Fix flaky dropout tests

[micmelesse]

Motivation

This pr fixes a few flaky tests that have been failing on main due to mis-compilation of tl.where. The tl.where calls are replaced with semantically equivalent code.

Technical Details

Test Plan

I tested locally on an MI350 machine.

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[github-actions]

🏷️ CI Guide

Runs automatically on every PR:

• ✅ Pre-checks (submodule verification, code formatting)
• ✅ Aiter op tests (gfx942 + gfx950)
• ✅ Triton tests on MI35X (only when aiter/ops/triton/** or related paths are changed)

Extended tests (opt-in via labels):

Label	Tests
ci:triton-300x	Run an additional Triton test job on MI300X in PRs; main branch always runs both MI35X and MI300X
ci:sglang	SGLang integration tests: DeepSeek-R1-MXFP4 accuracy, Qwen 3.5 accuracy
ci:atom	ATOM benchmark: DeepSeek-R1-0528, GPT-OSS-120B
ci:atom_full	ATOM accuracy suite for PR and main models from ATOM models_accuracy.json
ci:vllm	vLLM benchmark: GPT-OSS-120B, DeepSeek-R1-0528, Kimi-K2.5
ci:all	All standard extended tests (excludes ci:atom_full)

Only add ci:atom_full for FlyDSL or Triton upgrades.
Add labels via the sidebar or gh pr edit 3289 --add-label <label>

[Copilot]

Pull request overview

This PR aims to eliminate flakiness in dropout-related FlashAttention backward paths by removing tl.where usage in dropout masking (reported to miscompile) and replacing it with equivalent mask-multiply logic in Triton kernels.

Changes:

• Replaced tl.where(dropout_mask, x, 0.0) * dropout_scale with x * dropout_mask.to(x.dtype) * dropout_scale in multiple backward inner loops.
• Applied the change consistently across split-k and atomic variants of the backward kernels.

---

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