use DataCollatorWithFlattening when not sample packing (#2167)

tests/e2e/test_llama.py

@@ -104,3 +104,42 @@ class TestLlama:
 
         train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "model.safetensors").exists()
+
+    def test_batch_flattening(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "trust_remote_code": True,
+                "sequence_len": 512,
+                "val_set_size": 0.01,
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 5,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_8bit",
+                "lr_scheduler": "cosine",
+                "flash_attention": True,
+                "sample_packing": False,
+                "batch_flattening": True,
+                "bf16": True,
+                "save_safetensors": 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) / "model.safetensors").exists()

tests/patched/test_validation.py

@@ -1236,6 +1236,76 @@ class TestTorchCompileValidation(BaseValidation):
         assert updated_cfg.torch_compile is False
 
 
+class TestSampleOptimConfigValidation(BaseValidation):
+    """
+    test configurations for sample optimizations like batch flattening
+    """
+
+    def test_batch_flattening_auto_enables(self, minimal_cfg):
+        cfg = (
+            DictDefault(
+                {
+                    "flash_attention": True,
+                    "sample_packing": None,
+                    "micro_batch_size": 2,
+                    "batch_flattening": "auto",
+                }
+            )
+            | minimal_cfg
+        )
+
+        new_cfg = validate_config(cfg)
+        assert new_cfg["batch_flattening"] is True
+
+    def test_batch_flattening_auto_no_fa(self, minimal_cfg):
+        cfg = (
+            DictDefault(
+                {
+                    "flash_attention": False,
+                    "sample_packing": None,
+                    "micro_batch_size": 2,
+                    "batch_flattening": "auto",
+                }
+            )
+            | minimal_cfg
+        )
+
+        new_cfg = validate_config(cfg)
+        assert new_cfg["batch_flattening"] is False
+
+    def test_batch_flattening_auto_mbsz_1(self, minimal_cfg):
+        cfg = (
+            DictDefault(
+                {
+                    "flash_attention": True,
+                    "sample_packing": None,
+                    "micro_batch_size": 1,
+                    "batch_flattening": "auto",
+                }
+            )
+            | minimal_cfg
+        )
+
+        new_cfg = validate_config(cfg)
+        assert new_cfg["batch_flattening"] is False
+
+    def test_batch_flattening_auto_packing(self, minimal_cfg):
+        cfg = (
+            DictDefault(
+                {
+                    "flash_attention": True,
+                    "sample_packing": True,
+                    "micro_batch_size": 2,
+                    "batch_flattening": "auto",
+                }
+            )
+            | minimal_cfg
+        )
+
+        new_cfg = validate_config(cfg)
+        assert new_cfg["batch_flattening"] is False
+
+
 class TestValidationCheckModelConfig(BaseValidation):
     """
     Test the validation for the config when the model config is available