add e2e check for lora w/o flash attention for mixtral to check gate

axolotl-ai-cloud · Jan 12, 2024 · ff2a221 · ff2a221
1 parent 05f4555
commit ff2a221
Showing 1 changed file with 43 additions and 1 deletion.
diff --git a/tests/e2e/test_mixtral.py b/tests/e2e/test_mixtral.py
@@ -27,7 +27,7 @@ class TestMixtral(unittest.TestCase):
     """
 
     @with_temp_dir
-    def test_qlora(self, temp_dir):
+    def test_qlora_w_fa2(self, temp_dir):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
@@ -68,6 +68,48 @@ def test_qlora(self, temp_dir):
         train(cfg=cfg, cli_args=cli_args, dataset_meta=dataset_meta)
         assert (Path(temp_dir) / "adapter_model.bin").exists()
 
+    @with_temp_dir
+    def test_lora_wo_fa2(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "tokenizer_config": "mistralai/Mixtral-8x7B-v0.1",
+                "flash_attention": False,
+                "sequence_len": 1024,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "lora_r": 4,
+                "lora_alpha": 8,
+                "lora_dropout": 0.1,
+                "lora_target_linear": True,
+                "val_set_size": 0.1,
+                "special_tokens": {},
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 2,
+                "micro_batch_size": 2,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_bnb_8bit",
+                "lr_scheduler": "cosine",
+                "max_steps": 20,
+                "save_steps": 10,
+                "eval_steps": 10,
+            }
+        )
+        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()
+
     @with_temp_dir
     def test_ft(self, temp_dir):
         # pylint: disable=duplicate-code