add contiguous inside rmsnorm kernel

[jikunshang]

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

PLEASE FILL IN THE PR DESCRIPTION HERE ENSURING ALL CHECKLIST ITEMS ABOVE HAVE BEEN CONSIDERED.

Purpose

add contiguous inside rmsnorm kernel, vllm also add this in kernel in vllm-project/vllm#28103, and remove contiguous call in python level, this may bring potential accuracy issue.

Test Plan

Test Result

(Optional) Documentation Update

BEFORE SUBMITTING, PLEASE READ https://docs.vllm.ai/en/latest/contributing (anything written below this line will be removed by GitHub Actions)

[Copilot]

Pull request overview

This PR adds input tensor contiguity checks and enforcement directly within the RMSnorm kernel implementation, aligning with a similar change made in vLLM. The change moves contiguity handling from Python to C++ level to prevent potential accuracy issues.

Key Changes:

• Adds contiguity validation for output and weight tensors
• Implements automatic conversion to contiguous layout for input tensor when stride is not 1
• Adds runtime checks to ensure proper tensor layouts before kernel execution

---

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

[Copilot]

The check at line 211 is redundant. After the contiguous() call on line 209, the stride(-1) is guaranteed to be 1. This assertion will never fail and can be removed.

Suggested change

TORCH_CHECK(input.stride(-1) == 1);

[Copilot]

Modifying the input parameter silently may lead to unexpected behavior for callers. Consider either making input non-const to signal this modification is possible, or document this behavior clearly, or make a copy to avoid mutating the caller's reference.

Suggested change

	if (input.stride(-1) != 1) {
	input = input.contiguous();
	}
	TORCH_CHECK(input.stride(-1) == 1);
	TORCH_CHECK(weight.is_contiguous());
	VLLM_DISPATCH_FLOATING_TYPES(
	input.scalar_type(), "call_rms_norm_kernel", [&] {
	vllm::call_rms_norm_kernel<scalar_t>(out, input, weight, epsilon);
	auto input_ = input;
	if (input_.stride(-1) != 1) {
	input_ = input_.contiguous();
	}
	TORCH_CHECK(input_.stride(-1) == 1);
	TORCH_CHECK(weight.is_contiguous());
	VLLM_DISPATCH_FLOATING_TYPES(
	input_.scalar_type(), "call_rms_norm_kernel", [&] {
	vllm::call_rms_norm_kernel<scalar_t>(out, input_, weight, epsilon);

[mayuyuace]

Cuda may meet the same problem.
If possible, rms_norm kernel can handle the situation that last demission of input is not contiguous in the future.
Anyway, this PR LGTM.

[baodii]

LGTM