mul(sparse, sparse): extend autograd support to cases with broadcasted dense dims. #84929

nikitaved · 2022-09-13T14:32:00Z

Stack from ghstack (oldest at bottom):

…d dense dims. [ghstack-poisoned]

pytorch-bot · 2022-09-13T14:32:05Z

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/84929

📄 Preview Python docs built from this PR
📄 Preview C++ docs built from this PR
❓ Need help or want to give feedback on the CI? Visit our office hours

Note: Links to docs will display an error until the docs builds have been completed.

❌ 6 Failures

As of commit 5e9e221:

The following jobs have failed:

This comment was automatically generated by Dr. CI and updates every 15 minutes.

nikitaved · 2022-09-13T14:37:25Z

torch/csrc/autograd/FunctionsManual.cpp

+  // same shape as the inputs. However, sum(d=...) is not implemented
+  // for sparse tensors. So, instead, we explicitly reduce over grads'


Probably need to file a feature request for sum, maybe it is there already. The op should not be that hard to implement it seems...

amjames

This is super clear, and great work.

Question: If sum() was to support reduction over dense dimensions would that be sufficient to no longer need the special handling? If so it might be worth looking at.

cpuhrsch · 2022-09-16T16:17:07Z

Are the failures related?

nikitaved · 2022-09-16T17:27:08Z

OK, I am not sure, will try to rebase.

@pytorchbot rebase

pytorchmergebot · 2022-09-16T17:28:37Z

@pytorchbot successfully started a rebase job. Check the current status here

pytorchmergebot · 2022-09-16T17:28:41Z

Rebase failed due to Command git -C /home/runner/work/pytorch/pytorch fetch origin gh/nikitaved/4/orig:gh/nikitaved/4/orig returned non-zero exit code 128

fatal: couldn't find remote ref gh/nikitaved/4/orig

Raised by https://github.com/pytorch/pytorch/actions/runs/3069599253

amjames · 2022-09-16T18:12:37Z

I think you need to rebase locally

This is a combo PR of #84929 and ~#83428. Preliminary benchmarks (square matrices of shape (n, n)). <details> <summary>Script</summary> ```python import torch import math from IPython import get_ipython from itertools import product, repeat import pickle from torch.utils.benchmark import Timer, Compare torch.manual_seed(13) problem_dims = ( # n > nnz (10000, 100), (100000, 1000), (1000000, 10000), # n < nnz (10, 100), (10, 1000), (10, 10000), (100, 1000), (100, 10000), (1000, 10000), (1000, 100000), (1000, 1000000), #(1000000, 1000000000), ) name = "PR" device = "cuda" results = [] for n, nnz in problem_dims: def gen_tensor(coalesce=False): shape = (n, n) nrows, ncols = shape rowidx = torch.randint(low=0, high=nrows, size=(nnz,), device=device) colidx = torch.randint(low=0, high=ncols, size=(nnz,), device=device) itemidx = torch.vstack((rowidx, colidx)) xvalues = torch.randn(nnz, device=device) itemidx = torch.hstack((itemidx, itemidx)) xvalues = torch.hstack((xvalues, xvalues)) res = torch.sparse_coo_tensor(itemidx, xvalues, size=shape) if coalesce: return res.coalesce() else: return res for x_coalesce, y_coalesce in product(*repeat((True, False), 2)): x = gen_tensor(x_coalesce) y = gen_tensor(y_coalesce) smtp = "x * y" timer = Timer(smtp, globals=globals(), label="coo.mul", description=f"{name}: mul, device: {device}", sub_label=f"n={n}, nnz={nnz}, coalesce=({x_coalesce, y_coalesce})", num_threads=torch.get_num_threads()) results.append(timer.blocked_autorange()) compare = Compare(results) compare.trim_significant_figures() compare.print() with open(f"{name}_{device}_mul.pickle", 'wb') as f: pickle.dump(results, f) ``` </details> <details> <summary>Gather results</summary> ```python import pickle from torch.utils.benchmark import Timer, Compare files = [ "PR", "master" ] device = 'cuda' timers = [] for name in files: with open("{}_{}_mul.pickle".format(name, device), 'rb') as f: timers += pickle.load(f) compare = Compare(timers) compare.trim_significant_figures() compare.print() ``` </details> <details> <summary>CUDA</summary> ``` [------------------------------------------------- coo.mul -------------------------------------------------] | PR: mul, device: cuda | master: mul, device: cuda 24 threads: ------------------------------------------------------------------------------------------------- n=10000, nnz=100, coalesce=((True, True)) | 95 | 91 n=10000, nnz=100, coalesce=((True, False)) | 87 | 242 n=10000, nnz=100, coalesce=((False, True)) | 87 | 226 n=10000, nnz=100, coalesce=((False, False)) | 130 | 371 n=100000, nnz=1000, coalesce=((True, True)) | 100 | 521 n=100000, nnz=1000, coalesce=((True, False)) | 90 | 649 n=100000, nnz=1000, coalesce=((False, True)) | 100 | 659 n=100000, nnz=1000, coalesce=((False, False)) | 200 | 781 n=1000000, nnz=10000, coalesce=((True, True)) | 100 | 4861 n=1000000, nnz=10000, coalesce=((True, False)) | 100 | 5012 n=1000000, nnz=10000, coalesce=((False, True)) | 98 | 5010 n=1000000, nnz=10000, coalesce=((False, False)) | 384 | 5174 n=10, nnz=100, coalesce=((True, True)) | 100 | 79 n=10, nnz=100, coalesce=((True, False)) | 100 | 221 n=10, nnz=100, coalesce=((False, True)) | 100 | 221 n=10, nnz=100, coalesce=((False, False)) | 100 | 350 n=10, nnz=1000, coalesce=((True, True)) | 100 | 100 n=10, nnz=1000, coalesce=((True, False)) | 100 | 240 n=10, nnz=1000, coalesce=((False, True)) | 100 | 254 n=10, nnz=1000, coalesce=((False, False)) | 100 | 392 n=10, nnz=10000, coalesce=((True, True)) | 100 | 110 n=10, nnz=10000, coalesce=((True, False)) | 110 | 286 n=10, nnz=10000, coalesce=((False, True)) | 110 | 286 n=10, nnz=10000, coalesce=((False, False)) | 271 | 455 n=100, nnz=1000, coalesce=((True, True)) | 110 | 851 n=100, nnz=1000, coalesce=((True, False)) | 110 | 1000 n=100, nnz=1000, coalesce=((False, True)) | 110 | 990 n=100, nnz=1000, coalesce=((False, False)) | 140 | 1124 n=100, nnz=10000, coalesce=((True, True)) | 110 | 5137 n=100, nnz=10000, coalesce=((True, False)) | 110 | 5391 n=100, nnz=10000, coalesce=((False, True)) | 100 | 5405 n=100, nnz=10000, coalesce=((False, False)) | 249 | 5539 n=1000, nnz=10000, coalesce=((True, True)) | 100 | 8598 n=1000, nnz=10000, coalesce=((True, False)) | 100 | 8800 n=1000, nnz=10000, coalesce=((False, True)) | 100 | 8782 n=1000, nnz=10000, coalesce=((False, False)) | 255 | 8956 n=1000, nnz=100000, coalesce=((True, True)) | 120 | 84500 n=1000, nnz=100000, coalesce=((True, False)) | 200 | 88560 n=1000, nnz=100000, coalesce=((False, True)) | 160 | 89000 n=1000, nnz=100000, coalesce=((False, False)) | 373 | 89000 n=1000, nnz=1000000, coalesce=((True, True)) | 312 | 606400 n=1000, nnz=1000000, coalesce=((True, False)) | 1340 | 609200 n=1000, nnz=1000000, coalesce=((False, True)) | 1340 | 609100 n=1000, nnz=1000000, coalesce=((False, False)) | 4408 | 611400 Times are in microseconds (us). ``` </details> <details> <summary>CPU</summary> ``` [------------------------------------------------ coo.mul ------------------------------------------------] | PR: mul, device: cpu | master: mul, device: cpu 24 threads: ----------------------------------------------------------------------------------------------- n=10000, nnz=100, coalesce=((True, True)) | 8 | 8 n=10000, nnz=100, coalesce=((True, False)) | 32 | 34 n=10000, nnz=100, coalesce=((False, True)) | 32 | 34 n=10000, nnz=100, coalesce=((False, False)) | 41 | 56 n=100000, nnz=1000, coalesce=((True, True)) | 24 | 24 n=100000, nnz=1000, coalesce=((True, False)) | 90 | 100 n=100000, nnz=1000, coalesce=((False, True)) | 87 | 100 n=100000, nnz=1000, coalesce=((False, False)) | 231 | 255 n=1000000, nnz=10000, coalesce=((True, True)) | 190 | 200 n=1000000, nnz=10000, coalesce=((True, False)) | 908 | 2023 n=1000000, nnz=10000, coalesce=((False, True)) | 800 | 2036 n=1000000, nnz=10000, coalesce=((False, False)) | 3684 | 3989 n=10, nnz=100, coalesce=((True, True)) | 8 | 7 n=10, nnz=100, coalesce=((True, False)) | 34 | 30 n=10, nnz=100, coalesce=((False, True)) | 33 | 30 n=10, nnz=100, coalesce=((False, False)) | 44 | 50 n=10, nnz=1000, coalesce=((True, True)) | 8 | 7 n=10, nnz=1000, coalesce=((True, False)) | 100 | 100 n=10, nnz=1000, coalesce=((False, True)) | 130 | 100 n=10, nnz=1000, coalesce=((False, False)) | 746 | 210 n=10, nnz=10000, coalesce=((True, True)) | 8 | 7 n=10, nnz=10000, coalesce=((True, False)) | 1000 | 1500 n=10, nnz=10000, coalesce=((False, True)) | 1000 | 1510 n=10, nnz=10000, coalesce=((False, False)) | 3063 | 2457 n=100, nnz=1000, coalesce=((True, True)) | 25 | 25 n=100, nnz=1000, coalesce=((True, False)) | 180 | 130 n=100, nnz=1000, coalesce=((False, True)) | 200 | 130 n=100, nnz=1000, coalesce=((False, False)) | 271 | 255 n=100, nnz=10000, coalesce=((True, True)) | 100 | 100 n=100, nnz=10000, coalesce=((True, False)) | 2444 | 2290 n=100, nnz=10000, coalesce=((False, True)) | 2455 | 2357 n=100, nnz=10000, coalesce=((False, False)) | 5316 | 3783 n=1000, nnz=10000, coalesce=((True, True)) | 204 | 211 n=1000, nnz=10000, coalesce=((True, False)) | 2457 | 2480 n=1000, nnz=10000, coalesce=((False, True)) | 2448 | 2539 n=1000, nnz=10000, coalesce=((False, False)) | 3665 | 4801 n=1000, nnz=100000, coalesce=((True, True)) | 2293 | 2374 n=1000, nnz=100000, coalesce=((True, False)) | 9000 | 24620 n=1000, nnz=100000, coalesce=((False, True)) | 8000 | 25080 n=1000, nnz=100000, coalesce=((False, False)) | 26500 | 47650 n=1000, nnz=1000000, coalesce=((True, True)) | 10000 | 13000 n=1000, nnz=1000000, coalesce=((True, False)) | 80000 | 362200 n=1000, nnz=1000000, coalesce=((False, True)) | 78050 | 392600 n=1000, nnz=1000000, coalesce=((False, False)) | 312100 | 766900 Times are in microseconds (us). ``` </details> Pull Request resolved: #85336 Approved by: https://github.com/cpuhrsch

facebook-github-bot · 2022-10-04T00:22:53Z

/easycla

As part of the transition to the PyTorch Foundation, this project now requires contributions be covered under the new CLA. See #85559 for additional details.

This comment will trigger a new check of this PR. If you are already covered, you will simply see a new "EasyCLA" check that passes. If you are not covered, a bot will leave a new comment with a link to sign.

linux-foundation-easycla · 2022-10-04T00:22:57Z

The committers listed above are authorized under a signed CLA.

✅ login: nikitaved (8752cd0, 5e9e221)

pearu · 2022-10-04T17:49:45Z

torch/csrc/autograd/FunctionsManual.cpp

+    self_grad = at::_sparse_coo_tensor_unsafe(
+        self_grad_indices,
+        // Need to specify dtype because sum can promote.
+        self_grad_values.sum(
+            values_reduction_dims,
+            /*keepdim=*/true,
+            self_grad_values.scalar_type()),
+        self.sizes());


Could this be related to the failing autograd tests in the CI? If so, we could look into ways to change _values() in-place (via self_grad_values.set_(self_grad_values.sum(...))) as well as updating the dense part of self_grad shape in-place.

The failing tests are because self should not be inlined into the backward function in combination with copy_ used in the failing test.

mul_tensor_backward uses only self.sizes() and self.sparse_dim(). Does the above mean that self cannot be an argument to mul_tensor_backward but self.options() (includes layout and scalar type) and self shape information can be?

No, it cannot be used. Layout is not inlined yet, so I had to modify that part. Also, sizes are not implemented for nested tensors, hence separate dispatch branches in derivatives.yaml. The changes are not in this PR, unfortunately , they are in the mul(sparse, sparse) PR where we decided to revert all grad changes :(

OK, so this PR is blocked atm, right?

Yes, because probably it is best to upgrade autograd to handle shape reductions and propagation of different from inputs' layouts grads.

This is a combo PR of #84929 and ~#83428. Preliminary benchmarks (square matrices of shape (n, n)). <details> <summary>Script</summary> ```python import torch import math from IPython import get_ipython from itertools import product, repeat import pickle from torch.utils.benchmark import Timer, Compare torch.manual_seed(13) problem_dims = ( # n > nnz (10000, 100), (100000, 1000), (1000000, 10000), # n < nnz (10, 100), (10, 1000), (10, 10000), (100, 1000), (100, 10000), (1000, 10000), (1000, 100000), (1000, 1000000), #(1000000, 1000000000), ) name = "PR" device = "cuda" results = [] for n, nnz in problem_dims: def gen_tensor(coalesce=False): shape = (n, n) nrows, ncols = shape rowidx = torch.randint(low=0, high=nrows, size=(nnz,), device=device) colidx = torch.randint(low=0, high=ncols, size=(nnz,), device=device) itemidx = torch.vstack((rowidx, colidx)) xvalues = torch.randn(nnz, device=device) itemidx = torch.hstack((itemidx, itemidx)) xvalues = torch.hstack((xvalues, xvalues)) res = torch.sparse_coo_tensor(itemidx, xvalues, size=shape) if coalesce: return res.coalesce() else: return res for x_coalesce, y_coalesce in product(*repeat((True, False), 2)): x = gen_tensor(x_coalesce) y = gen_tensor(y_coalesce) smtp = "x * y" timer = Timer(smtp, globals=globals(), label="coo.mul", description=f"{name}: mul, device: {device}", sub_label=f"n={n}, nnz={nnz}, coalesce=({x_coalesce, y_coalesce})", num_threads=torch.get_num_threads()) results.append(timer.blocked_autorange()) compare = Compare(results) compare.trim_significant_figures() compare.print() with open(f"{name}_{device}_mul.pickle", 'wb') as f: pickle.dump(results, f) ``` </details> <details> <summary>Gather results</summary> ```python import pickle from torch.utils.benchmark import Timer, Compare files = [ "PR", "master" ] device = 'cuda' timers = [] for name in files: with open("{}_{}_mul.pickle".format(name, device), 'rb') as f: timers += pickle.load(f) compare = Compare(timers) compare.trim_significant_figures() compare.print() ``` </details> <details> <summary>CUDA</summary> ``` [------------------------------------------------- coo.mul -------------------------------------------------] | PR: mul, device: cuda | master: mul, device: cuda 24 threads: ------------------------------------------------------------------------------------------------- n=10000, nnz=100, coalesce=((True, True)) | 95 | 91 n=10000, nnz=100, coalesce=((True, False)) | 87 | 242 n=10000, nnz=100, coalesce=((False, True)) | 87 | 226 n=10000, nnz=100, coalesce=((False, False)) | 130 | 371 n=100000, nnz=1000, coalesce=((True, True)) | 100 | 521 n=100000, nnz=1000, coalesce=((True, False)) | 90 | 649 n=100000, nnz=1000, coalesce=((False, True)) | 100 | 659 n=100000, nnz=1000, coalesce=((False, False)) | 200 | 781 n=1000000, nnz=10000, coalesce=((True, True)) | 100 | 4861 n=1000000, nnz=10000, coalesce=((True, False)) | 100 | 5012 n=1000000, nnz=10000, coalesce=((False, True)) | 98 | 5010 n=1000000, nnz=10000, coalesce=((False, False)) | 384 | 5174 n=10, nnz=100, coalesce=((True, True)) | 100 | 79 n=10, nnz=100, coalesce=((True, False)) | 100 | 221 n=10, nnz=100, coalesce=((False, True)) | 100 | 221 n=10, nnz=100, coalesce=((False, False)) | 100 | 350 n=10, nnz=1000, coalesce=((True, True)) | 100 | 100 n=10, nnz=1000, coalesce=((True, False)) | 100 | 240 n=10, nnz=1000, coalesce=((False, True)) | 100 | 254 n=10, nnz=1000, coalesce=((False, False)) | 100 | 392 n=10, nnz=10000, coalesce=((True, True)) | 100 | 110 n=10, nnz=10000, coalesce=((True, False)) | 110 | 286 n=10, nnz=10000, coalesce=((False, True)) | 110 | 286 n=10, nnz=10000, coalesce=((False, False)) | 271 | 455 n=100, nnz=1000, coalesce=((True, True)) | 110 | 851 n=100, nnz=1000, coalesce=((True, False)) | 110 | 1000 n=100, nnz=1000, coalesce=((False, True)) | 110 | 990 n=100, nnz=1000, coalesce=((False, False)) | 140 | 1124 n=100, nnz=10000, coalesce=((True, True)) | 110 | 5137 n=100, nnz=10000, coalesce=((True, False)) | 110 | 5391 n=100, nnz=10000, coalesce=((False, True)) | 100 | 5405 n=100, nnz=10000, coalesce=((False, False)) | 249 | 5539 n=1000, nnz=10000, coalesce=((True, True)) | 100 | 8598 n=1000, nnz=10000, coalesce=((True, False)) | 100 | 8800 n=1000, nnz=10000, coalesce=((False, True)) | 100 | 8782 n=1000, nnz=10000, coalesce=((False, False)) | 255 | 8956 n=1000, nnz=100000, coalesce=((True, True)) | 120 | 84500 n=1000, nnz=100000, coalesce=((True, False)) | 200 | 88560 n=1000, nnz=100000, coalesce=((False, True)) | 160 | 89000 n=1000, nnz=100000, coalesce=((False, False)) | 373 | 89000 n=1000, nnz=1000000, coalesce=((True, True)) | 312 | 606400 n=1000, nnz=1000000, coalesce=((True, False)) | 1340 | 609200 n=1000, nnz=1000000, coalesce=((False, True)) | 1340 | 609100 n=1000, nnz=1000000, coalesce=((False, False)) | 4408 | 611400 Times are in microseconds (us). ``` </details> <details> <summary>CPU</summary> ``` [------------------------------------------------ coo.mul ------------------------------------------------] | PR: mul, device: cpu | master: mul, device: cpu 24 threads: ----------------------------------------------------------------------------------------------- n=10000, nnz=100, coalesce=((True, True)) | 8 | 8 n=10000, nnz=100, coalesce=((True, False)) | 32 | 34 n=10000, nnz=100, coalesce=((False, True)) | 32 | 34 n=10000, nnz=100, coalesce=((False, False)) | 41 | 56 n=100000, nnz=1000, coalesce=((True, True)) | 24 | 24 n=100000, nnz=1000, coalesce=((True, False)) | 90 | 100 n=100000, nnz=1000, coalesce=((False, True)) | 87 | 100 n=100000, nnz=1000, coalesce=((False, False)) | 231 | 255 n=1000000, nnz=10000, coalesce=((True, True)) | 190 | 200 n=1000000, nnz=10000, coalesce=((True, False)) | 908 | 2023 n=1000000, nnz=10000, coalesce=((False, True)) | 800 | 2036 n=1000000, nnz=10000, coalesce=((False, False)) | 3684 | 3989 n=10, nnz=100, coalesce=((True, True)) | 8 | 7 n=10, nnz=100, coalesce=((True, False)) | 34 | 30 n=10, nnz=100, coalesce=((False, True)) | 33 | 30 n=10, nnz=100, coalesce=((False, False)) | 44 | 50 n=10, nnz=1000, coalesce=((True, True)) | 8 | 7 n=10, nnz=1000, coalesce=((True, False)) | 100 | 100 n=10, nnz=1000, coalesce=((False, True)) | 130 | 100 n=10, nnz=1000, coalesce=((False, False)) | 746 | 210 n=10, nnz=10000, coalesce=((True, True)) | 8 | 7 n=10, nnz=10000, coalesce=((True, False)) | 1000 | 1500 n=10, nnz=10000, coalesce=((False, True)) | 1000 | 1510 n=10, nnz=10000, coalesce=((False, False)) | 3063 | 2457 n=100, nnz=1000, coalesce=((True, True)) | 25 | 25 n=100, nnz=1000, coalesce=((True, False)) | 180 | 130 n=100, nnz=1000, coalesce=((False, True)) | 200 | 130 n=100, nnz=1000, coalesce=((False, False)) | 271 | 255 n=100, nnz=10000, coalesce=((True, True)) | 100 | 100 n=100, nnz=10000, coalesce=((True, False)) | 2444 | 2290 n=100, nnz=10000, coalesce=((False, True)) | 2455 | 2357 n=100, nnz=10000, coalesce=((False, False)) | 5316 | 3783 n=1000, nnz=10000, coalesce=((True, True)) | 204 | 211 n=1000, nnz=10000, coalesce=((True, False)) | 2457 | 2480 n=1000, nnz=10000, coalesce=((False, True)) | 2448 | 2539 n=1000, nnz=10000, coalesce=((False, False)) | 3665 | 4801 n=1000, nnz=100000, coalesce=((True, True)) | 2293 | 2374 n=1000, nnz=100000, coalesce=((True, False)) | 9000 | 24620 n=1000, nnz=100000, coalesce=((False, True)) | 8000 | 25080 n=1000, nnz=100000, coalesce=((False, False)) | 26500 | 47650 n=1000, nnz=1000000, coalesce=((True, True)) | 10000 | 13000 n=1000, nnz=1000000, coalesce=((True, False)) | 80000 | 362200 n=1000, nnz=1000000, coalesce=((False, True)) | 78050 | 392600 n=1000, nnz=1000000, coalesce=((False, False)) | 312100 | 766900 Times are in microseconds (us). ``` </details> Pull Request resolved: #85336 Approved by: https://github.com/cpuhrsch

kit1980 · 2022-11-18T00:59:03Z

/easycla

soulitzer · 2022-12-13T04:59:05Z

tools/autograd/derivatives.yaml

-  self: mul_tensor_backward(grad, other, self.scalar_type())
-  other: mul_tensor_backward(grad, self, other.scalar_type())
+  self: mul_tensor_backward(grad, other, self, self.scalar_type())
+  other: mul_tensor_backward(grad, self, other, other.scalar_type())


We should be careful not to save an extra tensor for backward!

nikitaved · 2022-12-13T10:21:17Z

Closing now since it is probably best for the autograd to figure out how to reduce over sparse tensors. This could be open again if the feature is really needed.

mul(sparse, sparse): extend autograd support to cases with broadcaste…

8752cd0

…d dense dims. [ghstack-poisoned]

nikitaved requested review from albanD and soulitzer as code owners September 13, 2022 14:32

This was referenced Sep 13, 2022

coo binary_op intersection primitives #83427

Closed

Faster mul(sparse, sparse) with broadcasting in dense dims. #83428

Closed

facebook-github-bot added the cla signed label Sep 13, 2022

nikitaved requested review from cpuhrsch, pearu, amjames and bhosmer September 13, 2022 14:33

nikitaved added module: sparse Related to torch.sparse module: autograd Related to torch.autograd, and the autograd engine in general labels Sep 13, 2022

nikitaved commented Sep 13, 2022

View reviewed changes

pytorchbot added the open source label Sep 13, 2022

minor typo fix

5e9e221

amjames approved these changes Sep 16, 2022

View reviewed changes

nikitaved mentioned this pull request Sep 20, 2022

Faster mul(sparse, sparse) with broadcasting in dense dims. #85336

Closed

pearu reviewed Oct 4, 2022

View reviewed changes

soulitzer reviewed Dec 13, 2022

View reviewed changes

nikitaved closed this Dec 13, 2022

facebook-github-bot deleted the gh/nikitaved/4/head branch June 8, 2023 18:07

		// same shape as the inputs. However, sum(d=...) is not implemented
		// for sparse tensors. So, instead, we explicitly reduce over grads'

mul(sparse, sparse): extend autograd support to cases with broadcasted dense dims. #84929

mul(sparse, sparse): extend autograd support to cases with broadcasted dense dims. #84929

Uh oh!

Conversation

nikitaved commented Sep 13, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pytorch-bot bot commented Sep 13, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/84929

❌ 6 Failures

Uh oh!

nikitaved Sep 13, 2022

Choose a reason for hiding this comment

Uh oh!

amjames left a comment

Choose a reason for hiding this comment

Uh oh!

cpuhrsch commented Sep 16, 2022

Uh oh!

nikitaved commented Sep 16, 2022

Uh oh!

pytorchmergebot commented Sep 16, 2022

Uh oh!

pytorchmergebot commented Sep 16, 2022

Uh oh!

amjames commented Sep 16, 2022

Uh oh!

facebook-github-bot commented Oct 4, 2022

Uh oh!

linux-foundation-easycla bot commented Oct 4, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

pearu Oct 4, 2022

Choose a reason for hiding this comment

Uh oh!

nikitaved Oct 4, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pearu Oct 4, 2022

Choose a reason for hiding this comment

Uh oh!

nikitaved Oct 4, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pearu Oct 4, 2022

Choose a reason for hiding this comment

Uh oh!

nikitaved Oct 4, 2022 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kit1980 commented Nov 18, 2022

Uh oh!

soulitzer Dec 13, 2022

Choose a reason for hiding this comment

Uh oh!

nikitaved commented Dec 13, 2022

Uh oh!

Uh oh!

nikitaved commented Sep 13, 2022 •

edited

Loading

pytorch-bot bot commented Sep 13, 2022 •

edited

Loading

linux-foundation-easycla bot commented Oct 4, 2022 •

edited

Loading

nikitaved Oct 4, 2022 •

edited

Loading

nikitaved Oct 4, 2022 •

edited

Loading

nikitaved Oct 4, 2022 •

edited

Loading