[ExecuTorch][WebGPU] Route exact prefill shapes to the BK64 GEMM - #21649

Merged

JCNTH merged 13 commits into

gh/JCNTH/105/origfrom

gh/JCNTH/106/orig

Aug 7, 2026

Commits on Aug 7, 2026

[ExecuTorch][WebGPU] Route exact prefill shapes to the BK64 GEMM
JCNTH
committed
0d94262
[ExecuTorch][WebGPU] Replace broad QKV fusion with BK64 kernel
JCNTH
committed
710e9eb
[ExecuTorch][WebGPU] Add Llama K16 online causal attention
JCNTH
committed
8f138b6
[ExecuTorch][WebGPU] Select emdawnwebgpu for Emscripten builds
JCNTH
committed
203a49e
[ExecuTorch][WebGPU] Repair dynamic SDPA routing, add attestation
JCNTH
committed
a1a5f17
[ExecuTorch][WebGPU] Add HuggingFace rotate-half RoPE
JCNTH
committed
2246c4c
[ExecuTorch][WebGPU] Add Qwen3 K16 attention and fp16 KV cache
JCNTH
committed
e53fa18
[ExecuTorch][WebGPU] Make WGSL generation rollback-safe
JCNTH
committed
51eaf75
[ExecuTorch][WebGPU] Generate typed to_copy shader variants
JCNTH
committed
d6eefd8
[ExecuTorch][WebGPU] Generate extrema and unary shader variants
JCNTH
committed
977d553
[ExecuTorch][WebGPU] Generate logical and arithmetic binary shader variants
JCNTH
committed
be88c94
[ExecuTorch][WebGPU] Preserve Linear and Q4 embedding params across dynamic resize
JCNTH
committed
3a69f1c
[ExecuTorch][Vulkan] Preserve persistent buffer mutations (#21662)

pytorchbot
and
JCNTH
committed
0f65c9a