CUDA on one DGX Spark (GB10, 128 GB): Qwen3.8-Flash-Next NVFP4 with --parallel 4 --context 131072 — host swaps 22 GB from 3 concurrent streams although the startup estimate was 81 GiB #750

Line: Python server 0.6.3 (the CUDA engine; we plan to re-measure on 0.6.6 before you look at this —
say the word if that changes anything for you).
Hardware: one NVIDIA DGX Spark (GB10, 128 GB unified memory), weights on local NVMe.
Checkpoint: local-inference-lab/Qwen3.8-Flash-Next-NVFP4 (106 GB), the card made for this box.
Command (production):

tensorfold serve <checkpoint-dir> --host 127.0.0.1 --port <port> \
  --name qwen3.8-flash-next --backend cuda --parallel 4 --context 131072 --no-update-check

What we see

• One stream decodes ~28 tok/s, which is great.
• The server's own startup line estimates 81.09 GiB within an 84.36 GiB budget
  (CUDA rank 0 startup estimate … native 262144, allocated prompt/reply window 131072, cache slots …)
  — i.e. it believes it fits, and on a 128 GB box "fits" should mean fits together with everything else.
• From 3 concurrent streams the system starts swapping hard: free -h showed 111 GiB used of
  121 GiB and vmstat/swap showed ~22 GB swapped with 4 slots resident. At 3 streams a full
  agent session (several long contexts) thrashes; the four advertised slots are effectively 2.
• We therefore run the box at 2 real streams and advertise 4. The gap between the startup estimate and
  the 3-stream reality is the whole issue: the estimate covers the engine's device side, but on GB10 the
  host page cache, the NVFP4 tables kept memory-mapped (PLE n-gram tables stay mmap'd — see the
  --ple-on-ssd startup refusal) and the KV windows all land in the same 128 GB pool.

What we'd like to know / suggest

1. Is --parallel 4 --context 131072 on one 128 GB Spark a supported shape, or is the honest
   guidance --parallel 2 (or one --context reduction) there? A "what fits where" table for GB10
   (128 GB) per checkpoint class would make this answerable without burning nights — minimum NVIDIA card class & Apple RAM? (documentation request) #139 asked for a
   minimum-hardware table; this is the same question from the other end.
2. Could admission control watch host memory too, not just the engine's device budget? On unified
   memory, refusing a third stream when available RAM drops under a floor is much kinder than swapping:
   our clients see a 60–80 s TTFT and a 3× decode slowdown instead of a clean rejection.
3. A cheap start: the startup estimate line could print the host available memory alongside the
   device estimate (we had to free -h by hand to find the 9.7 GiB that were really left), and warn
   when parallel × context cannot plausibly coexist with it.

Close relative: #23 (closed) is the 2×Spark GLM-5.3-Flash OOM at context ≥ 65 536 — same unified-memory
theme, different family and single box. Happy to re-run the measurement matrix on 0.6.6 with your flags
of choice; we have the harness (a temperature-0 streaming probe measuring prefill, decode and 2-stream
aggregate per request).