support activation offloading via unsloth gc

Former-commit-id: fb72a3adb0916232cc9ac9f0c725c02d07b9354c

scripts/cal_mfu.py

@@ -109,6 +109,7 @@ def calculate_mfu(
     deepspeed_stage: int = 0,
     disable_gc: bool = False,
     liger_kernel: bool = False,
+    unsloth_gc: bool = False,
 ) -> float:
     r"""
     Calculates MFU for given model and hyper-params.
@@ -119,6 +120,7 @@ def calculate_mfu(
         "flash_attn": flash_attn,
         "disable_gradient_checkpointing": disable_gc,
         "enable_liger_kernel": liger_kernel,
+        "use_unsloth_gc": unsloth_gc,
         "stage": "pt",
         "do_train": True,
         "finetuning_type": finetuning_type,

src/llamafactory/hparams/model_args.py

@@ -215,6 +215,10 @@ class ModelArguments(QuantizationArguments, ProcessorArguments, ExportArguments,
         default=False,
         metadata={"help": "Whether or not to use unsloth's optimization for the LoRA training."},
     )
+    use_unsloth_gc: bool = field(
+        default=False,
+        metadata={"help": "Whether or not to use unsloth's gradient checkpointing."},
+    )
     enable_liger_kernel: bool = field(
         default=False,
         metadata={"help": "Whether or not to enable liger kernel for faster training."},

src/llamafactory/model/model_utils/checkpointing.py

@@ -1,8 +1,10 @@
-# Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
+# Copyright 2024 HuggingFace Inc., Daniel Han-Chen & the Unsloth team and the LlamaFactory team.
 #
-# This code is inspired by the HuggingFace's Transformers and PEFT library.
+# This code is inspired by the HuggingFace's Transformers and PEFT library,
 # https://github.com/huggingface/transformers/blob/v4.40.0/src/transformers/modeling_utils.py
 # https://github.com/huggingface/peft/blob/v0.10.0/src/peft/utils/other.py
+# and the Unsloth library.
+# https://github.com/unslothai/unsloth/blob/July-2024/unsloth/models/_utils.py
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -19,7 +21,7 @@
 import inspect
 from functools import partial
 from types import MethodType
-from typing import TYPE_CHECKING, Any, Dict, Optional, Tuple
+from typing import TYPE_CHECKING, Any, Dict, Optional, Tuple, Union
 
 import torch
 
@@ -36,8 +38,45 @@ if TYPE_CHECKING:
 logger = get_logger(__name__)
 
 
+class UnslothGradientCheckpointing(torch.autograd.Function):
+    r"""
+    Saves VRAM by smartly offloading to RAM.
+    """
+
+    @staticmethod
+    @torch.cuda.amp.custom_fwd
+    def forward(
+        ctx: "torch.autograd.Function",
+        forward_function: "torch.Module",
+        hidden_states: "torch.Tensor",
+        *args: Union["torch.Tensor", Any],
+    ) -> "torch.Tensor":
+        saved_hidden_states = hidden_states.to("cpu", non_blocking=True)
+        with torch.no_grad():
+            output = forward_function(hidden_states, *args)
+
+        ctx.save_for_backward(saved_hidden_states)
+        ctx.forward_function = forward_function
+        ctx.args = args
+        return output
+
+    @staticmethod
+    @torch.cuda.amp.custom_bwd
+    def backward(ctx: "torch.autograd.Function", grad_output: "torch.Tensor") -> "torch.Tensor":
+        (hidden_states,) = ctx.saved_tensors
+        hidden_states = hidden_states.to("cuda", non_blocking=True).detach()
+        hidden_states.requires_grad_(True)
+        with torch.enable_grad():
+            (output,) = ctx.forward_function(hidden_states, *ctx.args)
+
+        torch.autograd.backward(output, grad_output)
+        return (None, hidden_states.grad) + (None,) * len(ctx.args)
+
+
 def _gradient_checkpointing_enable(
-    self: "PreTrainedModel", gradient_checkpointing_kwargs: Optional[Dict[str, Any]] = None
+    self: "PreTrainedModel",
+    gradient_checkpointing_kwargs: Optional[Dict[str, Any]] = None,
+    use_unsloth_gc: bool = False,
 ) -> None:
     r"""
     Activates gradient checkpointing for the current model.
@@ -52,9 +91,12 @@ def _gradient_checkpointing_enable(
     if gradient_checkpointing_kwargs is None:
         gradient_checkpointing_kwargs = {"use_reentrant": True}
 
-    gradient_checkpointing_func = partial(checkpoint, **gradient_checkpointing_kwargs)
+    if use_unsloth_gc:
+        gradient_checkpointing_func = UnslothGradientCheckpointing.apply
+    else:
+        gradient_checkpointing_func = partial(checkpoint, **gradient_checkpointing_kwargs)
 
-    def custom_gradient_checkpointing_func(func, *args, **kwargs):
+    def custom_gradient_checkpointing_func(func, *args: Union["torch.Tensor", Any], **kwargs):
         module: "torch.nn.Module" = func.__self__
 
         if any(param.requires_grad for param in module.parameters()):
@@ -97,7 +139,10 @@ def prepare_model_for_training(model: "PreTrainedModel", model_args: "ModelArgum
         else:
             # use_reentrant=False might increase VRAM usage (have not been empirically verified yet)
             # According to: https://github.com/huggingface/transformers/issues/28339
-            model.gradient_checkpointing_enable = MethodType(_gradient_checkpointing_enable, model)
+            gradient_checkpointing_enable = partial(
+                _gradient_checkpointing_enable, use_unsloth_gc=model_args.use_unsloth_gc
+            )
+            model.gradient_checkpointing_enable = MethodType(gradient_checkpointing_enable, model)
             model.gradient_checkpointing_enable(gradient_checkpointing_kwargs={"use_reentrant": True})
             setattr(model.config, "use_cache", False)  # turn off when gradient checkpointing is enabled
             logger.info("Gradient checkpointing enabled.")