Repository navigation

[Kernel][Z-Image] Fuse QK RMSNorm + RoPE with CUDA-matching numerics - #7597

Open

yuweih205 wants to merge 14 commits into

vllm-project:mainfrom

yuweih205:perf/z-image-fused-qk-norm-rope

Open

[Kernel][Z-Image] Fuse QK RMSNorm + RoPE with CUDA-matching numerics#7597
yuweih205 wants to merge 14 commits into
vllm-project:mainfrom
yuweih205:perf/z-image-fused-qk-norm-rope

Commits on Oct 8, 2026
- [Kernel] Extend shared QK/RoPE ops for joint model paths
  yuweih205
  committed
  bf9acac
- [Kernel] Preserve 64-bit offsets in fused QK/RoPE kernels
  yuweih205
  committed
  24f9e25
- [Kernel][Z-Image] Fuse QK RMSNorm and RoPE
  yuweih205
  committed
  02984be
- docs: list Z-Image fused QK RoPE token gate default
  yuweih205
  committed
  ffdb710
Commits on Oct 12, 2026
- fix: keep Z-Image eager numerics by default
  yuweih205
  committed
  e16b7b3
- refactor: trim Z-Image fusion comments and tests
  yuweih205
  committed
  c6cfb84
Commits on Oct 13, 2026
- [Kernel] Select fused QK/RoPE paths automatically
  yuweih205
  committed
  d71bc30
Commits on Oct 16, 2026
- [Kernel][Z-Image] Gate the CUDA-matching fusion in the packer, not per block
  yuweih205
  committed
  27ec9bb
Commits on Oct 17, 2026
- [Kernel][Z-Image] Fuse under torch.compile as the other models do
  yuweih205
  committed
  6d7042a
Commits on Oct 19, 2026
- [Kernel] Use eight heads per program in the interleaved fused Q/K kernels
  yuweih205
  committed
  9aab8a8
Commits on Oct 20, 2026