trl/models/activation_offloading.py

# Copyright 2020-2025 The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of https://github.com/pytorch/torchtune.

import warnings

import psutil
import torch
from torch import nn
from torch.autograd.graph import saved_tensors_hooks


class OffloadActivations(saved_tensors_hooks):
    """
    Context manager under which activation tensors created in the forward pass will be offloaded.

    Enable the memory efficiency technique of activation offloading, where activations bigger than `min_offload_size`
    bytes will be offloaded to CPU in the forward and brought back in the backward. This is in contrast to maintaining
    the activation on GPU VRAM throughout the program.

    This manager contains the option of using one additional CUDA stream to handle the communication between CUDA and
    CPU, which is intended to overlap with the default computation stream to improve runtime. We designed
    synchronization with a few heuristics for optimizing the tradeoff between runtime vs memory usage.

    Args:
        use_pin_memory (`bool`, *optional*, defaults to `True`):
            Whether to offloaded Tensor will be placed in pinned memory on the CPU. Pinned memory allows the Tensor to
            be moved back onto GPU more quickly but is a limited resource.
        use_streams (`bool`, *optional*, defaults to `True`):
            Whether to use streams for performance optimization where the communications get overlapped with the
            computation. Requires a torch build after torch-2.5.0.
        min_offload_size (`int`, *optional*, defaults to `1024`):
            Minimum number of bytes a Tensor must be in order to qualify for offloading. If the tensor is too small, we
            do not want to waste bandwidth and resources moving it to CPU and back.
        max_fwd_stash_size (`int`, *optional*, defaults to `5`):
            Maximum size of the forward stash, or the maximum number of consecutive activations to keep alive during
            the forward pass. This number must be at least 1. Keeping alive more activations will potentially allow
            more overlap between the communication and compute streams at the cost of increasing memory usage. Keeping
            alive fewer activations will conserve memory, but may cause poor overlap between the streams, increasing
            runtime.

    Raises:
        ValueError: if `max_fwd_stash_size` is not at least `1`.

    Example:
        >>> with OffloadActivations():
        >>>     outputs = model(inputs, labels=labels)
        >>> loss = outputs.loss
        >>> loss.backward()
    """

    def __init__(
        self,
        use_pin_memory: bool = True,
        use_streams: bool = True,
        min_offload_size: int = 1024,
        max_fwd_stash_size: int = 5,
    ) -> None:
        self.use_streams = use_streams

        self.min_tensor_size_bytes = min_offload_size  # we don't want to bother with small tensors
        self.tracker = {}  # tensor_id => (new_tensor, if_modified)  ---> track what saved/offloaded tensors are where
        self.tensor_id = 0
        self.is_first_forward_call = True
        self.is_first_backward_call = True
        self.is_first_forward_pass = True

        # Managing cpu memory
        self.use_pin_memory = use_pin_memory
        self.virtual_memory_safe_pct = 60  # we should not exceed this percentage of memory

        self.accelerator_type = (
            torch.accelerator.current_accelerator().type if hasattr(torch, "accelerator") else "cuda"
        )
        # NOTE: xpu doesn't have `default_stream` API, use `current_stream` instead
        self.s0 = (
            torch.xpu.current_stream() if self.accelerator_type == "xpu" else torch.cuda.default_stream()
        )  # comp stream

        # For streaming
        if self.use_streams:
            self.s1 = torch.Stream() if self.accelerator_type == "xpu" else torch.cuda.Stream()  # comms stream
            self.fwd_stash = {}  # tensor_id => (activation, ev1)
            if max_fwd_stash_size < 1:
                raise ValueError(f"max_fwd_stash_size should be at least 1 but is {max_fwd_stash_size}")
            self.max_fwd_stash_size = max_fwd_stash_size
            self.bwd_tensor_stash = {}  # tensor_id => activation
            self.bwd_ev_stash = {}  # tensor_id => ev0
            self.curr_graph_id = None
            self.curr_autograd_node = None

        # -------- platform util functions -------- #
        def verify_sufficient_virtual_memory():
            curr_pct = get_cpu_ram_pct()
            if curr_pct > self.virtual_memory_safe_pct:
                warnings.warn(f"{curr_pct=}% > {self.virtual_memory_safe_pct=}% of virtual memory used")

        def get_cpu_ram_pct() -> float:
            # get the percentage of memory used by the system
            return psutil.virtual_memory().percent

        def get_tensor_id() -> int:
            # create a unique id for each tensor we are managing
            self.tensor_id += 1
            return self.tensor_id

        def get_num_bytes_tensor(x: torch.Tensor) -> int:
            # get the number of bytes in a tensor, for memory management purposes
            return x.element_size() * x.nelement()  # x.element_size() * x._base_storage().nbytes()

        # -------- core pack / unpack work -------- #
        def pack_tensor(activation: torch.Tensor) -> int:
            # activations are passed in during forward pass - from here we take over and return a unique id
            if self.is_first_forward_call:
                if len(self.tracker) != 0:
                    raise ValueError("Backward pass should have cleared tracker of all tensors")

                # set training phase trackers
                self.is_first_forward_call = False
                self.is_first_backward_call = True

            # query for basic tensor info
            num_bytes = get_num_bytes_tensor(activation)
            tensor_id = get_tensor_id()

            # only offload hefty bois if they're activations on CUDA (our heuristic
            # for that is to check if they're not params or buffers)!
            if (
                activation.device.type in ["cuda", "xpu"]
                and num_bytes >= self.min_tensor_size_bytes
                and (
                    not isinstance(activation, torch.nn.Parameter)
                    and not (hasattr(torch.nn, "Buffer") and isinstance(activation, torch.nn.Buffer))
                )
            ):
                if self.use_streams:
                    # First, sync back and dereference previously offloaded tensors
                    # as the offloading should be done sufficiently long ago.
                    for id in list(self.fwd_stash.keys()):
                        if id <= tensor_id - self.max_fwd_stash_size:
                            _, ev = self.fwd_stash[id]
                            self.s0.wait_event(ev)
                            del self.fwd_stash[id]
                        else:
                            break

                    # Sync in, offload, and add an event to sync back later
                    self.s1.wait_stream(self.s0)

                stream = self.s1 if self.use_streams else self.s0
                with stream if self.accelerator_type == "xpu" else torch.cuda.stream(stream):
                    cpu_tensor = torch.empty_like(activation, pin_memory=self.use_pin_memory, device="cpu")
                    cpu_tensor.copy_(activation, non_blocking=True)
                    self.tracker[tensor_id] = (
                        cpu_tensor,
                        True,  # True = (in future) modified
                    )

                if self.use_streams:
                    event = self.s1.record_event()

                    # Stash to keep activation alive til s1 is done
                    self.fwd_stash[tensor_id] = (activation, event)
            else:
                self.tracker[tensor_id] = (
                    activation,
                    False,
                )  # False = not modified, tensor is as is

            return tensor_id

        def unpack_tensor_single_stream(unpack_tensor_id: int) -> torch.Tensor:
            # backward pass - we are called with the tensor_id, which
            # we will use to retrieve the saved/offloaded tensor
            if self.is_first_backward_call:
                if self.is_first_forward_pass:
                    self.is_first_forward_pass = False
                    if self.use_pin_memory:
                        verify_sufficient_virtual_memory()

                self.is_first_backward_call = False
                self.is_first_forward_call = True

            if unpack_tensor_id not in self.tracker:
                raise ValueError(f"Untracked tensor with id {unpack_tensor_id}")

            maybe_accelerator_tensor, modified = self.tracker[unpack_tensor_id]
            if modified:
                accelerator_tensor = maybe_accelerator_tensor.to(self.accelerator_type, non_blocking=True)
                maybe_accelerator_tensor = accelerator_tensor

            # clear tensor from tracking
            del self.tracker[unpack_tensor_id]
            return maybe_accelerator_tensor

        def unpack_tensor_with_streams(unpack_tensor_id: int) -> torch.Tensor:
            # backward pass - we are called with the tensor_id, which
            # we will use to retrieve the saved/offloaded tensor
            if self.is_first_backward_call:
                self.curr_graph_id = torch._C._current_graph_task_id()

                def wait_and_del_remaining_references() -> None:
                    for id in list(self.bwd_tensor_stash.keys()):
                        event = self.bwd_ev_stash[id]
                        self.s1.wait_event(event)
                        del self.bwd_tensor_stash[id]

                # Register a callback to the end of autograd to clean everything up
                torch.autograd.variable.Variable._execution_engine.queue_callback(wait_and_del_remaining_references)

                if self.is_first_forward_pass:
                    self.is_first_forward_pass = False
                    if self.use_pin_memory:
                        verify_sufficient_virtual_memory()

                self.is_first_backward_call = False
                self.is_first_forward_call = True

            if unpack_tensor_id not in self.tracker:
                raise ValueError(f"untracked tensor with id {unpack_tensor_id}")

            maybe_accelerator_tensor, modified = self.tracker[unpack_tensor_id]
            if modified:
                # Get data on the current autograd node
                graph_id = torch._C._current_graph_task_id()
                node = torch._C._current_autograd_node()
                prev_node_ids = []

                # If we're on a new node, mark prev node's tensors to be freed later
                if graph_id == self.curr_graph_id and self.curr_autograd_node != node:
                    self.curr_autograd_node = node
                    prev_node_ids = list(self.bwd_tensor_stash.keys())

                brought_back_from_cpu = True
                if unpack_tensor_id in self.fwd_stash:
                    maybe_accelerator_tensor = self.fwd_stash[unpack_tensor_id][0]
                    brought_back_from_cpu = False
                else:
                    # Kick off the process to bring tensors back
                    with self.s1 if self.accelerator_type == "xpu" else torch.cuda.stream(self.s1):
                        accelerator_tensor = maybe_accelerator_tensor.to(self.accelerator_type, non_blocking=True)
                        maybe_accelerator_tensor = accelerator_tensor

                    # Tell comp stream to wait for the info to be loaded before executing
                    self.s0.wait_stream(self.s1)

                    # Stash the tensor to keep memory alive until compute stream is complete
                    self.bwd_tensor_stash[unpack_tensor_id] = maybe_accelerator_tensor

                    # Note: [Track views of the unpacked]
                    # Why do we get the use count of the unpacked tensor here? We want an
                    # initial count to compare to later, during the post-hook of the
                    # backward node, when we need to decide whether we're allowed to free
                    # the tensor yet. In what obscure cases must we delay freeing the
                    # tensor (and thus call record_stream)?
                    # 1. Any of the outputs of the backward node is a view of the unpacked
                    #    tensor.
                    # 2. In the case that this unpacked tensor will be used in a
                    #    checkpointed region, if one of the recomputed saved tensors ends
                    #    up as a view of the unpacked tensor.
                    # 3. The user abuses the system somehow and manually relies on the
                    #    unpacked tensor to exist after the backward node has executed.
                    storage_refcount = torch._C._storage_Use_Count(maybe_accelerator_tensor.untyped_storage()._cdata)

                def hook(outputs, inputs):
                    # create events for the current node inputs/outputs if they were streamed in
                    if brought_back_from_cpu:
                        # See Note: [Track views of the unpacked]
                        # IF any of the outputs is a view of the tensor, OR if a view of
                        # the tensor has been saved as a part of checkpoint's recompute
                        # process, OR the user has abusedly incurred a reference on the
                        # unpacked tensor, THEN the tensor might be used later and we
                        # cannot presume to delete it after only the current node is
                        # done! So we use our frenemy, record_stream, to ensure the
                        # Tensor stays unmessed with until it's done getting used in the
                        # compute stream (s0 here). Note that the con here is we introduce
                        # non-deterministic (thus higher) memory usage, but this case
                        # should not happen often.
                        unpacked_tensor = self.bwd_tensor_stash[unpack_tensor_id]
                        if torch._C._storage_Use_Count(unpacked_tensor.untyped_storage()._cdata) > storage_refcount:
                            unpacked_tensor.record_stream(self.s0)
                            del self.bwd_tensor_stash[unpack_tensor_id]
                        else:
                            event = self.s0.record_event()
                            self.bwd_ev_stash[unpack_tensor_id] = event

                    # if there are still things in the fwd_stash, get rid of them as we're in bwd now
                    for id in list(self.fwd_stash.keys()):
                        _, ev = self.fwd_stash[id]
                        self.s0.wait_event(ev)
                        del self.fwd_stash[id]

                    # wait on prev node's events and del those
                    for id in prev_node_ids:
                        event = self.bwd_ev_stash[id]
                        self.s1.wait_event(event)
                        del self.bwd_tensor_stash[id]

                    return outputs

                node.register_hook(hook)

            # clear tensor from tracking
            del self.tracker[unpack_tensor_id]
            return maybe_accelerator_tensor

        unpack_tensor = unpack_tensor_with_streams if self.use_streams else unpack_tensor_single_stream
        super().__init__(pack_tensor, unpack_tensor)


class NoOpManager(saved_tensors_hooks):
    """
    A `saved_tensors_hook` manager used to disable any other `saved_tensors_hook` manager applied before. This relies
    on the behavior that only the most recently registered `saved_tensors_hook` will run.

    One example usage is to opt a local region of code out of activations offloading, which is usually applied globally
    to best track state.
    """

    def __init__(self) -> None:
        def noop(tensor):
            return tensor

        super().__init__(noop, noop)


def get_act_offloading_ctx_manager(
    model: nn.Module,
    use_pin_memory: bool = True,
    use_streams: bool = True,
    min_offload_size: int = 1024,
    max_fwd_stash_size: int = 5,
    warn_if_no_head: bool = True,
) -> OffloadActivations:
    """
    Returns the activation offloading context manager for the model. All but the last output Linear in every step will
    be offloaded.

    If activation offloading is enabled, we return the OffloadActivations context manager.
    If activation offloading is disabled, we return a NoOpManager context manager.

    Args:
        model (`nn.Module`):
            Model to wrap with the activation offloading context manager.
        use_pin_memory (`bool`, *optional*, defaults to `True`):
            Whether to offloaded Tensor will be placed in pinned memory on the CPU. Pinned memory allows the Tensor to
            be moved back onto GPU more quickly but is a limited resource.
        use_streams (`bool`, *optional*, defaults to `True`):
            Whether to use streams for performance optimization where the communications get overlapped with the
            computation. Requires a torch build after torch-2.5.0.
        min_offload_size (`int`, *optional*, defaults to `1024`):
            Minimum number of bytes a Tensor must be in order to qualify for offloading. If the tensor is too small, we
            do not want to waste bandwidth and resources moving it to CPU and back.
        max_fwd_stash_size (`int`, *optional*, defaults to `5`):
            Maximum size of the forward stash, or the maximum number of consecutive activations to keep alive during
            the forward pass. This number must be at least 1. Keeping alive more activations will potentially allow
            more overlap between the communication and compute streams at the cost of increasing memory usage. Keeping
            alive fewer activations will conserve memory, but may cause poor overlap between the streams, increasing
            runtime.
        warn_if_no_head (`bool`, *optional*, defaults to `True`):
            Whether to warn if no output head is detected. If set to `False`, no warning will be raised if no output
            head is detected.

    Returns:
        `contextlib.ContextDecorator`:
            Activation offloading context manager for the model.
    """
    activations_handling_ctx = OffloadActivations(
        use_pin_memory=use_pin_memory,
        use_streams=use_streams,
        min_offload_size=min_offload_size,
        max_fwd_stash_size=max_fwd_stash_size,
    )

    # Below is our hack to disable offloading the last output Linear in every
    # step, as the cost for offloading the activation and then soon after bringing
    # it back is expensive.
    output_head_detected = False
    noop_ctx = NoOpManager()

    # Try to get the actual model if it's wrapped
    unwrapped_model = model
    if hasattr(unwrapped_model, "module"):
        unwrapped_model = unwrapped_model.module
    # check for PEFT models
    if hasattr(unwrapped_model, "base_model") and hasattr(unwrapped_model, "peft_config"):
        unwrapped_model = unwrapped_model.base_model

    # Check for different types of output heads
    if hasattr(unwrapped_model, "output"):
        if isinstance(unwrapped_model.output, nn.Module):
            unwrapped_model.output.register_forward_pre_hook(lambda *args: noop_ctx.__enter__())
            unwrapped_model.output.register_forward_hook(lambda *args: noop_ctx.__exit__(), always_call=True)
            output_head_detected = True
        elif hasattr(unwrapped_model.output, "linear") and isinstance(unwrapped_model.output.linear, nn.Module):
            unwrapped_model.output.linear.register_forward_pre_hook(lambda *args: noop_ctx.__enter__())
            unwrapped_model.output.linear.register_forward_hook(lambda *args: noop_ctx.__exit__(), always_call=True)
            output_head_detected = True

    # Check for HuggingFace model output heads
    elif hasattr(unwrapped_model, "lm_head"):
        unwrapped_model.lm_head.register_forward_pre_hook(lambda *args: noop_ctx.__enter__())
        unwrapped_model.lm_head.register_forward_hook(lambda *args: noop_ctx.__exit__(), always_call=True)
        output_head_detected = True

    # Check for decoder-based models
    elif hasattr(unwrapped_model, "decoder"):
        decoder = unwrapped_model.decoder
        if hasattr(decoder, "output"):
            decoder.output.register_forward_pre_hook(lambda *args: noop_ctx.__enter__())
            decoder.output.register_forward_hook(lambda *args: noop_ctx.__exit__(), always_call=True)
            output_head_detected = True
        # Some models have lm_head in the decoder
        elif hasattr(decoder, "lm_head"):
            decoder.lm_head.register_forward_pre_hook(lambda *args: noop_ctx.__enter__())
            decoder.lm_head.register_forward_hook(lambda *args: noop_ctx.__exit__(), always_call=True)
            output_head_detected = True

    # Check for transformer models with final layer norm
    elif hasattr(unwrapped_model, "final_layer_norm") or hasattr(unwrapped_model, "ln_f"):
        final_norm = getattr(unwrapped_model, "final_layer_norm", None) or unwrapped_model.ln_f
        final_norm.register_forward_pre_hook(lambda *args: noop_ctx.__enter__())
        final_norm.register_forward_hook(lambda *args: noop_ctx.__exit__(), always_call=True)
        output_head_detected = True

    # Check for models with head module
    elif hasattr(unwrapped_model, "head") and isinstance(unwrapped_model.head, nn.Module):
        unwrapped_model.head.register_forward_pre_hook(lambda *args: noop_ctx.__enter__())
        unwrapped_model.head.register_forward_hook(lambda *args: noop_ctx.__exit__(), always_call=True)
        output_head_detected = True

    if not output_head_detected and warn_if_no_head:
        warnings.warn(
            "During activation offloading, no output head was detected. If your model has an output head, it will be "
            "offloaded. This usually greatly slows training, given the large vocabulary size. To change this "
            "behavior, set your output head as model.output and make it an nn.Module. You can disable this warning by "
            "passing `warn_if_no_head=False`."
        )

    # Disable offloading for any Liger modules
    for name, module in unwrapped_model.named_modules():
        if "liger" in name.lower():
            module.register_forward_pre_hook(lambda *args: noop_ctx.__enter__())
            module.register_forward_hook(lambda *args: noop_ctx.__exit__(), always_call=True)

    return activations_handling_ctx