Add Premul Sum test

[Chao1Han]

No description provided.

[Copilot]

Pull request overview

This PR adds test coverage for the Premul Sum operation in the XCCL distributed communication backend. The Premul Sum operation performs element-wise multiplication by a factor before reduction, combining two operations into one for better performance.

• Adds Premul Sum tests to test_allreduce_ops covering multiple data types (half, float, double) and factor types (scalar and tensor)
• Includes commented-out test code for reduce operations with a note that Premul Sum is not supported for reduce ops
• Adds Premul Sum testing to test_reduce_scatter_ops comparing results against separate multiplication and sum operations

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

The variables output and tensor_lists are being modified in-place within the loop iteration. This means subsequent iterations will operate on already-converted float tensors, potentially causing errors or unexpected behavior. These conversions should either be moved outside the loop or use fresh copies for each iteration.

Suggested change

	for factor in (3.0, torch.tensor([5.0], device=self.rank)):
	if isinstance(factor, torch.Tensor):
	factor_ref = factor.cpu().item()
	else:
	factor_ref = factor
	output = [t.float() for t in output]
	tensor_lists = [[t.float() for t in tl] for tl in tensor_lists]
	output_ref = [t.float() for t in output]
	tensor_lists_ref = [
	[t.float() * factor_ref for t in tl] for tl in tensor_lists
	# Create baseline float tensors once, and clone them inside the loop to
	# avoid accumulating mutations across iterations.
	base_output = [t.float() for t in output]
	base_tensor_lists = [[t.float() for t in tl] for tl in tensor_lists]
	for factor in (3.0, torch.tensor([5.0], device=self.rank)):
	if isinstance(factor, torch.Tensor):
	factor_ref = factor.cpu().item()
	else:
	factor_ref = factor
	# Fresh copies for this iteration
	output = [t.clone() for t in base_output]
	tensor_lists = [[t.clone() for t in tl] for tl in base_tensor_lists]
	output_ref = [t.clone() for t in base_output]
	tensor_lists_ref = [
	[t.clone() * factor_ref for t in tl] for tl in base_tensor_lists

[Copilot]

This pattern of extracting a scalar reference from a factor (lines 773-776) is duplicated from earlier in the test file. Consider extracting this into a helper function to reduce code duplication and improve maintainability.