Update src/axolotl/common/datasets.py

Co-authored-by: Dan Saunders <danjsaund@gmail.com>

.coveragerc

@@ -0,0 +1,14 @@
+[run]
+source = axolotl
+omit =
+    */tests/*
+    setup.py
+
+[report]
+exclude_lines =
+    pragma: no cover
+    def __repr__
+    raise NotImplementedError
+    if __name__ == .__main__.:
+    pass
+    raise ImportError

.github/workflows/main.yml

@@ -29,7 +29,7 @@ jobs:
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.6.0
-            axolotl_extras:
+            axolotl_extras: vllm
             is_latest: true
     runs-on: axolotl-gpu-runner
     steps:

.github/workflows/tests.yml

@@ -102,9 +102,16 @@ jobs:
 
       - name: Run tests
         run: |
-          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/
+          pytest -v -n8 --dist loadfile --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli/ tests/ --cov=axolotl --cov-report=xml
-          pytest -v tests/patched/
+          pytest -v tests/patched/ --cov=axolotl --cov-append --cov-report=xml
-          pytest -v tests/cli/
+          pytest -v tests/cli/ --cov=axolotl --cov-append --cov-report=xml
+
+      - name: Upload coverage to Codecov
+        uses: codecov/codecov-action@v5
+        with:
+          files: ./coverage.xml
+          flags: unittests,pytorch-${{ matrix.pytorch_version }}
+          fail_ci_if_error: false
 
       - name: cleanup pip cache
         run: |

CNAME

@@ -0,0 +1 @@
+docs.axolotl.ai

README.md

@@ -9,6 +9,7 @@
 <p align="center">
     <img src="https://img.shields.io/github/license/axolotl-ai-cloud/axolotl.svg?color=blue" alt="GitHub License">
     <img src="https://github.com/axolotl-ai-cloud/axolotl/actions/workflows/tests.yml/badge.svg" alt="tests">
+    <a href="https://codecov.io/gh/axolotl-ai-cloud/axolotl"><img src="https://codecov.io/gh/axolotl-ai-cloud/axolotl/branch/main/graph/badge.svg" alt="codecov"></a>
     <a href="https://github.com/axolotl-ai-cloud/axolotl/releases"><img src="https://img.shields.io/github/release/axolotl-ai-cloud/axolotl.svg" alt="Releases"></a>
     <br/>
     <a href="https://github.com/axolotl-ai-cloud/axolotl/graphs/contributors"><img src="https://img.shields.io/github/contributors-anon/axolotl-ai-cloud/axolotl?color=yellow&style=flat-square" alt="contributors" style="height: 20px;"></a>
@@ -63,7 +64,7 @@ axolotl fetch examples
 axolotl fetch deepspeed_configs  # OPTIONAL
 ```
 
-Other installation approaches are described [here](https://axolotl-ai-cloud.github.io/axolotl/docs/installation.html).
+Other installation approaches are described [here](https://docs.axolotl.ai/docs/installation.html).
 
 ### Your First Fine-tune
 
@@ -78,7 +79,7 @@ axolotl fetch examples --dest path/to/folder
 axolotl train examples/llama-3/lora-1b.yml
 ```
 
-That's it! Check out our [Getting Started Guide](https://axolotl-ai-cloud.github.io/axolotl/docs/getting-started.html) for a more detailed walkthrough.
+That's it! Check out our [Getting Started Guide](https://docs.axolotl.ai/docs/getting-started.html) for a more detailed walkthrough.
 
 ## ✨ Key Features
 
@@ -91,20 +92,20 @@ That's it! Check out our [Getting Started Guide](https://axolotl-ai-cloud.github
 
 ## 📚 Documentation
 
-- [Installation Options](https://axolotl-ai-cloud.github.io/axolotl/docs/installation.html) - Detailed setup instructions for different environments
+- [Installation Options](https://docs.axolotl.ai/docs/installation.html) - Detailed setup instructions for different environments
-- [Configuration Guide](https://axolotl-ai-cloud.github.io/axolotl/docs/config.html) - Full configuration options and examples
+- [Configuration Guide](https://docs.axolotl.ai/docs/config.html) - Full configuration options and examples
-- [Dataset Guide](https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/) - Supported formats and how to use them
+- [Dataset Guide](https://docs.axolotl.ai/docs/dataset-formats/) - Supported formats and how to use them
-- [Multi-GPU Training](https://axolotl-ai-cloud.github.io/axolotl/docs/multi-gpu.html)
+- [Multi-GPU Training](https://docs.axolotl.ai/docs/multi-gpu.html)
-- [Multi-Node Training](https://axolotl-ai-cloud.github.io/axolotl/docs/multi-node.html)
+- [Multi-Node Training](https://docs.axolotl.ai/docs/multi-node.html)
-- [Multipacking](https://axolotl-ai-cloud.github.io/axolotl/docs/multipack.html)
+- [Multipacking](https://docs.axolotl.ai/docs/multipack.html)
-- [API Reference](https://axolotl-ai-cloud.github.io/axolotl/docs/api/) - Auto-generated code documentation
+- [API Reference](https://docs.axolotl.ai/docs/api/) - Auto-generated code documentation
-- [FAQ](https://axolotl-ai-cloud.github.io/axolotl/docs/faq.html) - Frequently asked questions
+- [FAQ](https://docs.axolotl.ai/docs/faq.html) - Frequently asked questions
 
 ## 🤝 Getting Help
 
 - Join our [Discord community](https://discord.gg/HhrNrHJPRb) for support
 - Check out our [Examples](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/) directory
-- Read our [Debugging Guide](https://axolotl-ai-cloud.github.io/axolotl/docs/debugging.html)
+- Read our [Debugging Guide](https://docs.axolotl.ai/docs/debugging.html)
 - Need dedicated support? Please contact [✉️wing@axolotl.ai](mailto:wing@axolotl.ai) for options
 
 ## 🌟 Contributing

cicd/cicd.sh

@@ -3,10 +3,59 @@ set -e
 
 python -c "import torch; assert '$PYTORCH_VERSION' in torch.__version__"
 
-pytest -v --durations=10 -n8 --ignore=tests/e2e/ --ignore=tests/patched/ --ignore=tests/cli /workspace/axolotl/tests/
+# Run unit tests with initial coverage report
-pytest -v --durations=10 /workspace/axolotl/tests/e2e/patched/lora_kernels  # running these with the other patches causes a failure
+pytest -v --durations=10 -n8 \
-pytest -v --durations=10 --ignore=tests/e2e/patched/lora_kernels /workspace/axolotl/tests/e2e/patched
+  --ignore=tests/e2e/ \
-pytest -v --durations=10 -n1 /workspace/axolotl/tests/e2e/solo/
+  --ignore=tests/patched/ \
-pytest -v --durations=10 /workspace/axolotl/tests/e2e/integrations/
+  --ignore=tests/cli \
-pytest -v --durations=10 /workspace/axolotl/tests/cli
+  /workspace/axolotl/tests/ \
-pytest -v --durations=10 --ignore=tests/e2e/solo/ --ignore=tests/e2e/patched/ --ignore=tests/e2e/multigpu/ --ignore=tests/e2e/integrations/ --ignore=tests/cli /workspace/axolotl/tests/e2e/
+  --cov=axolotl \
+  --cov-report=xml:coverage.xml
+
+# Run lora kernels tests with coverage append
+pytest -v --durations=10 \
+  /workspace/axolotl/tests/e2e/patched/lora_kernels \
+  --cov=axolotl \
+  --cov-append
+
+# Run patched tests excluding lora kernels with coverage append
+pytest -v --durations=10 \
+  --ignore=tests/e2e/patched/lora_kernels \
+  /workspace/axolotl/tests/e2e/patched \
+  --cov=axolotl \
+  --cov-append
+
+# Run solo tests with coverage append
+pytest -v --durations=10 -n1 \
+  /workspace/axolotl/tests/e2e/solo/ \
+  --cov=axolotl \
+  --cov-append
+
+# Run integration tests with coverage append
+pytest -v --durations=10 \
+  /workspace/axolotl/tests/e2e/integrations/ \
+  --cov=axolotl \
+  --cov-append
+
+pytest -v --durations=10 /workspace/axolotl/tests/cli \
+  --cov=axolotl \
+  --cov-append
+
+# Run remaining e2e tests with coverage append and final report
+pytest -v --durations=10 \
+  --ignore=tests/e2e/solo/ \
+  --ignore=tests/e2e/patched/ \
+  --ignore=tests/e2e/multigpu/ \
+  --ignore=tests/e2e/integrations/ \
+  --ignore=tests/cli \
+  /workspace/axolotl/tests/e2e/ \
+  --cov=axolotl \
+  --cov-append \
+  --cov-report=xml:coverage.xml
+
+# Upload coverage to Codecov
+if [ -f e2e-coverage.xml ]; then
+  codecov -f e2e-coverage.xml -F e2e,pytorch-${PYTORCH_VERSION}
+else
+  echo "Coverage file not found. Coverage report may have failed."
+fi

cicd/multigpu.py

@@ -68,7 +68,7 @@ def run_cmd(cmd: str, run_folder: str):
 @app.function(
     image=cicd_image,
     gpu=GPU_CONFIG,
-    timeout=60 * 60,
+    timeout=90 * 60,
     cpu=8.0,
     memory=131072 * N_GPUS,
     volumes=VOLUME_CONFIG,

cicd/multigpu.sh

@@ -4,3 +4,22 @@ set -e
 # only run one test at a time so as not to OOM the GPU
 pytest -v  --durations=10 -n2 /workspace/axolotl/tests/e2e/multigpu/ --ignore=/workspace/axolotl/tests/e2e/multigpu/solo/
 pytest -v  --durations=10 -n1 /workspace/axolotl/tests/e2e/multigpu/solo/
+
+# Only run two tests at a time to avoid OOM on GPU (with coverage collection)
+pytest -v -n2 \
+  --ignore=/workspace/axolotl/tests/e2e/multigpu/solo/
+  /workspace/axolotl/tests/e2e/multigpu/ \
+  --cov=axolotl \
+  --cov-report=xml:multigpu-coverage.xml
+
+pytest -v  --durations=10 -n1 /workspace/axolotl/tests/e2e/multigpu/solo/ \
+  --cov=axolotl \
+  --cov-append \
+  --cov-report=xml:multigpu-coverage.xml
+
+# Upload coverage to Codecov
+if [ -f multigpu-coverage.xml ]; then
+  codecov -f multigpu-coverage.xml -F multigpu,docker-tests,pytorch-${PYTORCH_VERSION}
+else
+  echo "Coverage file not found. Coverage report may have failed."
+fi

codecov.yml

@@ -0,0 +1,51 @@
+codecov:
+  require_ci_to_pass: yes
+
+coverage:
+  precision: 2
+  round: down
+  range: "70...100"
+  status:
+    project:
+      default:
+        # basic
+        target: auto
+        threshold: 0%
+        base: auto
+        # advanced
+        branches: null
+        if_no_uploads: error
+        if_not_found: success
+        if_ci_failed: error
+        only_pulls: false
+        flags: null
+        paths: null
+    patch:
+      default:
+        # basic
+        target: auto
+        threshold: 0%
+        base: auto
+        # advanced
+        branches: null
+        if_no_uploads: error
+        if_not_found: success
+        if_ci_failed: error
+        only_pulls: false
+        flags: null
+        paths: null
+
+parsers:
+  gcov:
+    branch_detection:
+      conditional: yes
+      loop: yes
+      method: no
+      macro: no
+
+comment:
+  layout: "reach,diff,flags,files,footer"
+  behavior: default
+  require_changes: no
+  require_base: no
+  require_head: yes

docs/config.qmd

@@ -90,7 +90,7 @@ lora_on_cpu: true
 
 # List[str]. Add plugins to extend the pipeline.
 # See `src/axolotl/integrations` for the available plugins or doc below for more details.
-# https://axolotl-ai-cloud.github.io/axolotl/docs/custom_integrations.html
+# https://docs.axolotl.ai/docs/custom_integrations.html
 plugins:
   # - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
 
@@ -394,7 +394,7 @@ lora_fan_in_fan_out: false
 
 # Apply custom LoRA autograd functions and activation function Triton kernels for
 # speed and memory savings
-# See: https://axolotl-ai-cloud.github.io/axolotl/docs/lora_optims.html
+# See: https://docs.axolotl.ai/docs/lora_optims.html
 lora_mlp_kernel: true
 lora_qkv_kernel: true
 lora_o_kernel: true
@@ -688,11 +688,14 @@ ddp_broadcast_buffers:
 # Use in long context training to prevent OOM when sequences cannot fit into a single GPU's VRAM.
 # E.g., if 4 GPUs are available, set this value to 2 to split each sequence into two equal-sized
 # subsequences, or set to 4 to split into four equal-sized subsequences.
-# See https://axolotl-ai-cloud.github.io/axolotl/docs/sequence_parallelism.html for more details.
+# See https://docs.axolotl.ai/docs/sequence_parallelism.html for more details.
 sequence_parallel_degree:
 # Optional; strides across the key dimension. Larger values use more memory but should make training faster.
 # Must evenly divide the number of KV heads in your model.
 heads_k_stride: 1
+# One of "varlen_llama3", "batch_ring", "batch_zigzag", "batch_stripe". Defaults to "varlen_llama3"
+# in the sample packing case, and "batch_ring" in the non-sample packing case.
+ring_attn_func:
 
 # Path to torch distx for optim 'adamw_anyprecision'
 torchdistx_path:

docs/dataset-formats/index.qmd

@@ -457,10 +457,7 @@ datasets:
     type: alpaca
 ```
 
-Axolotl supports many kinds of instruction dataset. All of them can be found here (https://axolotl-ai-cloud.github.io/axolotl/docs/dataset-formats/inst_tune.html) with their respective type and sample row format.
+Axolotl supports many kinds of instruction dataset. All of them can be found in the [Instruction Dataset Documentation](inst_tune.qmd) with their respective type and sample row format.
-
-
-Reference: [Instruction Dataset Documentation](inst_tune.qmd).
 
 #### Custom Instruct Prompt Format
 

docs/multi-gpu.qmd

@@ -36,6 +36,9 @@ deepspeed: deepspeed_configs/zero1.json
 ### Usage {#sec-deepspeed-usage}
 
 ```{.bash}
+# Fetch deepspeed configs (if not already present)
+axolotl fetch deepspeed_configs
+
 # Passing arg via config
 axolotl train config.yml
 
@@ -48,10 +51,20 @@ axolotl train config.yml --deepspeed deepspeed_configs/zero1.json
 We provide default configurations for:
 
 - ZeRO Stage 1 (`zero1.json`)
+- ZeRO Stage 1 with torch compile (`zero1_torch_compile.json`)
 - ZeRO Stage 2 (`zero2.json`)
 - ZeRO Stage 3 (`zero3.json`)
+- ZeRO Stage 3 with bf16 (`zero3_bf16.json`)
+- ZeRO Stage 3 with bf16 and CPU offload params(`zero3_bf16_cpuoffload_params.json`)
+- ZeRO Stage 3 with bf16 and CPU offload params and optimizer (`zero3_bf16_cpuoffload_all.json`)
 
-Choose based on your memory requirements and performance needs.
+::: {.callout-tip}
+
+Choose the configuration that offloads the least amount to memory while still being able to fit on VRAM for best performance.
+
+Start from Stage 1 -> Stage 2 -> Stage 3.
+
+:::
 
 ## FSDP {#sec-fsdp}
 

docs/multimodal.qmd

@@ -164,7 +164,7 @@ Here is an example of a multi-modal dataset:
         {
             "role": "user",
             "content": [
-                {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
+                {"type": "image", "image": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
                 {"type": "text", "text": "Describe this image in detail."}
             ]
         },

docs/rlhf.qmd

@@ -530,7 +530,7 @@ trl:
 ```
 
 ```bash
-CUDA_VISIBLE_DEVICES=2,3 axolotl vllm_serve grpo.yaml
+CUDA_VISIBLE_DEVICES=2,3 axolotl vllm-serve grpo.yaml
 ```
 
 Your `vLLM` instance will now attempt to spin up, and it's time to kick off training utilizing our remaining two GPUs. In another terminal, execute:

docs/sequence_parallelism.qmd

@@ -27,6 +27,9 @@ To enable sequence parallelism, add the following to your configuration file:
 sequence_parallel_degree: 4  # Split sequences across 4 GPUs
 # Optional; strides across the key dimension. Larger values use more memory but should make training faster.
 heads_k_stride: 1
+# Optional; one of "varlen_llama3", "batch_ring", "batch_zigzag", "batch_stripe". Defaults to
+# "varlen_llama3" when `sample_packing: true`, and "batch_ring" otherwise.
+ring_attn_func:
 ```
 
 The `sequence_parallel_degree` should be a divisor of the total number of GPUs. For example:

examples/cerebras/btlm-ft.yml

@@ -8,7 +8,6 @@ tokenizer_type: GPT2Tokenizer
 trust_remote_code: true
 tokenizer_use_fast: true
 tokenizer_legacy: true
-strict: false
 push_dataset_to_hub:
 hf_use_auth_token: true
 datasets:

examples/cerebras/qlora.yml

@@ -4,7 +4,6 @@ base_model: cerebras/Cerebras-GPT-1.3B
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned

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

@@ -7,7 +7,6 @@ tokenizer_type: CodeLlamaTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

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

@@ -7,7 +7,6 @@ tokenizer_type: CodeLlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

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

@@ -7,7 +7,6 @@ tokenizer_type: CodeLlamaTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

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

@@ -7,7 +7,6 @@ tokenizer_type: CodeLlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

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

@@ -7,7 +7,6 @@ tokenizer_type: CodeLlamaTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

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

@@ -7,7 +7,6 @@ tokenizer_type: CodeLlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

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

@@ -4,7 +4,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 # huggingface repo
 chat_template: cohere

examples/dbrx/16bit-lora.yaml

@@ -3,7 +3,6 @@ base_model: LnL-AI/dbrx-base-converted-v2
 # hub_model_id: username/custom_model_name
 
 trust_remote_code: true
-strict: false
 
 datasets:
   - path: tatsu-lab/alpaca

examples/dbrx/8bit-lora.yaml

@@ -6,7 +6,6 @@ trust_remote_code: true
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 datasets:
   - path: tatsu-lab/alpaca

examples/dbrx/fft-ds-zero3.yaml

@@ -3,7 +3,6 @@ base_model: LnL-AI/dbrx-base-converted-v2
 # hub_model_id: username/custom_model_name
 
 trust_remote_code: true
-strict: false
 
 datasets:
   - path: tatsu-lab/alpaca

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

@@ -0,0 +1,58 @@
+base_model: agentica-org/DeepCoder-14B-Preview
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+    field_messages: messages
+    message_property_mappings:
+      role: role
+      content: content
+
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/lora-out
+
+sequence_len: 4096
+sample_packing: true
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 2
+num_epochs: 4
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

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

@@ -0,0 +1,58 @@
+base_model: deepcogito/cogito-v1-preview-llama-3B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+    field_messages: messages
+    message_property_mappings:
+      role: role
+      content: content
+
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/lora-out
+
+sequence_len: 4096
+sample_packing: true
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

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

@@ -0,0 +1,58 @@
+base_model: deepcogito/cogito-v1-preview-qwen-14B
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+load_in_8bit: true
+load_in_4bit: false
+strict: false
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+    field_messages: messages
+    message_property_mappings:
+      role: role
+      content: content
+
+dataset_prepared_path:
+val_set_size: 0.05
+output_dir: ./outputs/lora-out
+
+sequence_len: 4096
+sample_packing: true
+eval_sample_packing: false
+pad_to_sequence_len: true
+
+adapter: lora
+lora_model_dir:
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0.05
+lora_target_linear: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 2
+micro_batch_size: 2
+num_epochs: 1
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: true
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 10
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:

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

@@ -2,7 +2,6 @@ base_model: deepseek-ai/DeepSeek-V2-Lite
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 trust_remote_code: true
-strict: false
 
 datasets:
   - path: tatsu-lab/alpaca

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

@@ -6,7 +6,6 @@ trust_remote_code: true
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 
 plugins:

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

@@ -11,7 +11,6 @@ trust_remote_code: true
 load_in_8bit: true
 load_in_4bit: false
 gptq: false
-strict: false
 push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned

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

@@ -15,7 +15,6 @@ load_in_8bit: false
 # enable 4bit for QLoRA
 load_in_4bit: true
 gptq: false
-strict: false
 push_dataset_to_hub:
 datasets:
   - path: QingyiSi/Alpaca-CoT

examples/falcon/config-7b.yml

@@ -8,7 +8,6 @@ tokenizer_type: AutoTokenizer
 # required by falcon custom model code: https://huggingface.co/tiiuae/falcon-7b/tree/main
 trust_remote_code: true
 gptq: false
-strict: false
 push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned

examples/gemma/qlora.yml

@@ -8,7 +8,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 # huggingface repo
 datasets:

examples/gemma2/qlora.yml

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 # huggingface repo
 chat_template: gemma

examples/gemma2/reward-model.yaml

@@ -5,7 +5,6 @@ num_labels: 1
 tokenizer_type: AutoTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-strict: false
 
 reward_model: true
 chat_template: gemma

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

@@ -10,7 +10,6 @@ ddp_find_unused_parameters: true
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 # huggingface repo
 chat_template: gemma3

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

@@ -1,5 +1,4 @@
 base_model: google/gemma-3-4b-it
-strict: false
 
 load_in_4bit: true
 

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

@@ -1,6 +1,5 @@
 base_model: google/gemma-3-4b-it
 processor_type: AutoProcessor
-strict: false
 
 load_in_4bit: true
 

examples/gptj/qlora.yml

@@ -4,7 +4,6 @@ base_model: EleutherAI/gpt-j-6b
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned

examples/jamba/qlora.yaml

@@ -6,7 +6,6 @@ trust_remote_code: true
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

examples/jamba/qlora_deepspeed.yaml

@@ -5,7 +5,6 @@ trust_remote_code: true
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

examples/jamba/qlora_fsdp_large.yaml

@@ -5,7 +5,6 @@ tokenizer_type: AutoTokenizer
 # hub_model_id: username/custom_model_name
 
 load_in_4bit: true
-strict: false
 use_tensorboard: true
 chat_template: jamba
 datasets:

examples/llama-2/fft_optimized.yml

@@ -4,7 +4,6 @@ model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

examples/llama-2/gptq-lora.yml

@@ -10,7 +10,6 @@ gptq_disable_exllama: true
 
 tokenizer_use_fast: true
 tokenizer_legacy: true
-strict: false
 push_dataset_to_hub:
 hf_use_auth_token: true
 datasets:

examples/llama-2/lisa.yml

@@ -4,7 +4,6 @@ model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-strict: false
 
 datasets:
   - path: teknium/GPT4-LLM-Cleaned

examples/llama-2/loftq.yml

@@ -4,7 +4,6 @@ model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

examples/llama-2/lora.yml

@@ -7,7 +7,6 @@ tokenizer_type: LlamaTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

examples/llama-2/qlora-fsdp.yml

@@ -7,7 +7,6 @@ tokenizer_type: LlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: yahma/alpaca-cleaned

examples/llama-2/qlora.yml

@@ -7,7 +7,6 @@ tokenizer_type: LlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

examples/llama-2/relora.yml

@@ -5,7 +5,6 @@ tokenizer_type: LlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: teknium/GPT4-LLM-Cleaned

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

@@ -4,7 +4,6 @@ processor_type: AutoProcessor
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
 
-strict: false
 
 # these 3 lines are needed for now to handle vision chat templates w images
 skip_prepare_dataset: true

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

@@ -9,7 +9,6 @@ liger_rms_norm: true
 liger_glu_activation: true
 liger_fused_linear_cross_entropy: true
 
-strict: false
 
 chat_template: llama3
 datasets:

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

@@ -1,7 +1,6 @@
 base_model: NousResearch/Meta-Llama-3.1-8B
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-strict: false
 
 datasets:
   - path: tatsu-lab/alpaca

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

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 chat_template: llama3
 rl: dpo

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

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 chat_template: llama3
 datasets:

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

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 chat_template: llama3
 rl: dpo

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

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

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

@@ -1,7 +1,6 @@
 base_model: NousResearch/Llama-3.2-1B
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-strict: false
 
 datasets:
   - path: teknium/GPT4-LLM-Cleaned

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

@@ -1,7 +1,6 @@
 base_model: NousResearch/Llama-3.2-1B
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-strict: false
 
 datasets:
   - path: teknium/GPT4-LLM-Cleaned

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

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

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

@@ -1,7 +1,6 @@
 base_model: NousResearch/Llama-3.2-1B
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-strict: false
 
 datasets:
   - path: teknium/GPT4-LLM-Cleaned

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

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

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

@@ -4,7 +4,6 @@ base_model: meta-llama/Llama-3.2-1B
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 rl: kto
 rl_beta: 0.5

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

@@ -4,7 +4,6 @@ base_model: NousResearch/Llama-3.2-1B
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: teknium/GPT4-LLM-Cleaned

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

@@ -5,7 +5,6 @@ tokenizer_type: AutoTokenizer
 # hub_model_id: username/custom_model_name
 
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: tatsu-lab/alpaca

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

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer  # PreTrainedTokenizerFast
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: tatsu-lab/alpaca

examples/llama-3/qlora.yml

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: aaditya/alpaca_subset_1

examples/llama-4/README.md

@@ -0,0 +1,28 @@
+# Llama 4 by Meta AI
+
+## Flash Attention vs Flex Attention
+
+While Flash Attention to support is "enabled" for Llama-4, the upstream implementation is not correct and usage of Flex Attention is recommended.
+
+## Available Examples
+
+### Llama 4 Scout 17Bx16Experts (109B)
+
+Flex Attention
+- [Text Single GPU (H100) QLoRA](./scout-qlora-single-h100-flex.yaml)
+- [Text Multi GPU QLoRA w/ FSDP2](./scout-qlora-flexattn-fsdp2.yaml)
+
+[//]: # (Flash Attention (Do not use))
+
+[//]: # (- [Multi-Modal/Vision QLoRA w/ FSDP1](./scout-vision-qlora-fsdp.yaml))
+
+[//]: # (- [Text Single GPU (H100) QLoRA](./scout-qlora-single-h100.yaml))
+
+[//]: # (- [Text Multi GPU QLoRA w/ FSDP1](./scout-qlora-fsdp1.yaml))
+
+Our Single H100 implementation for Llama 4 Scout uses only 64.5GB VRAM for post-training with 4k context length @ 519 tokens/second. [WandB logs here](https://wandb.ai/axolotl-ai/llama4-flexattn-qlora/runs/wpie7dkj)
+Multi-GPU (4xH100) for Llama 4 Scout uses 62.8GB VRAM/GPU @ 4k contenxt length @ 280tps/gpu, [WandB logs here](https://wandb.ai/axolotl-ai/llama4-flexattn-qlora/runs/2lkezdj8)
+
+### Llama 4 Maverick 17Bx128Experts (400B)
+
+Coming Soon

examples/llama-4/do-no-use-fa2/maverick-qlora-fsdp1.yaml

@@ -0,0 +1,88 @@
+base_model: axolotl-quants/Llama-4-Maverick-17B-128E-Linearized-bnb-nf4-bf16
+model_type: Llama4ForConditionalGeneration
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_glu_activation: true
+liger_rms_norm: true
+liger_layer_norm: true
+
+llama4_linearized_experts: true
+load_in_4bit: true
+adapter: qlora
+lora_r: 32
+lora_alpha: 64
+lora_target_modules:
+  - self_attn.q_proj
+  - self_attn.k_proj
+  - self_attn.v_proj
+  - self_attn.o_proj
+  - shared_expert.gate_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  # - experts.gate_projs.[0-9]+$
+  # - experts.up_projs.[0-9]+$
+  # - experts.down_projs.[0-9]+$
+lora_modules_to_save:
+# - lm_head
+# - embed_tokens
+
+chat_template: llama4
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 4096
+sample_packing: true
+pad_to_sequence_len: true
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 1e-4
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+flash_attention: true
+
+gradient_checkpointing: offload
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+
+warmup_steps: 20
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+fsdp:
+  - auto_wrap
+  - full_shard
+fsdp_config:
+  fsdp_transformer_layer_cls_to_wrap: Llama4TextDecoderLayer
+  fsdp_limit_all_gathers: true
+  fsdp_sync_module_states: true
+  fsdp_offload_params: true
+  fsdp_use_orig_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot|>

examples/llama-4/do-no-use-fa2/scout-qlora-fsdp1.yaml

@@ -3,7 +3,6 @@ model_type: Llama4ForConditionalGeneration
   # Automatically upload checkpoint and final model to HF
   # hub_model_id: username/custom_model_name
 
-strict: false
 
 # torch_compile: true
 plugins:

examples/llama-4/do-no-use-fa2/scout-qlora-single-h100.yaml

@@ -0,0 +1,85 @@
+base_model: axolotl-quants/Llama-4-Scout-17B-16E-Linearized-bnb-nf4-bf16
+model_type: Llama4ForConditionalGeneration
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_glu_activation: true
+liger_rms_norm: true
+liger_layer_norm: true
+
+llama4_linearized_experts: true
+load_in_4bit: true
+adapter: qlora
+lora_r: 32
+lora_alpha: 64
+lora_target_modules:
+  - self_attn.q_proj
+  - self_attn.k_proj
+  - self_attn.v_proj
+  - self_attn.o_proj
+  - shared_expert.gate_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  # - experts.gate_projs.[0-9]+$
+  # - experts.up_projs.[0-9]+$
+  # - experts.down_projs.[0-9]+$
+lora_modules_to_save:
+  # - lm_head
+  # - embed_tokens
+
+lora_mlp_kernel: true
+lora_qkv_kernel: true
+lora_o_kernel: true
+
+chat_template: llama4
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 4096  # up to 8k will work on a single H100
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 1e-4
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+flash_attention: true
+
+gradient_checkpointing: offload
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+
+warmup_steps: 20
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot|>

examples/llama-4/do-no-use-fa2/scout-vision-qlora-fsdp.yaml

@@ -0,0 +1,88 @@
+base_model: axolotl-quants/Llama-4-Scout-17B-16E-Linearized-bnb-nf4-bf16
+model_type: Llama4ForConditionalGeneration
+processor_type: Llama4Processor
+  # Automatically upload checkpoint and final model to HF
+  # hub_model_id: username/custom_model_name
+
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+sequence_len: 4096
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_glu_activation: true
+liger_rms_norm: true
+liger_layer_norm: true
+
+llama4_linearized_experts: true  # use Axolotl's customized model
+load_in_4bit: true
+adapter: qlora
+lora_r: 32
+lora_alpha: 64
+lora_target_modules:
+  - self_attn.q_proj
+  - self_attn.k_proj
+  - self_attn.v_proj
+  - self_attn.o_proj
+  - shared_expert.gate_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  - vision_adapter.mlp.fc1
+  - vision_adapter.mlp.fc2
+  # - experts.gate_projs.[0-9]+$
+  # - experts.up_projs.[0-9]+$
+  # - experts.down_projs.[0-9]+$
+lora_modules_to_save:
+  - lm_head
+  - embed_tokens
+
+chat_template: llama4
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+    field_messages: messages
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 100
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+fsdp:
+  - auto_wrap
+  - full_shard
+fsdp_config:
+  fsdp_transformer_layer_cls_to_wrap: Llama4TextDecoderLayer
+  fsdp_limit_all_gathers: true
+  fsdp_sync_module_states: true
+  fsdp_offload_params: true
+  fsdp_use_orig_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_state_dict_type: FULL_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_activation_checkpointing: true
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot|>

examples/llama-4/scout-qlora-flexattn-fsdp2.yaml

@@ -1,13 +1,18 @@
-base_model: meta-llama/Llama-4-Scout-17B-16E
+base_model: axolotl-quants/Llama-4-Scout-17B-16E-Linearized-bnb-nf4-bf16
 model_type: Llama4ForConditionalGeneration
-  # Automatically upload checkpoint and final model to HF
+# Automatically upload checkpoint and final model to HF
-  # hub_model_id: username/custom_model_name
+# hub_model_id: username/custom_model_name
 
-strict: false
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
 
-  # torch_compile: true
+liger_glu_activation: true
+liger_rms_norm: true
+liger_layer_norm: true
 
-adapter: lora
+llama4_linearized_experts: true
+load_in_4bit: true
+adapter: qlora
 lora_r: 32
 lora_alpha: 64
 lora_target_modules:
@@ -15,9 +20,15 @@ lora_target_modules:
   - self_attn.k_proj
   - self_attn.v_proj
   - self_attn.o_proj
+  - shared_expert.gate_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  # - experts.gate_projs.[0-9]+$
+  # - experts.up_projs.[0-9]+$
+  # - experts.down_projs.[0-9]+$
 lora_modules_to_save:
-  - lm_head
+  # - lm_head
-  - embed_tokens
+  # - embed_tokens
 
 chat_template: llama4
 datasets:
@@ -38,23 +49,23 @@ sample_packing: true
 pad_to_sequence_len: true
 
 gradient_accumulation_steps: 1
-micro_batch_size: 1
+micro_batch_size: 2
-num_epochs: 1
+num_epochs: 3
-optimizer: adamw_torch_8bit
+optimizer: adamw_torch_4bit
 lr_scheduler: cosine
-learning_rate: 2e-5
+learning_rate: 1e-4
 
 bf16: true
 tf32: true
 
-# gradient_checkpointing: true
-# gradient_checkpointing_kwargs:
-#   use_reentrant: false
 logging_steps: 1
-flash_attention: true
+flex_attention: true
+flex_attn_compile_kwargs:
+  dynamic: false
+  mode: max-autotune-no-cudagraphs
 
-warmup_steps: 100
+warmup_steps: 10
-evals_per_epoch: 2
+evals_per_epoch: 1
 saves_per_epoch: 1
 weight_decay: 0.0
 fsdp:

examples/llama-4/scout-qlora-single-h100-flex.yaml

@@ -0,0 +1,85 @@
+base_model: axolotl-quants/Llama-4-Scout-17B-16E-Linearized-bnb-nf4-bf16
+model_type: Llama4ForConditionalGeneration
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+liger_glu_activation: true
+liger_rms_norm: true
+liger_layer_norm: true
+cut_cross_entropy: true
+
+llama4_linearized_experts: true  # needed with custom linearized experts model
+load_in_4bit: true
+adapter: qlora
+lora_r: 32
+lora_alpha: 64
+lora_target_modules:
+  - self_attn.q_proj
+  - self_attn.k_proj
+  - self_attn.v_proj
+  - self_attn.o_proj
+  - shared_expert.gate_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  # - experts.gate_projs.[0-9]+$  # optionally train the moe experts
+  # - experts.up_projs.[0-9]+$
+  # - experts.down_projs.[0-9]+$
+lora_modules_to_save:
+  # - lm_head  # needed if modifying vocabulary
+  # - embed_tokens
+
+lora_mlp_kernel: true
+lora_qkv_kernel: true
+lora_o_kernel: true
+
+chat_template: llama4
+datasets:
+  - path: mlabonne/FineTome-100k
+    type: chat_template
+    split: train[:20%]
+    field_messages: conversations
+    message_property_mappings:
+      role: from
+      content: value
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+sequence_len: 4096  # up to 8k will work on a single H100
+sample_packing: true
+pad_to_sequence_len: true
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 1e-4
+
+bf16: true
+tf32: true
+
+torch_compile: true
+flex_attention: true
+flex_attn_compile_kwargs:
+  dynamic: false
+  mode: max-autotune-no-cudagraphs
+
+gradient_checkpointing: offload
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+
+logging_steps: 1
+warmup_steps: 20
+evals_per_epoch: 1
+saves_per_epoch: 1
+
+weight_decay: 0.0
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot|>

examples/llama-4/scout-vision-qlora-fsdp2-flex.yaml

@@ -0,0 +1,89 @@
+base_model: axolotl-quants/Llama-4-Scout-17B-16E-Linearized-bnb-nf4-bf16
+model_type: Llama4ForConditionalGeneration
+processor_type: Llama4Processor
+# Automatically upload checkpoint and final model to HF
+# hub_model_id: username/custom_model_name
+
+# these 3 lines are needed for now to handle vision chat templates w images
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+sequence_len: 4096
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+
+liger_glu_activation: true
+liger_rms_norm: true
+liger_layer_norm: true
+
+llama4_linearized_experts: true  # use Axolotl's customized model
+load_in_4bit: true
+adapter: qlora
+lora_r: 32
+lora_alpha: 64
+lora_target_modules:
+  - self_attn.q_proj
+  - self_attn.k_proj
+  - self_attn.v_proj
+  - self_attn.o_proj
+  - shared_expert.gate_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  - vision_adapter.mlp.fc1
+  - vision_adapter.mlp.fc2
+  # - experts.gate_projs.[0-9]+$
+  # - experts.up_projs.[0-9]+$
+  # - experts.down_projs.[0-9]+$
+lora_modules_to_save:
+  - lm_head
+  - embed_tokens
+
+chat_template: llama4
+datasets:
+  - path: HuggingFaceH4/llava-instruct-mix-vsft
+    type: chat_template
+    split: train[:1%]
+    field_messages: messages
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.0
+output_dir: ./outputs/out
+
+gradient_accumulation_steps: 1
+micro_batch_size: 1
+num_epochs: 1
+optimizer: adamw_torch_4bit
+lr_scheduler: cosine
+learning_rate: 1e-4
+
+bf16: true
+tf32: true
+
+logging_steps: 1
+flex_attention: true
+flex_attn_compile_kwargs:
+  dynamic: false
+  mode: max-autotune-no-cudagraphs
+
+warmup_steps: 10
+evals_per_epoch: 1
+saves_per_epoch: 1
+weight_decay: 0.0
+fsdp:
+  - auto_wrap
+  - full_shard
+fsdp_config:
+  fsdp_version: 2
+  fsdp_offload_params: false
+  fsdp_cpu_ram_efficient_loading: true
+  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
+  fsdp_transformer_layer_cls_to_wrap: Llama4TextDecoderLayer
+  fsdp_state_dict_type: SHARDED_STATE_DICT
+  fsdp_sharding_strategy: FULL_SHARD
+  fsdp_reshard_after_forward: true
+  fsdp_activation_checkpointing: true
+special_tokens:
+  pad_token: <|finetune_right_pad_id|>
+  eos_token: <|eot|>

examples/llava/lora-7b.yaml

@@ -1,6 +1,5 @@
 base_model: llava-hf/llava-1.5-7b-hf
 processor_type: AutoProcessor
-strict: false
 
 # these 3 lines are needed for now to handle vision chat templates w images
 skip_prepare_dataset: true

examples/mamba/config.yml

@@ -5,7 +5,6 @@ tokenizer_type: AutoTokenizer
 tokenizer_config: EleutherAI/gpt-neox-20b
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

examples/mistral/bigstral-ds-zero3.yaml

@@ -6,7 +6,6 @@ tokenizer_type: LlamaTokenizer
 # hub_model_id: username/custom_model_name
 
 trust_remote_code: true
-strict: false
 
 unfrozen_parameters:
   - ^lm_head.weight$

examples/mistral/config.yml

@@ -4,7 +4,6 @@ model_type: MistralForCausalLM
 tokenizer_type: LlamaTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

examples/mistral/lora-mps.yml

@@ -4,7 +4,6 @@ model_type: MistralForCausalLM
 tokenizer_type: LlamaTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

examples/mistral/lora.yml

@@ -7,7 +7,6 @@ tokenizer_type: LlamaTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

examples/mistral/mistral-dpo-qlora.yml

@@ -12,7 +12,6 @@ tokenizer_type: LlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 chat_template: chatml
 rl: dpo

examples/mistral/mistral-qlora-fsdp.yml

@@ -9,7 +9,6 @@ trust_remote_code: true
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: tatsu-lab/alpaca

examples/mistral/mistral-qlora-orpo.yml

@@ -7,7 +7,6 @@ tokenizer_type: LlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 rl: orpo
 orpo_alpha: 0.1

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

@@ -1,6 +1,5 @@
 base_model: mistralai/Mistral-Small-3.1-24B-Instruct-2503
 processor_type: AutoProcessor
-strict: false
 
 load_in_8bit: true
 

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

@@ -7,7 +7,6 @@ tokenizer_type: LlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: tatsu-lab/alpaca

examples/mistral/mixtral-qlora-fsdp.yml

@@ -9,7 +9,6 @@ trust_remote_code: true
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: tatsu-lab/alpaca

examples/mistral/mixtral.yml

@@ -9,7 +9,6 @@ trust_remote_code: true
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: tatsu-lab/alpaca

examples/mistral/mixtral_22.yml

@@ -6,7 +6,6 @@ tokenizer_type: LlamaTokenizer
 # hub_model_id: username/custom_model_name
 
 trust_remote_code: true
-strict: false
 
 unfrozen_parameters:
   - ^lm_head.weight$

examples/mistral/qlora.yml

@@ -7,7 +7,6 @@ tokenizer_type: LlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: mhenrichsen/alpaca_2k_test

examples/openllama-3b/config.yml

@@ -4,7 +4,6 @@ model_type: LlamaForCausalLM
 tokenizer_type: LlamaTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-strict: false
 push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned

examples/openllama-3b/lora.yml

@@ -7,7 +7,6 @@ tokenizer_type: LlamaTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned

examples/openllama-3b/qlora.yml

@@ -7,7 +7,6 @@ tokenizer_type: LlamaTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 push_dataset_to_hub:
 datasets:
   - path: teknium/GPT4-LLM-Cleaned

examples/phi/lora-3.5.yaml

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: true
 load_in_4bit: false
-strict: false
 
 chat_template: phi_3
 datasets:

examples/phi/phi-ft.yml

@@ -4,7 +4,6 @@ model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-strict: false
 
 datasets:
   - path: garage-bAInd/Open-Platypus

examples/phi/phi-qlora.yml

@@ -7,7 +7,6 @@ tokenizer_type: AutoTokenizer
 
 load_in_8bit: false
 load_in_4bit: true
-strict: false
 
 datasets:
   - path: garage-bAInd/Open-Platypus

examples/phi/phi2-ft.yml

@@ -4,7 +4,6 @@ model_type: AutoModelForCausalLM
 tokenizer_type: AutoTokenizer
 # Automatically upload checkpoint and final model to HF
 # hub_model_id: username/custom_model_name
-strict: false
 
 datasets:
   - path: garage-bAInd/Open-Platypus