Activation checkpointing with offloading to disk with prefetch (axolotl-ai-cloud#2663)

* offload activations to disk instead of CPU RAM

* add prefetch

* Disco :dance:

* include offload_disk in e2e test for AC

* document and make sure to cleanup

* fix annotation to match docs

* fix docs build

* address PR feedback

Author: winglian

_quarto.yml

@@ -139,7 +139,8 @@ quartodoc:
         - utils.optimizers.adopt
         - utils.data.pretraining
         - utils.data.sft
-        - utils.gradient_checkpointing.unsloth
+        - utils.gradient_checkpointing.offload_cpu
+        - utils.gradient_checkpointing.offload_disk
     - title: Schemas
       desc: Pydantic data models for Axolotl config
       contents:

docs/config.qmd

@@ -539,7 +539,7 @@ train_on_inputs: false
 # Note that training loss may have an oscillating pattern with this enabled.
 group_by_length: false
 
-# Whether to use gradient checkpointing. Available options are: true, false, "offload".
+# Whether to use gradient checkpointing. Available options are: true, false, "offload", "offload_disk".
 # https://huggingface.co/docs/transformers/v4.18.0/en/performance#gradient-checkpointing
 gradient_checkpointing: false
 # additional kwargs to pass to the trainer for gradient checkpointing

src/axolotl/utils/gradient_checkpointing/__init__.py

@@ -5,8 +5,11 @@
 
 from packaging import version
 
-from axolotl.utils.gradient_checkpointing.unsloth import (
-    Unsloth_Offloaded_Gradient_Checkpointer,
+from axolotl.utils.gradient_checkpointing.offload_cpu import (
+    CPU_Offloaded_Gradient_Checkpointer,
+)
+from axolotl.utils.gradient_checkpointing.offload_disk import (
+    Disco,
 )
 
 transformers_version = version.parse(importlib.metadata.version("transformers"))
@@ -26,12 +29,31 @@ def hf_grad_checkpoint_offload_wrapper(
     decoder_layer, *args, use_reentrant=None
 ):  # pylint: disable=unused-argument
     if uses_gc_layers(decoder_layer):
-        return Unsloth_Offloaded_Gradient_Checkpointer.apply(
+        return CPU_Offloaded_Gradient_Checkpointer.apply(
+            decoder_layer,
+            *args,
+        )
+
+    return CPU_Offloaded_Gradient_Checkpointer.apply(
+        (
+            decoder_layer.func.__self__
+            if isinstance(decoder_layer, partial)
+            else decoder_layer.__self__
+        ),
+        *args,
+    )
+
+
+def hf_grad_checkpoint_disk_offload_wrapper(
+    decoder_layer, *args, use_reentrant=None
+):  # pylint: disable=unused-argument
+    if uses_gc_layers(decoder_layer):
+        return Disco.apply(
             decoder_layer,
             *args,
         )
 
-    return Unsloth_Offloaded_Gradient_Checkpointer.apply(
+    return Disco.apply(
         (
             decoder_layer.func.__self__
             if isinstance(decoder_layer, partial)

src/axolotl/utils/gradient_checkpointing/unsloth.py renamed to src/axolotl/utils/gradient_checkpointing/offload_cpu.py

@@ -1,4 +1,4 @@
-"""Unsloth checkpointing"""
+"""CPU offloaded checkpointing"""
 
 # Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
 #
@@ -26,7 +26,7 @@
     torch_cuda_amp_custom_bwd = torch.amp.custom_bwd(device_type="cuda")
 
 
-class Unsloth_Offloaded_Gradient_Checkpointer(  # pylint: disable=invalid-name
+class CPU_Offloaded_Gradient_Checkpointer(  # pylint: disable=invalid-name
     torch.autograd.Function
 ):
     """