From 06b017483cb1cfc6aace3c60e94576fb667ec9bc Mon Sep 17 00:00:00 2001 From: Zhiyi Hong <2497491955@qq.com> Date: Thu, 30 Jul 2026 23:41:08 +0800 Subject: [PATCH] [Feat] add DSV4-Pro long-decode cases --- README.md | 4 +++ .../README.md | 35 ++++++++++++------- .../config.env | 2 +- .../quick_map_scenarios.tsv | 2 ++ .../tests/test_quick_map_results.py | 10 ++++-- 5 files changed, 38 insertions(+), 15 deletions(-) diff --git a/README.md b/README.md index 0fb13b0..4a7a715 100644 --- a/README.md +++ b/README.md @@ -1,5 +1,9 @@ # sskj — 多平台大模型推理性能基准测试项目 +> **更新（2026-07-30 23:38:42 CST）** +> +> 为 DeepSeek-V4-Pro 双机 Pro6000D SGLang TP16 Phase 1 增加两个最小长 Decode 补充场景：`1K → 4K, C=16` 用于观察持续长输出、KV 增长与稳态 Decode，`128K → 1K, C=1` 用于观察长上下文上的 Decode Attention。继续复用唯一 `run_quick_map.sh` 入口和 `CASE_IDS` 过滤，不新增启动脚本；场景验证单测由 9 点更新为 11 点，Shell、3 个 Python 单测和两点 Dry-run 已通过。 +> > **更新（2026-07-30 23:06:01 CST）** > > 完成 DeepSeek-V4-Pro 双机 Pro6000D SGLang TP16 Phase 1 正式 quick-map。Head 与 Worker 均通过 `mlx5_0/mlx5_3` 双 Rail `NET/IB + GDRDMA` 门禁；9 个固定点和 3 个混合 A/B 结果共 12/12 成功，总用时 28 分 36 秒。32K/128K 单请求 Prefill 输入吞吐为 2,652.76/2,710.16 token/s；在 C=32 Decode 中注入一个 128K Prefill 后，Output TPS 下降 24.08%，TPOT P95 增加 66.55%。阶段 HTML 已重写为只保留成功结果，并补充正式汇总、运行清单和 Phase 2 三个诊断负载；两节点容器和 16 张 GPU 已清理。 diff --git a/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/README.md b/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/README.md index 191d74a..9e9adf8 100644 --- a/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/README.md +++ b/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/README.md @@ -6,21 +6,16 @@ DeepSeek-V4-Pro on two RTX PRO 6000 Blackwell nodes. It does not modify or call ## Network precondition -The 2026-07-30 audit found that the current service container does not expose -`/dev/infiniband`, so NCCL falls back to `NET/Socket`. The quick-map previously -selected a low-speed non-compute interface and produced invalid 1K/32K cold -Prefill baselines. The Socket default is now `eth0`, one of the two deployment -compute interfaces (`eth0/eth3`). - -Declaring `NCCL_IB_HCA=mlx5_0,mlx5_3` does not by itself enable RDMA. The -launcher now exposes only the three device nodes needed by the two deployment +The launcher exposes only the three device nodes needed by the two deployment rails (`rdma_cm`, `uverbs0`, `uverbs3`) and fails before benchmarking unless both nodes' NCCL INFO logs prove `NET/IB` is using `mlx5_0` and `mlx5_3`. -`NCCL_CROSS_NIC` has no effect while the active transport is `NET/Socket`. +The 2026-07-30 Phase 1 run verified cross-node channels over +`NET/IB/*/GDRDMA`. `eth0` remains the NCCL bootstrap and Socket-fallback +interface. ## Scope -The fixed suite covers nine points: +The fixed suite covers eleven points: | Group | ISL | OSL | Concurrency | |---|---:|---:|---:| @@ -28,6 +23,8 @@ The fixed suite covers nine points: | Prefill throughput | 32K | 1 | 16 | | Decode latency | 1K | 1K | 1 | | Decode throughput | 1K | 1K | 16 / 32 / 64 | +| Long-output decode | 1K | 4K | 16 | +| Long-context decode | 128K | 1K | 1 | | Balanced pressure | 32K | 1K | 8 | The same main entry also runs a mixed-interference A/B: @@ -47,7 +44,7 @@ attribution belongs to the later scheduler/interference phase. |---|---| | `run_quick_map.sh` | The only Shell entry: service lifecycle, fixed map, mixed A/B, cleanup | | `config.env` | Machine, model, SGLang and benchmark configuration | -| `quick_map_scenarios.tsv` | Nine fixed workload points | +| `quick_map_scenarios.tsv` | Eleven fixed workload points | | `quick_map_results.py` | Validate output and generate CSV/JSONL/Markdown summaries | | `tests/test_quick_map_results.py` | Result parser regression tests | @@ -77,7 +74,21 @@ CASE_IDS="short_prefill_latency_1k_c1,mid_prefill_latency_32k_c1,decode_latency_ Remove `DRY_RUN=1` only after the TP16 service has passed the `NET/IB` startup gate. An unknown `CASE_IDS` value fails before any benchmark request is sent. -Leave `CASE_IDS` empty to run all nine fixed points. +Leave `CASE_IDS` empty to run all eleven fixed points. + +To run only the two long-decode supplements: + +```bash +export CASE_IDS="long_output_decode_1k_to_4k_c16,long_context_decode_128k_to_1k_c1" +export RUN_ID="dsv4pro-phase1-long-decode-$(date +%Y%m%d-%H%M%S)" + +trap 'bash run_quick_map.sh stop' EXIT INT TERM +bash run_quick_map.sh start +bash run_quick_map.sh fixed +``` + +This uses the existing single entry for startup, the selected fixed cases, and +trap-based cleanup. It does not run the mixed A/B suite. ## GPU run diff --git a/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/config.env b/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/config.env index 8370ab0..cee088b 100644 --- a/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/config.env +++ b/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/config.env @@ -71,7 +71,7 @@ RESULT_BASE="${RESULT_BASE:-${SCRIPT_DIR}/results}" RUNTIME_BASE="${RUNTIME_BASE:-${SCRIPT_DIR}/runtime}" # Optional comma- or space-separated subset of quick_map_scenarios.tsv. -# Leave empty for the complete nine-point fixed suite. +# Leave empty for the complete eleven-point fixed suite. CASE_IDS="${CASE_IDS:-}" DRY_RUN="${DRY_RUN:-0}" diff --git a/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/quick_map_scenarios.tsv b/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/quick_map_scenarios.tsv index b96ef69..9cae04e 100644 --- a/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/quick_map_scenarios.tsv +++ b/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/quick_map_scenarios.tsv @@ -7,4 +7,6 @@ decode_latency_1k_to_1k_c1 decode_latency 1024 1024 1 1 1 1 Single-request decod decode_throughput_1k_to_1k_c16 decode_throughput 1024 1024 16 1 1 1 Decode throughput at C16 decode_throughput_1k_to_1k_c32 decode_throughput 1024 1024 32 1 1 1 Decode throughput at C32 decode_throughput_1k_to_1k_c64 decode_throughput 1024 1024 64 1 1 1 Decode throughput at C64 +long_output_decode_1k_to_4k_c16 long_output_decode 1024 4096 16 1 1 1 Sustained long-output decode and KV growth +long_context_decode_128k_to_1k_c1 long_context_decode 131072 1024 1 1 1 0 Decode attention over a 128K context balanced_32k_to_1k_c8 balanced 32768 1024 8 1 1 0 Balanced prefill and decode pressure diff --git a/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/tests/test_quick_map_results.py b/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/tests/test_quick_map_results.py index 1e31230..e1890ba 100644 --- a/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/tests/test_quick_map_results.py +++ b/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/tests/test_quick_map_results.py @@ -16,13 +16,19 @@ import quick_map_results # noqa: E402 class QuickMapResultsTest(unittest.TestCase): def test_scenario_file_is_valid_and_unique(self) -> None: scenarios = quick_map_results.parse_scenarios(EXPERIMENT_DIR / "quick_map_scenarios.tsv") - self.assertEqual(len(scenarios), 9) - self.assertEqual(len({item["case_id"] for item in scenarios}), 9) + self.assertEqual(len(scenarios), 11) + self.assertEqual(len({item["case_id"] for item in scenarios}), 11) self.assertEqual(scenarios[2]["isl"], 131072) self.assertEqual(scenarios[2]["case_id"], "long_prefill_latency_128k_c1") self.assertEqual(scenarios[2]["prompt_multiplier"], 1) self.assertEqual(scenarios[2]["min_prompts"], 1) self.assertEqual(scenarios[2]["warmup_requests"], 0) + by_id = {item["case_id"]: item for item in scenarios} + self.assertEqual(by_id["long_output_decode_1k_to_4k_c16"]["osl"], 4096) + self.assertEqual(by_id["long_output_decode_1k_to_4k_c16"]["concurrency"], 16) + self.assertEqual(by_id["long_context_decode_128k_to_1k_c1"]["isl"], 131072) + self.assertEqual(by_id["long_context_decode_128k_to_1k_c1"]["osl"], 1024) + self.assertEqual(by_id["long_context_decode_128k_to_1k_c1"]["warmup_requests"], 0) def test_latency_percentiles_fall_back_to_request_arrays(self) -> None: data = {