Add a decomposition for torch.put #115306
Closed
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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/115306
Note: Links to docs will display an error until the docs builds have been completed. ❌ 5 New FailuresAs of commit 876e0fd with merge base e5f2ac1 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
peterbell10
reviewed
Dec 7, 2023
torch/_decomp/decompositions.py
Outdated
Comment on lines
4348
to
4352
| flattened = self.flatten().clone() | ||
| if accumulate: | ||
| flattened[index] += source | ||
| else: | ||
| flattened[index] = source |
There was a problem hiding this comment.
Mutation isn't legal inside decompositions of functional operators. If I'm not mistaken you could just use torch.index_put though.
Suggested change
| flattened = self.flatten().clone() | |
| if accumulate: | |
| flattened[index] += source | |
| else: | |
| flattened[index] = source | |
| flattened = self.flatten() | |
| flattened = torch.index_put(flattened, [index], source, accumulate) |
There was a problem hiding this comment.
It seems this fails because put supports NumPy-style "reverse broadcasting":
>>> x = torch.tensor(1.0)
>>> index = torch.tensor(0)
>>> source = torch.tensor([0.0])
>>> torch.put(x, index, source)
tensor(0.)(which is IMO a bad idea, but it is tested)
There was a problem hiding this comment.
I guess source.reshape(index.shape) should work, assuming the actual shape conditions are all tested separately.
Co-authored-by: peterbell10 <peterbell10@live.co.uk>
|
Benchmark script based on #114813 (comment) import torch
from torch.testing import make_tensor
from torch.fx.experimental.proxy_tensor import make_fx
from torch.utils.benchmark import Timer, Compare
from torch._inductor.compile_fx import compile_fx_inner, cudagraphify_impl
from torch._inductor.decomposition import decompositions
from itertools import product
from functools import partial
torch._logging.set_logs(output_code=True)
benchmark_name = "put"
Ss = [512]
def gen_inputs():
make_arg = partial(torch.randn, dtype=torch.float32, device="cuda")
make_source = partial(torch.randn, dtype=torch.float32, device="cuda")
def make_idx(n):
return make_tensor((n,), device="cuda", dtype=torch.int64, low=0, high=n)
for b, s, in product(Ss, Ss):
yield make_arg((b * s)), make_idx(b), make_source(b)
def benchmark(label, f, x, idx, source):
return Timer("f([x, idx, source])",
globals=locals(),
label=benchmark_name,
description=label,
sub_label=f"{tuple(x.shape)}",
num_threads=torch.get_num_threads()).blocked_autorange(min_run_time=2)
def compare(x, idx, source):
def f(args):
x, idx, source = args
val = torch.ops.aten.put(x, idx, source)
return (val,)
print(f"{tuple(x.shape)}")
args = [x, idx, source]
decomposed = make_fx(f, decomposition_table=decompositions, tracing_mode="fake")(args)
compiled_decomposed = compile_fx_inner(decomposed, args, cudagraphs=False)
yield benchmark("Decomposed", compiled_decomposed, *args)
non_decomposed = make_fx(f, tracing_mode="fake")(args)
compiled_nondecomposed = compile_fx_inner(non_decomposed, args, cudagraphs=False)
yield benchmark("Lowering", compiled_nondecomposed, *args)
# Just show the first two generated kernels
torch._logging.set_logs(output_code=False)
cuda_f = cudagraphify_impl(f, args, static_input_idxs=tuple(range(len(args))))
yield benchmark("Eager", cuda_f, *args)
results = []
for args in gen_inputs():
for res in compare(*args):
results.append(res)
compare = Compare(results)
compare.trim_significant_figures()
compare.print() |
peterbell10
reviewed
Dec 7, 2023
colesbury
added
the
triaged
Co-authored-by: peterbell10 <peterbell10@live.co.uk>
peterbell10
reviewed
Dec 13, 2023
torch/_inductor/lowering.py
Outdated
| make_fallback(aten.polygamma) | ||
| make_fallback(aten.put) | ||
| make_fallback(aten.reflection_pad1d) | ||
| make_fallback(aten.replication_pad1d) |
There was a problem hiding this comment.
I think your rebase has gone awry.
There was a problem hiding this comment.
You need to remove the aten.put fallback since we're adding a decomposition that replaces it.
| norm = x.norm(2, keep_dim, keepdim=True) | ||
| return x * (y / norm), norm | ||
|
|
||
| @register_decomposition(aten.put) |
There was a problem hiding this comment.
For the test_has_decomposition failure, you can run the test as
EXPECTTEST_ACCEPT=1 pytest test/test_decomp,py -k HasDecomp
and it will update the expected test output files
test/expect/HasDecompTest.test_has_decomposition.expect
test/expect/HasDecompTest.test_aten_core_opeartors.expect
pearu
added a commit
that referenced
this pull request
Feb 18, 2024
As in the title. It is an updated copy of #115306 . cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames [ghstack-poisoned]
pearu
added a commit
that referenced
this pull request
Feb 18, 2024
As in the title. It is an updated copy of #115306 . cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames [ghstack-poisoned]
pearu
added a commit
that referenced
this pull request
Feb 18, 2024
As in the title. It is an updated copy of #115306 . cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames [ghstack-poisoned]
pearu
added a commit
that referenced
this pull request
Feb 18, 2024
As in the title. It is an updated copy of #115306 . cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames [ghstack-poisoned]
pearu
added a commit
that referenced
this pull request
Feb 18, 2024
As in the title. It is an updated copy of #115306 . cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames [ghstack-poisoned]
pearu
added a commit
that referenced
this pull request
Feb 18, 2024
As in the title. It is an updated copy of #115306 . cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames [ghstack-poisoned]
pearu
added a commit
that referenced
this pull request
Feb 19, 2024
As in the title. It is an updated copy of #115306 . cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames [ghstack-poisoned]
pearu
added a commit
that referenced
this pull request
Feb 19, 2024
As in the title. It is an updated copy of #115306 . cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames [ghstack-poisoned]
pearu
added a commit
that referenced
this pull request
Feb 19, 2024
As in the title. It is an updated copy of #115306 . cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames [ghstack-poisoned]
pearu
added a commit
that referenced
this pull request
Feb 19, 2024
As in the title. It is an updated copy of #115306 . cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames [ghstack-poisoned]
pearu
added a commit
that referenced
this pull request
Feb 19, 2024
As in the title. It is an updated copy of #115306 . cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames [ghstack-poisoned]
pearu
added a commit
that referenced
this pull request
Feb 19, 2024
As in the title. It is an updated copy of #115306 . cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames [ghstack-poisoned]
pearu
added a commit
that referenced
this pull request
Feb 19, 2024
As in the title. It is an updated copy of #115306 . cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames [ghstack-poisoned]
pearu
added a commit
that referenced
this pull request
Feb 19, 2024
As in the title. It is an updated copy of #115306 . cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames [ghstack-poisoned]
pearu
added a commit
that referenced
this pull request
Mar 4, 2024
As in the title. It is an updated copy of #115306 . cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames [ghstack-poisoned]
pearu
added a commit
that referenced
this pull request
Mar 4, 2024
As in the title. It is an updated copy of #115306 . cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx peterbell10 ipiszy yf225 chenyang78 kadeng muchulee8 aakhundov ColinPeppler amjames [ghstack-poisoned]
pytorchmergebot
pushed a commit
that referenced
this pull request
Mar 4, 2024
As in the title. It is an updated copy of #115306 . Pull Request resolved: #120179 Approved by: https://github.com/lezcano, https://github.com/peterbell10, https://github.com/jgong5
|
Looks like this PR hasn't been updated in a while so we're going to go ahead and mark this as |
|
landed in #120179 |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
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.
Getting a test failure here that I don't understand
However, from pure Python (CPU)
torch.putand my decomposition seem to give the same thing:cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @peterbell10 @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @aakhundov @ColinPeppler @amjames @desertfire @chauhang