bump flash attention 2.5.8 -> 2.6.1 (#1738)

* bump flash attention 2.5.8 -> 2.6.1

* use triton implementation of cross entropy from flash attn

* add smoke test for flash attn cross entropy patch

* fix args to xentropy.apply

* handle tuple from triton loss fn

* ensure the patch tests run independently

* use the wrapper already built into flash attn for cross entropy

* mark pytest as forked for patches

* use pytest xdist instead of forked, since cuda doesn't like forking

* limit to 1 process and use dist loadfile for pytest

* change up pytest for fixture to reload transformers w monkeypathc

cicd/Dockerfile.jinja

@@ -30,7 +30,7 @@ RUN if [ "$AXOLOTL_EXTRAS" != "" ] ; then \
     fi
 
 # So we can test the Docker image
-RUN pip install pytest
+RUN pip install -r requirements-tests.txt
 
 # fix so that git fetch/pull from remote works
 RUN git config remote.origin.fetch "+refs/heads/*:refs/remotes/origin/*" && \

requirements-tests.txt

@@ -1 +1,2 @@
 pytest
+pytest-xdist

requirements.txt

@@ -12,7 +12,7 @@ fire
 PyYAML>=6.0
 requests
 datasets==2.19.1
-flash-attn==2.5.8
+flash-attn==2.6.1
 sentencepiece
 wandb
 einops

setup.py

@@ -80,10 +80,10 @@ setup(
     dependency_links=dependency_links,
     extras_require={
         "flash-attn": [
-            "flash-attn==2.5.8",
+            "flash-attn==2.6.1",
         ],
         "fused-dense-lib": [
-            "fused-dense-lib  @ git+https://github.com/Dao-AILab/flash-attention@v2.5.8#subdirectory=csrc/fused_dense_lib",
+            "fused-dense-lib  @ git+https://github.com/Dao-AILab/flash-attention@v2.6.1#subdirectory=csrc/fused_dense_lib",
         ],
         "deepspeed": [
             "deepspeed @ git+https://github.com/microsoft/DeepSpeed.git@bc48371c5e1fb8fd70fc79285e66201dbb65679b",

src/axolotl/monkeypatch/llama_attn_hijack_flash.py

@@ -104,17 +104,12 @@ def replace_llama_attn_with_flash_attn(
 
     # skip only if explicitly disabled
     if cross_entropy:
-        try:
-            from flash_attn.losses.cross_entropy import CrossEntropyLoss
+        from flash_attn.losses.cross_entropy import CrossEntropyLoss
 
-            LOG.info("patching with flash_attn.losses.cross_entropy")
-            transformers.models.llama.modeling_llama.CrossEntropyLoss = partial(
-                CrossEntropyLoss, inplace_backward=True
-            )
-        except ImportError:
-            LOG.warning(
-                "optimized flash-attention CrossEntropyLoss not found (run `pip install 'git+https://github.com/Dao-AILab/flash-attention.git#egg=xentropy_cuda_lib&subdirectory=csrc/xentropy'`)"
-            )
+        LOG.info("patching with flash_attn.losses.cross_entropy")
+        transformers.models.llama.modeling_llama.CrossEntropyLoss = partial(
+            CrossEntropyLoss, inplace_backward=True
+        )
 
     # skip only if explicitly disabled
     if rms_norm:

src/axolotl/utils/models.py

@@ -371,6 +371,12 @@ def load_model(
                     rms_norm=cfg.flash_attn_rms_norm,
                     use_shifted_sparse_attn=True,
                 )
+            elif cfg.flash_attn_cross_entropy or cfg.flash_attn_rms_norm:
+                replace_llama_attn_with_flash_attn(
+                    packed=False,
+                    cross_entropy=cfg.flash_attn_cross_entropy,
+                    rms_norm=cfg.flash_attn_rms_norm,
+                )
         elif cfg.xformers_attention:
             from axolotl.monkeypatch.llama_attn_hijack_xformers import (
                 hijack_llama_attention,

tests/e2e/patched/test_fa_xentropy.py

@@ -0,0 +1,87 @@
+"""
+E2E tests for lora llama
+"""
+
+import logging
+import os
+import unittest
+from importlib import reload
+from pathlib import Path
+
+import pytest
+from transformers.utils import is_torch_bf16_gpu_available
+
+from axolotl.cli import load_datasets
+from axolotl.common.cli import TrainerCliArgs
+from axolotl.train import train
+from axolotl.utils.config import normalize_config
+from axolotl.utils.dict import DictDefault
+
+from ..utils import with_temp_dir
+
+LOG = logging.getLogger("axolotl.tests.e2e")
+os.environ["WANDB_DISABLED"] = "true"
+
+
+@pytest.fixture(autouse=True)
+def reload_transformers():
+    import transformers.models.llama.modeling_llama
+
+    yield
+    reload(transformers.models.llama.modeling_llama)
+
+
+class TestFAXentropyLlama(unittest.TestCase):
+    """
+    Test case for Llama models using LoRA w multipack
+    """
+
+    @with_temp_dir
+    def test_lora_packing_fa_cross_entropy(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
+                "sequence_len": 1024,
+                "sample_packing": True,
+                "flash_attention": True,
+                "flash_attn_cross_entropy": True,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "lora_r": 32,
+                "lora_alpha": 64,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "val_set_size": 0.2,
+                "special_tokens": {
+                    "unk_token": "<unk>",
+                    "bos_token": "<s>",
+                    "eos_token": "</s>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 8,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+            }
+        )
+        if is_torch_bf16_gpu_available():
+            cfg.bf16 = True
+        else:
+            cfg.fp16 = True
+
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
+        assert (Path(temp_dir) / "adapter_model.bin").exists()