[release/2.10] [Upstream cherry-pick] Add partitioned scatter approach with optimizations#2894
Open
jataylo wants to merge 15 commits intoROCm:release/2.10from
Open
[release/2.10] [Upstream cherry-pick] Add partitioned scatter approach with optimizations#2894jataylo wants to merge 15 commits intoROCm:release/2.10from
jataylo wants to merge 15 commits intoROCm:release/2.10from
Conversation
This PR fixes the unit test,
test/test_cuda.py::TestCuda::test_set_per_process_memory_fraction FAILED
[0.1163s]
```
Traceback (most recent call last):
File "/var/lib/jenkins/pytorch/test/test_cuda.py", line 471, in test_set_per_process_memory_fraction
tmp_tensor = torch.empty(application, dtype=torch.int8, device="cuda")
RuntimeError: Trying to create tensor with negative dimension -5681285432: [-5681285432]
```
This error occurs only on gfx1101 arch.
This error is coming from an integer overflow when another unit test,
test/test_cuda.py::TestCuda::test_randint_generation_for_large_numel
creates a tensor with a huge numel, which overflows into a higher
torch.cuda.max_memory_reserved() when you call
test/test_cuda.py::TestCuda::test_set_per_process_memory_fraction
afterward. To avoid this we introduced torch.cuda.empty_cache() and
torch.cuda.reset_peak_memory_stats() to clean up CUDA states.
JIRA: https://ontrack-internal.amd.com/browse/SWDEV-535295
(cherry picked from commit f86d184)
(cherry picked from commit 1b44228)
These changes are currently in progress of being upstreamed. Bring into release 2.9 for customer model perf improvement --------- Co-authored-by: Nichols A. Romero <nick.romero@amd.com> Co-authored-by: Sampsa Riikonen <sriikone@amd.com> Co-authored-by: Nichols A. Romero <165712832+naromero77amd@users.noreply.github.com> Co-authored-by: AmdSampsa <sampsa.riikonen@amd.com>
…m#2742) hipblaslt should provide better performance in general
…d_memory_with_allocator (ROCm#2811) Use try/finally block. This follows a similar pattern elsewhere in test_cuda.py. Fixes #ROCm/TheRock#2118.
…config. (ROCm#2861) In support of [SWDEV-566103](https://ontrack-internal.amd.com/browse/SWDEV-566103)
…_GROUP_GEMM_CK and default to fallback path (ROCm#2865) On ROCm fast path routes to group_gemm_ck and slow path to _grouped_mm_fallback. By default, fast path = False route is activated since CK path is not performant yet. To activate CK path, use ROCM_ALLOW_GROUP_GEMM_CK=1 env variable. --------- Signed-off-by: Jagadish Krishnamoorthy <jagadish.krishnamoorthy@amd.com>
…s_121625 Cherry-picks from release/2.9 into release/2.10
…and linting fixes (cherry picked from commit 4cb344f)
|
Jenkins build for 913cce10b0702def5f47d91a9f217f70ad1ae339 commit finished as FAILURE |
bda7875 to
e809b45
Compare
cb2ee2f to
4b515e6
Compare
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
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.
From pytorch#168073
It has been observed that in the case of heavy contended atomics poor performance is being achieved.
To solve this problem while minimizing kernel overhead this PR proposes an fx pass which will replace the index_put operation with an alternative scatter approach.
Algorithm:
This will reduce atomic contention at the cost of memory usage. In order to combat this we have built heuristics around the total number of partitions for the expanded buffer, as well as setting a cap on how large these expanded tensors can be (currently 10% of GPU memory)
Note the heuristic cannot be perfect as we do not know the true indices data at compile time, in real world models the indices will have duplicates and not be uniformly distributed which increases atomic contention, currently this cannot be modelled and we have to estimate contention based on input and output buffer sizes.
Benchmark code: https://gist.github.com/jataylo/dd3a6353ad2859efd65fa87b28aa3ebd
This code executes 3 index_add ops to 3 seperate buffers.
N = 1000000
D = 100
n = 501
values = float32 [N,D]
indices = int64 [N]
output = float32 [n, D]
For each run we modify the range of randint to simulate various levels of atomic contention
Gathered two sets of results, one with partitioned_scatter_enabled=True, the other partitioned_scatter_enabled=False
MI300
H100
We can see this could potentially benefit H100 on worst-case examples but would degrade perf in the best case, the atomic add cost on MI300 is heavier meaning this is more beneficial.
On MI300 we can see a mixed bag of e2e model improvements
https://hud.pytorch.org/benchmark/v3/dashboard/compiler_inductor?renderGroupId=main&time.start=2025-11-05T00%3A00%3A00.000Z&time.end=2025-12-04T02%3A00%3A00.000Z&filters.repo=pytorch%2Fpytorch&filters.benchmarkName=compiler&filters.mode=training&filters.dtype=amp&filters.deviceName=rocm+%28mi300x%29&filters.device=rocm&filters.suite=all&filters.compiler=default&lcommit.commit=38c42c575d342a7ea6f4a555bf845071e03b5f35&lcommit.workflow_id=19635538449&lcommit.date=2025-11-24T14%3A00%3A00Z&lcommit.branch=refs%2Ftags%2Fciflow%2Finductor-perf-test-nightly-rocm-mi300%2F168073&rcommit.commit=fedb7f15d177a259bf25c94e888137e0a9a69a81&rcommit.workflow_id=19856622912&rcommit.date=2025-12-02T12%3A00%3A00Z&rcommit.branch=refs%2Ftags%2Fciflow%2Finductor-perf-test-nightly-rocm-mi300%2F168073&lbranch=refs%2Ftags%2Fciflow%2Finductor-perf-test-nightly-rocm-mi300%2F168073&rbranch=refs%2Ftags%2Fciflow%2Finductor-perf-test-nightly-rocm-mi300%2F168073&maxSampling=110
Due to mixed-bag of results we will initially enable this as non default feature but testing passed CI with this enabled here
https://hud.pytorch.org/pytorch/pytorch/pull/168073?sha=fedb7f15d177a259bf25c94e888137e0a9a69a81
Note there are improvements to make after this lands:
cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @hongxiayang @naromero77amd @pragupta @jerrymannil @xinyazhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @dllehr-amd @chenyang78