DarthAffe/stable-diffusion.cpp
1
0 Fork 0
mirror of https://github.com/leejet/stable-diffusion.cpp.git synced 2026-11-10 19:06:33 +00:00
Code Issues Packages Projects Releases Wiki Activity

perf(cuda): fuse RMSNorm and RoPE in Sol-H3-VAE (#2323)

Browse Source
This commit is contained in:
losewayy authored and GitHub committed 2026-11-11 01:06:53 +08:00
1 parent 465adf389f
commit f8417d02e1
3 files changed
+58 -10
Download Patch File Download Diff File Expand all files Collapse all files

No files matched your search

docs/ggml_extensions.md
+5 -3
View File
@@ -215,11 +215,13 @@ HIP, MUSA, and Vulkan. Other backends can use the scheduler's CPU fallback.
The former `ggml_rope_apply` API, builtin opcode, and CPU implementation have
been removed from ggml.
`src/core/cuda/rope.cu` supplies standalone CUDA RoPE
and `RMSNorm -> Mul -> RoPE [-> Scale -> FP16 Copy]` fusion. CuTe attention shares
`src/core/cuda/rope.cu` supplies standalone CUDA RoPE,
`RMSNorm -> RoPE` fusion for non-affine normalization, and
`RMSNorm -> Mul -> RoPE [-> Scale -> FP16 Copy]` fusion. CuTe attention shares
its RMSNorm/RoPE kernel. Partial rotation and arbitrary float-aligned input and
position-encoding strides are supported by the standalone kernel; RMSNorm fusion
requires contiguous position encoding and weights. Copy fusion uses workspace
requires contiguous position encoding and, when present, weights. The non-affine
fusion uses 64-channel F32 inputs and split-half RoPE. Copy fusion uses workspace
when the destination overlaps an input. These registrations do not require
CUTLASS. The old CUDA/HIP/MUSA RoPE dispatch is removed.