This change is to match the behavior of _record_memory_history which was
recently changed to enable history recording on all devices rather than
the current one. It prevents confusing situations where the observer
was registered before the device was set for the training run.
It also ensures the allocators have been initialized in the python binding just in case this is the first call to the CUDA API.
Fixes#107330
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107399
Approved by: https://github.com/eellison
ghstack dependencies: #107171
Previously when we recorded a free action in a memory trace, we would provide
the stack for when the block was allocated. This is faster because we do not
have to record stacks for free, which would otherwise double the number of stacks
collected. However, sometimes knowing the location of a free is useful for
figuring out why a tensor was live. So this PR adds this behavior. If
performance ends up being a concern the old behavior is possible by passing
"alloc" to the context argument rather than "all".
Also refactors some of glue logic to be consistent across C++ and Python and
routes the Python API through the C++ version.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/106758
Approved by: https://github.com/albanD
We want to display the stack for the original cudaMalloc that created a segment.
Previously we could only report the last time the segment memory was used,
or the record of the segment_alloc could appear in the list of allocator actions.
This PR ensure regardless of whether we still have the segment_alloc action,
the context for a segment is still available. The visualizer is updated to
be able to incorporate this information.
This PR adds a new field to Block. However the previous stacked cleanup PR
removed a field of the same size, making the change to Block size-neutral.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/106113
Approved by: https://github.com/aaronenyeshi
For free blocks of memory in the allocator, we previously kept a linked list
of the stack frames of previous allocations that lived there. This was only
ever used in one flamegraph visualization and never proved useful at
understanding what was going on. When memory history tracing was added, it
became redundant, since we can see the history of the free space from recording
the previous actions anyway.
This patch removes this functionality and simplifies the snapshot format:
allocated blocks directly have a 'frames' attribute rather than burying stack frames in the history.
Previously the memory history tracked the real size of allocations before rounding.
Since history was added, 'requested_size' has been added directly to the block which records the same information,
so this patch also removes that redundancy.
None of this functionality has been part of a PyTorch release with BC guarentees, so it should be safe to alter
this part of the format.
This patch also updates our visualization tools to work with the simplified format. Visualization tools keep
support for the old format in `_legacy` functions so that during the transition old snapshot files can still be read.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/106079
Approved by: https://github.com/eellison
When we run cudagraph trees we are not allowed to have permanent workspace allocations like in cublas because we might need to reclaim that memory for a previous cudagraph recording, and it is memory that is not accounted for in output weakrefs so it does not work with checkpointing. Previously, I would check that we didn't have any additional allocations through snapshotting. This was extremely slow so I had to turn it off.
This PR first does the quick checking to see if we are in an error state, then if we are does the slow logic of creating snapshot. Also turns on history recording so we get a stacktrace of where the bad allocation came from.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99985
Approved by: https://github.com/zdevito
In this stack of PRs we adding caching to output tensors for cudagraph trees after we've done initial recording. On initial recording we do not cache tensor outputs because this prevents memory from being reclaimed. On subsequent exeuctions we do cache them to avoid overhead. However, because there is an extra reference around, this caused divergent recording & execution behavior in both autocast caching and autograd gradient stealing. Divergent recording & execution would keep on re-recording and eventually stabilize, but it's not what you want to see happen.
This pr makes the autocast cache and buffer stealing aware of the cudagraph static output tensors.
I will add this to the other cudagraph impl in another pr.
Not sure if this should be in autograd or in autocast since it affects both.. Or somewhere else
Pull Request resolved: https://github.com/pytorch/pytorch/pull/99368
Approved by: https://github.com/albanD, https://github.com/ezyang
Caches output tensors for the common case when the output Tensor storage is unaliased for all graph outputs in all paths. For these persisted tensors we adjust the liveness tracking by also checking that the output tensor does not have an additional python reference.
I limit cached output tensors to be unaliased. If a descendent node discovers it has an alias of a prior output, then the aliased output will no longer be persisted in the ancestor.
The large majority of tensors are unaliased, and preserving aliased output tensors would add significant additional complexity with marginal gains. For instance, when do checkpointing and re-recordings, we need to remove the persisted tensors otherwise it would prevent memory from being reclaimed. If a single persisted tensor was present in multiple paths then that would create an inter-path dependence which adds complexity. Additionally, each further caching of the output would affect the reference count of the other caches, and that reference count would also need to be adjusted depending on if a node was checkpointed.
Still need to do a complete a run but for the models I tried makes the performance extremely close between trees and non trees impl.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98944
Approved by: https://github.com/jansel, https://github.com/ngimel
Common advice we give for handling memory fragmentation issues is to
allocate a big block upfront to reserve memory which will get split up later.
For programs with changing tensor sizes this can be especially helpful to
avoid OOMs that happen the first time we see a new largest input and would
otherwise have to allocate new segments.
However the issue with allocating a block upfront is that is nearly impossible
to correctly estimate the size of that block. If too small, space in the block
will run out and the allocator will allocate separate blocks anyway. Too large,
and other non-PyTorch libraries might stop working because they cannot allocate
any memory.
This patch provides the same benefits as using a pre-allocating block but
without having to choose its size upfront. Using the cuMemMap-style APIs,
it adds the ability to expand the last block in a segment when more memory is
needed.
Compared to universally using cudaMallocAsync to avoid fragmentation,
this patch can fix this common fragmentation issue while preserving most
of the existing allocator behavior. This behavior can be enabled and disabled dynamically.
This should allow users to, for instance, allocate long-lived parameters and state in individual buffers,
and put temporary state into the large expandable blocks, further reducing
fragmentation.
See inline comments for information about the implementation and its limitations.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/96995
Approved by: https://github.com/eellison
Significantly reduces overhead of constructing Tensors and Storages and checking Storage Liveness. Removes the regression for HF models that I tested and removes 75% of overhead of the extremely overhead bound resnet50 training we have in torchbench. (.91x base commit, 1.02x torchinductor default, 1.16x this PR, 1.25 previous cudagraphs impl).
This PR takes care of all of the lower hanging fruit.
- Computes storage aliasing at record time instead of during at runtime. We no longer need to use a runtime storage cache, and can instead index directly into the existing alias if there is one, or construct a new Storage
- Moves the heavyweight C++ calls into a batch - getting storage weakrefs and constructing tensors
Pull Request resolved: https://github.com/pytorch/pytorch/pull/98529
Approved by: https://github.com/jansel, https://github.com/ngimel
This method has to be accessible from `c10` to enable CUDA-12 integration.
Implemented by providing private `c10::cuda:_internal::setHasPrimaryContext` that passes the pointer to the implementation (in `torch_cuda`) back to c10.
Use global class constructor/destructor to guarantee RAII.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/96800
Approved by: https://github.com/ngimel
CUDA Graph Trees
Design doc: https://docs.google.com/document/d/1ZrxLGWz7T45MSX6gPsL6Ln4t0eZCSfWewtJ_qLd_D0E/edit
Not currently implemented :
- Right now, we are using weak tensor refs from outputs to check if a tensor has dies. This doesn't work because a) aliasing, and b) aot_autograd detaches tensors (see note [Detaching saved tensors in AOTAutograd]). Would need either https://github.com/pytorch/pytorch/issues/91395 to land to use storage weak refs or manually add a deleter fn that does what I want. This is doable but theres some interactions with the caching allocator checkpointing so saving for a stacked pr.
- Reclaiming memory from the inputs during model recording. This isn't terribly difficult but deferring to another PR. You would need to write over the input memory during warmup, and therefore copy the inputs to cpu. Saving for a stacked pr.
- Warning on overwriting previous generation outputs. and handling nested torch.compile() calls in generation tracking
Differential Revision: [D43999887](https://our.internmc.facebook.com/intern/diff/D43999887)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/89146
Approved by: https://github.com/ezyang
Previously the allocator would query whether a stream was recording a graph,
and look up the pool associated with a graph. This change has the allocator
directly associate a stream with a mempool, decoupling "record this stream to a pool"
from the action of "record all actions to a cuda graph".
Pull Request resolved: https://github.com/pytorch/pytorch/pull/96542
Approved by: https://github.com/eellison
This refactors the stack trace facility specific to memory profiling
in python+cuda to make a generic facility to generate combined stack
traces.
The generic facility (combined_traceback.h) does not require
python to be around to work, but will return python stacks if it is
present.
This facility is then used to add support for stack trace gathering in memory profiling that
happens directly from C++.
It is also used to expose a python API for gathering and symbolizing
combineds stacks.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95541
Approved by: https://github.com/ezyang
When we checkpoint the state of the private pool allocator, we will need to make sure that its current live allocated blocks will get properly cleaned up when the tensors they correspond to die. Return DataPtrs for these new allocated blocks that the callee can swap onto live Tensors.
The exact api for setting the checkpoint can be manipulated after this as the cudagraph implementation is built out, but this at least shows its sufficiently general.
This should be the last PR touching cuda caching allocator necessary for new cudagraphs integration.
Differential Revision: [D43999888](https://our.internmc.facebook.com/intern/diff/D43999888)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95020
Approved by: https://github.com/zdevito
Copying note from cuda caching allocator:
```
* Note [Checkpointing PrivatePoolState]
*
* Refer above to Note [Interaction with CUDA graph capture]. Allocations made
* during graph capture are made from a separate private pool. During graph
* capture allocations behave as usual. During graph replay the allocator
* state does not change even as new tensors are created. The private pool
* will not free its blocks to the main caching allocator until cuda graph use
* is finished to prevent an allocation from eager clobbering the memory from
* a live but unaccounted for tensor that was created during replay.
*
* `make_graphed_callables`, a series of separate callables chained in
* successive cuda graphs, can share a memory pool because after a cuda graph
* recording the allocations in the shared private pool exactly reflect the
* tensors that are allocated.
*
* We would like to extend callable chaining to support a graphed callable
* tree. In this scenario, we have a tree of callable chains which will be
* captured with cuda graphs. In the diagram below, we have a tree with four
* callables, A, B, C, and D. Suppose we have captured, and subsequently
* replayed, A, B, and C. Then on a new invocation, we replay A and B, but
* would now like to record D. At this point the private pool will not reflect
* any of the live tensors created during graph replay. Allocations made
* during a new recording with the pool could overwrite those live tensors.
*
* In order to record a new graph capture after replaying prior callables in
* the tree, we need the allocator to reflect the state of the live tensors.
* We checkpoint the state of the private after each recording, and then
* reapply it when we are starting a new recording chain. Additionally, we
* must free the allocations for any tensors that died between the end of our
* previous graph replaying and our new recording (TODO). All of the allocated
* segments that existed in the checkpointed state must still exist in the
* pool. There may also exist new segments, which we will free (TODO : link
* note [live tensors between iterations] when it exists).
*
*
* ---------------> A ---------------> B ---------------> C
* |
* |
* |
* |
* ---------------> D
```
A few TODOs:
- need to add logic for freeing tensors that have died between a last replay and current new recording
- Add logic for free that might be called on a pointer multiple times (because we are manually freeing live tensors)
The two scenarios above have not been exercised in the tests yet.
Differential Revision: [D43999889](https://our.internmc.facebook.com/intern/diff/D43999889)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/94653
Approved by: https://github.com/zdevito
Adds the ability to quickly generate stack traces for C++,
and combine Python, TorchScript, and C++ frames into a single trace.
This makes it possible for the memory tracer to record allocations inside
C++ code (e.g. convolution temporaries, backward operators).
The unwinder code is ~10x faster than execinfo.h's backward because it
cache fast unwinder routines for instruction pointers that have already been seen.
It is also only 1.2--2x slower than copying the entire stack (the approach perf takes),
while using 2 orders of magnitude less space per stack.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95357
Approved by: https://github.com/bertmaher
Summary:
The caching allocator can be configured to round memory allocations in order to reduce fragmentation. Sometimes however, the overhead from rounding can be higher than the fragmentation it helps reduce.
We have added a new stat to CUDA caching allocator stats to help track if rounding is adding too much overhead and help tune the roundup_power2_divisions flag:
- "requested_bytes.{current,peak,allocated,freed}": memory requested by client code, compare this with allocated_bytes to check if allocation rounding adds too much overhead
Test Plan: Added test case in caffe2/test/test_cuda.py
Differential Revision: D40810674
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88575
Approved by: https://github.com/zdevito
This adds `torch.cuda._DeviceGuard` which is a stripped down version of
`torch.cuda.device` with lower overhead. To do this, it only accepts `int` as
the device so we don't need to call `_get_device_index` and is implemented
with a new C++ helper `torch._C._cuda_exchangeDevice` that allows
`_DeviceGuard.__enter__` to be just a single function call. On my machine,
I see a drop from 3.8us of overhead to 0.94 us with this simple benchmark:
```python
def set_device():
with torch.cuda.device(0):
pass
%timeit set_device()
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/91045
Approved by: https://github.com/ngimel, https://github.com/anijain2305
#75854
A naive attempt at working around the limitations of using a single 64-bit integer to pack `stream_id`, `device_index`, and `device_type`.
Stills needs sanity checks, testing, and minimization of BC-breaking changes.
Currently a Holder for the `StreamData3` struct is used for `IValue` compatibility. While doing this seems to work for `ivalue.h` and `ivalue_inl.h`, this doesn't seem to be naively working for the JIT CUDA stream wrapper? (Something about ambiguous calls if an `intrusive_ptr` to `c10::ivalue::StreamData3Holder` is used as the return type for `pack()`. It turns out that the methods required to access the fields for rematerializing a CUDA Stream are basically already present anyway, so `pack` is simply removed in the wrapper for now and the methods to access the required fields are called directly.
CC @ptrblck
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81596
Approved by: https://github.com/ezyang
