fix optimizer reset for relora sft (#1414)

* fix optimizer reset

* set states to reset for 8bit optimizers and handle quantile runtime error for embeddings

* fix relora test to check grad_norm

* use flash attn for relora and tweak hyperparams for test

* fix messages field for test dataset

requirements-dev.txt

@@ -2,4 +2,3 @@ pre-commit
 black
 mypy
 types-requests
-tbparse

requirements-tests.txt

@@ -2,3 +2,4 @@ pytest
 pytest-xdist
 pytest-retry
 pytest-sugar
+tbparse

src/axolotl/monkeypatch/relora.py

@@ -46,9 +46,10 @@ def reset_optimizer(
     *,
     reset_params: List[str],  # where str is the key to a torch.nn.Parameter
     optimizer_state_keys: List[str],
-    prune_ratio: float = 0.9,
+    optimizer_magnitude_pruning: float = 0.9,
 ):
-    pruning_fn = partial(magnitude_pruning_, prune_ratio=prune_ratio)
+    # pylint:disable=unused-argument
+    pruning_fn = partial(magnitude_pruning_, prune_ratio=optimizer_magnitude_pruning)
     n_zeros = 0
     n_total = 0
 
@@ -56,16 +57,22 @@ def reset_optimizer(
     if isinstance(optimizer, ZeroRedundancyOptimizer):
         optimizer_state = optimizer.optim.state
 
-    for param in reset_params:
-        param_state = optimizer_state[param]
-        if len(param_state) == 0:  # no state for this param, happens for ZeRo optimizer
-            continue
-        for key in optimizer_state_keys:
-            pruning_fn(
-                param_state[key]
-            )  # pruning fn has to be inplace to keep the same keys in the dict
-            n_total += param_state[key].numel()
-            n_zeros += torch.sum(param_state[key] == 0).item()
+    for group in optimizer.param_groups:
+        for param in group["params"]:
+            state = optimizer_state[param]
+            for key, value in state.items():
+                if key not in optimizer_state_keys:
+                    continue
+                if torch.is_tensor(value):
+                    try:
+                        pruning_fn(value)
+                        n_total += value.numel()
+                        n_zeros += torch.sum(value == 0).item()
+                    except RuntimeError as exc:
+                        if "quantile() input tensor is too large" in str(exc):
+                            pass
+                        else:
+                            raise exc
 
     _zeroed = n_zeros / (1e-7 + n_total) * 100
     LOG.info(f"Percent of optimizer states zeroed: {_zeroed:.2f}")
@@ -129,6 +136,9 @@ class ReLoRACallback(TrainerCallback):
 
             if "adam" in args.optim.lower():
                 optimizer_state_keys = ["exp_avg", "exp_avg_sq"]
+                if "8bit" in args.optim.lower():
+                    optimizer_state_keys.append("state1")
+                    optimizer_state_keys.append("state2")
             else:
                 raise ValueError(f"Optimizer {args.optim} not supported with ReLoRA")
 
@@ -160,7 +170,7 @@ class ReLoRACallback(TrainerCallback):
                     optimizer,
                     reset_params=lora_params,
                     optimizer_state_keys=optimizer_state_keys,
-                    prune_ratio=args.relora_prune_ratio,
+                    optimizer_magnitude_pruning=args.relora_prune_ratio,
                 )
 
             if self.quantized:

tests/e2e/test_relora_llama.py

@@ -7,13 +7,15 @@ import os
 import unittest
 from pathlib import Path
 
+from tbparse import SummaryReader
+
 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
+from .utils import most_recent_subdir, with_temp_dir
 
 LOG = logging.getLogger("axolotl.tests.e2e")
 os.environ["WANDB_DISABLED"] = "true"
@@ -29,36 +31,48 @@ class TestReLoraLlama(unittest.TestCase):
         # pylint: disable=duplicate-code
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
-                "tokenizer_type": "LlamaTokenizer",
-                "sequence_len": 1024,
+                "base_model": "HuggingFaceTB/SmolLM2-135M",
+                "sequence_len": 2048,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
+                "flash_attention": True,
                 "load_in_8bit": True,
                 "adapter": "lora",
-                "lora_r": 32,
+                "lora_r": 8,
                 "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_steps": 100,
+                "relora_warmup_steps": 20,
+                "relora_anneal_steps": 10,
+                "relora_prune_ratio": 0.9,
                 "relora_cpu_offload": True,
                 "val_set_size": 0.0,
-                "special_tokens": {},
+                "special_tokens": {
+                    "pad_token": "<|endoftext|>",
+                },
+                "chat_template": "chatml",
                 "datasets": [
                     {
-                        "path": "mhenrichsen/alpaca_2k_test",
-                        "type": "alpaca",
+                        "path": "mlabonne/FineTome-100k",
+                        "type": "chat_template",
+                        "split": "train[:10%]",
+                        "field_messages": "conversations",
+                        "message_field_role": "from",
+                        "message_field_content": "value",
                     },
                 ],
-                "warmup_steps": 15,
+                "warmup_steps": 20,
                 "num_epochs": 2,
-                "max_steps": 51,  # at least 2x relora_steps
-                "micro_batch_size": 4,
+                "max_steps": 205,  # at least 2x relora_steps
+                "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
-                "optimizer": "adamw_torch",
+                "optimizer": "adamw_8bit",
                 "lr_scheduler": "cosine",
+                "save_safetensors": True,
+                "use_tensorboard": True,
             }
         )
         normalize_config(cfg)
@@ -66,4 +80,14 @@ class TestReLoraLlama(unittest.TestCase):
         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()
+        assert (
+            Path(temp_dir) / "checkpoint-100/adapter/adapter_model.safetensors"
+        ).exists()
+        assert (Path(temp_dir) / "checkpoint-100/relora/model.safetensors").exists()
+
+        tb_log_path = most_recent_subdir(temp_dir + "/runs")
+        event_file = os.path.join(tb_log_path, sorted(os.listdir(tb_log_path))[0])
+        reader = SummaryReader(event_file)
+        df = reader.scalars  # pylint: disable=invalid-name
+        df = df[(df.tag == "train/grad_norm")]  # pylint: disable=invalid-name
+        assert df.value.values[-1] < 0.2, "grad_norm is too high"