Fuse rms_norm, mul, quantize_q8_1#22710

Open

lnigam wants to merge 3 commits into

ggml-org:masterfrom

lnigam:fuse_rms_norm_mul_qunatize_q8_1

Commits on May 11, 2026

ggml-cuda : fuse RMS_NORM+MUL into Q8_1 output for MMVQ consumers - patch mul->type=Q8_1 in graph_optimize so gallocr allocates Q8_1 memory when all consumers are MMVQ-eligible; fuse norm+mul+quant…
lnigam
committed
clang formatting changes
lnigam
committed
register pressure in rms_norm_mul_q8_1_f32 kernel is getting increased to 64 registers for 1024 blocks failing test case on SM75. __launch_bounds__will take care of that
lnigam
committed