Based on the [conversation](https://github.com/pytorch/pytorch/issues/121791), we plan to drop the "highest, high, medium" to represent fp32 internal computation data types . Instead, we will directly use the algorithm to represent it.
### Design Choice: Directly use algorithms name like "TF32", "BF16".
#### Pros
- The names are more informative. 'tf32' is more informative than a simple "high".
- Easier to extend new algorithm like `tf32x3`
#### Cons
- "HIGHEST, HIGH, MEDIUM" indicated the relative precision between different algorithms. However, we can have more documents to discuss them.
### We provide a layered structure for backends/operators.
('f32' is short for 'fp32_precision')
### We provide 3 fp32 compute precision can be set:
- **"ieee"**: Not allowed to use any other internal computation data types .
- **"tf32"**: Allowed to use tf32 as internal computation data types.
- **"bf16"**: Allowed to use bf16 as internal computation data types.
- **"none"**: Precision's are not set. Can be override by its father node.
### Overriding Precision Settings
Child node can be override by its father node if it is set to default.
For current default settings:
```
backend = generic, op = all, precision setting = none
backend = cuda, op = all, precision setting = none
backend = cuda, op = conv, precision setting = tf32
backend = cuda, op = rnn, precision setting = tf32
backend = cuda, op = matmul, precision setting = none
backend = matmul, op = all, precision setting = none
backend = matmul, op = conv, precision setting = none
backend = matmul, op = rnn, precision setting = none
backend = matmul, op = matmul, precision setting = none
```
- If the user set `torch.backends.mkldnn.fp32_precision="bf16"`, his child nodes `torch.backends.mkldnn.matmul.fp32_precision` / `torch.backends.mkldnn.conv.fp32_precision` / `torch.backends.mkldnn.rnn.fp32_precision` will also be override to "bf16".
- If the user set `torch.backends.fp32_precision="bf16"`, `torch.backends.mkldnn.fp32_precision` and his child nodes will also we override to "bf16".
### Backward Compatible
Since new API allow user to have more fine-grained control. There will be some conflict. For example, previous `torch.backends.cudnn.allow_tf32` are not enough to represent the status for `torch.backends.cudnn.rnn.fp32_precision="ieee"` and `torch.backends.cudnn.conv.fp32_precision="tf32"`. Therefore, our goal for backward compatible is
- If the user only uses previous APIs, it will work as previous expectations.
- If the user use **new** API to change the status to an **un-representable** status for old API, and try to access the status by **old** API. We will raise Runtime Error and point the document for user.
### Test Plan
```
python test/test_cuda.py -k test_fp32_precision_with_tf32
python test/test_cuda.py -k test_fp32_precision_with_float32_matmul_precision
python test/test_cuda.py -k test_invalid_status_for_legacy_api
python test/test_mkldnn.py -k test_mlkdnn_get_set
python test/test_mkldnn.py -k test_generic_precision
python test/test_mkldnn.py -k test_invalid
python test/test_mkldnn.py -k test_default_use_parent
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/125888
Approved by: https://github.com/jgong5, https://github.com/albanD
Co-authored-by: Jiang, Yanbing <yanbing.jiang@intel.com>
MemPool is a separate pool of memory handled by the caching allocator. This PR adds the option let the caching allocator try to use this pool as a last resort instead of OOMing by associating a use_on_oom bool with each MemPool.
Usage:
Users can optionally specify a ``use_on_oom`` bool (which is False by default) during MemPool creation. If true, then the CUDACachingAllocator will be able to use memory in this pool as a last resort instead of OOMing.
```
pool = torch.cuda.MemPool(allocator, use_on_oom=True)
with torch.cuda.use_mem_pool(pool):
a = torch.randn(40 * 1024 * 1024, dtype=torch.uint8, device="cuda")
del a
# at the memory limit, this will succeed by using pool's memory in order to avoid the oom
b = torch.randn(40 * 1024 * 1024, dtype=torch.uint8, device="cuda")
```
Testing:
```
python test/test_cuda.py -k test_mempool_limited_memory_with_allocator
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/151487
Approved by: https://github.com/eqy, https://github.com/syed-ahmed, https://github.com/ngimel
Summary: Currently we process events in the regular allocation path and we call cudaEventQuery to check on the events and this path can take some locks in libcuda driver. Its not entirely needed to do process events in the allocation path, we could move this to a background thread and keep processing events regularly and put the freed block to the free list.
Differential Revision: D62396585
Pull Request resolved: https://github.com/pytorch/pytorch/pull/135524
Approved by: https://github.com/zyan0
Summary:
This diff adds an option to round the non-split blocks in caching allocator so that they can be reused without causing lots of fragmentation for large memory segments.
For example, if we specify max_split memory size as 400MB, then all allocations more than 400MB will not be split. Lets say, we allocated some 1024MB blocks and these are cached in the allocator blocks. If we request a new 500MB block, we round it to nearest power-2-division, thats 512MB, we add default kLargeBuffer of 20MB, that will be 532MB and since 532MB is less than existing 1024MB block, the 1024MB will not be used for this allocation, instead a new 512MB block will be created. In this diff, we provide an option to cofigure the kLargeBuffer for rounding and expose as a configurable option, so 512MB + max_non_split_rounding_size and if thats greater than 1024MB, we will use te 1024MB and we wont create a new 512MB block using cudaMalloc. This option is added so that we can pre-allocate some large blocks so that we can reuse them as much as possible and we dont stall on calling cudaMalloc.
Differential Revision: D62758758
Pull Request resolved: https://github.com/pytorch/pytorch/pull/136174
Approved by: https://github.com/zyan0
See #113541
The PR allows for registering and controlling multiple RNG states using indices, ensuring cudagraph-safe operations, and includes both C++ and Python API changes to support this functionality.
cc @eellison @anijain2305 @jansel @ezyang @ptrblck @csarofeen @mcarilli
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114068
Approved by: https://github.com/ezyang
Adding to the docs for now, hopefully we can move to `cudaMallocAsync`-backed cuBLAS workspaces soon which should alleviate the recent confusion around `cuBLAS` "leaking" memory through workspaces.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/100919
Approved by: https://github.com/ngimel
Essentially the same change as #67946, except that the default is to disallow reduced precision reductions in `BFloat16` GEMMs (for now). If performance is severely regressed, we can change the default, but this option appears to be necessary to pass some `addmm` `BFloat16` tests on H100.
CC @ptrblck @ngimel
Pull Request resolved: https://github.com/pytorch/pytorch/pull/89172
Approved by: https://github.com/ngimel
Fixes#43144
This uses the Backend system added by [82682](https://github.com/pytorch/pytorch/pull/82682) to change allocators dynamically during the code execution. This will allow us to use RMM, use CUDA managed memory for some portions of the code that do not fit in GPU memory. Write static memory allocators to reduce fragmentation while training models and improve interoperability with external DL compilers/libraries.
For example, we could have the following allocator in c++
```c++
#include <sys/types.h>
#include <cuda_runtime_api.h>
#include <iostream>
extern "C" {
void* my_malloc(ssize_t size, int device, cudaStream_t stream) {
void *ptr;
std::cout<<"alloc "<< size<<std::endl;
cudaMalloc(&ptr, size);
return ptr;
}
void my_free(void* ptr) {
std::cout<<"free "<<std::endl;
cudaFree(ptr);
}
}
```
Compile it as a shared library
```
nvcc allocator.cc -o alloc.so -shared --compiler-options '-fPIC'
```
And use it from PyTorch as follows
```python
import torch
# Init caching
# b = torch.zeros(10, device='cuda')
new_alloc = torch.cuda.memory.CUDAPluggableAllocator('alloc.so', 'my_malloc', 'my_free')
old = torch.cuda.memory.get_current_allocator()
torch.cuda.memory.change_current_allocator(new_alloc)
b = torch.zeros(10, device='cuda')
# This will error since the current allocator was already instantiated
torch.cuda.memory.change_current_allocator(old)
```
Things to discuss
- How to test this, needs compiling external code ...
Pull Request resolved: https://github.com/pytorch/pytorch/pull/86786
Approved by: https://github.com/albanD
Summary:
Improved roundup_power2_divisions knob so it allows better control of rouding in the PyTorch CUDA Caching Allocator.
This new version allows setting the number of divisions per power of two interval starting from 1MB and ending at 64GB and above. An example use case is when rouding is desirable for small allocations but there are also very large allocations which are persistent, thus would not benefit from rounding and take up extra space.
Test Plan: Tested locally
Differential Revision: D40103909
Pull Request resolved: https://github.com/pytorch/pytorch/pull/87290
Approved by: https://github.com/zdevito
Fixes#83973 (This is a substitute PR for https://github.com/pytorch/pytorch/pull/85024)
First of all, thanks for your invaluable contributions to PyTorch everyone!
Given how extensively `torch.cuda.is_available` is used in the PyTorch ecosystem, IMHO it's worthwhile to provide downstream libraries/frameworks/users the ability to alter the default behavior of `torch.cuda.is_available` in the context of their PyTorch usage.
I'm confident there are many current and future such use cases which could benefit from leveraging a weakened, NVML-based `torch.cuda.is_available` assessment at a downstream framework's explicit direction (thanks @malfet 81da50a972 !). Though one could always patch out the `torch.cuda.is_available` function with another implementation in a downstream library, I think this environmental variable based configuration option is more convenient and the cost to including the option is quite low.
As discussed in https://github.com/pytorch/pytorch/pull/85024#issuecomment-1261542045, this PR gates new non-default NVML-based CUDA behavior with an environmental variable (PYTORCH_NVML_BASED_CUDA_CHK) that allows a user/framework to invoke non-default, NVML-based `is_available()` assessments if desired.
Thanks again for your work everyone!
@ngimel @malfet @awaelchli
Pull Request resolved: https://github.com/pytorch/pytorch/pull/85951
Approved by: https://github.com/ngimel
Summary:
Added an additional roundup knob( ``roundup_bypass_threshold_mb``) to bypass rounding the requested allocation size, for allocation requests larger than the threshold value (in MB). This can help reduce the memory footprint when making large allocations that are expected to be persistent or have a large lifetime.
Differential Revision: D39868104
Pull Request resolved: https://github.com/pytorch/pytorch/pull/85940
Approved by: https://github.com/zdevito
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/74261
### Goal
Implement a cheap way to reclaim GPU memory (garbage collection) without incurring GPU sync.
### Why do we need this?
Currently, there are only two ways to reclaim GPU memory block already assigned to a particular stream.
- `release_available_cached_blocks(params)`: Free blocks exceeding the `CachingAllocatorConfig::max_split_size()` until we can satisfy the request.
Issue: If the `max_split_size` is unset (default), this function is a no-op. Even if this is set, the reclamation is quite conservative (e.g., never frees blocks under max_split_size).
- `release_cached_blocks()`: Waits for all the in-flight events and then reclaim blocks.
Issue: 'waiting for all event' is very expensive as it will likely stall all the GPU operations. Many GPU applications without a proper handling of potential GPU throttling would suffer/crash.
### Proposed idea
- If the garbage collection threshold is set, try to reclaim some memory blocks *without* synchronization. It should be safe to do so, as `release_available_cached_blocks` essentially does the same thing (but less aggressively).
- GC is triggered only when we fail to serve a `malloc` request from the block pool. No need to free blocks when the block pool is functioning just fine.
- Prioritize reclaiming blocks that weren't reused for long time. Reclamation stops once the used memory capacity < threshold.
- This code path is totally optional; by default it won't be invoked.
Test Plan:
- Unit tests
- Manually checked that the GPU memory usage stays as indicated by the garbage collector. If not the caching allocator at least tries to keep freeing the blocks.
Reviewed By: jianyuh
Differential Revision: D34482514
fbshipit-source-id: d5eae62ac60b94b0bca851f9d233a092d086e3c2
(cherry picked from commit 05780f1ed4b176f05e765b2411c9eaa2eaeb48b0)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/74213
In the current CUDACachingAllocator, the sizes are rounded up in multiple of blocks size of 512, so this works for smaller sizes. However for large sizes, we can have lots of different size blocks in the larger pool. This is problematic when we have variable batch sizes 1001, 1021, 1023 -> all will go to different block size and will create different size of blocks. This will create lots of unused blocks and will waste GPU memory capacity.
This diff adds a rounding approach to allocation size. It rounds up the size to nearest power-of-2 divisions and the power2-division can be changed with env variable setting.
For example, if we need to round-up size of1200 and if number of divisions is 4,
the size 1200 lies between 1024 and 2048 and if we do 4 divisions between
them, the values are 1024, 1280, 1536, and 1792. So the function will
return 1280 as the nearest ceiling of power-2 division.
env setting:
export PYTORCH_CUDA_ALLOC_CONF=roundup_power2_divisions:4
ghstack-source-id: 151446017
Reviewed By: ezyang
Differential Revision: D34868036
fbshipit-source-id: 494785add16e6b37c920dcb5a2b81d4c637b554a
(cherry picked from commit 548454ccacbd8700e7ffd2d762e40b4ba37abbae)
Summary:
https://github.com/pytorch/pytorch/issues/67578 disabled reduced precision reductions for FP16 GEMMs. After benchmarking, we've found that this has substantial performance impacts for common GEMM shapes (e.g., those found in popular instantiations of multiheaded-attention) on architectures such as Volta. As these performance regressions may come as a surprise to current users, this PR adds a toggle to disable reduced precision reductions
`torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = `
rather than making it the default behavior.
CC ngimel ptrblck
stas00 Note that the behavior after the previous PR can be replicated with
`torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction = False`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/67946
Reviewed By: zou3519
Differential Revision: D32289896
Pulled By: ngimel
fbshipit-source-id: a1ea2918b77e27a7d9b391e030417802a0174abe
Summary:
Powers have decided this API should be listed as beta.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/65247
Reviewed By: malfet
Differential Revision: D31057940
Pulled By: ngimel
fbshipit-source-id: 137b63cbd2c7409fecdc161a22135619bfc96bfa
Summary:
Puts memory sharing intro under Sharing memory... header, where it should have been all along.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/64996
Reviewed By: mruberry
Differential Revision: D30948619
Pulled By: ngimel
fbshipit-source-id: 5d9dd267b34e9d3fc499d4738377b58a22da1dc2
Summary:
Before https://github.com/pytorch/pytorch/pull/57833, calls to backward() or grad() synced only the calling thread's default stream with autograd leaf streams at the end of backward. This made the following weird pattern safe:
```python
with torch.cuda.stream(s):
# imagine forward used many streams, so backward leaf nodes may run on many streams
loss.backward()
# no sync
use grads
```
but a more benign-looking pattern was unsafe:
```python
with torch.cuda.stream(s):
# imagine forward used a lot of streams, so backward leaf nodes may run on many streams
loss.backward()
# backward() syncs the default stream with all the leaf streams, but does not sync s with anything,
# so counterintuitively (even though we're in the same stream context as backward()!)
# it is NOT SAFE to use grads here, and there's no easy way to make it safe,
# unless you manually sync on all the streams you used in forward,
# or move "use grads" back to default stream outside the context.
use grads
```
mruberry ngimel and I decided backward() should have the [same user-facing stream semantics as any cuda op](https://pytorch.org/docs/master/notes/cuda.html#stream-semantics-of-backward-passes).** In other words, the weird pattern should be unsafe, and the benign-looking pattern should be safe. Implementationwise, this meant backward() should sync its calling thread's current stream, not default stream, with the leaf streams.
After https://github.com/pytorch/pytorch/pull/57833, backward syncs the calling thread's current stream AND default stream with all leaf streams at the end of backward. The default stream syncs were retained for temporary backward compatibility.
This PR finishes https://github.com/pytorch/pytorch/pull/57833's work by deleting syncs on the default stream.
With this PR, graph-capturing an entire backward() call should be possible (see the [test_graph_grad_scaling diffs](https://github.com/pytorch/pytorch/compare/master...mcarilli:streaming_backwards_remove_default_syncs?expand=1#diff-893b1eea27352f336f4cd832919e48d721e4e90186e63400b8596db6b82e7450R3641-R3642)).
** first paragraph has a formatting error which this PR should also fix.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/60421
Reviewed By: albanD
Differential Revision: D29370344
Pulled By: ngimel
fbshipit-source-id: 3248bc5fb92fc517db0c15c897e5d7250f67d7fe
