Flex Attention + Packing with BlockMask support (#2363)

tests/e2e/multigpu/solo/test_flex.py

@@ -0,0 +1,92 @@
+"""
+E2E tests for multigpu lora tinyllama
+"""
+
+import logging
+import os
+from pathlib import Path
+
+import pytest
+import yaml
+from accelerate.test_utils import execute_subprocess_async
+from huggingface_hub import snapshot_download
+from transformers.testing_utils import get_torch_dist_unique_port
+from transformers.utils import is_torch_bf16_gpu_available
+
+from axolotl.utils.dict import DictDefault
+
+from tests.e2e.utils import check_tensorboard, require_torch_2_6_0
+
+LOG = logging.getLogger("axolotl.tests.e2e.multigpu")
+os.environ["WANDB_DISABLED"] = "true"
+
+AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_model():
+    # download the model
+    snapshot_download("HuggingFaceTB/SmolLM2-135M")
+
+
+class TestPackedFlex:
+    """
+    Test case for Packed training of llama models
+    """
+
+    @require_torch_2_6_0
+    def test_loss_llama(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "sequence_len": 1024,
+                "sample_packing": True,
+                "flex_attention": True,
+                "val_set_size": 0.0,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "datasets": [
+                    {
+                        "path": "vicgalle/alpaca-gpt4",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 4,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch_fused",
+                "lr_scheduler": "cosine",
+                "max_steps": 5,
+                "use_tensorboard": True,
+                "save_strategy": "no",
+            }
+        )
+        if is_torch_bf16_gpu_available():
+            cfg.bf16 = True
+        else:
+            cfg.fp16 = True
+
+        # write cfg to yaml file
+        Path(temp_dir).mkdir(parents=True, exist_ok=True)
+        with open(Path(temp_dir) / "config.yaml", "w", encoding="utf-8") as fout:
+            fout.write(yaml.dump(cfg.to_dict(), Dumper=yaml.Dumper))
+
+        execute_subprocess_async(
+            [
+                "axolotl",
+                "train",
+                str(Path(temp_dir) / "config.yaml"),
+                "--num-processes",
+                "2",
+                "--main-process-port",
+                f"{get_torch_dist_unique_port()}",
+            ]
+        )
+
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
+        )

tests/e2e/solo/test_flex.py

@@ -0,0 +1,73 @@
+"""
+E2E tests for packed training w/ flex attention
+"""
+
+import logging
+import os
+import unittest
+
+from transformers.utils import is_torch_bf16_gpu_available
+
+from axolotl.cli.args import TrainerCliArgs
+from axolotl.common.datasets import load_datasets
+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, require_torch_2_6_0, with_temp_dir
+
+LOG = logging.getLogger("axolotl.tests.e2e")
+os.environ["WANDB_DISABLED"] = "true"
+
+
+class TestPackedFlex(unittest.TestCase):
+    """
+    Test case for Packed training of llama models
+    """
+
+    @require_torch_2_6_0
+    @with_temp_dir
+    def test_loss_llama(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "sequence_len": 1024,
+                "sample_packing": True,
+                "flex_attention": True,
+                "val_set_size": 0.0,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "datasets": [
+                    {
+                        "path": "vicgalle/alpaca-gpt4",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 4,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch_fused",
+                "lr_scheduler": "cosine",
+                "max_steps": 5,
+                "use_tensorboard": True,
+            }
+        )
+        if is_torch_bf16_gpu_available():
+            cfg.bf16 = True
+        else:
+            cfg.fp16 = True
+
+        cfg = validate_config(cfg)
+        normalize_config(cfg)
+        cli_args = TrainerCliArgs()
+        dataset_meta = load_datasets(cfg=cfg, cli_args=cli_args)
+
+        train(cfg=cfg, dataset_meta=dataset_meta)
+
+        check_tensorboard(
+            temp_dir + "/runs", "train/train_loss", 2.0, "Train Loss is too high"
+        )

tests/e2e/utils.py

@@ -67,9 +67,21 @@ def require_torch_2_5_1(test_case):
     return unittest.skipUnless(is_min_2_5_1(), "test requires torch>=2.5.1")(test_case)
 
 
+def require_torch_2_6_0(test_case):
+    """
+    Decorator marking a test that requires torch >= 2.6.0
+    """
+
+    def is_min_2_6_0():
+        torch_version = version.parse(torch.__version__)
+        return torch_version >= version.parse("2.6.0")
+
+    return unittest.skipUnless(is_min_2_6_0(), "test requires torch>=2.6.0")(test_case)
+
+
 def require_torch_lt_2_6_0(test_case):
     """
-    Decorator marking a test that requires torch >= 2.5.1
+    Decorator marking a test that requires torch < 2.6.0
     """
 
     def is_max_2_6_0():