offload activations to disk instead of CPU RAM

* don't sort multipack sampler

* increased packing efficiency increases loss

---------

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

.github/workflows/tests-nightly.yml

@@ -18,9 +18,96 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
+  preload-cache:
+    name: Preload HF cache
+    runs-on: ubuntu-latest
+    strategy:
+      fail-fast: false
+      matrix:
+        python_version: ["3.11"]
+        pytorch_version: ["2.6.0"]
+    timeout-minutes: 20
+
+    env:
+      AXOLOTL_IS_CI_CACHE_PRELOAD: "1"
+
+    steps:
+      - name: Check out repository code
+        uses: actions/checkout@v4
+
+      - name: Restore HF cache
+        id: hf-cache-restore
+        uses: actions/cache/restore@v4
+        with:
+          path: |
+            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/models--*
+          key: ${{ runner.os }}-hf-hub-cache-v2
+
+      - name: Setup Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python_version }}
+          cache: 'pip' # caching pip dependencies
+
+      - name: upgrade pip
+        run: |
+          pip3 install --upgrade pip
+          pip3 install --upgrade packaging==23.2 setuptools==75.8.0 wheel
+
+      - name: Install PyTorch
+        run: |
+          pip3 install torch==${{ matrix.pytorch_version }}
+
+      - name: Install dependencies
+        run: |
+          pip3 show torch
+          pip3 install --no-build-isolation -U -e .
+          python scripts/unsloth_install.py | sh
+          python scripts/cutcrossentropy_install.py | sh
+          pip3 install -r requirements-dev.txt -r requirements-tests.txt
+
+      - name: Make sure PyTorch version wasn't clobbered
+        run: |
+          python -c "import torch; assert '${{ matrix.pytorch_version }}' in torch.__version__"
+
+      - name: Ensure axolotl CLI was installed
+        run: |
+          axolotl --help
+
+      - name: Pre-Download dataset fixture
+        run: |
+          huggingface-cli download --repo-type=dataset axolotl-ai-internal/axolotl-oss-dataset-fixtures
+
+      - name: Run tests
+        run: |
+          pytest -v tests/conftest.py
+
+      - name: Upload coverage to Codecov
+        uses: codecov/codecov-action@v5
+        with:
+          token: ${{ secrets.CODECOV_TOKEN }}
+          files: ./coverage.xml
+          flags: unittests,pytorch-${{ matrix.pytorch_version }}
+          fail_ci_if_error: false
+
+      - name: cleanup pip cache
+        run: |
+          find "$(pip cache dir)/http-v2" -type f -mtime +14 -exec rm {} \;
+
+      - name: Save HF cache
+        id: hf-cache
+        uses: actions/cache/save@v4
+        with:
+          path: |
+            /home/runner/.cache/huggingface/hub/datasets--*
+            /home/runner/.cache/huggingface/hub/models--*
+          key: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
+
   pytest:
     name: PyTest
     runs-on: ubuntu-latest
+    needs: [preload-cache]
     strategy:
       fail-fast: false
       max-parallel: 2

docs/config.qmd

@@ -612,6 +612,7 @@ lr_div_factor: # Learning rate div factor
 # - optimi_adamw
 # - ao_adamw_8bit
 # - ao_adamw_fp8
+# - came_pytorch
 optimizer:
 # Dictionary of arguments to pass to the optimizer
 optim_args:

examples/cerebras/btlm-ft.yml

@@ -59,7 +59,9 @@ gradient_checkpointing: false
 resume_from_checkpoint:
 logging_steps: 1
 
-attention: flash
+flash_attention: true
+sdp_attention:
+flash_optimum:
 
 gptq_groupsize:
 gptq_model_v1:

examples/cerebras/qlora.yml

@@ -39,7 +39,8 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: xformers
+xformers_attention: true
+flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10

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

@@ -45,8 +45,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -46,8 +46,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -45,8 +45,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -46,8 +46,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -45,8 +45,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -46,8 +46,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -49,8 +49,7 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_ratio: 0.1
 evals_per_epoch:

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

@@ -112,7 +112,9 @@
     "early_stopping_patience:\n",
     "resume_from_checkpoint:\n",
     "logging_steps: 1\n",
-    "attention: sdpa\n",
+    "xformers_attention:\n",
+    "flash_attention: false\n",
+    "sdp_attention: true\n",
     "\n",
     "warmup_steps: 1\n",
     "max_steps: 25\n",

examples/dbrx/16bit-lora.yaml

@@ -52,8 +52,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch:

examples/dbrx/8bit-lora.yaml

@@ -55,8 +55,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch:

examples/dbrx/fft-ds-zero3.yaml

@@ -39,8 +39,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch:

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

@@ -35,8 +35,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 100
 evals_per_epoch: 2

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

@@ -59,8 +59,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 100
 evals_per_epoch: 2

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

@@ -43,7 +43,8 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: xformers
+xformers_attention: true
+flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 40

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

@@ -73,7 +73,8 @@ early_stopping_patience: 3
 resume_from_checkpoint:
 auto_resume_from_checkpoints: true
 logging_steps: 1
-attention: xformers
+xformers_attention: true
+flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10

examples/falcon/config-7b.yml

@@ -40,7 +40,8 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: xformers
+xformers_attention: true
+flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 40

examples/gemma/qlora.yml

@@ -47,8 +47,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_ratio: 0.1
 evals_per_epoch: 4

examples/gemma2/qlora.yml

@@ -53,8 +53,7 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_ratio: 0.1
 evals_per_epoch:

examples/gemma2/reward-model.yaml

@@ -43,8 +43,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_ratio: 0.1
 evals_per_epoch:

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

@@ -57,8 +57,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_ratio: 0.1
 evals_per_epoch:

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

@@ -51,7 +51,8 @@ gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
 logging_steps: 1
-attention: flash
+flash_attention: true
+eager_attention:
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

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

@@ -53,7 +53,8 @@ gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: false
 logging_steps: 1
-attention: flash
+flash_attention: true
+eager_attention:
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

examples/gptj/qlora.yml

@@ -36,7 +36,8 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: xformers
+xformers_attention: true
+flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 10

examples/jamba/qlora.yaml

@@ -47,8 +47,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch:

examples/jamba/qlora_deepspeed.yaml

@@ -46,8 +46,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch:

examples/jamba/qlora_fsdp_large.yaml

@@ -45,8 +45,7 @@ gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: true
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 1

examples/jeopardy-bot/config.yml

@@ -37,7 +37,8 @@ bf16: auto
 tf32: true
 resume_from_checkpoint:
 logging_steps: 5
-attention: xformers
+xformers_attention: true
+flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20

examples/llama-2/fft_optimized.yml

@@ -42,8 +42,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 flash_attn_cross_entropy: false
 flash_attn_rms_norm: true
 flash_attn_fuse_qkv: false

examples/llama-2/gptq-lora.yml

@@ -53,7 +53,9 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention:
+sdp_attention:
+flash_optimum:
 warmup_steps: 100
 evals_per_epoch: 4
 saves_per_epoch: 1

examples/llama-2/lisa.yml

@@ -46,8 +46,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 flash_attn_cross_entropy: false
 flash_attn_rms_norm: true
 flash_attn_fuse_qkv: false

examples/llama-2/loftq.yml

@@ -45,8 +45,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-2/lora.yml

@@ -45,8 +45,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-2/qlora-fsdp.yml

@@ -48,8 +48,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-2/qlora.yml

@@ -46,8 +46,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-2/relora.yml

@@ -48,8 +48,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -50,7 +50,8 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-attention: flash
+flash_attention: true
+eager_attention:
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

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

@@ -49,8 +49,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 100
 evals_per_epoch: 2

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

@@ -34,8 +34,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 100
 evals_per_epoch: 2

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

@@ -61,8 +61,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -56,8 +56,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -77,8 +77,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -53,8 +53,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -54,8 +54,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

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

@@ -48,8 +48,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

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

@@ -55,8 +55,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -48,8 +48,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

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

@@ -49,8 +49,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -53,8 +53,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 20
 evals_per_epoch: 4

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

@@ -51,8 +51,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

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

@@ -39,8 +39,7 @@ gradient_checkpointing: true
 gradient_checkpointing_kwargs:
   use_reentrant: true
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

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

@@ -48,8 +48,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-3/qlora.yml

@@ -46,8 +46,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/llama-4/README.md

@@ -34,3 +34,5 @@ We provide a script to delinearize Llama 4 linearized models into regular Huggin
 ```bash
 axolotl delinearize-llama4 --model path/to/model_dir --output path/to/output_dir
 ```
+
+Note: This only works with the non-quantized linearized model. If you have an adapter, merge it with the *non-quantized linearized* model before delinearizing.

examples/llava/lora-7b.yaml

@@ -46,7 +46,8 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-attention: flash
+flash_attention: true
+eager_attention:
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

examples/mamba/config.yml

@@ -39,7 +39,7 @@ tf32: true
 gradient_checkpointing: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: eager
+flash_attention:
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/mistral/bigstral-ds-zero3.yaml

@@ -42,8 +42,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 save_total_limit: 1
 save_steps:

examples/mistral/config.yml

@@ -36,8 +36,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/mistral/lora-mps.yml

@@ -53,7 +53,8 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: sdpa
+flash_attention: false
+sdp_attention: true
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/mistral/lora.yml

@@ -54,8 +54,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/mistral/mistral-dpo-qlora.yml

@@ -71,7 +71,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: eager
+flash_attention: false
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/mistral/mistral-qlora-fsdp.yml

@@ -51,8 +51,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/mistral/mistral-qlora-orpo.yml

@@ -59,8 +59,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

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

@@ -48,7 +48,9 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-attention: eager  # PixtralVisionModel does not support Flash Attention 2.0 yet.
+flash_attention: false # PixtralVisionModel does not support Flash Attention 2.0 yet.
+eager_attention:
+
 warmup_ratio: 0.1
 evals_per_epoch: 1
 saves_per_epoch: 1

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

@@ -49,8 +49,7 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/mistral/mixtral-qlora-fsdp.yml

@@ -51,8 +51,7 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/mistral/mixtral.yml

@@ -69,8 +69,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/mistral/mixtral_22.yml

@@ -40,8 +40,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 save_total_limit: 1
 save_steps:

examples/mistral/qlora.yml

@@ -54,8 +54,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/mpt-7b/config.yml

@@ -39,7 +39,7 @@ bf16: auto
 tf32: true
 resume_from_checkpoint:
 logging_steps: 5
-attention: eager
+flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20

examples/openllama-3b/config.yml

@@ -39,8 +39,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20

examples/openllama-3b/lora.yml

@@ -47,8 +47,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20

examples/openllama-3b/qlora.yml

@@ -40,8 +40,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20

examples/phi/phi-ft.yml

@@ -48,8 +48,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: True
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 100
 evals_per_epoch: 4

examples/phi/phi-qlora.yml

@@ -51,8 +51,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: True
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 100
 evals_per_epoch: 4

examples/phi/phi2-ft.yml

@@ -48,8 +48,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: True
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 100
 evals_per_epoch: 4

examples/phi/phi3-ft-fsdp.yml

@@ -49,8 +49,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 100
 evals_per_epoch: 4

examples/phi/phi3-ft.yml

@@ -44,8 +44,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: True
 early_stopping_patience: 3
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 eval_steps: 1000
 save_steps: 5000

examples/pixtral/lora-12b.yml

@@ -46,7 +46,8 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-attention: eager  # PixtralVisionModel does not support Flash Attention 2.0 yet
+flash_attention: false # PixtralVisionModel does not support Flash Attention 2.0 yet
+eager_attention:
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

examples/qwen/lora.yml

@@ -47,7 +47,7 @@ tf32: false
 gradient_checkpointing: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: eager
+flash_attention:
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/qwen/qlora.yml

@@ -47,7 +47,7 @@ tf32: false
 gradient_checkpointing: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention:
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/qwen/qwen2-moe-lora.yaml

@@ -43,8 +43,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/qwen/qwen2-moe-qlora.yaml

@@ -46,8 +46,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/qwen2-vl/lora-7b.yaml

@@ -46,7 +46,8 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-attention: flash
+flash_attention: true
+eager_attention:
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

examples/qwen2/dpo.yaml

@@ -49,8 +49,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/qwen2/prm.yaml

@@ -47,8 +47,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_ratio: 0.1
 evals_per_epoch:

examples/qwen2/qlora-fsdp.yaml

@@ -47,8 +47,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/qwen2/reward-model.yaml

@@ -43,8 +43,7 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_ratio: 0.1
 evals_per_epoch:

examples/redpajama/config-3b.yml

@@ -40,7 +40,7 @@ bf16: auto
 tf32: true
 resume_from_checkpoint:
 logging_steps: 5
-attention: flash
+flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20

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

@@ -38,7 +38,7 @@ tf32: true
 gradient_checkpointing:
 resume_from_checkpoint:
 logging_steps: 1
-attention: eager
+flash_attention:
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20

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

@@ -44,8 +44,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 flash_attn_cross_entropy: false
 flash_attn_rms_norm: true
 flash_attn_fuse_qkv: false

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

@@ -47,8 +47,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 flash_attn_cross_entropy: false
 flash_attn_rms_norm: true
 

examples/starcoder2/qlora.yml

@@ -46,8 +46,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 20
 evals_per_epoch: 4

examples/tiny-llama/lora-mps.yml

@@ -47,7 +47,7 @@ tf32: true
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: eager
+flash_attention: false
 
 warmup_steps: 10
 evals_per_epoch: 0

examples/tiny-llama/lora.yml

@@ -45,8 +45,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/tiny-llama/pretrain.yml

@@ -36,8 +36,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch:

examples/tiny-llama/qlora.yml

@@ -47,8 +47,7 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-attention: flash
+flash_attention: true
-
 
 warmup_steps: 10
 evals_per_epoch: 4