relora: magnitude pruning of the optimizer (#1245)

* magnitude pruning of the optimizer

* add alpaca chat template and fix relora patch

* fix handling of lora adapter for relora

* fix merge and save call

* fixes for 8-bit lora merge

* save intermediate checkpoint adapters

* auto merge

* fix eval check

* handle relora annealing

* fix anneal step logic

* chore: lint

* misx fix

* fix types

* Update tests/e2e/test_relora_llama.py

* check for safetensors saved from relora

tests/e2e/patched/test_mistral_samplepack.py

@@ -7,8 +7,6 @@ import os
 import unittest
 from pathlib import Path
 
-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
@@ -63,6 +61,7 @@ class TestMistral(unittest.TestCase):
                 "max_steps": 20,
                 "save_steps": 10,
                 "eval_steps": 10,
+                "bf16": "auto",
             }
         )
         normalize_config(cfg)
@@ -103,12 +102,9 @@ class TestMistral(unittest.TestCase):
                 "max_steps": 20,
                 "save_steps": 10,
                 "eval_steps": 10,
+                "bf16": "auto",
             }
         )
-        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)

tests/e2e/test_relora_llama.py

@@ -0,0 +1,68 @@
+"""
+E2E tests for relora llama
+"""
+
+import logging
+import os
+import unittest
+from pathlib import Path
+
+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"
+
+
+class TestReLoraLlama(unittest.TestCase):
+    """
+    Test case for Llama models using LoRA
+    """
+
+    @with_temp_dir
+    def test_relora(self, temp_dir):
+        # pylint: disable=duplicate-code
+        cfg = DictDefault(
+            {
+                "base_model": "JackFram/llama-68m",
+                "tokenizer_type": "LlamaTokenizer",
+                "sequence_len": 1024,
+                "load_in_8bit": True,
+                "adapter": "lora",
+                "lora_r": 32,
+                "lora_alpha": 16,
+                "lora_dropout": 0.05,
+                "lora_target_modules": ["q_proj", "v_proj"],
+                "relora_steps": 25,
+                "relora_warmup_steps": 5,
+                "relora_anneal_steps": 5,
+                "relora_cpu_offload": True,
+                "val_set_size": 0.0,
+                "special_tokens": {},
+                "datasets": [
+                    {
+                        "path": "mhenrichsen/alpaca_2k_test",
+                        "type": "alpaca",
+                    },
+                ],
+                "warmup_steps": 15,
+                "num_epochs": 2,
+                "micro_batch_size": 4,
+                "gradient_accumulation_steps": 1,
+                "output_dir": temp_dir,
+                "learning_rate": 0.00001,
+                "optimizer": "adamw_torch",
+                "lr_scheduler": "cosine",
+            }
+        )
+        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()