From 15b548953590f2408e82e15bee6559fbc1f51990 Mon Sep 17 00:00:00 2001
From: yy-fighting <yy-fighting@noreply.gitee.com>
Date: Thu, 9 Jul 2026 10:01:51 +0000
Subject: [PATCH] dsv4_h200_vllm_tp_dp_matrix: use minimal official vLLM args,
 no EP, Y-only, exclude 1M

---
 .../tp2_dp4/results.json                      |  29 +++
 .../tp2_dp4/scenarios.tsv                     | 166 ++++++++++++++++++
 .../dsv4_h200_vllm_tp_dp_matrix/run_bench.sh  |  50 ++----
 .../start_vllm_dp.sh                          |  29 ++-
 4 files changed, 222 insertions(+), 52 deletions(-)
 create mode 100644 experiments/dsv4_h200_vllm_tp_dp_matrix/results/20260709-095536_smoke/tp2_dp4/results.json
 create mode 100644 experiments/dsv4_h200_vllm_tp_dp_matrix/results/20260709-095536_smoke/tp2_dp4/scenarios.tsv

diff --git a/experiments/dsv4_h200_vllm_tp_dp_matrix/results/20260709-095536_smoke/tp2_dp4/results.json b/experiments/dsv4_h200_vllm_tp_dp_matrix/results/20260709-095536_smoke/tp2_dp4/results.json
new file mode 100644
index 0000000..d101ad8
--- /dev/null
+++ b/experiments/dsv4_h200_vllm_tp_dp_matrix/results/20260709-095536_smoke/tp2_dp4/results.json
@@ -0,0 +1,29 @@
+{
+  "metadata": {
+    "experiment": "dsv4_h200_vllm_tp_dp_matrix_tp2_dp4",
+    "run_id": "20260709-095536_smoke",
+    "timestamp": "2026-07-09T09:55:43+00:00",
+    "model": "/data/models/DeepSeek-V4-Flash",
+    "backend": "vllm",
+    "engine": "vllm",
+    "hardware": "8x NVIDIA H200 143GB",
+    "accelerator": "NVIDIA H200",
+    "chip": "nvidia_h200",
+    "script": "experiments/dsv4_h200_vllm_tp_dp_matrix/run_bench.sh",
+    "env": "/data/user1/yy/envs/vllm",
+    "git_commit": "ec85ef1",
+    "git_dirty": "dirty",
+    "description": "H200 vLLM TP×DP matrix for DeepSeek-V4-Flash"
+  },
+  "config": {
+    "tp": 2,
+    "dp": 4,
+    "cuda_visible_devices": "0,1,2,3,4,5,6,7",
+    "backend": "vllm",
+    "server_start_script": "experiments/dsv4_h200_vllm_tp_dp_matrix/start_vllm_dp.sh",
+    "server_args_per_isl": {
+      "1024": "vllm serve /data/models/DeepSeek-V4-Flash --trust-remote-code --tensor-parallel-size 2 --kv-cache-dtype fp8 --max-model-len 5184 --max-num-seqs 136 --block-size 256 --gpu-memory-utilization 0.90 --tokenizer-mode deepseek_v4 --reasoning-parser deepseek_v4 --no-disable-hybrid-kv-cache-manager --disable-uvicorn-access-log --port 30030 --data-parallel-size 4 --data-parallel-size-local 4 --data-parallel-multi-port-external-lb "
+    }
+  },
+  "scenarios": []
+}
diff --git a/experiments/dsv4_h200_vllm_tp_dp_matrix/results/20260709-095536_smoke/tp2_dp4/scenarios.tsv b/experiments/dsv4_h200_vllm_tp_dp_matrix/results/20260709-095536_smoke/tp2_dp4/scenarios.tsv
new file mode 100644
index 0000000..a064e0a
--- /dev/null
+++ b/experiments/dsv4_h200_vllm_tp_dp_matrix/results/20260709-095536_smoke/tp2_dp4/scenarios.tsv
@@ -0,0 +1,166 @@
+mark	input_len	output_len	concurrency	num_prompts
+Y	1024	128	1	5
+Y	1024	128	22	110
+Y	1024	128	43	215
+Y	1024	128	65	325
+Y	1024	128	86	430
+Y	1024	128	107	535
+Y	1024	128	128	640
+Y	1024	256	1	5
+Y	1024	256	22	110
+Y	1024	256	43	215
+Y	1024	256	65	325
+Y	1024	256	86	430
+Y	1024	256	107	535
+Y	1024	256	128	640
+Y	1024	512	1	5
+Y	1024	512	22	110
+Y	1024	512	43	215
+Y	1024	512	65	325
+Y	1024	512	86	430
+Y	1024	512	107	535
+Y	1024	512	128	640
+Y	1024	1024	1	5
+Y	1024	1024	22	110
+Y	1024	1024	43	215
+Y	1024	1024	65	325
+Y	1024	1024	86	430
+Y	1024	1024	107	535
+Y	1024	1024	128	640
+Y	1024	2048	1	5
+Y	1024	2048	22	110
+Y	1024	2048	43	215
+Y	1024	2048	65	325
+Y	1024	2048	86	430
+Y	1024	2048	107	535
+Y	1024	2048	128	640
+Y	1024	4096	1	5
+Y	1024	4096	22	110
+Y	1024	4096	43	215
+Y	1024	4096	65	325
+Y	1024	4096	86	430
+Y	1024	4096	107	535
+Y	1024	4096	128	640
+Y	4096	128	1	5
+Y	4096	128	11	55
+Y	4096	128	22	110
+Y	4096	128	33	165
+Y	4096	128	43	215
+Y	4096	128	53	265
+Y	4096	128	64	320
+Y	4096	256	1	5
+Y	4096	256	11	55
+Y	4096	256	22	110
+Y	4096	256	33	165
+Y	4096	256	43	215
+Y	4096	256	53	265
+Y	4096	256	64	320
+Y	4096	512	1	5
+Y	4096	512	11	55
+Y	4096	512	22	110
+Y	4096	512	33	165
+Y	4096	512	43	215
+Y	4096	512	53	265
+Y	4096	512	64	320
+Y	4096	1024	1	5
+Y	4096	1024	11	55
+Y	4096	1024	22	110
+Y	4096	1024	33	165
+Y	4096	1024	43	215
+Y	4096	1024	53	265
+Y	4096	1024	64	320
+Y	4096	2048	1	5
+Y	4096	2048	11	55
+Y	4096	2048	22	110
+Y	4096	2048	33	165
+Y	4096	2048	43	215
+Y	4096	2048	53	265
+Y	4096	2048	64	320
+Y	4096	4096	1	5
+Y	4096	4096	11	55
+Y	4096	4096	22	110
+Y	4096	4096	33	165
+Y	4096	4096	43	215
+Y	4096	4096	53	265
+Y	4096	4096	64	320
+Y	16384	128	1	5
+Y	16384	128	6	30
+Y	16384	128	11	55
+Y	16384	128	17	85
+Y	16384	128	22	110
+Y	16384	128	27	135
+Y	16384	128	32	160
+Y	16384	256	1	5
+Y	16384	256	6	30
+Y	16384	256	11	55
+Y	16384	256	17	85
+Y	16384	256	22	110
+Y	16384	256	27	135
+Y	16384	256	32	160
+Y	16384	512	1	5
+Y	16384	512	6	30
+Y	16384	512	11	55
+Y	16384	512	17	85
+Y	16384	512	22	110
+Y	16384	512	27	135
+Y	16384	512	32	160
+Y	16384	1024	1	5
+Y	16384	1024	6	30
+Y	16384	1024	11	55
+Y	16384	1024	17	85
+Y	16384	1024	22	110
+Y	16384	1024	27	135
+Y	16384	1024	32	160
+Y	16384	2048	1	5
+Y	16384	2048	6	30
+Y	16384	2048	11	55
+Y	16384	2048	17	85
+Y	16384	2048	22	110
+Y	16384	2048	27	135
+Y	16384	2048	32	160
+Y	65536	128	1	5
+Y	65536	128	2	10
+Y	65536	128	3	15
+Y	65536	128	5	25
+Y	65536	128	6	30
+Y	65536	128	7	35
+Y	65536	128	8	40
+Y	65536	256	1	5
+Y	65536	256	2	10
+Y	65536	256	3	15
+Y	65536	256	5	25
+Y	65536	256	6	30
+Y	65536	256	7	35
+Y	65536	256	8	40
+Y	65536	512	1	5
+Y	65536	512	2	10
+Y	65536	512	3	15
+Y	65536	512	5	25
+Y	65536	512	6	30
+Y	65536	512	7	35
+Y	65536	512	8	40
+Y	65536	1024	1	5
+Y	65536	1024	2	10
+Y	65536	1024	3	15
+Y	65536	1024	5	25
+Y	65536	1024	6	30
+Y	65536	1024	7	35
+Y	65536	1024	8	40
+Y	131072	128	1	5
+Y	131072	128	2	10
+Y	131072	128	3	15
+Y	131072	128	4	20
+Y	131072	256	1	5
+Y	131072	256	2	10
+Y	131072	256	3	15
+Y	131072	256	4	20
+Y	131072	512	1	5
+Y	131072	512	2	10
+Y	131072	512	3	15
+Y	131072	512	4	20
+Y	262144	128	1	5
+Y	262144	128	2	10
+Y	262144	256	1	5
+Y	262144	256	2	10
+Y	524288	128	1	5
+Y	524288	128	2	10
diff --git a/experiments/dsv4_h200_vllm_tp_dp_matrix/run_bench.sh b/experiments/dsv4_h200_vllm_tp_dp_matrix/run_bench.sh
index 0df25b4..6e3ed6e 100755
--- a/experiments/dsv4_h200_vllm_tp_dp_matrix/run_bench.sh
+++ b/experiments/dsv4_h200_vllm_tp_dp_matrix/run_bench.sh
@@ -15,7 +15,7 @@ source "${SCRIPT_DIR}/config.env"
 RUN_ID="${RUN_ID:-$(date '+%Y%m%d-%H%M%S')}"
 RESULT_BASE="${SCRIPT_DIR}/results"
 MATRIX_FILE="${MATRIX_FILE:-${SCRIPT_DIR}/matrix.json}"
-MATRIX_MODE="${MATRIX_MODE:-Y+P}"
+MATRIX_MODE="${MATRIX_MODE:-Y}"
 SCENARIO_TIMEOUT_S="${SCENARIO_TIMEOUT_S:-1800}"
 GPU_MEM_SAMPLE_INTERVAL_S="${GPU_MEM_SAMPLE_INTERVAL_S:-1}"
 
@@ -72,22 +72,14 @@ stop_server() {
 build_server_args() {
   local tp="$1"
   local dp="$2"
-  local max_model_len="$3"
-  local max_num_seqs="$4"
 
   local args=(
     "vllm serve" "$MODEL_PATH"
     --trust-remote-code
-    --tensor-parallel-size "$tp"
     --kv-cache-dtype fp8
-    --max-model-len "$max_model_len"
-    --max-num-seqs "$max_num_seqs"
     --block-size 256
-    --gpu-memory-utilization 0.90
-    --tokenizer-mode deepseek_v4
-    --reasoning-parser deepseek_v4
-    --no-disable-hybrid-kv-cache-manager
-    --disable-uvicorn-access-log
+    --tensor-parallel-size "$tp"
+    --no-enable-flashinfer-autotune
     --port "$VLLM_PORT"
   )
   if [[ "$dp" -gt 1 ]]; then
@@ -103,11 +95,9 @@ build_server_args() {
 start_server() {
   local tp="$1"
   local dp="$2"
-  local max_model_len="$3"
-  local max_num_seqs="$4"
 
-  log "starting vllm server tp=${tp} dp=${dp} max_model_len=${max_model_len} max_num_seqs=${max_num_seqs}"
-  bash "${SCRIPT_DIR}/start_vllm_dp.sh" "$tp" "$dp" "$max_model_len" "$max_num_seqs" \
+  log "starting vllm server tp=${tp} dp=${dp}"
+  bash "${SCRIPT_DIR}/start_vllm_dp.sh" "$tp" "$dp" \
     >> "${log_dir_global}/vllm_tp${tp}_dp${dp}.server.outer.log" 2>&1
 
   local hport
@@ -244,22 +234,23 @@ run_parallel_config() {
     "$VENV_VLLM" \
     "H200 vLLM TP×DP matrix for DeepSeek-V4-Flash"
 
-  # Embed static config now; server_args will be updated per ISL group.
-  jq --arg tp "$tp" --arg dp "$dp" --arg cuda "$CUDA_VISIBLE_DEVICES" \
+  # Embed static config and the exact server command line used.
+  local server_args_str
+  server_args_str="$(build_server_args "$tp" "$dp")"
+  jq --arg tp "$tp" --arg dp "$dp" --arg cuda "$CUDA_VISIBLE_DEVICES" --arg args "$server_args_str" \
     '.config = {
        "tp": ($tp | tonumber),
        "dp": ($dp | tonumber),
        "cuda_visible_devices": $cuda,
        "backend": "vllm",
-       "server_start_script": "experiments/'${EXPERIMENT_NAME}'/start_vllm_dp.sh"
+       "server_start_script": "experiments/'${EXPERIMENT_NAME}'/start_vllm_dp.sh",
+       "server_args": $args
      }' "${result_root}/results.json" > "${result_root}/results.json.tmp" && \
     mv "${result_root}/results.json.tmp" "${result_root}/results.json"
 
-  # Group scenarios by input_len.  We start one server per input_len with a
-  # max_model_len large enough for the biggest output_len in that group.
+  # Group scenarios by input_len.  We restart the server for each ISL to keep
+  # memory state isolated and to warm up per context length.
   local current_isl=""
-  local max_dsl_for_isl=0
-  local max_conc_for_isl=0
   local group_started=false
 
   tail -n +2 "$scenario_tsv" | while IFS=$'\t' read -r mark isl dsl conc num; do
@@ -269,28 +260,15 @@ run_parallel_config() {
         stop_server "$tp" "$dp"
       fi
       current_isl="$isl"
-      max_dsl_for_isl="$(awk -F'\t' -v isl="$isl" '$2==isl {if($3>max) max=$3} END{print max+0}' "$scenario_tsv")"
-      max_conc_for_isl="$(awk -F'\t' -v isl="$isl" '$2==isl {if($4>max) max=$4} END{print max+0}' "$scenario_tsv")"
-
-      local max_model_len=$((isl + max_dsl_for_isl + 64))
-      local max_num_seqs=$((max_conc_for_isl + 8))
 
       stop_server "$tp" "$dp"
-      if ! start_server "$tp" "$dp" "$max_model_len" "$max_num_seqs"; then
+      if ! start_server "$tp" "$dp"; then
         log "ERROR: ${config_label} failed to start for ISL=${isl}; skipping this group"
         group_started=false
         continue
       fi
       group_started=true
 
-      # Update metadata with the actual server args for this ISL group.
-      local server_args_str
-      server_args_str="$(build_server_args "$tp" "$dp" "$max_model_len" "$max_num_seqs")"
-      jq --arg isl "$isl" --arg args "$server_args_str" \
-        '.config.server_args_per_isl += {($isl): $args}' \
-        "${result_root}/results.json" > "${result_root}/results.json.tmp" && \
-        mv "${result_root}/results.json.tmp" "${result_root}/results.json"
-
       # Warmup with the shortest output for this ISL.
       run_warmup "$tp" "$dp" "$isl" 128
     fi
diff --git a/experiments/dsv4_h200_vllm_tp_dp_matrix/start_vllm_dp.sh b/experiments/dsv4_h200_vllm_tp_dp_matrix/start_vllm_dp.sh
index 2ae2fee..a4e52e0 100755
--- a/experiments/dsv4_h200_vllm_tp_dp_matrix/start_vllm_dp.sh
+++ b/experiments/dsv4_h200_vllm_tp_dp_matrix/start_vllm_dp.sh
@@ -1,12 +1,18 @@
 #!/usr/bin/env bash
 # Start vLLM server for a given TP×DP configuration.
-# Usage: start_vllm_dp.sh <TP> <DP> <max_model_len> <max_num_seqs>
+# Usage: start_vllm_dp.sh <TP> <DP>
+#
+# Uses the minimal argument set from the official DeepSeek-V4-Flash recipe:
+#   --trust-remote-code
+#   --kv-cache-dtype fp8
+#   --block-size 256
+#   --tensor-parallel-size <TP>
+#   --no-enable-flashinfer-autotune
+# plus DP-related flags when DP > 1.
 set -e
 
 TP="${1}"
 DP="${2}"
-MAX_MODEL_LEN="${3}"
-MAX_NUM_SEQS="${4}"
 
 SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 # shellcheck source=/dev/null
@@ -26,20 +32,13 @@ PID_FILE="/data/user1/yy/${EXPERIMENT}_vllm_tp${TP}_dp${DP}.pid"
 
 rm -f "$PID_FILE"
 
-# Build the base command as an array.
 SERVER_ARGS=(
     vllm serve "$MODEL_PATH"
     --trust-remote-code
-    --tensor-parallel-size "$TP"
     --kv-cache-dtype fp8
-    --max-model-len "$MAX_MODEL_LEN"
-    --max-num-seqs "$MAX_NUM_SEQS"
     --block-size 256
-    --gpu-memory-utilization 0.90
-    --tokenizer-mode deepseek_v4
-    --reasoning-parser deepseek_v4
-    --no-disable-hybrid-kv-cache-manager
-    --disable-uvicorn-access-log
+    --tensor-parallel-size "$TP"
+    --no-enable-flashinfer-autotune
     --port "$VLLM_PORT"
 )
 
@@ -49,9 +48,7 @@ if [[ "$DP" -gt 1 ]]; then
         --data-parallel-size-local "$DP"
         --data-parallel-multi-port-external-lb
     )
-    # vLLM 0.24.0 hard-codes the DP supervisor port to 9256 in
-    # entrypoints/openai/cli_args.py.  The bench client and health checks must
-    # target that port when DP > 1.
+    # vLLM 0.24.0 hard-codes the DP supervisor port to 9256.
     HEALTH_PORT="$VLLM_DP_SUPERVISOR_PORT"
 else
     HEALTH_PORT="$VLLM_PORT"
@@ -59,7 +56,7 @@ fi
 
 SERVER_ARGS_STR="${SERVER_ARGS[*]}"
 
-echo "=== Starting vLLM server (TP=${TP}, DP=${DP}, max_model_len=${MAX_MODEL_LEN}, max_num_seqs=${MAX_NUM_SEQS}) ==="
+echo "=== Starting vLLM server (TP=${TP}, DP=${DP}) ==="
 echo "Model: $MODEL_PATH"
 echo "Health port: $HEALTH_PORT"
 echo "Command: $SERVER_ARGS_STR"