Release b12086 · ggml-org/llama.cpp

qwen4exp: write the conv state tail without an intermediate cont (#31443)

build_conv_state_at makes a contiguous copy of the conv input tail
before writing it into the recurrent state, so every GDN layer issues
two device copies per generated token where one is enough. On
Qwen3.8-Flash-Next (36 GDN layers) that is 72 copies per token, now 36.
Speculative decoding adds one more per rollback slot. ggml_cpy reads
the strided view directly, as the shared build_conv_state already does,
so the extra copy is dropped.

Website: