Built a system with four P100 GPUs.

<!-- SC_OFF --><div class="md"><p>Few days ago I posted about my <a href="https://www.reddit.com/r/LocalLLaMA/comments/1t94ito/i_have_deepseek_v4_pro_at_home/">DeepSeek V4 Pro</a>

<!-- 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/1sa060l/5950x_asrock_570m_pro_4_fully_headless_vllm_box/"> <img alt="5950x + ASrock 570m PRO 4 = fully

<!-- SC_OFF --><div class="md"><p>Got Qwen 3.5 27B to 1,103,941 tok/s on 12 nodes with 96 B200 GPUs.</p> <p>At that rate you process 50,000 insurance policy documents in hours

<!-- 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

<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>Every output token the model generates requires reading the entire KV cache from High Bandwidth Memory.</p> <p>Here's what that actually looks

<table> <tr><td> <a href="https://www.reddit.com/r/LocalLLaMA/comments/1sz0lyk/16x_dgx_sparks_what_should_i_run/"> <img alt="16x DGX Sparks - What should I run?"