perf: integrate fused AscendC LayerNormFwd.

[Fengfengst123]

中文说明

为什么修改

Qwen3.5 Gated DeltaNet 的 gated LayerNorm/RMSNorm 路径当前使用 Triton 实现。xllm_ops 已提供融合的 AscendC aclnnLayerNormFwd 算子，本 PR 将其接入 xLLM runtime，以减少 prefill 阶段的算子开销并改善 TTFT。

怎么修改

• 新增 layer_norm_fwd_aclnn C++ wrapper，处理输入 reshape、contiguous、optional bias/gate、输出和统计 tensor 分配。
• 将 wrapper 加入 NPU xllm_ops CMake 和 API 声明。
• NPU gated LayerNorm 默认调用融合 AscendC 算子。

性能

Qwen3.5-27B，BF16，TP=1，Ascend 910B3；启用 graph 和 schedule overlap，并使用相同模型、二进制、workload、NPU 和 CPU 绑核策略。

16 组交替顺序的 Triton/AscendC 配对 E2E 测试，input=2048、output=2048、parallel=4：

指标	Triton	AscendC	变化
TTFT	2189.237 ms	1961.203 ms	改善 9.833%

与 PR #1936 的 batch=1、input=2048、output=16 对比中，双方各取前 12 轮，并按 1.5×IQR 去除离群值：AscendC TTFT 为 591.911 ms，PR #1936 为 599.399 ms，AscendC 改善 1.249%。

验证

• xLLM runtime 接入静态检查：18/18 PASS。
• git diff --check：通过。
• AscendC 单算子精度测试：通过。
• 16 组 Triton/AscendC 配对 E2E 性能测试：通过，共 32 个有效服务 run。
• 固定提示 E2E smoke：两种后端均为 5/5 非空输出。
• 测试前清除了历史 xLLM/xllm_ops CMake、build 和 OPP 产物，并仅加载本轮新构建的 workspace OPP。

English Description

Why

The Qwen3.5 Gated DeltaNet gated LayerNorm/RMSNorm path currently uses the Triton implementation. xllm_ops now provides a fused AscendC aclnnLayerNormFwd operator. This PR integrates it into the xLLM runtime to reduce prefill operator overhead and improve TTFT.

What changed

• Add a layer_norm_fwd_aclnn C++ wrapper for reshaping, contiguous inputs, optional bias/gate tensors, and output/statistics allocation.
• Register the wrapper in the NPU xllm_ops CMake target and API header.
• Use the fused AscendC operator by default for NPU gated LayerNorm.

Performance

Qwen3.5-27B, BF16, TP=1 on Ascend 910B3, with graph mode and schedule overlap enabled. Both variants use the same model, binary, workload, NPU, and CPU affinity policy.

Across 16 alternating paired Triton/AscendC E2E rounds with input=2048, output=2048, and parallel=4:

Metric	Triton	AscendC	Change
TTFT	2189.237 ms	1961.203 ms	9.833% improvement

For the batch=1, input=2048, output=16 comparison against PR #1936, the first 12 rounds from each variant were evaluated with 1.5×IQR outlier filtering. AscendC achieved 591.911 ms TTFT versus 599.399 ms for PR #1936, a 1.249% improvement.

Validation

• xLLM runtime integration harness: 18/18 PASS.
• git diff --check: passed.
• AscendC single-op accuracy validation: passed.
• 16 paired Triton/AscendC E2E rounds: passed, with 32 valid service runs.
• Fixed-prompt E2E smoke: 5/5 non-empty outputs for both backends.
• Historical xLLM/xllm_ops CMake, build, and OPP artifacts were removed before testing; runtime loaded only the freshly built workspace OPP.

[Sinle4Cat]

为什么会有stringcmp操作？

[Fengfengst123]

嘿嘿，我把回退的代码给删掉吧