llama.cpp B9387 Significant AMD/ROCm PP Update

<!-- SC_OFF --><div class="md"><p>Might be another option of us local LLM folks. I am very curious on the price. </p> <p><a

<!-- SC_OFF --><div class="md"><p>Just ran some llama-bench comparisons between ROCm and Vulkan backends on my Strix Halo system. Vulkan came out ahead, which surprised me.</p>

<!-- SC_OFF --><div class="md"><p><a href="https://www.phoronix.com/news/ROCm-7.13-Released">https://www.phoronix.com/news/ROCm-7.13-Released</a> Quote: ...new optimizations for

<!-- SC_OFF --><div class="md"><p>We know that using ROCm is a disgrace compared to CUDA, especially if you don't have &quot;current&quot; hardware. </p> <p>Guess what? I just

<!-- SC_OFF --><div class="md"><p>So i recently updated LMstudio after a long pause and updated my llama.cpp runtimes too.. i was shocked.. i thought maybe something like

<!-- SC_OFF --><div class="md"><p>b9095 finally makes -sm tensor work on dual consumer Blackwell PCIe GPUs without NCCL</p> <p>If youre on dual Blackwell gpus this look like it

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1t4uj9h/mtp_on_strix_halo_with_llamacpp_pr_22673/"> <img alt="MTP on strix halo with llama.cpp (PR #22673)"

<!-- SC_OFF --><div class="md"><p>PR #22673 (commit 4f13cb7) landed MTP speculative decoding in mainline llama.cpp on May 16. I tested it on two separate rigs.</p> <p>Qwen3.6 27B,