perf: add native split gated RMSNorm for Qwen3.5 and Qwen3.6 on NPU

[wanghuanjun2113]

中文说明

为什么修改

Qwen3.5/Qwen3.6 Gated DeltaNet 的大 token prefill 场景当前使用融合 Gated RMSNorm 算子。将该路径拆分为 Ascend 原生 RmsNorm、SiLU 和逐元素乘后，端到端测试中 6 个场景的 TTFT 和 E2E 均有改善，其中 35B TP=2/TP=4 的 E2E 分别提升 7.35%/5.49%。

怎么修改

• 仅在 NPU 构建中增加原生拆分路径。
• 通过 XLLM_NATIVE_SPLIT_GATED_RMS_NORM=1 显式启用，默认仍走原融合实现。
• 当 gate 存在、输入为二维、token 行数不小于 1024 且 hidden size 与 RMSNorm 维度一致时，依次执行：
  1. NPU 原生 rms_norm；
  2. gate 原地 SiLU；
  3. normalized output 原地乘 gate。
• 其他设备、未设置环境变量或不满足形状条件时均回退原 gated_layer_norm 路径。
• NPU 专用头文件和 helper 均置于 USE_NPU 条件编译保护内，不影响其他后端。

性能测试配置

• 测试代码基线：95a3e499674900f6d08cce881d14e31479f14ca1
• workload：warmup 3，正式请求 20，统计 19 条稳态请求；平均输入约 4317 tokens，输出 25 tokens。
• MTP：3 speculative tokens；启用 graph、chunked prefill、schedule overlap 和 LCCL。
• A/B 使用相同模型、draft model、二进制构建选项、workload 和物理卡。

负数表示优化版延迟降低。

场景	TTFT 基线/优化 (ms)	TTFT Δ	TPOT 基线/优化 (ms)	TPOT Δ	E2E 基线/优化 (ms)	E2E Δ	MTP 接受率基线/优化	Δ (pp)
Qwen3.5 2B TP=1	109.684 / 102.474	-6.57%	4.4316 / 4.3895	-0.95%	216.011 / 208.111	-3.66%	72.95% / 75.86%	+2.92
Qwen3.5 2B TP=2	80.684 / 77.053	-4.50%	4.2158 / 4.1789	-0.87%	181.795 / 177.279	-2.48%	76.24% / 73.46%	-2.77
Qwen3.6 27B TP=2	656.368 / 620.105	-5.52%	15.0632 / 15.1211	+0.38%	1017.942 / 983.063	-3.43%	65.40% / 64.60%	-0.80
Qwen3.6 27B TP=4	396.421 / 382.316	-3.56%	9.8737 / 10.0158	+1.44%	633.568 / 622.616	-1.73%	67.31% / 64.90%	-2.41
Qwen3.6 35B TP=2	212.947 / 195.842	-8.03%	6.9526 / 6.5053	-6.43%	379.921 / 352.005	-7.35%	90.63% / 92.53%	+1.90
Qwen3.6 35B TP=4	159.316 / 149.895	-5.91%	5.8632 / 5.5632	-5.12%	300.053 / 283.584	-5.49%	92.35% / 92.53%	+0.18

验证

• clang-format 20.1.6：通过。
• git diff --check：通过。
• 6 个场景 A/B 性能测试：通过，每组均有 19 条稳态样本。

English Description

Why

The Qwen3.5/Qwen3.6 Gated DeltaNet path currently uses a fused Gated RMSNorm operator for large-token prefill workloads. After splitting this path into native Ascend RmsNorm, SiLU, and elementwise multiplication, all six end-to-end scenarios improve TTFT and E2E latency. The 35B TP=2 and TP=4 workloads improve E2E latency by 7.35% and 5.49%, respectively.

What changed

• Add the native split path only for NPU builds.
• Enable it explicitly with XLLM_NATIVE_SPLIT_GATED_RMS_NORM=1; the fused implementation remains the default.
• When a gate is present, the input is 2-D, the row count is at least 1024, and the hidden size matches the RMSNorm dimension, run:
  1. native NPU rms_norm;
  2. in-place SiLU on the gate;
  3. in-place multiplication of the normalized output and gate.
• Fall back to the existing gated_layer_norm path for other devices, when the environment variable is unset, or when the shape conditions are not met.
• Guard all NPU-only includes and helpers with USE_NPU so other backends are unaffected.

Performance setup

• Evaluated source baseline: 95a3e499674900f6d08cce881d14e31479f14ca1
• Workload: 3 warmup requests and 20 formal requests; metrics use 19 steady-state requests. Average input length is about 4,317 tokens and output length is 25 tokens.
• MTP uses 3 speculative tokens. Graph mode, chunked prefill, schedule overlap, and LCCL are enabled.
• Each A/B pair uses the same model, draft model, binary build options, workload, and physical devices.

Negative deltas mean lower latency.

Scenario	TTFT baseline/optimized (ms)	TTFT Δ	TPOT baseline/optimized (ms)	TPOT Δ	E2E baseline/optimized (ms)	E2E Δ	MTP acceptance baseline/optimized	Δ (pp)
Qwen3.5 2B TP=1	109.684 / 102.474	-6.57%	4.4316 / 4.3895	-0.95%	216.011 / 208.111	-3.66%	72.95% / 75.86%	+2.92
Qwen3.5 2B TP=2	80.684 / 77.053	-4.50%	4.2158 / 4.1789	-0.87%	181.795 / 177.279	-2.48%	76.24% / 73.46%	-2.77
Qwen3.6 27B TP=2	656.368 / 620.105	-5.52%	15.0632 / 15.1211	+0.38%	1017.942 / 983.063	-3.43%	65.40% / 64.60%	-0.80
Qwen3.6 27B TP=4	396.421 / 382.316	-3.56%	9.8737 / 10.0158	+1.44%	633.568 / 622.616	-1.73%	67.31% / 64.90%	-2.41
Qwen3.6 35B TP=2	212.947 / 195.842	-8.03%	6.9526 / 6.5053	-6.43%	379.921 / 352.005	-7.35%	90.63% / 92.53%	+1.90
Qwen3.6 35B TP=4	159.316 / 149.895	-5.91%	5.8632 / 5.5632	-5.12%	300.053 / 283.584	-5.49%	92.35% / 92.53%	+0.18

Validation

• clang-format 20.1.6: passed.
• git diff --check: passed.
• Six A/B performance scenarios: passed with 19 steady-state samples per run.