ci for fa3

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

@@ -32,21 +32,25 @@ jobs:
             pytorch: 2.6.0
             axolotl_extras: vllm
             num_gpus: 2
-            nightly_build: "true"
+          - cuda: 126
+            cuda_version: 12.6.3
+            python_version: "3.11"
+            pytorch: 2.6.0
+            axolotl_extras:
+            suffix: "-hopper"
+            num_gpus: 2
           - cuda: 124
             cuda_version: 12.4.1
             python_version: "3.11"
             pytorch: 2.5.1
             axolotl_extras:
             num_gpus: 2
-            nightly_build: "true"
           - cuda: 126
             cuda_version: 12.6.3
             python_version: "3.11"
             pytorch: 2.7.0
             axolotl_extras:
             num_gpus: 2
-            nightly_build: "true"
     runs-on: [self-hosted, modal]
     timeout-minutes: 120
     steps:
@@ -68,7 +72,6 @@ jobs:
           echo "AXOLOTL_EXTRAS=${{ matrix.axolotl_extras}}" >> $GITHUB_ENV
           echo "CUDA=${{ matrix.cuda }}" >> $GITHUB_ENV
           echo "N_GPUS=${{ matrix.num_gpus }}" >> $GITHUB_ENV
-          echo "NIGHTLY_BUILD=${{ matrix.nightly_build }}" >> $GITHUB_ENV
           echo "CODECOV_TOKEN=${{ secrets.CODECOV_TOKEN }}" >> $GITHUB_ENV
       - name: Run tests job on Modal
         run: |

cicd/Dockerfile.jinja

@@ -32,6 +32,11 @@ RUN if [ "$NIGHTLY_BUILD" = "true" ] ; then \
     fi
 
 RUN pip install packaging==23.2 setuptools==75.8.0
+RUN if [ "$PYTORCH_VERSION" = "2.6.0" ] && [ "$CUDA" = "126" ] ; then \
+        curl -L -O https://d1dttdx32dkk5p.cloudfront.net/fa3/cu${CUDA}/torch-${PYTORCH_VERSION}/flash_attn_3-3.0.0b1-cp311-cp311-linux_x86_64.whl; \
+        pip3 install --no-cache-dir flash_attn_3-3.0.0b1-cp311-cp311-linux_x86_64.whl; \
+        rm flash_attn_3-3.0.0b1-cp311-cp311-linux_x86_64.whl; \
+    fi
 RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
         pip install --no-build-isolation -e .[deepspeed,flash-attn,ring-flash-attn,optimizers,ray,$AXOLOTL_EXTRAS] $AXOLOTL_ARGS; \
     else \

src/axolotl/utils/models.py

@@ -629,6 +629,27 @@ class ModelLoader:
             )
 
         if self.cfg.flash_attention:
+            use_fa3 = False
+            if self.cfg.use_flash_attention_3 is True:
+                use_fa3 = True
+            elif self.cfg.use_flash_attention_3 == "auto":
+                if int(self.cfg.capabilities.compute_capability.split("_")[-1]) >= 90:
+                    # FA3 is only available on Hopper GPUs and newer
+                    use_fa3 = True
+            if use_fa3 and importlib.util.find_spec("flash_attn_interface") is not None:
+                from flash_attn_interface import flash_attn_func as flash_attn_func_v3
+                from flash_attn_interface import (
+                    flash_attn_varlen_func as flash_attn_varlen_func_v3,
+                )
+
+                transformers.modeling_flash_attention_utils.flash_attn_func = (
+                    flash_attn_func_v3
+                )
+                transformers.modeling_flash_attention_utils.flash_attn_varlen_func = (
+                    flash_attn_varlen_func_v3
+                )
+                LOG.info("Switched to Flash Attention v3")
+
             self.patch_attention()
 
         if self.cfg.sample_packing and self.cfg.s2_attention:
@@ -699,6 +720,7 @@ class ModelLoader:
 
                 patch_mllama()
 
+            # TODO deprecate soon
             if self.model_config.model_type == "btlm":
                 from axolotl.monkeypatch.btlm_attn_hijack_flash import (
                     replace_btlm_attn_with_flash_attn,
@@ -706,6 +728,7 @@ class ModelLoader:
 
                 replace_btlm_attn_with_flash_attn(self.cfg.base_model)
 
+            # TODO deprecate soon
             if (
                 self.model_config.model_type == "stablelm_epoch"
                 and self.cfg.sample_packing

src/axolotl/utils/schemas/config.py

@@ -233,6 +233,7 @@ class AxolotlInputConfig(
     flash_attn_fuse_qkv: bool | None = None
     flash_attn_fuse_mlp: bool | None = None
     flash_optimum: bool | None = None
+    use_flash_attention_3: Literal["auto"] | bool | None = "auto"
 
     eager_attention: bool | None = None
 

tests/conftest.py

@@ -458,6 +458,7 @@ def cleanup_monkeypatches():
         ("transformers.trainer",),
         ("transformers", ["Trainer"]),
         ("transformers.loss.loss_utils",),
+        ("transformers.modeling_flash_attention_utils",),
     ]
     for module_name_tuple in modules_to_reset:
         module_name = module_name_tuple[0]

tests/e2e/test_packing_loss.py

@@ -4,8 +4,8 @@ E2E tests for packed training
 
 import logging
 import os
-import unittest
 
+import pytest
 from transformers.utils import is_torch_bf16_gpu_available
 
 from axolotl.cli.args import TrainerCliArgs
@@ -14,19 +14,22 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from .utils import check_tensorboard, with_temp_dir
+from .utils import check_tensorboard
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
 
 
-class TestPackedLlama(unittest.TestCase):
+class TestPackedLlama:
     """
     Test case for Packed training of llama models
     """
 
-    @with_temp_dir
-    def test_loss_packed(self, temp_dir):
+    @pytest.mark.parametrize(
+        "use_flash_attention_3",
+        [False, "auto"],
+    )
+    def test_loss_packed(self, temp_dir, use_flash_attention_3):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
@@ -54,6 +57,7 @@ class TestPackedLlama(unittest.TestCase):
                 "lr_scheduler": "cosine",
                 "max_steps": 5,
                 "use_tensorboard": True,
+                "use_flash_attention_3": use_flash_attention_3,
             }
         )
         if is_torch_bf16_gpu_available():