This patch addresses the major limitations in our previous [PR #115981](https://github.com/pytorch/pytorch/pull/115981) through the new dedicated repository [AOTriton](https://github.com/ROCm/aotriton)
- [x] Only supports MI200 series GPU (i.e., `gcnArchName == gfx90a:sramecc+:xnack-`).
* MI300X is supported. More architectures will be added once Triton support them.
- [x] Only supports power of two sequence lengths.
* Now it support arbitrary sequence length
- [ ] No support for varlen APIs.
* varlen API will be supported in the next release of AOTriton
- [x] Only support head dimension 16,32,64,128.
* Now it support arbitrary head dimension <= 256
- [x] Performance is still being optimized.
* Kernel is selected according to autotune information from Triton.
Other improvements from AOTriton include
* Allow more flexible Tensor storage layout
* More flexible API
This is a more extensive fix to #112997
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121561
Approved by: https://github.com/malfet, https://github.com/atalman
Note about the Updates:
This PR:
1. skips more flash attention related UTs on MI200
2. Fix additional ATen compiling errors after hipification
3. Fix the author "root" of a specific commit
4. Includes the patch from Nikita in favor of block level static initialization.
CAVEAT: This revised PR has a commit that modifies the CI to force its running on MI200 nodes. That specific commit must be reverted before merge.
Original PR (https://github.com/pytorch/pytorch/pull/114309) Note:
This pull requests add initial Flash Attention support for AMD/ROCM platform. It added a specialized Triton repository/branch as a compile-time dependency for Flash Attention math library on AMD/ROCM. This triton submodule is not used at runtime and will not be shipped to the final pytorch package. We have the plan to release this specialized Triton as a separate project.
Know limitations:
- Only supports MI200 series GPU (i.e., `gcnArchName == gfx90a:sramecc+:xnack-`.
- Only supports power of two sequence lengths.
- No support for varlen APIs.
- Only support head dimension 16,32,64,128.
- Performance is still being optimized.
Fixes#112997
Pull Request resolved: https://github.com/pytorch/pytorch/pull/115981
Approved by: https://github.com/malfet
This pull requests add initial Flash Attention support for AMD/ROCM platform. It added a specialized Triton repository/branch as a compile-time dependency for Flash Attention math library on AMD/ROCM. This triton submodule is not used at runtime and will not be shipped to the final pytorch package. We have the plan to release this specialized Triton as a separate project.
Know limitations:
- [ ] Only supports MI200 series GPU (i.e., `gcnArchName == gfx90a:sramecc+:xnack-`.
- [ ] Only supports power of two sequence lengths.
- [ ] No support for varlen APIs.
- [ ] Only support head dimension 16,32,64,128.
- [ ] Performance is still being optimized.
Fixes https://github.com/pytorch/pytorch/issues/112997
Pull Request resolved: https://github.com/pytorch/pytorch/pull/114309
Approved by: https://github.com/jeffdaily, https://github.com/malfet
---------
Co-authored-by: Joseph Groenenboom <joseph.groenenboom@amd.com>
Should fix#13362 and fix#83790
I think I've discovered the root cause of the intermittent nccl link
failures. If we look at the variable name in the redefinition error:
```
_02021d91_11_sendrecv_cu_0bc7b9c8_11152
```
this is the name of the file being compiled + some form of unique ID.
As part of NCCL's build process, the same file is compiled multiple
times with different macro definitions depending on which operator and
dtype are being compiled, e.g.
```
nvcc -DNCCL_OP=0 -DNCCL_TYPE=0 -dc sendrecv.cu -o sendrecv_sum_i8.o
```
Since the filename parts are the same, then if the unique IDs also
happen to collide then the entire identifier will collide and the link
fails. So the fix here is to generate a unique `.cu` file for each
object file. I've implemented this as a `.patch` file that gets
applied from our cmake code, but if we instead fork nccl that would be
cleaner.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/84245
Approved by: https://github.com/janeyx99, https://github.com/malfet
Since #83173 was merged I have noticed some CI being slowed down by
the nccl building step. e.g. if there are no C++ changes then sccache
compiles everything else very quickly and nccl becomes the limiting
factor.
This re-enables parallel builds with some safeguards to protect
against oversubscription. When `make` is the parent build system, we
can use `$(MAKE)` and the `make` jobserver will coordinate job
allocation with the sub-process. For other build systems, this calls
`make` with the `-l` flag which should prevent it launching jobs when
the system load average is already too high.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/83696
Approved by: https://github.com/malfet
- Modifies the current cmake build definitions to use `find_package` to find UCX and UCC installed in the system
- Install UCX and UCC in CUDA dockers
- Build PyTorch with `USE_UCC=1` in pipelines
- Currently, we are not running unit tests with the UCC PG. Those tests will be added in future PRs.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/81583
Approved by: https://github.com/vtlam, https://github.com/malfet
By extending regex to match any character other than not just version
On Ubuntu version string looks as follows:
```
$ objcopy --version
GNU objcopy (GNU Binutils for Ubuntu) 2.30
```
And on some CentOSes it looks as
```
$ objcopy --version
GNU objcopy (GNU Binutils) 2.37
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/82774
Approved by: https://github.com/ngimel
Summary:
This diff integrates UCC process group as a native component of Pytorch Distributed core. It is based on the existing torch-ucc (https://github.com/facebookresearch/torch_ucc) as the wrapper for UCC collective communication library.
The environment and cmake variables are named in mirroring to the existing process groups such as NCCL and Gloo. Specifically,
- USE_UCC: enables UCC PG. This defaults to OFF, so there is no breakage of existing builds that do not have UCX/UCC external libraries.
- USE_SYSTEM_UCC: uses external UCX and UCC shared libraries that are set accordingly with UCX_HOME and UCC_HOME.
Currently, this diff only supports USE_SYSTEM_UCC=ON, i.e., requiring users to specify external libraries for UCX and UCC. In subsequent diffs, we will add UCX and UCC repos as third-party dependencies in pytorch/third-party.
Test Plan:
Passed Torch-UCC tests that invoke UCC process group. For example:
$ sh test/start_test.sh test/torch_allreduce_test.py --backend gloo --use-cuda
...
Test allreduce: succeeded
Differential Revision: D36973688
Pull Request resolved: https://github.com/pytorch/pytorch/pull/79918
Approved by: https://github.com/kwen2501, https://github.com/kingchc
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/55814
I don't really know if the original issue is resolved but let's just
check and see if this passes CI so that we can potentially get some
speed up on our builds
Signed-off-by: Eli Uriegas <eliuriegas@fb.com>
Test Plan: Imported from OSS
Reviewed By: walterddr
Differential Revision: D27715734
Pulled By: seemethere
fbshipit-source-id: a8f90774dfd25b0abf8e57283fe3591a8d8f3c4b
Summary:
Context: https://github.com/pytorch/pytorch/pull/53299#discussion_r587882857
These are the only hand-written parts of this diff:
- the addition to `.github/workflows/lint.yml`
- the file endings changed in these four files (to appease FB-internal land-blocking lints):
- `GLOSSARY.md`
- `aten/src/ATen/core/op_registration/README.md`
- `scripts/README.md`
- `torch/csrc/jit/codegen/fuser/README.md`
The rest was generated by running this command (on macOS):
```
git grep -I -l ' $' -- . ':(exclude)**/contrib/**' ':(exclude)third_party' | xargs gsed -i 's/ *$//'
```
I looked over the auto-generated changes and didn't see anything that looked problematic.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/53406
Test Plan:
This run (after adding the lint but before removing existing trailing spaces) failed:
- https://github.com/pytorch/pytorch/runs/2043032377
This run (on the tip of this PR) succeeded:
- https://github.com/pytorch/pytorch/runs/2043296348
Reviewed By: walterddr, seemethere
Differential Revision: D26856620
Pulled By: samestep
fbshipit-source-id: 3f0de7f7c2e4b0f1c089eac9b5085a58dd7e0d97
Summary:
Solves `the '-j' option requires a positive integer argument` error on some systems when MAX_JOBS is not defined
Pull Request resolved: https://github.com/pytorch/pytorch/pull/44557
Reviewed By: vkuzo
Differential Revision: D23653511
Pulled By: malfet
fbshipit-source-id: 7d86fb7fb6c946c34afdc81bf2c3168a74d00a1f
Summary:
Since NCCL makes calls to shm_open/shm_close it must depend on librt on Linux
This should fix `DSO missing from command line` error on some platforms
Pull Request resolved: https://github.com/pytorch/pytorch/pull/41978
Reviewed By: colesbury
Differential Revision: D22721430
Pulled By: malfet
fbshipit-source-id: d2ae08ce9da3979daaae599e677d5e4519b080f0
Summary:
This re-applies D21232894 (b9d3869df3) and D22162524, plus updates jni_deps in a few places
to avoid breaking host JNI tests.
Test Plan: `buck test @//fbandroid/mode/server //fbandroid/instrumentation_tests/com/facebook/caffe2:host-test`
Reviewed By: xcheng16
Differential Revision: D22199952
fbshipit-source-id: df13eef39c01738637ae8cf7f581d6ccc88d37d5
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/37243
*** Why ***
As it stands, we have two thread pool solutions concurrently in use in PyTorch mobile: (1) the open source pthreadpool library under third_party, and (2) Caffe2's implementation of pthreadpool under caffe2/utils/threadpool. Since the primary use-case of the latter has been to act as a drop-in replacement for the third party version so as to enable integration and usage from within NNPACK and QNNPACK, Caffe2's implementation is intentionally written to the exact same interface as the third party version.
The original argument in favor of C2's implementation has been improved performance as a result of using spin locks, as opposed to relinquishing the thread's time slot and putting it to sleep - a less expensive operation up to a point. That seems to have given C2's implementation the upper hand in performance, hence justifying the added maintenance complexity, until the third party version improved in parallel surpassing the efficiency of C2's implementation as I have verified in benchmarks. With that advantage gone, there is no reason to continue using C2's implementation in PyTorch mobile either from the perspective of performance or code hygiene. As a matter of fact, there is considerable performance benefit to be had as a result of using the third party version as it currently stands.
This is a tricky change though, mainly because in order to avoid potential performance regressions, of which I have witnessed none but just in abundance of caution, we have decided to continue using the internal C2's implementation whenever building for Caffe2. Again, this is mainly to avoid potential performance regressions in production C2 use cases even if doing so results in reduced performance as far as I can tell.
So to summarize, today, and as it currently stands, we are using C2's implementation for (1) NNPACK, (2) PyTorch QNNPACK, and (3) ATen parallel_for on mobile builds, while using the third party version of pthreadpool for XNNPACK as XNNPACK does not provide any build options to link against an external implementation unlike NNPACK and QNNPACK do.
The goal of this PR then, is to unify all usage on mobile to the third party implementation both for improved performance and better code hygiene. This applies to PyTorch's use of NNPACK, QNNPACK, XNNPACK, and mobile's implementation of ATen parallel_for, all getting routed to the
exact same third party implementation in this PR.
Considering that NNPACK, QNNPACK, and XNNPACK are not mobile specific, these benefits carry over to non-mobile builds of PyTorch (but not Caffe2) as well. The implementation of ATen parallel_for on non-mobile builds remains unchanged.
*** How ***
This is where things get tricky.
A good deal of the build system complexity in this PR arises from our desire to maintain C2's implementation intact for C2's use.
pthreadpool is a C library with no concept of namespaces, which means two copies of the library cannot exist in the same binary or symbol collision will occur violating ODR. This means that somehow, and based on some condition, we must decide on the choice of a pthreadpool implementation. In practice, this has become more complicated as a result of all the possible combinations that USE_NNPACK, USE_QNNPACK, USE_PYTORCH_QNNPACK, USE_XNNPACK, USE_SYSTEM_XNNPACK, USE_SYSTEM_PTHREADPOOL and other variables can result in. Having said that, I have done my best in this PR to surgically cut through this complexity in a way that minimizes the side effects, considering the significance of the performance we are leaving on the table, yet, as a result of this combinatorial explosion explained above I cannot guarantee that every single combination will work as expected on the first try. I am heavily relying on CI to find any issues as local testing can only go that far.
Having said that, this PR provides a simple non mobile-specific C++ thread pool implementation on top of pthreadpool, namely caffe2::PThreadPool that automatically routes to C2's implementation or the third party version depending on the build configuration. This simplifies the logic at the cost of pushing the complexity to the build scripts. From there on, this thread pool is used in aten parallel_for, and NNPACK and family, again, routing all usage of threading to C2 or third party pthreadpool depending on the build configuration.
When it is all said or done, the layering will look like this:
a) aten::parallel_for, uses
b) caffe2::PThreadPool, which uses
c) pthreadpool C API, which delegates to
c-1) third_party implementation of pthreadpool if that's what the build has requested, and the rabbit hole ends here.
c-2) C2's implementation of pthreadpool if that's what the build has requested, which itself delegates to
c-2-1) caffe2::ThreadPool, and the rabbit hole ends here.
NNPACK, and (PyTorch) QNNPACK directly hook into (c). They never go through (b).
Differential Revision: D21232894
Test Plan: Imported from OSS
Reviewed By: dreiss
Pulled By: AshkanAliabadi
fbshipit-source-id: 8b3de86247fbc3a327e811983e082f9d40081354
Summary:
NCCL library is built using [CUDA separate compilation](https://devblogs.nvidia.com/separate-compilation-linking-cuda-device-code/), which consists of building intermediate CUDA binaries and then linking them into GPU code that could be executed on device. Intermediate CUDA code is stored in `__nv_relfatbin` section, and code that can be launched is stored in `.nv_fatbin`. When `nvcc` is used to link executable/shared library, it removes those intermediate binaries, but default host linker is not aware of that and therefore it is kept inside host executable. Help compiler by removing `__nv_relfatbin` sections from object file inside `libncc_static.a`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/35843
Test Plan: Build pytorch with CUDA and run `test_distributed.py`
Differential Revision: D20882224
Pulled By: malfet
fbshipit-source-id: f23dd4aa416518324cb38b9bd6846e73a1c7dd21
Summary:
Ignore mixed upper-case/lower-case style for now
Fix space between function and its arguments violation
Pull Request resolved: https://github.com/pytorch/pytorch/pull/35574
Test Plan: CI
Differential Revision: D20712969
Pulled By: malfet
fbshipit-source-id: 0012d430aed916b4518599a0b535e82d15721f78
Summary:
---
How does the current code subsume all detections in the deleted `nccl.py`?
- The dependency of `USE_NCCL` on the OS and `USE_CUDA` is handled as dependency options in `CMakeLists.txt`.
- The main NCCL detection happens in [FindNCCL.cmake](8377d4b32c/cmake/Modules/FindNCCL.cmake), which is called by [nccl.cmake](8377d4b32c/cmake/External/nccl.cmake). When `USE_SYSTEM_NCCL` is false, the previous Python code defer the detection to `find_package(NCCL)`. The change in `nccl.cmake` retains this.
- `USE_STATIC_NCCL` in the previous Python code simply changes the name of the detected library. This is done in `IF (USE_STATIC_NCCL)`.
- Now we only need to look at how the lines below line 20 in `nccl.cmake` are subsumed. These lines list paths to header and library directories that NCCL headers and libraries may reside in and try to search these directories for the key header and library files in turn. These are done by `find_path` for headers and `find_library` for the library files in `FindNCCL.cmake`.
* The call of [find_path](https://cmake.org/cmake/help/v3.8/command/find_path.html) (Search for `NO_DEFAULT_PATH` in the link) by default searches for headers in `<prefix>/include` for each `<prefix>` in `CMAKE_PREFIX_PATH` and `CMAKE_SYSTEM_PREFIX_PATH`. Like the Python code, this commit sets `CMAKE_PREFIX_PATH` to search for `<prefix>` in `NCCL_ROOT_DIR` and home to CUDA. `CMAKE_SYSTEM_PREFIX_PATH` includes the standard directories such as `/usr/local` and `/usr`. `NCCL_INCLUDE_DIR` is also specifically handled.
* Similarly, the call of [find_library](https://cmake.org/cmake/help/v3.8/command/find_library.html) (Search for `NO_DEFAULT_PATH` in the link) by default searches for libraries in directories including `<prefix>/lib` for each `<prefix>` in `CMAKE_PREFIX_PATH` and `CMAKE_SYSTEM_PREFIX_PATH`. But it also handles the edge cases intended to be solved in the Python code more properly:
- It only searches for `<prefix>/lib64` (and `<prefix>/lib32`) if it is appropriate on the system.
- It only searches for `<prefix>/lib/<arch>` for the right `<arch>`, unlike the Python code searches for `lib/<arch>` in a generic way (e.g., the Python code searches for `/usr/lib/x86_64-linux-gnu` but in reality systems have `/usr/lib/x86_64-some-customized-name-linux-gnu`, see https://unix.stackexchange.com/a/226180/38242 ).
---
Regarding for relevant issues:
- https://github.com/pytorch/pytorch/issues/12063 and https://github.com/pytorch/pytorch/issues/2877: These are properly handled, as explained in the updated comment.
- https://github.com/pytorch/pytorch/issues/2941 does not changes NCCL detection specifically for Windows (it changed CUDA detection).
- b7e258f81e A versioned library detection is added, but the order is reversed: The unversioned library becomes preferred. This is because normally unversioned libraries are linked to versioned libraries and preferred by users, and local installation by users are often unversioned. Like the document of [find_library](https://cmake.org/cmake/help/v3.8/command/find_library.html) suggests:
> When using this to specify names with and without a version suffix, we recommend specifying the unversioned name first so that locally-built packages can be found before those provided by distributions.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22930
Differential Revision: D16440275
Pulled By: ezyang
fbshipit-source-id: 11fe80743d4fe89b1ed6f96d5d996496e8ec01aa
Summary:
This tests the water for adding back NNPACK in PyTorch, it's a lot better than the fallback THNN versions.
In #6151, we (ezyang and soumith) removed NNPACK support from PyTorch. Of course Maratyszcza might have advice, too. (Or an opinion on the CMake changes.)
The only functional changes are to use NNPack more aggressively on mobile and a .contiguous() to match NNPack's assumption (I stumbled over that while using NNPack for style transfer.)
The CMake changes try to use the NNPack we already have in git.
In terms of lines of code this is a large part of the diff of https://lernapparat.de/pytorch-jit-android/ . As far as I can tell, we don't have MKLDNN on mobile and the native THNN implementation are prohibitively expensive in terms of both CPU and memory.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15924
Differential Revision: D13709576
Pulled By: ezyang
fbshipit-source-id: f2e287739909451c173abf046588209a7450ca2c
Summary:
This has 4 changes
1) propagate USE_SYSTEM_NCCL. Previously it was ignored and cmake always did a FindPackage
2) respect SCCACHE_DISABLE in our caffe2 sccache wrapper for circleci
3) use SCCACHE_DISABLE when building nccl, because it triggers the same bug as when using CCACHE (already tracked in https://github.com/pytorch/pytorch/issues/13362). This was hidden because we weren't respecting USE_SYSTEM_NCCL, and were never building nccl ourselves in CI
4) In one particular CI configuration (caffe2, cuda 8, cudnn 7), force USE_SYSTEM_NCCL=1. Building the bundled nccl triggers a bug in nvlink. I've done some investigation, but this looks like a tricky, preexisting bug, so rather than hold up this diff I'm tracking it separately in https://github.com/pytorch/pytorch/issues/14486
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14195
Differential Revision: D13237502
Pulled By: anderspapitto
fbshipit-source-id: 1100ac1269c7cd39e2e0b3ba12a56a3ce8977c55
Summary:
I don't have a full analysis, but ccache appears to often fail while
nccl. To work around this, run the NCCL build with CCACHE_DISABLE.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13340
Differential Revision: D12855467
Pulled By: anderspapitto
fbshipit-source-id: 63eb12183ab9d03dd22090f084688ae6390fe8bd
Summary:
always build nccl from within the main cmake build, rather than via a separate invocation in build_pytorch_libs.sh. Use the existing caffe2 codepaths
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13150
Differential Revision: D12815674
Pulled By: anderspapitto
fbshipit-source-id: a710b6f242d159b9816911a25ee2c4b8c3f855aa
Summary:
- Removed the old nccl file
- Make open-source NCCL a submodule
- CMake to make NCCL itself
NCCL2 now is in the default build.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/12359
Reviewed By: orionr, yns88
Differential Revision: D10219665
Pulled By: teng-li
fbshipit-source-id: 134ff47057512ba617b48bf390c1c816fff3f881
