Summary:
We need to register all cache segments allocated by allocator, so that NCCL can apply zero copy algorithms at collective and point-to-point operations.
How to track and register all cache segments:
- It registers a register and a deregister hook to cache allocator as action tracker callbacks, tracking SEGMENT_ALLOC and SEGMENT_FREE trace entries, respectively. When SEGMENT_ALLOC is tracked, the register hook will register to the PG's communicators on the same device. Similarly, when SEGMENT_FREE is tracked, the deregister hook handles deregistration before cudaFree.
- When a new NCCL communicator is created, it dumps the snapspot from cache allocator to register all existing cache segments at once.
- When a NCCL communicator is aborted, it deregisters all segments that have been registered by this communicator
Test Plan: See test in D50726971
Reviewed By: wconstab
Differential Revision: D50726970
Pull Request resolved: https://github.com/pytorch/pytorch/pull/112850
Approved by: https://github.com/wconstab
NCCL 2.17+ introduces some user configurable parameters for NCCL communicators using [ncclConfig_t](https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/api/types.html#c.ncclConfig_t) datatype and [ncclCommInitRankConfig](https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/api/comms.html#ncclcomminitrankconfig). This PR enables that feature.
A user can tune the parameters as follows:
```
import torch.distributed as dist
nccl_options = dist.ProcessGroupNCCL.Options()
nccl_options.config.max_ctas = 32
nccl_options.config.min_ctas = 8
nccl_options.config.cga_cluster_size = 2
dist.init_process_group(backend='nccl', init_method='env://', pg_options=nccl_options)
my_group = dist.new_group(pg_options=nccl_options)
```
The default values of these parameters are what is initialized by `NCCL_CONFIG_INITIALIZER`. Only for DistributedDataParallel, this PR sets the default value of cga_cluster_size to 2 (a heuristic that works well especially for DDP workloads).
Tuning these parameters can lead to improvement in end-to-end performance, since it affects the communication-computation overlap for NCCL kernels.
CC: @ptrblck @kwen2501
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97394
Approved by: https://github.com/kwen2501
Support for nonblocking NCCL communicators/fault tolerance/checking which was added in 2.14 as an experimental feature.
Enabled via the environment variable:
```
TORCH_NCCL_USE_COMM_NONBLOCKING=1
```
CC @ptrblck
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95715
Approved by: https://github.com/kwen2501
Support for nonblocking NCCL communicators/fault tolerance/checking which was added in 2.14 as an experimental feature.
Enabled via the environment variable:
```
TORCH_NCCL_USE_COMM_NONBLOCKING=1
```
CC @ptrblck
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95715
Approved by: https://github.com/kwen2501
Currently all of the distributed errors are thrown from the `TORCH_CHECK` macro which throws a generic `RuntimeError`. This change introduced a new error type `DistBackendError` which derives from `RuntimeError` to signify there was an error with the backend communication library. This allows for better error handling and analysis at higher levels in the stack. Motivation: https://docs.google.com/document/d/1j6VPOkC6znscliFuiDWMuMV1_fH4Abgdq7TCHMcXai4/edit#heading=h.a9rc38misyx8
Changes:
- introduce new error type
- Update `C10D_NCCL_CHECK`
Sample script to demonstrate new error type
```python
# python -m torch.distributed.run --nproc_per_node=2 <script>.py
import torch
import torch.distributed as dist
if __name__ == "__main__":
dist.init_process_group("nccl")
dist.broadcast(torch.tensor([1, 2, 3]).cuda(), 0)
```
Differential Revision: [D40998803](https://our.internmc.facebook.com/intern/diff/D40998803)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/88134
Approved by: https://github.com/rohan-varma
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/73897
add a debug hint that async system error can be caused by unexpected exit of
a remote process if not an actual network issue. For example, the exit of the remote process
can cause a closed network connection error at a local process. The hint helps to direct
the debug focus to the remote process.
Test Plan: unit tests
Reviewed By: pritamdamania87, rohan-varma
Differential Revision: D34702348
fbshipit-source-id: d19f9116e9efe5f6d76c0158a7a447616437ca69
(cherry picked from commit 005e74b7b6764ecd832b3410063285bff2411b56)
Summary:
Patch bfloat16 support in NCCL, PR https://github.com/pytorch/pytorch/issues/63260 adds bfloat16 support but is
still not complete to enable bfloat16 for allreduce in end-to-end training.
This patch does the followings:
* fix minimum NCCL version from 2.9.7 to 2.10, NCCL adds bf16 support in
v2.10.3-1 (commit 7e51592)
* update bfloat16 datatype flag in `csrc/cuda/nccl.cpp` so that NCCL
operations like all reduce can use it
* enable unit tests for bfloat16 datatype if possible
cc pietern mrshenli pritamdamania87 zhaojuanmao satgera rohan-varma gqchen aazzolini osalpekar jiayisuse SciPioneer H-Huang
Pull Request resolved: https://github.com/pytorch/pytorch/pull/67843
Reviewed By: H-Huang
Differential Revision: D32248132
Pulled By: mrshenli
fbshipit-source-id: 081e96e725af3b933dd65ec157c5ad11c6873525
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/65610
- Replace HIP_PLATFORM_HCC with USE_ROCM
- Dont rely on CUDA_VERSION or HIP_VERSION and use USE_ROCM and ROCM_VERSION.
- In the next PR
- Will be removing the mapping from CUDA_VERSION to HIP_VERSION and CUDA to HIP in hipify.
- HIP_PLATFORM_HCC is deprecated, so will add HIP_PLATFORM_AMD to support HIP host code compilation on gcc.
cc jeffdaily sunway513 jithunnair-amd ROCmSupport amathews-amd
Reviewed By: jbschlosser
Differential Revision: D30909053
Pulled By: ezyang
fbshipit-source-id: 224a966ebf1aaec79beccbbd686fdf3d49267e06
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/64241
When things go wrong PG NCCL aborts nccl communicators via `ncclCommAbort`, but one issues is that often the error can be set to `ncclSystemError` (see https://github.com/pytorch/pytorch/blob/master/torch/csrc/distributed/c10d/NCCLUtils.hpp#L176) when that might not be the true cause of the issue and the actual issue is that some prior work timed out, communicator was aborted on other rank, etc.
This results in a lot of confusion when debugging jobs with a large no. of processes as the current message for ncclSystemError is not very informative: https://github.com/pytorch/pytorch/blob/master/torch/csrc/distributed/c10d/NCCLUtils.hpp#L22
The fix here is to pass in a string exception message from PG NCCL down to `NCCLUtils` which will aim to raise that as the actual issue and not the confusing `ncclSystemError` message.
Test Plan: CI
Reviewed By: pallab-zz, cbalioglu
Differential Revision: D30658855
fbshipit-source-id: 17661dbe0a1bb8cc5b87b637c47634b1f52f54e1
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/63260
Add BF16 all-reduce communication hook. Skip if CUDA version < 11 or NCCL version < 2.9.7.
Reviewed By: SciPioneer
Differential Revision: D30238317
fbshipit-source-id: bad35bf7d43f10f1c40997a282b831b61ef592bb
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/60543
Since now c10d is part of libtorch, it would also be nice if the sources lived all in one place.
ghstack-source-id: 132306292
Test Plan: It builds
Reviewed By: cbalioglu
Differential Revision: D29062002
fbshipit-source-id: d9e1301e9d73e1643fa0f0119cd2d618f1ad52e6
