From 303e61e02a831baa94a5d802a5f5285880d9e674 Mon Sep 17 00:00:00 2001
From: chenfeiz0326 <203214996+chenfeiz0326@users.noreply.github.com>
Date: Sun, 25 Oct 2026 08:06:25 -0700
Subject: [PATCH] [https://nvbugs/6826000][fix] Size the DeepSeek-V4-Pro AgentX
 ctx KV pools for the compressed history

The ctx servers of both DeepSeek-V4-Pro DSpark AgentX e2e lanes split their KV
cache 0.55 / 0.22 / 0.23 over the DeepSeek-V4 life cycles: the SWA windows, the
compressed history with the indexer keys, and the compressor state windows.
With per_conversation reuse only the history grows with the retained
conversations, so its pools filled and dropped cached prefixes while ~150 GiB
per rank of the windowed host pools stayed unused. Attention-DP ranks whose
queue wait outlived that short retention re-prefilled every waiting prompt and
stayed collapsed; each run collapsed a random 8-10 of 16 ranks (con1156:
6.90-7.70 req/s, CV 4.8 %).

Give the history 64 % of the pools. con1156 over four node sets: 29.97-30.41
req/s (CV 0.7 %), prefix reuse 85 % -> 95 %, no collapsed rank.

Signed-off-by: chenfeiz0326 <203214996+chenfeiz0326@users.noreply.github.com>
---
 ..._con1156_ctx2_dep8_gen1_dep8_eplb0_dspark3_ccb-NIXL.yaml | 6 +++++-
 ...con1456_ctx3_dep8_gen1_dep16_eplb0_dspark5_ccb-NIXL.yaml | 6 +++++-
 2 files changed, 10 insertions(+), 2 deletions(-)

diff --git a/tests/scripts/perf-sanity/disaggregated/gb300_deepseek-v4-pro-dspark_agentx_con1156_ctx2_dep8_gen1_dep8_eplb0_dspark3_ccb-NIXL.yaml b/tests/scripts/perf-sanity/disaggregated/gb300_deepseek-v4-pro-dspark_agentx_con1156_ctx2_dep8_gen1_dep8_eplb0_dspark3_ccb-NIXL.yaml
index 63dfd8b70166..0999a242dc14 100644
--- a/tests/scripts/perf-sanity/disaggregated/gb300_deepseek-v4-pro-dspark_agentx_con1156_ctx2_dep8_gen1_dep8_eplb0_dspark3_ccb-NIXL.yaml
+++ b/tests/scripts/perf-sanity/disaggregated/gb300_deepseek-v4-pro-dspark_agentx_con1156_ctx2_dep8_gen1_dep8_eplb0_dspark3_ccb-NIXL.yaml
@@ -114,7 +114,11 @@ worker_config:
       block_reuse_config:
         policy: per_conversation
         max_num_turns: 5
-      pool_ratio: [0.55, 0.22, 0.23]
+      # DeepSeek-V4 KV cache manager v2 life cycles, in order: SWA windows, the compressed
+      # history with the indexer keys, compressor state windows. With per_conversation
+      # reuse only the history grows with the retained conversations, so it takes most of
+      # the GPU and host pools.
+      pool_ratio: [0.18, 0.64, 0.18]
       host_cache_size: 193273528320
     moe_config:
       backend: CUTEDSL
diff --git a/tests/scripts/perf-sanity/disaggregated/gb300_deepseek-v4-pro-dspark_agentx_con1456_ctx3_dep8_gen1_dep16_eplb0_dspark5_ccb-NIXL.yaml b/tests/scripts/perf-sanity/disaggregated/gb300_deepseek-v4-pro-dspark_agentx_con1456_ctx3_dep8_gen1_dep16_eplb0_dspark5_ccb-NIXL.yaml
index ddef64d8567e..f94a1aa283b3 100644
--- a/tests/scripts/perf-sanity/disaggregated/gb300_deepseek-v4-pro-dspark_agentx_con1456_ctx3_dep8_gen1_dep16_eplb0_dspark5_ccb-NIXL.yaml
+++ b/tests/scripts/perf-sanity/disaggregated/gb300_deepseek-v4-pro-dspark_agentx_con1456_ctx3_dep8_gen1_dep16_eplb0_dspark5_ccb-NIXL.yaml
@@ -114,7 +114,11 @@ worker_config:
       block_reuse_config:
         policy: per_conversation
         max_num_turns: 5
-      pool_ratio: [0.55, 0.22, 0.23]
+      # DeepSeek-V4 KV cache manager v2 life cycles, in order: SWA windows, the compressed
+      # history with the indexer keys, compressor state windows. With per_conversation
+      # reuse only the history grows with the retained conversations, so it takes most of
+      # the GPU and host pools.
+      pool_ratio: [0.18, 0.64, 0.18]
       host_cache_size: 193273528320
     moe_config:
       backend: CUTEDSL