feat(bench/pd): 支持 --flush-cache 透传 + PD profile 加 --disable-radix-cache

- sskj.bench: 新增 --flush-cache 参数，透传给 sglang.bench_serving
  （warmup 后、main run 前 flush 服务器 KV/prefix cache，保证 TTFT/TPOT 测量纯净）
- P/D PD profiles: LAUNCH_ARGS 加 --disable-radix-cache（关闭 RadixAttention 前缀缓存）

deploy/profiles/pro6000/kimi3_pro6000_pd_decode.env

@@ -51,4 +51,4 @@ PATCH_MOUNTS="/tmp/patch_k3_sm120.py:/tmp/patch_k3_sm120.py:ro /tmp/flash_kda-0.
 # BOOTSTRAP: 打补丁 → 装 flashkda + mooncake wheel → 按 rank 设 SGLANG_HOST_IP → 启动 decode。
 BOOTSTRAP="python3 /tmp/patch_k3_sm120.py && pip install /flash_kda-0.0.1-cp312-cp312-linux_x86_64.whl --no-deps -q && pip install /mc_wheels/mooncake_transfer_engine_cuda13-0.3.12.post1-cp312-cp312-manylinux_2_28_x86_64.whl --no-deps -q && export SGLANG_HOST_IP=\"174.1.60.$((5 + ${NODE_RANK}))\" && exec python3 -m sglang.launch_server ${LAUNCH_ARGS}"
 
-LAUNCH_ARGS="--model-path ${MODEL_PATH} --served-model-name ${SERVED_MODEL_NAME} --tp-size ${TP} --ep-size 32 --nnodes ${NNODES} --node-rank ${NODE_RANK} --dist-init-addr ${MASTER_IP}:${DIST_PORT} --trust-remote-code --moe-runner-backend marlin --mem-fraction-static 0.88 --cuda-graph-max-bs-decode 16 --mamba-radix-cache-strategy extra_buffer --dist-timeout 3600 --mamba-full-memory-ratio 0.36 --disaggregation-transfer-backend mooncake --disaggregation-bootstrap-port 28800 --disaggregation-ib-device mlx5_0,mlx5_1,mlx5_2,mlx5_3 --disaggregation-mode decode --host 0.0.0.0 --port ${PORT}"
+LAUNCH_ARGS="--model-path ${MODEL_PATH} --served-model-name ${SERVED_MODEL_NAME} --tp-size ${TP} --ep-size 32 --nnodes ${NNODES} --node-rank ${NODE_RANK} --dist-init-addr ${MASTER_IP}:${DIST_PORT} --trust-remote-code --moe-runner-backend marlin --mem-fraction-static 0.88 --cuda-graph-max-bs-decode 16 --mamba-radix-cache-strategy extra_buffer --disable-radix-cache --dist-timeout 3600 --mamba-full-memory-ratio 0.36 --disaggregation-transfer-backend mooncake --disaggregation-bootstrap-port 28800 --disaggregation-ib-device mlx5_0,mlx5_1,mlx5_2,mlx5_3 --disaggregation-mode decode --host 0.0.0.0 --port ${PORT}"

deploy/profiles/pro6000/kimi3_pro6000_pd_prefill.env

@@ -57,4 +57,4 @@ PATCH_MOUNTS="/tmp/patch_k3_sm120.py:/tmp/patch_k3_sm120.py:ro /tmp/flash_kda-0.
 # BOOTSTRAP: 打补丁 → 装 flashkda + mooncake wheel → 按 rank 设 SGLANG_HOST_IP → 启动 prefill。
 BOOTSTRAP="python3 /tmp/patch_k3_sm120.py && pip install /flash_kda-0.0.1-cp312-cp312-linux_x86_64.whl --no-deps -q && pip install /mc_wheels/mooncake_transfer_engine_cuda13-0.3.12.post1-cp312-cp312-manylinux_2_28_x86_64.whl --no-deps -q && export SGLANG_HOST_IP=\"174.1.60.$((1 + ${NODE_RANK}))\" && exec python3 -m sglang.launch_server ${LAUNCH_ARGS}"
 
-LAUNCH_ARGS="--model-path ${MODEL_PATH} --served-model-name ${SERVED_MODEL_NAME} --tp-size ${TP} --ep-size 32 --nnodes ${NNODES} --node-rank ${NODE_RANK} --dist-init-addr ${MASTER_IP}:${DIST_PORT} --trust-remote-code --moe-runner-backend marlin --mem-fraction-static 0.88 --cuda-graph-max-bs-decode 16 --mamba-radix-cache-strategy extra_buffer --dist-timeout 3600 --mamba-full-memory-ratio 0.36 --disaggregation-transfer-backend mooncake --disaggregation-bootstrap-port 28800 --disaggregation-ib-device mlx5_0,mlx5_1,mlx5_2,mlx5_3 --disaggregation-mode prefill --linear-attn-prefill-backend flashkda --host 0.0.0.0 --port ${PORT}"
+LAUNCH_ARGS="--model-path ${MODEL_PATH} --served-model-name ${SERVED_MODEL_NAME} --tp-size ${TP} --ep-size 32 --nnodes ${NNODES} --node-rank ${NODE_RANK} --dist-init-addr ${MASTER_IP}:${DIST_PORT} --trust-remote-code --moe-runner-backend marlin --mem-fraction-static 0.88 --cuda-graph-max-bs-decode 16 --mamba-radix-cache-strategy extra_buffer --disable-radix-cache --dist-timeout 3600 --mamba-full-memory-ratio 0.36 --disaggregation-transfer-backend mooncake --disaggregation-bootstrap-port 28800 --disaggregation-ib-device mlx5_0,mlx5_1,mlx5_2,mlx5_3 --disaggregation-mode prefill --linear-attn-prefill-backend flashkda --host 0.0.0.0 --port ${PORT}"

src/sskj/bench/cli.py

@@ -48,6 +48,7 @@ def _build_parser() -> argparse.ArgumentParser:
     run_parser.add_argument("--output-dir", type=Path, default=None)
     run_parser.add_argument("--timeout", type=int, default=1800)
     run_parser.add_argument("--health-wait", type=int, default=600)
+    run_parser.add_argument("--flush-cache", action="store_true", help="Pass --flush-cache to sglang.bench_serving (flush server KV/prefix cache after warmup, before main run)")
     run_parser.add_argument("--dry-run", action="store_true")
     return parser
 
@@ -198,6 +199,7 @@ def _cmd_run(args: argparse.Namespace, root: Path) -> int:
             container_dataset_path=experiment_env.get("CONTAINER_DATASET_PATH"),
             random_range_ratio=float(experiment_env.get("RANDOM_RANGE_RATIO", "1.0")),
             warmup_max_requests=int(experiment_env.get("BENCH_WARMUP_MAX_REQUESTS", "0") or 0),
+            flush_cache=args.flush_cache,
             client_mode=client_mode,
             client_image=client_image,
             client_python=client_python,

src/sskj/bench/runner.py

@@ -24,6 +24,7 @@ class BenchClientOptions:
     container_dataset_path: str | None = None
     random_range_ratio: float = 1.0
     warmup_max_requests: int = 0
+    flush_cache: bool = False
     client_mode: str = "auto"
     client_image: str | None = None
     client_python: str | None = None
@@ -68,6 +69,8 @@ def _bench_args(scenario: dict[str, Any], options: BenchClientOptions) -> list[s
     ]
     if options.warmup_max_requests and options.warmup_max_requests > 0:
         args += ["--warmup-requests", str(options.warmup_max_requests)]
+    if options.flush_cache:
+        args += ["--flush-cache"]
 
     dataset_path = (
         options.container_dataset_path