[zaneni6]

Summary

Follow-up to #922. These 4 kernels were still Chess-built, but Peano builds of them now exist. #922 missed them because they are listed under different names.

File	Before	After
Qwen2.5-VL-3B-Instruct-NPU2/vision_attention.xclbin	Chess	Peano build of the Qwen2.5-VL whole-array vision attention (same 32 AIE cores)
Qwen3.5-OMNI-NPU2/attn.xclbin	Chess (same kernel Qwen3.5 used before #922)	byte-identical to Qwen3.5-*-NPU2/attn.xclbin on main
Qwen3.5-OMNI-NPU2/conv.xclbin	Chess (same kernel Qwen3.5 used before #922)	byte-identical to Qwen3.5-*-NPU2/conv.xclbin on main
Qwen3.5-OMNI-NPU2/GateDeltaNet_prefill.xclbin	Chess (same kernel Qwen3.5 used before #922)	byte-identical to Qwen3.5-*-NPU2/GateDeltaNet_prefill.xclbin on main

src/xclbins goes from 226 Peano / 87+4 Chess to 230 Peano / 87 Chess.

The remaining Qwen3.5-OMNI kernels (layer, lm_head, mm, mm_high, vision_attn, vision_mm, window_attention) are unchanged, because no Peano build of them exists yet.

Verification

• All 4 files fingerprint as Peano (AIE core startup code), with the same core count as before.
• The three OMNI files are byte-identical to the Qwen3.5 kernels that chore(xclbins): vendor the current Peano kernel builds #922 shipped and that passed NPU CI.
• To check before merging: CI on qwen2.5vl-it:3b (image input) and on Qwen3.5-OMNI.

Please squash-merge.

🤖 Generated with Claude Code