Bump vendored XNNPACK to include qd8_bf16_qb4w subgraph FC (#21400)
## Summary Bumps the vendored XNNPACK submodule (`backends/xnnpack/third-party/XNNPACK`) from `1adaa7c` to `92a7ad5` (current `google/XNNPACK` master). The key change picked up in this range is [google/XNNPACK#10818](https://github.com/google/XNNPACK/pull/10818), which adds the **`qd8_bf16_qb4w` fully-connected path to the subgraph layer**. Before this, XNNPACK's subgraph layer could build `qd8_f32_qb4w` and `qd8_f16_qb4w` dynamic-quant FC subgraphs but had no bf16-scale variant, so a bf16 activation feeding an `8da4w` (int8-dynamic-activation / int4-weight, blockwise) linear could not be lowered/run through XNNPACK. This bump is a prerequisite for bf16 dynamic-quant (`8da4w`) delegation to XNNPACK; the ExecuTorch-side backend changes that consume it are in a follow-up PR. ## Test plan - Built ExecuTorch with the bumped submodule (x86, XNNPACK backend). - Exported `google/gemma-3-1b-it` with `--dtype bfloat16 --qlinear 8da4w --qembedding 8w` via `optimum-executorch` (XNNPACK recipe, custom SDPA + KV cache) — lowering succeeds and the FC nodes delegate as `qd8_bf16_qb4w`. - Runtime forward on the resulting `.pte` produces finite `bfloat16` logits with argmax matching the fp32 and bf16-no-quant baselines. cc @GregoryComer @digantdesai @cbilgin
J
Jacob Stevens committed
d1df0926bfea7cb435e6428951fbcb410a274d21
Parent: a88ef35
Committed by GitHub <noreply@github.com>
on 7/28/2026, 4:56:20 PM