[system] You are a tool-using agent on DGX Spark. Ground answers in retrieved context, traces, and repo paths. Cite span IDs and sources when retrieval or tools are used. Prefer dense technical output; no filler. [user] INTELLECT-3-MoE FP4 is loaded on my Spark; nvtop shows about 52/128 GB at idle after weights. How much headroom is left for a 32k context prefill batch, and does KV or weights dominate first? [context: 4 chunks retrieved, 2 run configs] configs/intellect-3-moe-fp4.yaml — max_seq_len: 131072, tensor_parallel: 1, kv_cache_dtype: fp8 docs/spark-memory.md — unified 128 GB pool; note GB10 bandwidth vs HBM-class rigs experiments/bench-a8f2/moe-spark.json — prefill tok/s vs batch for MoE FP4 on Spark nous-docs/hermes-70b-kv.md — KV sizing heuristics (compare dense 70B vs MoE active params) [lab notes] Prime Intellect Lab: treat published run configs as authoritative for MoE expert routing overhead GPU MODE thread #4821: measured KV growth vs seq_len on GB10 (community numbers, verify locally)

At idle after load, the ~52 GB you see is overwhelmingly static weights (MoE + FP4 weights and allocator slack), not KV. KV is allocated with context; it scales roughly linearly with batch and sequence length once you start prefilling. Back-of-envelope for 32k prefill on one batch: Use your measured bytes/token from experiments/bench-a8f2 or run a short cudaMemGetInfo sweep after a single long prefill. With FP8 KV in the retrieved config, KV is often the crossing point before you exhaust 128 GB unified memory, but MoE active-parameter count sets attention FLOPs and effective residency. Order of operations: weights fixed (~52 GB here) plus KV growth as prefill advances, plus activation peaks during the long matmul stack. On Spark, bandwidth to unified memory matters as much as raw capacity — treat 273 GB/s as the ceiling called out in dgxmode hardware notes when you sanity-check sustained prefill. [4 sources grounded PASS]