[tawei-amd]

What

Replaces the two Qwen3.8-27B-MTP layer xclbins with builds of the q4_K projection engine on aie::mmul (internal IRON PR #135). Decode goes 1.77 -> 2.84 tok/s on a local NPU at unchanged MTP acceptance. The engine library is unchanged: no .so/.dll change, only the two kernels.

file	sha256	IRON artifact
src/xclbins/Qwen3.8-27B-NPU2/deltanet_layer.xclbin	5881abeed388f52ecb7c4be5f1b71c6580876bb81d02dae5906f8a52fda8d162	xclbins-Qwen3_8_gate_delta_net_verify
src/xclbins/Qwen3.8-27B-NPU2/attention_layer.xclbin	1caee645f17fc942911e629ba75d10209fed2a479f86f62c2a102aacbce576d3	xclbins-Qwen3_8_gated_attention_verify (layer.xclbin, renamed to the name the engine registers)

IRON PR Check run: 41264160514 (the vendored files are exactly that run's artifacts; no local builds committed).

Verification

• IRON CI green: xclbin builds, perf, GSM8K, DocVQA on XRT and HRX. CI decode for qwen3.8-mtp:27b: XRT +20..+80%, HRX +44..+115%; GSM8K 98.0 -> 98.0 (XRT), 96.0 -> 98.0 (HRX).
• Local, these exact files: layer bench vs the stock kernel on identical input, delta-net 23.7 -> 12.3 ms (cos 0.99995), attention 20.6 -> 9.5 ms (cos 0.9998).
• Local, production engine lib + these xclbins, six prompts each, temp 0, 160 tokens:

prompt set	stock tok/s (MTP accept len)	this PR tok/s (accept len)
English / code	1.77 (2.78)	2.84 (2.72)
multilingual / code / math	1.83 (2.65)	2.87 (2.56)

Output text diverges from stock after a few hundred characters (greedy decode with ~1-2% numeric noise in the layer), stays coherent.

Order

The IRON PR is still a draft. Per the usual order this should land after IRON merges; the kernel sources are not touched by the merge, but I will re-verify the merged run's artifacts against the sha256 above before this leaves draft.

🤖 Generated with Claude Code

[tawei-amd]

IRON#135 merged (5014861). Re-verified: the merge commit has no file differences from the PR head cf425f4 that run 41264160514 built, so these xclbins are the merged kernels. sha256 of the vendored files still match that run's artifacts (5881abee... deltanet, 1caee645... attention). Marking ready.

[tawei-amd]

Added the docs benchmark update: docs/docs/benchmarks/qwen3.8_results.md decode row 1.06/1.86/1.50/1.32 -> 2.04/2.23/2.70/2.31 tok/s (1k/2k/4k/8k) and a regenerated qwen38_decoding.png in the same style. Numbers are IRON PR Check run 41264160514 XRT perf-bench for this PR's kernels (the same bench the page's previous numbers match: its baseline gave 1.50/1.32 at 4k/8k and prefill within 0.2%). Prefill and TTFT are unchanged by this kernel and untouched.