add ds zero3 to multigpu biweekly tests (#1900)

* add ds zero3 to multigpu biweekly tests

* fix for upstream api change

* use updated accelerate and fix deepspeed tests

* stringify the Path, and run multigpu tests if the multigpu tests change for a PR

* use correct json rather than yaml

* revert accelerate for deepspeed

requirements.txt

@@ -4,7 +4,7 @@ peft==0.13.2
 transformers==4.45.2
 tokenizers>=0.20.1
 bitsandbytes==0.44.1
-accelerate==1.0.0
+accelerate==0.34.2
 datasets==3.0.1
 deepspeed==0.14.4
 pydantic==2.6.3

tests/e2e/multigpu/test_llama.py

@@ -19,6 +19,8 @@ from ..utils import with_temp_dir
 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():
@@ -346,3 +348,115 @@ class TestMultiGPULlama(unittest.TestCase):
                 str(Path(temp_dir) / "config.yaml"),
             ]
         )
+
+    @with_temp_dir
+    def test_ds_zero3_packed(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "TinyLlama/TinyLlama_v1.1",
+                "tokenizer_type": "LlamaTokenizer",
+                "sample_packing": True,
+                "eval_sample_packing": False,
+                "pad_to_sequence_len": True,
+                "sequence_len": 2048,
+                "val_set_size": 0.05,
+                "special_tokens": {
+                    "unk_token": "<unk>",
+                    "bos_token": "<s>",
+                    "eos_token": "</s>",
+                },
+                "datasets": [
+                    {
+                        "path": "tatsu-lab/alpaca",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 100,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 4,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+                "flash_attention": True,
+                "deepspeed": str(AXOLOTL_ROOT / "deepspeed_configs/zero3_bf16.json"),
+            }
+        )
+
+        # 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(
+            [
+                "accelerate",
+                "launch",
+                "--num-processes",
+                "2",
+                "-m",
+                "axolotl.cli.train",
+                str(Path(temp_dir) / "config.yaml"),
+            ]
+        )
+
+    @with_temp_dir
+    def test_ds_zero3_qlora_packed(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "TinyLlama/TinyLlama_v1.1",
+                "tokenizer_type": "LlamaTokenizer",
+                "load_in_4bit": True,
+                "adapter": "qlora",
+                "lora_r": 8,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_linear": True,
+                "sample_packing": True,
+                "eval_sample_packing": False,
+                "pad_to_sequence_len": True,
+                "sequence_len": 2048,
+                "val_set_size": 0.05,
+                "special_tokens": {
+                    "unk_token": "<unk>",
+                    "bos_token": "<s>",
+                    "eos_token": "</s>",
+                },
+                "datasets": [
+                    {
+                        "path": "tatsu-lab/alpaca",
+                        "type": "alpaca",
+                    },
+                ],
+                "num_epochs": 1,
+                "max_steps": 100,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 4,
+                "output_dir": temp_dir,
+                "learning_rate": 0.0001,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+                "flash_attention": True,
+                "deepspeed": str(AXOLOTL_ROOT / "deepspeed_configs/zero3_bf16.json"),
+            }
+        )
+
+        # 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(
+            [
+                "accelerate",
+                "launch",
+                "--num-processes",
+                "2",
+                "-m",
+                "axolotl.cli.train",
+                str(Path(temp_dir) / "config.yaml"),
+            ]
+        )