feat(ascend): add SwiGLU forward and backward kernels

[erfgss]

Summary

• Add fused SwiGLU (gate * sigmoid(gate) * up) forward and backward kernels written in Ascend C (csrc/ascend/activation.asc), computing in FP32 and returning the input dtype (fp16 / bf16 / fp32)
• Register the kernels on the NPU extension via a new csrc/ascend/bindings.asc module initializer (swiglu_forward / swiglu_backward) with .pyi stubs
• Add SwiGLUAscendOp Python wrapper with autograd support (once_differentiable), input validation, arbitrary shapes / empty tensors / strided-view handling, and an FP32-output variant
• Wire ASCEND_SWIGLU into KernelRegistry NPU dispatch priority (ascend -> pytorch fallback) and update gtest operator specs
• Extend tests/test_swiglu.py (precision vs reference, dtype/shape/strided/empty coverage, autograd checks) and update docs/operators/activation.md

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Team

Run ID: 1bd775d9-4a75-4199-b759-afbc8f2ab30f

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

_{Comment @coderabbitai help to get the list of available commands.}

[zhangj1an]

Thank you for your contribution! The gtest contract passes for fp32/bf16/fp16 (forward + backward) and the pytest suite passes.

Gtest results (scripts/check_operator.py, all three dtypes, --check-grad):

dtype	forward	gradient:gate	gradient:up	verdict
fp32	0.0 (bitwise)	9.54e-7 (~1 ulp)	0.0 (bitwise)	✅ PASS
bf16	3.01e-2	6.10e-5	0.0 (bitwise)	✅ PASS
fp16	3.83e-3	4.88e-4	0.0 (bitwise)	✅ PASS

Thresholds are (atol, rtol) = (2e-2, 1.6e-2) for bf16 and (1e-3, 1e-3) for fp16, with per-element allclose semantics (|diff| ≤ atol + rtol·|ref|). So bf16 forward max_abs 3.01e-2 and fp16 forward max_abs 3.83e-3 exceed the bare atol but pass legitimately via the rtol term — that is the intended gtest contract behavior, not a fudge.

Command used:

ASCEND_RT_VISIBLE_DEVICES=1 python scripts/check_operator.py \
  --op swiglu --candidate ascend --device npu \
  --dtype {fp32,bf16,fp16} --batch 2 --seq 16 --vocab 257 --normalized-dim 4096 \
  --check-grad

Optional minor Issues

1. Flaky test (must fix): test_autograd_wrapper_contiguity_and_gradient_routing fails in ~2/5 full-suite runs (different parametrization each time, always passes in isolation): torch.testing.assert_close(result, ref, rtol=0, atol=0) with 1/35 elements off by 1 fp32 ULP (5.96e-8). The wrapper path feeds contiguous copies to NativeSwiGLUOp while the reference runs on the original strided .t() tensors; torch's elementwise kernels can differ by 1 ULP between layouts. CPU-only wrapper test — fix by feeding the same layout to both paths or allowing 1 ULP.
2. Naming inconsistent with #378: the binding file is csrc/ascend/bindings.asc while #378 uses csrc/ascend/ops_ascend.cpp (other Ascend branches use npu_module.cpp) — three names for the same role; also a host-only pybind TU named .asc rides through the bisheng kernel glob. The bound symbols swiglu_forward / swiglu_backward lack the _ascend suffix used by rope_apply_ascend and batch_invariant_logp_ascend in the same _C_npu module.