[WebGPU] LinearAttention: increase tile_v when subgroups are available

[Jianhui Dai (daijh)]

Description

• Scale tile_v by 4x when subgroup is enabled and the vectorized dimension has enough columns, improving data reuse.
• Gate the tile_v expansion on seq_length >= 16, where the prefill benefit outweighs increased register pressure.
• Remove redundant zero-initialization of the state tile (WGSL default- initializes vars to zero).

Intel Panther Lake (xe-3lpg)

Model	Prefill	Baseline (TPS)	Optimized (TPS)	Change
Qwen3.5-0.8B	128	1534.30	1681.00	9.56%
Qwen3.5-0.8B	1024	3267.30	3917.60	19.90%
Qwen3.5-0.8B	4096	2864.50	3563.40	24.40%
Qwen3.5-2B	128	1295.60	1344.60	3.78%
Qwen3.5-2B	1024	2177.10	2344.00	7.67%
Qwen3.5-2B	4096	1942.60	2247.00	15.67%
Qwen3.5-4B	128	701.30	736.20	4.98%
Qwen3.5-4B	1024	946.90	1036.30	9.44%
Qwen3.5-4B	4096	824.10	912.00	10.67%

Motivation and Context

See above.

[Jianhui Dai (daijh)]

Guenther Schmuelling (@guschmue) Jiajia Qin (@qjia7) PTAL.

[Copilot]

Pull request overview

This PR optimizes the WebGPU LinearAttention kernel by increasing the V-dimension tiling when subgroup reductions are available (to improve data reuse), gating that expansion to longer sequences, and removing redundant per-invocation state zeroing in the WGSL shader.

Changes:

• Adjust tile_v selection logic and expand it (up to 4×) when subgroups are enabled and seq_length >= 16.
• Move subgroup_min_size determination earlier and pass it into the shader program configuration.
• Remove explicit zero-initialization of the per-thread state tile in the WGSL template.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
onnxruntime/contrib_ops/webgpu/bert/linear_attention.cc	Adjusts tile selection/expansion heuristics and subgroup configuration passed into the shader.
onnxruntime/contrib_ops/webgpu/bert/linear_attention.wgsl.template	Removes explicit state zero-initialization (relies on WGSL default initialization).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Jianhui Dai (daijh)]

Jiajia Qin (@qjia7) Thanks. Fixed the comment, please take another look.