chore: refactor normalize_attn to use mapping and loop

* fsdp embeddings should be float32 per comment

* patch peft to not upcast everything

* add tabs back to code check

* fix import

* add configurable option and fix check

* add check for dtypes

* move embeddings test to patch dir

* fix test

* fix comment and logic

.github/workflows/tests.yml

@@ -44,12 +44,98 @@ jobs:
         env:
           SKIP: no-commit-to-branch
 
-  pytest:
-    name: PyTest
+  preload-cache:
+    name: Preload HF cache
     runs-on: ubuntu-latest
     strategy:
       fail-fast: false
-      max-parallel: 2
+      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: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
+
+  pytest:
+    name: PyTest
+    runs-on: ubuntu-latest
+    needs: [preload-cache]
+    strategy:
+      fail-fast: false
       matrix:
         python_version: ["3.11"]
         pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
@@ -121,21 +207,12 @@ jobs:
         run: |
           find "$(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: ${{ steps.hf-cache-restore.outputs.cache-primary-key }}
-
   pytest-sdist:
     name: PyTest from Source Dist
     runs-on: ubuntu-latest
+    needs: [preload-cache]
     strategy:
       fail-fast: false
-      max-parallel: 1
       matrix:
         python_version: ["3.11"]
         pytorch_version: ["2.5.1", "2.6.0", "2.7.0"]
@@ -199,15 +276,6 @@ jobs:
         run: |
           find "$(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 }}
-
   docker-e2e-tests-1st:
     if: ${{ ! contains(github.event.commits[0].message, '[skip e2e]') && github.repository_owner == 'axolotl-ai-cloud' }}
     # this job needs to be run on self-hosted GPU runners...

docs/config.qmd

@@ -32,6 +32,8 @@ tokenizer_legacy:
 resize_token_embeddings_to_32x:
 # Optional[bool] Whether to shrink the embeddings to len(tokenizer). By default, we won't shrink.
 shrink_embeddings:
+# Optional[bool] Don't upcast the embeddings to float32 when using PEFT. Useful for low-VRAM GPUs
+embeddings_skip_upcast:
 # Whether to load the model with randomly initialized weights. Useful for
 # pre-training a model from scratch or debugging purposes.
 random_init_weights:
@@ -73,11 +75,12 @@ load_in_8bit: true
 load_in_4bit:
 
 # Use CUDA bf16
-bf16: true # bool or 'full' for `bf16_full_eval`. require >=ampere
+bf16: true # bool or 'full' for `bf16_full_eval`, or 'auto' for automatic detection. require >=ampere
 # Use CUDA fp16
 fp16: true
 # Use CUDA tf32
 tf32: true # require >=ampere
+# Note: if bf16 is set to 'auto', and fp16 is set to true, we will prefer the explict fp16 setting
 
 # No AMP (automatic mixed precision)
 bfloat16: true # require >=ampere
@@ -184,8 +187,8 @@ datasets:
     # adding a system turn with empty content.
     drop_system_message:
 
-    # Optional[bool]. Whether to split the assistant turn based on a reasoning trace inside delimited tags
-    # defaults to False
+    # Optional[bool]. (for Qwen3 template only) Whether to split the assistant content based on a reasoning trace inside delimited tags
+    # See example at `docs/dataset-formats/conversation.qmd`
     split_thinking:
 
     # IMPORTANT: The following fields determine which parts of the conversation to train on.
@@ -547,7 +550,7 @@ gradient_checkpointing: false
 early_stopping_patience: 3
 
 # Specify a scheduler and kwargs to use with the optimizer
-lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | empty for cosine
+lr_scheduler: # 'one_cycle' | 'rex' | 'log_sweep' | 'linear' | 'cosine_with_restarts' | 'polynomial' | 'constant' | 'constant_with_warmup' | 'inverse_sqrt' | 'reduce_lr_on_plateau' | 'cosine_with_min_lr' | 'warmup_stable_decay' | empty for cosine
 lr_scheduler_kwargs:
 cosine_min_lr_ratio: # decay lr to some percentage of the peak lr, e.g. cosine_min_lr_ratio=0.1 for 10% of peak lr
 cosine_constant_lr_ratio: # freeze lr at some percentage of the step, e.g. cosine_constant_lr_ratio=0.8 means start cosine_min_lr at 80% of training step (https://arxiv.org/pdf/2308.04014.pdf)

docs/dataset-formats/conversation.qmd

@@ -196,6 +196,34 @@ datasets:
 It is not necessary to set both `message_field_training` and `message_field_training_detail` at once.
 :::
 
+8. (For Qwen3 template only) Enable reasoning split, where the reasoning is split from the content and passed as a separate field into the template.
+
+```yaml
+datasets:
+  - path: ...
+    type: chat_template
+    chat_template: qwen3
+    split_thinking: true
+```
+
+For example, a content can look like:
+
+```json
+{
+  "content": "<think>Some thinking outputs</think>Output after thinking."
+}
+```
+
+After split, it will look like:
+
+```json
+{
+  "reasoning_content": "Some thinking outputs",
+  "content": "Output after thinking..."
+}
+```
+
+
 ## sharegpt
 
 ::: {.callout-important}

examples/cerebras/btlm-ft.yml

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

examples/cerebras/qlora.yml

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

examples/dbrx/16bit-lora.yaml

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

examples/dbrx/8bit-lora.yaml

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

examples/dbrx/fft-ds-zero3.yaml

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

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

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

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

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

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

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

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

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

examples/falcon/config-7b.yml

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

examples/gemma/qlora.yml

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

examples/gemma2/qlora.yml

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

examples/gemma2/reward-model.yaml

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

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

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

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

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

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

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

examples/gptj/qlora.yml

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

examples/jamba/qlora.yaml

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

examples/jamba/qlora_deepspeed.yaml

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

examples/jamba/qlora_fsdp_large.yaml

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

examples/jeopardy-bot/config.yml

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

examples/llama-2/fft_optimized.yml

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

examples/llama-2/gptq-lora.yml

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

examples/llama-2/lisa.yml

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

examples/llama-2/loftq.yml

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

examples/llama-2/lora.yml

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

examples/llama-2/qlora-fsdp.yml

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

examples/llama-2/qlora.yml

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

examples/llama-2/relora.yml

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

examples/llama-3/qlora.yml

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

examples/llava/lora-7b.yaml

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

examples/mistral/bigstral-ds-zero3.yaml

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

examples/mistral/config.yml

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

examples/mistral/lora-mps.yml

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

examples/mistral/lora.yml

@@ -54,7 +54,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 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
-flash_attention: false
+attention: eager
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/mistral/mistral-qlora-fsdp.yml

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

examples/mistral/mistral-qlora-orpo.yml

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

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

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

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

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

examples/mistral/mixtral-qlora-fsdp.yml

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

examples/mistral/mixtral.yml

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

examples/mistral/mixtral_22.yml

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

examples/mistral/qlora.yml

@@ -54,7 +54,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 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
-flash_attention:
+attention: eager
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20

examples/openllama-3b/config.yml

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

examples/openllama-3b/lora.yml

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

examples/openllama-3b/qlora.yml

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

examples/orpheus/README.md

@@ -0,0 +1,341 @@
+# Finetuning LLMs to output audio
+
+In this example, we finetune Orpcanopylabs/orpheus-tts-0.1-pretrained (a LLaMA 3.2 3b model) to output audio.
+
+The `finetune.yml` withe current settings will run on any Nvidia GPU with 45GB VRAM or more. If you adjust the batch size it can easily run on any GPU under 24GB.
+
+## Dataset pre-processing for pre-training
+If you are adding another voice in English, please jump ahead to finetuning pre-processing.
+
+For this to work, we need to preprocess our dataset. Since we are expecting to output audio, we will need to add tokens to the tokenizer.
+
+Using this code, it will download the SNAC model and add the correct tokens and upload the final dataset.
+
+```python
+import torch
+from snac import SNAC
+from datasets import load_dataset
+from huggingface_hub import snapshot_download
+from datasets import load_dataset
+import random
+import torchaudio.transforms as T
+from transformers import AutoTokenizer
+import os
+
+my_original_dataset_name = "<huggingface-id-of-dataset-that-we-want-to-preprocess>"
+name_to_push_dataset_to = "<huggingface-id-of-where-to-save-dataset>"
+
+dsn = my_original_dataset_name
+
+snapshot_download(
+    repo_id=dsn,
+    repo_type="dataset",
+    revision="main",
+    max_workers=64,
+)
+
+
+ds = load_dataset(dsn, split="train")
+ds_sample_rate = ds[0]["audio"]["sampling_rate"]
+
+model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")
+model = model.to("mps")
+
+def tokenise_audio(waveform):
+  waveform = torch.from_numpy(waveform).unsqueeze(0)
+  waveform = waveform.to(dtype=torch.float32)
+  resample_transform = T.Resample(orig_freq=ds_sample_rate, new_freq=24000)
+  waveform = resample_transform(waveform)
+
+  waveform = waveform.unsqueeze(0).to("cuda")
+
+  #generate the codes from snac
+  with torch.inference_mode():
+    codes = model.encode(waveform)
+
+  all_codes = []
+  for i in range(codes[0].shape[1]):
+    all_codes.append(codes[0][0][i].item()+128266)
+    all_codes.append(codes[1][0][2*i].item()+128266+4096)
+    all_codes.append(codes[2][0][4*i].item()+128266+(2*4096))
+    all_codes.append(codes[2][0][(4*i)+1].item()+128266+(3*4096))
+    all_codes.append(codes[1][0][(2*i)+1].item()+128266+(4*4096))
+    all_codes.append(codes[2][0][(4*i)+2].item()+128266+(5*4096))
+    all_codes.append(codes[2][0][(4*i)+3].item()+128266+(6*4096))
+
+
+  return all_codes
+
+def add_codes(example):
+    # Always initialize codes_list to None
+    codes_list = None
+
+    try:
+        answer_audio = example.get("audio")
+        # If there's a valid audio array, tokenise it
+        if answer_audio and "array" in answer_audio:
+            audio_array = answer_audio["array"]
+            codes_list = tokenise_audio(audio_array)
+    except Exception as e:
+        print(f"Skipping row due to error: {e}")
+        # Keep codes_list as None if we fail
+    example["codes_list"] = codes_list
+
+    return example
+
+ds = ds.map(add_codes, remove_columns=["audio"])
+
+#@title Load Tokenizer
+tokeniser_length = 128256
+start_of_text = 128000
+end_of_text = 128009
+
+start_of_speech = tokeniser_length + 1
+end_of_speech = tokeniser_length + 2
+
+start_of_human = tokeniser_length + 3
+end_of_human = tokeniser_length + 4
+
+start_of_ai = tokeniser_length + 5
+end_of_ai =  tokeniser_length + 6
+pad_token = tokeniser_length + 7
+
+audio_tokens_start = tokeniser_length + 10
+
+tokenizer_name = "canopylabs/orpheus-3b-0.1-pretrained"
+
+
+tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
+num_proc = os.cpu_count() - 2
+
+ds = ds.filter(lambda x: x["codes_list"] is not None)
+ds = ds.filter(lambda x: len(x["codes_list"]) > 0)
+
+#@title Create Input Ids
+def remove_duplicate_frames(example):
+    vals = example["codes_list"]
+    if len(vals) % 7 != 0:
+        raise ValueError("Input list length must be divisible by 7")
+
+    result = vals[:7]
+
+    removed_frames = 0
+
+    for i in range(7, len(vals), 7):
+        current_first = vals[i]
+        previous_first = result[-7]
+
+        if current_first != previous_first:
+            result.extend(vals[i:i+7])
+        else:
+            removed_frames += 1
+
+    example["codes_list"] = result
+
+    return example
+
+ds = ds.map(remove_duplicate_frames, num_proc=num_proc)
+
+
+def create_input_ids(example):
+    text_ids = tokenizer.encode({example['text']},  add_special_tokens=True)
+    text_ids.append(end_of_text)
+    example["text_tokens"] = text_ids
+    input_ids = (
+        [start_of_human]
+        + example["text_tokens"]
+        + [end_of_human]
+        + [start_of_ai]
+        + [start_of_speech]
+        + example["codes_list"]
+        + [end_of_speech]
+        + [end_of_ai]
+    )
+    example["input_ids"] = input_ids
+    example["labels"] = input_ids
+    example["attention_mask"] = [1] * len(input_ids)
+
+    return example
+
+ds = ds.map(create_input_ids, num_proc=num_proc, remove_columns=["text", "codes_list"])
+
+#@title Remove unnecessary columns
+columns_to_keep = ["input_ids", "labels", "attention_mask"]
+columns_to_remove = [col for col in ds.column_names if col not in columns_to_keep]
+
+ds = ds.remove_columns(columns_to_remove)
+
+ds.push_to_hub(name_to_push_dataset_to)
+```
+
+
+## Finetune pre-processing
+Use this code to add a new voice.
+
+```python
+import torch
+from snac import SNAC
+from datasets import load_dataset
+from huggingface_hub import snapshot_download
+from datasets import load_dataset
+import random
+import torchaudio.transforms as T
+from transformers import AutoTokenizer
+import os
+
+my_original_dataset_name = "<huggingface-id-of-dataset-that-we-want-to-preprocess>"
+name_to_push_dataset_to = "<huggingface-id-of-where-to-save-dataset>"
+
+dsn = my_original_dataset_name
+
+snapshot_download(
+    repo_id=dsn,
+    repo_type="dataset",
+    revision="main",
+    max_workers=64,
+)
+
+
+ds = load_dataset(dsn, split="train")
+ds_sample_rate = ds[0]["audio"]["sampling_rate"]
+
+model = SNAC.from_pretrained("hubertsiuzdak/snac_24khz")
+model = model.to("mps")
+
+def tokenise_audio(waveform):
+  waveform = torch.from_numpy(waveform).unsqueeze(0)
+  waveform = waveform.to(dtype=torch.float32)
+  resample_transform = T.Resample(orig_freq=ds_sample_rate, new_freq=24000)
+  waveform = resample_transform(waveform)
+
+  waveform = waveform.unsqueeze(0).to("cuda")
+
+  #generate the codes from snac
+  with torch.inference_mode():
+    codes = model.encode(waveform)
+
+  all_codes = []
+  for i in range(codes[0].shape[1]):
+    all_codes.append(codes[0][0][i].item()+128266)
+    all_codes.append(codes[1][0][2*i].item()+128266+4096)
+    all_codes.append(codes[2][0][4*i].item()+128266+(2*4096))
+    all_codes.append(codes[2][0][(4*i)+1].item()+128266+(3*4096))
+    all_codes.append(codes[1][0][(2*i)+1].item()+128266+(4*4096))
+    all_codes.append(codes[2][0][(4*i)+2].item()+128266+(5*4096))
+    all_codes.append(codes[2][0][(4*i)+3].item()+128266+(6*4096))
+
+
+  return all_codes
+
+def add_codes(example):
+    # Always initialize codes_list to None
+    codes_list = None
+
+    try:
+        answer_audio = example.get("audio")
+        # If there's a valid audio array, tokenise it
+        if answer_audio and "array" in answer_audio:
+            audio_array = answer_audio["array"]
+            codes_list = tokenise_audio(audio_array)
+    except Exception as e:
+        print(f"Skipping row due to error: {e}")
+        # Keep codes_list as None if we fail
+    example["codes_list"] = codes_list
+
+    return example
+
+ds = ds.map(add_codes, remove_columns=["audio"])
+
+#@title Load Tokenizer
+tokeniser_length = 128256
+start_of_text = 128000
+end_of_text = 128009
+
+start_of_speech = tokeniser_length + 1
+end_of_speech = tokeniser_length + 2
+
+start_of_human = tokeniser_length + 3
+end_of_human = tokeniser_length + 4
+
+start_of_ai = tokeniser_length + 5
+end_of_ai =  tokeniser_length + 6
+pad_token = tokeniser_length + 7
+
+audio_tokens_start = tokeniser_length + 10
+
+tokenizer_name = "canopylabs/orpheus-3b-0.1-pretrained"
+
+
+tokenizer = AutoTokenizer.from_pretrained(tokenizer_name)
+num_proc = os.cpu_count() - 2
+
+ds = ds.filter(lambda x: x["codes_list"] is not None)
+ds = ds.filter(lambda x: len(x["codes_list"]) > 0)
+
+#@title Create Input Ids
+def remove_duplicate_frames(example):
+    vals = example["codes_list"]
+    if len(vals) % 7 != 0:
+        raise ValueError("Input list length must be divisible by 7")
+
+    result = vals[:7]
+
+    removed_frames = 0
+
+    for i in range(7, len(vals), 7):
+        current_first = vals[i]
+        previous_first = result[-7]
+
+        if current_first != previous_first:
+            result.extend(vals[i:i+7])
+        else:
+            removed_frames += 1
+
+    example["codes_list"] = result
+
+    return example
+
+ds = ds.map(remove_duplicate_frames, num_proc=num_proc)
+
+tok_info = '''*** HERE you can modify the text prompt
+i.e. if you wanted a multispeaker model like canopylabs/orpheus-3b-0.1-ft, you can pass:
+f"{example["source"]}:  {example["text"]}", as is passed.
+'''
+print(tok_info)
+
+def create_input_ids(example):
+    text_ids = tokenizer.encode(f"{example['speaker_id']}: {example['text']}",  add_special_tokens=True)
+    text_ids.append(end_of_text)
+    example["text_tokens"] = text_ids
+    input_ids = (
+        [start_of_human]
+        + example["text_tokens"]
+        + [end_of_human]
+        + [start_of_ai]
+        + [start_of_speech]
+        + example["codes_list"]
+        + [end_of_speech]
+        + [end_of_ai]
+    )
+    example["input_ids"] = input_ids
+    example["labels"] = input_ids
+    example["attention_mask"] = [1] * len(input_ids)
+
+    return example
+
+ds = ds.map(create_input_ids, num_proc=num_proc, remove_columns=["text", "codes_list"])
+
+#@title Remove unnecessary columns
+columns_to_keep = ["input_ids", "labels", "attention_mask"]
+columns_to_remove = [col for col in ds.column_names if col not in columns_to_keep]
+
+ds = ds.remove_columns(columns_to_remove)
+
+ds.push_to_hub(name_to_push_dataset_to)
+```
+
+## Training
+After preprocessing is done, fill out the blanks in finetune.yml and simply run `axolotl train finetune.yml`
+
+## Inference
+For inference, please refer to the original [orpheus github](https://github.com/canopyai/Orpheus-TTS/tree/main).

examples/orpheus/finetune.yml

@@ -0,0 +1,52 @@
+base_model: canopylabs/orpheus-3b-0.1-pretrained
+
+hub_model_id: <your-hub-model-id>
+
+plugins:
+  - axolotl.integrations.liger.LigerPlugin
+liger_rope: true
+liger_rms_norm: true
+liger_glu_activation: true
+liger_fused_linear_cross_entropy: true
+
+datasets:
+  - path: <your-hf-dataset-id>
+    type:  # leave empty to load pre-tokenized
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.01
+output_dir: ./outputs/out
+
+sequence_len: 8192
+sample_packing: true
+pad_to_sequence_len: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 8
+micro_batch_size: 4
+num_epochs: 3
+optimizer: adamw_torch_fused
+lr_scheduler: cosine
+learning_rate: 2e-5
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+gradient_checkpointing_kwargs:
+  use_reentrant: false
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_steps: 20
+evals_per_epoch: 5
+saves_per_epoch: 5
+weight_decay: 0.05
+
+special_tokens:
+  pad_token: <custom_token_7>

examples/phi/phi-ft.yml

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

examples/phi/phi-qlora.yml

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

examples/phi/phi2-ft.yml

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

examples/phi/phi3-ft-fsdp.yml

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

examples/phi/phi3-ft.yml

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

examples/pixtral/lora-12b.yml

@@ -46,8 +46,7 @@ tf32: true
 
 gradient_checkpointing: true
 logging_steps: 1
-flash_attention: false # PixtralVisionModel does not support Flash Attention 2.0 yet
-eager_attention:
+attention: eager  # PixtralVisionModel does not support Flash Attention 2.0 yet
 
 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
-flash_attention:
+attention: eager
 
 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
-flash_attention:
+attention: flash
 
 warmup_steps: 10
 evals_per_epoch: 4

examples/qwen/qwen2-moe-lora.yaml

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

examples/qwen/qwen2-moe-qlora.yaml

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

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

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

examples/qwen2/dpo.yaml

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

examples/qwen2/prm.yaml

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

examples/qwen2/qlora-fsdp.yaml

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

examples/qwen2/reward-model.yaml

@@ -43,7 +43,8 @@ gradient_checkpointing_kwargs:
   use_reentrant: false
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 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
-flash_attention:
+attention: flash
 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
-flash_attention:
+attention: eager
 gptq_groupsize:
 gptq_model_v1:
 warmup_steps: 20

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

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

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

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

examples/starcoder2/qlora.yml

@@ -46,7 +46,8 @@ tf32: false
 gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
-flash_attention: true
+attention: flash
+
 
 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
-flash_attention: false
+attention: eager
 
 warmup_steps: 10
 evals_per_epoch: 0

examples/tiny-llama/lora.yml

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