fix(910c/glm52): start_vllm_docker.sh 加per-TP参数覆盖 + 修printf转义破坏JSON

问题: add16实际启动走start_vllm_docker.sh(用config.env全局MAX_MODEL_LEN=131072),
而非engine_build_server_args(有per-TP覆盖但只写server_cmd.txt不影响启动).
TP=8用131072上下文+无per-TP覆盖,且engine_build_server_args的printf %q转义
破坏了JSON(--additional-config的enable_dsa_cp等未生效),致MoE tiling失败.

修复:
1. start_vllm_docker.sh: 加per-TP case覆盖(TP=8:16384/64/0.95, TP=16:131072/256/0.92)
2. run_adaptive_concurrency_add16.sh: engine_build_server_args 用单引号包裹替代
   printf %q,避免JSON被反斜杠转义(仅影响server_cmd.txt记录,实际启动走start_vllm_docker.sh)

experiments/910c/glm52_910c_vllm_tp_dp_matrix/run_adaptive_concurrency_add16.sh

@@ -92,6 +92,12 @@ engine_build_server_args() {
       ;;
   esac
   log "server args tp=${tp} dp=${dp} gpu_mem_util=${mem_util} max_model_len=${max_len} max_num_seqs=${max_seqs}" >&2
+  # MoE / DSA-CP args required by GLM-5.2 on vllm-ascend (must match
+  # start_vllm_docker.sh). --enable-expert-parallel is mandatory for
+  # FlashComm v1 (VLLM_ASCEND_ENABLE_FLASHCOMM1=1); without it MoE
+  # MoeDistributeDispatchV2 tiling fails at engine init.
+  local compilation_config='{"cudagraph_mode": "FULL_DECODE_ONLY"}'
+  local additional_config='{"enable_dsa_cp": true,"enable_sparse_sfa_c8": false, "enable_sparse_li_c8": true,"enable_balance_scheduling": true,"multistream_overlap_shared_expert":true}'
   local -a args=(
     vllm serve "$MODEL_PATH"
     --served-model-name "$SERVED_MODEL_NAME"
@@ -99,16 +105,26 @@ engine_build_server_args() {
     --kv-cache-dtype "$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 "$max_seqs"
+    --compilation-config "$compilation_config"
+    --additional-config "$additional_config"
     --host 0.0.0.0
     --port "$ENGINE_PORT"
   )
   if (( dp > 1 )); then
     args+=(--data-parallel-size "$dp")
   fi
-  printf '%q ' "${args[@]}"
+  # Emit args space-separated. JSON config strings are single-quoted
+  # verbatim (not %q-escaped) so the recorded command stays re-runnable.
+  local out=""
+  local a
+  for a in "${args[@]}"; do
+    out+="'$a' "
+  done
+  printf '%s' "$out"
 }
 
 engine_start_server() {

experiments/910c/glm52_910c_vllm_tp_dp_matrix/start_vllm_docker.sh

@@ -31,6 +31,21 @@ docker rm -f "$NAME" >/dev/null 2>&1 || true
 # vLLM-Ascend launch args. Differences vs NVIDIA vLLM:
 #\n  - no --no-enable-flashinfer-autotune (Ascend uses its own attention path)
 #\n  - --kv-cache-dtype may need to be fp16 if the image rejects fp8 on 910C
+# Per-TP overrides: TP=8 has tight KV cache on 64GB/die (cap context/seqs),
+# TP=16 uses all 16 dies and can serve the full 128K context.
+case "$TP" in
+  8)
+    GPU_MEMORY_UTILIZATION="${TP8_GPU_MEMORY_UTILIZATION:-0.95}"
+    MAX_MODEL_LEN="${TP8_MAX_MODEL_LEN:-16384}"
+    MAX_NUM_SEQS="${TP8_MAX_NUM_SEQS:-64}"
+    ;;
+  16)
+    GPU_MEMORY_UTILIZATION="${TP16_GPU_MEMORY_UTILIZATION:-0.92}"
+    MAX_MODEL_LEN="${TP16_MAX_MODEL_LEN:-131072}"
+    MAX_NUM_SEQS="${TP16_MAX_NUM_SEQS:-256}"
+    ;;
+esac
+
 SERVER_ARGS=(
   "$MODEL_PATH"
   --served-model-name "$SERVED_MODEL_NAME"