xnnpack moe: support activation='silu'

[internetoftim]

What

Adds silu (x * sigmoid(x)) as a third accepted value for the moe custom op's flexbuffer activation option in the XNNPACK delegate kernel, alongside gelu and gelu_tanh. Four small hunks in tflite/delegates/xnnpack/moe_delegate_kernel.cc; purely additive — models not passing silu are byte-for-byte unaffected.

Why

DeepSeek-V3-architecture MoE checkpoints (e.g. Moonshot's Moonlight-16B-A3B and the Kimi-VL text tower) use SiLU expert MLPs. Today the kernel rejects activation='silu' at prepare time, which forces such models onto the sequential dense-fallback path (all experts computed per token) — roughly an order of magnitude more weight traffic per decode step for a 64-expert/top-6 model.

Everything else about DeepSeek-style routing already works with this kernel: its sigmoid-scaled, non-renormalized top weights are used exactly as passed (verified against a litert-torch export of a DeepseekV3ForCausalLM toy model — with a flexbuffer-patched activation, logits matched a GELU-substituted eager reference at 1.7e-5, confirming activation is the only semantic gap on the CPU path).

The SiLU helper follows the file's existing pattern of host-side scalar activation helpers (Gelu, GeluTanh), including the same TODO trajectory of eventually lowering to an xnn unary op.

The ML-Drift/GPU parser (ml_drift_delegate/delegate/composite/moe_experts_parser.cc) still accepts gelu only; this PR deliberately does not touch it since I can't validate the GPU expert body's activation from outside. Happy to extend if a maintainer confirms the GPU path can take SiLU.

Testing

• Verified against the flexbuffer parse path and activation dispatch; no test target exists for this kernel today.
• End-to-end: a deepseek_v3 model extension for litert-torch's export_hf path (parity-tested vs transformers at fp32) drives this op; with this patch, a SiLU MoE toy export runs on the stock Python interpreter. That work lives in https://github.com/internetoftim/litert-torch/tree/kimi-moonlight-port and is being prepared for a separate litert-torch PR.

🤖 Generated with Claude Code

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.