2 台 DGX Spark 跑 V4.1 Flash 1017 四会话：248k 提示词零报错
本地部署记录：V4.1 Flash 1017（flash17 2.8bpw）在 2 台 DGX Spark 上开 4 个 524k 上下文会话，调低 prefill 分块后 248k 提示词跑通、0 错误。
No problem! Keen to contribute where I can.. I had some issues with other components taking up RAM on Node 1. I disabled hermes and this gave me some more room. But, straight from my agent: WHAT IT TOOK — DeepSeek-V4.1-Flash-1017 (flash17 2.8bpw) on 2x DGX Spark, 4 sessions Config (config/prod-zig-flash17.env) — diffs from the shipped profile: TF_DSV41_CONTEXT=524288 # was 1048576 — THIS is the four-session change TF_DSV41_SLOTS=4 TF_DSV41_POOL_TOKENS=2100000 # 4 x 524,288 = 2,097,152, so 4 full windows TF_DSV41_PREFILL_CHUNK=2048 # was 4096 TF_DSV41_PREFILL_ROWS=2048 # was 4096 TF_DSV41_PF_4K=0 # drop the 4K prefill workspace TF_DSV41_FLOOR_GIB=7 # was 5 TF_DSV41_GRAPHS_MAX=24 # default 48; I had seen it hold 96 TF_DSV41_VISION_MAX_IMAGES=32 On both hosts: earlyoom -m 2,1 -r 60 # wedge insurance no-swap cgroup slice for the rank containers (MemorySwapMax=0) keep the serving node otherwise idle — no long agent turns on the head Why the memory lines matter: The low point is the PREFILL WORKSPACE, not the weights or the KV pool. A 4K chunk allocates a ~2.27 GiB transient on BOTH ranks; at 2K it's ~1.1 GiB. Result: 248k-token prompt OK — MemAvailable bottoms at 7.93 GiB (3.49 GiB when it OOM'd), 0 errors, 2.1M KV pool intact.