[LeON-Nie-code]

Overview

Add WebGPU support for F32 MEAN with contiguous rows, including offset views and strided higher dimensions. Reuse the row-sum shader and correct its output indexing for 2D dispatch.

Includes correctness and performance tests, plus updated operation support documentation. Related to #14909.

Additional information

Tested on NVIDIA L4 with native WebGPU.

test-backend-ops test -b WebGPU -o MEAN,SUM_ROWS,SUM
test-backend-ops perf -b WebGPU -o MEAN

• All 17 MEAN cases passed against the CPU reference; all 38 supported reduction cases passed together.
• Covers sliced/permuted inputs, reduction boundaries and 65,537-row dispatch. The large-row SUM_ROWS regression fails on upstream and passes with this change.
• Release build, local regression tests, clang-format 22 and EditorConfig checks passed.
• WikiText perplexity, 32 chunks at context 512: E2B 29.6885 -> 29.6811; E4B 27.7562 -> 27.7657.

Full-model llama-bench results with Gemma3n E2B/E4B Q4_K_M. Five runs per version, alternating execution order, with in-process warmup. Both versions used the same weights, two CPU threads, batch/microbatch 512 and flash attention off. Model loading and warmup are excluded. Times are mean +/- SD for separate prompt-processing and decoding workloads.

Model	Test	Upstream (s)	PR (s)	Time reduction
Gemma3n E2B	pp512	0.744 +/- 0.004	0.674 +/- 0.003	9.4%
Gemma3n E2B	tg128	4.201 +/- 0.112	3.669 +/- 0.248	12.7%
Gemma3n E4B	pp512	1.137 +/- 0.005	0.975 +/- 0.005	14.3%
Gemma3n E4B	tg128	4.769 +/- 0.168	3.992 +/- 0.250	16.3%

Both models execute MEAN in the first 10 FFN layers. Traces confirm all 1,280 calls in 128 decode steps move from CPU to WebGPU.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.

[yomaytk]

WebGPU changes look fine, thanks! Also, fixing the existing bug in sum_rows shader is nice.