fix(910c/glm52): 对齐官方A3教程参数(修DP die分配问题,同dsv4 99a22f0)

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

@@ -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

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

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"}'