FP8 GEMM Optimization on AMD CDNA4 Architecture

<!-- SC_OFF --><div class="md"><p>New blog post by Daniel Vega-Myhre (Meta/PyTorch) illustrating GEMM design for FP8, including deep-dives into all the constraints and design

<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)"

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1tjdz92/amd_powers_nextgeneration_agent_computers_with/"> <img alt="AMD Powers Next-Generation Agent

<!-- SC_OFF --><div class="md"><p>You call torch.matmul(A, B) every day. A naive version of it barely uses ~1% of your GPU. This series shows why, and how to fix it. </p>

<!-- SC_OFF --><div class="md"><p>Wrote a deep dive on <strong>FlashAttention-4 (03/05/2026)</strong> that's relevant for anyone thinking about inference performance.</p>

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1s71d3f/optimize_moe_gemv_kernel_for_bs_1_by_gaugargnv/"> <img alt="Optimize MOE GEMV kernel for BS &gt; 1.