fix losses

.github/CONTRIBUTING.md

@@ -31,11 +31,10 @@ PRs are **greatly welcome**!
 
 Please run below to setup env
 ```bash
-# Install axolotl + dev and test dependencies
+# Install axolotl + dev and test dependencies from lockfile
 export UV_TORCH_BACKEND=cu128  # or cu130
-uv venv --no-project --relocatable
+uv sync --extra flash-attn --extra deepspeed --group dev --group test
 source .venv/bin/activate
-uv pip install --no-build-isolation -e '.[deepspeed]' --group dev --group test
 pre-commit install
 
 # test

.github/workflows/base.yml

@@ -30,6 +30,14 @@ jobs:
       fail-fast: false
       matrix:
         include:
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.9.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-base"
+            platforms: "linux/amd64,linux/arm64"
           - cuda: "128"
             cuda_version: 12.8.1
             cudnn_version: ""
@@ -160,6 +168,14 @@ jobs:
             torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
             dockerfile: "Dockerfile-uv-base"
             platforms: "linux/amd64,linux/arm64"
+          - cuda: "128"
+            cuda_version: 12.8.1
+            cudnn_version: ""
+            python_version: "3.11"
+            pytorch: 2.9.0
+            torch_cuda_arch_list: "7.0 7.5 8.0 8.6 8.7 8.9 9.0+PTX"
+            dockerfile: "Dockerfile-uv-base"
+            platforms: "linux/amd64,linux/arm64"
           - cuda: "128"
             cuda_version: 12.8.1
             cudnn_version: ""

.github/workflows/main.yml

@@ -18,6 +18,12 @@ jobs:
       fail-fast: false
       matrix:
         include:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.9.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"
@@ -174,6 +180,12 @@ jobs:
       fail-fast: false
       matrix:
         include:
+          - cuda: 128
+            cuda_version: 12.8.1
+            python_version: "3.11"
+            pytorch: 2.9.0
+            axolotl_extras:
+            platforms: "linux/amd64,linux/arm64"
           - cuda: 128
             cuda_version: 12.8.1
             python_version: "3.11"

VERSION

@@ -1 +1 @@
-0.16.2.dev0
+0.16.0.dev0

docker/Dockerfile

@@ -24,9 +24,9 @@ WORKDIR /workspace/axolotl
 # If AXOLOTL_EXTRAS is set, append it in brackets; don't install deepspeed with arm64
 RUN pip uninstall -y causal_conv1d
 RUN if [ "$TARGETARCH" = "arm64" ]; then \
-        BASE_EXTRAS="optimizers,ray"; \
+        BASE_EXTRAS="flash-attn,ring-flash-attn,optimizers,ray"; \
     else \
-        BASE_EXTRAS="deepspeed,optimizers,ray"; \
+        BASE_EXTRAS="deepspeed,flash-attn,ring-flash-attn,optimizers,ray"; \
     fi && \
     if [ "$AXOLOTL_EXTRAS" != "" ]; then \
         pip install --no-build-isolation -e .[$BASE_EXTRAS,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \

docker/Dockerfile-base

@@ -58,3 +58,19 @@ RUN git lfs install --skip-repo && \
     # The base image ships with `pydantic==1.8.2` which is not working
     pip3 install -U --no-cache-dir pydantic==1.10.10 && \
     pip3 cache purge
+
+# Map Python version (e.g., 3.12 -> cp312)
+RUN PYTHON_CP="cp$(echo $PYTHON_VERSION | tr -d '.')" && \
+    # Map PyTorch version (e.g., 2.9.1 -> torch2.9, 2.10.0 -> torch2.10)
+    TORCH_TAG="torch$(echo $PYTORCH_VERSION | grep -oP '^\d+\.\d+')" && \
+    # Map architecture
+    case "$TARGETARCH" in \
+        amd64) ARCH_TAG="x86_64" ;; \
+        arm64) ARCH_TAG="aarch64" ;; \
+        *) echo "Unsupported architecture: $TARGETARCH"; exit 1 ;; \
+    esac && \
+    WHL_VERSION="v0.7.16" && \
+    WHL_FILE="flash_attn-2.8.3+cu${CUDA}${TORCH_TAG}-${PYTHON_CP}-${PYTHON_CP}-linux_${ARCH_TAG}.whl" && \
+    wget -nv "https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/${WHL_VERSION}/${WHL_FILE}" && \
+    pip3 install --no-cache-dir "${WHL_FILE}" && \
+    rm "${WHL_FILE}"

docker/Dockerfile-tests

@@ -24,9 +24,9 @@ RUN git fetch origin +$GITHUB_REF && \
 
 # If AXOLOTL_EXTRAS is set, append it in brackets
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
-        pip install --no-build-isolation -e .[deepspeed,mamba-ssm,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,mamba-ssm,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \
-        pip install --no-build-isolation -e .[deepspeed,mamba-ssm] $AXOLOTL_ARGS; \
+        pip install --no-build-isolation -e .[deepspeed,flash-attn,mamba-ssm] $AXOLOTL_ARGS; \
     fi
 
 # So we can test the Docker image

docker/Dockerfile-uv

@@ -24,9 +24,9 @@ WORKDIR /workspace/axolotl
 # If AXOLOTL_EXTRAS is set, append it in brackets; don't install deepspeed with arm64
 RUN uv pip uninstall causal_conv1d
 RUN if [ "$TARGETARCH" = "arm64" ]; then \
-        BASE_EXTRAS="optimizers,ray"; \
+        BASE_EXTRAS="flash-attn,ring-flash-attn,optimizers,ray"; \
     else \
-        BASE_EXTRAS="deepspeed,optimizers,ray"; \
+        BASE_EXTRAS="deepspeed,flash-attn,ring-flash-attn,optimizers,ray"; \
     fi && \
     if [ "$AXOLOTL_EXTRAS" != "" ]; then \
         uv pip install --no-build-isolation -e .[$BASE_EXTRAS,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \

docker/Dockerfile-uv-base

@@ -38,3 +38,20 @@ RUN uv pip install packaging setuptools wheel psutil \
 RUN if [ "$TARGETARCH" = "amd64" ]; then \
         MAMBA_SKIP_CUDA_BUILD=TRUE CAUSAL_CONV1D_SKIP_CUDA_BUILD=TRUE uv pip install --no-build-isolation mamba_ssm causal_conv1d; \
     fi
+
+# Map Python version (e.g., 3.12 -> cp312)
+RUN PYTHON_CP="cp$(echo $PYTHON_VERSION | tr -d '.')" && \
+    # Map PyTorch version (e.g., 2.9.1 -> torch2.9, 2.10.0 -> torch2.10)
+    TORCH_TAG="torch$(echo $PYTORCH_VERSION | grep -oP '^\d+\.\d+')" && \
+    LINUX_TAG="manylinux_" && \
+    # Map architecture
+    case "$TARGETARCH" in \
+        amd64) ARCH_TAG="2_24_x86_64.manylinux_2_28_x86_64" ;; \
+        arm64) ARCH_TAG="2_34_aarch64" ;; \
+        *) echo "Unsupported architecture: $TARGETARCH"; exit 1 ;; \
+    esac && \
+    WHL_VERSION="v0.7.16" && \
+    WHL_FILE="flash_attn-2.8.3+cu${CUDA}${TORCH_TAG}-${PYTHON_CP}-${PYTHON_CP}-${LINUX_TAG}${ARCH_TAG}.whl" && \
+    wget -nv "https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/download/${WHL_VERSION}/${WHL_FILE}" && \
+    uv pip install --no-cache-dir "${WHL_FILE}" && \
+    rm "${WHL_FILE}"

docs/debugging.qmd

@@ -77,9 +77,8 @@ Make sure you have an [editable install](https://setuptools.pypa.io/en/latest/us
 
 ```bash
 export UV_TORCH_BACKEND=cu128  # or cu130
-uv venv --no-project --relocatable
+uv sync --extra flash-attn --extra deepspeed --group dev --group test
 source .venv/bin/activate
-uv pip install --no-build-isolation -e '.[deepspeed]' --group dev --group test
 ```
 
 #### Remote Hosts
@@ -219,9 +218,8 @@ docker run --privileged --gpus '"all"' --shm-size 10g --rm -it --name axolotl --
 You will now be in the container.  Next, install Axolotl with dev dependencies:
 
 ```bash
-uv venv --no-project --relocatable
+uv sync --extra flash-attn --extra deepspeed --group dev --group test
 source .venv/bin/activate
-uv pip install --no-build-isolation -e '.[deepspeed]' --group dev --group test
 ```
 
 ### Attach To Container

docs/docker.qmd

@@ -10,16 +10,13 @@ This section describes the different Docker images that are released by AxolotlA
 [Docker Hub](https://hub.docker.com/u/axolotlai).
 
 ::: {.callout-important}
-### Switch to the `-uv` images
+For Blackwell GPUs, please use the tags with PyTorch 2.9.1 and CUDA 12.8.
+:::
 
-Each image below ships a **uv variant** that uses [uv](https://docs.astral.sh/uv/) with a relocatable venv
+::: {.callout-tip}
-(`/workspace/axolotl-venv`) instead of Miniconda + pip. Append `-uv` to the image name
+Each image below is available in a **uv variant** that uses [uv](https://docs.astral.sh/uv/) with
-(e.g. `axolotlai/axolotl-uv`, `axolotlai/axolotl-base-uv`, `axolotlai/axolotl-cloud-uv`). Tags follow the
+a relocatable venv (`/workspace/axolotl-venv`) instead of Miniconda + pip. Append `-uv` to the image name
-same format as their non-uv counterparts.
+(e.g. `axolotlai/axolotl-base-uv`). Tags follow the same format. We recommend the uv images for new deployments.
-
-**We recommend switching to the `-uv` images early.** In the near future we will publish the uv-based
-build to the non-uv tags as well. The non-uv names will continue to work, but they will start serving
-the uv image.
 :::
 
 ## Base
@@ -88,7 +85,7 @@ Tags examples:
 - `main-py3.12-cu130-2.10.0`
 - `main-latest`
 - `main-20260315-py3.11-cu128-2.9.1`
-- `0.16.1`
+- `0.12.0`
 
 ## Cloud
 

docs/faq.qmd

@@ -57,7 +57,7 @@ description: Frequently asked questions
 
 **Q: vLLM is not working with Axolotl**
 
-> A: We currently recommend torch 2.10 for use with `vllm`. Please ensure you use the right version. For Docker, please use the `main-py3.12-cu128-2.10.0` tag (note: torch 2.10 images are built with Python 3.12).
+> A: We currently recommend torch 2.6.0 for use with `vllm`. Please ensure you use the right version. For Docker, please use the `main-py3.11-cu124-2.6.0` tag.
 
 **Q: FA2 2.8.0 `undefined symbol` runtime error on CUDA 12.4**
 

docs/installation.qmd

@@ -15,7 +15,7 @@ This guide covers all the ways you can install and set up Axolotl for your envir
 
 - NVIDIA GPU (Ampere architecture or newer for `bf16` and Flash Attention) or AMD GPU
 - Python ≥3.11
-- PyTorch ≥2.9.1
+- PyTorch ≥2.9.0
 
 ## Installation {#sec-installation}
 
@@ -36,9 +36,9 @@ source $HOME/.local/bin/env
 Choose your CUDA version (e.g. `cu128`, `cu130`), create a venv, and install:
 ```{.bash}
 export UV_TORCH_BACKEND=cu128  # or cu130
-uv venv
+uv venv --no-project --relocatable
 source .venv/bin/activate
-uv pip install --no-build-isolation axolotl[deepspeed]
+uv pip install --no-build-isolation axolotl[flash-attn,deepspeed]
 ```
 
 ### Edge/Development Build {#sec-edge-build}
@@ -49,11 +49,12 @@ For the latest features between releases:
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
 export UV_TORCH_BACKEND=cu128  # or cu130
-uv venv
+uv sync --extra flash-attn --extra deepspeed
 source .venv/bin/activate
-uv pip install --no-build-isolation -e '.[deepspeed]'
 ```
 
+`uv sync` creates a `.venv`, installs exact pinned versions from `uv.lock`, and sets up an editable install automatically.
+
 ### Docker {#sec-docker}
 
 ```{.bash}
@@ -131,11 +132,11 @@ source $HOME/.local/bin/env
 
 # Create a fresh venv (recommended for a clean start)
 export UV_TORCH_BACKEND=cu128  # or cu130
-uv venv
+uv venv --no-project --relocatable
 source .venv/bin/activate
 
 # Reinstall axolotl
-uv pip install --no-build-isolation axolotl[deepspeed]
+uv pip install --no-build-isolation axolotl[flash-attn,deepspeed]
 ```
 
 ## Using pip (Alternative) {#sec-pip}
@@ -150,13 +151,13 @@ Follow the instructions at: [https://pytorch.org/get-started/locally/](https://p
 
 ```{.bash}
 pip3 install -U packaging setuptools wheel ninja
-pip3 install --no-build-isolation axolotl[deepspeed]
+pip3 install --no-build-isolation axolotl[flash-attn,deepspeed]
 ```
 
 For editable/development installs:
 ```{.bash}
 pip3 install -U packaging setuptools wheel ninja
-pip3 install --no-build-isolation -e '.[deepspeed]'
+pip3 install --no-build-isolation -e '.[flash-attn,deepspeed]'
 ```
 
 ## Troubleshooting {#sec-troubleshooting}

examples/LiquidAI/README.md

@@ -15,7 +15,7 @@ Thanks to the team at LiquidAI for giving us early access to prepare for these r
     Here is an example of how to install from pip:
     ```bash
     # Ensure you have a compatible version of Pytorch installed
-    uv pip install --no-build-isolation 'axolotl>=0.16.1'
+    uv pip install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
     ```
 
 2.  Run one of the finetuning examples below.

examples/apertus/README.md

@@ -11,11 +11,11 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
     Here is an example of how to install from main for pip:
 
 ```bash
-# Ensure you have Pytorch installed (Pytorch 2.9.1 min)
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
 
-uv pip install --no-build-isolation -e '.'
+uv pip 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

examples/arcee/README.md

@@ -13,11 +13,11 @@ Thanks to the team at Arcee.ai for using Axolotl in supervised fine-tuning the A
     Here is an example of how to install from main for pip:
 
 ```bash
-# Ensure you have Pytorch installed (Pytorch 2.9.1 min)
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
 
-uv pip install --no-build-isolation -e '.'
+uv pip 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

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

@@ -36,7 +36,12 @@
     "id": "msOCO4NRmRLa"
    },
    "outputs": [],
-   "source": "%%capture\n# This step can take ~5-10 minutes to install dependencies\n!pip install --no-build-isolation \"axolotl>=0.16.1\"\n!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@fec1a88\""
+   "source": [
+    "%%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@fec1a88\""
+   ]
   },
   {
    "cell_type": "markdown",

examples/devstral/README.md

@@ -15,8 +15,8 @@ Thanks to the team at MistralAI for giving us early access to prepare for this r
     Here is an example of how to install from pip:
 
 ```bash
-# Ensure you have Pytorch installed (Pytorch 2.9.1 min)
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-uv pip install --no-build-isolation 'axolotl>=0.16.1'
+uv pip install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage

examples/gemma3n/README.md

@@ -9,8 +9,8 @@ Gemma-3n is a family of multimodal models from Google found on [HuggingFace](htt
     Here is an example of how to install from pip:
 
 ```bash
-# Ensure you have Pytorch installed (Pytorch 2.9.1 min)
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-uv pip install --no-build-isolation 'axolotl>=0.16.1'
+uv pip install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. In addition to Axolotl's requirements, Gemma-3n requires:

examples/gpt-oss/README.md

@@ -13,8 +13,8 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
     Here is an example of how to install from pip:
 
 ```bash
-# Ensure you have Pytorch installed (Pytorch 2.9.1 min)
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-uv pip install --no-build-isolation 'axolotl>=0.16.1'
+uv pip install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. Choose one of the following configs below for training the 20B model. (for 120B, see [below](#training-120b))

examples/granite4/README.md

@@ -11,11 +11,11 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
     Here is an example of how to install from main for pip:
 
 ```bash
-# Ensure you have Pytorch installed (Pytorch 2.9.1 min)
+# Ensure you have Pytorch installed (Pytorch 2.7.1 min)
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
 
-uv pip install --no-build-isolation -e '.'
+uv pip 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

examples/hunyuan/README.md

@@ -9,11 +9,11 @@ Tencent released a family of opensource models called HunYuan with varying param
     Here is an example of how to install from main for pip:
 
 ```bash
-# Ensure you have Pytorch installed (Pytorch 2.9.1 min)
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
 git clone https://github.com/axolotl-ai-cloud/axolotl.git
 cd axolotl
 
-uv pip install --no-build-isolation -e '.'
+uv pip 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

examples/magistral/README.md

@@ -13,8 +13,8 @@ Thanks to the team at MistralAI for giving us early access to prepare for these
     Here is an example of how to install from pip:
 
 ```bash
-# Ensure you have Pytorch installed (Pytorch 2.9.1 min)
+# Ensure you have Pytorch installed (Pytorch 2.7.0 min)
-uv pip install --no-build-isolation 'axolotl>=0.16.1'
+uv pip install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage

examples/seed-oss/README.md

@@ -11,7 +11,7 @@ This guide shows how to fine-tune it with Axolotl with multi-turn conversations
     Here is an example of how to install from pip:
     ```bash
     # Ensure you have a compatible version of Pytorch installed
-    uv pip install --no-build-isolation 'axolotl>=0.16.1'
+    uv pip install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 
     # Install Cut Cross Entropy
     python scripts/cutcrossentropy_install.py | sh

examples/smolvlm2/README.md

@@ -13,7 +13,7 @@ This guide shows how to fine-tune SmolVLM2 models with Axolotl.
     Here is an example of how to install from pip:
     ```bash
     # Ensure you have a compatible version of Pytorch installed
-    uv pip install --no-build-isolation 'axolotl>=0.16.1'
+    uv pip install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
     ```
 
 2. Install an extra dependency:

examples/voxtral/README.md

@@ -11,8 +11,8 @@ Thanks to the team at MistralAI for giving us early access to prepare for this r
     Here is an example of how to install from pip:
 
 ```bash
-# Ensure you have Pytorch installed (Pytorch 2.9.1 min)
+# Ensure you have Pytorch installed (Pytorch 2.6.0 min)
-uv pip install --no-build-isolation 'axolotl>=0.16.1'
+uv pip install --no-build-isolation 'axolotl[flash-attn]>=0.12.0'
 ```
 
 2. Please install the below.

pyproject.toml

@@ -12,7 +12,7 @@ requires-python = ">=3.10"
 
 dependencies = [
     # Core ML stack
-    "torch>=2.9.1",
+    "torch>=2.6.0",
     "packaging==26.0",
     "huggingface_hub>=1.1.7",
     "peft>=0.19.1,<0.20.0",
@@ -79,7 +79,7 @@ dependencies = [
     # Platform-specific (Linux only)
     "bitsandbytes==0.49.1 ; sys_platform != 'darwin'",
     "triton>=3.4.0 ; sys_platform != 'darwin'",
-    "xformers>=0.0.33.post2 ; sys_platform != 'darwin' and platform_machine != 'aarch64'",
+    "xformers>=0.0.23.post1 ; sys_platform != 'darwin'",
     "liger-kernel==0.7.0 ; sys_platform != 'darwin'",
     "torchao==0.17.0 ; sys_platform != 'darwin' and platform_machine != 'aarch64'",
 

src/axolotl/core/builders/causal.py

@@ -370,7 +370,7 @@ class HFCausalTrainerBuilder(TrainerBuilderBase):
         data_collator_kwargs = {
             "padding": True,  # True/"longest" is the default
         }
-        multiple = getattr(self.cfg, "pad_to_multiple_of", None) or 64
+        multiple = 64
         if self.cfg.pad_to_sequence_len:
             data_collator_kwargs["pad_to_multiple_of"] = multiple * math.ceil(
                 self.cfg.sequence_len / multiple

src/axolotl/core/builders/rl.py

@@ -228,47 +228,9 @@ class HFRLTrainerBuilder(TrainerBuilderBase):
 
         return training_args, trainer_kwargs
 
-    def build_collator(self, **kwargs):
-        """Build a data collator for preference-tuning trainers.
-
-        Returns None for RL types that provide their own collator (e.g. GRPO,
-        KTO), letting the trainer construct its default. For DPO/IPO/ORPO/SIMPO
-        returns an ``AxolotlDPODataCollatorWithPadding`` when
-        ``pad_to_multiple_of`` is set, otherwise None (so the trainer
-        falls back to the TRL default).
-        """
-        if self.cfg.rl not in (
-            RLType.DPO,
-            RLType.IPO,
-            RLType.ORPO,
-            RLType.SIMPO,
-        ):
-            return None
-
-        pad_to_multiple_of = getattr(self.cfg, "pad_to_multiple_of", None)
-        if not pad_to_multiple_of:
-            return None
-
-        from axolotl.utils.collators.dpo import AxolotlDPODataCollatorWithPadding
-
-        LOG.info(
-            f"Using AxolotlDPODataCollatorWithPadding with pad_to_multiple_of="
-            f"{pad_to_multiple_of}"
-        )
-        is_enc_dec = getattr(self.model.config, "is_encoder_decoder", False)
-        return AxolotlDPODataCollatorWithPadding(
-            pad_token_id=self.tokenizer.pad_token_id,
-            is_encoder_decoder=is_enc_dec,
-            pad_to_multiple_of=pad_to_multiple_of,
-            **kwargs,
-        )
-
     def build(self, total_num_steps):
         training_args, trainer_kwargs = self._build_training_arguments(total_num_steps)
 
-        if (data_collator := self.build_collator()) is not None:
-            trainer_kwargs["data_collator"] = data_collator
-
         if self.eval_dataset:
             trainer_kwargs["eval_dataset"] = self.eval_dataset
         if (

src/axolotl/integrations/kd/README.md

@@ -11,7 +11,7 @@ kd_ce_alpha: 0.1
 kd_alpha: 0.9
 kd_temperature: 1.0
 
-torch_compile: True  # recommended to reduce vram
+torch_compile: True  # torch>=2.6.0, recommended to reduce vram
 
 datasets:
   - path: ...

src/axolotl/monkeypatch/trainer/utils.py

@@ -407,10 +407,7 @@ def selective_log_softmax(logits, index) -> torch.Tensor:
     K = index.shape[-1]
     original_index_shape = index.shape
 
-    try:
+    flat_logits = logits.reshape(-1, V).contiguous()
-        flat_logits = logits.view(-1, V)
-    except RuntimeError:
-        flat_logits = logits.reshape(-1, V).contiguous()
     flat_index = index.reshape(-1, K).contiguous()
 
     BLOCK_V = 4096

src/axolotl/utils/collators/__init__.py

@@ -6,7 +6,6 @@ from .batching import (
     PretrainingBatchSamplerDataCollatorForSeq2Seq,
     V2BatchSamplerDataCollatorForSeq2Seq,
 )
-from .dpo import AxolotlDPODataCollatorWithPadding
 from .mamba import MambaDataCollator
 
 __all__ = [
@@ -14,6 +13,5 @@ __all__ = [
     "BatchSamplerDataCollatorForSeq2Seq",
     "V2BatchSamplerDataCollatorForSeq2Seq",
     "PretrainingBatchSamplerDataCollatorForSeq2Seq",
-    "AxolotlDPODataCollatorWithPadding",
     "MambaDataCollator",
 ]

src/axolotl/utils/collators/dpo.py

@@ -1,128 +0,0 @@
-"""DPO/ORPO/IPO/KTO data collator with pad_to_multiple_of support.
-
-Extends TRL's DPODataCollatorWithPadding to round padded sequence lengths
-up to a fixed multiple. This stabilizes Triton autotune caches for kernels
-that key on sequence length (e.g. fla's linear attention kernels used by
-Qwen3.5), which otherwise re-autotune on every distinct batch length.
-"""
-
-from __future__ import annotations
-
-from dataclasses import dataclass
-from typing import Any
-
-import torch
-from torch.nn.utils.rnn import pad_sequence
-from trl.experimental.utils import DPODataCollatorWithPadding
-from trl.trainer.utils import pad
-
-
-def _round_up(length: int, multiple: int) -> int:
-    return ((length + multiple - 1) // multiple) * multiple
-
-
-@dataclass
-class AxolotlDPODataCollatorWithPadding(DPODataCollatorWithPadding):
-    """DPO data collator that pads to a multiple of ``pad_to_multiple_of``.
-
-    Args:
-        pad_token_id: Tokenizer pad token id (inherited).
-        is_encoder_decoder: Whether the model is encoder-decoder (inherited).
-        pad_to_multiple_of: If set, padded lengths are rounded up to this
-            multiple. Helps stabilize Triton autotune caches.
-    """
-
-    pad_to_multiple_of: int | None = None
-
-    def __call__(self, features: list[dict[str, Any]]) -> dict[str, Any]:
-        pad_to_mult = self.pad_to_multiple_of
-
-        padded_batch: dict[str, Any] = {}
-        for k in features[0].keys():
-            if k.endswith(
-                ("_input_ids", "_attention_mask", "_labels", "_pixel_values")
-            ):
-                if self.is_encoder_decoder:
-                    if k.endswith("_pixel_values"):
-                        to_pad = [
-                            torch.tensor(ex[k], dtype=torch.float32) for ex in features
-                        ]
-                    else:
-                        to_pad = [torch.LongTensor(ex[k]) for ex in features]
-
-                    if k.startswith("prompt") and k.endswith("input_ids"):
-                        if self.pad_token_id is None:
-                            raise ValueError(
-                                "Padding is enabled, but the tokenizer is not configured with a padding token."
-                            )
-                        padding_value = self.pad_token_id
-                    elif k.endswith("_attention_mask"):
-                        padding_value = 0
-                    elif k.endswith("_pixel_values"):
-                        padding_value = 0
-                    elif (
-                        k.startswith(("chosen", "rejected", "completion"))
-                        or "decoder" in k
-                    ):
-                        padding_value = -100
-                    else:
-                        raise ValueError(f"Unexpected key in batch '{k}'")
-
-                    padded = pad_sequence(
-                        to_pad, batch_first=True, padding_value=padding_value
-                    )
-                    if pad_to_mult:
-                        cur = padded.shape[1]
-                        target = _round_up(cur, pad_to_mult)
-                        if target > cur:
-                            extra = target - cur
-                            pad_shape = list(padded.shape)
-                            pad_shape[1] = extra
-                            filler = torch.full(
-                                pad_shape,
-                                padding_value,
-                                dtype=padded.dtype,
-                                device=padded.device,
-                            )
-                            padded = torch.cat([padded, filler], dim=1)
-                    padded_batch[k] = padded
-                else:
-                    if k.endswith("_input_ids"):
-                        if self.pad_token_id is None:
-                            raise ValueError(
-                                "Padding is enabled, but the tokenizer is not configured with a padding token."
-                            )
-                        padding_value = self.pad_token_id
-                    elif k.endswith("_labels"):
-                        padding_value = -100
-                    elif k.endswith("_attention_mask"):
-                        padding_value = 0
-                    elif k.endswith("_pixel_values"):
-                        padding_value = 0
-                    else:
-                        raise ValueError(f"Unexpected key in batch '{k}'")
-
-                    padding_side = (
-                        "left"
-                        if k in ("prompt_input_ids", "prompt_attention_mask")
-                        else "right"
-                    )
-
-                    dtype = (
-                        torch.float32 if k.endswith("_pixel_values") else torch.int64
-                    )
-                    to_pad = [torch.tensor(ex[k], dtype=dtype) for ex in features]
-
-                    # trl.pad() natively supports pad_to_multiple_of
-                    padded_batch[k] = pad(
-                        to_pad,
-                        padding_value=padding_value,
-                        padding_side=padding_side,
-                        pad_to_multiple_of=pad_to_mult,
-                    )
-            elif k.endswith("_logps"):
-                padded_batch[k] = torch.tensor([ex[k] for ex in features])
-            else:
-                padded_batch[k] = [ex[k] for ex in features]
-
-        return padded_batch

src/axolotl/utils/schemas/config.py

@@ -673,12 +673,6 @@ class AxolotlInputConfig(
             "description": "Pad inputs so each step uses constant sized buffers. This will reduce memory fragmentation and may prevent OOMs, by re-using memory more efficiently. Defaults to True if `sample_packing` enabled"
         },
     )
-    pad_to_multiple_of: int | None = Field(
-        default=None,
-        json_schema_extra={
-            "description": ("Pad each batch to a multiple of this value.")
-        },
-    )
     curriculum_sampling: bool | None = Field(
         default=None,
         json_schema_extra={
@@ -1016,7 +1010,7 @@ class AxolotlInputConfig(
     torch_compile: Literal["auto"] | bool | None = Field(
         default=None,
         json_schema_extra={
-            "description": "Whether to use torch.compile and which backend to use."
+            "description": "Whether to use torch.compile and which backend to use. setting to `auto` will enable torch compile when torch>=2.6.0"
         },
     )
     torch_compile_backend: str | None = Field(

tests/conftest.py

@@ -119,15 +119,49 @@ def download_smollm2_135m_gptq_model():
 
 
 @pytest.fixture(scope="session", autouse=True)
-def download_qwen_2_5_half_billion_model():
+def download_qwen3_half_billion_model():
-    # download the model
+    # download the model (still used as the KD teacher in tests/e2e/integrations/test_kd.py)
-    snapshot_download_w_retry("Qwen/Qwen2.5-0.5B", repo_type="model")
+    snapshot_download_w_retry("Qwen/Qwen3-0.6B", repo_type="model")
 
 
 @pytest.fixture(scope="session", autouse=True)
-def download_qwen3_half_billion_model():
+def download_tiny_llama_model():
-    # download the model
+    snapshot_download_w_retry("axolotl-ai-co/tiny-llama-50m", repo_type="model")
-    snapshot_download_w_retry("Qwen/Qwen3-0.6B", repo_type="model")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_tiny_mistral_model():
+    snapshot_download_w_retry("axolotl-ai-co/tiny-mistral-25m", repo_type="model")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_tiny_mixtral_model():
+    snapshot_download_w_retry("axolotl-ai-co/tiny-mixtral-30m", repo_type="model")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_tiny_phi_model():
+    snapshot_download_w_retry("axolotl-ai-co/tiny-phi-64m", repo_type="model")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_tiny_falcon_model():
+    snapshot_download_w_retry("axolotl-ai-co/tiny-falcon-42m", repo_type="model")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_tiny_qwen2_model():
+    snapshot_download_w_retry("axolotl-ai-co/tiny-qwen2-129m", repo_type="model")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_tiny_qwen3_model():
+    snapshot_download_w_retry("axolotl-ai-co/tiny-qwen3-129m", repo_type="model")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_tiny_gemma2_model():
+    snapshot_download_w_retry("axolotl-ai-co/tiny-gemma2-137m", repo_type="model")
 
 
 @pytest.fixture(scope="session", autouse=True)
@@ -620,7 +654,15 @@ def fixture_min_base_cfg():
 )
 def test_load_fixtures(
     download_smollm2_135m_model,
-    download_qwen_2_5_half_billion_model,
+    download_qwen3_half_billion_model,
+    download_tiny_llama_model,
+    download_tiny_mistral_model,
+    download_tiny_mixtral_model,
+    download_tiny_phi_model,
+    download_tiny_falcon_model,
+    download_tiny_qwen2_model,
+    download_tiny_qwen3_model,
+    download_tiny_gemma2_model,
     download_tatsu_lab_alpaca_dataset,
     download_mhenrichsen_alpaca_2k_dataset,
     download_mhenrichsen_alpaca_2k_w_revision_dataset,

tests/e2e/integrations/test_cut_cross_entropy.py

@@ -10,7 +10,10 @@ from axolotl.utils import get_pytorch_version
 from axolotl.utils.config import normalize_config, prepare_plugins, validate_config
 from axolotl.utils.dict import DictDefault
 
-from tests.e2e.utils import check_model_output_exists
+from tests.e2e.utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+)
 
 
 @pytest.fixture()
@@ -35,13 +38,16 @@ def min_cfg(temp_dir):
         "num_epochs": 1,
         "micro_batch_size": 8,
         "gradient_accumulation_steps": 1,
-        "learning_rate": 0.00001,
+        "learning_rate": 5e-4,
         "optimizer": "adamw_torch_fused",
         "output_dir": temp_dir,
         "lr_scheduler": "cosine",
-        "max_steps": 10,
+        "max_steps": 40,
+        "warmup_steps": 5,
         "bf16": "auto",
         "save_first_step": False,
+        "use_tensorboard": True,
+        "seed": 42,
     }
 
 
@@ -64,11 +70,18 @@ class TestCutCrossEntropyIntegration:
         else:
             train(cfg=cfg, dataset_meta=dataset_meta)
             check_model_output_exists(temp_dir, cfg)
+            check_tensorboard_loss_decreased(
+                temp_dir + "/runs",
+                initial_window=5,
+                final_window=5,
+                max_initial=2.2,
+                max_final=2.0,
+            )
 
     def test_qwen2_w_cce(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "plugins": [
                     "axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin",
                 ],
@@ -87,13 +100,15 @@ class TestCutCrossEntropyIntegration:
                 "num_epochs": 1,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "output_dir": temp_dir,
                 "lr_scheduler": "cosine",
-                "max_steps": 10,
+                "max_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -108,6 +123,13 @@ class TestCutCrossEntropyIntegration:
         else:
             train(cfg=cfg, dataset_meta=dataset_meta)
             check_model_output_exists(temp_dir, cfg)
+            check_tensorboard_loss_decreased(
+                temp_dir + "/runs",
+                initial_window=5,
+                final_window=5,
+                max_initial=5.0,
+                max_final=4.7,
+            )
 
     @pytest.mark.parametrize(
         "attention_type",
@@ -136,3 +158,10 @@ class TestCutCrossEntropyIntegration:
         else:
             train(cfg=cfg, dataset_meta=dataset_meta)
             check_model_output_exists(temp_dir, cfg)
+            check_tensorboard_loss_decreased(
+                temp_dir + "/runs",
+                initial_window=5,
+                final_window=5,
+                max_initial=2.2,
+                max_final=2.0,
+            )

tests/e2e/kernels/test_lora_features.py

@@ -24,7 +24,7 @@ from axolotl.monkeypatch.lora_kernels import (
 )
 from axolotl.utils.dict import DictDefault
 
-MODEL_NAME = "Qwen/Qwen3-0.6B"
+MODEL_NAME = "axolotl-ai-co/tiny-qwen3-129m"
 DEVICE = "cuda"
 DTYPE = torch.bfloat16
 

tests/e2e/multigpu/test_dist_muon_fsdp2.py

@@ -1,23 +1,22 @@
 """Test module for DistMuon optimizer with FSDP2 multi-GPU functionality."""
 
-import os
 from pathlib import Path
 
-import torch
 import yaml
 from accelerate.test_utils import execute_subprocess_async
-from tbparse import SummaryReader
 from transformers.testing_utils import get_torch_dist_unique_port
 
 from axolotl.utils.dict import DictDefault
 
-from tests.e2e.utils import most_recent_subdir, require_torch_2_7_0
+from tests.e2e.utils import check_tensorboard_loss_decreased, require_torch_2_7_0
 
 AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
 
 
 def verify_training_success(temp_dir):
-    """Verify that training completed successfully by checking artifacts and loss."""
+    """Verify that training completed successfully — artifacts, no-NaN, loss
+    stayed in qwen2-pretraining scale (tiny-qwen2-129m final pretrain CE ~3.92).
+    """
     output_path = Path(temp_dir)
 
     model_files = list(output_path.glob("*.bin")) + list(
@@ -30,19 +29,13 @@ def verify_training_success(temp_dir):
         "No checkpoint files found - training may have failed"
     )
 
-    tb_log_path = most_recent_subdir(temp_dir + "/runs")
+    check_tensorboard_loss_decreased(
-    if tb_log_path:
+        temp_dir + "/runs",
-        event_files = sorted(os.listdir(tb_log_path))
+        initial_window=10,
-        if event_files:
+        final_window=10,
-            event_file = os.path.join(tb_log_path, event_files[0])
+        max_initial=5.0,
-            reader = SummaryReader(event_file)
+        max_final=4.7,
-            df = reader.scalars
+    )
-            train_loss_df = df[df.tag == "train/train_loss"]
-            if len(train_loss_df) > 0:
-                final_loss = train_loss_df.value.values[-1]
-                assert not torch.isnan(torch.tensor(final_loss)), (
-                    f"Training loss is NaN: {final_loss}"
-                )
 
 
 class TestDistMuon:
@@ -52,7 +45,7 @@ class TestDistMuon:
     def test_fft_sft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -63,11 +56,12 @@ class TestDistMuon:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.02,
+                "learning_rate": 2e-3,
                 "optimizer": "muon",
                 "weight_decay": 0.01,
                 "lr_scheduler": "cosine",
@@ -82,6 +76,9 @@ class TestDistMuon:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
             }
         )
@@ -109,7 +106,7 @@ class TestDistMuon:
     def test_lora_sft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -122,14 +119,15 @@ class TestDistMuon:
                 "adapter": "lora",
                 "lora_r": 8,
                 "lora_alpha": 16,
-                "lora_dropout": 0.05,
+                "lora_dropout": 0.0,
                 "lora_target_linear": True,
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.02,
+                "learning_rate": 2e-3,
                 "optimizer": "muon",
                 "weight_decay": 0.01,
                 "lr_scheduler": "cosine",
@@ -144,6 +142,9 @@ class TestDistMuon:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
             }
         )

tests/e2e/multigpu/test_fsdp1.py

@@ -1,24 +1,23 @@
 """Test module for FSDP1 multi-GPU functionality."""
 
-import os
 from pathlib import Path
 
 import pytest
-import torch
 import yaml
 from accelerate.test_utils import execute_subprocess_async
-from tbparse import SummaryReader
 from transformers.testing_utils import get_torch_dist_unique_port
 
 from axolotl.utils.dict import DictDefault
 
-from tests.e2e.utils import most_recent_subdir
+from tests.e2e.utils import check_tensorboard_loss_decreased
 
 AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
 
 
 def verify_training_success(temp_dir):
-    """Verify that training completed successfully by checking artifacts and loss."""
+    """Verify that training completed successfully — artifacts, no-NaN, loss
+    stayed in qwen2-pretraining scale (tiny-qwen2-129m final pretrain CE ~3.92).
+    """
     output_path = Path(temp_dir)
 
     model_files = list(output_path.glob("*.bin")) + list(
@@ -31,19 +30,13 @@ def verify_training_success(temp_dir):
         "No checkpoint files found - training may have failed"
     )
 
-    tb_log_path = most_recent_subdir(temp_dir + "/runs")
+    check_tensorboard_loss_decreased(
-    if tb_log_path:
+        temp_dir + "/runs",
-        event_files = sorted(os.listdir(tb_log_path))
+        initial_window=10,
-        if event_files:
+        final_window=10,
-            event_file = os.path.join(tb_log_path, event_files[0])
+        max_initial=5.0,
-            reader = SummaryReader(event_file)
+        max_final=4.7,
-            df = reader.scalars
+    )
-            train_loss_df = df[df.tag == "train/train_loss"]
-            if len(train_loss_df) > 0:
-                final_loss = train_loss_df.value.values[-1]
-                assert not torch.isnan(torch.tensor(final_loss)), (
-                    f"Training loss is NaN: {final_loss}"
-                )
 
 
 class TestFSDP1:
@@ -56,7 +49,7 @@ class TestFSDP1:
     def test_fft_sft(self, temp_dir, fsdp_cpu_ram_efficient_loading):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -67,11 +60,12 @@ class TestFSDP1:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -87,6 +81,9 @@ class TestFSDP1:
                     "fsdp_use_orig_params": False,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
             }
         )
@@ -126,7 +123,7 @@ class TestFSDP1:
     def test_lora_sft(self, temp_dir, adapter_config):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -140,14 +137,15 @@ class TestFSDP1:
                 "load_in_4bit": adapter_config["load_in_4bit"],
                 "lora_r": 8,
                 "lora_alpha": 16,
-                "lora_dropout": 0.05,
+                "lora_dropout": 0.0,
                 "lora_target_linear": True,
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 1e-3,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -163,6 +161,9 @@ class TestFSDP1:
                     "fsdp_use_orig_params": False,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
             }
         )
@@ -190,7 +191,7 @@ class TestFSDP1:
     def test_dpo_fft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "rl": "dpo",
@@ -203,11 +204,11 @@ class TestFSDP1:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 20,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -223,6 +224,9 @@ class TestFSDP1:
                     "fsdp_use_orig_params": False,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
             }
         )
 
@@ -262,7 +266,7 @@ class TestFSDP1:
     def test_dpo_lora(self, temp_dir, adapter_config):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "load_in_4bit": adapter_config["load_in_4bit"],
                 "rl": "dpo",
                 "chat_template": "chatml",
@@ -281,11 +285,11 @@ class TestFSDP1:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 20,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 1e-3,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -301,6 +305,9 @@ class TestFSDP1:
                     "fsdp_use_orig_params": False,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": "auto",
                 "tf32": True,
             }

tests/e2e/multigpu/test_fsdp2.py

@@ -1,24 +1,23 @@
 """Test module for FSDP2 multi-GPU functionality."""
 
-import os
 from pathlib import Path
 
 import pytest
-import torch
 import yaml
 from accelerate.test_utils import execute_subprocess_async
-from tbparse import SummaryReader
 from transformers.testing_utils import get_torch_dist_unique_port
 
 from axolotl.utils.dict import DictDefault
 
-from tests.e2e.utils import most_recent_subdir, require_torch_2_7_0
+from tests.e2e.utils import check_tensorboard_loss_decreased, require_torch_2_7_0
 
 AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
 
 
 def verify_training_success(temp_dir):
-    """Verify that training completed successfully by checking artifacts and loss."""
+    """Verify that training completed successfully — artifacts, no-NaN, loss
+    stayed in qwen2-pretraining scale (tiny-qwen2-129m final pretrain CE ~3.92).
+    """
     output_path = Path(temp_dir)
 
     model_files = list(output_path.glob("*.bin")) + list(
@@ -31,19 +30,13 @@ def verify_training_success(temp_dir):
         "No checkpoint files found - training may have failed"
     )
 
-    tb_log_path = most_recent_subdir(temp_dir + "/runs")
+    check_tensorboard_loss_decreased(
-    if tb_log_path:
+        temp_dir + "/runs",
-        event_files = sorted(os.listdir(tb_log_path))
+        initial_window=10,
-        if event_files:
+        final_window=10,
-            event_file = os.path.join(tb_log_path, event_files[0])
+        max_initial=5.0,
-            reader = SummaryReader(event_file)
+        max_final=4.7,
-            df = reader.scalars
+    )
-            train_loss_df = df[df.tag == "train/train_loss"]
-            if len(train_loss_df) > 0:
-                final_loss = train_loss_df.value.values[-1]
-                assert not torch.isnan(torch.tensor(final_loss)), (
-                    f"Training loss is NaN: {final_loss}"
-                )
 
 
 class TestFSDP2:
@@ -57,7 +50,7 @@ class TestFSDP2:
     def test_fft_sft(self, temp_dir, fsdp_cpu_ram_efficient_loading):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -68,11 +61,12 @@ class TestFSDP2:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -86,6 +80,9 @@ class TestFSDP2:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
             }
         )
@@ -114,7 +111,7 @@ class TestFSDP2:
     def test_lora_sft(self, temp_dir, peft_use_dora):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -128,14 +125,15 @@ class TestFSDP2:
                 "adapter": "lora",
                 "lora_r": 8,
                 "lora_alpha": 16,
-                "lora_dropout": 0.05,
+                "lora_dropout": 0.0,
                 "lora_target_linear": True,
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 1e-3,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -149,6 +147,9 @@ class TestFSDP2:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
                 # explicitly disable LORA kernels, as they may be auto-enabled
                 "lora_mlp_kernel": False,
@@ -180,7 +181,7 @@ class TestFSDP2:
     def test_lora_sft_kernels(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -195,11 +196,12 @@ class TestFSDP2:
                 "lora_alpha": 16,
                 "lora_target_linear": True,
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 1e-3,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -213,6 +215,9 @@ class TestFSDP2:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
                 "lora_mlp_kernel": True,
                 "lora_qkv_kernel": True,
@@ -243,7 +248,7 @@ class TestFSDP2:
     def test_qlora_sft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -257,14 +262,15 @@ class TestFSDP2:
                 "adapter": "qlora",
                 "lora_r": 8,
                 "lora_alpha": 16,
-                "lora_dropout": 0.05,
+                "lora_dropout": 0.0,
                 "lora_target_linear": True,
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 1e-3,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -278,6 +284,9 @@ class TestFSDP2:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
             }
         )
@@ -305,7 +314,7 @@ class TestFSDP2:
     def test_qlora_sft_kernels(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -321,11 +330,12 @@ class TestFSDP2:
                 "lora_alpha": 16,
                 "lora_target_linear": True,
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 1e-3,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -339,6 +349,9 @@ class TestFSDP2:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
                 "lora_mlp_kernel": True,
                 "lora_qkv_kernel": True,
@@ -370,7 +383,7 @@ class TestFSDP2:
     def test_dpo_fft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "rl": "dpo",
@@ -383,11 +396,11 @@ class TestFSDP2:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 20,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -401,6 +414,9 @@ class TestFSDP2:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
             }
         )
 
@@ -428,7 +444,7 @@ class TestFSDP2:
     def test_dpo_lora(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "rl": "dpo",
                 "chat_template": "chatml",
@@ -445,11 +461,11 @@ class TestFSDP2:
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 20,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 1e-3,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -463,6 +479,9 @@ class TestFSDP2:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
             }
         )
 

tests/e2e/multigpu/test_fsdp2_lora_kernels.py

@@ -40,7 +40,7 @@ def _run_training(temp_dir, cfg):
 def _base_lora_fsdp2_config(temp_dir, **overrides):
     """Base config for LoRA + FSDP2 + kernel tests."""
     cfg = {
-        "base_model": "Qwen/Qwen3-0.6B",
+        "base_model": "axolotl-ai-co/tiny-qwen3-129m",
         "sequence_len": 512,
         "val_set_size": 0.0,
         "datasets": [

tests/e2e/multigpu/test_tp.py

@@ -8,7 +8,7 @@ from accelerate.test_utils import execute_subprocess_async, get_torch_dist_uniqu
 
 from axolotl.utils.dict import DictDefault
 
-from tests.e2e.utils import check_tensorboard, require_torch_2_7_0
+from tests.e2e.utils import check_tensorboard_loss_decreased, require_torch_2_7_0
 
 
 class TestTensorParallel:
@@ -21,7 +21,7 @@ class TestTensorParallel:
     def test_fft_sft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -63,6 +63,6 @@ class TestTensorParallel:
             ]
         )
 
-        check_tensorboard(
+        check_tensorboard_loss_decreased(
-            temp_dir + "/runs", "train/train_loss", 1.0, "Train Loss (%s) is too high"
+            temp_dir + "/runs", max_initial=5.0, max_final=4.7
         )

tests/e2e/patched/lora_kernels/test_lora_kernel_patching.py

@@ -32,12 +32,12 @@ from axolotl.utils.dict import DictDefault
 
 MODEL_CONFIGS = [
     {
-        "name": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+        "name": "axolotl-ai-co/tiny-mistral-25m",
         "expected_activation": apply_lora_mlp_swiglu,
         "dtype": torch.float16,
     },
     {
-        "name": "trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
+        "name": "axolotl-ai-co/tiny-qwen2-129m",
         "expected_activation": apply_lora_mlp_swiglu,
         "dtype": torch.float16,
     },
@@ -47,7 +47,7 @@ MODEL_CONFIGS = [
         "dtype": torch.float32,
     },
     {
-        "name": "trl-internal-testing/tiny-Gemma2ForCausalLM",
+        "name": "axolotl-ai-co/tiny-gemma2-137m",
         "expected_activation": apply_lora_mlp_geglu,
         "dtype": torch.float16,
     },
@@ -159,7 +159,7 @@ def test_swiglu_mlp_integration(small_llama_model):
 def test_geglu_model_integration():
     """Test GeGLU activation with Gemma model."""
     model = AutoModelForCausalLM.from_pretrained(
-        "trl-internal-testing/tiny-Gemma2ForCausalLM",
+        "axolotl-ai-co/tiny-gemma2-137m",
         dtype=torch.float16,
         device_map="cuda:0",
     )

tests/e2e/patched/test_falcon_samplepack.py

@@ -4,14 +4,16 @@ E2E tests for falcon
 
 import unittest
 
-import pytest
-
 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 ..utils import check_model_output_exists, with_temp_dir
+from ..utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+    with_temp_dir,
+)
 
 
 class TestFalconPatched(unittest.TestCase):
@@ -19,13 +21,12 @@ class TestFalconPatched(unittest.TestCase):
     Test case for Falcon models
     """
 
-    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_qlora(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "illuin/tiny-random-FalconForCausalLM",
+                "base_model": "axolotl-ai-co/tiny-falcon-42m",
-                "flash_attention": True,
+                "flash_attention": False,
                 "sample_packing": True,
                 "sequence_len": 2048,
                 "load_in_4bit": True,
@@ -47,17 +48,20 @@ class TestFalconPatched(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -66,14 +70,20 @@ class TestFalconPatched(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=6.0,
+            max_final=4.7,
+        )
 
-    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_ft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "illuin/tiny-random-FalconForCausalLM",
+                "base_model": "axolotl-ai-co/tiny-falcon-42m",
-                "flash_attention": True,
+                "flash_attention": False,
                 "sample_packing": True,
                 "sequence_len": 2048,
                 "val_set_size": 0.05,
@@ -88,17 +98,20 @@ class TestFalconPatched(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -107,3 +120,10 @@ class TestFalconPatched(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=6.0,
+            max_final=4.7,
+        )

tests/e2e/patched/test_mistral_samplepack.py

@@ -9,7 +9,12 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import check_model_output_exists, require_torch_2_6_0, with_temp_dir
+from ..utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+    require_torch_2_6_0,
+    with_temp_dir,
+)
 
 
 class TestMistral(unittest.TestCase):
@@ -22,7 +27,7 @@ class TestMistral(unittest.TestCase):
     def test_lora_packing(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "axolotl-ai-co/tiny-mistral-25m",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 1024,
@@ -45,17 +50,20 @@ class TestMistral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 5,
+                "max_steps": 50,
-                "save_steps": 3,
+                "logging_steps": 1,
-                "eval_steps": 4,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -64,12 +72,19 @@ class TestMistral(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.5,
+            max_final=4.3,
+        )
 
     @with_temp_dir
     def test_ft_packing(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "axolotl-ai-co/tiny-mistral-25m",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 1024,
@@ -86,17 +101,20 @@ class TestMistral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 5,
+                "max_steps": 50,
-                "save_steps": 3,
+                "logging_steps": 1,
-                "eval_steps": 4,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -105,3 +123,10 @@ class TestMistral(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.5,
+            max_final=4.3,
+        )

tests/e2e/patched/test_mixtral_samplepack.py

@@ -9,7 +9,11 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import check_model_output_exists, with_temp_dir
+from ..utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+    with_temp_dir,
+)
 
 
 class TestMixtral(unittest.TestCase):
@@ -21,8 +25,7 @@ class TestMixtral(unittest.TestCase):
     def test_qlora(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "base_model": "axolotl-ai-co/tiny-mixtral-30m",
-                "tokenizer_config": "LoneStriker/Mixtral-8x7B-v0.1-HF",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 2048,
@@ -30,7 +33,7 @@ class TestMixtral(unittest.TestCase):
                 "adapter": "qlora",
                 "lora_r": 16,
                 "lora_alpha": 32,
-                "lora_dropout": 0.1,
+                "lora_dropout": 0.0,
                 "lora_target_linear": True,
                 "val_set_size": 0.05,
                 "special_tokens": {},
@@ -41,17 +44,21 @@ class TestMixtral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 3e-3,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 5,
+                "max_steps": 80,
-                "save_steps": 3,
+                "warmup_steps": 5,
-                "eval_steps": 4,
+                "logging_steps": 1,
+                "save_steps": 80,
+                "eval_steps": 80,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -60,13 +67,19 @@ class TestMixtral(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=10,
+            final_window=10,
+            max_initial=6.0,
+            max_final=4.7,
+        )
 
     @with_temp_dir
     def test_ft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "base_model": "axolotl-ai-co/tiny-mixtral-30m",
-                "tokenizer_config": "LoneStriker/Mixtral-8x7B-v0.1-HF",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 2048,
@@ -79,17 +92,21 @@ class TestMixtral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 5e-4,
-                "optimizer": "adamw_bnb_8bit",
+                "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 5,
+                "max_steps": 80,
-                "save_steps": 3,
+                "warmup_steps": 5,
-                "eval_steps": 4,
+                "logging_steps": 1,
+                "save_steps": 80,
+                "eval_steps": 80,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -98,3 +115,10 @@ class TestMixtral(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=6.0,
+            max_final=4.7,
+        )

tests/e2e/patched/test_model_patches.py

@@ -22,8 +22,7 @@ class TestModelPatches(unittest.TestCase):
     def test_mixtral_multipack(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "base_model": "axolotl-ai-co/tiny-mixtral-30m",
-                "tokenizer_config": "LoneStriker/Mixtral-8x7B-v0.1-HF",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 2048,
@@ -57,7 +56,7 @@ class TestModelPatches(unittest.TestCase):
     def test_mistral_multipack(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "axolotl-ai-co/tiny-mistral-25m",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 2048,

tests/e2e/patched/test_phi_multipack.py

@@ -9,7 +9,11 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import check_model_output_exists, with_temp_dir
+from ..utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+    with_temp_dir,
+)
 
 
 class TestPhiMultipack(unittest.TestCase):
@@ -21,7 +25,7 @@ class TestPhiMultipack(unittest.TestCase):
     def test_ft_packed(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "microsoft/phi-1_5",
+                "base_model": "axolotl-ai-co/tiny-phi-64m",
                 "model_type": "PhiForCausalLM",
                 "tokenizer_type": "AutoTokenizer",
                 "sequence_len": 1024,
@@ -43,17 +47,20 @@ class TestPhiMultipack(unittest.TestCase):
                 "dataset_shard_num": 10,
                 "dataset_shard_idx": 0,
                 "num_epochs": 1,
-                "micro_batch_size": 1,
+                "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
-                "optimizer": "adamw_bnb_8bit",
+                "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 5,
+                "max_steps": 50,
-                "eval_steps": 3,
+                "logging_steps": 1,
-                "save_steps": 4,
+                "eval_steps": 50,
+                "save_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
 
@@ -63,12 +70,19 @@ class TestPhiMultipack(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=6.0,
+            max_final=4.7,
+        )
 
     @with_temp_dir
     def test_qlora_packed(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "microsoft/phi-1_5",
+                "base_model": "axolotl-ai-co/tiny-phi-64m",
                 "model_type": "PhiForCausalLM",
                 "tokenizer_type": "AutoTokenizer",
                 "sequence_len": 1024,
@@ -94,17 +108,20 @@ class TestPhiMultipack(unittest.TestCase):
                 "dataset_shard_num": 10,
                 "dataset_shard_idx": 0,
                 "num_epochs": 1,
-                "micro_batch_size": 1,
+                "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 5,
+                "max_steps": 50,
-                "eval_steps": 3,
+                "logging_steps": 1,
-                "save_steps": 4,
+                "eval_steps": 50,
+                "save_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
 
@@ -114,3 +131,10 @@ class TestPhiMultipack(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=6.0,
+            max_final=4.7,
+        )

tests/e2e/solo/test_batch_flattening.py

@@ -18,7 +18,7 @@ from transformers import AutoModelForCausalLM
 # Import the actual trainer methods we want to test
 from axolotl.core.trainers.grpo.async_trainer import AsyncGRPOTrainer
 
-MODEL_NAME = "Qwen/Qwen3-0.6B"
+MODEL_NAME = "axolotl-ai-co/tiny-qwen3-129m"
 
 
 def _fix_patched_attention(model):

tests/e2e/test_falcon.py

@@ -4,14 +4,16 @@ E2E tests for falcon
 
 import unittest
 
-import pytest
-
 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 .utils import check_model_output_exists, with_temp_dir
+from .utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+    with_temp_dir,
+)
 
 
 class TestFalcon(unittest.TestCase):
@@ -19,13 +21,12 @@ class TestFalcon(unittest.TestCase):
     Test case for falcon
     """
 
-    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_lora(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "illuin/tiny-random-FalconForCausalLM",
+                "base_model": "axolotl-ai-co/tiny-falcon-42m",
-                "flash_attention": True,
+                "flash_attention": False,
                 "sequence_len": 1024,
                 "load_in_8bit": True,
                 "adapter": "lora",
@@ -49,17 +50,21 @@ class TestFalcon(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "warmup_steps": 5,
-                "eval_steps": 10,
+                "logging_steps": 1,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
 
@@ -69,14 +74,20 @@ class TestFalcon(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )
 
-    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_lora_added_vocab(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "illuin/tiny-random-FalconForCausalLM",
+                "base_model": "axolotl-ai-co/tiny-falcon-42m",
-                "flash_attention": True,
+                "flash_attention": False,
                 "sequence_len": 1024,
                 "load_in_8bit": True,
                 "adapter": "lora",
@@ -104,17 +115,21 @@ class TestFalcon(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "warmup_steps": 5,
-                "eval_steps": 10,
+                "logging_steps": 1,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
 
@@ -124,14 +139,20 @@ class TestFalcon(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )
 
-    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_ft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "illuin/tiny-random-FalconForCausalLM",
+                "base_model": "axolotl-ai-co/tiny-falcon-42m",
-                "flash_attention": True,
+                "flash_attention": False,
                 "sequence_len": 1024,
                 "val_set_size": 0.02,
                 "special_tokens": {
@@ -145,17 +166,23 @@ class TestFalcon(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 5e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 80,
-                "save_steps": 10,
+                "warmup_steps": 5,
-                "eval_steps": 10,
+                "logging_steps": 1,
+                "save_steps": 80,
+                "eval_steps": 80,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
 
@@ -165,3 +192,10 @@ class TestFalcon(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=10,
+            final_window=10,
+            max_initial=5.0,
+            max_final=4.7,
+        )

tests/e2e/test_mistral.py

@@ -11,7 +11,11 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from .utils import check_model_output_exists, with_temp_dir
+from .utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+    with_temp_dir,
+)
 
 
 class TestMistral(unittest.TestCase):
@@ -23,7 +27,7 @@ class TestMistral(unittest.TestCase):
     def test_lora(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "axolotl-ai-co/tiny-mistral-25m",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "load_in_8bit": True,
@@ -45,16 +49,18 @@ class TestMistral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "save_first_step": False,
+                "use_tensorboard": True,
             }
         )
 
@@ -64,12 +70,19 @@ class TestMistral(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=4.5,
+            max_final=4.3,
+        )
 
     @with_temp_dir
     def test_ft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "axolotl-ai-co/tiny-mistral-25m",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "val_set_size": 0.02,
@@ -85,16 +98,18 @@ class TestMistral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "save_first_step": False,
+                "use_tensorboard": True,
             }
         )
         if is_torch_bf16_gpu_available():
@@ -108,3 +123,10 @@ class TestMistral(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=4.5,
+            max_final=4.3,
+        )

tests/e2e/test_mixtral.py

@@ -12,7 +12,11 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from .utils import check_model_output_exists, with_temp_dir
+from .utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+    with_temp_dir,
+)
 
 
 class TestMixtral(unittest.TestCase):
@@ -24,8 +28,7 @@ class TestMixtral(unittest.TestCase):
     def test_qlora_w_fa2(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "base_model": "axolotl-ai-co/tiny-mixtral-30m",
-                "tokenizer_config": "LoneStriker/Mixtral-8x7B-v0.1-HF",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "load_in_4bit": True,
@@ -51,16 +54,18 @@ class TestMixtral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "save_first_step": False,
+                "use_tensorboard": True,
             }
         )
 
@@ -74,13 +79,19 @@ class TestMixtral(unittest.TestCase):
             == torch.float32
         )
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )
 
     @with_temp_dir
     def test_qlora_wo_fa2(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "base_model": "axolotl-ai-co/tiny-mixtral-30m",
-                "tokenizer_config": "LoneStriker/Mixtral-8x7B-v0.1-HF",
                 "flash_attention": False,
                 "sequence_len": 1024,
                 "load_in_4bit": True,
@@ -106,16 +117,18 @@ class TestMixtral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "save_first_step": False,
+                "use_tensorboard": True,
             }
         )
 
@@ -129,13 +142,19 @@ class TestMixtral(unittest.TestCase):
             == torch.float32
         )
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )
 
     @with_temp_dir
     def test_16bit_lora_w_fa2(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "base_model": "axolotl-ai-co/tiny-mixtral-30m",
-                "tokenizer_config": "LoneStriker/Mixtral-8x7B-v0.1-HF",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "adapter": "lora",
@@ -160,16 +179,18 @@ class TestMixtral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "save_first_step": False,
+                "use_tensorboard": True,
             }
         )
         if is_torch_bf16_gpu_available():
@@ -187,13 +208,19 @@ class TestMixtral(unittest.TestCase):
             == torch.float32
         )
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )
 
     @with_temp_dir
     def test_16bit_lora_wo_fa2(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "base_model": "axolotl-ai-co/tiny-mixtral-30m",
-                "tokenizer_config": "LoneStriker/Mixtral-8x7B-v0.1-HF",
                 "flash_attention": False,
                 "sequence_len": 1024,
                 "adapter": "lora",
@@ -218,16 +245,18 @@ class TestMixtral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "save_first_step": False,
+                "use_tensorboard": True,
             }
         )
 
@@ -245,13 +274,19 @@ class TestMixtral(unittest.TestCase):
             == torch.float32
         )
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )
 
     @with_temp_dir
     def test_ft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "base_model": "axolotl-ai-co/tiny-mixtral-30m",
-                "tokenizer_config": "LoneStriker/Mixtral-8x7B-v0.1-HF",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "val_set_size": 0.02,
@@ -263,16 +298,18 @@ class TestMixtral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "save_first_step": False,
+                "use_tensorboard": True,
             }
         )
         if is_torch_bf16_gpu_available():
@@ -286,3 +323,10 @@ class TestMixtral(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )

tests/e2e/test_optimizers.py

@@ -13,6 +13,7 @@ from axolotl.utils.dict import DictDefault
 
 from .utils import (
     check_model_output_exists,
+    check_tensorboard_loss_decreased,
     require_torch_2_5_1,
     require_torch_2_6_0,
     require_torch_2_7_0,
@@ -243,20 +244,18 @@ class TestCustomOptimizers(unittest.TestCase):
     def test_came_pytorch(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "axolotl-ai-co/tiny-llama-50m",
-                "tokenizer_type": "LlamaTokenizer",
+                "tokenizer_type": "AutoTokenizer",
                 "sequence_len": 1024,
                 "load_in_8bit": True,
                 "adapter": "lora",
                 "lora_r": 8,
                 "lora_alpha": 16,
-                "lora_dropout": 0.05,
+                "lora_dropout": 0.0,
                 "lora_target_linear": True,
                 "val_set_size": 0.1,
                 "special_tokens": {
-                    "unk_token": "<unk>",
+                    "pad_token": "<|endoftext|>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
                 },
                 "datasets": [
                     {
@@ -265,16 +264,22 @@ class TestCustomOptimizers(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "micro_batch_size": 8,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 1e-4,
                 "optimizer": "came_pytorch",
                 "adam_beta3": 0.9999,
                 "adam_epsilon2": 1e-16,
-                "max_steps": 5,
+                "max_steps": 80,
+                "warmup_steps": 5,
+                "logging_steps": 1,
                 "lr_scheduler": "cosine",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
 
@@ -284,6 +289,13 @@ class TestCustomOptimizers(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=10,
+            final_window=10,
+            max_initial=4.0,
+            max_final=3.0,
+        )
 
 
 @require_torch_2_7_0

tests/e2e/test_phi.py

@@ -9,7 +9,11 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from .utils import check_model_output_exists, with_temp_dir
+from .utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+    with_temp_dir,
+)
 
 
 class TestPhi(unittest.TestCase):
@@ -21,7 +25,7 @@ class TestPhi(unittest.TestCase):
     def test_phi_ft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "microsoft/phi-1_5",
+                "base_model": "axolotl-ai-co/tiny-phi-64m",
                 "model_type": "AutoModelForCausalLM",
                 "tokenizer_type": "AutoTokenizer",
                 "sequence_len": 2048,
@@ -41,18 +45,22 @@ class TestPhi(unittest.TestCase):
                 "dataset_shard_num": 10,
                 "dataset_shard_idx": 0,
                 "num_epochs": 1,
-                "micro_batch_size": 1,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
-                "optimizer": "paged_adamw_8bit",
+                "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
-                "max_steps": 10,
+                "max_steps": 50,
-                "save_steps": 10,
+                "warmup_steps": 5,
-                "eval_steps": 10,
+                "logging_steps": 1,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -61,12 +69,19 @@ class TestPhi(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )
 
     @with_temp_dir
     def test_phi_qlora(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "microsoft/phi-1_5",
+                "base_model": "axolotl-ai-co/tiny-phi-64m",
                 "model_type": "AutoModelForCausalLM",
                 "tokenizer_type": "AutoTokenizer",
                 "sequence_len": 2048,
@@ -90,18 +105,22 @@ class TestPhi(unittest.TestCase):
                 "dataset_shard_num": 10,
                 "dataset_shard_idx": 0,
                 "num_epochs": 1,
-                "micro_batch_size": 1,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "paged_adamw_8bit",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
-                "max_steps": 10,
+                "max_steps": 50,
-                "save_steps": 10,
+                "warmup_steps": 5,
-                "eval_steps": 10,
+                "logging_steps": 1,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -110,3 +129,10 @@ class TestPhi(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )

tests/e2e/test_preprocess.py

@@ -18,7 +18,7 @@ class TestPreprocess:
 
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [

tests/e2e/test_quantization.py

@@ -45,7 +45,7 @@ def _get_fake_quant_config_dtype(config):
 @pytest.fixture()
 def model():
     dummy_model = AutoModelForCausalLM.from_pretrained(
-        "Qwen/Qwen2-0.5B",
+        "axolotl-ai-co/tiny-qwen2-129m",
         device_map="auto",
         dtype=torch.bfloat16,
     )

tests/e2e/test_qwen.py

@@ -17,7 +17,7 @@ class TestE2eQwen:
     Test cases for qwen models
     """
 
-    @pytest.mark.parametrize("base_model", ["Qwen/Qwen2-0.5B", "Qwen/Qwen2.5-0.5B"])
+    @pytest.mark.parametrize("base_model", ["axolotl-ai-co/tiny-qwen2-129m"])
     def test_dpo(self, base_model, temp_dir):
         cfg = DictDefault(
             {

tests/e2e/utils.py

@@ -199,6 +199,106 @@ def check_tensorboard(
         assert df.value.values[-1] > 1e-5, "Expected loss to be greater than zero"
 
 
+def check_tensorboard_loss_decreased(
+    temp_run_dir: str,
+    tag: str | None = None,
+    initial_window: int = 1,
+    final_window: int = 1,
+    min_delta: float | None = None,
+    max_initial: float | None = None,
+    max_final: float | None = None,
+    max_loss_ratio: float = 0.95,
+) -> None:
+    """Check that training actually learned — loss went down and stayed in
+    a sensible range.
+
+    Used with the tiny ``axolotl-ai-co/tiny-*`` CI models, where pretraining
+    was brief enough that final loss won't clear the absolute thresholds used
+    for 135M+ models — but the training pipeline should still behave.
+
+    ``train/train_loss`` is only logged once (end-of-training aggregate). The
+    per-step tag is ``train/loss`` for SFT/LM trainers and may vary across
+    trainers (e.g. DPO). When ``tag`` is None we try common per-step tags in
+    order and use the first with enough samples.
+
+    Two kinds of regression we guard against:
+
+    1. **Loss blew up.** A silent bug (e.g. broken label masking) can start
+       training at an absurdly high loss. ``max_initial`` / ``max_final``
+       assert the measured means stay at-or-below bounds measured from a
+       known-good run. Both are optional but strongly encouraged — loss
+       going *down* from a bad starting scale still looks like "learning."
+
+    2. **Loss didn't go down enough.** ``max_loss_ratio`` (default 0.95)
+       requires ``final <= initial * ratio``. A default below 1.0 means the
+       final window mean must sit at least 5% below the initial window mean
+       — real learning, not noise that happened to land below start. Only
+       raise this for configs where a smaller drop is expected *and*
+       documented (e.g. DPO with near-trivial pairs); in that case you are
+       intentionally weakening the test.
+
+    ``min_delta`` is optional; when set, additionally requires
+    ``final + min_delta <= initial`` — use for configs with enough signal
+    to demand a specific minimum absolute drop.
+    """
+    tb_log_path = most_recent_subdir(temp_run_dir)
+    event_file = os.path.join(tb_log_path, sorted(os.listdir(tb_log_path))[0])
+    reader = SummaryReader(event_file)
+    df = reader.scalars
+
+    if tag is None:
+        candidates = ["train/loss", "train/train_loss"]
+    else:
+        candidates = [tag]
+
+    required = initial_window + final_window
+    chosen_tag, values = None, None
+    for candidate in candidates:
+        sub = df[df.tag == candidate]
+        if len(sub) >= required:
+            chosen_tag = candidate
+            values = sub.value.values
+            break
+
+    available = sorted({t for t in df.tag.unique() if "loss" in t.lower()})
+    assert values is not None, (
+        f"None of the tags {candidates} had ≥{required} logged steps. "
+        f"Loss tags present: {available}"
+    )
+
+    initial = float(values[:initial_window].mean())
+    final = float(values[-final_window:].mean())
+    print(
+        f"[check_tensorboard_loss_decreased] tag={chosen_tag} n={len(values)} "
+        f"initial_mean{initial_window}={initial:.4f} final_mean{final_window}={final:.4f}"
+    )
+    assert final > 1e-5, "Expected loss to be greater than zero"
+    assert final <= initial * max_loss_ratio, (
+        f"Loss did not decrease for {chosen_tag}: "
+        f"initial(mean of first {initial_window})={initial:.4f}, "
+        f"final(mean of last {final_window})={final:.4f}, "
+        f"ratio={final / initial:.4f} (max allowed {max_loss_ratio}). "
+        f"Expected final <= initial — training did not learn."
+    )
+    if min_delta is not None:
+        assert final + min_delta <= initial, (
+            f"Expected loss to decrease by at least {min_delta} for {chosen_tag}: "
+            f"initial={initial:.4f}, final={final:.4f}, delta={initial - final:.4f}"
+        )
+    if max_initial is not None:
+        assert initial <= max_initial, (
+            f"Initial loss {initial:.4f} is above the expected max {max_initial}. "
+            f"Absolute scale is wrong — probably a silent regression "
+            f"(e.g. bad label masking) that bumped the starting point."
+        )
+    if max_final is not None:
+        assert final <= max_final, (
+            f"Final loss {final:.4f} is above the expected max {max_final}. "
+            f"Absolute scale is wrong — probably a silent regression "
+            f"(e.g. bad label masking) that bumped the endpoint."
+        )
+
+
 def check_model_output_exists(temp_dir: str, cfg: DictDefault) -> None:
     """
     helper function to check if a model output file exists after training