]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
CUDA: fuse rms_norm + mul + rope (+ view + set_rows) (#26767)
authorRafail Giavrimis <redacted>
Sat, 8 Aug 2026 16:32:37 +0000 (17:32 +0100)
committerGitHub <redacted>
Sat, 8 Aug 2026 16:32:37 +0000 (00:32 +0800)
commit687e7789271ec1276e3470f158428e11a4f80b6f
treeaa97b9a3475b39a4efafdef3540af94452c08e4a
parent18f7ad7fc912444acc0f51995a4b8e45fd9a0cd4
CUDA: fuse rms_norm + mul + rope (+ view + set_rows) (#26767)

* CUDA: fuse rms_norm + mul + rope (+ view + set_rows)

* tests: add broadcast weight case to rms_norm_mul_rope

* CUDA: check memory ranges before rms_norm rope fusion

* CUDA: check memory ranges in rope set_rows fusion
ggml/src/ggml-cuda/ggml-cuda.cu
ggml/src/ggml-cuda/rope.cu
ggml/src/ggml-cuda/rope.cuh
tests/test-backend-ops.cpp