fix(910c): bench client 可用化 - 本地 vllm-ascend-sglang 镜像 + torch_npu 自动加载禁用

- runner.py docker client 注入 TORCH_DEVICE_BACKEND_AUTOLOAD=0（bench 纯 HTTP
  client 不需要 NPU backend，跳过 torch_npu 加载失败）
- dsv4/glm52 config.env: DOCKER_CLIENT_IMAGE 指向本地定制镜像
  (local/vllm-ascend:0.23-a3-dsv4-sglang，内置 sglang 0.5.2 bench_serving)，
  USE_DOCKER_CLIENT=1，sskj.bench/run_bench.sh 的 docker client 分支可用
- run_bench.sh docker client 分支同步注入 AUTOLOAD=0
- glm52 profile 修正 DOCKER_IMAGE 为本机存在的 glm5.2-a3-openeuler

deploy/profiles/910c/glm52_910c_vllm_tp_dp_matrix.env

@@ -9,7 +9,7 @@ EXPERIMENT=glm52_910c_vllm_tp_dp_matrix
 MODEL_NAME=GLM-5.2
 ENGINE=vllm
 RUNTIME=docker
-DOCKER_IMAGE=local/vllm-ascend:0.23-a3-20260718-sglang
+DOCKER_IMAGE=quay.io/ascend/vllm-ascend:glm5.2-a3-openeuler
 CONTAINER_NAME=vllm-ascend-glm52-910c_tp${TP}_dp${DP}
 MODEL_PATH=/mnt/models/GLM-5.2-w4a8c8
 SERVED_MODEL_NAME=glm-5.2

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/config.env

@@ -31,12 +31,12 @@ CONTAINER_PYTHON="/usr/local/python3.12.13/bin/python3"
 USE_DOCKER="${USE_DOCKER:-1}"
 DOCKER_IMAGE="${DOCKER_IMAGE:-local/vllm-ascend:0.23-a3-dsv4-sglang}"
 
-# Benchmark client Docker image. vLLM's image does not include sglang.bench_serving;
+# Benchmark client Docker image. The vllm-ascend images built on this host
-# reuse the vllm-ascend container itself for the client via `docker exec` (see
+# (local/vllm-ascend:0.23-a3-*-sglang, see envs/ASCEND_910C_ENV_SETUP.md §8.1)
-# run_adaptive_concurrency_add16.sh), so this is only used if USE_DOCKER_CLIENT=1
+# carry sglang 0.5.2 bench_serving, so the client runs as a standalone container
-# with an external sglang image. Default off on 910c.
+# (TORCH_DEVICE_BACKEND_AUTOLOAD=0 injected by the runner, no NPU needed).
-DOCKER_CLIENT_IMAGE="${DOCKER_CLIENT_IMAGE:-lmsysorg/sglang:latest}"
+DOCKER_CLIENT_IMAGE="${DOCKER_CLIENT_IMAGE:-local/vllm-ascend:0.23-a3-dsv4-sglang}"
-USE_DOCKER_CLIENT="${USE_DOCKER_CLIENT:-0}"
+USE_DOCKER_CLIENT="${USE_DOCKER_CLIENT:-1}"
 
 # Device selection. We mount all 8 cards (16 dies) via --device /dev/davinci0..15
 # in start_vllm_docker.sh. ASCEND_VISIBLE_DEVICES is kept for parity with the

experiments/910c/dsv4_910c_vllm_tp_dp_matrix/run_bench.sh

@@ -124,6 +124,7 @@ run_bench_serving() {
       -e HF_HUB_OFFLINE=1 \
       -e TRANSFORMERS_OFFLINE=1 \
       -e HF_DATASETS_OFFLINE=1 \
+      -e TORCH_DEVICE_BACKEND_AUTOLOAD=0 \
       "${DOCKER_CLIENT_IMAGE}" \
       python -m sglang.bench_serving "$@"
   else

experiments/910c/glm52_910c_vllm_tp_dp_matrix/config.env

@@ -29,12 +29,12 @@ CONTAINER_PYTHON="/usr/local/python3.12.13/bin/python3"
 USE_DOCKER="${USE_DOCKER:-1}"
 DOCKER_IMAGE="${DOCKER_IMAGE:-local/vllm-ascend:0.23-a3-20260718-sglang}"
 
-# Benchmark client Docker image. vLLM's image does not include sglang.bench_serving;
+# Benchmark client Docker image. The vllm-ascend images built on this host
-# reuse the vllm-ascend container itself for the client via `docker exec` (see
+# (local/vllm-ascend:0.23-a3-*-sglang, see envs/ASCEND_910C_ENV_SETUP.md §8.1)
-# run_adaptive_concurrency_add16.sh), so this is only used if USE_DOCKER_CLIENT=1
+# carry sglang 0.5.2 bench_serving, so the client runs as a standalone container
-# with an external sglang image. Default off on 910c.
+# (TORCH_DEVICE_BACKEND_AUTOLOAD=0 injected by the runner, no NPU needed).
-DOCKER_CLIENT_IMAGE="${DOCKER_CLIENT_IMAGE:-lmsysorg/sglang:latest}"
+DOCKER_CLIENT_IMAGE="${DOCKER_CLIENT_IMAGE:-local/vllm-ascend:0.23-a3-dsv4-sglang}"
-USE_DOCKER_CLIENT="${USE_DOCKER_CLIENT:-0}"
+USE_DOCKER_CLIENT="${USE_DOCKER_CLIENT:-1}"
 
 # Device selection. ASCEND_VISIBLE_DEVICES selects NPU cards 0..7; the Ascend
 # Docker Runtime (default runtime on this host) injects the matching dies.

experiments/910c/glm52_910c_vllm_tp_dp_matrix/run_bench.sh

@@ -115,6 +115,7 @@ run_bench_serving() {
       -e HF_HUB_OFFLINE=1 \
       -e TRANSFORMERS_OFFLINE=1 \
       -e HF_DATASETS_OFFLINE=1 \
+      -e TORCH_DEVICE_BACKEND_AUTOLOAD=0 \
       "${DOCKER_CLIENT_IMAGE}" \
       python -m sglang.bench_serving "$@"
   else

src/sskj/bench/runner.py

@@ -113,7 +113,16 @@ def run_scenario(scenario: dict[str, Any], options: BenchClientOptions) -> int:
             cmd += ["-v", f"{options.dataset_path}:{options.dataset_path}:ro"]
         if options.output_file:
             cmd += ["-v", f"{options.output_file.parent}:{options.output_file.parent}"]
-        cmd += ["-e", "HF_HUB_OFFLINE=1", "-e", "TRANSFORMERS_OFFLINE=1", "-e", "HF_DATASETS_OFFLINE=1"]
+        cmd += [
+            "-e",
+            "HF_HUB_OFFLINE=1",
+            "-e",
+            "TRANSFORMERS_OFFLINE=1",
+            "-e",
+            "HF_DATASETS_OFFLINE=1",
+            "-e",
+            "TORCH_DEVICE_BACKEND_AUTOLOAD=0",
+        ]
         cmd += [
             options.client_image,
             "python",