ReLoRA implementation (with quantization) (#322)

* Experimental ReLoRA (+qlora) implementation

* Add CPU offload

* Remove local config

* Fix saving logic

* Remove redundant assert

* Fix logic errors

* Move ReLoRA into its own trainer class with a method override to create the proper scheduler

* Formatting & typing fixes

* Use safe_serialization

* Don't allow fsdp/deepspeed with ReLoRA

* Fix cpu-offload logic, enable multi gpu

* Document parameters and add comment

* Fix merge issue

* Smooth over some sharp edges

* Implement resume from checkpoint for relora

* Address review comments

* Fix saving logic

* Add necessary metadata to safetensors

---------

Co-authored-by: Wing Lian <wing.lian@gmail.com>

README.md

@@ -493,6 +493,12 @@ lora_modules_to_save:
 lora_out_dir:
 lora_fan_in_fan_out: false
 
+# ReLoRA configuration
+# must use either 'lora' or 'qlora' adapter, and does not support fsdp or deepspeed
+relora_steps: # number of steps per ReLoRA restart
+relora_warmup_steps: # number of per-restart warmup steps
+relora_cpu_offload: # true to perform lora weight merges on cpu during restarts, for modest gpu memory savings
+
 # wandb configuration if you're using it
 wandb_mode: # "offline" to save run metadata locally and not sync to the server, "disabled" to turn off wandb
 wandb_project: # your wandb project name

scripts/finetune.py

@@ -242,6 +242,21 @@ def train(
         model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
         return
 
+    if cfg.resume_from_checkpoint is None and cfg.auto_resume_from_checkpoints:
+        possible_checkpoints = [
+            str(cp) for cp in Path(cfg.output_dir).glob("checkpoint-*")
+        ]
+        if len(possible_checkpoints) > 0:
+            sorted_paths = sorted(
+                possible_checkpoints,
+                key=lambda path: int(path.split("-")[-1]),
+            )
+            cfg.resume_from_checkpoint = sorted_paths[-1]
+            LOG.info(
+                f"Using Auto-resume functionality to start with checkpoint at {cfg.resume_from_checkpoint}"
+            )
+    resume_from_checkpoint = cfg.resume_from_checkpoint
+
     trainer = setup_trainer(
         cfg, train_dataset, eval_dataset, model, tokenizer, total_num_steps
     )
@@ -273,20 +288,6 @@ def train(
     LOG.info("Starting trainer...")
     if cfg.group_by_length:
         LOG.info("hang tight... sorting dataset for group_by_length")
-    resume_from_checkpoint = cfg.resume_from_checkpoint
-    if cfg.resume_from_checkpoint is None and cfg.auto_resume_from_checkpoints:
-        possible_checkpoints = [
-            str(cp) for cp in Path(cfg.output_dir).glob("checkpoint-*")
-        ]
-        if len(possible_checkpoints) > 0:
-            sorted_paths = sorted(
-                possible_checkpoints,
-                key=lambda path: int(path.split("-")[-1]),
-            )
-            resume_from_checkpoint = sorted_paths[-1]
-            LOG.info(
-                f"Using Auto-resume functionality to start with checkpoint at {resume_from_checkpoint}"
-            )
 
     if not Path(cfg.output_dir).is_dir():
         os.makedirs(cfg.output_dir, exist_ok=True)
@@ -301,6 +302,13 @@ def train(
 
     LOG.info(f"Training Completed!!! Saving pre-trained model to {cfg.output_dir}")
 
+    if cfg.relora_steps:
+        if cfg.adapter == "lora" and not (cfg.load_in_4bit or cfg.load_in_8bit):
+            model = model.merge_and_unload()
+        else:
+            # final model weights have already been saved by `ReLoRACallback.on_train_end`
+            return
+
     # TODO do we need this fix? https://huggingface.co/docs/accelerate/usage_guides/fsdp#saving-and-loading
     # only save on rank 0, otherwise it corrupts output on multi-GPU when multiple processes attempt to write the same file
     if cfg.fsdp:
@@ -308,6 +316,7 @@ def train(
     elif cfg.local_rank == 0:
         if cfg.flash_optimum:
             model = BetterTransformer.reverse(model)
+
         model.save_pretrained(cfg.output_dir, safe_serialization=safe_serialization)
 
 

src/axolotl/monkeypatch/relora.py

@@ -0,0 +1,393 @@
+"""Implements the ReLoRA training procedure from https://arxiv.org/abs/2307.05695, minus the initial full fine-tune."""
+import glob
+import json
+import logging
+import os.path
+import shutil
+from pathlib import Path
+from typing import Dict, List, Sequence
+
+import bitsandbytes as bnb
+import peft
+import safetensors.torch as st
+import torch
+from huggingface_hub import snapshot_download
+from torch.optim.lr_scheduler import LRScheduler
+from torch.optim.optimizer import Optimizer
+from transformers import (
+    TrainerCallback,
+    TrainerControl,
+    TrainerState,
+    TrainingArguments,
+)
+from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
+
+from axolotl.utils.dict import DictDefault
+from axolotl.utils.distributed import is_main_process
+
+LOG = logging.getLogger("axolotl.relora")
+
+
+def reset_optimizer(optimizer: torch.optim.Optimizer):
+    for group in optimizer.param_groups:
+        for param in group["params"]:
+            param_state = optimizer.state[param]
+            for key in param_state:
+                if "qmap" in key:
+                    continue
+
+                if key == "step" and isinstance(param_state[key], int):
+                    param_state[key] = 0
+                else:
+                    param_state[key] = torch.zeros_like(param_state[key])
+
+
+class ReLoRACallback(TrainerCallback):
+    """Callback to merge LoRA weights into the base model and save full-weight checkpoints"""
+
+    def __init__(self, cfg: DictDefault):
+        self.relora_steps = cfg.relora_steps
+        self.cpu_offload = cfg.relora_cpu_offload
+        self.quantized = cfg.load_in_4bit or cfg.load_in_8bit
+        self.last_full_model = cfg.base_model
+        self.resume_from_checkpoint = cfg.resume_from_checkpoint
+
+        if not os.path.exists(self.last_full_model):
+            self.last_full_model = str(Path(snapshot_download(cfg.base_model)))
+
+        assert os.path.exists(
+            self.last_full_model
+        ), "for ReLORA base_model must be a local path"
+
+        self.num_lora_restarts = 0
+        self.need_full_save = False
+
+    def on_train_begin(
+        self,
+        _args: TrainingArguments,
+        _state: TrainerState,
+        control: TrainerControl,
+        model: peft.LoraModel,
+        **_kwargs,
+    ):
+        if self.resume_from_checkpoint:
+            weight_path = os.path.join(self.resume_from_checkpoint, "relora")
+            if not os.path.exists(weight_path):
+                LOG.warning(
+                    "Resuming ReLoRA from checkpoint, but no full-weight save found"
+                )
+            else:
+                LOG.info(f"Loading adjusted base weights from {weight_path}")
+                load_weight_checkpoint(model, weight_path)
+        return control
+
+    def on_step_begin(
+        self,
+        args: TrainingArguments,
+        state: TrainerState,
+        control: TrainerControl,
+        model: peft.LoraModel,
+        optimizer: torch.optim.Optimizer,
+        **_kwargs,
+    ):
+        if state.global_step > 0 and state.global_step % self.relora_steps == 0:
+            checkpoint_folder = os.path.join(
+                args.output_dir,
+                f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}",
+                "relora",
+            )
+
+            with torch.no_grad():
+                merge_and_save(
+                    model,
+                    self.last_full_model,
+                    checkpoint_folder,
+                    reinit=True,
+                    quantized=self.quantized,
+                    actually_save=is_main_process(),
+                    cpu_offload=self.cpu_offload,
+                )
+                reset_optimizer(optimizer)
+
+            if self.quantized:
+                self.last_full_model = checkpoint_folder
+            self.num_lora_restarts += 1
+
+        return control
+
+    def on_save(
+        self,
+        args: TrainingArguments,
+        state: TrainerState,
+        control: TrainerControl,
+        model: peft.LoraModel,
+        **_kwargs,
+    ):
+        checkpoint_folder = os.path.join(
+            args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}", "relora"
+        )
+        if (
+            state.global_step >= self.relora_steps
+            and state.global_step % self.relora_steps != 0
+        ):
+            if self.quantized:
+                if self.last_full_model != checkpoint_folder:
+                    # ensure the latest full parameter save is in the latest checkpoint
+                    # folder, so that automatic pruning of checkpoints does not remove it
+                    LOG.info(f"moving last full parameter save to {checkpoint_folder}")
+                    os.makedirs(checkpoint_folder, exist_ok=True)
+                    chunks = glob.glob(
+                        f"{self.last_full_model}/model*.safetensors"
+                    ) + glob.glob(f"{self.last_full_model}/model*.index.json")
+                    for path in chunks:
+                        new_path = os.path.abspath(shutil.move(path, checkpoint_folder))
+                        try:
+                            os.symlink(new_path, path)
+                        except OSError:
+                            # probably on windows without permission to symlink
+                            pass
+
+                    self.last_full_model = checkpoint_folder
+            else:
+                model.model.save_pretrained(checkpoint_folder, safe_serialization=True)
+
+        return control
+
+    def on_log(
+        self,
+        _args: TrainingArguments,
+        _state: TrainerState,
+        control: TrainerControl,
+        logs: Dict[str, float],
+        **_kwargs,
+    ):
+        logs["num_lora_restarts"] = self.num_lora_restarts
+        return control
+
+    def on_train_end(
+        self,
+        args: TrainingArguments,
+        _state: TrainerState,
+        control: TrainerControl,
+        model: peft.LoraModel,
+        **_kwargs,
+    ):
+        if self.quantized:
+            # perform final merge and save
+            with torch.no_grad():
+                merge_and_save(
+                    model,
+                    self.last_full_model,
+                    args.output_dir,
+                    reinit=False,
+                    quantized=self.quantized,
+                    actually_save=is_main_process(),
+                    cpu_offload=self.cpu_offload,
+                )
+        # no need to save if unquantized, as finetune.py will call merge_and_unload()
+        return control
+
+
+class ReLoRAScheduler(LRScheduler):
+    """Wraps another scheduler to apply per-lora-restart learning rate warmups."""
+
+    def __init__(
+        self,
+        optimizer: Optimizer,
+        inner_schedule: LRScheduler,
+        relora_steps: int,
+        warmup_steps: int,
+        min_lr_scale: float = 0.001,
+    ) -> None:
+        self.inner_schedule = inner_schedule
+        self.relora_steps = relora_steps
+        self.warmup_steps = warmup_steps
+        self.min_lr_scale = min_lr_scale
+        super().__init__(optimizer, inner_schedule.last_epoch, inner_schedule.verbose)
+
+    def get_lr(self) -> float:
+        self.inner_schedule.last_epoch = self.last_epoch
+
+        original = self.inner_schedule.get_lr()
+        step = self.last_epoch
+        if step < self.relora_steps:
+            scale = 1
+        else:
+            cycle_t = min(1.0, (step % self.relora_steps) / self.warmup_steps)
+            scale = cycle_t * (1 - self.min_lr_scale) + self.min_lr_scale
+
+        if isinstance(original, Sequence):
+            return [lr * scale for lr in original]
+        return original * scale
+
+
+def sharded_paths(path: str, module_names: List[str]) -> Dict[str, str]:
+    model_name = "model.safetensors"
+    if not os.path.exists(str(Path(path) / model_name)) and not os.path.exists(
+        str(Path(path) / f"{model_name}.index.json")
+    ):
+        model_name = "pytorch_model.bin"
+
+    index_path = str(Path(path) / f"{model_name}.index.json")
+    if os.path.exists(index_path):
+        with open(index_path, "r", encoding="utf-8") as file:
+            data = json.load(file)
+        return data["weight_map"]
+    return {(module_name + ".weight"): model_name for module_name in module_names}
+
+
+def lora_delta_weight(layer: peft.tuners.lora.LoraLayer, device) -> torch.Tensor:
+    if isinstance(layer, (peft.tuners.lora.Linear8bitLt, peft.tuners.lora.Linear4bit)):
+        adapter = layer.active_adapter
+        return (
+            peft.utils.transpose(
+                layer.lora_B[adapter].weight.detach().to(device)
+                @ layer.lora_A[adapter].weight.detach().to(device),
+                getattr(layer, "fan_in_fan_out", False),
+            )
+            * layer.scaling[adapter]
+        )
+
+    return layer.get_delta_weight().to(device)
+
+
+def find_lora_modules(model: peft.LoraModel) -> Dict[str, peft.tuners.lora.LoraLayer]:
+    modules: Dict[str, peft.tuners.lora.LoraLayer] = {}
+
+    key_list = [key for key, _ in model.model.named_modules() if "lora" not in key]
+    for key in key_list:
+        try:
+            # pylint: disable=protected-access
+            _parent, target, _target_name = peft.utils._get_submodules(model.model, key)
+        except AttributeError:
+            continue
+
+        if isinstance(target, peft.tuners.lora.LoraLayer):
+            modules[key] = target
+
+    return modules
+
+
+def update_weights(
+    target: peft.tuners.lora.LoraLayer, new_weight: torch.Tensor, reinit: bool, device
+):
+    if reinit:
+        for adapter_name in target.lora_A:
+            target.reset_lora_parameters(adapter_name)
+        for adapter_name in target.lora_embedding_A:
+            target.reset_lora_parameters(adapter_name)
+
+    if isinstance(target, peft.tuners.lora.Linear4bit):
+        # This could be faster, but the quantization of Linear4bit weights occurs
+        # when the module is moved from cpu to gpu. Without meddling *too* deeply in
+        # PEFT's innards or maintaining a duplicate of that codepath, this is good
+        # enough for now.
+        target.weight.quant_state = None
+        target.weight.data = new_weight.cpu()
+        target.to(device)
+    elif isinstance(target, peft.tuners.lora.Linear8bitLt):
+        target.weight = bnb.nn.Int8Params(new_weight, requires_grad=False).to(device)
+    else:
+        target.weight.data = new_weight.to(device)
+
+
+def merge_and_save(
+    model: peft.LoraModel,
+    model_src: str,
+    model_dst: str,
+    reinit: bool = False,
+    quantized: bool = False,
+    cpu_offload: bool = False,
+    actually_save: bool = True,
+):
+    modules = find_lora_modules(model)
+
+    if not quantized:
+        for module_name, target in modules.items():
+            update = target.get_delta_weight(target.active_adapter).detach()
+            target.weight.data += update
+
+            if reinit:
+                for adapter_name in target.lora_A:
+                    target.reset_lora_parameters(adapter_name)
+                for adapter_name in target.lora_embedding_A:
+                    target.reset_lora_parameters(adapter_name)
+        return
+
+    os.makedirs(model_dst, exist_ok=True)
+    shard_paths = sharded_paths(model_src, modules.keys())
+    out_shard_paths = {}
+
+    unique_shards = list(set(shard_paths.values()))
+    for shard_path in unique_shards:
+        out_tensors = {}
+        if shard_path.endswith(".safetensors"):
+            in_tensors = st.load_file(str(Path(model_src) / shard_path))
+        else:
+            in_tensors = torch.load(Path(model_src) / shard_path)
+            if "state_dict" in in_tensors:
+                in_tensors = in_tensors["state_dict"]
+
+        for module_name, target in modules.items():
+            key = module_name + ".weight"
+            if key not in shard_paths or shard_paths[key] != shard_path:
+                continue
+
+            orig_weight = in_tensors[key]
+            old_dev = target.weight.device
+            math_dev = "cpu" if cpu_offload else old_dev
+
+            delta_weight = lora_delta_weight(target, math_dev)
+            new_weight = orig_weight.to(math_dev) + delta_weight
+            del delta_weight
+
+            if actually_save:
+                out_tensors[key] = new_weight.half().cpu()
+
+            update_weights(target, new_weight, reinit=reinit, device=old_dev)
+
+        if actually_save:
+            out_shard_name = shard_path
+            if out_shard_name.startswith("pytorch_model"):
+                out_shard_name = (
+                    out_shard_name.replace("pytorch_model", "model").rstrip(".bin")
+                    + ".safetensors"
+                )
+
+            for module_name in in_tensors:
+                if module_name not in out_tensors:
+                    out_tensors[module_name] = in_tensors[module_name].half()
+                out_shard_paths[module_name] = out_shard_name
+
+            shard_fn = str(Path(model_dst) / out_shard_name)
+            LOG.info(f"saving tensors to {shard_fn}")
+            st.save_file(out_tensors, shard_fn, metadata={"format": "pt"})
+
+        del in_tensors
+        del out_tensors
+        torch.cuda.empty_cache()
+
+    if actually_save and len(unique_shards) > 1:
+        with open(
+            str(Path(model_dst, "model.safetensors.index.json")), "w", encoding="utf-8"
+        ) as file:
+            json.dump({"metadata": {}, "weight_map": out_shard_paths}, file)
+
+
+def load_weight_checkpoint(model: peft.LoraModel, checkpoint_path: str):
+    modules = find_lora_modules(model)
+    shard_paths = sharded_paths(checkpoint_path, modules.keys())
+    unique_shards = list(set(shard_paths.values()))
+
+    for shard_path in unique_shards:
+        tensors = st.load_file(os.path.join(checkpoint_path, shard_path))
+
+        for module_name, target in modules.items():
+            key = module_name + ".weight"
+            if key not in shard_paths or shard_paths[key] != shard_path:
+                continue
+
+            new_weight = tensors[key]
+            update_weights(
+                target, new_weight, reinit=False, device=target.weight.device
+            )

src/axolotl/utils/callbacks.py

@@ -33,7 +33,9 @@ class SavePeftModelCallback(TrainerCallback):  # pylint: disable=too-few-public-
         )
 
         peft_model_path = os.path.join(checkpoint_folder, "adapter_model")
-        kwargs["model"].save_pretrained(peft_model_path)
+        kwargs["model"].save_pretrained(
+            peft_model_path, save_safetensors=args.save_safetensors
+        )
 
         return control
 

src/axolotl/utils/config.py

@@ -126,6 +126,19 @@ def validate_config(cfg):
     if not cfg.load_in_8bit and cfg.adapter == "lora":
         LOG.warning("We recommend setting `load_in_8bit: true` for LORA finetuning")
 
+    if cfg.relora_steps:
+        if cfg.adapter not in ("lora", "qlora"):
+            raise ValueError("cfg.adapter must be lora or qlora to use ReLoRA")
+
+        if cfg.fsdp:
+            raise ValueError("fsdp not supported with ReLoRA")
+
+        if cfg.deepspeed:
+            raise ValueError("deepspeed not supported with ReLoRA")
+
+        if cfg.lr_scheduler == "one_cycle":
+            raise ValueError("ReLoRA is not compatible with the one_cycle scheduler")
+
     if cfg.trust_remote_code:
         LOG.warning(
             "`trust_remote_code` is set to true. Please make sure that you reviewed the remote code/model."

src/axolotl/utils/trainer.py

@@ -24,6 +24,7 @@ from transformers.trainer_pt_utils import (
     get_parameter_names,
 )
 
+from axolotl.monkeypatch.relora import ReLoRACallback, ReLoRAScheduler
 from axolotl.utils.callbacks import (
     GPUStatsCallback,
     SaveBetterTransformerModelCallback,
@@ -127,6 +128,14 @@ class AxolotlTrainingArguments(TrainingArguments):
         default=1,
         metadata={"help": "the multiplier for the max len for packed sequences"},
     )
+    relora_steps: Optional[int] = field(
+        default=None,
+        metadata={"help": "how often to reset for ReLoRA"},
+    )
+    relora_warmup_steps: Optional[int] = field(
+        default=None,
+        metadata={"help": "how many warmup steps to take after reset for ReLoRA"},
+    )
 
 
 class AxolotlTrainer(Trainer):
@@ -265,6 +274,39 @@ class OneCycleLRSchedulerTrainer(AxolotlTrainer):
         return self.lr_scheduler
 
 
+class ReLoRATrainer(AxolotlTrainer):
+    """
+    Trainer subclass that uses the OneCycleLR scheduler
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self.lr_scheduler = None
+
+    def create_scheduler(
+        self,
+        num_training_steps: int,
+        optimizer: Optional[torch.optim.Optimizer] = None,
+    ):
+        optimizer = self.optimizer if optimizer is None else optimizer
+        lr_scheduler = super().create_scheduler(num_training_steps, optimizer)
+
+        if self.args.relora_steps:
+            warmup_steps = (
+                self.args.relora_warmup_steps if self.args.relora_warmup_steps else 10
+            )
+            self.lr_scheduler = ReLoRAScheduler(
+                optimizer,
+                lr_scheduler,
+                self.args.relora_steps,
+                warmup_steps,
+            )
+        else:
+            self.lr_scheduler = lr_scheduler
+
+        return self.lr_scheduler
+
+
 def add_position_ids(sample):
     sample["position_ids"] = torch.arange(len(sample["input_ids"]))
     return sample
@@ -517,6 +559,8 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
         weight_decay=cfg.weight_decay if cfg.weight_decay is not None else 0.0,
         sample_packing=cfg.sample_packing if cfg.sample_packing else False,
         sample_packing_seq_len_multiplier=cfg.micro_batch_size,
+        relora_steps=cfg.relora_steps,
+        relora_warmup_steps=cfg.relora_warmup_steps,
         **training_arguments_kwargs,
     )
 
@@ -589,6 +633,10 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
 
     callbacks = []
     callbacks.append(GPUStatsCallback(cfg))
+
+    if cfg.relora_steps:
+        callbacks.append(ReLoRACallback(cfg))
+
     # TODO on_save callback to sync checkpoints to GCP/AWS in background
     if cfg.early_stopping_patience:
         early_stop_cb = EarlyStoppingCallback(
@@ -633,11 +681,11 @@ def setup_trainer(cfg, train_dataset, eval_dataset, model, tokenizer, total_num_
                 num_proc=32,
             )
 
-    trainer_cls = (
-        OneCycleLRSchedulerTrainer
-        if cfg.lr_scheduler == "one_cycle" and (cfg.fsdp or cfg.adapter == "qlora")
-        else AxolotlTrainer
-    )
+    trainer_cls = AxolotlTrainer
+    if cfg.lr_scheduler == "one_cycle" and (cfg.fsdp or cfg.adapter == "qlora"):
+        trainer_cls = OneCycleLRSchedulerTrainer
+    elif cfg.relora_steps:
+        trainer_cls = ReLoRATrainer
     trainer = trainer_cls(
         model=model,
         train_dataset=train_dataset,