[adenzhou1350]

Summary

Fix incorrect activation values and gradients for one-dimensional strided views, such as x[::3]. The layout guard previously accepted every rank-one tensor even though these kernels use unit column offsets; the existing contiguous fallback now also handles strided vector inputs and upstream gradients for sigmoid, logsigmoid, swish, SwiGLU, SigmoidGLU and PowGLU.

Contiguous vectors and the higher-dimensional row-stride path are unchanged. This follows the non-contiguous support in #765; it does not extend scalar support or change the separate Ascend implementation.

Test plan

• RTX 5090, Python 3.12, PyTorch 2.11.0+cu130, Triton 3.6.0, default Triton backend, FLA_CI_ENV=0.
• New FP32 cases cross contiguous/stride-3 inputs with contiguous/stride-3 incoming gradients for all six functions, checking outputs and every supported gradient against the existing differentiable references at the existing 1e-3 tolerance.
• On baseline f3bec720f7847d4f176872b34f002cc580e2cc6e, the identical regression tests give 6 passes and 19 failures (18 numerical failures plus the layout-helper assertion). On this head all 25 pass. Independent public-call probes also reproduce the numerical failures before the fix and pass after it.
• FLA_CI_ENV=0 pytest tests/modules/test_activation.py -q --tb=short: 103 cases passed across two bounded runs. The first completed 72 cases before a 900-second cold-compilation limit; the next deselected exactly those 72 and passed the remaining 31. This is not a single uninterrupted full-suite pass. No cases failed or skipped; the second run reported three PyTorch/pytest setup advisories, not tolerance warnings.
• pre-commit run --files fla/modules/activations/ops.py tests/modules/test_activation.py: passed.
• Ran python scripts/find_dependent_tests.py fla/modules/activations/ops.py and audited shared callers. Its broad transitive layer/model test list was not run; this change is limited to rank-one layout admission, and the activation suite covers the existing dense/chunked and compiled paths. No full-model or non-NVIDIA validation claim.

Benchmark / NCU

N/A for kernel benchmarks: no kernel code or higher-rank dispatch changed. This restores correctness through the existing copy fallback for previously misread vector views, so the incorrect no-copy baseline is not a valid performance comparison for those views. No performance-neutrality, throughput or speedup claim.

Breaking changes

None. Public signatures, dtypes and tolerances are unchanged.

Checklist

• I have read CONTRIBUTING.md and follow its conventions (code style, docstrings, commit prefixes).
• I have read AGENTS.md and, where my change matches its scope, the relevant skill under .agents/skills.
• Dependent tests pass locally or in CI, and new behavior is covered by tests where applicable (tick as N/A for changes with no testable code, e.g. docs-only).
• Kernel changes include same-hardware before/after benchmark numbers, dense + varlen where applicable (tick as N/A when no kernel code changed).
• This PR is minor/cosmetic-only (typo, formatting, style-only tweaks) — tick only if it is, and justify below.
• I understand this PR exceeds 500 changed lines and have explained its single purpose and why it should stay together below.