Breeze-TTS 2 performance: weight packing + fused bf16 rounding - #393

Merged

0xShug0 merged 9 commits into

0xShug0:devfrom

IIIIIllllIIIIIlllll:breeze-perf-upstream

Sep 3, 2026

Commits on Sep 3, 2026

breeze: pack qkv and gate/up projection weights, document weight_type

IIIIIllllIIIIIlllll
authored and
0xShug0
committed
d6ed331
ggml, qwen_decoder: fuse bf16 activation rounding into a single kernel

IIIIIllllIIIIIlllll
authored and
0xShug0
committed
98797aa
ggml, breeze: support row-strided inputs in fused bf16 rounding

IIIIIllllIIIIIlllll
authored and
0xShug0
committed
2c74a99
ggml-cuda: allow CUDA graphs on pre-Ampere GPUs via GGML_CUDA_GRAPHS_PRE_AMPERE

IIIIIllllIIIIIlllll
authored and
0xShug0
committed
6d7d591
ggml, breeze: generalize fused bf16 rounding to f16/bf16 inputs

IIIIIllllIIIIIlllll
authored and
0xShug0
committed
db84ed1
conv_transpose1d: enable col2im fast path on Vulkan

IIIIIllllIIIIIlllll
authored and
0xShug0
committed
0d66e26
breeze: skip the unconditional branch when guidance_scale == 1

IIIIIllllIIIIIlllll
authored and
0xShug0
committed
6a3eb1f
ggml-vulkan: add bf16<->f32/f16 cpy pipelines

IIIIIllllIIIIIlllll
authored and
0xShug0
committed
a0cb169
breeze: round activations to bf16 on GPU backends to match reference

IIIIIllllIIIIIlllll
authored and
0xShug0
committed
330e347