[ExecuTorch][WebGPU] Replace broad QKV fusion with BK64 kernel - #21650

Merged

JCNTH merged 12 commits into

gh/JCNTH/106/origfrom

gh/JCNTH/107/orig

Aug 7, 2026

Commits on Aug 7, 2026

[ExecuTorch][WebGPU] Replace broad QKV fusion with BK64 kernel
JCNTH
committed
cb607c5
[ExecuTorch][WebGPU] Add Llama K16 online causal attention
JCNTH
committed
2bd9d69
[ExecuTorch][WebGPU] Select emdawnwebgpu for Emscripten builds
JCNTH
committed
4c723f5
[ExecuTorch][WebGPU] Repair dynamic SDPA routing, add attestation
JCNTH
committed
e535bd3
[ExecuTorch][WebGPU] Add HuggingFace rotate-half RoPE
JCNTH
committed
26d4fb6
[ExecuTorch][WebGPU] Add Qwen3 K16 attention and fp16 KV cache
JCNTH
committed
2c9ca64
[ExecuTorch][WebGPU] Make WGSL generation rollback-safe
JCNTH
committed
f98d92e
[ExecuTorch][WebGPU] Generate typed to_copy shader variants
JCNTH
committed
4c55df4
[ExecuTorch][WebGPU] Generate extrema and unary shader variants
JCNTH
committed
1b73aa5
[ExecuTorch][WebGPU] Generate logical and arithmetic binary shader variants
JCNTH
committed
f8110c8
[ExecuTorch][WebGPU] Preserve Linear and Q4 embedding params across dynamic resize
JCNTH
committed
07b93e5
[ExecuTorch][Vulkan] Preserve persistent buffer mutations (#21662)

pytorchbot
and
JCNTH
committed
509ea95