llama.cpp fuses matmuls for Qualcomm's Hexagon HMX chip

Two llama.cpp builds released September 2 target Qualcomm's Hexagon HMX chip and Vulkan batch inference. Build b10758 fuses the QKV and FFN matmuls that run on Hexagon HMX and folds MUL_MAT_ID into a new MUL_MAT_ID_NX operation. It also adds VA defragmentation, so large-dimension runs stop aborting on fragmented address space, per the b10758 release notes.

Build b10757, released the same day, fixes IQ3_S mat-vec kernels that previously handled only up to 4 columns. That cap had throttled batch inference for quantized models on phones. With that fixed, batch size 8 runs 5x faster.

Kernel fusion like this cut per-token overhead on CUDA hardware two years ago. b10758 and b10757 bring a version of that to phone silicon instead of datacenter GPUs. The release notes don't name which Snapdragon chips see the Hexagon HMX gains. They also don't say whether the VA defragmentation fix changes behavior on GPUs outside Qualcomm's line. It's not clear yet whether other Vulkan-based mobile runtimes will adopt the same batch-fusion pattern.