Differential Revision: D63206258
This diff introduces a mechanism to generate a json-compatible deserializer in cpp using nlohmann json (already being used by AOTI).
Why we need this? Because there will be a lot of cases where people don't want to use Python to load the graph (e.g. cpp runtime), and instead they can use this header to deserialize the JSON graph.
Every time we call update_schema.py to update the schema, the header will be auto generated and included into the source files.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/136398
Approved by: https://github.com/angelayi
This PR introduces the following:
### torch.ops.symm_mem._async_input_mm
`_async_input_mm(Tensor a, Tensor b, Tensor a_chunk_signals, int a_chunk_pivot) -> Tensor`
An mm impl that supports consuming asynchronous input. It guarantees the following rasterization order, and that the corresponding signal arrives before an input chunk is consumed.
```
num_chunks = a_chunks_signals.numel()
for chunk_idx in range(a_chunk_pivot, num_chunks + a_chunk_pivot):
chunk_idx = chunk_idx % num_chunks
wait_signal(a_chunk_signals, chunk_idx)
# Compute output tiles that consumes the input chunk
```
### PersistentAsyncInputScheduler
This is a forked version of PersistentScheduler that supports consuming asynchronous input. This tile scheduler introduces the following arguments:
- `tiles_per_chunk_m` – Specifies the size of an M chunk. Chunks are the granularity at which the asynchronous input becomes ready. It must be an interger multiple of the size of an M tile.
- `chunk_signals` – `chunk_signals[i] == 1` indicates that chunk i is ready. Before returning a work tile, get_current_work() waits for the signal to ensure that the corresponding chunk is ready.
- `tile_idx_pivot_m` – After applying swizzling, apply `pivot(m) => (m + tile_idx_pivot_m) % tiles_m` to `m`. In a distributed setting, this allows different ranks to process different m indices at the same time, thus avoiding communication hotspots.
Note that this scheduler currently only supports the `KernelTmaWarpSpecializedCooperative` kernel schedule. This is enforced via the template argument `KernelSchedule`.
Usage:
```
using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
Shape<int, int, int, int>,
CollectiveMainloop,
CollectiveEpilogue,
cutlass::gemm::PersistentAsyncInputScheduler<KernelSchedule>>;
```
### _fused_all_gather_matmul_native
An ag-mm impl that combines `torch.ops.symm_mem._async_input_mm` and progress-aware all-gather. This is not yet enabled via the async-tp passes. We will use it as a backend to optimize the current decomposition-based async-tp impl.
## Benchmarks
### 4096x3584x8192
- cublas + nccl: 539us
- decomp-based async-tp w/o cuda graph: 694us
- decomp-based async-tp w/ cuda graph: 478us
- new cutlass kernel: 408us
<img width="478" alt="image" src="https://github.com/user-attachments/assets/39f316ab-36c5-4b41-af77-07854a385dfc">
### 2048x3584x8192
- cublas + nccl: 301us
- decomp-based async-tp w/o cuda graph: 687us
- decomp-based async-tp w/ cuda graph: 356us
- new cutlass kernel: 276us
<img width="441" alt="image" src="https://github.com/user-attachments/assets/9e23ce21-863b-43dd-a562-fb05d3a5a144">
## Next Steps
- Add tuning logic
- Use `_fused_all_gather_matmul_native` as a backend for the decomp-based async-tp impl
Differential temp Revision: [D65623152](https://our.internmc.facebook.com/intern/diff/D65623152)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/139227
Approved by: https://github.com/weifengpy, https://github.com/Chillee
This PR introduces the following:
### torch.ops.symm_mem._async_input_mm
`_async_input_mm(Tensor a, Tensor b, Tensor a_chunk_signals, int a_chunk_pivot) -> Tensor`
An mm impl that supports consuming asynchronous input. It guarantees the following rasterization order, and that the corresponding signal arrives before an input chunk is consumed.
```
num_chunks = a_chunks_signals.numel()
for chunk_idx in range(a_chunk_pivot, num_chunks + a_chunk_pivot):
chunk_idx = chunk_idx % num_chunks
wait_signal(a_chunk_signals, chunk_idx)
# Compute output tiles that consumes the input chunk
```
### PersistentAsyncInputScheduler
This is a forked version of PersistentScheduler that supports consuming asynchronous input. This tile scheduler introduces the following arguments:
- `tiles_per_chunk_m` – Specifies the size of an M chunk. Chunks are the granularity at which the asynchronous input becomes ready. It must be an interger multiple of the size of an M tile.
- `chunk_signals` – `chunk_signals[i] == 1` indicates that chunk i is ready. Before returning a work tile, get_current_work() waits for the signal to ensure that the corresponding chunk is ready.
- `tile_idx_pivot_m` – After applying swizzling, apply `pivot(m) => (m + tile_idx_pivot_m) % tiles_m` to `m`. In a distributed setting, this allows different ranks to process different m indices at the same time, thus avoiding communication hotspots.
Note that this scheduler currently only supports the `KernelTmaWarpSpecializedCooperative` kernel schedule. This is enforced via the template argument `KernelSchedule`.
Usage:
```
using GemmKernel = cutlass::gemm::kernel::GemmUniversal<
Shape<int, int, int, int>,
CollectiveMainloop,
CollectiveEpilogue,
cutlass::gemm::PersistentAsyncInputScheduler<KernelSchedule>>;
```
### _fused_all_gather_matmul_native
An ag-mm impl that combines `torch.ops.symm_mem._async_input_mm` and progress-aware all-gather. This is not yet enabled via the async-tp passes. We will use it as a backend to optimize the current decomposition-based async-tp impl.
## Benchmarks
### 4096x3584x8192
- cublas + nccl: 539us
- decomp-based async-tp w/o cuda graph: 694us
- decomp-based async-tp w/ cuda graph: 478us
- new cutlass kernel: 408us
<img width="478" alt="image" src="https://github.com/user-attachments/assets/39f316ab-36c5-4b41-af77-07854a385dfc">
### 2048x3584x8192
- cublas + nccl: 301us
- decomp-based async-tp w/o cuda graph: 687us
- decomp-based async-tp w/ cuda graph: 356us
- new cutlass kernel: 276us
<img width="441" alt="image" src="https://github.com/user-attachments/assets/9e23ce21-863b-43dd-a562-fb05d3a5a144">
## Next Steps
- Add tuning logic
- Use `_fused_all_gather_matmul_native` as a backend for the decomp-based async-tp impl
Pull Request resolved: https://github.com/pytorch/pytorch/pull/139227
Approved by: https://github.com/weifengpy, https://github.com/Chillee
# Motivation
According to [[RFC]A device-agnostic Python runtime API design for stream-based accelerators](https://github.com/pytorch/pytorch/issues/128403), this PR intends to introduce a device-agnostic runtime API design.
I personally prefer the **Simple Version** APIs that no longer accept the device type as an input argument. It means we will leverage `getAccelerator` to fetch the current accelerator. And it is flexible to expand these APIs to handle multiple types of accelerator scenarios. The design does **NOT** break the previous design philosophies.
I also believe that namespace torch.accelerator is better. It lets users know that the APIs they are calling are running on an accelerator rather than CPU. This is important. Meanwhile, we can follow a simple API design principle:
1. Device-agnostic APIs should be placed under the torch.accelerator namespace and not accept a device_type optional parameter.
2. Device-specific APIs should be placed under device-specific submodules.
3. APIS required by both CPU and accelerators should be placed under the torch namespace and accept a device_type optional parameter.
Also, I list the pros and cons of **Simple Version** here:
Pros:
- `torch.accelerator.foo` will have the same input argument as `torch.xxx.foo`, bringing a better user experience;
- more concise, facilitate the developer to write a device-agnostic code.
Cons:
- no obvious drawbacks.
# Additional Context
I list the new APIs here:
```python
torch.accelerator.is_available() -> bool:
torch.accelerator.current_accelerator() -> torch.device:
torch.accelerator.device_count() -> int:
torch.accelerator.current_device_idx() -> int:
torch.accelerator.set_device_idx(device: Union[torch.device, str, int, None]) -> None:
torch.accelerator.current_stream(device: Union[torch.device, str, int, None]) -> torch.Stream:
torch.accelerator.set_stream(stream: torch.Stream) -> None:
torch.accelerator.synchronize(device: Union[torch.device, str, int, None]) -> None:
```
According to the discussion with Alban, we decide to change the API name `set_device` to `set_device_idx` and `current_device` to `current_device_idx` for more explicit. And will submit other PR to support device and stream context manager.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/132204
Approved by: https://github.com/EikanWang, https://github.com/abhilash1910, https://github.com/gujinghui, https://github.com/albanD
* Added a cpp loader, AOTIModelPackageLoader, which can load the .pt2, build the .so, and create a runner. The python-facing API is that users can directly call the `run` function, whereas in cpp users can directly access the `runner_` if they are more familiar with that. I couldn't figure out how to bind the `get_runner()` function to python...
* Added a new config, `aot_inductor.package_cpp_only` which will **not** package the so. This means that whenever the package is loaded, we will need to build the so. This is turned off by default so that new environments do not need to rebuild their so. The `package_cpp_only` is a feature which torchchat intends to use to provide flexibility to users.
* Added a new config, `aot_inductor.metadata` which stores user-provided metadata, serialized to the pt2 as a json file. It also stores the device used when exporting, "cuda" or "cpu", so that during load time, we can use that data to determine which AOTIModelContainerRunner to use. The metadata can be accessed through `loader.get_metadata()`. TODO is to move this metadata to the toplevel `package_aoti` function so that we can remove the metadata as a config.
* Separated out `package_aoti` as a standalone function, instead of it automatically being called in inductor. This is to prepare for the case where users will compile multiple models, and want to bundle it in one package. The specific use case is in torchchat, where we want to package the separately-exported encoder and decoder layers. An example of how to use this is in `test_multiple_methods`.
* `load_package` will load a singular model, given the model name.
* The loader doesn't support windows for now, I think I need to add some more casing to make the build commands work on windows?
Differential Revision: [D62329906](https://our.internmc.facebook.com/intern/diff/D62329906)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/135374
Approved by: https://github.com/desertfire, https://github.com/malfet
So that the tensor's lifetime management is the same as the management built for the NCCL, pre and post kernels.
Also so that on visualizers, they show up in the NCCL stream line. Otherwise if they show up in the compute line, user may get confused (my code does not have these kernels).
The check is thus moved after the point where we depend NCCL stream from the last compute kernel.
Also moved declaration of `checkForNan` from Utils.hpp to NCCLUtils.hpp, and renamed Utils.cu to NCCLUtils.cu.
Differential Revision: [D61957573](https://our.internmc.facebook.com/intern/diff/D61957573)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134300
Approved by: https://github.com/shuqiangzhang, https://github.com/wconstab
This adds logs if we can't acquire locks in NCCLUtils and ProcessGroupNCCL for 30s.
This is motivated by some deadlocks were seeing and it's unclear if it's in NCCL or on the PyTorch side of things.
This required replacing most `std::mutex` with `std::timed_mutex` and `std::condition_variable_any` as appropriate.
Test plan:
existing CI for regressions
will add unit tests on `C10D_LOCK_GUARD`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134131
Approved by: https://github.com/c-p-i-o, https://github.com/fduwjj
So that the tensor's lifetime management is the same as the management built for the NCCL, pre and post kernels.
Also so that on visualizers, they show up in the NCCL stream line. Otherwise if they show up in the compute line, user may get confused (my code does not have these kernels).
The check is thus moved after the point where we depend NCCL stream from the last compute kernel.
Also moved declaration of `checkForNan` from Utils.hpp to NCCLUtils.hpp, and renamed Utils.cu to NCCLUtils.cu.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/134300
Approved by: https://github.com/shuqiangzhang, https://github.com/wconstab
Summary:
This PR adds in cuSPARSELt as a backend to PyTorch.
It is now possible to see if cuSPARSELt is available and the version if
it is with
```
torch.backends.cusparselt.is_available()
torch.backends.cusparselt.version()
```
Test Plan:
```
python test/test_sparse_semi_structured.py -k test_cusparselt_backend
```
Reviewers:
Subscribers:
Tasks:
Tags:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128534
Approved by: https://github.com/cpuhrsch, https://github.com/eqy, https://github.com/syed-ahmed
This PR only adds the execution of the benchmarks on this PR and print results, following diffs will add checking out head~1 and running it and comparing.
to access results goto test pr_time_benchmarks and inspect logs:
you should see
```
+ echo 'benchmark results on current PR: '
benchmark results on current PR:
+ cat /var/lib/jenkins/workspace/test/test-reports/pr_time_benchmarks_before.txt
update_hint_regression,instruction_count,27971461254
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/131475
Approved by: https://github.com/ezyang
Summary: Since WaitCounter frontend itself has minimal depdendencies it's fine to be moved into c10. Specific backends can be registered/linked separately.
Test Plan: unit test
Reviewed By: jamesperng, asiab4, c-p-i-o
Differential Revision: D59842868
Pull Request resolved: https://github.com/pytorch/pytorch/pull/131021
Approved by: https://github.com/asiab4
Summary: Since WaitCounter frontend itself has minimal depdendencies it's fine to be moved into c10. Specific backends can be registered/linked separately.
Test Plan: unit test
Reviewed By: jamesperng, asiab4, c-p-i-o
Differential Revision: D59842868
Pull Request resolved: https://github.com/pytorch/pytorch/pull/131021
Approved by: https://github.com/asiab4
We should be able to create multiple CUDAPluggableAllocators in the same pytorch program (see https://github.com/pytorch/pytorch/issues/124807, https://github.com/pytorch/pytorch/pull/125722 for context). When mixing CUDAPluggableAllocators in the same pytorch program, we need to make sure that the deleter passed in through the CUDAPluggableAllocator gets "attached" to the data_ptr and persist until program exit (when it's called to free the memory).
Currently, CUDAPluggableAllocator maintains a global `current_custom_allocator`. When creating the `DataPtr`, `raw_deleter` attaches `custom_raw_deleter` to the DataPtr which calls `current_custom_allocator->raw_delete(...)`. This approach is fine when using only one allocator, however for multiple allocator use case, DataPtr would be using the deleter of whatever is in the `current_custom_allocator`. For example, if allocation 1 was done with `cudaMalloc` and allocation 2 was done with `ncclMemAlloc`, and if `current_custom_allocator` is currently pointing to the CUDAPluggableAllocator with `ncclMemAlloc` - when cleaning up the allocation 1, we'd be using `ncclMemFree` instead of `cudaFree`.
In this PR, we solve the above problem by remembering the `free_fn_` using a deleter context. Hence, there is no need to go through an allocator object to find the deleter.
CC: @zdevito @ptrblck @eqy
Pull Request resolved: https://github.com/pytorch/pytorch/pull/130472
Approved by: https://github.com/eqy, https://github.com/ezyang
Construct frame localsplus in 3.12+ using our own simplified way rather than copypasting from CPython.
This is necessary for 3.13 since we can no longer generate frame `f_locals` before executing the interpreter frame.
We also enable this for 3.12 since the `f_locals` construction between 3.12 and 3.13 is the same, so we can test for correctness with 3.12.
This is also one of the first steps to completing https://github.com/pytorch/pytorch/issues/93753 - we will implement simplified f_locals generation of previous Python versions in the future.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/129185
Approved by: https://github.com/jansel
Construct frame localsplus in 3.12+ using our own simplified way rather than copypasting from CPython.
This is necessary for 3.13 since we can no longer generate frame `f_locals` before executing the interpreter frame.
We also enable this for 3.12 since the `f_locals` construction between 3.12 and 3.13 is the same, so we can test for correctness with 3.12.
This is also one of the first steps to completing https://github.com/pytorch/pytorch/issues/93753 - we will implement simplified f_locals generation of previous Python versions in the future.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/129185
Approved by: https://github.com/jansel
This PR introduces `_detect_dma_connectivity` - a utility for detecting DMA connectivity among devices.
The "DMA connectivity" in this context is more stringent than the ability to perform memory copy without CPU involvement. We define it as the ability for a device to issue load/store instructions and perform atomic operations on memory that resides on connected devices. The ability translates to the ability to run most aten GPU operations with operands backed by remote memory. `_detect_dma_connectivity` can help PyTorch and its users to determine whether certain DMA-based optimizations are possible.
`_detect_dma_connectivity` takes a `(device_type, connection_type)` pair and returns a matrix describing the connectivity. Connectivity detectors are statically registered on a `(device_type, connection_type)` basis. This PR implements the detector for `(CUDA, "nvlink")`. Later, detectors for pairs such as `(ROCM, "infinity_fabric")` can be introduced.
Example:
```python3
>>> from torch._C._autograd import DeviceType
>>> from torch._C._distributed_c10d import _detect_dma_connectivity
>>> connectivity = _detect_dma_connectivity(DeviceType.CUDA, "nvlink")
>>> for row in connectivity.matrix:
... print(row)
...
[0, 18, 18, 18, 18, 18, 18, 18]
[18, 0, 18, 18, 18, 18, 18, 18]
[18, 18, 0, 18, 18, 18, 18, 18]
[18, 18, 18, 0, 18, 18, 18, 18]
[18, 18, 18, 18, 0, 18, 18, 18]
[18, 18, 18, 18, 18, 0, 18, 18]
[18, 18, 18, 18, 18, 18, 0, 18]
[18, 18, 18, 18, 18, 18, 18, 0]
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/129510
Approved by: https://github.com/weifengpy
This PR removes `ProcessGroupCudaP2P` and changes async-TP to use `SymmetricMemory`. The async-TP implementation is still workspace-based, but it now doesn't require a buffer size to be specified upfront.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128762
Approved by: https://github.com/wanchaol
We've been facing issues where TCPStore can successfully connect but then fail in the validate() function due to resets from listen backlog queue overflow when combined with reset enabled as well as long init times.
This PR does a few things:
* Retry that connect and validate up to the specified timeout.
* Use exponential backoff for the retry logic with jitter instead of a fixed 1s sleep.
* Eliminate the `sleep(std::chrono::milliseconds(numWorkers))` on init which can add significant delays to startup. This is no longer necessary per @XilunWu https://github.com/pytorch/pytorch/pull/116141
Test plan:
```
python test/distributed/test_store.py -v
./build/bin/BackoffTest
```
Will do internal testing with some large scale jobs to ensure TCPStore works correctly.
At 4k scale: 4x improvement
```
tristanr@devvm4382 ~/pt_tests [SIGABRT]> time TORCH_SHOW_CPP_STACKTRACES=1 python tcpstore_large_test.py (pytorch-3.10)
started 0
init 0
set 0
joined all
________________________________________________________
Executed in 1.98 secs fish external
usr time 0.93 secs 91.00 micros 0.93 secs
sys time 1.98 secs 954.00 micros 1.97 secs
tristanr@devvm4382 ~/pt_tests> conda activate torchdrive-3.10 (pytorch-3.10)
tristanr@devvm4382 ~/pt_tests> time TORCH_SHOW_CPP_STACKTRACES=1 python tcpstore_large_test.py (torchdrive-3.10)
started 0
init 0
set 0
joined all
________________________________________________________
Executed in 8.20 secs fish external
usr time 2.15 secs 0.00 micros 2.15 secs
sys time 2.76 secs 843.00 micros 2.76 secs
```
```py
import time
import os
import threading
from multiprocessing import Pool
WORLD_SIZE = 10000
import torch.distributed as dist
def run(rank):
should_log = rank % (WORLD_SIZE // 10) == 0
if should_log:
print(f"started {rank}")
store = dist.TCPStore(
host_name="devvm4382.nao0.facebook.com",
port=29500,
world_size=WORLD_SIZE,
is_master=rank == 0,
use_libuv=True,
)
if should_log:
print(f"init {rank}")
store.set(f"key{rank}", "1234")
if should_log:
print(f"set {rank}")
del store
def noop(rank):
pass
print("starting pool")
with Pool(WORLD_SIZE) as pool:
pool.map(noop, range(WORLD_SIZE), 1)
print("pool hot")
start = time.time()
pool.map(run, range(WORLD_SIZE), 1)
print("run finished", time.time()-start)
```
```
tristanr@devvm4382 ~/pt_tests> python tcpstore_large_test.py (pytorch-3.10)
starting pool
pool hot
started 0
[W624 16:58:09.086081750 TCPStore.cpp:343] [c10d] Starting store with 10000 workers but somaxconn is 4096.This might cause instability during bootstrap, consider increasing it.
started 1000
init 1000
set 1000
started 2000
init 2000
set 2000
started 3000
init 3000
set 3000
started 4000
init 4000
set 4000
started 5000
init 5000
set 5000
started 6000
init 6000
set 6000
started 7000
init 7000
set 7000
started 8000
init 8000
set 8000
started 9000
init 9000
set 9000
init 0
set 0
run finished 0.705092191696167
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/129261
Approved by: https://github.com/rsdcastro, https://github.com/wconstab, https://github.com/kurman, https://github.com/XilunWu, https://github.com/c-p-i-o
Currently if `x` is a CUDA tensor, calling `x.untyped_storage().resize_()` seems to always go into the `built without cuda` branch of `resize_storage_bytes_()` regardless of whether PyTorch is built with CUDA. I suspect this is because `inductor_ops.cpp` is only included in `libtorch_cpu.so` thus doesn't have the `USE_CUDA` information or ability to link to CUDA-related functions.
This PR moves `resize_storage_bytes_()` related custom op functions out of `inductor_ops.cpp` into its standalone file `resize_storage_bytes.cpp` to be included in `libtorch_python.so` instead. This mimics the setup for `StorageMethods.cpp`. This way, `resize_storage_bytes_()` can have access to the CUDA-related functions, which passes the CUDA unit test.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/129215
Approved by: https://github.com/jansel
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
This PR introduces a prototype for `SymmetricMemory` (including a CUDA implementation) - a remote-memory access-based communication primitive. It allows for user-defined communication patterns/kernels and is designed to be torch.compile-friendly. It addresses the major limitations of `IntraNodeComm` and `ProcessGroupCudaP2p` and serves as a replacement for them.
### SymmetricMemory
`SymmetricMemory` represents symmetric allocations across a group of devices. The allocations represented by a `SymmetricMemory` object are accessible by all devices in the group. The class can be used for **op-level custom communication patterns** (via the get_buffer APIs and the synchronization primitives), as well as **custom communication kernels** (via the buffer and signal_pad device pointers).
### Python API Example
```python
from torch._C.distributed_c10d import _SymmetricMemory
# Set a store for rendezvousing symmetric allocations on a group of devices
# identified by group_name. The concept of groups is logical; users can
# utilize predefined groups (e.g., a group of device identified by a
# ProcessGroup) or create custom ones. Note that a SymmetricMemoryAllocator
# backends might employ a more efficient communication channel for the actual
# rendezvous process and only use the store for bootstrapping purposes.
_SymmetricMemory.set_group_info(group_name, rank, world_size, store)
# Identical to empty_strided, but allows symmetric memory access to be
# established for the allocated tensor via _SymmetricMemory.rendezvous().
# This function itself is not a collective operation.
t = _SymmetricMemory.empty_strided_p2p((64, 64), (64, 1), torch.float32, group_name)
# Users can write Python custom ops that leverages the symmetric memory access.
# Below are examples of things users can do (assuming the group's world_size is 2).
# Establishes symmetric memory access on tensors allocated via
# _SymmetricMemory.empty_strided_p2p(). rendezvous() is a one-time process,
# and the mapping between a local memory region and the associated SymmetricMemory
# object is unique. Subsequent calls to rendezvous() with the same tensor will receive
# the cached SymmetricMemory object.
#
# The function has a collective semantic and must be invoked simultaneously
# from all rendezvous participants.
symm_mem = _SymmetricMemory.rendezvous(t)
# This represents the allocation on rank 0 and is accessible from all devices.
buf = symm_mem.get_buffer(0, (64, 64), torch.float32)
if symm_mem.rank == 0:
symm_mem.wait_signal(src_rank=1)
assert buf.eq(42).all()
else:
# The remote buffer can be used as a regular tensor
buf.fill_(42)
symm_mem.put_signal(dst_rank=0)
symm_mem.barrier()
if symm_mem.rank == 0:
symm_mem.barrier()
assert buf.eq(43).all()
else:
new_val = torch.empty_like(buf)
new_val.fill_(43)
# Contiguous copies to/from a remote buffer utilize copy engines
# which bypasses SMs (i.e. no need to load the data into registers)
buf.copy_(new_val)
symm_mem.barrier()
```
### Custom CUDA Comm Kernels
Given a tensor, users can access the associated `SymmetricMemory` which provides pointer to remote buffers/signal_pads needed for custom communication kernels.
```cpp
TORCH_API c10::intrusive_ptr<SymmetricMemory> get_symmetric_memory(
const at::Tensor& tensor);
class TORCH_API SymmetricMemory : public c10::intrusive_ptr_target {
public:
...
virtual std::vector<void*> get_buffer_ptrs() = 0;
virtual std::vector<void*> get_signal_pad_ptrs() = 0;
virtual void** get_buffer_ptrs_dev() = 0;
virtual void** get_signal_pad_ptrs_dev() = 0;
virtual size_t get_buffer_size() = 0;
virtual size_t get_signal_pad_size() = 0;
virtual int get_rank() = 0;
virtual int get_world_size() = 0;
...
};
```
### Limitations of IntraNodeComm and ProcessGroupCudaP2p
Both `IntraNodeComm` (used by `ProcessGroupCudaP2p`) manages a single fixed-size workspace. This approach:
- Leads to awkward UX in which the required workspace needs to be specified upfront.
- Can not avoid extra copies for some algorithms in eager mode (e.g., custom/multimem all-reduce, reduce-scatter, all-gather).
- Prevents torch.compile from eliminating all copies.
In addition, they only offer out-of-the-box communication kernels and don't expose required pointers for user-defined, custom CUDA comm kernels.
* __->__ #128582
Differential Revision: [D58849033](https://our.internmc.facebook.com/intern/diff/D58849033)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128582
Approved by: https://github.com/wanchaol
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
This PR introduces a prototype for `SymmetricMemory` (including a CUDA implementation) - a remote-memory access-based communication primitive. It allows for user-defined communication patterns/kernels and is designed to be torch.compile-friendly. It addresses the major limitations of `IntraNodeComm` and `ProcessGroupCudaP2p` and serves as a replacement for them.
### SymmetricMemory
`SymmetricMemory` represents symmetric allocations across a group of devices. The allocations represented by a `SymmetricMemory` object are accessible by all devices in the group. The class can be used for **op-level custom communication patterns** (via the get_buffer APIs and the synchronization primitives), as well as **custom communication kernels** (via the buffer and signal_pad device pointers).
### Python API Example
```python
from torch._C.distributed_c10d import _SymmetricMemory
# Set a store for rendezvousing symmetric allocations on a group of devices
# identified by group_name. The concept of groups is logical; users can
# utilize predefined groups (e.g., a group of device identified by a
# ProcessGroup) or create custom ones. Note that a SymmetricMemoryAllocator
# backends might employ a more efficient communication channel for the actual
# rendezvous process and only use the store for bootstrapping purposes.
_SymmetricMemory.set_group_info(group_name, rank, world_size, store)
# Identical to empty_strided, but allows symmetric memory access to be
# established for the allocated tensor via _SymmetricMemory.rendezvous().
# This function itself is not a collective operation.
t = _SymmetricMemory.empty_strided_p2p((64, 64), (64, 1), torch.float32, group_name)
# Users can write Python custom ops that leverages the symmetric memory access.
# Below are examples of things users can do (assuming the group's world_size is 2).
# Establishes symmetric memory access on tensors allocated via
# _SymmetricMemory.empty_strided_p2p(). rendezvous() is a one-time process,
# and the mapping between a local memory region and the associated SymmetricMemory
# object is unique. Subsequent calls to rendezvous() with the same tensor will receive
# the cached SymmetricMemory object.
#
# The function has a collective semantic and must be invoked simultaneously
# from all rendezvous participants.
symm_mem = _SymmetricMemory.rendezvous(t)
# This represents the allocation on rank 0 and is accessible from all devices.
buf = symm_mem.get_buffer(0, (64, 64), torch.float32)
if symm_mem.rank == 0:
symm_mem.wait_signal(src_rank=1)
assert buf.eq(42).all()
else:
# The remote buffer can be used as a regular tensor
buf.fill_(42)
symm_mem.put_signal(dst_rank=0)
symm_mem.barrier()
if symm_mem.rank == 0:
symm_mem.barrier()
assert buf.eq(43).all()
else:
new_val = torch.empty_like(buf)
new_val.fill_(43)
# Contiguous copies to/from a remote buffer utilize copy engines
# which bypasses SMs (i.e. no need to load the data into registers)
buf.copy_(new_val)
symm_mem.barrier()
```
### Custom CUDA Comm Kernels
Given a tensor, users can access the associated `SymmetricMemory` which provides pointer to remote buffers/signal_pads needed for custom communication kernels.
```cpp
TORCH_API c10::intrusive_ptr<SymmetricMemory> get_symmetric_memory(
const at::Tensor& tensor);
class TORCH_API SymmetricMemory : public c10::intrusive_ptr_target {
public:
...
virtual std::vector<void*> get_buffer_ptrs() = 0;
virtual std::vector<void*> get_signal_pad_ptrs() = 0;
virtual void** get_buffer_ptrs_dev() = 0;
virtual void** get_signal_pad_ptrs_dev() = 0;
virtual size_t get_buffer_size() = 0;
virtual size_t get_signal_pad_size() = 0;
virtual int get_rank() = 0;
virtual int get_world_size() = 0;
...
};
```
### Limitations of IntraNodeComm and ProcessGroupCudaP2p
Both `IntraNodeComm` (used by `ProcessGroupCudaP2p`) manages a single fixed-size workspace. This approach:
- Leads to awkward UX in which the required workspace needs to be specified upfront.
- Can not avoid extra copies for some algorithms in eager mode (e.g., custom/multimem all-reduce, reduce-scatter, all-gather).
- Prevents torch.compile from eliminating all copies.
In addition, they only offer out-of-the-box communication kernels and don't expose required pointers for user-defined, custom CUDA comm kernels.
* __->__ #128582
Pull Request resolved: https://github.com/pytorch/pytorch/pull/128582
Approved by: https://github.com/wanchaol
