From 455a78161b3645b10991da7c18bcaffa578a0e19 Mon Sep 17 00:00:00 2001
From: shishi <shishi@meta-stone.com>
Date: Wed, 29 Jul 2026 15:28:54 +0800
Subject: [PATCH] =?UTF-8?q?fix(910c/glm52):=20=E5=AF=B9=E9=BD=90=E5=AE=98?=
 =?UTF-8?q?=E6=96=B9A3=E6=95=99=E7=A8=8B=E5=8F=82=E6=95=B0(=E4=BF=AEDP=20d?=
 =?UTF-8?q?ie=E5=88=86=E9=85=8D=E9=97=AE=E9=A2=98,=E5=90=8Cdsv4=2099a22f0)?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

dsv4实验发现DP副本绑定到同一组die的问题(99a22f0),glm52存在相同问题:
缺--max-num-batched-tokens和--api-server-count导致DP worker设备分配异常
(不加api-server-count时vllm为N个DP rank启动N个API server)。

对齐docs.vllm.ai GLM5.2 A3官方教程:
- 加 --max-num-batched-tokens 8192 (官方值,影响DP调度)
- 加 --api-server-count 1 (官方值,避免多API server干扰设备分配)
- 去掉 --kv-cache-dtype fp8 (官方不指定,用默认bfloat16;且此镜像fp8本就未生效)
- 保留 --trust-remote-code / --enable-expert-parallel / enable_dsa_cp (官方有)
---
 experiments/910c/glm52_910c_vllm_tp_dp_matrix/config.env   | 7 ++++++-
 .../run_adaptive_concurrency_add16.sh                      | 3 ++-
 .../910c/glm52_910c_vllm_tp_dp_matrix/start_vllm_docker.sh | 3 ++-
 3 files changed, 10 insertions(+), 3 deletions(-)

diff --git a/experiments/910c/glm52_910c_vllm_tp_dp_matrix/config.env b/experiments/910c/glm52_910c_vllm_tp_dp_matrix/config.env
index 681a2dd..e2ab628 100644
--- a/experiments/910c/glm52_910c_vllm_tp_dp_matrix/config.env
+++ b/experiments/910c/glm52_910c_vllm_tp_dp_matrix/config.env
@@ -73,10 +73,15 @@ fi
 #   - MAX_MODEL_LEN: GLM-5.2 supports up to 128K context; cap at 131072.
 #   - gpu-memory-utilization maps to NPU HBM fraction on vllm-ascend (0.9 mirrors H20).
 GPU_MEMORY_UTILIZATION="${GPU_MEMORY_UTILIZATION:-0.95}"
-KV_CACHE_DTYPE="${KV_CACHE_DTYPE:-fp8}"
+KV_CACHE_DTYPE="${KV_CACHE_DTYPE:-}"
 BLOCK_SIZE="${BLOCK_SIZE:-128}"
 MAX_MODEL_LEN="${MAX_MODEL_LEN:-131072}"
 MAX_NUM_SEQS="${MAX_NUM_SEQS:-256}"
+# Official A3 tutorial params (docs.vllm.ai GLM5.2): max-num-batched-tokens=8192,
+# api-server-count=1 (without it, vllm spawns N API servers for N DP ranks,
+# disturbing DP worker device placement -- same issue as dsv4 fix 99a22f0).
+MAX_NUM_BATCHED_TOKENS="${MAX_NUM_BATCHED_TOKENS:-8192}"
+API_SERVER_COUNT="${API_SERVER_COUNT:-1}"
 
 # Per-TP parameter overrides (applied in start_vllm_docker.sh via case $TP).
 # GLM-5.2-w4a8c8 ~391 GiB total; with expert-parallel expert weights are sharded
diff --git a/experiments/910c/glm52_910c_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh b/experiments/910c/glm52_910c_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh
index bc0cc76..8312b3c 100755
--- a/experiments/910c/glm52_910c_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh
+++ b/experiments/910c/glm52_910c_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh
@@ -107,13 +107,14 @@ engine_build_server_args() {
     vllm serve "$MODEL_PATH"
     --served-model-name "$SERVED_MODEL_NAME"
     --trust-remote-code
-    --kv-cache-dtype "$KV_CACHE_DTYPE"
     --block-size "$BLOCK_SIZE"
     --tensor-parallel-size "$tp"
     --enable-expert-parallel
     --gpu-memory-utilization "$mem_util"
     --max-model-len "$max_len"
     --max-num-seqs "$max_seqs"
+    --max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS"
+    --api-server-count "$API_SERVER_COUNT"
     --compilation-config "$compilation_config"
     --additional-config "$additional_config"
     --host 0.0.0.0
diff --git a/experiments/910c/glm52_910c_vllm_tp_dp_matrix/start_vllm_docker.sh b/experiments/910c/glm52_910c_vllm_tp_dp_matrix/start_vllm_docker.sh
index 0bcad29..4641b38 100755
--- a/experiments/910c/glm52_910c_vllm_tp_dp_matrix/start_vllm_docker.sh
+++ b/experiments/910c/glm52_910c_vllm_tp_dp_matrix/start_vllm_docker.sh
@@ -55,13 +55,14 @@ SERVER_ARGS=(
   "$MODEL_PATH"
   --served-model-name "$SERVED_MODEL_NAME"
   --trust-remote-code
-  --kv-cache-dtype "$KV_CACHE_DTYPE"
   --block-size "$BLOCK_SIZE"
   --tensor-parallel-size "$TP"
   --enable-expert-parallel
   --gpu-memory-utilization "$GPU_MEMORY_UTILIZATION"
   --max-model-len "$MAX_MODEL_LEN"
   --max-num-seqs "$MAX_NUM_SEQS"
+  --max-num-batched-tokens "$MAX_NUM_BATCHED_TOKENS"
+  --api-server-count "$API_SERVER_COUNT"
   --host 0.0.0.0
   --port "$PORT"
   --compilation-config '{"cudagraph_mode": "FULL_DECODE_ONLY"}'