From 3034c8d389edaf4a131c30304eee76bb9aef5022 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Nicol=C3=B2=20Lucchesi?= <nicolo.lucchesi@mistral.ai>
Date: Wed, 15 Jul 2026 23:04:54 +0200
Subject: [PATCH] [CI][PD] Add optional/nightly DSv4 Disaggregated eval
 (#42310)

Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai>
---
 .buildkite/test_areas/disaggregated.yaml      | 22 +++++++++++++++++++
 .../nixl_integration/run_accuracy_test.sh     | 13 ++++++-----
 .../nixl_integration/test_accuracy.py         |  1 +
 3 files changed, 31 insertions(+), 5 deletions(-)

diff --git a/.buildkite/test_areas/disaggregated.yaml b/.buildkite/test_areas/disaggregated.yaml
index b25884a96f..4012f1ebd5 100644
--- a/.buildkite/test_areas/disaggregated.yaml
+++ b/.buildkite/test_areas/disaggregated.yaml
@@ -203,3 +203,25 @@ steps:
   commands:
     - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
     - bash v1/kv_connector/nixl_integration/run_multi_connector_edge_case_test.sh
+
+# P TP 4 - D DPEP 4 test case for DSv4-Flash
+- label: DSv4-Flash Disaggregated DP EP
+  key: dsv4-flash-disaggregated
+  timeout_in_minutes: 60
+  device: h200
+  optional: true
+  working_dir: "/vllm-workspace/tests"
+  num_devices: 8
+  env:
+    ENABLE_HMA_FLAG: "1"
+    DP_EP: "1"
+    GPU_MEMORY_UTILIZATION: "0.85"
+    PREFILLER_TP_SIZE: "4"
+    DECODER_TP_SIZE: "4"
+    PREFILL_BLOCK_SIZE: "256"
+    DECODE_BLOCK_SIZE: "256"
+    MODEL_NAMES: "deepseek-ai/DeepSeek-V4-Flash"
+    VLLM_SERVE_EXTRA_ARGS: "--trust-remote-code,--kv-cache-dtype,fp8"
+  commands:
+    - bash /vllm-workspace/.buildkite/scripts/install-kv-connectors.sh
+    - bash v1/kv_connector/nixl_integration/run_accuracy_test.sh
diff --git a/tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh b/tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh
index a357128d3c..77235582fe 100755
--- a/tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh
+++ b/tests/v1/kv_connector/nixl_integration/run_accuracy_test.sh
@@ -140,12 +140,13 @@ run_tests_for_model() {
   # Start prefill instances
   for i in $(seq 0 $((NUM_PREFILL_INSTANCES-1))); do
     # Calculate GPU ID - we'll distribute across available GPUs
-    GPU_ID=$((i % $(get_num_gpus)))
-    NEXT_GPU=${GPU_ID}
+    GPU_START=$((i % $(get_num_gpus)))
+    GPU_ID=$GPU_START
+    NEXT_GPU=$GPU_START
     # Reserve TP*PP GPUs for the prefiller (TP shards across PP stages).
     PREFILLER_WORLD_SIZE=$((PREFILLER_TP_SIZE * PREFILLER_PP_SIZE))
     for (( j=1; j < PREFILLER_WORLD_SIZE; j++ )); do
-      NEXT_GPU=$(((GPU_ID + j) % $(get_num_gpus)))
+      NEXT_GPU=$(((GPU_START + j) % $(get_num_gpus)))
       GPU_ID="${GPU_ID},${NEXT_GPU}"
     done
 
@@ -195,10 +196,12 @@ run_tests_for_model() {
   # Start decode instances
   for i in $(seq 0 $((NUM_DECODE_INSTANCES-1))); do
     # Calculate GPU ID - we'll distribute across available GPUs, starting from after prefill GPUs
-    GPU_ID=$(((i + NEXT_GPU + 1) % $(get_num_gpus)))
+    DECODE_START=$(((i + NEXT_GPU + 1) % $(get_num_gpus)))
+    GPU_ID=$DECODE_START
+    NEXT_GPU=$DECODE_START
     # If DECODER_TP_SIZE is more than 1
     for (( j=1; j < DECODER_TP_SIZE; j++ )); do
-      NEXT_GPU=$(((GPU_ID + j) % $(get_num_gpus)))
+      NEXT_GPU=$(((DECODE_START + j) % $(get_num_gpus)))
       GPU_ID="${GPU_ID},${NEXT_GPU}"
     done
     # Calculate port number (base port + instance number)
diff --git a/tests/v1/kv_connector/nixl_integration/test_accuracy.py b/tests/v1/kv_connector/nixl_integration/test_accuracy.py
index bb68b7a572..711f27d209 100644
--- a/tests/v1/kv_connector/nixl_integration/test_accuracy.py
+++ b/tests/v1/kv_connector/nixl_integration/test_accuracy.py
@@ -26,6 +26,7 @@ EXPECTED_VALUES = {
     "Qwen/Qwen3.5-0.8B": 0.33,
     "google/gemma-4-E2B-it": 0.485,
     "ai21labs/AI21-Jamba2-3B": 0.74,
+    "deepseek-ai/DeepSeek-V4-Flash": 0.95,
 }
 
 SIMPLE_PROMPT = (