llama.cpp really memory inefficient for qwen context?

<!-- SC_OFF --><div class="md"><p>Hey all, </p> <p>im having a crisis that i just cant figure... </p> <p>i used Qwen3.6-27B-GGUF:UD-Q8_K_XL ever since it came out (on a DGX SPARK)

<!-- SC_OFF --><div class="md"><p>I have been using llama.cpp to run some models recently. For example, I've been running GLM-4.7-Flash with this command <code>.\llama-server.exe

<!-- 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

<!-- SC_OFF --><div class="md"><p>Running Llama-3-70B with long contexts and hitting HBM limits constantly. 128K context burns approx 40GB just for KV cache — half the H100 gone

<!-- SC_OFF --><div class="md"><p>I’ve been testing LM Studio on my RTX 5070 Ti (16GB) and Ryzen 9800X3D, running Unsloth Qwen3.5 35B (UD Q4_K_XL).</p> <p>Initially, I thought LM

<!-- SC_OFF --><div class="md"><p>I'm genuinely unsure if I'm doing something wrong, or if the model is just worse, despite the community enthusiasm. I tend to use it with pi

<!-- SC_OFF --><div class="md"><blockquote> <p>Wait, I need to be careful about the &quot;no_think&quot; tag in the system prompt. The system prompt says <code>/no_think</code>.

<!-- 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