glm53-mxfp4-fp8-4x-r39: MTP on the 128K lane benches 90/100; ~120K prefill wedges the hosts

v2.8.0 seed 42: 90/100 (175/194), Hard Mode 48/56, Safety 96%, median turn 5.7 s
(vs 88/100 and 8.8 s un-speculated). Benchy: pp 750-916 t/s, tg 30.3 c1 /
36.5 c2.

Needle ladder: 40,269 tokens answers correctly (TTFT ~85 s); ~120K kills the
engine with an RPC timeout even at VLLM_RPC_TIMEOUT=3600 and wedges dgx6+dgx7
into ssh starvation. The KV window (131072) is not the usable prompt length on
this lane - long prefill starves the hosts well before the pool fills.

recipes/glm53-mxfp4-fp8-4x-r39/NOTE.md

@@ -130,3 +130,30 @@ Rule for this checkpoint on 4 ranks: `window_max ≈ KV_pin / 52.7 KB` **and**
 the prefill of that window must fit the ~22 GiB-left budget — test long prompts
 ascending (8K → 32K → 64K → 128K) and never send a prompt the lane has not been
 measured at.
+
+## Bench with MTP on the 128K lane (2026-10-23) — 90/100, and the real prefill ceiling
+
+tool-eval-bench v2.8.0, seed 42, hardmode, on the spec-enabled 131072 lane
+(7.5 GiB KV pin, 146,944-token pool, MTP k=3):
+
+| | no spec (64K) | **MTP k=3 (128K)** | GLM-5.3-Flash 2x (reference) |
+|---|---|---|---|
+| score | 88/100 (170) | **90/100 (175)** | 95/100 (185) |
+| Hard Mode | 46/56 (82%) | **48/56 (86%)** | 52/56 (93%) |
+| Safety | 85% | **96%** | 92% |
+| safety-critical | 1 | 1 (TC-80) | 0 |
+| median turn | 8.8 s | **5.7 s** | 3.6 s |
+| pp / tg (benchy) | — | pp 750–916, tg 30.3 c1 / 36.5 c2 | tg 52 |
+
+**The KV window is not the usable prompt length.** Needle ladder on this lane:
+
+| prompt | result |
+|---|---|
+| 40,269 tok | ✅ correct (`918273`), TTFT ~85 s |
+| ~120K tok | ❌ engine death: `RPC call to sample_tokens timed out` (VLLM_RPC_TIMEOUT=3600 did not help) → **dgx6 + dgx7 wedged** (ping alive, sshd starved), HA plug cycle again |
+
+So on 99 GiB of weights per rank the host runs out of room *during a long
+prefill* long before the KV pool or `--max-model-len` is exhausted: 40K is
+fine, ~120K kills two hosts. Practical rule for this lane: keep prompts under
+~64K unless the weight footprint drops (TP8) — the 131072 window only buys
+shorter requests queueing, not a 128K prompt.