[ROCm] Improve softmax performance. #1740
Open
+2,209
−443
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
- Fortran package installation moved after gcc - Update libtinfo search code in cmake1 - Install libstdc++.so
Co-authored-by: Wang, Yanyao <[email protected]>
Reversed the condition as required
…pired (#1399) * Skip certificate check only for CentOS7 since certificate expired * Naming
* Triton build conditionalized on ROCM_VERSION (cherry picked from commit 1a7e1fa) * Update pinned commit for rocm6.1 conditionalisation --------- Co-authored-by: Pruthvi Madugundu <[email protected]>
…t date This logic was present until release/2.2 (https://github.com/ROCm/pytorch/blob/4cd7f3ac9078ed449b8ae096887125f9b3b30659/.circleci/scripts/binary_populate_env.sh#L14) but was removed in release/2.3
…MENTS since we plan to use builder repo to set it for older and newer branches. Otherwise we end up with duplicate triton dependency specification eg. PYTORCH_EXTRA_INSTALL_REQUIREMENTS='pytorch-triton-rocm==2.3.0+rocm6.1.0.4804a0dd4a | pytorch-triton-rocm==2.3.0+rocm6.1.0.4804a0dd4a'
(cherry picked from commit 5783935)
… sync (#1455) (#1472) * [SWDEV-469514] hipGraphExecDestroy requires an explicit sync There is a new hip feature where they do not free hipGraph memory as soon as hipGraphExecDestroy is called. This is to support async work on the GPU. See this for more details: https://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#cuda-user-objects We noticed this issue when an allreduce op inside a hipGraph hung. Essentially, ncclCommAbort was waiting for all GPU activity to finish. However, since hipGraph memory was technically still in use, we had an infinite hang. So, I added an extra hipDeviceSynchronize in CUDAGraph's destructor to esure that memory is freed and got test_allreduce_in_cudagraph UT to pass. However, when I ran this on CUDA machine, I noticed that they did not require this extra sync in order to successfully run the UT. It seems that they were calling cudaGraphInstantiateWithFlags with cudaGraphInstantiateFlagAutoFreeOnLaunch, which aggressively frees memory after graph lauch. There is support for this API in our ROCm stack, but we were missing cuda to hip mappings in PyTorch. So, I brought them in and added the necesary conditions to call this API in HIP case also. * Update comments * Use USE_ROCM in keeping with convention * Use USE_ROCM to match convention --------- Co-authored-by: Jithun Nair <[email protected]> (cherry picked from commit e752b4f)
(cherry picked from commit 1feb1a8) Co-authored-by: Jithun Nair <[email protected]>
…tree (#1494) * Skip test__int_mm in 6.0 (cherry picked from commit bf4c478) * [release/2.3] fix test_vmapvjpvjp and skip test_profiler_experimental_tree (#1460) * fix test_vmapvjpvjp and skip test_profiler_experimental_tree (#1444) (cherry picked from commit 7e96391) * remove trailing spaces --------- Co-authored-by: Ramana Cherukuri <[email protected]> (cherry picked from commit 0766b9c) * Reformat test_float8_basics for current rocm support (#1415) (cherry picked from commit cb0e9ad) --------- Co-authored-by: Jack Taylor <[email protected]> Co-authored-by: Andres Lugo <[email protected]>
[release/2.4] Cherry-picks from release/2.3
[MPS][TYPE_PROMOTION] Fix Clamp (pytorch#130226) Summary: 1. Fixed pytorch#130201 by adding type promotion. 2. Added proper tests. 3. Found torch's type promotion is different from numpy as follows: ```python import torch import numpy as np np.clip(np.array([1], dtype=np.float32), np.array([1], dtype=np.int32), None).dtype # dtype('float64') torch.clamp(torch.tensor([1], dtype=torch.float32), torch.tensor([1], dtype=torch.int32)).dtype # torch.float32 ``` ~Not sure the proper way to handle it, it causes numpy ref tests to fail.~ Reason here, so think I'm gonna xfail it: https://github.com/pytorch/pytorch/blob/3c1cf03fde145bdbe1f5ffb81765d076c10b4c04/test/test_ops.py#L260-L264 Pull Request resolved: pytorch#130226 Approved by: https://github.com/malfet (cherry picked from commit 99967e1) Co-authored-by: Li-Huai (Allan) Lin <[email protected]>
[Doc] update guide install mkl-static from conda to pip (pytorch#130026) <img width="619" alt="image" src="https://github.com/pytorch/pytorch/assets/8433590/4ac3ca68-57dc-42c7-ac7a-876dc377ebcf"> Conda intel channel is not avaliable now. Use `pip` install instead of `conda`. `Windows` and `Linux` are avaliable: Binary list: https://pypi.org/project/mkl-static/#files `MacOS` is avaliable for old version: https://pypi.org/project/mkl-static/2021.3.0/#files TODO: 1. cherry-pick to `release/2.4` branch, @atalman . 2. fix it also in `release/2.3` branch: pytorch#131853 Pull Request resolved: pytorch#130026 Approved by: https://github.com/jgong5, https://github.com/atalman (cherry picked from commit 484852c) Co-authored-by: Xu Han <[email protected]>
pytorch#133346) fix for launching kernel invalid config error when calling embedding … (pytorch#130994) …with large index Fixes pytorch#130806 When an output size of 2147483648 (=131072*16384) is expected in the above issue, it throwed out the following error: RuntimeError: HIP error: invalid configuration argument What happened was that the second parameter passed to hipLaunchKernel was crazy {2147483648,1,1}. Found two issues in the Indexing.cu: 1: ptrdiff_t was used but it is signed int, outTotalSize >= 2147483648 can cause overflow when doing [this](https://github.com/pytorch/pytorch/blame/39493aa93419532957e6e5ee97cae842b53b8b59/aten/src/ATen/native/cuda/Indexing.cu#L1367): 2: On ROCm, std::min -> ::min did not work as expected when outTotalSize>=2147483648 As the result, 2147483648 was sent to hipLaunchKernel which the GPU does not support such a huge number since this number specifies the number of threads per block. The original code intended to set 128 threads per block, though this is debatable as the perf would not good for latest powerful GPUs (a TODO item to update for perf maybe?) , but at least it would not cause `invalid configuration argument` error. [Test] Run the same code snippet in the [issue](pytorch#130806), and print the output, its dim and numel(), which looks like below now: ``` output=tensor([[ 0.4044, -0.0244, -0.6865, ..., -0.7800, 0.1175, 1.6726], [-1.0866, -0.1609, 0.3538, ..., 1.9105, 0.7882, 1.1583], [-2.2079, 0.3736, 0.3610, ..., -0.2658, -0.0459, 1.3077], ..., [ 0.8753, -0.7482, -0.1978, ..., 0.9016, 1.1501, -0.5178], [-1.5845, -0.6277, 1.4520, ..., 0.5733, -2.1198, -0.0915], [-0.6310, -1.0239, -0.1910, ..., 0.4309, 0.1630, 0.3239]], device='cuda:0'), dim=2, numel=2147483648 ``` Added a large tensor unit test too. ``` /pytorch# pytest test/nn/test_embedding.py -k test_large_tensors ================================================================================== test session starts =================================================================================== platform linux -- Python 3.9.19, pytest-7.3.2, pluggy-1.4.0 rootdir: /dockerx/development/pytorch configfile: pytest.ini plugins: flakefinder-1.1.0, rerunfailures-14.0, xdist-3.3.1, xdoctest-1.1.0, cpp-2.3.0, hypothesis-5.35.1 collected 288 items / 287 deselected / 1 selected Running 1 items in this shard test/nn/test_embedding.py . [100%] =========================================================================== 1 passed, 287 deselected in 3.16s ============================================================================ ``` Pull Request resolved: pytorch#130994 Approved by: https://github.com/jeffdaily, https://github.com/xw285cornell (cherry picked from commit 637ab85) Co-authored-by: hongxyan <[email protected]>
…nd for scaled dot product function in ROCm (#1818) Fixes [10211](ROCm/frameworks-internal#10211) Deselecting the efficient attention backend for scaled dot product function for ROCm architecture when casual parameter is True. Reference - ROCm/aotriton#25 We are aware of this issue but the precise ETA has not been determined at the moment due to various missing functionalities in AOTriton, (varlen, MQA, causal variants, etc.) Co-authored-by: root <[email protected]>
|
Jenkins build for 70a25e21a46ec6f8685e0b86d1a975f4aa248409 commit finished as FAILURE |
This PR adds a meta_registration for miopen_batch_norm in release/2.4 to resolve this issue ``` NotImplementedError: aten::miopen_batch_norm: attempted to run this operator with Meta tensors, but there was no fake impl or Meta kernel registered. ``` cherry-picked from upstream pytorch@4e4182d Co-authored-by: Jack Taylor <[email protected]> (cherry picked from commit 0ca1e5b)
|
Jenkins build for 70a25e21a46ec6f8685e0b86d1a975f4aa248409 commit finished as FAILURE |
|
Jenkins build for 70a25e21a46ec6f8685e0b86d1a975f4aa248409 commit finished as FAILURE |
|
Jenkins build for 70a25e21a46ec6f8685e0b86d1a975f4aa248409 commit finished as ABORTED |
Update related_commits to pick up apex version update
This PR enables: * using MIOpen OCL_mix backend for bf16 batchnorm with fp32 weights (using torch autocast). This was required and tested for customer workload using NCHW (which is the only memory_layout enabled). * logging for MIOpen batchnorm using `PYTORCH_MIOPEN_EXTRA_LOGGING` env var. TODO in separate PR: Need to implement PyTorch unit tests for this bf16/fp16 inputs + fp32 weights case. (cherry picked from commit abbfe77)
|
Jenkins build for 70a25e21a46ec6f8685e0b86d1a975f4aa248409 commit finished as FAILURE |
… (#1839) Improve performance of reduce sum for 3D shapes. Pull Request resolved: pytorch#143137 Approved by: https://github.com/jeffdaily, https://github.com/eqy Co-authored-by: Doru Bercea <[email protected]>
|
Jenkins build for 65d155ce2b085af1a1e9605ec28faae2628cc8cf commit finished as FAILURE |
doru1004
force-pushed
the
improve-softmax-performance-2.4
branch
from
70a25e2 to
1dfe9db
Compare
|
Jenkins build for 65d155ce2b085af1a1e9605ec28faae2628cc8cf commit finished as FAILURE |
|
Jenkins build for 65d155ce2b085af1a1e9605ec28faae2628cc8cf commit finished as FAILURE |
|
Jenkins build for 65d155ce2b085af1a1e9605ec28faae2628cc8cf commit finished as FAILURE |
|
Jenkins build for 65d155ce2b085af1a1e9605ec28faae2628cc8cf commit finished as FAILURE |
|
Jenkins build for 65d155ce2b085af1a1e9605ec28faae2628cc8cf commit finished as FAILURE |
|
Jenkins build for 65d155ce2b085af1a1e9605ec28faae2628cc8cf commit finished as ABORTED |
|
Jenkins build for 65d155ce2b085af1a1e9605ec28faae2628cc8cf commit finished as FAILURE |
|
Jenkins build for 65d155ce2b085af1a1e9605ec28faae2628cc8cf commit finished as FAILURE |
|
Jenkins build for 65d155ce2b085af1a1e9605ec28faae2628cc8cf commit finished as FAILURE |
jithunnair-amd
force-pushed
the
release/2.4
branch
from
9f6c415 to
2f00d85
Compare
|
Jenkins build for 65d155ce2b085af1a1e9605ec28faae2628cc8cf commit finished as FAILURE |
|
Jenkins build for 65d155ce2b085af1a1e9605ec28faae2628cc8cf commit finished as FAILURE |
|
Jenkins build for 65d155ce2b085af1a1e9605ec28faae2628cc8cf commit finished as FAILURE |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This patch improves the performance of softmax for 2D tensors by:
The impact on numerical accuracy is within a 1e-5 for half precision and 1e-7 for full precision.
The impact on performance for MI300X is between 22% and 50% percentage improvement over current runtimes.