Skip dist all2all related case

[Chao1Han]

No description provided.

[Copilot]

Pull Request Overview

This PR disables several distributed all-to-all tests by commenting them out, effectively skipping these tests during execution.

• Commented out the test for alltoall operations with xpufree race
• Commented out multiple tests for all-to-all single operations and variants

[Copilot]

[nitpick] Disabling the 'test_alltoall_ops_with_xpufree_race' by commenting out the test may lead to code clutter; consider using a skip decorator or removing the test if it's no longer required.

Suggested change

	# @requires_xccl()
	# @skip_but_pass_in_sandcastle_if(not TEST_MULTIGPU, "XCCL test requires 2+ GPUs")
	# def test_alltoall_ops_with_xpufree_race(self):
	# pg = self.pg
	# opts = c10d.AllToAllOptions()
	# local_device = f"xpu:{self.rank_to_GPU[self.rank][0]}"
	# torch.xpu.set_device(local_device)
	# input = torch.rand(1000, 1000, device=local_device)
	# output = torch.rand(1000, 1000, device=local_device)
	# race_tensors = []
	# # create some tensors to race with alltoall collective
	# for _ in range(10):
	# tmp = []
	# for i in range(5):
	@unittest.skip("Skipping test_alltoall_ops_with_xpufree_race due to known issues with XPU free race conditions.")
	@requires_xccl()
	@skip_but_pass_in_sandcastle_if(not TEST_MULTIGPU, "XCCL test requires 2+ GPUs")
	def test_alltoall_ops_with_xpufree_race(self):
	pg = self.pg
	opts = c10d.AllToAllOptions()
	local_device = f"xpu:{self.rank_to_GPU[self.rank][0]}"
	torch.xpu.set_device(local_device)
	input = torch.rand(1000, 1000, device=local_device)
	output = torch.rand(1000, 1000, device=local_device)
	race_tensors = []
	# create some tensors to race with alltoall collective
	for _ in range(10):
	tmp = []

[Copilot]

[nitpick] The block of tests for various all-to-all operations has been commented out rather than formally skipped; consider refactoring with proper skip annotations or removing this code to improve maintainability.

Suggested change

	# @requires_xccl()
	# @skip_but_pass_in_sandcastle_if(not TEST_MULTIGPU, "XCCL test requires 2+ GPUs")
	# def test_all_to_all_single(self):
	# device = self.rank_to_GPU[self.rank][0]
	# row = self.world_size * (self.rank + 1) * (self.world_size + 1) / 2
	# x = torch.ones(int(row), 5, device=device) * (self.rank + 1)
	# x.requires_grad = True
	# y = torch.empty_like(x)
	# split_sizes = [(i + 1) * (self.rank + 1) for i in range(self.world_size)]
	# y = torch.distributed.nn.all_to_all_single(
	# y, x, output_split_sizes=split_sizes, input_split_sizes=split_sizes
	# )
	# expected = []
	# for idx, tensor in enumerate(torch.split(x, split_sizes)):
	# expected.append(torch.full_like(tensor, (idx + 1)))
	# expected = torch.cat(expected)
	@requires_xccl()
	@skip_but_pass_in_sandcastle_if(not TEST_MULTIGPU, "XCCL test requires 2+ GPUs")
	def test_all_to_all_single(self):
	device = self.rank_to_GPU[self.rank][0]
	row = self.world_size * (self.rank + 1) * (self.world_size + 1) / 2
	x = torch.ones(int(row), 5, device=device) * (self.rank + 1)
	x.requires_grad = True
	y = torch.empty_like(x)
	split_sizes = [(i + 1) * (self.rank + 1) for i in range(self.world_size)]
	y = torch.distributed.nn.all_to_all_single(
	y, x, output_split_sizes=split_sizes, input_split_sizes=split_sizes
	)
	expected = []
	for idx, tensor in enumerate(torch.split(x, split_sizes)):
	expected.append(torch.full_like(tensor, (idx + 1)))
	expected = torch.cat(expected)