utils.collators import (
     BatchSamplerDataCollatorForSeq2Seq,
@@ -74,6 +76,12 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         if self.cfg.qat:
             callbacks.append(QATCallback(self.cfg.qat))
 
+        if self.cfg.include_tkps:
+            callbacks.append(
+                TokensPerSecondCallback(
+                    self.cfg.tensor_parallel_size, self.cfg.context_parallel_size
+                )
+            )
         return callbacks
 
     def get_post_trainer_create_callbacks(self, trainer):
@@ -340,6 +348,10 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
 
         if self.cfg.reward_model:
             training_args_cls = AxolotlRewardConfig
+            if self.cfg.center_rewards_coefficient is not None:
+                training_arguments_kwargs["center_rewards_coefficient"] = (
+                    self.cfg.center_rewards_coefficient
+                )
         elif self.cfg.process_reward_model:
             training_args_cls = AxolotlPRMConfig
         else:
@@ -383,10 +395,11 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
                 **data_collator_kwargs,
             )
         sig = inspect.signature(trainer_cls)
-        if "processing_class" in sig.parameters:
+        if "processing_class" in sig.parameters or issubclass(trainer_cls, Trainer):
             trainer_kwargs["processing_class"] = self.tokenizer
         elif "tokenizer" in sig.parameters:
             trainer_kwargs["tokenizer"] = self.tokenizer
+
         if (
             trainer_cls not in [AxolotlRewardTrainer, AxolotlPRMTrainer]
             and self.cfg.datasets is not None
@@ -404,6 +417,9 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
             **trainer_kwargs,
         )
         trainer = self.hook_post_create_trainer(trainer)
+        # if the trainer has the `axolotl_cfg` property, set it
+        if hasattr(trainer, "axolotl_cfg"):
+            trainer.axolotl_cfg = self.cfg
         for callback in self.get_post_trainer_create_callbacks(trainer):
             trainer.add_callback(callback)
 

src/axolotl/core/trainers/base.py

@@ -42,12 +42,20 @@ from axolotl.core.trainers.utils import (
 )
 from axolotl.utils import get_not_null
 from axolotl.utils.bench import get_gpu_memory_usage
+from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import is_main_process
 from axolotl.utils.logging import get_logger
 from axolotl.utils.samplers import MultipackBatchSampler, get_dataset_lengths
 
 LOG = get_logger(__name__)
 
+REDUCTION_FNS = {
+    "mean": torch.mean,
+    "min": torch.min,
+    "max": torch.max,
+    "sum": torch.sum,
+}
+
 
 class AxolotlTrainer(
     PackingMixin,
@@ -63,6 +71,15 @@ class AxolotlTrainer(
 
     args = None  # type: "AxolotlTrainingArguments"  # type: ignore[name-defined]
     tag_names = ["axolotl"]
+    _axolotl_cfg: DictDefault | None = None
+
+    @property
+    def axolotl_cfg(self):
+        return self._axolotl_cfg
+
+    @axolotl_cfg.setter
+    def axolotl_cfg(self, cfg):
+        self._axolotl_cfg = cfg
 
     def __init__(
         self,
@@ -78,9 +95,10 @@ class AxolotlTrainer(
         self._signature_columns = None  # workaround for pylint
 
         super().__init__(*_args, **kwargs)
-
         self.train_data_collator = self.data_collator
-        self._stored_metrics = defaultdict(lambda: defaultdict(list))
+        self._stored_metrics = defaultdict(
+            lambda: defaultdict(lambda: {"values": [], "reduction": "mean"})
+        )
         if self.args.orpo_alpha:
             self.loss_fct = torch.nn.CrossEntropyLoss(reduction="none")
 
@@ -327,6 +345,17 @@ class AxolotlTrainer(
         #     outputs = model(**inputs)
         #     loss = trainer_weighted_loss(outputs, labels, shift_labels=True)
         #     return (loss, outputs) if return_outputs else loss
+
+        # track number of tokens for tokens per second calculation
+        if self.args.include_tkps:
+            inputs_key = "labels" if "labels" in inputs else "input_ids"
+            if hasattr(self.state, "num_tokens"):
+                self.state.num_tokens = (
+                    self.state.num_tokens + (inputs[inputs_key] != -100).sum().cpu()
+                )
+            else:
+                self.state.num_tokens = (inputs[inputs_key] != -100).sum().cpu()
+
         if self.args.orpo_alpha:
             return self.orpo_compute_loss(
                 model,
@@ -342,6 +371,11 @@ class AxolotlTrainer(
             num_items_in_batch=num_items_in_batch,
         )
 
+    @override
+    def evaluate(self, *args, **kwargs):
+        LOG.info("Running evaluation step...")
+        return super().evaluate(*args, **kwargs)
+
     @staticmethod
     def orpo_concatenate_inputs(inputs, label_pad_token=-100, pad_token=0, device=None):
         concatenated_batch = {}
@@ -526,9 +560,6 @@ class AxolotlTrainer(
 
         super().create_accelerator_and_postprocess()
 
-        # now we need to put parallelism_config back on the PartialState since we rely on that info in other places
-        # PartialState().parallelism_config = self.accelerator.state.parallelism_config
-
         if self.is_fsdp_enabled:
             if (
                 "limit_all_gathers" in self.args.fsdp_config
@@ -568,29 +599,61 @@ class AxolotlTrainer(
         """
         # logs either has 'loss' or 'eval_loss'
         train_eval = "train" if "loss" in logs else "eval"
-        # Add averaged stored metrics to logs
-        for key, metrics in self._stored_metrics[train_eval].items():
-            logs[key] = torch.tensor(metrics).mean().item()
+
+        for key, metric_data in self._stored_metrics[train_eval].items():
+            values = torch.tensor(metric_data["values"])  # type: ignore[arg-type]
+            reduction_type = metric_data["reduction"]
+
+            fn = REDUCTION_FNS.get(reduction_type)
+            if fn is None:
+                raise NotImplementedError(
+                    "Metric reduction must be one of [mean, min, max, sum]"
+                )
+            logs[key] = round(fn(values).item(), 4)
 
         if is_main_process():
             # Add memory usage
             try:
                 active, allocated, reserved = get_gpu_memory_usage()
-                logs["memory/max_mem_active(gib)"] = round(active, 2)
-                logs["memory/max_mem_allocated(gib)"] = round(allocated, 2)
-                logs["memory/device_mem_reserved(gib)"] = round(reserved, 2)
+                logs["memory/max_active (GiB)"] = round(active, 2)
+                logs["memory/max_allocated (GiB)"] = round(allocated, 2)
+                logs["memory/device_reserved (GiB)"] = round(reserved, 2)
             except (ValueError, TypeError, FileNotFoundError):
                 pass
 
+        if self.args.include_tkps and train_eval == "train":
+            # each rank will log its own tokens per second
+            # for logging_steps > 1 we obtain a moving average of this metric
+            logs["tokens_per_second_per_gpu"] = round(
+                self.state.last_tokens_per_second.item() / self.args.logging_steps, 2
+            )
+
         del self._stored_metrics[train_eval]
 
         return super().log(logs, start_time)
 
     def store_metrics(
-        self, metrics: dict[str, float], train_eval: Literal["train", "eval"] = "train"
+        self,
+        metrics: dict[str, float] | dict[str, tuple[int | float, str]],
+        train_eval: Literal["train", "eval"] = "train",
+        reduction: Literal["mean", "min", "max", "sum"] = "mean",
     ) -> None:
+        """
+        Store metrics with specified reduction type.
+
+        Args:
+            metrics: Dictionary of metric names to values, or metric names to (value,
+                reduction_type) tuples.
+            train_eval: Whether this is for training or evaluation.
+        """
         for key, value in metrics.items():
-            self._stored_metrics[train_eval][key].append(value)
+            if isinstance(value, tuple):
+                value, _reduction = value  # type: ignore[assignment]
+            else:
+                value, _reduction = value, reduction
+
+            self._stored_metrics[train_eval][key]["values"].append(value)
+            self._stored_metrics[train_eval][key]["reduction"] = _reduction
 
     def _save_checkpoint(self, model, trial, **kwargs):
         # make sure the checkpoint dir exists, since trainer is flakey
@@ -657,6 +720,11 @@ class AxolotlTrainer(
                 LOG.info(
                     "Saving Trainer.data_collator.tokenizer by default as Trainer.processing_class is `None`"
                 )
-                self.data_collator.tokenizer.save_pretrained(output_dir)
+                save_jinja_files = True
+                if self.axolotl_cfg:
+                    save_jinja_files = self.axolotl_cfg.tokenizer_save_jinja_files
+                self.data_collator.tokenizer.save_pretrained(
+                    output_dir, save_jinja_files=save_jinja_files
+                )
             # Good practice: save your training arguments together with the trained model
             torch.save(self.args, os.path.join(output_dir, TRAINING_ARGS_NAME))

src/axolotl/core/training_args_base.py

@@ -49,6 +49,12 @@ class AxolotlTrainingMixins:
         default=False,
         metadata={"help": "Use real batches for efficient training."},
     )
+    include_tkps: bool = field(
+        default=True,
+        metadata={
+            "help": "Whether to include tokens per second in the training metrics."
+        },
+    )
     eval_sample_packing: Optional[bool] = field(
         default=None,
         metadata={"help": "Use sample packing for efficient evals."},

src/axolotl/datasets.py

@@ -1,18 +1,17 @@
-"""Module containing Dataset functionality"""
+"""
+Module containing dataset functionality.
+
+We want this to be a wrapper for an existing dataset that we have loaded. Lets use the
+concept of middlewares to wrap each dataset. We'll use the collators later on to pad the
+datasets.
+"""
 
-import torch
 from datasets import Dataset, IterableDataset
 
 from axolotl.utils.logging import get_logger
 
 from .prompt_tokenizers import PromptTokenizingStrategy
 
-# We want this to be a wrapper for an existing dataset that we have loaded
-# lets use the concept of middlewares to wrap each dataset, for example
-# ConstantLengthDataset(ShuffledDataset([TokenizedPromptDataset(alpaca_dataset)]))
-# let's check to ensure we don't truncate an item in the middle, we'll use
-# the collators later on to pad the datasets
-
 LOG = get_logger(__name__)
 
 
@@ -86,133 +85,3 @@ def wrap_dataset_for_tokenized_prompt(
             **map_kwargs,
         )
     return TokenizedPromptDataset(prompt_tokenizer, dataset, **kwargs)
-
-
-# TODO this isn't the best since it can't interleave datasets
-class ConstantLengthDataset(IterableDataset):
-    """Iterable dataset that returns constant length chunks of tokens from stream of
-    text files.
-
-    Args:
-        tokenizer: The processor used for processing the data.
-        dataset: Dataset with text files.
-        seq_length: Length of token sequences to return.
-    """
-
-    def __init__(
-        self,
-        tokenizer,
-        datasets,
-        seq_length=2048,
-    ):
-        self.tokenizer = tokenizer
-        self.concat_token_id = tokenizer.eos_token_id
-        self.datasets: list[IterableDataset] = datasets
-        self.seq_length = seq_length
-
-        vocab_size = len(tokenizer.get_vocab())
-
-        if vocab_size <= torch.iinfo(torch.int16).max:
-            self.tokens_dtype = torch.int16
-        elif vocab_size <= torch.iinfo(torch.int32).max:
-            self.tokens_dtype = torch.int32
-        else:
-            self.tokens_dtype = torch.int64
-
-    def __iter__(self):
-        buffer = {
-            "input_ids": [],
-            "attention_mask": [],
-            "labels": [],
-            "position_ids": [],
-        }
-        buffer_len = 0
-        for dataset in self.datasets:
-            idx = 0
-            iterator = iter(dataset)
-            more_examples = True
-            while more_examples:
-                try:
-                    example = next(iterator)
-                    idx += 1
-                except StopIteration:
-                    more_examples = False
-                    example = None
-
-                add_concat_token = False
-                if example:
-                    example_len = len(example["input_ids"])
-                    add_concat_token = example["input_ids"][-1] != self.concat_token_id
-                else:
-                    example_len = 0
-
-                if not example_len or (
-                    buffer_len + int(add_concat_token) + example_len > self.seq_length
-                ):
-                    if buffer["input_ids"]:
-                        input_ids = torch.cat(buffer["input_ids"], dim=-1)[
-                            : self.seq_length
-                        ]
-                        attention_mask = torch.cat(buffer["attention_mask"], dim=-1)[
-                            : self.seq_length
-                        ]
-                        position_ids = torch.cat(buffer["position_ids"], dim=-1)[
-                            : self.seq_length
-                        ]
-                        labels = torch.cat(buffer["labels"], dim=-1)[: self.seq_length]
-                        if labels.size() == input_ids.size() and (
-                            attention_mask.size() == input_ids.size()
-                        ):
-                            yield {
-                                "input_ids": input_ids,
-                                "labels": labels,
-                                "attention_mask": attention_mask,
-                                "position_ids": position_ids,
-                            }
-                        else:
-                            LOG.warning(
-                                "Dropping batch due to tensor size mismatch "
-                                f"input_ids: {input_ids.size()}, "
-                                f"labels: {labels.size()}, "
-                                f"attention_mask: {attention_mask.size()}"
-                            )
-                    buffer = {
-                        "input_ids": [],
-                        "attention_mask": [],
-                        "labels": [],
-                        "position_ids": [],
-                    }
-                    buffer_len = 0
-                    idx = 1
-
-                if example:
-                    # FIXME
-                    # just going to drop data points that are too long
-                    if len(example["input_ids"]) <= self.seq_length:
-                        input_ids = example["input_ids"]
-                        attention_mask = example["attention_mask"]
-                        labels = example["labels"]
-
-                        if add_concat_token:
-                            input_ids.append(self.concat_token_id)
-                            attention_mask.append(1)
-                            labels.append(self.concat_token_id)
-
-                        input_ids_with_concat = torch.tensor(
-                            input_ids, dtype=self.tokens_dtype
-                        )
-                        attention_mask_with_concat = torch.tensor(
-                            [idx * m for m in attention_mask], dtype=torch.int16
-                        )
-                        labels_with_concat = torch.tensor(
-                            labels, dtype=self.tokens_dtype
-                        )
-                        position_ids = torch.arange(
-                            len(input_ids), dtype=self.tokens_dtype
-                        )
-
-                        buffer["input_ids"].append(input_ids_with_concat)
-                        buffer["attention_mask"].append(attention_mask_with_concat)
-                        buffer["labels"].append(labels_with_concat)
-                        buffer["position_ids"].append(position_ids)
-                        buffer_len += len(input_ids)

src/axolotl/integrations/base.py

@@ -142,7 +142,7 @@ class BasePlugin:
             model: The loaded model.
         """
 
-    def get_trainer_cls(self, cfg: DictDefault) -> Trainer | None:
+    def get_trainer_cls(self, cfg: DictDefault) -> type[Trainer] | None:
         """Returns a custom class for the trainer.
 
         Args:

src/axolotl/integrations/config.py

@@ -20,8 +20,8 @@ from typing import Any, Dict, List, Type
 
 from axolotl.utils.schemas.config import (
     AxolotlConfigWCapabilities as AxolotlConfigWCapabilitiesBase,
+    AxolotlInputConfig as AxolotlInputConfigBase,
 )
-from axolotl.utils.schemas.config import AxolotlInputConfig as AxolotlInputConfigBase
 
 
 def merge_input_args():

src/axolotl/integrations/cut_cross_entropy/README.md

@@ -19,7 +19,7 @@ python scripts/cutcrossentropy_install.py | sh
 
 - If you are installing from pip
 ```bash
-pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@0ee9ee8"
+pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@c6a32c5"
 ```
 
 ## Usage
@@ -34,6 +34,7 @@ plugins:
 - arcee
 - cohere
 - cohere2
+- deepseek_v3
 - gemma
 - gemma2
 - gemma3
@@ -42,6 +43,7 @@ plugins:
 - gemma3n_text
 - glm
 - glm4
+- glm4_moe
 - gpt_oss
 - granite
 - granitemoe
@@ -64,6 +66,7 @@ plugins:
 - qwen3
 - qwen3_moe
 - smollm3
+- seed_oss
 - voxtral
 
 ## Citation

src/axolotl/integrations/cut_cross_entropy/__init__.py

@@ -35,7 +35,7 @@ LOG = get_logger(__name__)
 
 _CCE_INSTALL_MESSAGE = (
     "Please install Axolotl's fork of cut_cross_entropy with transformers support using "
-    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@0ee9ee8"`'
+    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@c6a32c5"`'
 )
 
 

src/axolotl/integrations/diffusion/README.md

@@ -0,0 +1,154 @@
+# Diffusion LM Training Plugin for Axolotl
+
+This plugin enables diffusion language model training using an approach inspired by
+LLaDA (Large Language Diffusion Models) within Axolotl.
+
+## Overview
+
+LLaDA is a diffusion-based approach to language model training that uses:
+- **Random token masking** during training instead of next-token prediction
+- **Bidirectional attention** to allow the model to attend to the full context
+- **Importance weighting** based on masking probabilities for stable training
+
+This approach can lead to more robust language models with better understanding of
+bidirectional context.
+
+## Installation
+
+The plugin is included with Axolotl. See our
+[installation docs](https://docs.axolotl.ai/docs/installation.html).
+
+## Quickstart
+
+Train with an example config (Llama‑3.2 1B):
+   - Pretrain: `axolotl train examples/llama-3/diffusion-3.2-1b-pretrain.yaml`
+   - SFT: `axolotl train examples/llama-3/diffusion-3.2-1b-sft.yaml`
+
+### Basic Configuration
+
+You can also modify your existing configs to enable / customize diffusion training.
+
+Add the following to your Axolotl config:
+
+```yaml
+# Enable diffusion LM training plugin
+plugins:
+  - axolotl.integrations.diffusion.DiffusionPlugin
+```
+
+And, configure the nested `diffusion` block (defaults shown):
+
+```yaml
+diffusion:
+  noise_schedule: linear  # or "cosine"
+  min_mask_ratio: 0.1
+  max_mask_ratio: 0.9
+  num_diffusion_steps: 128
+  eps: 1e-3
+  importance_weighting: true
+
+  # Mask token (training auto-adds if missing, avoid pad/eos)
+  mask_token_str: "<|diffusion_mask|>"
+  # Or use an existing special token id (e.g., 128002 for Llama-3.x)
+  # mask_token_id: 128002
+
+  # Sample generation during training (optional)
+  generate_samples: true
+  generation_interval: 100
+  num_generation_samples: 3
+  generation_steps: 128
+  generation_temperature: 0.0
+  generation_max_length: 100
+```
+
+## Supported Models
+
+Any models that support 4D attention masks should work out of the box. If not, please
+create an [issue](https://github.com/axolotl-ai-cloud/axolotl/issues) or open a
+[PR](https://github.com/axolotl-ai-cloud/axolotl/compare)!
+
+## How It Works
+
+### Random Masking
+During training, tokens are randomly masked:
+- Sample timestep `t` uniformly from [0, 1]
+- Calculate masking probability: `p = (1 - eps) * t + eps`
+- Randomly mask tokens with probability `p`
+
+### Diffusion Loss
+
+Loss is computed only on masked tokens with (optional) importance weighting:
+
+```python
+loss = sum(cross_entropy(pred, target) / p_mask) / total_tokens
+```
+
+## Sample Generation
+
+When `diffusion.generate_samples: true`, the plugin generates samples during training:
+
+```
+Sample 1:
+   Original (45 tokens): The quick brown fox jumps over the lazy dog...
+   Masked (18/45 tokens, 40.0%): The [MASK] [MASK] fox [MASK] over [MASK] lazy [MASK]...
+   Generated: The quick brown fox jumps over the lazy dog...
+```
+
+Samples are logged to console and wandb (if enabled).
+
+## Inference
+
+Diffusion inference is integrated into the standard Axolotl CLI. Use the same config
+you trained with and run:
+
+```
+axolotl inference path/to/your-config.yaml
+```
+
+Optionally, pass `--gradio` to use a simple web interface.
+
+Interactive controls (prefix the prompt with commands):
+- `:complete N` → completion mode with N new masked tokens appended (default 64)
+- `:mask R` → random masking mode with target mask ratio R in [0.0, 1.0]
+
+Example session:
+
+```
+================================================================================
+Commands:
+:complete N -> completion mode with N tokens (default 64)
+:mask R     -> random masking with ratio R (0.0–1.0)
+================================================================================
+Give me an instruction (Ctrl + D to submit):
+
+:mask 0.4 The quick brown fox jumps over the lazy dog
+
+Masked (40.0%):
+The [MASK] brown [MASK] jumps over the [MASK] dog
+
+Generated:
+The quick brown fox jumps over the loud dog
+```
+
+## Metrics and Monitoring
+
+The plugin adds (or modifies) several metrics to track diffusion training:
+
+- `train/loss`: Weighted diffusion loss
+- `train/accuracy`: Accuracy on masked tokens
+- `train/mask_ratio`: Average fraction of tokens masked
+- `train/num_masked_tokens`: Number of tokens masked
+- `train/avg_p_mask`: Average masking probability
+- `train/ce_loss`: Unweighted cross-entropy loss
+- `train/importance_weight_avg`: Average importance weight
+
+## Limitations
+
+- No flash attention support
+- No RL training support
+
+## References
+
+- [LLaDA Paper](https://arxiv.org/abs/2404.10406)
+- [Axolotl Documentation](https://docs.axolotl.ai/)
+- [API reference for plugin](https://docs.axolotl.ai/docs/api/integrations.diffusion.args.html#axolotl.integrations.diffusion.args)

src/axolotl/integrations/diffusion/__init__.py

@@ -0,0 +1,19 @@
+"""Diffusion LM training plugin init."""
+
+from .args import DiffusionArgs, DiffusionConfig
+from .callbacks import DiffusionGenerationCallback
+from .generation import generate
+from .plugin import DiffusionPlugin
+from .trainer import DiffusionTrainer
+from .utils import create_bidirectional_attention_mask, resolve_mask_token_id
+
+__all__ = [
+    "DiffusionArgs",
+    "DiffusionPlugin",
+    "DiffusionTrainer",
+    "generate",
+    "resolve_mask_token_id",
+    "create_bidirectional_attention_mask",
+    "DiffusionGenerationCallback",
+    "DiffusionConfig",
+]

src/axolotl/integrations/diffusion/args.py

@@ -0,0 +1,95 @@
+"""Config args for diffusion LM training (nested under `diffusion:`)."""
+
+from __future__ import annotations
+
+from typing import Literal
+
+from pydantic import BaseModel, Field, model_validator
+
+
+class DiffusionConfig(BaseModel):
+    """Nested diffusion configuration available under the `diffusion` key."""
+
+    # Noise schedule config
+    noise_schedule: Literal["linear", "cosine"] = Field(
+        default="linear", description="Type of noise schedule for diffusion training"
+    )
+    min_mask_ratio: float = Field(
+        default=0.1,
+        ge=0.0,
+        le=1.0,
+        description="Minimum masking ratio for diffusion noise schedule",
+    )
+    max_mask_ratio: float = Field(
+        default=0.9,
+        ge=0.0,
+        le=1.0,
+        description="Maximum masking ratio for diffusion noise schedule",
+    )
+    num_diffusion_steps: int = Field(
+        default=128, ge=1, description="Number of diffusion timesteps"
+    )
+    eps: float = Field(
+        default=1e-3,
+        ge=0.0,
+        le=1.0,
+        description="Epsilon value for minimum masking probability in forward process",
+    )
+
+    # Training config
+    importance_weighting: bool = Field(
+        default=True,
+        description="Apply importance weighting to loss based on masking probability",
+    )
+    mask_token_id: int | None = Field(
+        default=None,
+        description=(
+            "Token ID to use for masking. Unset by default; can use one of the "
+            "tokenizer's special tokens here."
+        ),
+    )
+    mask_token_str: str | None = Field(
+        default=None,
+        description=(
+            "Token string to use as a mask. If `mask_token_id` is invalid or unset, "
+            "this token will be ensured to exist as an additional special token and "
+            "used. If absent, a default '<|diffusion_mask|>' will be added."
+        ),
+    )
+
+    # Sample generation config
+    generate_samples: bool = Field(
+        default=True, description="Enable sample generation during training"
+    )
+    generation_interval: int = Field(
+        default=100, ge=1, description="Generate samples every N steps"
+    )
+    num_generation_samples: int = Field(
+        default=3, ge=1, description="Number of samples to generate each time"
+    )
+    generation_steps: int = Field(
+        default=128, ge=1, description="Number of diffusion steps for generation"
+    )
+    generation_temperature: float = Field(
+        default=0.0,
+        ge=0.0,
+        description="Temperature for generation sampling (0.0 = deterministic)",
+    )
+    generation_max_length: int = Field(
+        default=100, ge=1, description="Maximum sequence length for generation"
+    )
+
+    @model_validator(mode="after")
+    def _validate_mask_ratios(self) -> "DiffusionConfig":
+        if self.min_mask_ratio > self.max_mask_ratio:
+            raise ValueError("min_mask_ratio must be ≤ max_mask_ratio")
+        return self
+
+
+class DiffusionArgs(BaseModel):
+    """Plugin entry that exposes the nested `diffusion` block to the core config."""
+
+    diffusion: DiffusionConfig = Field(
+        default_factory=DiffusionConfig,
+        description="Diffusion training configuration. Only nested block is supported.",
+    )

src/axolotl/integrations/diffusion/callbacks.py

@@ -0,0 +1,174 @@
+"""Callbacks for diffusion training."""
+
+import logging
+import sys
+
+import wandb
+from colorama import Fore, Style
+from transformers.trainer_callback import TrainerCallback, TrainerControl, TrainerState
+from transformers.training_args import TrainingArguments
+
+from .generation import generate_samples
+
+# Simpler logger for more readable sample generation
+logger = logging.getLogger(__name__)
+if not logger.handlers:
+    handler = logging.StreamHandler(sys.stdout)
+    handler.setFormatter(logging.Formatter("%(message)s"))
+    logger.addHandler(handler)
+    logger.propagate = False
+logger.setLevel(logging.INFO)
+
+
+class DiffusionGenerationCallback(TrainerCallback):
+    """Callback for generating samples during diffusion training."""
+
+    def __init__(self, trainer):
+        self.trainer = trainer
+
+    def on_step_end(
+        self,
+        args: TrainingArguments,
+        state: TrainerState,
+        control: TrainerControl,
+        **kwargs,
+    ):
+        """Generate samples at specified intervals."""
+        if (
+            state.global_step > 0
+            and state.global_step % self.trainer.cfg.diffusion.generation_interval == 0
+        ):
+            if not self.trainer.state.is_world_process_zero:
+                return
+
+            # Use eval dataloader if available, otherwise use train dataloader
+            dataloader = None
+            try:
+                if getattr(self.trainer, "eval_dataset", None) is not None:
+                    dataloader = self.trainer.get_eval_dataloader()
+            except Exception:
+                dataloader = None
+            if dataloader is None:
+                dataloader = self.trainer.get_train_dataloader()
+
+            # Generate samples
+            diffusion_cfg = self.trainer.cfg.diffusion
+            samples = generate_samples(
+                model=self.trainer.model,
+                tokenizer=self.trainer.processing_class,
+                dataloader=dataloader,
+                num_generation_samples=diffusion_cfg.num_generation_samples,
+                max_length=diffusion_cfg.generation_max_length,
+                num_diffusion_steps=diffusion_cfg.generation_steps,
+                temperature=diffusion_cfg.generation_temperature,
+                mask_token_id=diffusion_cfg.mask_token_id,
+            )
+
+            # Log samples
+            self._log_samples(samples, state.global_step)
+
+    def _log_samples(self, samples: list, step: int):
+        """Log generated samples."""
+        if not samples:
+            return
+
+        logger.info("=" * 60)
+        logger.info("GENERATED SAMPLES")
+        logger.info("=" * 60)
+
+        for i, sample_data in enumerate(samples, 1):
+            original = sample_data["original"]
+            masked = sample_data["masked"]
+            generated = sample_data["generated"]
+            mask_ratio = sample_data["mask_ratio"]
+            masked_tokens = sample_data["masked_tokens"]
+            total_tokens = sample_data["total_tokens"]
+
+            logger.info(f"\nSample {i}:")
+            logger.info(f"\tOriginal ({total_tokens} tokens): {original}")
+            logger.info(
+                f"\tMasked ({masked_tokens}/{total_tokens} tokens, "
+                f"{mask_ratio:.1%}): {masked}"
+            )
+
+            try:
+                gen_ids = sample_data.get("generated_ids")
+                orig_ids = sample_data.get("orig_ids")
+                masked_positions = set(sample_data.get("masked_positions") or [])
+                if isinstance(gen_ids, list) and isinstance(orig_ids, list):
+                    styles: list[str] = []
+                    for i, tid in enumerate(gen_ids):
+                        if i in masked_positions:
+                            if i < len(orig_ids) and tid == orig_ids[i]:
+                                styles.append("green")
+                            elif i < len(orig_ids):
+                                styles.append("red")
+                            else:
+                                styles.append("normal")
+                        else:
+                            same = i < len(orig_ids) and tid == orig_ids[i]
+                            styles.append("dim" if same else "normal")
+
+                    spans: list[tuple[str, int, int]] = []
+                    if gen_ids:
+                        cur = styles[0]
+                        start = 0
+                        for i in range(1, len(gen_ids)):
+                            s = styles[i]
+                            if s != cur:
+                                spans.append((cur, start, i))
+                                cur, start = s, i
+                        spans.append((cur, start, len(gen_ids)))
+
+                    parts = []
+                    for style_name, a, b in spans:
+                        chunk_text = self.trainer.processing_class.decode(
+                            gen_ids[a:b], skip_special_tokens=False
+                        )
+                        if style_name == "green":
+                            parts.append(Fore.GREEN + chunk_text + Style.RESET_ALL)
+                        elif style_name == "red":
+                            parts.append(Fore.RED + chunk_text + Style.RESET_ALL)
+                        else:
+                            if style_name == "dim":
+                                parts.append(Style.DIM + chunk_text + Style.RESET_ALL)
+                            else:
+                                parts.append(chunk_text)
+                    logger.info("\tGenerated:\n%s", "".join(parts))
+                else:
+                    logger.info(f"\tGenerated: {generated}")
+            except Exception:
+                logger.info(f"\tGenerated: {generated}")
+
+        logger.info("=" * 60)
+
+        if self.trainer.cfg.use_wandb:
+            if wandb.run is not None:
+                wandb.log(
+                    {
+                        "generated_samples": wandb.Table(
+                            columns=[
+                                "step",
+                                "original",
+                                "masked",
+                                "generated",
+                                "mask_ratio",
+                                "masked_tokens",
+                                "total_tokens",
+                            ],
+                            data=[
+                                [
+                                    step,
+                                    sample["original"],
+                                    sample["masked"],
+                                    sample["generated"],
+                                    f"{sample['mask_ratio']:.1%}",
+                                    sample["masked_tokens"],
+                                    sample["total_tokens"],
+                                ]
+                                for sample in samples
+                            ],
+                        )
+                    },
+                    step=step,
+                )

src/axolotl/integrations/diffusion/generation.py

@@ -0,0 +1,409 @@
+"""Sample generation utilities for diffusion training."""
+
+import re
+from typing import Any, List, Literal, Optional
+
+import torch
+
+from axolotl.utils.logging import get_logger
+
+from .utils import create_bidirectional_attention_mask
+
+LOG = get_logger(__name__)
+
+
+def generate_samples(
+    model: torch.nn.Module,
+    tokenizer: Any,
+    dataloader: Optional[Any] = None,
+    num_generation_samples: int = 3,
+    max_length: int = 100,
+    num_diffusion_steps: int = 128,
+    temperature: float = 0.0,
+    mask_token_id: int = 32000,
+    mode: Literal["random", "completion"] = "random",
+    completion_tokens: int = 0,
+    target_mask_ratio: Optional[float] = None,
+) -> List[dict]:
+    """
+    Generate text samples using the diffusion model by randomly masking sequences from
+    the given dataset and running the reverse diffusion process.
+
+    Args:
+        model: The wrapped or unwrapped model
+        tokenizer: Tokenizer for encoding/decoding
+        dataloader: Validation dataloader (for sampling sequences)
+        num_generation_samples: Number of samples to generate
+        max_length: Maximum length of sequences to use
+        num_diffusion_steps: Number of diffusion steps for generation
+        temperature: Temperature for sampling (0.0 = deterministic)
+        mask_token_id: Token ID used for masking
+
+    Returns:
+        List of dictionaries with original text, masked text, and generated text
+    """
+    if dataloader is None:
+        LOG.warning("No validation dataloader provided, cannot generate samples")
+        return []
+
+    unwrapped_model = model.module if hasattr(model, "module") else model
+    training = unwrapped_model.training
+    unwrapped_model.eval()
+
+    # Resolve device robustly (some modules don't expose `.device`)
+    device = getattr(unwrapped_model, "device", None)
+    if device is None:
+        try:
+            device = next(unwrapped_model.parameters()).device
+        except StopIteration:
+            device = torch.device("cpu")
+    generations = []
+
+    # Sample sequences from validation dataset
+    sampled_sequences = _sample_sequences_from_dataloader(
+        dataloader, num_generation_samples, max_length, device
+    )
+    LOG.info(f"Sampled {len(sampled_sequences)} sequences from validation dataset")
+
+    # Generate samples using reverse diffusion process
+    with torch.no_grad():
+        for sample in sampled_sequences:
+            if isinstance(sample, dict):
+                original_sequence = sample.get("input_ids")
+                labels_seq = sample.get("labels")
+                attn_seq = sample.get("attention_mask")
+            else:
+                original_sequence = sample
+                labels_seq = None
+                attn_seq = None
+            generation_result = generate(
+                unwrapped_model,
+                tokenizer,
+                original_sequence,
+                num_diffusion_steps,
+                temperature,
+                mask_token_id,
+                mode=mode,
+                completion_tokens=completion_tokens,
+                target_mask_ratio=target_mask_ratio,
+                labels=labels_seq,
+                attention_mask=attn_seq,
+            )
+            generations.append(generation_result)
+
+    # Restore prior training state
+    if training:
+        unwrapped_model.train()
+    else:
+        unwrapped_model.eval()
+
+    return generations
+
+
+def _sample_sequences_from_dataloader(
+    dataloader: Any, num_samples: int, max_length: int, device: torch.device
+) -> List[Any]:
+    """Sample sequences from validation dataloader."""
+    sampled_sequences: list[dict[str, torch.Tensor] | torch.Tensor] = []
+    sample_count = 0
+
+    # Skip a random number of batches (we could be more clever about this)
+    skip_batches = torch.randint(0, 10, (1,)).item()
+    batch_count = 0
+
+    for batch in dataloader:
+        # Skip some batches for variety
+        if batch_count < skip_batches:
+            batch_count += 1
+            continue
+
+        if sample_count >= num_samples:
+            break
+
+        batch_count += 1
+        input_ids = batch["input_ids"]
+        attention_mask = batch.get("attention_mask")
+        labels = batch.get("labels")
+
+        # Randomly sample from sequences in this batch
+        batch_indices = torch.randperm(input_ids.size(0)).tolist()
+
+        for i in batch_indices:
+            if sample_count >= num_samples:
+                break
+
+            # Get actual sequence length (non-padded)
+            if attention_mask is not None:
+                seq_len = attention_mask[i].sum().item()
+            else:
+                seq_len = input_ids.size(1)
+
+            if seq_len < 10:
+                continue
+
+            # Determine truncation length
+            max_total = min(seq_len, max_length)
+            if labels is not None:
+                labels_i = labels[i][:seq_len]
+                answer_mask = labels_i != -100
+                if not answer_mask.any():
+                    # No answer tokens; skip for SFT masking
+                    continue
+                first_ans_idx = int(
+                    torch.nonzero(answer_mask, as_tuple=False)[0].item()
+                )
+                prompt_len = first_ans_idx
+                if prompt_len >= max_total:
+                    # Prompt alone reaches cap; cannot include any answer
+                    continue
+                remaining_answer = int(answer_mask[prompt_len:].sum().item())
+                allowed_answer = max_total - prompt_len
+                take_answer = min(remaining_answer, allowed_answer)
+                if take_answer <= 0:
+                    continue
+                actual_length = prompt_len + take_answer
+            else:
+                actual_length = max_total
+
+            # Extract the (possibly truncated) sequence
+            sequence = input_ids[i][:actual_length].unsqueeze(0).to(device)
+            attn_seq = (
+                attention_mask[i][:actual_length].unsqueeze(0).to(device)
+                if attention_mask is not None
+                else None
+            )
+            if labels is not None:
+                labels_seq = labels[i][:actual_length].unsqueeze(0).to(device)
+                sampled_sequences.append(
+                    {
+                        "input_ids": sequence,
+                        "labels": labels_seq,
+                        "attention_mask": attn_seq,
+                    }
+                )
+            else:
+                if attn_seq is not None:
+                    sampled_sequences.append(
+                        {"input_ids": sequence, "attention_mask": attn_seq}
+                    )
+                else:
+                    sampled_sequences.append(sequence)
+            sample_count += 1
+
+    return sampled_sequences
+
+
+def generate(
+    model: torch.nn.Module,
+    tokenizer: Any,
+    original_sequence: torch.Tensor,
+    num_diffusion_steps: int,
+    temperature: float,
+    mask_token_id: int,
+    *,
+    mode: Literal["random", "completion"] = "random",
+    completion_tokens: int = 0,
+    target_mask_ratio: Optional[float] = None,
+    labels: Optional[torch.Tensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+) -> dict:
+    """Generate a single sample using reverse diffusion."""
+    # Get original text for comparison
+    original_text = tokenizer.decode(
+        original_sequence[0].cpu(), skip_special_tokens=True
+    )
+
+    # Build masked sequence
+    if (
+        labels is not None
+        and labels.numel() > 0
+        and (labels == -100).any()
+        and (labels != -100).any()
+    ):
+        # SFT case: completely mask all answer tokens (labels != -100)
+        total_tokens = original_sequence.size(1)
+        masked_indices = (labels != -100).to(dtype=torch.bool)
+        masked_sequence = original_sequence.clone()
+        masked_sequence[masked_indices] = mask_token_id
+        masked_tokens = int(masked_indices.sum().item())
+        mask_ratio = masked_tokens / max(int(total_tokens), 1)
+    elif mode == "completion" and completion_tokens > 0:
+        # Append mask tokens to the right for completion
+        total_tokens = original_sequence.size(1) + int(completion_tokens)
+        masked_indices = torch.zeros(
+            1, total_tokens, dtype=torch.bool, device=original_sequence.device
+        )
+        masked_indices[0, -int(completion_tokens) :] = True
+
+        append = torch.full(
+            (1, int(completion_tokens)), mask_token_id, device=original_sequence.device
+        )
+        masked_sequence = torch.cat([original_sequence, append], dim=1)
+        masked_tokens = int(completion_tokens)
+        mask_ratio = masked_tokens / total_tokens
+    else:
+        # Apply random masking with optional fixed ratio
+        total_tokens = original_sequence.size(1)
+        if target_mask_ratio is None:
+            min_ratio, max_ratio = 0.1, 0.7
+            target_mask_ratio = (
+                torch.rand(1).item() * (max_ratio - min_ratio) + min_ratio
+            )
+        target_masked_tokens = max(1, int(total_tokens * float(target_mask_ratio)))
+
+        # Create random mask indices
+        mask_positions = torch.randperm(total_tokens)[:target_masked_tokens]
+        masked_indices = torch.zeros(
+            1, total_tokens, dtype=torch.bool, device=original_sequence.device
+        )
+        masked_indices[0, mask_positions] = True
+
+        # Create masked sequence
+        masked_sequence = original_sequence.clone()
+        masked_sequence[masked_indices] = mask_token_id
+
+        # Calculate actual mask ratio
+        masked_tokens = masked_indices.sum().item()
+        mask_ratio = masked_tokens / total_tokens
+
+    # Get masked text for comparison
+    masked_text = tokenizer.decode(masked_sequence[0].cpu(), skip_special_tokens=False)
+    masked_text = _clean_masked_text(masked_text, tokenizer, mask_token_id)
+
+    # Run reverse diffusion process
+    sequence = masked_sequence.clone()
+    attention_mask = create_bidirectional_attention_mask(
+        sequence, attention_mask, sample_packing=attention_mask is not None
+    )
+    for step in range(num_diffusion_steps):
+        sequence = _diffusion_step(
+            model,
+            sequence,
+            step,
+            num_diffusion_steps,
+            temperature,
+            mask_token_id,
+            attention_mask,
+        )
+    generated_text = tokenizer.decode(sequence[0].cpu(), skip_special_tokens=True)
+
+    # Collect diagnostic info
+    final_ids = sequence[0].detach().cpu().tolist()
+    orig_ids_for_render = original_sequence[0].detach().cpu().tolist()
+    if masked_indices is not None:
+        masked_positions = (
+            torch.where(masked_indices[0])[0].detach().cpu().tolist()
+            if masked_indices.ndim == 2
+            else []
+        )
+    else:
+        masked_positions = []
+
+    result = {
+        "original": original_text,
+        "masked": masked_text,
+        "generated": generated_text,
+        "mask_ratio": mask_ratio,
+        "masked_tokens": masked_tokens,
+        "total_tokens": total_tokens,
+        "generated_ids": final_ids,
+        "masked_positions": masked_positions,
+        "orig_ids": orig_ids_for_render,
+        "formatted": (
+            f"Original: '{original_text}' → Masked: '{masked_text}' "
+            f"({mask_ratio:.1%}) → Generated: '{generated_text}'"
+        ),
+    }
+
+    return result
+
+
+def _clean_masked_text(masked_text: str, tokenizer: Any, mask_token_id: int) -> str:
+    """Clean up masked text for display."""
+    mask_token_repr = tokenizer.decode([mask_token_id], skip_special_tokens=False)
+    cleaned = masked_text.replace(mask_token_repr, "[MASK]")
+
+    # Remove literal special token strings
+    if hasattr(tokenizer, "special_tokens_map"):
+        for token_value in tokenizer.special_tokens_map.values():
+            if token_value and isinstance(token_value, str):
+                cleaned = cleaned.replace(token_value, "")
+
+    # Normalize whitespace but preserve newlines
+    cleaned = cleaned.replace("\r\n", "\n").replace("\r", "\n")
+    cleaned = re.sub(r"[ \t]+", " ", cleaned)
+    cleaned = "\n".join(line.rstrip() for line in cleaned.split("\n")).strip()
+    return cleaned
+
+
+def _diffusion_step(
+    model: torch.nn.Module,
+    sequence: torch.Tensor,
+    step: int,
+    num_diffusion_steps: int,
+    temperature: float,
+    mask_token_id: int,
+    attention_mask: torch.Tensor | None = None,
+) -> torch.Tensor:
+    """Perform a single diffusion step with remasking."""
+    # Only process if there are masked tokens remaining
+    current_mask = sequence == mask_token_id
+    if not current_mask.any():
+        return sequence
+
+    # Create or use provided attention mask
+    if attention_mask is None:
+        batch_size, seq_len = sequence.shape
+        attention_mask = torch.ones(
+            batch_size, 1, seq_len, seq_len, dtype=torch.bool, device=sequence.device
+        )
+
+    # Forward pass
+    outputs = model(input_ids=sequence, attention_mask=attention_mask)
+    logits = outputs.logits
+
+    # Only sample at currently masked positions
+    if current_mask.any():
+        masked_logits = logits[current_mask]
+
+        # Apply temperature scaling
+        if temperature > 0:
+            scaled_logits = masked_logits / temperature
+        else:
+            scaled_logits = masked_logits
+
+        # Suppress mask token in outputs
+        scaled_logits[:, mask_token_id] = -float("inf")
+
+        if temperature > 0:
+            # Add Gumbel noise for sampling
+            gumbel_noise = -torch.log(
+                -torch.log(torch.rand_like(scaled_logits, dtype=torch.float32))
+            )
+            gumbel_logits = scaled_logits + gumbel_noise
+            predicted_tokens = torch.argmax(gumbel_logits, dim=-1)
+        else:
+            predicted_tokens = torch.argmax(scaled_logits, dim=-1)
+
+        # Calculate probabilities for confidence scoring
+        probs = torch.softmax(scaled_logits, dim=-1)
+        predicted_token_probs = probs[range(len(predicted_tokens)), predicted_tokens]
+
+        # Determine how many tokens to unmask this step
+        remaining_masked = current_mask.sum().item()
+        if step == num_diffusion_steps - 1:
+            num_to_unmask = remaining_masked
+        else:
+            unmask_ratio = 1.0 / (num_diffusion_steps - step)
+            num_to_unmask = max(1, int(remaining_masked * unmask_ratio))
+
+        # Select highest confidence predictions to unmask
+        if num_to_unmask >= remaining_masked:
+            sequence[current_mask] = predicted_tokens
+        else:
+            _, top_indices = predicted_token_probs.topk(num_to_unmask)
+            mask_positions = torch.where(current_mask)[1]
+            positions_to_unmask = mask_positions[top_indices]
+            sequence[0, positions_to_unmask] = predicted_tokens[top_indices]
+
+    return sequence

src/axolotl/integrations/diffusion/plugin.py

@@ -0,0 +1,41 @@
+"""Diffusion LM training plugin for Axolotl."""
+
+from peft import PeftModel
+from transformers import PreTrainedModel
+
+from axolotl.integrations.base import BasePlugin
+from axolotl.utils.dict import DictDefault
+from axolotl.utils.logging import get_logger
+
+from .trainer import DiffusionTrainer
+
+LOG = get_logger(__name__)
+
+
+class DiffusionPlugin(BasePlugin):
+    """
+    Plugin for diffusion language model training.
+
+    This plugin enables diffusion-based training using the LLaDA approach, which uses
+    random masking and bidirectional attention to train language models.
+    """
+
+    def __init__(self):
+        super().__init__()
+        self.cfg = None
+
+    def get_input_args(self) -> str:
+        """Returns the pydantic model for LLaDA plugin arguments."""
+        return "axolotl.integrations.diffusion.DiffusionArgs"
+
+    def post_model_load(self, cfg: DictDefault, model: PreTrainedModel | PeftModel):
+        """Perform actions after model is loaded."""
+        self.cfg = cfg
+
+    def get_trainer_cls(self, cfg: DictDefault) -> type[DiffusionTrainer] | None:
+        """Return custom trainer class for diffusion training."""
+        return DiffusionTrainer
+
+    def post_trainer_create(self, cfg: DictDefault, trainer: DiffusionTrainer):
+        """Configure trainer after creation."""
+        trainer.set_config(cfg)

src/axolotl/integrations/diffusion/trainer.py

@@ -0,0 +1,301 @@
+"""Custom trainer for diffusion LM training."""
+
+from typing import Any, Literal
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from axolotl.core.trainers.base import AxolotlTrainer
+from axolotl.utils.dict import DictDefault
+from axolotl.utils.logging import get_logger
+
+from .callbacks import DiffusionGenerationCallback
+from .utils import create_bidirectional_attention_mask
+
+LOG = get_logger(__name__)
+
+
+class DiffusionTrainer(AxolotlTrainer):
+    """Custom trainer for diffusion LM training that overrides loss computation."""
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.cfg = None
+        self._special_token_ids = None
+
+    def set_config(self, config: DictDefault):
+        """Set config for diffusion training."""
+        self.cfg = config
+        self._cache_special_token_ids()
+        self._resolve_mask_token_id()
+
+        token_id = int(getattr(self.cfg.diffusion, "mask_token_id", 0))
+        LOG.info(f"Diffusion: using mask_token_id={token_id}")
+
+        if getattr(config.diffusion, "generate_samples", True):
+            generation_callback = DiffusionGenerationCallback(self)
+            self.add_callback(generation_callback)
+
+    def _resolve_mask_token_id(self) -> None:
+        """Ensure mask_token_id is valid for the current tokenizer."""
+        from .utils import resolve_mask_token_id
+
+        tokenizer = getattr(self, "processing_class", None)
+        if tokenizer is None:
+            return
+
+        mid = resolve_mask_token_id(
+            tokenizer,
+            self.cfg,
+            allow_add=True,
+            model=getattr(self, "model", None),
+        )
+        try:
+            self.cfg.diffusion.mask_token_id = int(mid)
+        except Exception:
+            pass
+
+    def compute_loss(
+        self,
+        model: nn.Module,
+        inputs: dict[str, torch.Tensor],
+        return_outputs: bool = False,
+        num_items_in_batch: torch.Tensor | None = None,
+    ) -> torch.Tensor | tuple[torch.Tensor, dict[str, torch.Tensor]]:
+        """Override compute_loss to use diffusion loss."""
+        input_ids = inputs.get("input_ids")
+        attention_mask = inputs.get("attention_mask")
+        labels = inputs.get("labels")
+
+        if input_ids is None:
+            raise ValueError("input_ids is required for diffusion training")
+
+        loss, outputs = self._compute_diffusion_loss(
+            model, input_ids, attention_mask, labels
+        )
+
+        if return_outputs:
+            return loss, outputs
+        return loss
+
+    def _cache_special_token_ids(self):
+        """Cache special token IDs to avoid repeated tokenizer access."""
+        if self.processing_class is None:
+            self._special_token_ids = set()
+            return
+
+        tokenizer = self.processing_class
+        special_tokens = set()
+
+        if hasattr(tokenizer, "bos_token_id") and tokenizer.bos_token_id is not None:
+            special_tokens.add(tokenizer.bos_token_id)
+        if hasattr(tokenizer, "eos_token_id") and tokenizer.eos_token_id is not None:
+            special_tokens.add(tokenizer.eos_token_id)
+        if hasattr(tokenizer, "pad_token_id") and tokenizer.pad_token_id is not None:
+            special_tokens.add(tokenizer.pad_token_id)
+
+        self._special_token_ids = special_tokens
+
+    def _forward_process(
+        self,
+        input_ids: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        labels: torch.Tensor | None = None,
+        eps: float = 1e-3,
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Forward noising process. A timestep is sampled along the process, and tokens are
+        masked with probability determined by the configured noise schedule.
+
+        Args:
+            input_ids: Input token ids [batch_size, seq_len].
+            attention_mask: Attention mask [batch_size, seq_len].
+            labels: Labels for SFT training [batch_size, seq_len].
+            eps: Small epsilon value for minimum masking probability.
+
+        Returns:
+            noisy_batch: Input with some tokens masked.
+            masked_indices: Boolean mask indicating which tokens were masked.
+            p_mask: Masking probabilities for each token [batch_size, seq_len].
+        """
+        batch_size, seq_len = input_ids.shape
+        device = input_ids.device
+
+        # Sample random timesteps for each sample in batch
+        t = torch.rand(batch_size, device=device)
+        p_mask = (1 - eps) * t + eps  # [batch_size]
+        p_mask = p_mask[:, None].repeat(1, seq_len)  # [batch_size, seq_len]
+
+        # Don't mask padding tokens if attention_mask is provided
+        if attention_mask is not None:
+            valid_mask = attention_mask.bool()
+            p_mask = p_mask * valid_mask.float()
+
+        # Create mask to exclude special tokens
+        special_token_mask = torch.zeros_like(input_ids, dtype=torch.bool)
+        if self._special_token_ids:
+            for token_id in self._special_token_ids:
+                special_token_mask |= input_ids == token_id
+
+        # Create random mask based on p_mask
+        masked_indices = torch.rand((batch_size, seq_len), device=device) < p_mask
+        masked_indices = masked_indices & ~special_token_mask
+        if attention_mask is not None:
+            masked_indices = masked_indices & attention_mask.bool()
+
+        # For SFT data, only mask answer tokens
+        if labels is not None:
+            answer_mask = labels != -100
+            masked_indices = masked_indices & answer_mask
+
+        # Create masked input
+        mask_token_id = int(self.cfg.diffusion.mask_token_id)
+        mask_value = torch.full_like(input_ids, mask_token_id)
+        noisy_batch = torch.where(masked_indices, mask_value, input_ids)
+
+        return noisy_batch, masked_indices, p_mask
+
+    def _compute_diffusion_loss(
+        self,
+        model: nn.Module,
+        input_ids: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
+        labels: torch.Tensor | None = None,
+    ) -> tuple[torch.Tensor, torch.Tensor | Any]:
+        """
+        Compute diffusion loss.
+
+        Args:
+            model: The model to compute loss for.
+            input_ids: Ground truth token ids [batch_size, seq_len].
+            attention_mask: Attention mask [batch_size, seq_len].
+            labels: Labels for SFT training [batch_size, seq_len].
+
+        Returns:
+            loss: Cross-entropy loss.
+            metrics: Dictionary of metrics.
+        """
+        # Short-circuit empty sequences
+        if input_ids is None or input_ids.numel() == 0 or input_ids.shape[1] == 0:
+            zero = torch.tensor(
+                0.0,
+                device=(input_ids.device if input_ids is not None else None),
+                requires_grad=True,
+            )
+            return zero, {}
+
+        # If an attention_mask is provided and all positions are padding for every
+        # sample in this batch, skip the step.
+        if attention_mask is not None:
+            if attention_mask.dim() == 2 and (attention_mask.sum(dim=1) == 0).all():
+                zero = torch.tensor(0.0, device=input_ids.device, requires_grad=True)
+                return zero, {}
+
+        # Apply forward process
+        noisy_batch, masked_indices, p_mask = self._forward_process(
+            input_ids, attention_mask, labels, self.cfg.diffusion.eps
+        )
+
+        # Create bidirectional attention mask
+        bidirectional_mask = create_bidirectional_attention_mask(
+            input_ids, attention_mask, sample_packing=self.cfg.sample_packing
+        )
+
+        # Forward pass
+        outputs = model(
+            input_ids=noisy_batch.long(),
+            attention_mask=bidirectional_mask,
+        )
+        logits = outputs.logits
+
+        if masked_indices.sum() > 0:
+            valid_indices = torch.where(masked_indices)
+            batch_indices, seq_indices = valid_indices
+
+            masked_logits = logits[batch_indices, seq_indices]
+            masked_targets = input_ids[batch_indices, seq_indices]
+            masked_p_mask = p_mask[batch_indices, seq_indices]
+
+            # Compute cross-entropy loss without reduction
+            token_loss = F.cross_entropy(
+                masked_logits.float(), masked_targets, reduction="none"
+            )
+
+            if self.cfg.diffusion.importance_weighting:
+                masked_p_mask = masked_p_mask.float()
+                weighted_loss = token_loss / masked_p_mask
+            else:
+                weighted_loss = token_loss
+
+            if labels is not None:
+                # For SFT data: normalize by answer token count per sample
+                answer_mask = labels != -100
+                answer_lengths = answer_mask.sum(dim=1).float()  # [batch_size]
+
+                # Get batch indices for masked tokens
+                masked_batch_indices = batch_indices
+
+                # Sum losses per sample and divide by answer length
+                batch_size = input_ids.shape[0]
+                loss_per_sample = torch.zeros(batch_size, device=input_ids.device)
+                for i in range(batch_size):
+                    sample_mask = masked_batch_indices == i
+                    if sample_mask.sum() > 0:
+                        sample_loss = weighted_loss[sample_mask].sum()
+                        denom = answer_lengths[i].clamp(min=1.0)
+                        loss_per_sample[i] = sample_loss / denom
+
+                loss = loss_per_sample.mean()
+            else:
+                # Non-SFT: when importance weighting is enabled, use unbiased estimator
+                # (sum(loss/p) / total_tokens). Otherwise, average over masked tokens
+                # for stable scaling across varying mask ratios.
+                if self.cfg.diffusion.importance_weighting:
+                    loss = weighted_loss.sum() / (
+                        input_ids.shape[0] * input_ids.shape[1]
+                    )
+                else:
+                    loss = weighted_loss.mean()
+
+            ce_loss = token_loss.mean()
+
+            # Compute accuracy on masked tokens
+            with torch.no_grad():
+                pred_tokens = masked_logits.argmax(dim=-1)
+                accuracy = (pred_tokens == masked_targets).float().mean()
+        else:
+            loss = torch.tensor(0.0, device=input_ids.device, requires_grad=True)
+            accuracy = torch.tensor(0.0, device=input_ids.device)
+            ce_loss = torch.tensor(0.0, device=input_ids.device)
+            masked_p_mask = torch.tensor(1.0, device=input_ids.device)
+
+        avg_p_mask = (
+            p_mask[masked_indices].mean().item() if masked_indices.any() else 0.0
+        )
+        metrics = {
+            "loss": loss.item(),
+            "accuracy": accuracy.item(),
+            "mask_ratio": masked_indices.float().mean().item(),
+            "num_masked_tokens": (masked_indices.sum().item(), "sum"),
+            "avg_p_mask": avg_p_mask,
+            "ce_loss": ce_loss.item(),
+        }
+
+        # If doing SFT training, log answer-specific metrics
+        if self.cfg.datasets is not None:
+            with torch.no_grad():
+                answer_mask = labels != -100
+                answer_lengths = answer_mask.sum(dim=1).float()  # type: ignore
+                total_answer_tokens = answer_mask.sum().item()  # type: ignore
+                total_tokens = labels.numel()  # type: ignore
+                metrics["answer_ratio"] = total_answer_tokens / max(total_tokens, 1)
+                metrics["avg_answer_length"] = answer_lengths.mean().item()
+
+        if self.cfg.diffusion.importance_weighting:
+            metrics["importance_weight_avg"] = (1.0 / masked_p_mask).mean().item()
+
+        train_eval: Literal["train", "eval"] = "train" if model.training else "eval"
+        self.store_metrics(metrics, train_eval=train_eval)
+
+        return loss, outputs

src/axolotl/integrations/diffusion/utils.py

@@ -0,0 +1,159 @@
+"""Shared utilities for diffusion integration."""
+
+from __future__ import annotations
+
+from typing import Any, Optional
+
+import torch
+
+from axolotl.utils.dict import DictDefault
+
+
+def resolve_mask_token_id(
+    tokenizer: Any,
+    cfg: DictDefault,
+    *,
+    allow_add: bool,
+    model: Any | None = None,
+    default_token: str = "<|diffusion_mask|>",
+) -> int:
+    """Resolve mask token id. Training may add a new special token; inference won't."""
+    # Determine vocab size if available
+    vocab_size = None
+    if tokenizer is not None:
+        if hasattr(tokenizer, "vocab_size") and tokenizer.vocab_size is not None:
+            try:
+                vocab_size = int(tokenizer.vocab_size)  # type: ignore[arg-type]
+            except Exception:
+                vocab_size = None
+        elif hasattr(tokenizer, "__len__"):
+            try:
+                vocab_size = int(len(tokenizer))
+            except Exception:
+                vocab_size = None
+
+    # Use explicit id from config if provided
+    diffusion_cfg = getattr(cfg, "diffusion", None)
+    # Fallback to top-level attr names only if nested missing (shouldn't happen)
+    cfg_id = (
+        getattr(diffusion_cfg, "mask_token_id", None)
+        if diffusion_cfg is not None
+        else getattr(cfg, "diffusion_mask_token_id", None)
+    )
+    if isinstance(cfg_id, int) and cfg_id >= 0:
+        if vocab_size is None or cfg_id < vocab_size:
+            return int(cfg_id)
+
+    def _existing_special_token_id(token_str: str | None) -> int | None:
+        """Attempt to resolve an existing special token string to a real ID."""
+        if not token_str or not hasattr(tokenizer, "convert_tokens_to_ids"):
+            return None
+        try:
+            token_id = tokenizer.convert_tokens_to_ids(token_str)
+        except Exception:
+            return None
+
+        if not isinstance(token_id, int) or token_id < 0:
+            return None
+
+        # Ensure it's registered as special and not UNK, and within vocab
+        unk_id = getattr(tokenizer, "unk_token_id", None)
+        specials = set(getattr(tokenizer, "all_special_tokens", []) or [])
+        addl = set(getattr(tokenizer, "additional_special_tokens", []) or [])
+        is_special = token_str in specials or token_str in addl
+        in_vocab = vocab_size is None or token_id < vocab_size
+        if (
+            (unk_id is not None and token_id == unk_id)
+            or not is_special
+            or not in_vocab
+        ):
+            return None
+        return token_id
+
+    # Try mask token string if provided
+    token_str = (
+        getattr(diffusion_cfg, "mask_token_str", None)
+        if diffusion_cfg is not None
+        else getattr(cfg, "diffusion_mask_token_str", None)
+    )
+    for candidate in (token_str, default_token):
+        token_id = _existing_special_token_id(candidate)
+        if isinstance(token_id, int):
+            try:
+                if diffusion_cfg is None:
+                    cfg.diffusion_mask_token_id = int(token_id)  # legacy fallback
+                else:
+                    diffusion_cfg.mask_token_id = int(token_id)
+            except Exception:
+                pass
+            return int(token_id)
+
+    # Optionally add and return a dedicated special token during training
+    if allow_add and hasattr(tokenizer, "add_special_tokens"):
+        token_to_add = token_str or default_token
+        try:
+            tokenizer.add_special_tokens({"additional_special_tokens": [token_to_add]})
+
+            # Resize embeddings if possible
+            if (
+                model is not None
+                and hasattr(tokenizer, "__len__")
+                and hasattr(model, "resize_token_embeddings")
+            ):
+                try:
+                    model.resize_token_embeddings(len(tokenizer))
+                except Exception:
+                    pass
+            new_id = tokenizer.convert_tokens_to_ids(token_to_add)
+            if isinstance(new_id, int) and new_id >= 0:
+                try:
+                    if diffusion_cfg is None:
+                        cfg.diffusion_mask_token_id = int(new_id)  # legacy fallback
+                    else:
+                        diffusion_cfg.mask_token_id = int(new_id)
+                except Exception:
+                    pass
+                return int(new_id)
+        except Exception:
+            pass
+
+    # Fallback to unk or 0 (do not update cfg)
+    fallback = getattr(tokenizer, "unk_token_id", 0) or 0
+    return int(fallback)
+
+
+def create_bidirectional_attention_mask(
+    input_ids: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    sample_packing: bool = False,
+) -> torch.Tensor:
+    """
+    Create bidirectional attention mask to override default causal masking.
+    Handles sample-packed sequences where different samples are identified
+    by different attention mask values.
+
+    Args:
+        input_ids: Input token ids [batch_size, seq_len]
+        attention_mask: Attention mask [batch_size, seq_len]
+        sample_packing: Whether sample packing is enabled
+
+    Returns:
+        bidirectional_mask: 4D attention mask [batch_size, 1, seq_len, seq_len]
+    """
+    batch_size, seq_len = input_ids.shape
+    device = input_ids.device
+
+    if attention_mask is None or not sample_packing:
+        return torch.ones(
+            batch_size, 1, seq_len, seq_len, dtype=torch.bool, device=device
+        )
+
+    # Handle sample packing: tokens can only attend within their sample
+    mask_i = attention_mask.unsqueeze(2)  # [batch_size, seq_len, 1]
+    mask_j = attention_mask.unsqueeze(1)  # [batch_size, 1, seq_len]
+
+    # Tokens can attend to each other if they have the same non-zero sample ID
+    bidirectional_mask = (mask_i == mask_j) & (mask_i > 0)
+
+    # Add head dimension: [batch_size, 1, seq_len, seq_len]
+    return bidirectional_mask.unsqueeze(1)

src/axolotl/loaders/adapter.py

@@ -14,6 +14,7 @@ from peft import (
     PeftConfig,
     PeftMixedModel,
     PeftModel,
+    TaskType,
     get_peft_model,
 )
 from transformers import PreTrainedModel
@@ -98,6 +99,17 @@ def load_lora(
         lora_config_kwargs["use_rslora"] = cfg.peft_use_rslora
     if cfg.peft_layer_replication:
         lora_config_kwargs["layer_replication"] = cfg.peft_layer_replication
+    if cfg.peft_trainable_token_indices:
+        lora_config_kwargs["trainable_token_indices"] = cfg.peft_trainable_token_indices
+
+    # Determine the correct PEFT task type
+    model_cls = type(model).__name__
+    if "SequenceClassification" in model_cls:
+        task_type = TaskType.SEQ_CLS
+    elif "TokenClassification" in model_cls:
+        task_type = TaskType.TOKEN_CLS
+    else:
+        task_type = TaskType.CAUSAL_LM
 
     lora_config = LoraConfig(
         r=cfg.lora_r,
@@ -110,7 +122,7 @@ def load_lora(
         fan_in_fan_out=cfg.lora_fan_in_fan_out,
         modules_to_save=cfg.lora_modules_to_save if cfg.lora_modules_to_save else None,
         bias="none",
-        task_type="CAUSAL_LM",
+        task_type=task_type,
         **lora_config_kwargs,
     )
 

src/axolotl/loaders/model.py

@@ -673,6 +673,33 @@ class ModelLoader:
 
         return hf_ds_cfg
 
+    def _load_model_from_config(self, model_loader_class=None) -> PreTrainedModel:
+        """
+        Load model with random initialization using from_config.
+
+        Uses the selected loader when provided; otherwise falls back to the auto loader.
+        """
+        loader = model_loader_class or self.auto_model_loader
+        if loader in [AutoModelForCausalLM, AutoModelForVision2Seq]:
+            model = loader.from_config(
+                config=self.model_config,
+                trust_remote_code=self.cfg.trust_remote_code or False,
+            )
+        else:
+            model = loader(config=self.model_config)
+
+        return model
+
+    def _load_model_from_pretrained(self, model_loader_class=None) -> PreTrainedModel:
+        """Load model from pretrained weights."""
+        loader = model_loader_class or self.auto_model_loader
+        kwargs = {
+            "config": self.model_config,
+            "trust_remote_code": self.cfg.trust_remote_code or False,
+            **self.model_kwargs,
+        }
+        return loader.from_pretrained(self.base_model, **kwargs)
+
     def _build_model(self) -> bool:
         """Load model, with load strategy depending on config."""
         skip_move_to_device = False
@@ -687,7 +714,8 @@ class ModelLoader:
         if self.is_fsdp_enabled:
             if self.cfg.fsdp_config.cpu_ram_efficient_loading:
                 skip_move_to_device = True
-                # Don't delete device_map for QLoRA + FSDP - it was set correctly in _set_device_map
+                # Don't delete device_map for QLoRA + FSDP - it was set correctly in
+                # _set_device_map
                 if (
                     "device_map" in self.model_kwargs
                     and not self.is_qlora_and_fsdp_enabled
@@ -716,6 +744,11 @@ class ModelLoader:
                 or self.cfg.qlora_sharded_model_loading
             )
         ):
+            if self.cfg.reinit_weights:
+                LOG.warning(
+                    "reinit_weights is not supported with sharded quantized loading. "
+                    "Loading from pretrained weights instead."
+                )
             quant_storage = self.cfg.torch_dtype
             quantization_config = getattr(
                 self.model_config, "quantization_config", None
@@ -731,33 +764,12 @@ class ModelLoader:
                 quantization_config=quantization_config,
             )
             skip_move_to_device = True
-        elif (
-            self.model_config.model_type in ["llama", "llama4"]
-            and not self.cfg.trust_remote_code
-            and not self.cfg.gptq
-        ):
-            # Please don't remove underscore binding without reading the fn docstring.
-            _ = self._configure_zero3_memory_efficient_loading()
-
-            # Load model with random initialization if specified
-            if self.cfg.random_init_weights:
-                # AutoModel classes support the from_config method
-                if self.auto_model_loader in [
-                    AutoModelForCausalLM,
-                    AutoModelForVision2Seq,
-                ]:
-                    self.model = self.auto_model_loader.from_config(
-                        config=self.model_config,
-                    )
-                else:
-                    self.model = self.auto_model_loader(config=self.model_config)
-            else:
-                self.model = self.auto_model_loader.from_pretrained(
-                    self.base_model,
-                    config=self.model_config,
-                    **self.model_kwargs,
-                )
         elif self.model_type == "MambaLMHeadModel":
+            if self.cfg.reinit_weights:
+                LOG.warning(
+                    "reinit_weights is not supported with MambaLMHeadModel. "
+                    "Loading from pretrained weights instead."
+                )
             # FIXME this is janky at best and hacked together to make it work
             MambaLMHeadModel = fix_mamba_attn_for_loss()
 
@@ -770,41 +782,27 @@ class ModelLoader:
                 self.base_model,
                 **self.model_kwargs,
             )
-        elif (
-            self.model_type
-            and self.model_type != "AutoModelForCausalLM"
-            and not self.cfg.trust_remote_code
-        ):
-            if self.cfg.gptq:
-                self.model = self.auto_model_loader.from_pretrained(
-                    self.base_model,
-                    config=self.model_config,
-                    trust_remote_code=self.cfg.trust_remote_code or False,
-                    **self.model_kwargs,
-                )
-            else:
-                self.model = getattr(transformers, self.model_type).from_pretrained(
-                    self.base_model,
-                    config=self.model_config,
-                    trust_remote_code=self.cfg.trust_remote_code or False,
-                    **self.model_kwargs,
-                )
-        elif self.cfg.gptq:
-            self.model = self.auto_model_loader.from_pretrained(
-                self.base_model,
-                config=self.model_config,
-                trust_remote_code=self.cfg.trust_remote_code or False,
-                **self.model_kwargs,
-            )
         else:
-            # Please don't remove underscore binding without reading the fn docstring.
+            # Please don't remove underscore binding without reading the fn docstring
             _ = self._configure_zero3_memory_efficient_loading()
-            self.model = self.auto_model_loader.from_pretrained(
-                self.base_model,
-                config=self.model_config,
-                trust_remote_code=self.cfg.trust_remote_code or False,
-                **self.model_kwargs,
-            )
+
+            if (
+                self.model_type
+                and self.model_type != "AutoModelForCausalLM"
+                and not self.cfg.trust_remote_code
+                and not self.cfg.gptq
+            ):
+                # Use model type from transformers
+                model_loader_class = getattr(transformers, self.model_type)
+            else:
+                # Use auto model loader (handles gptq and default cases)
+                model_loader_class = self.auto_model_loader
+
+            if self.cfg.reinit_weights:
+                self.model = self._load_model_from_config(model_loader_class)
+            else:
+                self.model = self._load_model_from_pretrained(model_loader_class)
+
         if is_deepspeed_zero3_enabled():
             skip_move_to_device = True
 

src/axolotl/loaders/patch_manager.py

@@ -4,6 +4,7 @@ Applies pre- and post-model load patches for various fixes and optimizations.
 """
 
 import importlib.util
+import os
 from functools import cached_property
 
 import addict
@@ -66,6 +67,7 @@ class PatchManager:
         self._apply_mistral_cross_entropy_patch()
         self._apply_self_attention_lora_patch()
         self._apply_fsdp2_bnb_patches()
+        self._apply_patch_deepspeed_zero3()
 
     def apply_post_plugin_pre_model_load_patches(self):
         """Apply post plugin-pre_model_load load patches based on config."""
@@ -78,13 +80,7 @@ class PatchManager:
             patch_maybe_log_save_evaluate,
         )
 
-        patch_fsdp2 = (
-            self.cfg.torch_compile
-            and self.cfg.fsdp_config
-            and self.cfg.fsdp_version == 2
-        )
-
-        patch_evaluation_loop(patch_fsdp2)
+        patch_evaluation_loop()
         patch_maybe_log_save_evaluate()
 
     def apply_post_model_load_patches(self, model: PreTrainedModel):
@@ -147,14 +143,12 @@ class PatchManager:
     def _apply_flex_attention_patches(self):
         """Apply patches for flexible attention."""
         if self.cfg.flex_attention:
-            # from axolotl.monkeypatch.attention.flex_attn import (
-            #     patch_flex_make_mask,
-            #     patch_flex_wrapper,
-            # )
-            #
-            # flex_attn_compile_kwargs = self.cfg.flex_attn_compile_kwargs or {}
-            # patch_flex_wrapper(**flex_attn_compile_kwargs)
-            # patch_flex_make_mask()
+            from axolotl.monkeypatch.attention.flex_attn import (
+                patch_flex_wrapper,
+            )
+
+            flex_attn_compile_kwargs = self.cfg.flex_attn_compile_kwargs or {}
+            patch_flex_wrapper(**flex_attn_compile_kwargs)
             if self.cfg.sample_packing:
                 from axolotl.core.attention.flex_block_mask import (
                     patch_create_causal_mask,
@@ -471,3 +465,17 @@ class PatchManager:
             from axolotl.monkeypatch.lora_kernels import apply_lora_kernel_patches
 
             apply_lora_kernel_patches(model=model, cfg=self.cfg)
+
+    def _apply_patch_deepspeed_zero3(self):
+        try:
+            from transformers.integrations.deepspeed import is_deepspeed_zero3_enabled
+
+            from axolotl.monkeypatch.deepspeed_utils import apply_deepspeed_patches
+
+            if self.cfg.activation_offloading is True and (
+                is_deepspeed_zero3_enabled()
+                or os.getenv("ACCELERATE_DEEPSPEED_ZERO_STAGE") == "3"
+            ):
+                apply_deepspeed_patches()
+        except ImportError as e:
+            LOG.warning(f"DeepSpeed patches not applied: {e}")

src/axolotl/loaders/tokenizer.py

@@ -296,7 +296,7 @@ def load_tokenizer(cfg: DictDefault) -> PreTrainedTokenizer:
             )
 
         tokenizer.chat_template = chat_template_string
-    else:
+    elif getattr(tokenizer, "chat_template", None) is None:
         LOG.info(
             "No Chat template selected. Consider adding a chat template for easier inference."
         )

src/axolotl/monkeypatch/accelerate/fsdp2.py

@@ -160,9 +160,11 @@ def get_state_dict(self, model, unwrap=True):
                 state_dict[param_name] = param.cpu()
             torch.distributed.barrier()
     elif self.distributed_type == DistributedType.FSDP:
-        from torch.distributed.fsdp import FullStateDictConfig
-        from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
-        from torch.distributed.fsdp import StateDictType
+        from torch.distributed.fsdp import (
+            FullStateDictConfig,
+            FullyShardedDataParallel as FSDP,
+            StateDictType,
+        )
 
         full_state_dict_config = FullStateDictConfig(
             offload_to_cpu=True, rank0_only=True
@@ -178,6 +180,38 @@ def get_state_dict(self, model, unwrap=True):
 
     return state_dict
 
+def cast_lora_module(module):
+    base_layer_dtype = module.base_layer.weight.dtype
+    # Linear4Bit will keep it's bias term in fp32. If the weight dtype is in bf16 we are not able to
+    # wrap this. Therefore we must ensure the bias has the same dtype as the weight
+    if hasattr(module.base_layer, "bias") and module.base_layer.bias is not None:
+        if module.base_layer.weight.dtype != module.base_layer.bias.dtype:
+            log_bias_dtype_mismatch = True
+            module.base_layer.bias.data = module.base_layer.bias.data.to(
+                module.base_layer.weight.dtype
+            )
+
+    for active_adapter in module.active_adapters:
+        if module.lora_A:
+            module.lora_A[active_adapter] = module.lora_A[active_adapter].to(base_layer_dtype)
+            if hasattr(module.lora_A[active_adapter], 'bias') and module.lora_A[active_adapter].bias is not None:
+                module.lora_A[active_adapter].bias.data = module.lora_A[active_adapter].bias.data.to(base_layer_dtype)
+        if module.lora_B:
+           module.lora_B[active_adapter] = module.lora_B[active_adapter].to(base_layer_dtype)
+           if hasattr(module.lora_B[active_adapter], 'bias') and module.lora_B[active_adapter].bias is not None:
+               module.lora_B[active_adapter].bias.data = module.lora_B[active_adapter].bias.data.to(base_layer_dtype)
+        if module.lora_embedding_A:
+            module.lora_embedding_A[active_adapter] = module.lora_embedding_A[active_adapter].to(base_layer_dtype)
+            if hasattr(module.lora_embedding_A[active_adapter], 'bias') and module.lora_embedding_A[active_adapter].bias is not None:
+                module.lora_embedding_A[active_adapter].bias.data = module.lora_embedding_A[active_adapter].bias.data.to(base_layer_dtype)
+        if module.lora_embedding_B:
+            module.lora_embedding_B[active_adapter] = module.lora_embedding_B[active_adapter].to(base_layer_dtype)
+            if hasattr(module.lora_embedding_B[active_adapter], 'bias') and module.lora_embedding_B[active_adapter].bias is not None:
+                module.lora_embedding_B[active_adapter].bias.data = module.lora_embedding_B[active_adapter].bias.data.to(base_layer_dtype)
+        if module.lora_magnitude_vector:
+            module.lora_magnitude_vector[active_adapter] = module.lora_magnitude_vector[active_adapter].to(base_layer_dtype)
+            if hasattr(module.lora_magnitude_vector[active_adapter], 'bias') and module.lora_magnitude_vector[active_adapter].bias is not None:
+                module.lora_magnitude_vector[active_adapter].bias.data = module.lora_magnitude_vector[active_adapter].bias.data.to(base_layer_dtype)
 
 def _process_lora_module_for_fsdp(module, fsdp2_kwargs):
     """Helper function to process LoRA modules for FSDP2."""
@@ -193,18 +227,37 @@ def _process_lora_module_for_fsdp(module, fsdp2_kwargs):
             module.base_layer.bias.data = module.base_layer.bias.data.to(
                 module.base_layer.weight.dtype
             )
-
-    for active_adapter in module.active_adapters:
-        if module.lora_A:
-            fully_shard(module.lora_A[active_adapter], **fsdp2_kwargs)
-        if module.lora_B:
-            fully_shard(module.lora_B[active_adapter], **fsdp2_kwargs)
-        if module.lora_embedding_A:
-            fully_shard(module.lora_embedding_A[active_adapter], **fsdp2_kwargs)
-        if module.lora_embedding_B:
-            fully_shard(module.lora_embedding_B[active_adapter], **fsdp2_kwargs)
-        if module.lora_magnitude_vector:
-            fully_shard(module.lora_magnitude_vector[active_adapter], **fsdp2_kwargs)
+    fully_shard(module, **fsdp2_kwargs)
+    module.set_reshard_after_forward(False)
+    module.set_reshard_after_backward(False)
+    # for active_adapter in module.active_adapters:
+    #     for adapter_name in [
+    #         "lora_A",
+    #         "lora_B",
+    #         "lora_embedding_A",
+    #         "lora_embedding_B",
+    #         "lora_magnitude_vector",
+    #     ]:
+    #         adapter_module = getattr(module, adapter_name, None)
+    #         # print(adapter_module, adapter_name)
+    #         # torch.distributed.breakpoint()
+    #         if not adapter_module:
+    #             continue
+    #         fsdp_adapter_module = fully_shard(adapter_module[active_adapter], **fsdp2_kwargs)
+    #         # fsdp_adapter_module.unshard()
+    #         fsdp_adapter_module.set_reshard_after_backward(False)
+    #         fsdp_adapter_module.set_reshard_after_forward(False)
+            # torch.distributed.breakpoint()
+        # if module.lora_A:
+        #     fully_shard(module.lora_A[active_adapter], **fsdp2_kwargs)
+        # if module.lora_B:
+        #     fully_shard(module.lora_B[active_adapter], **fsdp2_kwargs)
+        # if module.lora_embedding_A:
+        #     fully_shard(module.lora_embedding_A[active_adapter], **fsdp2_kwargs)
+        # if module.lora_embedding_B:
+        #     fully_shard(module.lora_embedding_B[active_adapter], **fsdp2_kwargs)
+        # if module.lora_magnitude_vector:
+            # fully_shard(module.lora_magnitude_vector[active_adapter], **fsdp2_kwargs)
     return log_bias_dtype_mismatch
 
 
@@ -318,16 +371,26 @@ def fsdp2_prepare_model(accelerator, model: torch.nn.Module) -> torch.nn.Module:
             model.tie_weights()
 
     is_peft_model = isinstance(model, PeftModel)
-
+    # TODO - this doesn't actually do anything
+    for name, module in model.named_children():
+        if name == "experts":
+            # torch.distributed.breakpoint()
+            for expert in module.children():
+                # torch.distributed.breakpoint()
+                print(f"expert: {expert}")
+                for lora_module in expert.children():
+                    print(f"lora {lora_module}")
+                    # torch.distributed.breakpoint()
+                    cast_lora_module(lora_module)
+                    _process_lora_module_for_fsdp(lora_module, fsdp2_kwargs)
     auto_wrap_policy = fsdp2_prepare_auto_wrap_policy(fsdp2_plugin, model)
     log_bias_dtype_mismatch = False
     if auto_wrap_policy is not None:
         for module in get_module_children_bottom_up(model)[:-1]:
-            if is_peft_model and isinstance(module, LoraLayer):
-                module_log_bias_mismatch = _process_lora_module_for_fsdp(
-                    module, fsdp2_kwargs
-                )
-                log_bias_dtype_mismatch |= module_log_bias_mismatch
+            if is_peft_model and isinstance(module, LoraLayer) and not isinstance(module, FSDPModule):
+                # torch.distributed.breakpoint()
+                cast_lora_module(module)
+                # torch.distributed.breakpoint()
             if auto_wrap_policy(module) and not isinstance(module, FSDPModule):
                 fully_shard(module, **fsdp2_kwargs)
 
@@ -344,6 +407,9 @@ def fsdp2_prepare_model(accelerator, model: torch.nn.Module) -> torch.nn.Module:
             accelerator, model, original_sd, offload_to_cpu=offload_to_cpu
         )
 
+    # for module in model.named_modules():
+    #     if "Lora" in 
+
     if fsdp2_plugin.cpu_ram_efficient_loading and not model_has_params4bit:
         # We re-register the buffers, as they may not be in the state_dict
         for fqn, buffer_tensor in original_non_persistent_buffers.items():

src/axolotl/monkeypatch/attention/flex_attn.py

@@ -1,10 +1,11 @@
 """Flex attention monkey patch"""
 
 import sys
-from typing import Optional, Tuple, Union
 
 import torch
 import transformers
+from packaging import version
+from transformers.utils.import_utils import _torch_version, is_torch_less_or_equal
 
 from axolotl.utils.logging import get_logger
 
@@ -46,19 +47,33 @@ def patch_flex_wrapper(**flex_attn_compile_kwargs):
             """
             self.training = None
             if not self._is_flex_compiled or training != self.training:
+                self.training = training
+                if is_torch_less_or_equal("2.5.1"):
+                    self._compiled_flex_attention = torch.compile(
+                        flex_attention, dynamic=False
+                    )
                 # In PyTorch 2.6.0, there's a known issue with flex attention compilation which may
                 # cause errors. The suggested fix is to compile with "max-autotune-no-cudagraphs"
                 # see https://github.com/pytorch/pytorch/issues/146260 for training
-                self.training = training
-                LOG.info(
-                    "Compiling flex attention with kwargs: %s. This may take a while...",
-                    flex_attn_compile_kwargs,
-                )
-                self._compiled_flex_attention = torch.compile(
-                    flex_attention,
-                    **flex_attn_compile_kwargs,
-                )
-                LOG.info("Flex attention compiled successfully.")
+                elif version.parse(_torch_version).base_version == "2.6.0" and training:
+                    self._compiled_flex_attention = torch.compile(
+                        flex_attention, dynamic=False, mode="max-autotune-no-cudagraphs"
+                    )
+                # Fallback, usually the most recent torch 2.7.x+ versions
+                else:
+                    LOG.info(
+                        "Compiling flex attention with kwargs: %s. This may take a while...",
+                        flex_attn_compile_kwargs,
+                        main_process_only=True,
+                    )
+                    self._compiled_flex_attention = torch.compile(
+                        flex_attention,
+                        **flex_attn_compile_kwargs,
+                    )
+                    LOG.info(
+                        "Flex attention compiled successfully.", main_process_only=True
+                    )
+
                 self._is_flex_compiled = True
 
         def __call__(self):
@@ -68,139 +83,3 @@ def patch_flex_wrapper(**flex_attn_compile_kwargs):
     sys.modules[
         "transformers.integrations.flex_attention"
     ].WrappedFlexAttention = WrappedFlexAttention
-
-
-def patch_flex_make_mask():
-    is_torch_2_6 = torch.__version__.startswith("2.6")
-
-    if not is_torch_2_6:
-        return
-
-    from torch.nn.attention.flex_attention import (
-        _DEFAULT_SPARSE_BLOCK_SIZE as flex_default_block_size,
-    )
-    from torch.nn.attention.flex_attention import (
-        BlockMask,
-    )
-    from torch.nn.attention.flex_attention import (
-        create_block_mask as create_block_causal_mask_flex,
-    )
-
-    Offset = Union[torch.Tensor, int]
-
-    def patched_make_flex_block_causal_mask(
-        attention_mask_2d: torch.Tensor,
-        attention_chunk_size: Optional[int] = None,
-        query_length=None,
-        key_length=None,
-        offsets: Optional[Tuple[Offset, Offset]] = None,
-    ) -> "BlockMask":
-        """
-        Create a block causal document mask for a batch of sequences, both packed and unpacked.
-        Create Block causal logic and passing it into :func:`torch.nn.attention.flex_attention.create_block_mask`.
-        The resultant BlockMask is a compressed representation of the full block causal
-        mask. BlockMask is essential for performant computation of flex attention.
-        See: https://pytorch.org/blog/flexattention/
-
-        Args:
-            attention_mask_2d (torch.Tensor): Attention mask for packed and padded sequences
-            of shape (batch_size, total_seq_len). e.g.
-
-            For unpacked sequence:
-            [[1, 1, 1, 1, 0, 0, 0],
-             [1, 1, 1, 1, 1, 0, 0]]
-
-            For packed sequence:
-            [[1, 1, 1, 2, 2, 2, 0],
-             [1, 1, 2, 2, 2, 3, 3]]
-
-        Returns:
-            BlockMask
-        """
-
-        batch_size, total_seq_len = attention_mask_2d.shape
-        if not key_length:
-            key_length = total_seq_len
-        if not query_length:
-            query_length = total_seq_len
-        attention_mask_2d = torch.nn.functional.pad(
-            attention_mask_2d,
-            value=0,
-            pad=(0, abs(total_seq_len - max(key_length, flex_default_block_size))),
-        )
-        device = attention_mask_2d.device
-        document_ids = attention_mask_2d.clone()
-
-        if attention_chunk_size is not None:
-            # we create an arange, then we just // by chunk size to get [0, 0, 0, 1, 1, 1, 2, 2, 2, 3, 3, 3]
-            chunk_idxs = (document_ids.clone().fill_(1).cumsum(-1) - 1) // (
-                attention_chunk_size
-            )
-
-        # Instead of passing a tensor mask, flex attention requires a mask_mod function
-        # that determines which elements of QK^T should be included in the attention
-        # computation prior to the softmax. For sample packing, we need both the
-        # logic for both causal mask and document mask. See PyTorch's official
-        # blog post for more details: https://pytorch.org/blog/flexattention/#mask-mods
-        def causal_mask_mod(batch_idx, head_idx, q_idx, kv_idx):
-            """
-            Defines the logic of a block causal mask by combining both a standard causal mask
-            and a block diagonal document mask.
-
-            See :func:`~torchtune.modules.attention_utils.create_block_causal_mask`
-            for an illustration.
-            """
-            causal_mask = q_idx >= kv_idx  # not valid when decoding
-            document_mask = (
-                document_ids[batch_idx, q_idx] == document_ids[batch_idx, kv_idx]
-            )
-            padding_mask = attention_mask_2d[batch_idx, q_idx] > 0
-            final_mask = causal_mask & padding_mask & document_mask
-            return final_mask
-
-        def chunk_causal_mask_mod(batch_idx, head_idx, q_idx, kv_idx):
-            """
-            Combines the chunk mask with the causal mask for chunked attention.
-            """
-            chunk_mask = chunk_idxs[batch_idx, q_idx] == chunk_idxs[batch_idx, kv_idx]
-            causal_doc_mask = causal_mask_mod(batch_idx, head_idx, q_idx, kv_idx)
-            return chunk_mask & causal_doc_mask
-
-        mask_mod_maybe_combined = (
-            causal_mask_mod if attention_chunk_size is None else chunk_causal_mask_mod
-        )
-
-        if offsets is not None:
-            q_offset = offsets[0]
-            kv_offset = offsets[1]
-
-            def mask_mod(batch_idx, head_idx, q_idx, kv_idx):
-                offset_q = q_idx + q_offset
-                offset_kv = kv_idx + kv_offset
-                return mask_mod_maybe_combined(batch_idx, head_idx, offset_q, offset_kv)
-
-        else:
-            mask_mod = mask_mod_maybe_combined
-        return create_block_causal_mask_flex(
-            mask_mod=mask_mod,
-            B=batch_size,
-            H=None,  # attention head
-            Q_LEN=query_length,
-            KV_LEN=key_length,
-            device=device,
-            _compile=True,
-        )
-
-    for n in tuple(sys.modules):
-        if ".modeling_" in n:
-            if hasattr(sys.modules[n], "make_flex_block_causal_mask"):
-                sys.modules[
-                    n
-                ].make_flex_block_causal_mask = patched_make_flex_block_causal_mask
-                sys.modules[
-                    n
-                ].make_flex_block_causal_mask = patched_make_flex_block_causal_mask
-
-    transformers.integrations.flex_attention.make_flex_block_causal_mask = (
-        patched_make_flex_block_causal_mask
-    )

src/axolotl/monkeypatch/deepspeed_utils.py

@@ -0,0 +1,67 @@
+import importlib
+import importlib.util
+
+from axolotl.utils.logging import get_logger
+
+LOG = get_logger(__name__)
+
+
+def patch_checkpoint_wrapper_setattr():
+    """
+    Patch CheckpointWrapper to properly forward DeepSpeed attributes to wrapped modules.
+
+    This fixes the issue where CheckpointWrapper doesn't forward ds_* attributes
+    (like ds_grads_remaining) to the actual wrapped module, causing DeepSpeed
+    ZeRO-3 to fail when gradient checkpointing is enabled.
+
+    This issue occurs specifically with:
+    - QLoRA + DeepSpeed ZeRO-3
+    - gradient_checkpointing: true
+    - activation_offloading: true
+
+    References:
+    - https://github.com/deepspeedai/DeepSpeed/issues/7203
+    - https://github.com/deepspeedai/DeepSpeed/blob/38d1a9eb64c9e01e32eccc50b25ba18925287441/deepspeed/runtime/zero/parameter_offload.py#L424-L458
+    - https://github.com/axolotl-ai-cloud/axolotl/pull/3102
+    """
+
+    try:
+        from torch.distributed.algorithms._checkpoint.checkpoint_wrapper import (
+            CheckpointWrapper,
+        )
+
+        # Check if already patched
+        if hasattr(CheckpointWrapper, "_axolotl_setattr_patched"):
+            LOG.debug("CheckpointWrapper already patched")
+            return
+
+        original_setattr = CheckpointWrapper.__setattr__
+
+        def new_setattr(self, name: str, value) -> None:
+            if name.startswith("ds_") and hasattr(self, "_checkpoint_wrapped_module"):
+                setattr(self._checkpoint_wrapped_module, name, value)
+                LOG.debug(
+                    f"Forwarded {name} to wrapped module {type(self._checkpoint_wrapped_module).__name__}"
+                )
+            else:
+                original_setattr(self, name, value)
+
+        CheckpointWrapper.__setattr__ = new_setattr
+        CheckpointWrapper._axolotl_setattr_patched = True
+
+        LOG.info("CheckpointWrapper patched to forward DeepSpeed attributes")
+
+    except ImportError as e:
+        LOG.debug(f"CheckpointWrapper not available: {e}")
+    except Exception as e:
+        LOG.warning(f"Failed to patch CheckpointWrapper: {e}")
+
+
+def apply_deepspeed_patches():
+    """
+    Apply DeepSpeed-related patches
+    """
+    if importlib.util.find_spec("deepspeed") is not None:
+        patch_checkpoint_wrapper_setattr()
+    else:
+        LOG.debug("DeepSpeed not available, skipping patches")

src/axolotl/monkeypatch/lora_kernels.py

@@ -149,6 +149,11 @@ def get_attention_cls_from_config(cfg: DictDefault) -> Type[nn.Module]:
 
         return MistralAttention
 
+    if model_type == "gemma3_text":
+        from transformers.models.gemma3.modeling_gemma3 import Gemma3Attention
+
+        return Gemma3Attention
+
     try:
         # Dynamically import the module and attention class
         module_path = f"transformers.models.{model_type}.modeling_{model_type}"

src/axolotl/monkeypatch/multipack.py

@@ -36,8 +36,13 @@ SUPPORTED_MULTIPACK_MODEL_TYPES = [
     "glm",
     "glm4",
     "smollm3",
+    "granite",
+    "granitemoe",
+    "hunyuan_v1_dense",
+    "hunyuan_v1_moe",
     "gpt_oss",
     "arcee",
+    "seed_oss",
 ]
 
 

src/axolotl/monkeypatch/tiled_mlp/base.py

@@ -8,6 +8,94 @@ from typing import List
 import torch
 
 
+class DeepSpeedTiledMLPMoE(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        fn,
+        self,
+        x,
+        shards,
+        compute_params,
+    ) -> torch.Tensor:
+        ctx.fn = fn
+        ctx.self = self
+        ctx.shards = shards
+        ctx.compute_params = [p for p in compute_params if p.requires_grad]
+        ctx.save_for_backward(x)
+
+        x_shards = list(torch.chunk(x, chunks=shards, dim=1))
+        with torch.no_grad():
+            output_shards = [fn(self, x_shard) for x_shard in x_shards]
+
+        ctx.is_tuple_output = isinstance(output_shards[0], tuple)
+        if isinstance(output_shards[0], tuple):
+            tuple_dim_idx = [1, 0]
+            output_unsharded = tuple(
+                torch.cat(
+                    [output_shard[i] for output_shard in output_shards],
+                    dim=tuple_dim_idx[i],
+                )
+                for i in range(len(output_shards[0]))
+            )
+        else:
+            output_unsharded = torch.cat(output_shards, dim=1)
+
+        return output_unsharded
+
+    @staticmethod
+    def backward(ctx, *grads) -> torch.Tensor:
+        fn = ctx.fn
+        (x,) = ctx.saved_tensors
+        self = ctx.self
+        shards = ctx.shards
+        compute_params = ctx.compute_params
+        is_tuple_output = ctx.is_tuple_output
+
+        x_requires_grad = x.requires_grad
+        x = x.detach()
+        # detach() unsets `x.requires_grad`, so restore it
+        x.requires_grad_(x_requires_grad)
+
+        incoming_grad = grads[0]
+        x_grad = torch.zeros_like(x)
+        x_shards = list(torch.chunk(x, chunks=shards, dim=1))
+
+        shard_step = x_shards[0].numel()
+        for i, x_shard in enumerate(x_shards):
+            # Tell deepspeed not to add a new grad to its ipg bucket until the last shard is run
+            if compute_params is not None:
+                if i + 1 < shards:
+                    for param in compute_params:
+                        param.ds_grad_is_ready = False
+                else:
+                    # last shard, can add the grad
+                    for param in compute_params:
+                        param.ds_grad_is_ready = True
+
+            x_shard.requires_grad_(x_requires_grad)
+
+            shard_offset = i * shard_step
+            x_shard.grad = (
+                x_grad.view(-1)
+                .narrow(0, shard_offset, x_shard.numel())
+                .view_as(x_shard)
+            )
+            incoming_grad_shard = (
+                incoming_grad.view(-1)
+                .narrow(0, shard_offset, x_shard.numel())
+                .view_as(x_shard)
+            )
+            with torch.enable_grad():
+                output = fn(self, x_shard)
+            if is_tuple_output:
+                torch.autograd.backward(output[0], incoming_grad_shard)
+            else:
+                torch.autograd.backward(output, incoming_grad_shard)
+
+        return (None, None, x_grad, None, None)
+
+
 class TiledMLP(torch.autograd.Function):
     """
     TiledMLP implementation using gradient hooks
@@ -31,7 +119,18 @@ class TiledMLP(torch.autograd.Function):
         x_shards = list(torch.chunk(x, chunks=shards, dim=1))
         with torch.no_grad():
             output_shards = [fn(self, x_shard) for x_shard in x_shards]
-        output_unsharded = torch.cat(output_shards, dim=1)
+        ctx.is_tuple_output = isinstance(output_shards[0], tuple)
+        if isinstance(output_shards[0], tuple):
+            tuple_dim_idx = [1, 0]
+            output_unsharded = tuple(
+                torch.cat(
+                    [output_shard[i] for output_shard in output_shards],
+                    dim=tuple_dim_idx[i],
+                )
+                for i in range(len(output_shards[0]))
+            )
+        else:
+            output_unsharded = torch.cat(output_shards, dim=1)
 
         return output_unsharded
 
@@ -42,6 +141,7 @@ class TiledMLP(torch.autograd.Function):
         self = ctx.self
         shards = ctx.shards
         compute_params = ctx.compute_params
+        is_tuple_output = ctx.is_tuple_output
 
         x_requires_grad = x.requires_grad
         x = x.detach()
@@ -76,7 +176,10 @@ class TiledMLP(torch.autograd.Function):
 
             with torch.enable_grad():
                 output = fn(self, x_shard)
-            torch.autograd.backward(output, incoming_grad_shard)
+            if is_tuple_output:
+                torch.autograd.backward(output[0], incoming_grad_shard)
+            else:
+                torch.autograd.backward(output, incoming_grad_shard)
 
         # Clean up hooks
         grad_accumulator.cleanup()

src/axolotl/monkeypatch/tiled_mlp/patch.py

@@ -17,7 +17,7 @@ def patch_tiled_mlp(model_type, use_original_mlp=True, cfg_num_shards=None):
         TiledMLP as DeepSpeedTiledMLP,
     )
 
-    from axolotl.monkeypatch.tiled_mlp.base import TiledMLP
+    from axolotl.monkeypatch.tiled_mlp.base import DeepSpeedTiledMLPMoE, TiledMLP
 
     try:
         # Dynamically import the module and MLP class
@@ -64,7 +64,10 @@ def patch_tiled_mlp(model_type, use_original_mlp=True, cfg_num_shards=None):
                         for p in self._compute_params
                     )
                 ) or os.environ.get("ACCELERATE_USE_DEEPSPEED", "false") == "true":
-                    self._tiled_mlp_dist_impl = DeepSpeedTiledMLP
+                    if model_type == "gpt_oss":
+                        self._tiled_mlp_dist_impl = DeepSpeedTiledMLPMoE
+                    else:
+                        self._tiled_mlp_dist_impl = DeepSpeedTiledMLP
                 else:
                     self._tiled_mlp_dist_impl = TiledMLP
 

src/axolotl/monkeypatch/transformers/trainer_loss_calc.py

@@ -28,15 +28,6 @@ PATCHED_EVAL_CODE = {
     "array": 'metrics[f"{metric_key_prefix}_loss"] = np.nanmean(all_losses).item()',
 }
 
-ORIGINAL_FSDP2_CODE = """
-    model.eval()
-"""
-
-PATCHED_FSDP2_CODE = """
-    if hasattr(model, "eval") and callable(model.eval):
-        self.model.eval()
-"""
-
 ORIGINAL_MAYBE_CODE = "tr_loss_scalar = self._nested_gather(tr_loss).mean().item()"
 PATCHED_MAYBE_CODE = "tr_loss_scalar = self._nested_gather(tr_loss).nanmean().item()"
 
@@ -46,13 +37,7 @@ def check_evaluation_loop_is_patchable() -> bool:
     return all(value in evaluation_loop_source for value in ORIGINAL_EVAL_CODE.values())
 
 
-def check_evaluation_loop_is_fsdp2_patchable() -> bool:
-    evaluation_loop_source = inspect.getsource(Trainer.evaluation_loop)
-    evaluation_loop_source, _ = detab_code(evaluation_loop_source)
-    return ORIGINAL_FSDP2_CODE in evaluation_loop_source
-
-
-def patch_evaluation_loop(patch_fsdp2: bool):
+def patch_evaluation_loop():
     """Patch the evaluation_loop method."""
     # Check if already patched
     if hasattr(Trainer, "_original_evaluation_loop"):
@@ -75,13 +60,6 @@ def patch_evaluation_loop(patch_fsdp2: bool):
         ORIGINAL_EVAL_CODE["array"], PATCHED_EVAL_CODE["array"]
     )
 
-    # Apply FSDP2 eval guard patch if needed
-    if patch_fsdp2 and ORIGINAL_FSDP2_CODE in evaluation_loop_source:
-        evaluation_loop_source = evaluation_loop_source.replace(
-            ORIGINAL_FSDP2_CODE, PATCHED_FSDP2_CODE
-        )
-        LOG.info("Applied FSDP2 eval guard patch to evaluation_loop")
-
     # Rename the function to avoid conflicts
     evaluation_loop_source = evaluation_loop_source.replace(
         "def evaluation_loop(",

src/axolotl/prompt_tokenizers.py

@@ -75,7 +75,7 @@ class PromptTokenizingStrategy(abc.ABC):
     ) -> BatchEncoding:
         empty = BatchEncoding(data={"input_ids": [], "attention_mask": []})
         if not prompt:
-            LOG.warning("Empty text requested for tokenization.")
+            LOG.warning_once("Empty text requested for tokenization.")
             return empty
 
         result = self.tokenizer(

src/axolotl/train.py

@@ -30,11 +30,7 @@ from axolotl.contribs.lgpl import (  # pylint: disable = no-name-in-module
     fix_untrained_tokens,
 )
 from axolotl.integrations.base import PluginManager
-from axolotl.loaders import (
-    ModelLoader,
-    load_processor,
-    load_tokenizer,
-)
+from axolotl.loaders import ModelLoader, load_processor, load_tokenizer
 from axolotl.utils.ctx_managers.sequence_parallel import SequenceParallelContextManager
 from axolotl.utils.dict import DictDefault
 from axolotl.utils.distributed import cleanup_distributed
@@ -234,16 +230,15 @@ def save_trained_model(
 
     # handle QAT
     if cfg.qat:
-        from axolotl.utils.quantization import convert_qat_model_for_ptq
+        from axolotl.utils.quantization import convert_qat_model
 
-        LOG.info("Processing QAT model for saving...")
-        convert_qat_model_for_ptq(
+        convert_qat_model(
             model,
             quantize_embedding=cfg.qat.quantize_embedding,
         )
         LOG.info(
-            "QAT modules have been converted for PTQ. Please ensure you quantize "
-            "your model weights with `axolotl quantize`."
+            "QAT usage note: please ensure you quantize your model fine-tuned using QAT by running `axolotl quantize`"
+            " with the same config which you used for training."
         )
     # Handle ReLoRA early return case
     if cfg.relora:
@@ -337,9 +332,7 @@ def save_trained_model(
 
     if hasattr(cfg, "llmcompressor") and cfg.llmcompressor:
         # TODO: add integration support so this can be implemented completely within the plugin
-        from axolotl.integrations.llm_compressor.utils import (
-            save_compressed_model,
-        )
+        from axolotl.integrations.llm_compressor.utils import save_compressed_model
 
         save_compressed_model(
             model=model,
@@ -416,7 +409,9 @@ def save_initial_configs(
 
     # Pre-save the tokenizer and model configs
     LOG.info(f"Pre-saving tokenizer to {cfg.output_dir}...")
-    tokenizer.save_pretrained(str(output_dir))
+    tokenizer.save_pretrained(
+        str(Path(cfg.output_dir)), save_jinja_files=cfg.tokenizer_save_jinja_files
+    )
     if hasattr(model, "config"):
         LOG.info(f"Pre-saving model config to {cfg.output_dir}...")
         model.config.save_pretrained(str(output_dir))
@@ -592,6 +587,9 @@ def train(
 
     # Save the trained model and cleanup
     save_trained_model(cfg, trainer, model, safe_serialization)
+    tokenizer.save_pretrained(
+        str(Path(cfg.output_dir)), save_jinja_files=cfg.tokenizer_save_jinja_files
+    )
     create_model_card(cfg, trainer)
     if not cfg.use_ray:
         cleanup_distributed()

src/axolotl/utils/bench.py

@@ -60,13 +60,14 @@ def gpu_memory_usage_all(device=0):
     active = torch.cuda.memory_stats().get("active_bytes.all.peak", 0) / 1024.0**3
     allocated = torch.cuda.max_memory_allocated(device) / 1024.0**3
     reserved = torch.cuda.max_memory_reserved(device) / 1024.0**3
+    torch.cuda.reset_peak_memory_stats(device)
     return active, allocated, reserved
 
 
 def mps_memory_usage_all():
-    usage = torch.mps.current_allocated_memory() / 1024.0**3
-    reserved = torch.mps.driver_allocated_memory() / 1024.0**3
-    return usage, reserved - usage, 0
+    active = torch.mps.current_allocated_memory() / 1024.0**3
+    allocated = torch.mps.driver_allocated_memory() / 1024.0**3
+    return active, allocated, 0
 
 
 def npu_memory_usage_all(device=0):

src/axolotl/utils/callbacks/tokens_per_second.py

@@ -0,0 +1,64 @@
+"""A callback for calculating tokens per second during training."""
+
+import time
+
+import torch
+from transformers import (
+    TrainerCallback,
+    TrainerControl,
+    TrainerState,
+    TrainingArguments,
+)
+
+
+class TokensPerSecondCallback(TrainerCallback):
+    """
+    A callback to measure and log tokens per second during training.
+    """
+
+    def __init__(self, tensor_parallel_size, context_parallel_size):
+        super().__init__()
+        self.step_time = 0.0
+        self.start_time = 0.0
+        self.non_data_parallel_size = 1
+        if tensor_parallel_size is not None:
+            self.non_data_parallel_size *= tensor_parallel_size
+        if context_parallel_size is not None:
+            self.non_data_parallel_size *= context_parallel_size
+
+    def on_step_begin(
+        self,
+        args: TrainingArguments,
+        state: TrainerState,
+        control: TrainerControl,
+        **kwargs,
+    ):  # pylint: disable=unused-argument
+        self.start_time = time.perf_counter()
+        state.last_tokens_per_second = torch.zeros(1)
+
+    def on_step_end(
+        self,
+        args: TrainingArguments,
+        state: TrainerState,
+        control: TrainerControl,
+        **kwargs,
+    ):  # pylint: disable=unused-argument
+        if hasattr(state, "num_tokens"):
+            step_time = time.perf_counter() - self.start_time
+            num_tokens_per_device = state.num_tokens.clone()
+            # non data parallel groups have duplicated tokens, so we avoid double-counting
+            num_tokens_per_device = num_tokens_per_device / self.non_data_parallel_size
+            state.last_tokens_per_second = num_tokens_per_device / step_time
+
+    def on_log(
+        self,
+        args: TrainingArguments,
+        state: TrainerState,
+        control: TrainerControl,
+        logs=None,
+        **kwargs,
+    ):  # pylint: disable=unused-argument
+        # after logging, clear the running metrics
+        if hasattr(state, "last_tokens_per_second"):
+            state.last_tokens_per_second.zero_()
+            state.num_tokens = torch.zeros(1)

src/axolotl/utils/collators/__init__.py

@@ -1,11 +1,17 @@
-"""
-shared axolotl collators for multipack, mamba, multimodal
-"""
+"""Shared axolotl collators for multipacking, mamba, multimodal."""
 
-from .batching import (  # noqa: F401
+from .batching import (
     BatchSamplerDataCollatorForSeq2Seq,
     DataCollatorForSeq2Seq,
     PretrainingBatchSamplerDataCollatorForSeq2Seq,
     V2BatchSamplerDataCollatorForSeq2Seq,
 )
-from .mamba import MambaDataCollator  # noqa: F401
+from .mamba import MambaDataCollator
+
+__all__ = [
+    "DataCollatorForSeq2Seq",
+    "BatchSamplerDataCollatorForSeq2Seq",
+    "V2BatchSamplerDataCollatorForSeq2Seq",
+    "PretrainingBatchSamplerDataCollatorForSeq2Seq",
+    "MambaDataCollator",
+]

src/axolotl/utils/config/__init__.py

@@ -17,8 +17,8 @@ from axolotl.utils.dict import DictDefault
 from axolotl.utils.logging import get_logger
 from axolotl.utils.schemas.config import (
     AxolotlConfigWCapabilities as AxolotlConfigWCapabilitiesBase,
+    AxolotlInputConfig as AxolotlInputConfigBase,
 )
-from axolotl.utils.schemas.config import AxolotlInputConfig as AxolotlInputConfigBase
 from axolotl.utils.schemas.datasets import DPODataset, KTODataset, SFTDataset
 
 LOG = get_logger(__name__)
@@ -77,7 +77,7 @@ def resolve_dtype(cfg):
     if cfg.device == "mps":
         cfg.load_in_8bit = False
         cfg.tf32 = False
-        if cfg.bf16:
+        if cfg.bf16 and cfg.fp16 is not False:
             cfg.fp16 = True
         cfg.bf16 = False
     else:
@@ -273,7 +273,9 @@ def validate_config(
     # Convert datasets to proper format if needed
     if cfg.get("datasets"):
         for idx, ds_cfg in enumerate(cfg["datasets"]):
-            if cfg.get("rl") in ["dpo", "simpo"] and not isinstance(ds_cfg, DPODataset):
+            if cfg.get("rl") in ["dpo", "ipo", "simpo"] and not isinstance(
+                ds_cfg, DPODataset
+            ):
                 cfg["datasets"][idx] = DPODataset(**ds_cfg)
             elif cfg.get("rl") == "kto" and not isinstance(ds_cfg, KTODataset):
                 cfg["datasets"][idx] = KTODataset(**dict(ds_cfg))

src/axolotl/utils/ctx_managers/sequence_parallel.py

@@ -48,10 +48,10 @@ def apply_sequence_parallelism(
             - The original sequence length before padding.
             - The number of padding tokens added.
     """
-    original_seq_len = batch["input_ids"].size(1)
+    batch_size, original_seq_len = batch["input_ids"].shape
 
     # Update ring attention params if needed
-    if batch.get("position_ids") is not None:
+    if batch.get("position_ids") is not None and batch_size == 1:
         update_ring_attn_params(position_ids=batch["position_ids"])
     else:
         # If position_ids aren't already in the batch, create them

src/axolotl/utils/data/__init__.py

@@ -1,19 +1,19 @@
 """Init for `axolotl.utils.data` module."""
 
-from axolotl.utils.data.pretraining import (
-    encode_pretraining,
-    wrap_pretraining_dataset,
-)
 from axolotl.utils.data.rl import prepare_preference_datasets
 from axolotl.utils.data.sft import (
     get_dataset_wrapper,
     prepare_datasets,
 )
+from axolotl.utils.data.streaming import (
+    encode_streaming,
+    wrap_streaming_dataset,
+)
 from axolotl.utils.data.utils import md5
 
 __all__ = [
-    "encode_pretraining",
-    "wrap_pretraining_dataset",
+    "encode_streaming",
+    "wrap_streaming_dataset",
     "prepare_preference_datasets",
     "get_dataset_wrapper",
     "prepare_datasets",

src/axolotl/utils/data/sft.py

@@ -9,13 +9,13 @@ from datasets import (
     Dataset,
     DatasetDict,
     IterableDataset,
+    IterableDatasetDict,
     load_dataset,
 )
 from transformers import PreTrainedTokenizer, ProcessorMixin
 
 from axolotl.prompters import Prompter
 from axolotl.utils.data.lock import FileLockLoader
-from axolotl.utils.data.pretraining import wrap_pretraining_dataset
 from axolotl.utils.data.shared import (
     create_train_validation_split,
     datasets_with_name_generator,
@@ -26,6 +26,7 @@ from axolotl.utils.data.shared import (
     save_preprocessed_dataset,
     try_load_from_hub,
 )
+from axolotl.utils.data.streaming import wrap_streaming_dataset
 from axolotl.utils.data.utils import (
     deduplicate_and_log_datasets,
     handle_long_seq_in_dataset,
@@ -48,7 +49,6 @@ def prepare_datasets(
     cfg: DictDefault,
     tokenizer: PreTrainedTokenizer,
     processor: ProcessorMixin | None = None,
-    preprocess_iterable: bool = False,
 ) -> tuple[IterableDataset | Dataset, Dataset | None, int, list[Prompter | None]]:
     """Prepare training and evaluation datasets based on configuration.
 
@@ -56,23 +56,19 @@ def prepare_datasets(
         cfg: Dictionary mapping `axolotl` config keys to values.
         tokenizer: Tokenizer to use for processing text.
         processor: Optional processor for multimodal datasets.
-        preprocess_iterable: Whether to use iterable preprocessing.
 
     Returns:
         Tuple of (train_dataset, eval_dataset, total_steps, prompters).
     """
-    if cfg.pretraining_dataset:
-        return _prepare_pretraining_dataset(
-            cfg, tokenizer, processor, preprocess_iterable
-        )
-    return _prepare_standard_dataset(cfg, tokenizer, processor, preprocess_iterable)
+    if cfg.streaming or cfg.pretraining_dataset:
+        return _prepare_streaming_dataset(cfg, tokenizer, processor)
+    return _prepare_standard_dataset(cfg, tokenizer, processor)
 
 
 def _prepare_standard_dataset(
     cfg: DictDefault,
     tokenizer: PreTrainedTokenizer,
     processor: ProcessorMixin | None,
-    preprocess_iterable: bool,
 ) -> tuple[Dataset, Dataset | None, int, list[Prompter | None]]:
     """Prepare standard (non-pretraining) datasets."""
 
@@ -83,7 +79,6 @@ def _prepare_standard_dataset(
             cfg,
             split="train",
             processor=processor,
-            preprocess_iterable=preprocess_iterable,
         )
 
         # Overwrite eval_dataset if test data exists
@@ -93,7 +88,6 @@ def _prepare_standard_dataset(
                 cfg,
                 split="test",
                 processor=processor,
-                preprocess_iterable=preprocess_iterable,
             )
 
         return train_dataset, eval_dataset, prompters
@@ -128,22 +122,40 @@ def _prepare_standard_dataset(
     return train_dataset, eval_dataset, total_num_steps, prompters
 
 
-def _prepare_pretraining_dataset(
+def _prepare_streaming_dataset(
     cfg: DictDefault,
     tokenizer: PreTrainedTokenizer,
     processor: ProcessorMixin | None,
-    preprocess_iterable: bool,
 ) -> tuple[IterableDataset, Dataset | None, int, list[Prompter | None]]:
     """
-    Prepare dataset for pretraining mode.
+    Prepare dataset for streaming mode.
 
-    Note: Pre-training datasets are streamed from the HuggingFace Hub.
+    Note: Streaming datasets are loaded incrementally from the source.
     """
-    # Extract pretraining dataset configuration
-    pretraining_config = _extract_pretraining_config(cfg)
+    if cfg.pretraining_dataset:
+        dataset_config = _extract_pretraining_config(cfg)
+        train_dataset = _load_streaming_dataset(dataset_config, cfg, tokenizer)
+    elif cfg.sample_packing:
+        # TODO(djsaunde): Implement for multiple datasets
+        dataset_config = DictDefault(cfg.datasets[0])
 
-    # Load streaming dataset for training
-    train_dataset = _load_pretraining_dataset(pretraining_config, cfg, tokenizer)
+        # Ensure we have a split set - default to 'train' if not specified
+        if not hasattr(dataset_config, "split") or not dataset_config.split:
+            dataset_config.split = "train"
+        train_dataset = _load_streaming_dataset(dataset_config, cfg, tokenizer)
+    else:
+        # Use legacy loading function for non-packed streaming datasets
+        train_dataset, eval_dataset, prompters = _load_and_prepare_datasets(
+            tokenizer,
+            cfg,
+            split="train",
+            processor=processor,
+            streaming=True,
+        )
+
+        # Return early for non-packed streaming datasets
+        total_num_steps = cfg.max_steps if cfg.max_steps else -1
+        return train_dataset, eval_dataset, total_num_steps, prompters
 
     # Load evaluation dataset if specified
     eval_dataset = None
@@ -153,14 +165,12 @@ def _prepare_pretraining_dataset(
             cfg,
             split="test",
             processor=processor,
-            preprocess_iterable=preprocess_iterable,
+            streaming=False,
         )
 
-    if cfg.dataset_exact_deduplication:
-        LOG.info("Deduplication not available for pretrained datasets")
-
-    # For pretraining, we return max_steps directly from config
-    return train_dataset, eval_dataset, cfg.max_steps, []
+    # For streaming, we return max_steps directly from config or -1 if not set
+    total_num_steps = cfg.max_steps if cfg.max_steps else -1
+    return train_dataset, eval_dataset, total_num_steps, []
 
 
 def _extract_pretraining_config(cfg: DictDefault) -> DictDefault:
@@ -192,7 +202,7 @@ def _extract_pretraining_config(cfg: DictDefault) -> DictDefault:
     )
 
 
-def _load_pretraining_dataset(
+def _load_streaming_dataset(
     pretraining_config: DictDefault, cfg: DictDefault, tokenizer: PreTrainedTokenizer
 ) -> IterableDataset:
     """Load and prepare a streaming dataset for pretraining."""
@@ -227,15 +237,11 @@ def _load_pretraining_dataset(
         iter_dataset = iter_dataset.skip(pretraining_config["skip"])
 
     # Wrap the dataset for pretraining
-    train_dataset = wrap_pretraining_dataset(
+    train_dataset = wrap_streaming_dataset(
         iter_dataset,
         tokenizer,
         cfg,
         dataset_wrapper_partial,
-        max_tokens=cfg.sequence_len,
-        batch_size=cfg.micro_batch_size,
-        seed=cfg.seed,
-        buffer_size=cfg.pretrain_multipack_buffer_size or 10_000,
     )
 
     # Format for PyTorch
@@ -256,7 +262,7 @@ def _load_tokenized_prepared_datasets(
     cfg: DictDefault,
     split: Literal["train", "test"] = "train",
     processor: ProcessorMixin | None = None,
-    preprocess_iterable: bool = False,
+    streaming: bool = False,
 ) -> tuple[Dataset | DatasetDict, list[Prompter | None]]:
     """Load or create tokenized and prepared datasets for training or testing.
 
@@ -265,7 +271,7 @@ def _load_tokenized_prepared_datasets(
         cfg: Configuration object.
         split: Dataset split to load ('train' or 'test').
         processor: Optional processor for multimodal datasets.
-        preprocess_iterable: Whether to use iterable preprocessing.
+        streaming: Whether to use iterable preprocessing.
 
     Returns:
         Tuple of (dataset, prompters list).
@@ -296,7 +302,7 @@ def _load_tokenized_prepared_datasets(
             tokenizer,
             split,
             processor,
-            preprocess_iterable,
+            streaming,
         )
 
     return dataset, prompters
@@ -308,7 +314,7 @@ def _load_raw_datasets(
     tokenizer: PreTrainedTokenizer,
     split: str,
     processor: ProcessorMixin | None = None,
-    preprocess_iterable: bool = False,
+    streaming: bool = False,
 ) -> tuple[Dataset, list[Prompter | None]]:
     """Load, process, merge, and save raw datasets."""
     LOG.info("Loading raw datasets...", main_process_only=False)
@@ -329,7 +335,7 @@ def _load_raw_datasets(
             split=split,
             seed=cfg.seed,
             processor=processor,
-            preprocess_iterable=preprocess_iterable,
+            streaming=streaming,
         )
         datasets.append(dataset_wrapper)
         prompters.append(dataset_prompter)
@@ -337,7 +343,7 @@ def _load_raw_datasets(
     # Merge datasets
     dataset = merge_datasets(datasets, cfg)
 
-    if not cfg.skip_prepare_dataset:
+    if not cfg.skip_prepare_dataset and not streaming:
         if split == "test" and cfg.eval_sequence_len:
             dataset = handle_long_seq_in_dataset(dataset, cfg.eval_sequence_len, cfg)
         else:
@@ -361,19 +367,19 @@ def _load_and_process_single_dataset(
     split: str,
     seed: int,
     processor: ProcessorMixin | None = None,
-    preprocess_iterable: bool = False,
+    streaming: bool = False,
 ) -> tuple[Dataset | IterableDataset, Prompter | None]:
     """Load and process a single dataset based on the passed config."""
     # Load the dataset
     dataset = load_dataset_with_config(
-        dataset_config, cfg.hf_use_auth_token, streaming=preprocess_iterable
+        dataset_config, cfg.hf_use_auth_token, streaming=streaming
     )
 
     # Parse dataset type
     d_base_type, d_prompt_style = _parse_dataset_type(dataset_config.type)
 
     # Select the appropriate split
-    if isinstance(dataset, DatasetDict):
+    if isinstance(dataset, (DatasetDict, IterableDatasetDict)):
         if dataset_config.split and dataset_config.split in dataset:
             dataset = dataset[dataset_config.split]
         elif split in dataset:
@@ -479,7 +485,7 @@ def _load_and_prepare_datasets(
     cfg: DictDefault,
     split: Literal["train", "test"] = "train",
     processor: ProcessorMixin | None = None,
-    preprocess_iterable: bool = False,
+    streaming: bool = False,
 ) -> tuple[Dataset | None, Dataset | None, list[Prompter | None]]:
     """Load and prepare datasets with optional validation split and sharding.
 
@@ -488,7 +494,7 @@ def _load_and_prepare_datasets(
         cfg: Configuration object.
         split: Dataset split to load ('train' or 'test').
         processor: Optional processor for multimodal datasets.
-        preprocess_iterable: Whether to use iterable preprocessing.
+        streaming: Whether to use iterable preprocessing.
 
     Returns:
         Tuple of (train_dataset, eval_dataset, prompters).
@@ -499,7 +505,7 @@ def _load_and_prepare_datasets(
         cfg,
         split=split,
         processor=processor,
-        preprocess_iterable=preprocess_iterable,
+        streaming=streaming,
     )
 
     # Apply dataset sharding if configured using shared function

src/axolotl/utils/data/shared.py

@@ -236,11 +236,9 @@ def _load_from_local_path(
         try:
             return load_from_disk(dataset_config.path)
         except FileNotFoundError:
-            load_dataset_kwargs["streaming"] = False
             return load_dataset(dataset_config.path, **load_dataset_kwargs)
     elif local_path.is_file():
         dataset_type = get_dataset_type(dataset_config)
-        load_dataset_kwargs["streaming"] = False
         return load_dataset(
             dataset_type,
             data_files=dataset_config.path,

src/axolotl/utils/data/streaming.py

@@ -1,4 +1,4 @@
-"""data handling specific to pretraining"""
+"""Data handling specific to streaming datasets."""
 
 import functools
 from collections import defaultdict
@@ -17,10 +17,10 @@ from axolotl.utils.trainer import process_pretraining_datasets_for_packing
 LOG = get_logger(__name__)
 
 
-def encode_pretraining(
+def encode_streaming(
+    examples: Dict[str, List],
     tokenizer: PreTrainedTokenizerBase,
     max_tokens: int,
-    examples: Dict[str, List],
     text_column: str = "text",
     concatenate: bool = True,
 ) -> Dict[str, List]:
@@ -176,45 +176,57 @@ def encode_pretraining(
     return ret
 
 
-def wrap_pretraining_dataset(
+def wrap_streaming_dataset(
     dataset,
     tokenizer,
     cfg,
     ds_wrapper_fn,
-    max_tokens=2048,
-    batch_size=1,
-    seed=42,
-    buffer_size=10_000,
 ):
     if cfg.sample_packing:
+        # For SFT (non-pretraining) datasets, always use multipack_attn=True to ensure
+        # attention isolation between packed sequences
+        multipack_attn = (
+            True if not cfg.pretraining_dataset else cfg.pretrain_multipack_attn
+        )
+
         collate_fn = PretrainingBatchSamplerDataCollatorForSeq2Seq(
             tokenizer,
             return_tensors="pt",
             padding=True,
-            pad_to_multiple_of=max_tokens,
-            multipack_attn=cfg.pretrain_multipack_attn,
+            pad_to_multiple_of=cfg.sequence_len,
+            multipack_attn=multipack_attn,
         )
         encode = functools.partial(
-            encode_packed_pretraining,
+            encode_packed_streaming,
             collate_fn,
             ds_wrapper_fn,
-            max_seq_length=max_tokens,
-            batch_size=batch_size,
-            multipack_attn=cfg.pretrain_multipack_attn,
+            max_seq_length=cfg.sequence_len,
+            batch_size=cfg.micro_batch_size,
+            multipack_attn=multipack_attn,
         )
-        # set this to 1 so downstream data_loader doesn't try to increase the batch again
+
+        # Set this to 1 so downstream data_loader doesn't try to increase the batch size
+        # again
         cfg.micro_batch_size = 1
     else:
+        # NOTE: This is not reachable for SFT datasets since we use the pre-existing
+        # loading function for non-packed streaming datasets. Refer to
+        # _prepare_streaming_datasets in sft.py for that code path.
+        text_column = (
+            getattr(cfg.pretraining_dataset[0], "text_column", "text") or "text"
+        )
         encode = functools.partial(
-            encode_pretraining,
-            tokenizer,
-            max_tokens,
-            text_column=cfg.pretraining_dataset[0].text_column or "text",
+            encode_streaming,
+            tokenizer=tokenizer,
+            max_tokens=cfg.sequence_len,
+            text_column=text_column,
             concatenate=cfg.pretraining_sample_concatenation is True,
         )
 
     if cfg.shuffle_merged_datasets:
-        dataset = dataset.shuffle(seed=seed, buffer_size=buffer_size)
+        dataset = dataset.shuffle(
+            seed=cfg.seed, buffer_size=cfg.streaming_multipack_buffer_size
+        )
     else:
         LOG.debug("NOT shuffling merged pretraining datasets")
 
@@ -232,14 +244,13 @@ def wrap_pretraining_dataset(
     dataset = dataset.map(
         encode,
         batched=True,
-        batch_size=buffer_size,
-        # input_columns="text",
+        batch_size=cfg.streaming_multipack_buffer_size,
         remove_columns=remove_columns,
     )
     return dataset
 
 
-def encode_packed_pretraining(
+def encode_packed_streaming(
     collate_fn,
     ds_wrapper: Callable,
     examples: Dict[str, List],
@@ -274,8 +285,6 @@ def encode_packed_pretraining(
     for batch in sampler:
         for data in batch:
             features = train_dataset[data]
-            if "num_truncated_tokens" in features:
-                del features["num_truncated_tokens"]
             if "num_truncated_tokens" in features:
                 del features["num_truncated_tokens"]
             if "overflow_to_sample_mapping" in features:

src/axolotl/utils/data/utils.py

@@ -190,12 +190,21 @@ def handle_long_seq_in_dataset(
     Returns:
         Filtered dataset with long sequences removed.
     """
-    if "input_ids" not in dataset.column_names:
+    if (
+        hasattr(dataset, "column_names")
+        and dataset.column_names
+        and "input_ids" not in dataset.column_names
+    ):
         LOG.warning(
             "Dataset does not contain 'input_ids' column. Skip drop long seq. This is "
             "expected for reward modeling."
         )
         return dataset
+    elif not hasattr(dataset, "column_names") or dataset.column_names is None:
+        LOG.info(
+            "Dataset is streaming (IterableDataset), skipping long sequence handling"
+        )
+        return dataset
 
     drop_long = functools.partial(
         drop_long_seq,

src/axolotl/utils/environment.py

@@ -6,8 +6,6 @@ from importlib.metadata import version
 
 from accelerate.utils.environment import (
     check_cuda_p2p_ib_support as accelerate_check_cuda_p2p_ib_support,
-)
-from accelerate.utils.environment import (
     get_gpu_info,
 )
 from packaging.version import Version, parse

src/axolotl/utils/quantization.py

@@ -3,30 +3,47 @@ Utilities for quantization including QAT and PTQ using torchao.
 """
 
 import torch
-from torch import nn
+from packaging import version
 from torchao.core.config import AOBaseConfig
 from torchao.quantization import quantize_
 from torchao.quantization.qat import (
-    FakeQuantizeConfig,
-    FromIntXQuantizationAwareTrainingConfig,
-    IntXQuantizationAwareTrainingConfig,
+    QATConfig,
 )
 from torchao.quantization.quant_api import (
-    Int4DynamicActivationInt4WeightConfig,
-    Int4WeightOnlyConfig,
+    Float8DynamicActivationFloat8WeightConfig,
+    Float8DynamicActivationInt4WeightConfig,
     Int8DynamicActivationInt4WeightConfig,
-    Int8DynamicActivationInt8WeightConfig,
-    Int8WeightOnlyConfig,
-    UIntXWeightOnlyConfig,
-    _is_linear,
 )
 
-from axolotl.utils.schemas.enums import TorchIntDType
+from axolotl.utils.schemas.enums import TorchAOQuantDType
+
+quantization_config_to_str = {
+    Int8DynamicActivationInt4WeightConfig: "int8int4",
+    Float8DynamicActivationFloat8WeightConfig: "fp8fp8",
+    Float8DynamicActivationInt4WeightConfig: "fp8int4",
+}
+
+if version.parse(torch.__version__) >= version.parse("2.8.0"):
+    try:
+        from torchao.prototype.mx_formats import NVFP4InferenceConfig
+
+        quantization_config_to_str[NVFP4InferenceConfig] = "nvfp4"
+    except:
+        pass
+
+    # int4 weight config imports will fail on machines with fbgemm-gpu installed
+    # without a CUDA runtime available so we do this safely
+    try:
+        from torchao.quantization.quant_api import Int4WeightOnlyConfig
+
+        quantization_config_to_str[Int4WeightOnlyConfig] = "int4"
+    except:
+        pass
 
 
-def get_ptq_config(
-    weight_dtype: TorchIntDType,
-    activation_dtype: TorchIntDType | None = None,
+def get_quantization_config(
+    weight_dtype: TorchAOQuantDType,
+    activation_dtype: TorchAOQuantDType | None = None,
     group_size: int | None = None,
 ) -> AOBaseConfig:
     """
@@ -45,44 +62,101 @@ def get_ptq_config(
             or if the group size is not specified for int8 or int4 weight only quantization.
     """
     if activation_dtype is None:
-        if not weight_dtype.value.is_signed:  # type: ignore[attr-defined,union-attr]
-            return UIntXWeightOnlyConfig(
-                dtype=weight_dtype.value,
-                group_size=group_size,
-                set_inductor_config=False,
-            )
-        if weight_dtype == TorchIntDType.int8:
-            if group_size is None:
-                raise ValueError(
-                    "group_size must be specified for int8 weight only quantization"
-                )
-            return Int8WeightOnlyConfig(
-                group_size=group_size,
-            )
-        if weight_dtype == TorchIntDType.int4:
-            if group_size is None:
-                raise ValueError(
-                    "group_size must be specified for int4 weight only quantization"
-                )
-            return Int4WeightOnlyConfig(
-                group_size=group_size,
-            )
-    if activation_dtype == TorchIntDType.int4 and weight_dtype == TorchIntDType.int4:
-        return Int4DynamicActivationInt4WeightConfig()
-    if activation_dtype == TorchIntDType.int8 and weight_dtype == TorchIntDType.int8:
-        return Int8DynamicActivationInt8WeightConfig()
-    if activation_dtype == TorchIntDType.int8 and weight_dtype == TorchIntDType.int4:
-        return Int8DynamicActivationInt4WeightConfig()
+        if weight_dtype == TorchAOQuantDType.int8:
+            raise ValueError("Int8WeightOnlyConfig is not supported by torchao QAT.")
+        if weight_dtype == TorchAOQuantDType.int4:
+            from torchao.quantization.quant_api import Int4WeightOnlyConfig
+
+            if group_size is not None:
+                return Int4WeightOnlyConfig(group_size=group_size, version=2)
+            else:
+                return Int4WeightOnlyConfig(version=2)
+    if (
+        activation_dtype == TorchAOQuantDType.int4
+        and weight_dtype == TorchAOQuantDType.int4
+    ):
+        raise ValueError(
+            "Int4DynamicActivationInt4WeightConfig is not supported by torchao QAT."
+        )
+    if (
+        activation_dtype == TorchAOQuantDType.int8
+        and weight_dtype == TorchAOQuantDType.int8
+    ):
+        raise ValueError(
+            "Int8DynamicActivationInt8WeightConfig is not supported by torchao QAT."
+        )
+    if (
+        activation_dtype == TorchAOQuantDType.int8
+        and weight_dtype == TorchAOQuantDType.int4
+    ):
+        if group_size is not None:
+            return Int8DynamicActivationInt4WeightConfig(group_size=group_size)
+        else:
+            return Int8DynamicActivationInt4WeightConfig()
+    if (
+        activation_dtype == TorchAOQuantDType.float8_e4m3fn
+        and weight_dtype == TorchAOQuantDType.float8_e4m3fn
+    ):
+        return Float8DynamicActivationFloat8WeightConfig()
+    if (
+        activation_dtype == TorchAOQuantDType.float8_e4m3fn
+        and weight_dtype == TorchAOQuantDType.int4
+    ):
+        return Float8DynamicActivationInt4WeightConfig()
+    if weight_dtype == TorchAOQuantDType.nvfp4:
+        from torchao.prototype.mx_formats import NVFP4InferenceConfig
+
+        if group_size is not None and group_size != 16:
+            raise ValueError("NVFP4 quantization must use a group_size of 16")
+        return NVFP4InferenceConfig()
     raise ValueError(
         f"Invalid activation/weight dtype combination: {activation_dtype}/{weight_dtype}"
     )
 
 
+def quantize_model(
+    model,
+    weight_dtype: TorchAOQuantDType,
+    group_size: int | None = None,
+    activation_dtype: TorchAOQuantDType | None = None,
+    quantize_embedding: bool | None = None,
+):
+    """
+    This function is used to quantize a model.
+
+    Args:
+        model: The model to quantize.
+        weight_dtype: The dtype to use for weight quantization.
+        group_size: The group size to use for weight quantization.
+        activation_dtype: The dtype to use for activation quantization.
+        quantize_embedding: Whether to quantize the model's embedding weights.
+
+    """
+    linear_ptq_config = get_quantization_config(
+        weight_dtype=weight_dtype,
+        activation_dtype=activation_dtype,
+        group_size=group_size,
+    )
+    quantize_(model, linear_ptq_config)
+    if quantize_embedding:
+        # activation fake quantization is not supported for embedding layers
+        embedding_quantize_config = get_quantization_config(
+            weight_dtype=weight_dtype,
+            activation_dtype=None,
+            group_size=group_size,
+        )
+        quantize_(
+            model,
+            embedding_quantize_config,
+            filter_fn=lambda m, _: isinstance(m, torch.nn.Embedding),
+        )
+
+
 def prepare_model_for_qat(
     model,
-    weight_dtype: TorchIntDType,
-    group_size: int,
-    activation_dtype: TorchIntDType | None = None,
+    weight_dtype: TorchAOQuantDType,
+    group_size: int | None = None,
+    activation_dtype: TorchAOQuantDType | None = None,
     quantize_embedding: bool = False,
 ):
     """
@@ -100,86 +174,40 @@ def prepare_model_for_qat(
     Raises:
         ValueError: If the activation/weight dtype combination is invalid.
     """
-    if activation_dtype:
-        activation_config = FakeQuantizeConfig(
-            dtype=activation_dtype.value, granularity="per_token", is_symmetric=False
-        )
-    weight_config = FakeQuantizeConfig(dtype=weight_dtype.value, group_size=group_size)
-    linear_quantize_config = IntXQuantizationAwareTrainingConfig(
-        activation_config=None if activation_dtype is None else activation_config,
-        weight_config=weight_config,
-    )
-    quantize_(model, linear_quantize_config)
-    if quantize_embedding:
-        # activation fake quantization is not supported for embedding layers
-        embedding_quantize_config = IntXQuantizationAwareTrainingConfig(
-            activation_config=None,
-            weight_config=weight_config,
-        )
-        quantize_(
-            model,
-            embedding_quantize_config,
-            filter_fn=lambda m, _: isinstance(m, torch.nn.Embedding),
-        )
-
-
-def quantize_model_for_ptq(
-    model,
-    weight_dtype: TorchIntDType,
-    group_size: int | None = None,
-    activation_dtype: TorchIntDType | None = None,
-    quantize_embedding: bool | None = None,
-):
-    """
-    This function is used to quantize a model for post-training quantization.
-    It swaps the model's linear layers with fake quantized linear layers.
-    If `quantize_embedding` is True, it will also swap the model's embedding weights with fake quantized embedding weights.
-
-    Args:
-        model: The model to quantize.
-        weight_dtype: The dtype to use for weight quantization.
-        group_size: The group size to use for weight quantization.
-        activation_dtype: The dtype to use for activation quantization.
-        quantize_embedding: Whether to quantize the model's embedding weights.
-
-    """
-    linear_ptq_config = get_ptq_config(
+    base_config = get_quantization_config(
         weight_dtype=weight_dtype,
         activation_dtype=activation_dtype,
         group_size=group_size,
     )
-    quantize_(model, linear_ptq_config)
+    qat_config = QATConfig(base_config)
+    quantize_(model, qat_config)
     if quantize_embedding:
-        embedding_quantize_config = get_ptq_config(
+        # activation fake quantization is not supported for embedding layers
+        embedding_base_config = get_quantization_config(
             weight_dtype=weight_dtype,
             activation_dtype=None,
             group_size=group_size,
         )
+        embedding_qat_config = QATConfig(embedding_base_config)
         quantize_(
             model,
-            embedding_quantize_config,
+            embedding_qat_config,
             filter_fn=lambda m, _: isinstance(m, torch.nn.Embedding),
         )
 
 
-def convert_qat_model_for_ptq(
+def convert_qat_model(
     model,
-    *,
-    quantize_embedding: bool | None = None,
+    quantize_embedding: bool = False,
 ):
     """
-    This function is used to convert a swap fake-quantized modules in a model
-    which has been trained with QAT back to the original modules, ready for PTQ.
-
-    Args:
-        model: The model to convert.
-        quantize_embedding: Whether to quantize the model's embedding weights.
+    This function converts a QAT model which has fake quantized layers back to the original model.
     """
+    config = QATConfig(step="convert")
+    quantize_(model, config)
     if quantize_embedding:
-
-        def filter_fn(m, _):
-            return isinstance(m, nn.Embedding) or _is_linear(m)
-
-    else:
-        filter_fn = _is_linear
-    quantize_(model, FromIntXQuantizationAwareTrainingConfig(), filter_fn=filter_fn)
+        quantize_(
+            model,
+            config,
+            filter_fn=lambda m, _: isinstance(m, torch.nn.Embedding),
+        )

src/axolotl/utils/schemas/config.py

@@ -106,6 +106,12 @@ class AxolotlInputConfig(
             "description": "Don't upcast the embeddings to float32 when using PEFT. Useful for low-VRAM GPUs"
         },
     )
+    reinit_weights: bool | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Reinitialize model weights randomly instead of loading pretrained weights"
+        },
+    )
 
     trainer_cls: str | None = Field(
         default=None,
@@ -138,6 +144,12 @@ class AxolotlInputConfig(
             "description": "Process reward modelling: `True` or `False`"
         },
     )
+    center_rewards_coefficient: float | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "Coefficient to incentivize the reward model to output mean-zero rewards (proposed by https://huggingface.co/papers/2312.09244, Eq. 2). Recommended value: `0.01`."
+        },
+    )
     num_labels: int | None = None
     # Whether to use weighting in DPO trainer.
     # If `None`, default is `False` in the trainer.
@@ -475,12 +487,6 @@ class AxolotlInputConfig(
         },
     )
     multipack_real_batches: bool | None = None
-    pretraining_sample_concatenation: bool | None = Field(
-        default=None,
-        json_schema_extra={
-            "description": "whether to concatenate samples during pretraining",
-        },
-    )
 
     batch_flattening: Literal["auto"] | bool | None = Field(
         default=None,
@@ -495,13 +501,34 @@ class AxolotlInputConfig(
     pose_max_context_len: int | None = None
     pose_num_chunks: int | None = None
 
-    pretrain_multipack_buffer_size: int | None = 10_000
+    # Deprecated: Use streaming_multipack_buffer_size instead
+    pretrain_multipack_buffer_size: int | None = Field(
+        default=None,
+        deprecated="Deprecated in v0.13.0, will be removed in v0.14.0. Use streaming_multipack_buffer_size instead",
+    )
     pretrain_multipack_attn: bool | None = Field(
         default=True,
         json_schema_extra={
             "description": "whether to prevent cross attention for packed sequences during pretraining",
         },
     )
+    pretraining_sample_concatenation: bool | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": "whether to concatenate samples during pretraining",
+        },
+    )
+
+    streaming: bool | None = Field(
+        default=None,
+        json_schema_extra={"description": "Use streaming mode for loading datasets"},
+    )
+    streaming_multipack_buffer_size: int | None = Field(
+        default=10_000,
+        json_schema_extra={
+            "description": "Buffer size for multipack streaming datasets"
+        },
+    )
 
     xformers_attention: bool | None = Field(
         default=None,
@@ -830,10 +857,15 @@ class AxolotlInputConfig(
     include_tokens_per_second: bool | None = Field(
         default=None,
         json_schema_extra={
-            "description": "bool of whether to include tokens trainer per second in the training metrics. This iterates over the entire dataset once, so it takes some time."
+            "description": "bool of whether to report tokens per second at the end of training. This is not supported with pre-training datasets."
+        },
+    )
+    include_tkps: bool | None = Field(
+        default=True,
+        json_schema_extra={
+            "description": "bool of whether to report tokens per second per-gpu during training by measuring throughput of non-padding tokens."
         },
     )
-
     neftune_noise_alpha: float | None = Field(
         default=None,
         json_schema_extra={
@@ -927,7 +959,15 @@ class AxolotlInputConfig(
         },
     )
 
-    fix_untrained_tokens: int | list[int] | None = None
+    fix_untrained_tokens: int | list[int] | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": (
+                "Token index or indices to adjust embedding weights to the mean of the other tokens. "
+                "This is useful when the model has untrained embeddings."
+            )
+        },
+    )
 
     # INTERNALS - document for now, generally not set externally
     is_preprocess: bool | None = None
@@ -986,6 +1026,26 @@ class AxolotlInputConfig(
             return [ds_config.model_dump(exclude_none=True) for ds_config in ds_configs]
         return None
 
+    @model_validator(mode="before")
+    @classmethod
+    def warn_peft_trainable_token_to_fix_untrained(cls, data):
+        if (
+            peft_trainable_token_indices := data.get("peft_trainable_token_indices")
+        ) and (fix_untrained_tokens := data.get("fix_untrained_tokens")):
+            if isinstance(fix_untrained_tokens, int):
+                fix_untrained_tokens = (fix_untrained_tokens,)
+
+            if isinstance(peft_trainable_token_indices, int):
+                peft_trainable_token_indices = (peft_trainable_token_indices,)
+
+            for untrained_token_id in fix_untrained_tokens:
+                if untrained_token_id not in peft_trainable_token_indices:
+                    LOG.warning_once(
+                        f"Token {untrained_token_id} is fixed via `fix_untrained_tokens`, yet not in `peft_trainable_token_indices: ` list. "
+                        "Please add it, otherwise the token won't be trained on."
+                    )
+        return data
+
 
 class AxolotlConfigWCapabilities(AxolotlInputConfig):
     """wrapper to valdiate GPU capabilities with the configured options"""
@@ -1259,3 +1319,14 @@ class AxolotlConfigWCapabilities(AxolotlInputConfig):
             data["dataset_processes"] = get_default_process_count()
 
         return data
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_deduplication_with_streaming(cls, data):
+        if data.get("dataset_exact_deduplication") and (
+            data.get("streaming") or data.get("pretraining_dataset")
+        ):
+            raise NotImplementedError(
+                "dataset_exact_deduplication is not available for streaming datasets. "
+            )
+        return data

src/axolotl/utils/schemas/enums.py

@@ -5,18 +5,21 @@ from enum import Enum
 import torch
 
 
-class TorchIntDType(Enum):
-    """Torch integer data types - `getattr` guards against torch < 2.6 which does not support int4"""
+class TorchAOQuantDType(Enum):
+    int4 = torch.int4
+    int8 = torch.int8
+    float8_e4m3fn = torch.float8_e4m3fn
+    nvfp4 = "nvfp4"
 
-    uint1 = getattr(torch, "uint1", None)
-    uint2 = getattr(torch, "uint2", None)
-    uint3 = getattr(torch, "uint3", None)
-    uint4 = getattr(torch, "uint4", None)
-    uint5 = getattr(torch, "uint5", None)
-    uint6 = getattr(torch, "uint6", None)
-    uint7 = getattr(torch, "uint7", None)
-    int4 = getattr(torch, "int4", None)
-    int8 = getattr(torch, "int8", None)
+    def from_string(str):
+        if str == "int4":
+            return TorchAOQuantDType.int4
+        if str == "int8":
+            return TorchAOQuantDType.int8
+        if str in ["float8_e4m3fn", "fp8", "float8"]:
+            return TorchAOQuantDType.float8_e4m3fn
+        if str == "nvfp4":
+            return TorchAOQuantDType.nvfp4
 
 
 class RLType(str, Enum):

src/axolotl/utils/schemas/model.py

@@ -59,16 +59,21 @@ class ModelInputConfig(BaseModel):
     processor_type: str | None = Field(
         default=None, json_schema_extra={"description": "transformers processor class"}
     )
+    tokenizer_save_jinja_files: bool | None = Field(
+        default=True,  # match the default behavior from transformers
+        json_schema_extra={
+            "description": "Whether to save jinja files for tokenizer, transformers default is True"
+        },
+    )
     trust_remote_code: bool | None = Field(
         default=None,
         json_schema_extra={"description": "Trust remote code for untrusted source"},
     )
 
     experimental_skip_move_to_device: bool | None = Field(
-        default=None,
+        default=True,
         json_schema_extra={
-            "description": "Don't move the model to the device before sharding. "
-            "This is an experimental feature that may be included in the future as the default."
+            "description": "Don't move the model to the device before sharding. Set to `false` to revert to legacy behavior."
         },
     )
 

src/axolotl/utils/schemas/peft.py

@@ -90,6 +90,16 @@ class LoraConfig(BaseModel):
             "description": "How to initialize LoRA weights. Default to True which is MS original implementation."
         },
     )
+    peft_trainable_token_indices: list[int] | dict[str, list[int]] | None = Field(
+        default=None,
+        json_schema_extra={
+            "description": (
+                "A list of token indices to fine-tune on the `embed_tokens` layer.\n"
+                "Otherwise, a dict mapping an embedding layer name to its trainable token indices.\n"
+                "See https://huggingface.co/docs/peft/v0.17.0/en/developer_guides/lora#efficiently-train-tokens-alongside-lora"
+            )
+        },
+    )
 
     qlora_sharded_model_loading: bool | None = Field(
         default=False,

src/axolotl/utils/schemas/quantization.py

@@ -6,7 +6,23 @@ from typing import Any
 
 from pydantic import BaseModel, Field, field_validator
 
-from axolotl.utils.schemas.enums import TorchIntDType
+from axolotl.utils.schemas.enums import TorchAOQuantDType
+
+
+def validate_ao_dtype(v: Any) -> TorchAOQuantDType | None:
+    if v is None:
+        return None
+    if v == "int4":
+        return TorchAOQuantDType.int4
+    if v == "int8":
+        return TorchAOQuantDType.int8
+    if v in ["float8_e4m3fn", "fp8", "float8"]:
+        return TorchAOQuantDType.float8_e4m3fn
+    if v == "nvfp4":
+        return TorchAOQuantDType.nvfp4
+    raise ValueError(
+        f"Invalid dtype: '{v}'. Must be one of: {[e.name for e in TorchAOQuantDType] + ['fp8', 'float8']}"
+    )
 
 
 class QATConfig(BaseModel):
@@ -14,13 +30,13 @@ class QATConfig(BaseModel):
     QAT Config Schema
     """
 
-    activation_dtype: TorchIntDType | None = Field(
+    activation_dtype: TorchAOQuantDType | None = Field(
         default=None,
-        description='Fake quantization layout to use for activation quantization. Valid options are "int4" and "int8"',
+        description="Fake quantization layout to use for activation quantization.",
     )
-    weight_dtype: TorchIntDType = Field(
-        default=TorchIntDType.int8,
-        description='Fake quantization layout to use for weight quantization. Valid options are "int4" and "int8"',
+    weight_dtype: TorchAOQuantDType = Field(
+        default=TorchAOQuantDType.int8,
+        description="Fake quantization layout to use for weight quantization.",
     )
     quantize_embedding: bool | None = Field(
         default=False, description="Quantize embedding"
@@ -35,12 +51,8 @@ class QATConfig(BaseModel):
 
     @field_validator("activation_dtype", "weight_dtype", mode="before")
     @classmethod
-    def validate_dtype(cls, v: Any) -> TorchIntDType | None:
-        if v == "int4":
-            return TorchIntDType.int4
-        if v == "int8":
-            return TorchIntDType.int8
-        raise ValueError(f"Invalid dtype: '{v}'. Must be one of: ['int4', 'int8']")
+    def validate_dtype(cls, v: Any) -> TorchAOQuantDType | None:
+        return validate_ao_dtype(v)
 
 
 class PTQConfig(BaseModel):
@@ -48,13 +60,13 @@ class PTQConfig(BaseModel):
     PTQ Config Schema
     """
 
-    weight_dtype: TorchIntDType = Field(
-        default=TorchIntDType.int8,
-        description="Fake quantization layout to use for weight quantization. Valid options are uintX for X in [1, 2, 3, 4, 5, 6, 7], or int4, or int8",
+    weight_dtype: TorchAOQuantDType = Field(
+        default=TorchAOQuantDType.int8,
+        description="Fake quantization layout to use for weight quantization.",
     )
-    activation_dtype: TorchIntDType | None = Field(
+    activation_dtype: TorchAOQuantDType | None = Field(
         default=None,
-        description='Fake quantization layout to use for activation quantization. Valid options are "int4" and "int8"',
+        description="Fake quantization layout to use for activation quantization.",
     )
     quantize_embedding: bool | None = Field(
         default=None, description="Whether to quantize the embedding layer."
@@ -66,9 +78,5 @@ class PTQConfig(BaseModel):
 
     @field_validator("activation_dtype", "weight_dtype", mode="before")
     @classmethod
-    def validate_dtype(cls, v: Any) -> TorchIntDType | None:
-        if v == "int4":
-            return TorchIntDType.int4
-        if v == "int8":
-            return TorchIntDType.int8
-        raise ValueError(f"Invalid dtype: '{v}'. Must be one of: ['int4', 'int8']")
+    def validate_dtype(cls, v: Any) -> TorchAOQuantDType | None:
+        return validate_ao_dtype(v)

src/axolotl/utils/schemas/validation.py

@@ -14,7 +14,6 @@ from transformers.utils.import_utils import is_torch_npu_available
 from axolotl.utils.logging import get_logger
 from axolotl.utils.schemas.enums import ChatTemplate, RingAttnFunc, RLType
 
-
 LOG = get_logger(__name__)
 
 SUPPORTED_METRICS = {"sacrebleu", "comet", "ter", "chrf", "perplexity"}
@@ -60,6 +59,20 @@ class DatasetValidationMixin:
             raise ValueError("either datasets or pretraining_dataset is required")
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def check_pretraining_streaming_deprecation(cls, data):
+        # TODO(djsaunde): remove this check + implement change for 0.13.0 release
+        if data.get("pretraining_dataset") and not data.get("streaming"):
+            LOG.warning(
+                "Setting `pretraining_dataset` without explicitly setting `streaming: "
+                "true` is deprecated. In a future release, streaming will not be "
+                "automatically enabled when using pretraining_dataset. Please "
+                "explicitly set `streaming: true` in your configuration to maintain "
+                "current behavior."
+            )
+        return data
+
     @model_validator(mode="before")
     @classmethod
     def check_push_ds_auth(cls, data):
@@ -340,6 +353,30 @@ class TrainingValidationMixin:
             )
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def check_multipack_buffer_size(cls, data):
+        if data.get("pretrain_multipack_buffer_size") and not data.get(
+            "streaming_multipack_buffer_size"
+        ):
+            LOG.warning(
+                "`pretrain_multipack_buffer_size` is deprecated in v0.13.0, will be "
+                "removed in v0.14.0. Use `streaming_multipack_buffer_size` instead."
+            )
+            data["streaming_multipack_buffer_size"] = data[
+                "pretrain_multipack_buffer_size"
+            ]
+            del data["pretrain_multipack_buffer_size"]
+        elif data.get("pretrain_multipack_buffer_size") and data.get(
+            "streaming_multipack_buffer_size"
+        ):
+            raise ValueError(
+                "pretrain_multipack_buffer_size is deprecated, use "
+                "streaming_multipack_buffer_size; both are set, please remove the "
+                "deprecated pretrain_multipack_buffer_size setting"
+            )
+        return data
+
     @model_validator(mode="after")
     def check_fft_possible_bad_config(self):
         if (
@@ -1074,6 +1111,50 @@ class PretrainingValidationMixin:
                     data["accelerator_config"]["dispatch_batches"] = False
         return data
 
+    @model_validator(mode="before")
+    @classmethod
+    def check_pretraining_w_val_set_size(cls, data):
+        if data.get("pretraining_dataset") and data.get("val_set_size"):
+            raise ValueError(
+                "val_set_size is not supported with pretraining_dataset. "
+                "Use test_datasets to specify evaluation datasets for pretraining."
+            )
+        return data
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_streaming_w_val_set_size(cls, data):
+        if data.get("streaming") and data.get("val_set_size"):
+            raise ValueError(
+                "val_set_size is not supported with streaming datasets. "
+                "Use test_datasets to specify evaluation datasets when streaming is enabled."
+            )
+        return data
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_streaming_w_max_steps(cls, data):
+        if data.get("streaming") and not data.get("max_steps"):
+            raise ValueError(
+                "max_steps must be set when using streaming datasets. "
+                "Trainer cannot infer dataset length for iterable datasets."
+            )
+        return data
+
+    @model_validator(mode="before")
+    @classmethod
+    def check_streaming_w_multiple_datasets(cls, data):
+        if (
+            data.get("streaming")
+            and data.get("sample_packing")
+            and data.get("datasets")
+            and len(data.get("datasets")) > 1
+        ):
+            raise NotImplementedError(
+                "Sample packing with multiple streaming datasets is not yet supported"
+            )
+        return data
+
 
 class ModelCompatibilityValidationMixin:
     """Validation methods for specific model compatibility."""

src/axolotl/utils/trainer.py

@@ -475,7 +475,9 @@ def calculate_total_num_steps(cfg, train_dataset, update=True):
                 train_dataset.remove_columns(["length"]),
                 batch_sampler=sampler,
             )
-            data_loader_len = len(data_loader) * cfg.micro_batch_size // cfg.batch_size
+            data_loader_len = max(
+                1, len(data_loader) * cfg.micro_batch_size // cfg.batch_size
+            )
             LOG.debug(f"data_loader_len: {data_loader_len}")
             # FIXME: is there a bug here somewhere? the total num steps depends
             # on the agreed on value for sample_packing_eff_est
@@ -547,6 +549,13 @@ def setup_deepspeed_env(cfg, stage=None):
         if stage == 3:
             os.environ["ACCELERATE_DEEPSPEED_ZERO3_INIT"] = "true"
 
+    device_count = torch.cuda.device_count()
+    if device_count == 1:
+        os.environ.setdefault("WORLD_SIZE", "1")
+        os.environ.setdefault("LOCAL_RANK", "0")
+        os.environ.setdefault("MASTER_ADDR", "0.0.0.0")  # nosec B104
+        os.environ.setdefault("MASTER_PORT", "29500")
+
     # NOTE(djsaunde): The distribued state cannot be initialized prior to the
     # ACCELERATE_USE_DEEPSPEED assignment, but it must be initialized some time prior
     # to model load.

tests/e2e/integrations/test_kd.py

@@ -25,7 +25,7 @@ def min_cfg(temp_dir):
         "liger_rms_norm": True,
         "liger_glu_activation": True,
         "torch_compile": True,
-        "chat_template": "llama3",
+        "chat_template": "qwen3",
         "kd_trainer": True,
         "kd_ce_alpha": 0.1,
         "kd_alpha": 0.9,

tests/e2e/test_diffusion.py

@@ -0,0 +1,139 @@
+"""E2E smoke test for diffusion training plugin."""
+
+from axolotl.common.datasets import load_datasets
+from axolotl.train import train
+from axolotl.utils.config import normalize_config, validate_config
+from axolotl.utils.dict import DictDefault
+
+from tests.e2e.utils import check_model_output_exists
+
+
+class TestDiffusion:
+    """Test case for diffusion training plugin."""
+
+    def test_diffusion_smoke_test(self, temp_dir):
+        """
+        Smoke test for diffusion training to ensure the plugin loads and trains without
+        error.
+        """
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "tokenizer_type": "AutoTokenizer",
+                "trust_remote_code": True,
+                "sequence_len": 256,
+                "val_set_size": 0.1,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 3,
+                "micro_batch_size": 1,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.0001,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+                "bf16": True,
+                "save_safetensors": True,
+                "save_first_step": False,
+                "logging_steps": 1,
+                "eval_steps": 3,
+                # Diffusion-specific config
+                "plugins": ["axolotl.integrations.diffusion.DiffusionPlugin"],
+                "diffusion": {
+                    # sample generation
+                    "generate_samples": True,
+                    "generation_interval": 1,
+                    "num_generation_samples": 1,
+                    "generation_steps": 2,
+                    "generation_max_length": 32,
+                    "generation_temperature": 0.0,
+                    # training-specific
+                    "mask_token_id": 16,
+                    "eps": 1e-3,
+                    "importance_weighting": False,
+                },
+            }
+        )
+
+        cfg = validate_config(cfg)
+        normalize_config(cfg)
+        dataset_meta = load_datasets(cfg=cfg)
+
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)
+
+    def test_diffusion_sft_labels(self, temp_dir):
+        """Test that diffusion training properly handles SFT data with labels."""
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "tokenizer_type": "AutoTokenizer",
+                "trust_remote_code": True,
+                "sequence_len": 256,
+                "val_set_size": 0.1,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 3,
+                "micro_batch_size": 1,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.0001,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+                "bf16": True,
+                "save_safetensors": True,
+                "save_first_step": False,
+                "logging_steps": 1,
+                "eval_steps": 2,
+                # Diffusion-specific config
+                "plugins": ["axolotl.integrations.diffusion.DiffusionPlugin"],
+                "diffusion": {
+                    # sample generation
+                    "generate_samples": True,
+                    "generation_interval": 1,
+                    "num_generation_samples": 1,
+                    "generation_steps": 2,
+                    "generation_max_length": 32,
+                    "generation_temperature": 0.0,
+                    # training-specific
+                    "mask_token_id": 16,
+                    "eps": 1e-3,
+                    "importance_weighting": True,
+                },
+                # Ensure we have proper SFT labels
+                "train_on_inputs": False,
+            }
+        )
+
+        cfg = validate_config(cfg)
+        normalize_config(cfg)
+        dataset_meta = load_datasets(cfg=cfg)
+
+        # Verify that the dataset has labels
+        sample = dataset_meta.train_dataset[0]
+        assert "labels" in sample, "SFT dataset should have labels"
+
+        # Check that some labels are -100 (prompt tokens)
+        labels = sample["labels"]
+        if hasattr(labels, "tolist"):
+            labels = labels.tolist()
+        assert -100 in labels, "SFT dataset should have -100 labels for prompt tokens"
+
+        train(cfg=cfg, dataset_meta=dataset_meta)
+        check_model_output_exists(temp_dir, cfg)