diff --git a/README.md b/README.md
index ea0f4a1..0fb13b0 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 23:06:01 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。
+>
> **更新(2026-07-30 18:40:53 CST)**
>
> 为 DeepSeek-V4-Pro 双机 TP16 quick-map 的唯一入口新增 `CASE_IDS` 场景过滤和未知 Case 预检,可在完整九点实验前先跑 1K/32K Prefill 与 C1/C32 Decode 四点 Sanity;运行清单会记录实际过滤条件。同步精简阶段档案:正文只保留最终成功 Run 与有效结论,历史失败压缩到末尾经验教训。
diff --git a/docs/dsv4pro_pro6000d_2node_sglang/phase1_dsv4pro_pro6000d_2node_sglang_quick_map.html b/docs/dsv4pro_pro6000d_2node_sglang/phase1_dsv4pro_pro6000d_2node_sglang_quick_map.html
index 1d3f41c..31cf0e0 100644
--- a/docs/dsv4pro_pro6000d_2node_sglang/phase1_dsv4pro_pro6000d_2node_sglang_quick_map.html
+++ b/docs/dsv4pro_pro6000d_2node_sglang/phase1_dsv4pro_pro6000d_2node_sglang_quick_map.html
@@ -112,6 +112,12 @@
color: #075e58;
}
+ .decision {
+ padding: 14px 18px;
+ background: var(--teal-soft);
+ border-left: 4px solid var(--teal);
+ }
+
h2 {
margin: 48px 0 16px;
padding-bottom: 9px;
@@ -228,7 +234,7 @@
节点:174.1.51.5 + 174.1.51.7
拓扑:SGLang TP16 / EP2
- 更新:2026-07-30 17:54 CST
+ 更新:2026-07-30 22:55:37 CST
@@ -237,9 +243,10 @@
返回推理优化主计划
- 阶段状态:等待真实双 Rail RDMA 重跑。
- 启动器与静态验证已完成;历史无效 Run 不进入最终分析。下一次仅在
- NCCL 日志通过 NET/IB 门禁后执行 Sanity 和完整 quick map。
+ 阶段状态:已完成。
+ 正式 Run dsv4pro-phase1-full-20260730-220916 在双 Rail
+ NET/IB + GDRDMA 下完成 9 个固定点和 3 个混合 A/B 结果,
+ 共 12/12 成功,用时 28 分 36 秒;服务、容器与 16 张 GPU 已清理。
1. 目标与边界
@@ -530,34 +537,96 @@ wait "${background_pid}"
| 完整 Dry-run | 通过 | 服务、九个固定点、混合 A/B、清理均展开成功 |
| 真实旧 Bench JSON 解析 | 通过 | 成功解析 P50/P95/P99 与吞吐字段 |
| 项目精简 | 通过 | 实验目录顶层仅保留一个 Shell 入口 |
- | RDMA fail-closed | 静态验证通过 | 非法网卡、缺失设备或未出现 NET/IB 时禁止 benchmark |
+ | 双 Rail 传输门禁 | 通过 | Head 与 Worker 均识别 mlx5_0/mlx5_3,跨节点 Channel 使用 NET/IB/*/GDRDMA |
+ | 四点 Sanity | 4/4 通过 | 1K/32K Prefill 与 C1/C32 Decode 均恢复到合理量级 |
+ | 冷 Prefix 口径 | 通过 | 正式测量请求的 Head 日志显示 #cached-token: 0 |
+ | 完整真机 Run | 12/12 通过 | 固定矩阵 9/9,混合 A/B 3/3,运行期失败 0 |
+ | 资源清理 | 通过 | 两节点相关容器与计算进程为 0,16 张 GPU 显存占用为 0 |
9. 最终真机结果
-
- 正确双 Rail RDMA 下的正式结果尚未生成。本节只接受通过传输门禁、
- 冷缓存口径和结果校验的最终 Run;完成后将写入四点 Sanity、九个固定点、
- 混合 A/B 与阶段结论。
+
+ 本节只使用正式成功 Run。Profiler 与投机解码均关闭,每个 Case 只做一次快速测量,
+ 所以它适合决定下一步 Profile 对象,不作为需要统计置信度的最终容量认证。
+
+ 9.1 执行摘要
- | 门禁或实验 | 状态 | 最终证据 |
+ | 项目 | 结果 | 证据 |
- 双 Rail NET/IB | 待真机验证 | 头、Worker 日志均需识别 mlx5_0/mlx5_3 |
- | 四点 Sanity | 待执行 | 1K/32K Prefill 与 C1/C32 Decode |
- | 九个固定点 | 待执行 | 最终 summary.csv 与 report.md |
- | 混合 A/B | 待执行 | Control 与 Treatment 的 Decode 指标变化 |
+ | Run ID | dsv4pro-phase1-full-20260730-220916 | COMPLETED |
+ | 运行时间 | 28 分 36 秒 | 22:09:47 至 22:38:22 CST |
+ | 网络路径 | 双 Rail NET/IB + GDRDMA | mlx5_0 与 mlx5_3 |
+ | 结果完整性 | 12/12 成功 | 固定点 9/9;混合 A/B 3/3 |
+ 9.2 Prefill
+
+
+ | 场景 | Input TPS | TTFT P95 | 观察 |
+
+
+ | 1K → 1,C=1 | 1,969.66 tok/s | 0.502 s | 短请求固定开销占比更高 |
+ | 32K → 1,C=1 | 2,652.76 tok/s | 12.335 s | 单请求吞吐进入稳定区间 |
+ | 128K → 1,C=1 | 2,710.16 tok/s | 48.344 s | 长 Prefill 代表点 |
+ | 32K → 1,C=16 | 3,112.77 tok/s | 162.087 s | 聚合吞吐仅比 C=1 高 17.3%,排队时延显著增加 |
+
+
+
+ 9.3 Decode
+
+
+ | 1K → 1K | Output TPS | TTFT P95 | TPOT P95 | E2E P95 |
+
+
+ | C=1 | 31.41 tok/s | 0.363 s | 31.47 ms | 32.555 s |
+ | C=16 | 295.29 tok/s | 4.950 s | 50.02 ms | 55.444 s |
+ | C=32 | 461.68 tok/s | 8.022 s | 63.31 ms | 70.933 s |
+ | C=64 | 647.42 tok/s | 12.716 s | 93.44 ms | 101.163 s |
+
+
+
+ Decode 吞吐到 C=64 仍在上升,但增益递减且 TPOT 明显变差。综合场景
+ 32K → 1K,C=8 的 Input/Output TPS 为
+ 2,038.00 / 63.69,TTFT P95 为 82.084 s,
+ 说明 Prefill 与 Decode 同时存在时干扰很强。
+
+
+ 9.4 混合 Prefill/Decode A/B
+
+
+ | 指标 | A:仅 Decode | B:注入 128K Prefill | 变化 |
+
+
+ | Output TPS | 455.68 tok/s | 345.95 tok/s | -24.08% |
+ | TTFT P95 | 9.443 s | 10.194 s | +7.96% |
+ | TPOT P95 | 65.88 ms | 109.73 ms | +66.55% |
+ | E2E P95 | 72.008 s | 117.630 s | +63.36% |
+
+
+
+ Phase 2 优先采集三类代表负载:128K → 1,C=1 的纯长 Prefill、
+ 32K → 1,C=16 的并发 Prefill,以及
+ 1K → 1K,C=32 在有无 128K 注入时的混合 A/B。
+ 目标是区分计算、显存带宽、调度排队、跨机通信和节点不均衡。
+
+
+ 完整产物:
+ 报告、
+ 逐点汇总、
+ 聚合表、
+ 运行清单。
+
+
10. 经验教训
- 启动参数不等于实际传输路径;开始性能测试前必须由 NCCL 日志证明
NET/IB。
- Warm-up、固定随机种子和跨 Case Prefix Cache 会改变 TTFT,冷缓存与热缓存必须分开报告。
- - 传输门禁或测量口径不成立时,性能结果不能用于归因模型、调度器或 Kernel。
- - 完整实验前先跑少量 Sanity 点,能更早发现环境和口径问题。
+ - 先跑四点 Sanity 再启动完整矩阵,可以在几分钟内验证环境、口径和数量级。
历史排查细节保存在
@@ -570,11 +639,11 @@ wait "${background_pid}"
11. 运行命令
cd /data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map
-tmux new-session -d -s dsv4pro-pro6000d-2node-sglang-quick-map -c "$PWD"
-tmux send-keys -t dsv4pro-pro6000d-2node-sglang-quick-map \
- 'RUN_ID=dsv4pro-pro6000d-2node-sglang-quick-v2-20260730-143625 bash run_quick_map.sh all' Enter
+tmux new-session -d -s dsv4pro-phase1-full \
+ "RUN_ID=dsv4pro-phase1-full-20260730-220916 bash run_quick_map.sh all \
+ 2>&1 | tee /data/hzy/dsv4pro_phase1_full_20260730-220916.log"
-tmux attach -t dsv4pro-pro6000d-2node-sglang-quick-map
+tmux attach -t dsv4pro-phase1-full
下一阶段:
diff --git a/docs/dsv4pro_pro6000d_2node_sglang/phase2_dsv4pro_pro6000d_2node_sglang_prefill_hardware_attribution.html b/docs/dsv4pro_pro6000d_2node_sglang/phase2_dsv4pro_pro6000d_2node_sglang_prefill_hardware_attribution.html
index b75a520..03bd8e3 100644
--- a/docs/dsv4pro_pro6000d_2node_sglang/phase2_dsv4pro_pro6000d_2node_sglang_prefill_hardware_attribution.html
+++ b/docs/dsv4pro_pro6000d_2node_sglang/phase2_dsv4pro_pro6000d_2node_sglang_prefill_hardware_attribution.html
@@ -135,7 +135,7 @@
节点:174.1.51.5 + 174.1.51.7
拓扑:SGLang TP16 / EP2
- 更新:2026-07-30 17:54 CST
+ 更新:2026-07-30 22:55:37 CST
@@ -144,55 +144,35 @@
返回推理优化主计划
- 当前状态:网络前置条件未满足,Phase 2 暂停,代码尚未开始。
- 本页从第一行 Phase 2 代码开始同步维护。每次代码改动、静态验证、真机运行和
- 结果判断都会在对应小节留下文件路径、命令和证据,不在阶段结束后凭记忆补写。
+ 当前状态:Phase 1 前置条件已通过,等待阶段汇报确认后开始实现。
+ Phase 2 代码尚未创建。本页已经根据最终 Phase 1 结果选择诊断 Case;
+ 后续代码改动、静态验证、真机运行和结果判断会同步写入本页。
- 1. 为什么现在不能直接进入 Phase 2
-
- Phase 1 在没有 Profiler、没有 Prefix Cache 命中的条件下得到以下结果,
- 但这些数值后来确认受错误 Socket 网络路径污染:
-
+ 1. Phase 1 交接结果
- | ISL / OSL / C | 输入 TPS | TTFT | 结果 |
+ | 代表负载 | 关键结果 | Phase 2 用途 |
- | 1K / 1 / 1 | 64.44 tok/s | 15.88 s | 错误网络证据,不作为基线 |
- | 32K / 1 / 1 | 64.96 tok/s | 504.44 s | 错误网络证据,不作为基线 |
- | 128K / 1 / 1 | 65.20 tok/s | 2010.38 s | 错误网络证据,不作为基线 |
+ | 128K → 1,C=1 | Input TPS 2,710.16;TTFT P95 48.344 s | 纯长 Prefill 的计算、显存与通信归因 |
+ | 32K → 1,C=16 | Input TPS 3,112.77;TTFT P95 162.087 s | 并发 Prefill 的排队、Chunk 调度与节点均衡 |
+ | 1K → 1K,C=32 + 128K 注入 | Output TPS -24.08%;TPOT P95 +66.55% | Prefill 干扰 Decode 时的硬件资源竞争 |
- quick-map 容器没有 /dev/infiniband,NCCL 回退
- NET/Socket;脚本又误选低速非计算网。原网络配置冷请求控制组中,
- 1K/32K TTFT 分别只有 1.458s/38.062s,比 quick-map 快约 10.9×/13.25×。
- 因此约 65 token/s 不是待 profile 的模型现象,而是已定位的部署配置错误。
-
-
- 1.1 Phase 2 前置审计
-
- 旧脚本较短的 TTFT 包含两个因素。其一,旧脚本固定执行 16 条同 Prompt
- Warm-up,从不清 Prefix Cache,并按固定 Seed 递增长度;17:40 的失败 Run
- 还在 18:01 正式 Run 前预热了同一批 1K 请求。其二,新旧模型参数虽相同,
- NCCL Socket 接口不同,而二者容器都没有形成真实 RDMA 数据面。
-
-
- 网络控制组明确打印 NET/IB : No device found 和
- Using network Socket。Phase 1 入口现已加入 RDMA 设备透传与
- NET/IB fail-closed 校验,但真机验证尚未运行。Phase 2 只有在
- 双 Rail RDMA 得到运行时证据并重跑 Phase 1 后才会开始。Warm Prefix/Prefix Cache
- 收益仍另立 A/B,不与冷 Prefill 混算。
+ 最终基线已由 Head 与 Worker 日志证明使用
+ mlx5_0/mlx5_3 双 Rail NET/IB + GDRDMA,
+ 正式测量请求为冷 Prefix,12/12 结果成功。Phase 2 保持相同服务配置和请求口径。
2. 本阶段的边界
- 只测试 SGLang,不测试 vLLM。
- - 保留模型、镜像、TP16、EP2 和显存比例;NCCL 参数必须使用修正并验证后的版本。
+ - 保留模型、镜像、TP16、EP2、显存比例和已验证的双 Rail NCCL 配置。
- 不启用 Nsight Systems、PyTorch Profiler、NCCL DEBUG 或投机解码。
- 不调参,不尝试优化;先获得足以区分瓶颈类别的硬件证据。
- - 第一轮仍只重放
32K → 1, C=1,但必须与修正后的 Phase 1 结果对齐。
+ - 只重放本页选出的三个代表负载,不重复 Phase 1 全矩阵。
- 采集器从请求开始前启动,到请求结束后停止,不能中途补采后声称完整。
@@ -216,15 +196,16 @@
保存两节点静态快照:GPU/NIC/NUMA 拓扑、驱动、CUDA、镜像与服务命令。
复用 Phase 1 已验证的 run_quick_map.sh start 启动同配置双机服务。
在 Head 和 Worker 同时启动 GPU、CPU、网卡与 RDMA 采样,先记录 15 秒空闲基线。
- 清空 Prefix Cache,发送一条 32K → 1, C=1 请求,Seed 与 Phase 1 一致。
+ 依次重放 128K → 1, C=1 与 32K → 1, C=16,请求参数和 Seed 与 Phase 1 一致。
+ 重放 1K → 1K, C=32 Control 与 128K Prefill 注入 Treatment,保留相同注入时序。
请求结束后继续采样 15 秒,再停止采集器和服务。
- 按时间戳将请求、8K Chunk、GPU、CPU 和 Rail 指标对齐,生成摘要与判定。
+ 按时间戳将请求、GPU、CPU 和双 Rail 指标对齐,生成摘要与判定。
- idle 15s │──────── 32K Prefill:4 × 8K Chunk ────────│ cooldown 15s
- ↑ request_start ↑ request_end
- Head 与 Worker 的所有采集器覆盖完整时间窗
+ idle 15s │ 128K C1 │ 32K C16 │ Decode Control │ Decode + Prefill │ cooldown 15s
+ Head 与 Worker 的所有采集器覆盖完整诊断窗口
- 预计服务加载约 6 分钟、请求约 8.5 分钟,连同快照和清理应在 20 分钟左右完成。
+ Phase 1 中服务加载约 5 分 30 秒,三个诊断负载合计为分钟级;
+ 连同静态快照、采样和清理,目标仍控制在 30 分钟内。
5. 采集指标
@@ -251,7 +232,7 @@ dsv4pro_pro6000d_2node_sglang_prefill_hardware_attribution/
run_prefill_hardware_attribution.sh | 唯一 Shell 入口;服务启停、双节点采集器、单 Case、Trap 清理 | 待实现 |
- config.env | Phase 1 入口路径、32K Case、采样间隔、结果路径 | 待实现 |
+ config.env | Phase 1 入口路径、三个诊断 Case、采样间隔、结果路径 | 待实现 |
hardware_attribution.py | 结构化解析、时间对齐、统计摘要与报告生成 | 待实现 |
tests/test_hardware_attribution.py | 计数器差分、单位换算、统计与缺失工具回退测试 | 待实现 |
README.md | 入口命令、环境变量和结果目录说明 | 待实现 |
@@ -259,7 +240,7 @@ dsv4pro_pro6000d_2node_sglang_prefill_hardware_attribution/
Phase 2 不复制双机 Docker 启停实现。唯一入口通过环境变量调用 Phase 1 的
- run_quick_map.sh start/stop,只新增硬件采集和 32K 请求编排。
+ run_quick_map.sh start/stop,只新增硬件采集和三个代表负载的编排。
顶层仍只保留一个 Shell 文件,不创建额外 tmux、launch、start 或 stop 脚本。
@@ -296,7 +277,7 @@ dsv4pro_pro6000d_2node_sglang_prefill_hardware_attribution/
8. 验收条件
- - Bench 的 ISL、OSL、并发、Seed、缓存状态与 Phase 1 的 32K Case 一致。
+ - Bench 的 ISL、OSL、并发、Seed、缓存状态与 Phase 1 对应 Case 一致。
- 两节点采集器均覆盖请求开始前 15 秒到结束后 15 秒。
- 每份时间序列有节点名、墙钟时间和单调时钟,能与服务 Chunk 日志对齐。
- 采集器不可用时记录
UNAVAILABLE 和原因,不静默跳过。
@@ -311,17 +292,15 @@ dsv4pro_pro6000d_2node_sglang_prefill_hardware_attribution/
| 时间 | 代码或运行 | 结果 |
- | 2026-07-30 15:46 CST | 创建 Phase 2 设计与档案 | 代码尚未开始,等待按本页设计实现 |
- | 2026-07-30 16:35 CST | 完成旧脚本与 quick-map 同口径审计 | 排除服务参数、OSL=1 和一次性 JIT;确认旧产物被 Warm-up、跨 Case 与前一轮 Prefix Cache 污染 |
- | 2026-07-30 17:30 CST | 完成原网络配置冷请求控制组 | 确认 quick-map 误入低速非计算网 Socket;Phase 2 暂停,先修正并重跑 Phase 1 |
- | 2026-07-30 17:54 CST | 完成 Phase 1 RDMA fail-closed 代码与 dry-run | 只允许 eth0/eth3 与 mlx5_0/mlx5_3;等待真机 NET/IB 证据 |
+ | 2026-07-30 15:46 CST | 创建 Phase 2 设计与档案 | 采用单入口和独立轻量采样,不复制双机服务启动逻辑 |
+ | 2026-07-30 22:55:37 CST | 完成 Phase 1 阶段交接 | 双 Rail 门禁和 12/12 正式结果通过;选定纯 Prefill、并发 Prefill、混合干扰三个诊断负载 |
10. 真机结果
- 尚未运行,也不应立即运行。先获得修正网络后的 Phase 1 冷 32K 基线;
- Phase 2 代码实现、静态验证和 Dry-run 完成后,再向用户说明具体改动并等待阶段门。
+ 尚未运行。按阶段门约定,先完成 Phase 1 汇报;用户确认进入 Phase 2 后,
+ 再实现代码、完成静态验证和 Dry-run,并在启动真机诊断前说明具体改动。
返回 Phase 1 实施记录
diff --git a/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/aggregate.csv b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/aggregate.csv
new file mode 100644
index 0000000..712308b
--- /dev/null
+++ b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/aggregate.csv
@@ -0,0 +1,13 @@
+suite,case_id,role,stage,isl,osl,concurrency,planned_repetitions,completed_repetitions,statuses,total_tps_mean,total_tps_cv_pct,output_tps_mean,output_tps_cv_pct,ttft_p95_mean_ms,tpot_p95_mean_ms,itl_p95_mean_ms,e2e_p95_mean_ms
+fixed,balanced_32k_to_1k_c8,,balanced,32768,1024,8,1,1,COMPLETED,2101.686722481554,,63.68747643883498,,82083.64989476977,109.32309288714016,43.84165157971437,128588.8987001963
+fixed,decode_latency_1k_to_1k_c1,,decode_latency,1024,1024,1,1,1,COMPLETED,62.82990203738489,,31.414951018692445,,363.2723209448159,31.468245639333286,31.659404194215313,32555.28760998277
+fixed,decode_throughput_1k_to_1k_c16,,decode_throughput,1024,1024,16,1,1,COMPLETED,590.5818337765435,,295.29091688827174,,4950.267374995747,50.01878498413362,49.8873026604997,55443.82167501317
+fixed,decode_throughput_1k_to_1k_c32,,decode_throughput,1024,1024,32,1,1,COMPLETED,923.3692532200665,,461.68462661003326,,8021.852347906679,63.30661669023356,62.237933481810614,70933.18627287517
+fixed,decode_throughput_1k_to_1k_c64,,decode_throughput,1024,1024,64,1,1,COMPLETED,1294.833542313111,,647.4167711565555,,12716.491229846724,93.44459645826957,87.80462378927041,101163.1970004586
+fixed,long_prefill_latency_128k_c1,,prefill_latency,131072,1,1,1,1,COMPLETED,2710.1829473156417,,0.020676897204730506,,48344.461318978574,0.0,0.0,48344.52949295519
+fixed,mid_prefill_latency_32k_c1,,prefill_latency,32768,1,1,1,1,COMPLETED,2652.8390294198,,0.08095575175988892,,12334.68782599084,0.0,0.0,12334.764264000114
+fixed,mid_prefill_throughput_32k_c16,,prefill_throughput,32768,1,16,1,1,COMPLETED,3112.8645802960464,,0.09499418902914482,,162087.46085499297,0.0,0.0,162087.50443853205
+fixed,short_prefill_latency_1k_c1,,prefill_latency,1024,1,1,1,1,COMPLETED,1971.5883577957954,,1.9235008368739468,,501.89953204244375,0.0,0.0,501.94522901438177
+mixed,decode_control_1k_to_1k_c32,control,mixed_interference,1024,1024,32,1,1,COMPLETED,911.3699336194111,,455.68496680970554,,9442.636363586644,65.88327712923635,61.753195300116204,72007.63789927005
+mixed,decode_with_128k_prefill_1k_to_1k_c32,decode_background,mixed_interference,1024,1024,32,1,1,COMPLETED,691.8932985710949,,345.94664928554744,,10193.896457596566,109.7299457727384,61.61956858995836,117629.63820034638
+mixed,long_prefill_injection_128k_to_1_c1,prefill_injection,mixed_interference,131072,1,1,1,1,COMPLETED,2859.7533639992707,,0.021818020217735695,,45752.13837000774,0.0,0.0,45752.21362197772
diff --git a/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/report.md b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/report.md
new file mode 100644
index 0000000..d3b3f8e
--- /dev/null
+++ b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/report.md
@@ -0,0 +1,31 @@
+# DeepSeek-V4-Pro Pro6000D Two-Node SGLang Quick Map
+
+Profiler: disabled. Speculative decoding: disabled.
+
+## Aggregate results
+
+| Case | Suite / role | Stage | ISL | OSL | C | Reps | Total TPS | CV | Output TPS | TTFT P95 | TPOT P95 | E2E P95 | Status |
+|---|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|---|
+| balanced_32k_to_1k_c8 | fixed / - | balanced | 32768 | 1024 | 8 | 1/1 | 2101.69 | -% | 63.69 | 82083.65 ms | 109.32 ms | 128588.90 ms | COMPLETED |
+| decode_latency_1k_to_1k_c1 | fixed / - | decode_latency | 1024 | 1024 | 1 | 1/1 | 62.83 | -% | 31.41 | 363.27 ms | 31.47 ms | 32555.29 ms | COMPLETED |
+| decode_throughput_1k_to_1k_c16 | fixed / - | decode_throughput | 1024 | 1024 | 16 | 1/1 | 590.58 | -% | 295.29 | 4950.27 ms | 50.02 ms | 55443.82 ms | COMPLETED |
+| decode_throughput_1k_to_1k_c32 | fixed / - | decode_throughput | 1024 | 1024 | 32 | 1/1 | 923.37 | -% | 461.68 | 8021.85 ms | 63.31 ms | 70933.19 ms | COMPLETED |
+| decode_throughput_1k_to_1k_c64 | fixed / - | decode_throughput | 1024 | 1024 | 64 | 1/1 | 1294.83 | -% | 647.42 | 12716.49 ms | 93.44 ms | 101163.20 ms | COMPLETED |
+| long_prefill_latency_128k_c1 | fixed / - | prefill_latency | 131072 | 1 | 1 | 1/1 | 2710.18 | -% | 0.02 | 48344.46 ms | 0.00 ms | 48344.53 ms | COMPLETED |
+| mid_prefill_latency_32k_c1 | fixed / - | prefill_latency | 32768 | 1 | 1 | 1/1 | 2652.84 | -% | 0.08 | 12334.69 ms | 0.00 ms | 12334.76 ms | COMPLETED |
+| mid_prefill_throughput_32k_c16 | fixed / - | prefill_throughput | 32768 | 1 | 16 | 1/1 | 3112.86 | -% | 0.09 | 162087.46 ms | 0.00 ms | 162087.50 ms | COMPLETED |
+| short_prefill_latency_1k_c1 | fixed / - | prefill_latency | 1024 | 1 | 1 | 1/1 | 1971.59 | -% | 1.92 | 501.90 ms | 0.00 ms | 501.95 ms | COMPLETED |
+| decode_control_1k_to_1k_c32 | mixed / control | mixed_interference | 1024 | 1024 | 32 | 1/1 | 911.37 | -% | 455.68 | 9442.64 ms | 65.88 ms | 72007.64 ms | COMPLETED |
+| decode_with_128k_prefill_1k_to_1k_c32 | mixed / decode_background | mixed_interference | 1024 | 1024 | 32 | 1/1 | 691.89 | -% | 345.95 | 10193.90 ms | 109.73 ms | 117629.64 ms | COMPLETED |
+| long_prefill_injection_128k_to_1_c1 | mixed / prefill_injection | mixed_interference | 131072 | 1 | 1 | 1/1 | 2859.75 | -% | 0.02 | 45752.14 ms | 0.00 ms | 45752.21 ms | COMPLETED |
+
+## Mixed-interference A/B
+
+| Metric | Control | With 128K prefill | Change |
+|---|---:|---:|---:|
+| Output TPS | 455.68 | 345.95 | -24.08% |
+| TTFT P95 (ms) | 9442.64 | 10193.90 | 7.96% |
+| TPOT P95 (ms) | 65.88 | 109.73 | 66.55% |
+| E2E P95 (ms) | 72007.64 | 117629.64 | 63.36% |
+
+The fixed map is descriptive and never stops on SLO. With one repetition, CV is intentionally unavailable.
diff --git a/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/run.log b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/run.log
new file mode 100644
index 0000000..7dc0446
--- /dev/null
+++ b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/run.log
@@ -0,0 +1,53 @@
+[2026-07-30 22:09:49] Starting worker node=10.101.0.13 rank=1 container=dsv4pro_pro6000d_2node_sglang_tp16_quick_map_worker
+[2026-07-30 22:09:55] Starting head node=10.101.0.11 rank=0 container=dsv4pro_pro6000d_2node_sglang_tp16_quick_map_head
+[2026-07-30 22:09:56] Waiting for SGLang health at 10.101.0.11:30002
+[2026-07-30 22:10:55] Still starting: checks=12/600
+[2026-07-30 22:12:00] Still starting: checks=24/600
+[2026-07-30 22:13:05] Still starting: checks=36/600
+[2026-07-30 22:14:09] Still starting: checks=48/600
+[2026-07-30 22:15:14] Still starting: checks=60/600
+[2026-07-30 22:15:20] SGLang is healthy after 61 checks
+[2026-07-30 22:15:20] Verified worker NCCL transport: NET/IB with mlx5_0 and mlx5_3
+[2026-07-30 22:15:21] Verified head NCCL transport: NET/IB with mlx5_0 and mlx5_3
+[2026-07-30 22:15:21] Validating scenario file: /data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/quick_map_scenarios.tsv
+short_prefill_latency_1k_c1: stage=prefill_latency isl=1024 osl=1 c=1 prompts=1 warmup=1
+mid_prefill_latency_32k_c1: stage=prefill_latency isl=32768 osl=1 c=1 prompts=1 warmup=0
+long_prefill_latency_128k_c1: stage=prefill_latency isl=131072 osl=1 c=1 prompts=1 warmup=0
+mid_prefill_throughput_32k_c16: stage=prefill_throughput isl=32768 osl=1 c=16 prompts=16 warmup=0
+decode_latency_1k_to_1k_c1: stage=decode_latency isl=1024 osl=1024 c=1 prompts=1 warmup=1
+decode_throughput_1k_to_1k_c16: stage=decode_throughput isl=1024 osl=1024 c=16 prompts=16 warmup=1
+decode_throughput_1k_to_1k_c32: stage=decode_throughput isl=1024 osl=1024 c=32 prompts=32 warmup=1
+decode_throughput_1k_to_1k_c64: stage=decode_throughput isl=1024 osl=1024 c=64 prompts=64 warmup=1
+balanced_32k_to_1k_c8: stage=balanced isl=32768 osl=1024 c=8 prompts=8 warmup=0
+[2026-07-30 22:15:21] START case=short_prefill_latency_1k_c1 rep=1 isl=1024 osl=1 c=1
+[2026-07-30 22:15:55] DONE case=short_prefill_latency_1k_c1 rep=1 elapsed=34s
+[2026-07-30 22:16:00] START case=mid_prefill_latency_32k_c1 rep=1 isl=32768 osl=1 c=1
+[2026-07-30 22:16:46] DONE case=mid_prefill_latency_32k_c1 rep=1 elapsed=46s
+[2026-07-30 22:16:51] START case=long_prefill_latency_128k_c1 rep=1 isl=131072 osl=1 c=1
+[2026-07-30 22:18:13] DONE case=long_prefill_latency_128k_c1 rep=1 elapsed=81s
+[2026-07-30 22:18:18] START case=mid_prefill_throughput_32k_c16 rep=1 isl=32768 osl=1 c=16
+[2026-07-30 22:21:39] DONE case=mid_prefill_throughput_32k_c16 rep=1 elapsed=201s
+[2026-07-30 22:21:44] START case=decode_latency_1k_to_1k_c1 rep=1 isl=1024 osl=1024 c=1
+[2026-07-30 22:22:51] DONE case=decode_latency_1k_to_1k_c1 rep=1 elapsed=66s
+[2026-07-30 22:22:56] START case=decode_throughput_1k_to_1k_c16 rep=1 isl=1024 osl=1024 c=16
+[2026-07-30 22:24:25] DONE case=decode_throughput_1k_to_1k_c16 rep=1 elapsed=89s
+[2026-07-30 22:24:30] START case=decode_throughput_1k_to_1k_c32 rep=1 isl=1024 osl=1024 c=32
+[2026-07-30 22:26:15] DONE case=decode_throughput_1k_to_1k_c32 rep=1 elapsed=105s
+[2026-07-30 22:26:20] START case=decode_throughput_1k_to_1k_c64 rep=1 isl=1024 osl=1024 c=64
+[2026-07-30 22:28:36] DONE case=decode_throughput_1k_to_1k_c64 rep=1 elapsed=136s
+[2026-07-30 22:28:41] START case=balanced_32k_to_1k_c8 rep=1 isl=32768 osl=1024 c=8
+[2026-07-30 22:31:23] DONE case=balanced_32k_to_1k_c8 rep=1 elapsed=161s
+[2026-07-30 22:31:28] Fixed quick map complete: completed=9/9 failed=0
+[2026-07-30 22:31:28] START case=decode_control_1k_to_1k_c32 rep=1 isl=1024 osl=1024 c=32
+[2026-07-30 22:34:26] DONE case=decode_control_1k_to_1k_c32 rep=1 elapsed=178s
+[2026-07-30 22:34:26] START mixed background rep=1; injection delay=10s
+[2026-07-30 22:34:26] START case=decode_with_128k_prefill_1k_to_1k_c32 rep=1 isl=1024 osl=1024 c=32
+[2026-07-30 22:35:06] START case=long_prefill_injection_128k_to_1_c1 rep=1 isl=131072 osl=1 c=1
+[2026-07-30 22:36:24] DONE case=long_prefill_injection_128k_to_1_c1 rep=1 elapsed=78s
+[2026-07-30 22:38:09] DONE case=decode_with_128k_prefill_1k_to_1k_c32 rep=1 elapsed=223s
+[2026-07-30 22:38:14] Mixed-interference A/B complete: failed_repetitions=0
+[2026-07-30 22:38:20] Stopped head container=dsv4pro_pro6000d_2node_sglang_tp16_quick_map_head on 10.101.0.11
+[2026-07-30 22:38:22] Stopped worker container=dsv4pro_pro6000d_2node_sglang_tp16_quick_map_worker on 10.101.0.13
+summarized 12 case runs into /data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916
+[2026-07-30 22:38:22] Quick performance map complete: status=COMPLETED
+[2026-07-30 22:38:22] Results: /data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916
diff --git a/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/run_manifest.json b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/run_manifest.json
new file mode 100644
index 0000000..04c3d48
--- /dev/null
+++ b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/run_manifest.json
@@ -0,0 +1,43 @@
+{
+ "schema_version": 1,
+ "workflow_stage": "quick_performance_map",
+ "run_id": "dsv4pro-phase1-full-20260730-220916",
+ "status": "COMPLETED",
+ "started_at": "2026-07-30T22:09:47+08:00",
+ "updated_at": "2026-07-30T22:38:22+08:00",
+ "suites": [
+ "fixed",
+ "mixed"
+ ],
+ "engine": "sglang",
+ "model_name": "DeepSeek-V4-Pro",
+ "model_path": "/data/hf_models/DeepSeek-V4-Pro",
+ "docker_image": "lmsysorg/sglang:nightly-dev-cu13-20260720-b3570a45",
+ "head_node": "10.101.0.11",
+ "worker_node": "10.101.0.13",
+ "head_ip": "10.101.0.11",
+ "sglang_port": 30002,
+ "dist_init_port": 20002,
+ "tp_size": 16,
+ "ep_size": 2,
+ "nnodes": 2,
+ "mem_fraction_static": 0.9,
+ "cuda_graph_max_bs_decode": 64,
+ "max_running_requests": 256,
+ "nccl_socket_ifname": "eth0",
+ "nccl_ib_hca": "=mlx5_0:1,mlx5_3:1",
+ "nccl_cross_nic": "1",
+ "enable_rdma": true,
+ "require_nccl_ib": true,
+ "rdma_device_paths": "/dev/infiniband/rdma_cm,/dev/infiniband/uverbs0,/dev/infiniband/uverbs3",
+ "git_commit": "75182c6ededf94511c6678552aaf489b9fc6fba6",
+ "git_dirty": false,
+ "scenario_file": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/quick_map_scenarios.tsv",
+ "case_ids": "",
+ "notes": [
+ "The fixed quick map does not stop on SLO.",
+ "Profiler is disabled; these results are eligible for performance comparison.",
+ "Speculative decoding is not enabled."
+ ],
+ "ended_at": "2026-07-30T22:38:22+08:00"
+}
diff --git a/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/server/head_nccl_transport.log b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/server/head_nccl_transport.log
new file mode 100644
index 0000000..47315f9
--- /dev/null
+++ b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/server/head_nccl_transport.log
@@ -0,0 +1,1012 @@
+/usr/local/lib/python3.12/dist-packages/torchao/quantization/quant_api.py:1731: SyntaxWarning: invalid escape sequence '\.'
+ """Configuration class for applying different quantization configs to modules or parameters based on their fully qualified names (FQNs).
+/sgl-workspace/sglang/python/sglang/launch_server.py:56: UserWarning: 'python -m sglang.launch_server' is still supported, but 'sglang serve' is the recommended entrypoint.
+ Example: sglang serve --model-path [options]
+ warnings.warn(
+[2026-07-30 14:10:19] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:19] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:19] Breakable CUDA graph is incompatible with DeepSeek-V4 (heavy capture-pool memory pressure); disabling prefill CUDA graph.
+[2026-07-30 14:10:19] Use dsv4 attention backend for DeepseekV4ForCausalLM, setting page_size to 256.
+[2026-07-30 14:10:19] Setting swa_full_tokens_ratio to 0.1 for DeepseekV4ForCausalLM.
+[2026-07-30 14:10:19] Setting KV cache dtype to fp8_e4m3 for DeepseekV4ForCausalLM.
+[2026-07-30 14:10:19] Use flashinfer_mxfp4 as MoE runner backend on SM120 for DeepseekV4
+[2026-07-30 14:10:22] server_args=ServerArgs(model_path='/data/hf_models/DeepSeek-V4-Pro', tokenizer_path='/data/hf_models/DeepSeek-V4-Pro', tokenizer_mode='auto', tokenizer_backend='huggingface', tokenizer_worker_num=1, detokenizer_worker_num=1, skip_tokenizer_init=False, load_format='auto', model_loader_extra_config='{}', trust_remote_code=True, context_length=None, is_embedding=False, enable_multimodal=None, revision=None, model_impl='auto', model_config_parser='auto', json_model_override_args='{}', dtype='auto', quantization=None, quantization_param_path=None, kv_cache_dtype='fp8_e4m3', enable_fp32_lm_head=False, modelopt_quant=None, modelopt_checkpoint_restore_path=None, modelopt_checkpoint_save_path=None, modelopt_export_path=None, quantize_and_serve=False, rl_quant_profile=None, enable_tf32_matmul=False, mem_fraction_static=0.9, max_running_requests=256, max_queued_requests=None, max_total_tokens=None, chunked_prefill_size=8192, enable_dynamic_chunking=False, max_prefill_tokens=16384, prefill_max_requests=None, schedule_policy='fcfs', enable_priority_scheduling=False, disable_priority_preemption=False, default_priority_value=None, abort_on_priority_when_disabled=False, schedule_low_priority_values_first=False, priority_scheduling_preemption_threshold=10, retraction_policy='length', schedule_conservativeness=1.0, page_size=256, swa_full_tokens_ratio=0.1, disable_hybrid_swa_memory=False, radix_eviction_policy='lru', prefill_only_disable_kv_cache=False, disable_radix_cache=False, enable_page_major_kv_layout=False, enable_unified_memory=False, disable_chunked_prefix_cache=False, disable_overlap_schedule=False, num_continuous_decode_steps=1, scheduler_recv_interval=1, enable_mixed_chunk=False, nccl_port=None, dist_timeout=None, dist_init_addr='10.101.0.11:20002', nnodes=2, node_rank=0, tp_size=16, dcp_size=1, pp_size=1, pp_max_micro_batch_size=None, pp_async_batch_depth=0, dp_size=1, load_balance_method='round_robin', attn_cp_size=1, moe_dp_size=1, enable_prefill_cp=False, cp_strategy=None, enable_dsa_cache_layer_split=False, enable_dsa_prefill_context_parallel=False, dsa_prefill_cp_mode='round-robin-split', enable_prefill_context_parallel=False, prefill_cp_mode='in-seq-split', enable_dp_attention=False, enable_dp_attention_local_control_broadcast=False, enable_dp_lm_head=False, enable_attn_tp_input_scattered=False, disable_attn_tp_gather=False, enable_p2p_check=False, device='cuda', base_gpu_id=0, gpu_id_step=1, random_seed=115970149, watchdog_timeout=300, soft_watchdog_timeout=None, sleep_on_idle=False, use_ray=False, custom_sigquit_handler=None, numa_node=None, gc_threshold=None, host='0.0.0.0', port=30002, fastapi_root_path='', smg_grpc_mode=False, grpc_mode=False, grpc_port=None, skip_server_warmup=False, warmups=None, enable_http2=False, ssl_keyfile=None, ssl_certfile=None, ssl_ca_certs=None, ssl_keyfile_password=None, enable_ssl_refresh=False, api_key=None, admin_api_key=None, served_model_name='/data/hf_models/DeepSeek-V4-Pro', weight_version='default', chat_template=None, hf_chat_template_name=None, completion_template=None, file_storage_path='sglang_storage', enable_cache_report=False, reasoning_parser=None, default_chat_template_kwargs=None, strip_thinking_cache=False, enable_strict_thinking=False, tool_call_parser=None, tool_server=None, sampling_defaults='model', asr_max_buffer_seconds=60, asr_max_concurrent_sessions=32, preferred_sampling_params=None, allow_auto_truncate=False, stream_interval=1, batch_notify_size=16, stream_response_default_include_usage=False, incremental_streaming_output=False, enable_streaming_session=False, enable_session_radix_cache=False, log_level='info', log_level_http=None, log_requests=False, log_requests_level=2, log_requests_format='text', log_requests_target=None, uvicorn_access_log_exclude_prefixes=[], crash_dump_folder=None, show_time_cost=False, enable_metrics=False, smg_http_sidecar_port=None, enable_mfu_metrics=False, enable_metrics_for_all_schedulers=False, load_snapshot_publish_interval=15, tokenizer_metrics_custom_labels_header='x-custom-labels', tokenizer_metrics_allowed_custom_labels=None, extra_metric_labels=None, bucket_time_to_first_token=None, bucket_inter_token_latency=None, bucket_e2e_request_latency=None, prompt_tokens_buckets=None, generation_tokens_buckets=None, gc_warning_threshold_secs=0.0, decode_log_interval=40, enable_request_time_stats_logging=False, kv_events_config=None, enable_forward_pass_metrics=False, forward_pass_metrics_worker_id='', forward_pass_metrics_ipc_name=None, enable_trace=False, trace_modules='request', otlp_traces_endpoint='localhost:4317', export_metrics_to_file=False, export_metrics_to_file_dir=None, stat_loggers=None, constrained_json_whitespace_pattern=None, constrained_json_disable_any_whitespace=False, attention_backend='dsv4', decode_attention_backend=None, prefill_attention_backend=None, sampling_backend='flashinfer', grammar_backend='xgrammar', radix_cache_backend=None, mm_attention_backend=None, fp8_gemm_runner_backend='auto', fp4_gemm_runner_backend='auto', bf16_gemm_backend='auto', dsa_prefill_backend=None, dsa_decode_backend=None, dsa_paged_mqa_logits_backend='auto', dsa_topk_backend='sgl-kernel', disable_flashinfer_autotune=False, mamba_backend='triton', cuda_graph_config=CudaGraphConfig(decode=PhaseConfig(backend='full', max_bs=64, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], tc_compiler='eager', full_prefill_max_req=None), prefill=PhaseConfig(backend='disabled', max_bs=8192, bs=[4, 8, 12, 16, 20, 24, 28, 32, 48, 64, 80, 96, 112, 128, 144, 160, 176, 192, 208, 224, 240, 256, 288, 320, 352, 384, 416, 448, 480, 512, 576, 640, 704, 768, 832, 896, 960, 1024, 1280, 1536, 1792, 2048, 2304, 2560, 2816, 3072, 3328, 3584, 3840, 4096, 4608, 5120, 5632, 6144, 6656, 7168, 7680, 8192], tc_compiler='eager', full_prefill_max_req=None)), cuda_graph_backend_decode=None, cuda_graph_backend_prefill=None, cuda_graph_max_bs_decode=64, cuda_graph_max_bs_prefill=None, cuda_graph_bs_decode=None, cuda_graph_bs_prefill=None, cuda_graph_tc_compiler=None, disable_prefill_cuda_graph=False, disable_decode_cuda_graph=False, disable_cuda_graph=False, disable_cuda_graph_padding=False, enable_profile_cuda_graph=False, enable_cudagraph_gc=False, debug_cuda_graph=False, enable_layerwise_nvtx_marker=False, enable_nccl_nvls=False, enable_symm_mem=False, triton_attention_reduce_in_fp32=False, triton_attention_num_kv_splits=8, triton_attention_split_tile_size=None, flashinfer_mla_disable_ragged=False, enable_fused_qk_norm_rope=False, enable_precise_embedding_interpolation=False, enable_fused_moe_sum_all_reduce=False, enable_deepseek_v4_fp4_indexer=False, disable_custom_all_reduce=False, enable_mscclpp=False, enable_torch_symm_mem=False, pre_warm_nccl=False, enable_quant_communications=False, enable_flashinfer_allreduce_fusion=False, enforce_disable_flashinfer_allreduce_fusion=False, flashinfer_allreduce_fusion_backend=None, enable_aiter_allreduce_fusion=False, enable_torch_compile=False, enable_torch_compile_debug_mode=False, torch_compile_max_bs=32, torchao_config='', speculative_algorithm=None, speculative_draft_model_path=None, speculative_draft_model_revision=None, speculative_draft_load_format=None, speculative_num_steps=None, speculative_eagle_topk=None, speculative_num_draft_tokens=None, speculative_dflash_block_size=None, speculative_dspark_block_size=None, speculative_dspark_sps_table_path=None, speculative_dspark_confidence_sts_path=None, speculative_dspark_align_verify_tokens_to_graph_tier=False, speculative_accept_threshold_single=1.0, speculative_accept_threshold_acc=1.0, speculative_use_rejection_sampling=False, speculative_token_map=None, speculative_attention_mode='prefill', speculative_draft_attention_backend=None, speculative_draft_window_size=None, speculative_moe_runner_backend='flashinfer_mxfp4', speculative_moe_a2a_backend=None, speculative_draft_model_quantization=None, speculative_skip_dp_mlp_sync=False, enable_multi_layer_eagle=False, speculative_adaptive=False, speculative_adaptive_config=None, decoupled_spec_bind_endpoint=None, decoupled_spec_connect_endpoints=None, decoupled_spec_rank=None, decoupled_spec_role='null', spec_trace_dir=None, speculative_ngram_min_bfs_breadth=1, speculative_ngram_max_bfs_breadth=10, speculative_ngram_match_type='BFS', speculative_ngram_max_trie_depth=18, speculative_ngram_capacity=10000000, speculative_ngram_external_corpus_path=None, speculative_ngram_external_sam_budget=0, speculative_ngram_external_corpus_max_tokens=10000000, ep_size=2, moe_a2a_backend='none', moe_runner_backend='flashinfer_mxfp4', flashinfer_mxfp4_moe_precision='default', deepep_mode='auto', fuseep_mode=2, deepep_dispatcher_output_dtype='auto', ep_num_redundant_experts=0, ep_dispatch_algorithm=None, init_expert_location='trivial', enable_eplb=False, eplb_algorithm='auto', eplb_rebalance_num_iterations=1000, eplb_rebalance_layers_per_chunk=None, eplb_min_rebalancing_utilization_threshold=1.0, expert_distribution_recorder_mode=None, expert_distribution_recorder_buffer_size=1000, enable_expert_distribution_metrics=False, deepep_config=None, moe_dense_tp_size=None, elastic_ep_backend=None, enable_elastic_expert_backup=False, mooncake_ib_device=None, enable_waterfill=False, ep_join_mode=None, ep_join_rank_offset=0, elastic_ep_initial_size=None, max_ep_size=None, elastic_ep_scale_timeout=600, elastic_ep_rejoin=False, disable_flashinfer_cutlass_moe_fp4_allgather=False, disable_shared_experts_fusion=False, enforce_shared_experts_fusion=False, max_mamba_cache_size=None, mamba_ssm_dtype=None, enable_mamba_cache_stochastic_rounding=False, mamba_cache_philox_rounds=0, mamba_full_memory_ratio=0.9, mamba_radix_cache_strategy='auto', uses_mamba_radix_cache=False, mamba_track_interval=256, enable_int8_mamba_checkpoint=False, int8_mamba_ckpt_size=None, linear_attn_backend='triton', linear_attn_decode_backend=None, linear_attn_prefill_backend=None, enable_linear_replayssm=False, linear_replayssm_cache_len=16, enable_hierarchical_cache=False, hicache_ratio=2.0, hicache_size=0, hicache_write_policy='write_through', hicache_io_backend='kernel', hicache_mem_layout='page_first', hicache_storage_backend=None, hicache_storage_prefetch_policy='timeout', hicache_storage_backend_extra_config=None, enable_hisparse=False, hisparse_config=None, enable_broadcast_mm_inputs_process=False, enable_prefix_mm_cache=False, mm_enable_dp_encoder=False, mm_process_config={}, limit_mm_data_per_request=None, enable_mm_global_cache=False, disable_fast_image_processor=False, mm_feature_transport='cpu', keep_mm_feature_on_device=False, enable_lora=None, enable_lora_overlap_loading=None, max_lora_rank=None, lora_target_modules=None, lora_paths=None, max_loaded_loras=None, max_loras_per_batch=8, lora_eviction_policy='lru', lora_backend='csgmv', max_lora_chunk_size=16, experts_shared_outer_loras=None, lora_use_virtual_experts=False, lora_strict_loading=False, lora_drain_wait_threshold=0.0, enable_two_batch_overlap=False, enable_single_batch_overlap=False, tbo_token_distribution_threshold=0.48, cpu_offload_gb=0, offload_group_size=-1, offload_num_in_group=1, offload_prefetch_step=1, offload_mode='cpu', enable_lmcache=False, lmcache_config_file=None, enable_flexkv=False, flexkv_config_file=None, kt_weight_path=None, kt_method='AMXINT4', kt_cpuinfer=None, kt_threadpool_count=2, kt_num_gpu_experts=None, kt_max_deferred_experts_per_token=None, dllm_algorithm=None, dllm_algorithm_config=None, dllm_fdfo=True, disaggregation_mode='null', disaggregation_transfer_backend='mooncake', disaggregation_bootstrap_port=8998, disaggregation_ib_device=None, disaggregation_decode_enable_radix_cache=False, disaggregation_decode_enable_offload_kvcache=False, num_reserved_decode_tokens=512, disaggregation_decode_extra_slots=None, disaggregation_decode_polling_interval=1, optimistic_prefill_attempts=0, encoder_only=False, language_only=False, encoder_transfer_backend='zmq_to_scheduler', encoder_urls=[], encoder_bootstrap_port=8997, encoder_register_urls=[], enable_adaptive_dispatch_to_encoder=False, enable_pdmux=False, pdmux_config_path=None, sm_group_num=8, custom_weight_loader=[], weight_loader_disable_mmap=False, weight_loader_prefetch_checkpoints=False, weight_loader_prefetch_num_threads=4, weight_loader_drop_cache_after_load=False, remote_instance_weight_loader_seed_instance_ip=None, remote_instance_weight_loader_seed_instance_service_port=None, remote_instance_weight_loader_send_weights_group_ports=None, remote_instance_weight_loader_backend='nccl', remote_instance_weight_loader_start_seed_via_transfer_engine=False, engine_info_bootstrap_port=6789, modelexpress_config=None, download_dir=None, model_checksum=None, delete_ckpt_after_loading=False, decrypted_config_file=None, decrypted_draft_config_file=None, checkpoint_engine_wait_weights_before_ready=False, enable_prefill_delayer=False, prefill_delayer_max_delay_passes=30, prefill_delayer_token_usage_low_watermark=None, prefill_delayer_forward_passes_buckets=None, prefill_delayer_wait_seconds_buckets=None, prefill_delayer_queue_min_ratio=None, prefill_delayer_max_delay_ms=None, min_free_slots_delay=None, enable_deterministic_inference=False, rl_on_policy_target=None, kv_canary='none', kv_canary_real_data='none', kv_canary_sweep_interval=0, enable_dynamic_batch_tokenizer=False, dynamic_batch_tokenizer_batch_size=32, dynamic_batch_tokenizer_batch_timeout=0.002, enable_tokenizer_batch_encode=False, disable_tokenizer_batch_decode=False, debug_tensor_dump_output_folder=None, debug_tensor_dump_layers=None, debug_tensor_dump_input_file=None, enable_memory_saver=False, enable_weights_cpu_backup=False, enable_draft_weights_cpu_backup=False, enable_custom_logit_processor=False, enable_return_hidden_states=False, enable_return_routed_experts=False, enable_return_indexer_topk=False, disable_outlines_disk_cache=False, enable_mis=False, forward_hooks=None, msprobe_dump_config=None)
+[2026-07-30 14:10:22] User lacks permission to set NUMA affinity, skipping NUMA node configuration for GPU. If using docker, try adding --cap-add SYS_NICE to your docker run command.
+[2026-07-30 14:10:22] User lacks permission to set NUMA affinity, skipping NUMA node configuration for GPU. If using docker, try adding --cap-add SYS_NICE to your docker run command.
+[2026-07-30 14:10:22] User lacks permission to set NUMA affinity, skipping NUMA node configuration for GPU. If using docker, try adding --cap-add SYS_NICE to your docker run command.
+[2026-07-30 14:10:22] User lacks permission to set NUMA affinity, skipping NUMA node configuration for GPU. If using docker, try adding --cap-add SYS_NICE to your docker run command.
+[2026-07-30 14:10:22] User lacks permission to set NUMA affinity, skipping NUMA node configuration for GPU. If using docker, try adding --cap-add SYS_NICE to your docker run command.
+[2026-07-30 14:10:22] User lacks permission to set NUMA affinity, skipping NUMA node configuration for GPU. If using docker, try adding --cap-add SYS_NICE to your docker run command.
+[2026-07-30 14:10:22] User lacks permission to set NUMA affinity, skipping NUMA node configuration for GPU. If using docker, try adding --cap-add SYS_NICE to your docker run command.
+[2026-07-30 14:10:22] User lacks permission to set NUMA affinity, skipping NUMA node configuration for GPU. If using docker, try adding --cap-add SYS_NICE to your docker run command.
+[2026-07-30 14:10:22] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:22] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:24] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:24] No HuggingFace chat template found
+[2026-07-30 14:10:24] No chat template found, defaulting to 'string' content format
+[2026-07-30 14:10:35 TP1 EP0] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:35 TP1 EP0] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:37 TP0 EP0] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:37 TP0 EP0] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:37 TP3 EP0] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:37 TP3 EP0] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:37 TP7 EP0] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:37 TP7 EP0] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:37 TP6 EP0] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:37 TP6 EP0] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:37 TP5 EP0] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:37 TP5 EP0] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:37 TP2 EP0] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:37 TP2 EP0] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:37 TP4 EP0] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:37 TP4 EP0] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:38 TP1 EP0] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:38 TP1 EP0] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:38 TP1 EP0] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:38 TP1 EP0] Init torch distributed begin.
+[2026-07-30 14:10:39] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:39 TP0 EP0] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:39 TP3 EP0] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:39 TP0 EP0] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:39 TP0 EP0] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:39 TP6 EP0] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:39 TP5 EP0] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:39 TP2 EP0] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:39 TP7 EP0] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:39 TP0 EP0] Init torch distributed begin.
+[2026-07-30 14:10:39 TP3 EP0] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:39 TP3 EP0] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:39 TP3 EP0] Init torch distributed begin.
+[2026-07-30 14:10:39 TP4 EP0] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:39 TP6 EP0] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:39 TP6 EP0] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:39 TP5 EP0] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:39 TP5 EP0] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:39 TP2 EP0] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:39 TP2 EP0] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:40 TP7 EP0] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:40 TP7 EP0] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:40 TP6 EP0] Init torch distributed begin.
+[2026-07-30 14:10:40 TP5 EP0] Init torch distributed begin.
+[2026-07-30 14:10:40 TP2 EP0] Init torch distributed begin.
+[2026-07-30 14:10:40 TP7 EP0] Init torch distributed begin.
+[2026-07-30 14:10:40 TP4 EP0] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:40 TP4 EP0] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:40 TP4 EP0] Init torch distributed begin.
+[2026-07-30 14:10:47 TP0 EP0] sglang is using nccl==2.28.9
+gpu1:237:237 [0] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so
+gpu1:237:237 [0] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu1:237:237 [0] NCCL INFO Bootstrap: Using eth0:10.101.0.11<0>
+gpu1:238:238 [1] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so
+gpu1:238:238 [1] NCCL INFO cudaDriverVersion 13000
+gpu1:242:242 [5] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so
+gpu1:242:242 [5] NCCL INFO cudaDriverVersion 13000
+gpu1:238:238 [1] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu1:242:242 [5] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu1:238:238 [1] NCCL INFO Bootstrap: Using eth0:10.101.0.11<0>
+gpu1:242:242 [5] NCCL INFO Bootstrap: Using eth0:10.101.0.11<0>
+gpu1:238:238 [1] NCCL INFO NCCL version 2.28.9+cuda13.0
+gpu1:242:242 [5] NCCL INFO NCCL version 2.28.9+cuda13.0
+gpu1:237:237 [0] NCCL INFO cudaDriverVersion 13000
+gpu1:243:243 [6] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so
+gpu1:243:243 [6] NCCL INFO cudaDriverVersion 13000
+gpu1:237:237 [0] NCCL INFO NCCL version 2.28.9+cuda13.0
+gpu1:244:244 [7] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so
+gpu1:239:239 [2] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so
+gpu1:244:244 [7] NCCL INFO cudaDriverVersion 13000
+gpu1:239:239 [2] NCCL INFO cudaDriverVersion 13000
+gpu1:240:240 [3] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so
+gpu1:240:240 [3] NCCL INFO cudaDriverVersion 13000
+gpu1:241:241 [4] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so
+gpu1:241:241 [4] NCCL INFO cudaDriverVersion 13000
+gpu1:244:244 [7] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu1:243:243 [6] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu1:239:239 [2] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu1:240:240 [3] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu1:241:241 [4] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu1:244:244 [7] NCCL INFO Bootstrap: Using eth0:10.101.0.11<0>
+gpu1:244:244 [7] NCCL INFO NCCL version 2.28.9+cuda13.0
+gpu1:243:243 [6] NCCL INFO Bootstrap: Using eth0:10.101.0.11<0>
+gpu1:243:243 [6] NCCL INFO NCCL version 2.28.9+cuda13.0
+gpu1:239:239 [2] NCCL INFO Bootstrap: Using eth0:10.101.0.11<0>
+gpu1:239:239 [2] NCCL INFO NCCL version 2.28.9+cuda13.0
+gpu1:240:240 [3] NCCL INFO Bootstrap: Using eth0:10.101.0.11<0>
+gpu1:241:241 [4] NCCL INFO Bootstrap: Using eth0:10.101.0.11<0>
+gpu1:240:240 [3] NCCL INFO NCCL version 2.28.9+cuda13.0
+gpu1:241:241 [4] NCCL INFO NCCL version 2.28.9+cuda13.0
+gpu1:242:242 [5] NCCL INFO NET/Plugin: Could not find: libnccl-net.so
+gpu1:242:242 [5] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu1:238:238 [1] NCCL INFO NET/Plugin: Could not find: libnccl-net.so
+gpu1:242:242 [5] NCCL INFO NCCL_IB_HCA set to =mlx5_0:1,mlx5_3:1
+gpu1:238:238 [1] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu1:238:238 [1] NCCL INFO NCCL_IB_HCA set to =mlx5_0:1,mlx5_3:1
+gpu1:238:238 [1] NCCL INFO NET/IB : Using [0]mlx5_0:1/RoCE [1]mlx5_3:1/RoCE [RO]; OOB eth0:10.101.0.11<0>
+gpu1:238:238 [1] NCCL INFO Initialized NET plugin IB
+gpu1:238:238 [1] NCCL INFO Assigned NET plugin IB to comm
+gpu1:238:238 [1] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu1:238:238 [1] NCCL INFO Using network IB
+gpu1:238:238 [1] NCCL INFO DMA-BUF is available on GPU device 1
+gpu1:238:238 [1] NCCL INFO ncclCommInitRank comm 0x4dcaf9f0 rank 1 nranks 16 cudaDev 1 nvmlDev 1 busId 1a000 commId 0x92ed644083b4525 - Init START
+gpu1:243:243 [6] NCCL INFO NET/Plugin: Could not find: libnccl-net.so
+gpu1:237:237 [0] NCCL INFO NET/Plugin: Could not find: libnccl-net.so
+gpu1:243:243 [6] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu1:239:239 [2] NCCL INFO NET/Plugin: Could not find: libnccl-net.so
+gpu1:243:243 [6] NCCL INFO NCCL_IB_HCA set to =mlx5_0:1,mlx5_3:1
+gpu1:237:237 [0] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu1:239:239 [2] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu1:237:237 [0] NCCL INFO NCCL_IB_HCA set to =mlx5_0:1,mlx5_3:1
+gpu1:239:239 [2] NCCL INFO NCCL_IB_HCA set to =mlx5_0:1,mlx5_3:1
+gpu1:242:242 [5] NCCL INFO NET/IB : Using [0]mlx5_0:1/RoCE [1]mlx5_3:1/RoCE [RO]; OOB eth0:10.101.0.11<0>
+gpu1:242:242 [5] NCCL INFO Initialized NET plugin IB
+gpu1:242:242 [5] NCCL INFO Assigned NET plugin IB to comm
+gpu1:242:242 [5] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu1:242:242 [5] NCCL INFO Using network IB
+gpu1:244:244 [7] NCCL INFO NET/Plugin: Could not find: libnccl-net.so
+gpu1:244:244 [7] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu1:243:243 [6] NCCL INFO NET/IB : Using [0]mlx5_0:1/RoCE [1]mlx5_3:1/RoCE [RO]; OOB eth0:10.101.0.11<0>
+gpu1:243:243 [6] NCCL INFO Initialized NET plugin IB
+gpu1:243:243 [6] NCCL INFO Assigned NET plugin IB to comm
+gpu1:243:243 [6] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu1:243:243 [6] NCCL INFO Using network IB
+gpu1:244:244 [7] NCCL INFO NCCL_IB_HCA set to =mlx5_0:1,mlx5_3:1
+gpu1:239:239 [2] NCCL INFO NET/IB : Using [0]mlx5_0:1/RoCE [1]mlx5_3:1/RoCE [RO]; OOB eth0:10.101.0.11<0>
+gpu1:239:239 [2] NCCL INFO Initialized NET plugin IB
+gpu1:239:239 [2] NCCL INFO Assigned NET plugin IB to comm
+gpu1:239:239 [2] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu1:239:239 [2] NCCL INFO Using network IB
+gpu1:244:244 [7] NCCL INFO NET/IB : Using [0]mlx5_0:1/RoCE [1]mlx5_3:1/RoCE [RO]; OOB eth0:10.101.0.11<0>
+gpu1:244:244 [7] NCCL INFO Initialized NET plugin IB
+gpu1:244:244 [7] NCCL INFO Assigned NET plugin IB to comm
+gpu1:244:244 [7] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu1:244:244 [7] NCCL INFO Using network IB
+gpu1:239:239 [2] NCCL INFO DMA-BUF is available on GPU device 2
+gpu1:239:239 [2] NCCL INFO ncclCommInitRank comm 0x20454e60 rank 2 nranks 16 cudaDev 2 nvmlDev 2 busId 1b000 commId 0x92ed644083b4525 - Init START
+gpu1:242:242 [5] NCCL INFO DMA-BUF is available on GPU device 5
+gpu1:242:242 [5] NCCL INFO ncclCommInitRank comm 0x26b11980 rank 5 nranks 16 cudaDev 5 nvmlDev 5 busId ab000 commId 0x92ed644083b4525 - Init START
+gpu1:237:237 [0] NCCL INFO NET/IB : Using [0]mlx5_0:1/RoCE [1]mlx5_3:1/RoCE [RO]; OOB eth0:10.101.0.11<0>
+gpu1:237:237 [0] NCCL INFO Initialized NET plugin IB
+gpu1:237:237 [0] NCCL INFO Assigned NET plugin IB to comm
+gpu1:237:237 [0] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu1:237:237 [0] NCCL INFO Using network IB
+gpu1:244:244 [7] NCCL INFO DMA-BUF is available on GPU device 7
+gpu1:244:244 [7] NCCL INFO ncclCommInitRank comm 0x470123c0 rank 7 nranks 16 cudaDev 7 nvmlDev 7 busId af000 commId 0x92ed644083b4525 - Init START
+gpu1:243:243 [6] NCCL INFO DMA-BUF is available on GPU device 6
+gpu1:243:243 [6] NCCL INFO ncclCommInitRank comm 0x3461a390 rank 6 nranks 16 cudaDev 6 nvmlDev 6 busId ae000 commId 0x92ed644083b4525 - Init START
+gpu1:244:244 [7] NCCL INFO RAS client listening socket at 127.0.0.1<28028>
+gpu1:243:243 [6] NCCL INFO RAS client listening socket at 127.0.0.1<28028>
+gpu1:237:237 [0] NCCL INFO DMA-BUF is available on GPU device 0
+gpu1:237:237 [0] NCCL INFO ncclCommInitRank comm 0x52b5b9e0 rank 0 nranks 16 cudaDev 0 nvmlDev 0 busId 19000 commId 0x92ed644083b4525 - Init START
+gpu1:237:237 [0] NCCL INFO RAS client listening socket at 127.0.0.1<28028>
+gpu1:238:238 [1] NCCL INFO RAS client listening socket at 127.0.0.1<28028>
+gpu1:241:241 [4] NCCL INFO NET/Plugin: Could not find: libnccl-net.so
+gpu1:241:241 [4] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu1:241:241 [4] NCCL INFO NCCL_IB_HCA set to =mlx5_0:1,mlx5_3:1
+gpu1:241:241 [4] NCCL INFO NET/IB : Using [0]mlx5_0:1/RoCE [1]mlx5_3:1/RoCE [RO]; OOB eth0:10.101.0.11<0>
+gpu1:241:241 [4] NCCL INFO Initialized NET plugin IB
+gpu1:241:241 [4] NCCL INFO Assigned NET plugin IB to comm
+gpu1:241:241 [4] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu1:241:241 [4] NCCL INFO Using network IB
+gpu1:240:240 [3] NCCL INFO NET/Plugin: Could not find: libnccl-net.so
+gpu1:240:240 [3] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu1:240:240 [3] NCCL INFO NCCL_IB_HCA set to =mlx5_0:1,mlx5_3:1
+gpu1:241:241 [4] NCCL INFO DMA-BUF is available on GPU device 4
+gpu1:241:241 [4] NCCL INFO ncclCommInitRank comm 0x38d5fbf0 rank 4 nranks 16 cudaDev 4 nvmlDev 4 busId aa000 commId 0x92ed644083b4525 - Init START
+gpu1:242:242 [5] NCCL INFO RAS client listening socket at 127.0.0.1<28028>
+gpu1:240:240 [3] NCCL INFO NET/IB : Using [0]mlx5_0:1/RoCE [1]mlx5_3:1/RoCE [RO]; OOB eth0:10.101.0.11<0>
+gpu1:240:240 [3] NCCL INFO Initialized NET plugin IB
+gpu1:240:240 [3] NCCL INFO Assigned NET plugin IB to comm
+gpu1:240:240 [3] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu1:240:240 [3] NCCL INFO Using network IB
+gpu1:240:240 [3] NCCL INFO DMA-BUF is available on GPU device 3
+gpu1:240:240 [3] NCCL INFO ncclCommInitRank comm 0x4dcdd4f0 rank 3 nranks 16 cudaDev 3 nvmlDev 3 busId 1d000 commId 0x92ed644083b4525 - Init START
+gpu1:239:239 [2] NCCL INFO RAS client listening socket at 127.0.0.1<28028>
+gpu1:241:241 [4] NCCL INFO RAS client listening socket at 127.0.0.1<28028>
+gpu1:240:240 [3] NCCL INFO RAS client listening socket at 127.0.0.1<28028>
+gpu1:242:242 [5] NCCL INFO Bootstrap timings total 0.103585 (create 0.000061, send 0.000249, recv 0.003079, ring 0.011514, delay 0.000003)
+gpu1:240:240 [3] NCCL INFO Bootstrap timings total 0.001619 (create 0.000064, send 0.000269, recv 0.000206, ring 0.000503, delay 0.000002)
+gpu1:239:239 [2] NCCL INFO Bootstrap timings total 0.112140 (create 0.000028, send 0.000110, recv 0.110768, ring 0.000959, delay 0.000001)
+gpu1:242:242 [5] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0.
+gpu1:239:239 [2] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0.
+gpu1:240:240 [3] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0.
+gpu1:243:243 [6] NCCL INFO Bootstrap timings total 0.100568 (create 0.000036, send 0.000132, recv 0.000249, ring 0.099803, delay 0.000001)
+gpu1:238:238 [1] NCCL INFO Bootstrap timings total 0.140545 (create 0.000033, send 0.000130, recv 0.028539, ring 0.090217, delay 0.000001)
+gpu1:241:241 [4] NCCL INFO Bootstrap timings total 0.012608 (create 0.000059, send 0.000268, recv 0.000095, ring 0.000495, delay 0.000003)
+gpu1:237:237 [0] NCCL INFO Bootstrap timings total 0.095306 (create 0.000036, send 0.000134, recv 0.000271, ring 0.094008, delay 0.000002)
+gpu1:243:243 [6] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0.
+gpu1:238:238 [1] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0.
+gpu1:244:244 [7] NCCL INFO Bootstrap timings total 0.101825 (create 0.000037, send 0.000121, recv 0.000139, ring 0.099829, delay 0.000001)
+gpu1:237:237 [0] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0.
+gpu1:244:244 [7] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0.
+gpu1:241:241 [4] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0.
+gpu1:237:237 [0] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 0 is 0-31,64-95. (GPU affinity = 0-31,64-95 ; CPU affinity = 0-127).
+gpu1:237:237 [0] NCCL INFO NCCL_NVLS_ENABLE set by environment to 0.
+gpu1:237:237 [0] NCCL INFO NCCL_CROSS_NIC set by environment to 1.
+gpu1:244:244 [7] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 7 is 32-63,96-127. (GPU affinity = 32-63,96-127 ; CPU affinity = 0-127).
+gpu1:244:244 [7] NCCL INFO NCCL_NVLS_ENABLE set by environment to 0.
+gpu1:244:244 [7] NCCL INFO NCCL_CROSS_NIC set by environment to 1.
+gpu1:238:238 [1] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 1 is 0-31,64-95. (GPU affinity = 0-31,64-95 ; CPU affinity = 0-127).
+gpu1:238:238 [1] NCCL INFO NCCL_NVLS_ENABLE set by environment to 0.
+gpu1:238:238 [1] NCCL INFO NCCL_CROSS_NIC set by environment to 1.
+gpu1:239:239 [2] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 2 is 0-31,64-95. (GPU affinity = 0-31,64-95 ; CPU affinity = 0-127).
+gpu1:239:239 [2] NCCL INFO NCCL_NVLS_ENABLE set by environment to 0.
+gpu1:239:239 [2] NCCL INFO NCCL_CROSS_NIC set by environment to 1.
+gpu1:243:243 [6] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 6 is 32-63,96-127. (GPU affinity = 32-63,96-127 ; CPU affinity = 0-127).
+gpu1:243:243 [6] NCCL INFO NCCL_NVLS_ENABLE set by environment to 0.
+gpu1:243:243 [6] NCCL INFO NCCL_CROSS_NIC set by environment to 1.
+gpu1:242:242 [5] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 5 is 32-63,96-127. (GPU affinity = 32-63,96-127 ; CPU affinity = 0-127).
+gpu1:242:242 [5] NCCL INFO NCCL_NVLS_ENABLE set by environment to 0.
+gpu1:242:242 [5] NCCL INFO NCCL_CROSS_NIC set by environment to 1.
+gpu1:240:240 [3] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 3 is 0-31,64-95. (GPU affinity = 0-31,64-95 ; CPU affinity = 0-127).
+gpu1:241:241 [4] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 4 is 32-63,96-127. (GPU affinity = 32-63,96-127 ; CPU affinity = 0-127).
+gpu1:241:241 [4] NCCL INFO NCCL_NVLS_ENABLE set by environment to 0.
+gpu1:241:241 [4] NCCL INFO NCCL_CROSS_NIC set by environment to 1.
+gpu1:240:240 [3] NCCL INFO NCCL_NVLS_ENABLE set by environment to 0.
+gpu1:240:240 [3] NCCL INFO NCCL_CROSS_NIC set by environment to 1.
+gpu1:242:242 [5] NCCL INFO comm 0x26b11980 rank 5 nRanks 16 nNodes 4 localRanks 4 localRank 1 MNNVL 0
+gpu1:244:244 [7] NCCL INFO comm 0x470123c0 rank 7 nRanks 16 nNodes 4 localRanks 4 localRank 3 MNNVL 0
+gpu1:243:243 [6] NCCL INFO comm 0x3461a390 rank 6 nRanks 16 nNodes 4 localRanks 4 localRank 2 MNNVL 0
+gpu1:241:241 [4] NCCL INFO comm 0x38d5fbf0 rank 4 nRanks 16 nNodes 4 localRanks 4 localRank 0 MNNVL 0
+gpu1:244:244 [7] NCCL INFO Trees [0] -1/-1/-1->7->6 [1] 5/-1/-1->7->11 [2] -1/-1/-1->7->6 [3] 5/11/3->7->15
+gpu1:243:243 [6] NCCL INFO Trees [0] 7/-1/-1->6->5 [1] 4/-1/-1->6->5 [2] 7/-1/-1->6->5 [3] 4/-1/-1->6->5
+gpu1:244:244 [7] NCCL INFO P2P Chunksize set to 131072
+gpu1:240:240 [3] NCCL INFO comm 0x4dcdd4f0 rank 3 nRanks 16 nNodes 4 localRanks 4 localRank 3 MNNVL 0
+gpu1:243:243 [6] NCCL INFO P2P Chunksize set to 131072
+gpu1:242:242 [5] NCCL INFO Trees [0] 6/-1/-1->5->4 [1] 6/-1/-1->5->7 [2] 6/-1/-1->5->4 [3] 6/-1/-1->5->7
+gpu1:239:239 [2] NCCL INFO comm 0x20454e60 rank 2 nRanks 16 nNodes 4 localRanks 4 localRank 2 MNNVL 0
+gpu1:242:242 [5] NCCL INFO P2P Chunksize set to 131072
+gpu1:241:241 [4] NCCL INFO Trees [0] 5/-1/-1->4->8 [1] -1/-1/-1->4->6 [2] 5/8/0->4->12 [3] -1/-1/-1->4->6
+gpu1:241:241 [4] NCCL INFO P2P Chunksize set to 131072
+gpu1:237:237 [0] NCCL INFO comm 0x52b5b9e0 rank 0 nRanks 16 nNodes 4 localRanks 4 localRank 0 MNNVL 0
+gpu1:238:238 [1] NCCL INFO comm 0x4dcaf9f0 rank 1 nRanks 16 nNodes 4 localRanks 4 localRank 1 MNNVL 0
+gpu1:239:239 [2] NCCL INFO Trees [0] 3/-1/-1->2->1 [1] 0/-1/-1->2->1 [2] 3/-1/-1->2->1 [3] 0/-1/-1->2->1
+gpu1:239:239 [2] NCCL INFO P2P Chunksize set to 131072
+gpu1:240:240 [3] NCCL INFO Trees [0] -1/-1/-1->3->2 [1] 1/11/-1->3->-1 [2] -1/-1/-1->3->2 [3] 1/-1/-1->3->7
+gpu1:237:237 [0] NCCL INFO Channel 00/04 : 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+gpu1:240:240 [3] NCCL INFO P2P Chunksize set to 131072
+gpu1:237:237 [0] NCCL INFO Channel 01/04 : 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+gpu1:238:238 [1] NCCL INFO Trees [0] 2/-1/-1->1->0 [1] 2/-1/-1->1->3 [2] 2/-1/-1->1->0 [3] 2/-1/-1->1->3
+gpu1:237:237 [0] NCCL INFO Channel 02/04 : 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+gpu1:238:238 [1] NCCL INFO P2P Chunksize set to 131072
+gpu1:237:237 [0] NCCL INFO Channel 03/04 : 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+gpu1:237:237 [0] NCCL INFO Trees [0] 1/8/-1->0->-1 [1] -1/-1/-1->0->2 [2] 1/-1/-1->0->4 [3] -1/-1/-1->0->2
+gpu1:237:237 [0] NCCL INFO P2P Chunksize set to 131072
+gpu1:244:244 [7] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so
+gpu1:243:243 [6] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so
+gpu1:242:242 [5] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so
+gpu1:244:244 [7] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu1:243:243 [6] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu1:242:242 [5] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu1:241:241 [4] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so
+gpu1:241:241 [4] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu1:244:1587 [0] NCCL INFO [Proxy Service] Device 7 CPU core 120
+gpu1:243:1586 [0] NCCL INFO [Proxy Service] Device 6 CPU core 119
+gpu1:242:1588 [0] NCCL INFO [Proxy Service] Device 5 CPU core 102
+gpu1:240:240 [3] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so
+gpu1:243:1590 [0] NCCL INFO [Proxy Service UDS] Device 6 CPU core 105
+gpu1:239:239 [2] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so
+gpu1:242:1592 [0] NCCL INFO [Proxy Service UDS] Device 5 CPU core 40
+gpu1:244:1591 [0] NCCL INFO [Proxy Service UDS] Device 7 CPU core 104
+gpu1:240:240 [3] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu1:239:239 [2] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu1:241:1589 [0] NCCL INFO [Proxy Service] Device 4 CPU core 107
+gpu1:241:1593 [0] NCCL INFO [Proxy Service UDS] Device 4 CPU core 44
+gpu1:237:237 [0] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so
+gpu1:238:238 [1] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so
+gpu1:237:237 [0] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu1:238:238 [1] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu1:239:1594 [0] NCCL INFO [Proxy Service] Device 2 CPU core 16
+gpu1:240:1597 [0] NCCL INFO [Proxy Service UDS] Device 3 CPU core 78
+gpu1:239:1596 [0] NCCL INFO [Proxy Service UDS] Device 2 CPU core 14
+gpu1:237:1598 [0] NCCL INFO [Proxy Service] Device 0 CPU core 2
+gpu1:240:1595 [0] NCCL INFO [Proxy Service] Device 3 CPU core 16
+gpu1:238:1600 [0] NCCL INFO [Proxy Service UDS] Device 1 CPU core 67
+gpu1:237:1601 [0] NCCL INFO [Proxy Service UDS] Device 0 CPU core 67
+gpu1:238:1599 [0] NCCL INFO [Proxy Service] Device 1 CPU core 2
+gpu1:242:242 [5] NCCL INFO Channel 00/0 : 5[5] -> 6[6] via P2P/IPC
+gpu1:242:242 [5] NCCL INFO Channel 01/0 : 5[5] -> 6[6] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 00/0 : 6[6] -> 7[7] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 01/0 : 6[6] -> 7[7] via P2P/IPC
+gpu1:242:242 [5] NCCL INFO Channel 02/0 : 5[5] -> 6[6] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 02/0 : 6[6] -> 7[7] via P2P/IPC
+gpu1:242:242 [5] NCCL INFO Channel 03/0 : 5[5] -> 6[6] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 03/0 : 6[6] -> 7[7] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 00/0 : 2[2] -> 3[3] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 01/0 : 2[2] -> 3[3] via P2P/IPC
+gpu1:238:238 [1] NCCL INFO Channel 00/0 : 1[1] -> 2[2] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 02/0 : 2[2] -> 3[3] via P2P/IPC
+gpu1:238:238 [1] NCCL INFO Channel 01/0 : 1[1] -> 2[2] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 03/0 : 2[2] -> 3[3] via P2P/IPC
+gpu1:238:238 [1] NCCL INFO Channel 02/0 : 1[1] -> 2[2] via P2P/IPC
+gpu1:238:238 [1] NCCL INFO Channel 03/0 : 1[1] -> 2[2] via P2P/IPC
+gpu1:244:244 [7] NCCL INFO Channel 00/0 : 7[7] -> 8[0] [send] via NET/IB/1/GDRDMA
+gpu1:244:1602 [0] NCCL INFO [Proxy Progress] Device 7 CPU core 46
+gpu1:244:244 [7] NCCL INFO Channel 01/0 : 7[7] -> 8[0] [send] via NET/IB/1/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 02/0 : 7[7] -> 8[0] [send] via NET/IB/1/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 03/0 : 7[7] -> 8[0] [send] via NET/IB/1/GDRDMA
+gpu1:241:1603 [0] NCCL INFO [Proxy Progress] Device 4 CPU core 112
+gpu1:241:241 [4] NCCL INFO Channel 00/0 : 3[3] -> 4[4] [receive] via NET/IB/1/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 01/0 : 3[3] -> 4[4] [receive] via NET/IB/1/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 02/0 : 3[3] -> 4[4] [receive] via NET/IB/1/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 03/0 : 3[3] -> 4[4] [receive] via NET/IB/1/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 00/0 : 4[4] -> 5[5] via P2P/IPC
+gpu1:241:241 [4] NCCL INFO Channel 01/0 : 4[4] -> 5[5] via P2P/IPC
+gpu1:241:241 [4] NCCL INFO Channel 02/0 : 4[4] -> 5[5] via P2P/IPC
+gpu1:237:1604 [0] NCCL INFO [Proxy Progress] Device 0 CPU core 69
+gpu1:237:237 [0] NCCL INFO Channel 00/0 : 15[7] -> 0[0] [receive] via NET/IB/0/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 03/0 : 4[4] -> 5[5] via P2P/IPC
+gpu1:237:237 [0] NCCL INFO Channel 01/0 : 15[7] -> 0[0] [receive] via NET/IB/0/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 02/0 : 15[7] -> 0[0] [receive] via NET/IB/0/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 03/0 : 15[7] -> 0[0] [receive] via NET/IB/0/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 00/0 : 0[0] -> 1[1] via P2P/IPC
+gpu1:237:237 [0] NCCL INFO Channel 01/0 : 0[0] -> 1[1] via P2P/IPC
+gpu1:240:1605 [0] NCCL INFO [Proxy Progress] Device 3 CPU core 81
+gpu1:240:240 [3] NCCL INFO Channel 00/0 : 3[3] -> 4[4] [send] via NET/IB/0/GDRDMA
+gpu1:240:240 [3] NCCL INFO Channel 01/0 : 3[3] -> 4[4] [send] via NET/IB/0/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 02/0 : 0[0] -> 1[1] via P2P/IPC
+gpu1:240:240 [3] NCCL INFO Channel 02/0 : 3[3] -> 4[4] [send] via NET/IB/0/GDRDMA
+gpu1:240:240 [3] NCCL INFO Channel 03/0 : 3[3] -> 4[4] [send] via NET/IB/0/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 03/0 : 0[0] -> 1[1] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu1:242:242 [5] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu1:244:244 [7] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu1:241:241 [4] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu1:242:242 [5] NCCL INFO Channel 01/0 : 5[5] -> 7[7] via P2P/IPC
+gpu1:240:240 [3] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu1:241:241 [4] NCCL INFO Channel 01/0 : 4[4] -> 6[6] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu1:238:238 [1] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu1:237:237 [0] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu1:238:238 [1] NCCL INFO Channel 01/0 : 1[1] -> 3[3] via P2P/IPC
+gpu1:237:237 [0] NCCL INFO Channel 01/0 : 0[0] -> 2[2] via P2P/IPC
+gpu1:242:242 [5] NCCL INFO Channel 03/0 : 5[5] -> 7[7] via P2P/IPC
+gpu1:238:238 [1] NCCL INFO Channel 03/0 : 1[1] -> 3[3] via P2P/IPC
+gpu1:241:241 [4] NCCL INFO Channel 03/0 : 4[4] -> 6[6] via P2P/IPC
+gpu1:237:237 [0] NCCL INFO Channel 03/0 : 0[0] -> 2[2] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 01/0 : 6[6] -> 4[4] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 01/0 : 2[2] -> 0[0] via P2P/IPC
+gpu1:244:244 [7] NCCL INFO Channel 03/0 : 3[3] -> 7[7] [receive] via NET/IB/1/GDRDMA
+gpu1:240:240 [3] NCCL INFO Channel 03/0 : 3[3] -> 7[7] [send] via NET/IB/0/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 02/0 : 0[0] -> 4[4] [receive] via NET/IB/1/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 02/0 : 0[0] -> 4[4] [send] via NET/IB/0/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 01/0 : 7[7] -> 11[3] [send] via NET/IB/1/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 03/0 : 7[7] -> 11[3] [send] via NET/IB/1/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 00/0 : 4[4] -> 8[0] [send] via NET/IB/1/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 02/0 : 4[4] -> 8[0] [send] via NET/IB/1/GDRDMA
+gpu1:240:240 [3] NCCL INFO Channel 01/0 : 11[3] -> 3[3] [receive] via NET/IB/0/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 03/0 : 15[7] -> 7[7] [receive] via NET/IB/1/GDRDMA
+gpu1:240:240 [3] NCCL INFO Channel 01/0 : 3[3] -> 11[3] [send] via NET/IB/0/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 03/0 : 7[7] -> 15[7] [send] via NET/IB/1/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 00/0 : 8[0] -> 0[0] [receive] via NET/IB/0/GDRDMA
+gpu1:239:239 [2] NCCL INFO Channel 03/0 : 2[2] -> 0[0] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 03/0 : 6[6] -> 4[4] via P2P/IPC
+gpu1:241:241 [4] NCCL INFO Channel 02/0 : 12[4] -> 4[4] [receive] via NET/IB/1/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 00/0 : 0[0] -> 8[0] [send] via NET/IB/0/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 02/0 : 4[4] -> 12[4] [send] via NET/IB/1/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 01/0 : 11[3] -> 7[7] [receive] via NET/IB/1/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 03/0 : 11[3] -> 7[7] [receive] via NET/IB/1/GDRDMA
+gpu1:240:240 [3] NCCL INFO Channel 03/0 : 7[7] -> 3[3] [receive] via NET/IB/0/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 03/0 : 7[7] -> 3[3] [send] via NET/IB/1/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 00/0 : 8[0] -> 4[4] [receive] via NET/IB/1/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 02/0 : 4[4] -> 0[0] [receive] via NET/IB/0/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 02/0 : 8[0] -> 4[4] [receive] via NET/IB/1/GDRDMA
+gpu1:240:240 [3] NCCL INFO Channel 01/0 : 3[3] -> 1[1] via P2P/IPC
+gpu1:244:244 [7] NCCL INFO Channel 01/0 : 7[7] -> 5[5] via P2P/IPC
+gpu1:241:241 [4] NCCL INFO Channel 02/0 : 4[4] -> 0[0] [send] via NET/IB/1/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 03/0 : 7[7] -> 5[5] via P2P/IPC
+gpu1:240:240 [3] NCCL INFO Channel 03/0 : 3[3] -> 1[1] via P2P/IPC
+gpu1:244:244 [7] NCCL INFO Channel 00/0 : 7[7] -> 6[6] via P2P/IPC
+gpu1:240:240 [3] NCCL INFO Channel 00/0 : 3[3] -> 2[2] via P2P/IPC
+gpu1:244:244 [7] NCCL INFO Channel 02/0 : 7[7] -> 6[6] via P2P/IPC
+gpu1:240:240 [3] NCCL INFO Channel 02/0 : 3[3] -> 2[2] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 00/0 : 6[6] -> 5[5] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 00/0 : 2[2] -> 1[1] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 01/0 : 2[2] -> 1[1] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 01/0 : 6[6] -> 5[5] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 02/0 : 2[2] -> 1[1] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 02/0 : 6[6] -> 5[5] via P2P/IPC
+gpu1:242:242 [5] NCCL INFO Channel 00/0 : 5[5] -> 4[4] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 03/0 : 6[6] -> 5[5] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 03/0 : 2[2] -> 1[1] via P2P/IPC
+gpu1:238:238 [1] NCCL INFO Channel 00/0 : 1[1] -> 0[0] via P2P/IPC
+gpu1:242:242 [5] NCCL INFO Channel 02/0 : 5[5] -> 4[4] via P2P/IPC
+gpu1:238:238 [1] NCCL INFO Channel 02/0 : 1[1] -> 0[0] via P2P/IPC
+gpu1:244:244 [7] NCCL INFO Connected all trees
+gpu1:237:237 [0] NCCL INFO Connected all trees
+gpu1:241:241 [4] NCCL INFO Connected all trees
+gpu1:240:240 [3] NCCL INFO Connected all trees
+gpu1:243:243 [6] NCCL INFO Connected all trees
+gpu1:242:242 [5] NCCL INFO Connected all trees
+gpu1:238:238 [1] NCCL INFO Connected all trees
+gpu1:239:239 [2] NCCL INFO Connected all trees
+gpu1:244:244 [7] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so
+gpu1:244:244 [7] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu1:244:244 [7] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu1:241:241 [4] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so
+gpu1:241:241 [4] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu1:241:241 [4] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu1:240:240 [3] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so
+gpu1:240:240 [3] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu1:240:240 [3] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu1:237:237 [0] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so
+gpu1:237:237 [0] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu1:237:237 [0] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu1:237:237 [0] NCCL INFO CC Off, workFifoBytes 1048576
+gpu1:243:1606 [0] NCCL INFO [Proxy Progress] Device 6 CPU core 122
+gpu1:242:1607 [0] NCCL INFO [Proxy Progress] Device 5 CPU core 103
+gpu1:242:242 [5] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so
+gpu1:242:242 [5] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu1:242:242 [5] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu1:243:243 [6] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so
+gpu1:243:243 [6] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu1:243:243 [6] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu1:242:242 [5] NCCL INFO ncclCommInitRank comm 0x26b11980 rank 5 nranks 16 cudaDev 5 nvmlDev 5 busId ab000 commId 0x92ed644083b4525 - Init COMPLETE
+gpu1:244:244 [7] NCCL INFO ncclCommInitRank comm 0x470123c0 rank 7 nranks 16 cudaDev 7 nvmlDev 7 busId af000 commId 0x92ed644083b4525 - Init COMPLETE
+gpu1:242:242 [5] NCCL INFO Init timings - ncclCommInitRank: rank 5 nranks 16 total 0.73 (kernels 0.31, alloc 0.05, bootstrap 0.10, allgathers 0.01, topo 0.04, graphs 0.00, connections 0.22, rest 0.00)
+gpu1:244:244 [7] NCCL INFO Init timings - ncclCommInitRank: rank 7 nranks 16 total 0.71 (kernels 0.32, alloc 0.02, bootstrap 0.10, allgathers 0.01, topo 0.03, graphs 0.00, connections 0.17, rest 0.05)
+gpu1:241:241 [4] NCCL INFO ncclCommInitRank comm 0x38d5fbf0 rank 4 nranks 16 cudaDev 4 nvmlDev 4 busId aa000 commId 0x92ed644083b4525 - Init COMPLETE
+gpu1:241:241 [4] NCCL INFO Init timings - ncclCommInitRank: rank 4 nranks 16 total 0.70 (kernels 0.37, alloc 0.06, bootstrap 0.01, allgathers 0.00, topo 0.04, graphs 0.00, connections 0.17, rest 0.05)
+gpu1:243:243 [6] NCCL INFO ncclCommInitRank comm 0x3461a390 rank 6 nranks 16 cudaDev 6 nvmlDev 6 busId ae000 commId 0x92ed644083b4525 - Init COMPLETE
+gpu1:243:243 [6] NCCL INFO Init timings - ncclCommInitRank: rank 6 nranks 16 total 0.70 (kernels 0.31, alloc 0.03, bootstrap 0.10, allgathers 0.01, topo 0.03, graphs 0.00, connections 0.22, rest 0.00)
+gpu1:238:1608 [0] NCCL INFO [Proxy Progress] Device 1 CPU core 70
+gpu1:239:1609 [0] NCCL INFO [Proxy Progress] Device 2 CPU core 19
+gpu1:238:238 [1] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so
+gpu1:238:238 [1] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu1:238:238 [1] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu1:239:239 [2] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so
+gpu1:239:239 [2] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu1:239:239 [2] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu1:238:238 [1] NCCL INFO ncclCommInitRank comm 0x4dcaf9f0 rank 1 nranks 16 cudaDev 1 nvmlDev 1 busId 1a000 commId 0x92ed644083b4525 - Init COMPLETE
+gpu1:238:238 [1] NCCL INFO Init timings - ncclCommInitRank: rank 1 nranks 16 total 0.74 (kernels 0.31, alloc 0.01, bootstrap 0.14, allgathers 0.00, topo 0.03, graphs 0.01, connections 0.22, rest 0.00)
+gpu1:240:240 [3] NCCL INFO ncclCommInitRank comm 0x4dcdd4f0 rank 3 nranks 16 cudaDev 3 nvmlDev 3 busId 1d000 commId 0x92ed644083b4525 - Init COMPLETE
+gpu1:240:240 [3] NCCL INFO Init timings - ncclCommInitRank: rank 3 nranks 16 total 0.71 (kernels 0.42, alloc 0.02, bootstrap 0.00, allgathers 0.00, topo 0.04, graphs 0.01, connections 0.17, rest 0.05)
+gpu1:237:237 [0] NCCL INFO ncclCommInitRank comm 0x52b5b9e0 rank 0 nranks 16 cudaDev 0 nvmlDev 0 busId 19000 commId 0x92ed644083b4525 - Init COMPLETE
+gpu1:239:239 [2] NCCL INFO ncclCommInitRank comm 0x20454e60 rank 2 nranks 16 cudaDev 2 nvmlDev 2 busId 1b000 commId 0x92ed644083b4525 - Init COMPLETE
+gpu1:237:237 [0] NCCL INFO Init timings - ncclCommInitRank: rank 0 nranks 16 total 0.71 (kernels 0.31, alloc 0.03, bootstrap 0.10, allgathers 0.01, topo 0.03, graphs 0.01, connections 0.17, rest 0.05)
+gpu1:239:239 [2] NCCL INFO Init timings - ncclCommInitRank: rank 2 nranks 16 total 0.71 (kernels 0.31, alloc 0.02, bootstrap 0.11, allgathers 0.01, topo 0.03, graphs 0.01, connections 0.22, rest 0.00)
+[2026-07-30 14:10:48 TP0 EP0] CustomAllreduce is disabled because this process group spans across nodes.
+[2026-07-30 14:10:48 TP1 EP0] CustomAllreduce is disabled because this process group spans across nodes.
+[2026-07-30 14:10:48 TP2 EP0] CustomAllreduce is disabled because this process group spans across nodes.
+[2026-07-30 14:10:48 TP3 EP0] CustomAllreduce is disabled because this process group spans across nodes.
+[2026-07-30 14:10:48 TP4 EP0] CustomAllreduce is disabled because this process group spans across nodes.
+[2026-07-30 14:10:48 TP5 EP0] CustomAllreduce is disabled because this process group spans across nodes.
+[2026-07-30 14:10:48 TP6 EP0] CustomAllreduce is disabled because this process group spans across nodes.
+[2026-07-30 14:10:48 TP7 EP0] CustomAllreduce is disabled because this process group spans across nodes.
+[2026-07-30 14:10:48 TP0 EP0] Init torch distributed ends. elapsed=8.67 s, mem usage=0.42 GB
+[2026-07-30 14:10:48 TP7 EP0] Init torch distributed ends. elapsed=8.25 s, mem usage=0.48 GB
+[2026-07-30 14:10:48 TP6 EP0] Init torch distributed ends. elapsed=8.35 s, mem usage=0.39 GB
+[2026-07-30 14:10:48 TP5 EP0] Init torch distributed ends. elapsed=8.33 s, mem usage=0.41 GB
+[2026-07-30 14:10:48 TP4 EP0] Init torch distributed ends. elapsed=8.08 s, mem usage=0.46 GB
+[2026-07-30 14:10:48 TP3 EP0] Init torch distributed ends. elapsed=8.55 s, mem usage=0.44 GB
+[2026-07-30 14:10:48 TP2 EP0] Init torch distributed ends. elapsed=8.32 s, mem usage=0.39 GB
+[2026-07-30 14:10:48 TP1 EP0] Init torch distributed ends. elapsed=9.77 s, mem usage=0.41 GB
+[2026-07-30 14:10:51 TP4 EP0] Load weight begin. avail mem=82.04 GB
+[2026-07-30 14:10:51 TP7 EP0] Load weight begin. avail mem=82.03 GB
+[2026-07-30 14:10:51 TP2 EP0] Load weight begin. avail mem=82.11 GB
+[2026-07-30 14:10:51 TP0 EP0] Load weight begin. avail mem=82.08 GB
+[2026-07-30 14:10:51 TP0 EP0] Detected fp8 checkpoint.
+[2026-07-30 14:10:51 TP0 EP0] Config does not support fused shared expert(s). Shared experts fusion optimization is disabled.
+[2026-07-30 14:10:51 TP1 EP0] Load weight begin. avail mem=82.10 GB
+[2026-07-30 14:10:51 TP3 EP0] Load weight begin. avail mem=82.07 GB
+[2026-07-30 14:10:51 TP5 EP0] Load weight begin. avail mem=82.10 GB
+[2026-07-30 14:10:51 TP6 EP0] Load weight begin. avail mem=82.11 GB
+[2026-07-30 14:10:51 TP4 EP0] FlashInfer TRTLLM MoE deferred finalize is disabled (moe_runner_backend=flashinfer_mxfp4, quant_method=Mxfp4FlashinferCutlassMoEMethod).
+[2026-07-30 14:10:51 TP7 EP0] FlashInfer TRTLLM MoE deferred finalize is disabled (moe_runner_backend=flashinfer_mxfp4, quant_method=Mxfp4FlashinferCutlassMoEMethod).
+[2026-07-30 14:10:51 TP2 EP0] FlashInfer TRTLLM MoE deferred finalize is disabled (moe_runner_backend=flashinfer_mxfp4, quant_method=Mxfp4FlashinferCutlassMoEMethod).
+[2026-07-30 14:10:51 TP0 EP0] FlashInfer TRTLLM MoE deferred finalize is disabled (moe_runner_backend=flashinfer_mxfp4, quant_method=Mxfp4FlashinferCutlassMoEMethod).
+[2026-07-30 14:10:51 TP1 EP0] FlashInfer TRTLLM MoE deferred finalize is disabled (moe_runner_backend=flashinfer_mxfp4, quant_method=Mxfp4FlashinferCutlassMoEMethod).
+[2026-07-30 14:10:51 TP3 EP0] FlashInfer TRTLLM MoE deferred finalize is disabled (moe_runner_backend=flashinfer_mxfp4, quant_method=Mxfp4FlashinferCutlassMoEMethod).
+[2026-07-30 14:10:51 TP6 EP0] FlashInfer TRTLLM MoE deferred finalize is disabled (moe_runner_backend=flashinfer_mxfp4, quant_method=Mxfp4FlashinferCutlassMoEMethod).
+[2026-07-30 14:10:51 TP5 EP0] FlashInfer TRTLLM MoE deferred finalize is disabled (moe_runner_backend=flashinfer_mxfp4, quant_method=Mxfp4FlashinferCutlassMoEMethod).
+[2026-07-30 14:10:57 TP0 EP0] multimem all-gather disabled because the TP group spans across nodes.
+[2026-07-30 14:10:57 TP3 EP0] multimem all-gather disabled because the TP group spans across nodes.
+[2026-07-30 14:10:57 TP1 EP0] multimem all-gather disabled because the TP group spans across nodes.
+[2026-07-30 14:10:57 TP2 EP0] multimem all-gather disabled because the TP group spans across nodes.
+[2026-07-30 14:10:57 TP4 EP0] multimem all-gather disabled because the TP group spans across nodes.
+[2026-07-30 14:10:57 TP7 EP0] multimem all-gather disabled because the TP group spans across nodes.
+[2026-07-30 14:10:57 TP5 EP0] multimem all-gather disabled because the TP group spans across nodes.
+[2026-07-30 14:10:57 TP6 EP0] multimem all-gather disabled because the TP group spans across nodes.
+
Multi-thread loading shards: 0% Completed | 0/64 [00:00, ?it/s]
Multi-thread loading shards: 3% Completed | 2/64 [00:02<01:12, 1.17s/it]
Multi-thread loading shards: 8% Completed | 5/64 [00:02<00:24, 2.45it/s]
Multi-thread loading shards: 16% Completed | 10/64 [00:05<00:25, 2.11it/s]
Multi-thread loading shards: 22% Completed | 14/64 [00:05<00:15, 3.33it/s]
Multi-thread loading shards: 27% Completed | 17/64 [00:05<00:10, 4.57it/s]
Multi-thread loading shards: 30% Completed | 19/64 [00:07<00:18, 2.46it/s]
Multi-thread loading shards: 33% Completed | 21/64 [00:07<00:13, 3.12it/s]
Multi-thread loading shards: 36% Completed | 23/64 [00:07<00:10, 3.96it/s]
Multi-thread loading shards: 41% Completed | 26/64 [00:09<00:15, 2.41it/s]
Multi-thread loading shards: 44% Completed | 28/64 [00:09<00:11, 3.10it/s]
Multi-thread loading shards: 48% Completed | 31/64 [00:10<00:07, 4.38it/s]
Multi-thread loading shards: 52% Completed | 33/64 [00:10<00:05, 5.37it/s]
Multi-thread loading shards: 55% Completed | 35/64 [00:12<00:13, 2.23it/s]
Multi-thread loading shards: 62% Completed | 40/64 [00:12<00:05, 4.13it/s]
Multi-thread loading shards: 67% Completed | 43/64 [00:14<00:08, 2.58it/s]
Multi-thread loading shards: 72% Completed | 46/64 [00:15<00:05, 3.49it/s]
Multi-thread loading shards: 77% Completed | 49/64 [00:15<00:03, 4.70it/s]
Multi-thread loading shards: 80% Completed | 51/64 [00:17<00:05, 2.48it/s]
Multi-thread loading shards: 84% Completed | 54/64 [00:17<00:02, 3.44it/s]
Multi-thread loading shards: 89% Completed | 57/64 [00:17<00:01, 4.72it/s][2026-07-30 14:11:16 TP6 EP0] Execute dequant fp8 wo_a
+
Multi-thread loading shards: 92% Completed | 59/64 [00:19<00:01, 2.96it/s]
Multi-thread loading shards: 100% Completed | 64/64 [00:19<00:00, 3.34it/s]
+[2026-07-30 14:11:16 TP0 EP0] Execute dequant fp8 wo_a
+[2026-07-30 14:11:16 TP7 EP0] Execute dequant fp8 wo_a
+[2026-07-30 14:11:17 TP3 EP0] Execute dequant fp8 wo_a
+[2026-07-30 14:11:17 TP1 EP0] Execute dequant fp8 wo_a
+[2026-07-30 14:11:17 TP4 EP0] Execute dequant fp8 wo_a
+[2026-07-30 14:11:17 TP5 EP0] Execute dequant fp8 wo_a
+[2026-07-30 14:11:17 TP2 EP0] Execute dequant fp8 wo_a
+[2026-07-30 14:13:00 TP7 EP0] Using FP8 KV cache but no scaling factors provided. Defaulting to scaling factors of 1.0. This may lead to less accurate results!
+[2026-07-30 14:13:00 TP7 EP0] Load weight end. elapsed=129.48 s, type=DeepseekV4ForCausalLM, quant=fp8, fmt=e4m3, avail mem=25.16 GB, mem usage=56.87 GB.
+[2026-07-30 14:13:01 TP1 EP0] Using FP8 KV cache but no scaling factors provided. Defaulting to scaling factors of 1.0. This may lead to less accurate results!
+[2026-07-30 14:13:01 TP1 EP0] Load weight end. elapsed=129.58 s, type=DeepseekV4ForCausalLM, quant=fp8, fmt=e4m3, avail mem=25.23 GB, mem usage=56.87 GB.
+[2026-07-30 14:13:01 TP6 EP0] Using FP8 KV cache but no scaling factors provided. Defaulting to scaling factors of 1.0. This may lead to less accurate results!
+[2026-07-30 14:13:01 TP6 EP0] Load weight end. elapsed=129.58 s, type=DeepseekV4ForCausalLM, quant=fp8, fmt=e4m3, avail mem=25.24 GB, mem usage=56.87 GB.
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.0.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.1.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.2.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.3.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.4.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.5.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.6.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.7.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.8.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.9.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.10.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.11.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.12.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.13.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.14.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.15.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.16.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.17.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.18.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.19.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.20.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.21.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.22.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.23.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.24.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.25.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.26.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.27.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.28.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.29.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.30.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.31.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.32.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.33.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.34.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.35.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.36.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.37.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.38.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.39.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.40.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.41.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.42.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.43.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.44.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.45.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.46.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.47.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.48.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.49.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.50.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.51.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.52.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.53.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.54.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.55.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.56.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.57.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.58.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.59.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Preparing DSv4 MXFP4 experts for FlashInfer SM120 CUTLASS (layer: model.layers.60.mlp.experts)...
+[2026-07-30 14:13:02 TP0 EP0] Using FP8 KV cache but no scaling factors provided. Defaulting to scaling factors of 1.0. This may lead to less accurate results!
+[2026-07-30 14:13:02 TP0 EP0] Load weight end. elapsed=130.87 s, type=DeepseekV4ForCausalLM, quant=fp8, fmt=e4m3, avail mem=25.21 GB, mem usage=56.87 GB.
+[2026-07-30 14:13:02 TP2 EP0] Using FP8 KV cache but no scaling factors provided. Defaulting to scaling factors of 1.0. This may lead to less accurate results!
+[2026-07-30 14:13:02 TP2 EP0] Load weight end. elapsed=131.14 s, type=DeepseekV4ForCausalLM, quant=fp8, fmt=e4m3, avail mem=25.24 GB, mem usage=56.87 GB.
+[2026-07-30 14:13:02 TP3 EP0] Using FP8 KV cache but no scaling factors provided. Defaulting to scaling factors of 1.0. This may lead to less accurate results!
+[2026-07-30 14:13:02 TP3 EP0] Load weight end. elapsed=131.22 s, type=DeepseekV4ForCausalLM, quant=fp8, fmt=e4m3, avail mem=25.19 GB, mem usage=56.87 GB.
+[2026-07-30 14:13:04 TP4 EP0] Using FP8 KV cache but no scaling factors provided. Defaulting to scaling factors of 1.0. This may lead to less accurate results!
+[2026-07-30 14:13:04 TP4 EP0] Load weight end. elapsed=133.32 s, type=DeepseekV4ForCausalLM, quant=fp8, fmt=e4m3, avail mem=25.17 GB, mem usage=56.87 GB.
+[2026-07-30 14:13:06 TP5 EP0] Using FP8 KV cache but no scaling factors provided. Defaulting to scaling factors of 1.0. This may lead to less accurate results!
+[2026-07-30 14:13:06 TP5 EP0] Load weight end. elapsed=135.37 s, type=DeepseekV4ForCausalLM, quant=fp8, fmt=e4m3, avail mem=25.23 GB, mem usage=56.87 GB.
+[2026-07-30 14:13:08 TP6 EP0] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:13:08 TP2 EP0] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:13:09 TP3 EP0] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:13:09 TP7 EP0] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:13:09 TP1 EP0] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:13:09 TP0 EP0] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:13:09 TP4 EP0] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:13:09 TP5 EP0] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:13:09 TP0 EP0] DSV4 memory calculation: bytes_per_full_token=10993.84, available_bytes=16.93 GB, c128_state_fixed=3.92 GB, full_token=1270016
+[2026-07-30 14:13:09 TP0 EP0] DSV4 pool sizes: full=1270016, swa=126976, c4=317504, c128=9922, c4_state=7936, c128_state=0
+[2026-07-30 14:13:09 TP5 EP0] DSV4 memory calculation: bytes_per_full_token=10993.84, available_bytes=16.93 GB, c128_state_fixed=3.92 GB, full_token=1270016
+[2026-07-30 14:13:09 TP5 EP0] DSV4 pool sizes: full=1270016, swa=126976, c4=317504, c128=9922, c4_state=7936, c128_state=0
+[2026-07-30 14:13:09 TP7 EP0] DSV4 memory calculation: bytes_per_full_token=10993.84, available_bytes=16.93 GB, c128_state_fixed=3.92 GB, full_token=1270016
+[2026-07-30 14:13:09 TP6 EP0] DSV4 memory calculation: bytes_per_full_token=10993.84, available_bytes=16.93 GB, c128_state_fixed=3.92 GB, full_token=1270016
+[2026-07-30 14:13:09 TP7 EP0] DSV4 pool sizes: full=1270016, swa=126976, c4=317504, c128=9922, c4_state=7936, c128_state=0
+[2026-07-30 14:13:09 TP6 EP0] DSV4 pool sizes: full=1270016, swa=126976, c4=317504, c128=9922, c4_state=7936, c128_state=0
+[2026-07-30 14:13:09 TP4 EP0] DSV4 memory calculation: bytes_per_full_token=10993.84, available_bytes=16.93 GB, c128_state_fixed=3.92 GB, full_token=1270016
+[2026-07-30 14:13:09 TP2 EP0] DSV4 memory calculation: bytes_per_full_token=10993.84, available_bytes=16.93 GB, c128_state_fixed=3.92 GB, full_token=1270016
+[2026-07-30 14:13:09 TP1 EP0] DSV4 memory calculation: bytes_per_full_token=10993.84, available_bytes=16.93 GB, c128_state_fixed=3.92 GB, full_token=1270016
+[2026-07-30 14:13:09 TP3 EP0] DSV4 memory calculation: bytes_per_full_token=10993.84, available_bytes=16.93 GB, c128_state_fixed=3.92 GB, full_token=1270016
+[2026-07-30 14:13:09 TP4 EP0] DSV4 pool sizes: full=1270016, swa=126976, c4=317504, c128=9922, c4_state=7936, c128_state=0
+[2026-07-30 14:13:09 TP2 EP0] DSV4 pool sizes: full=1270016, swa=126976, c4=317504, c128=9922, c4_state=7936, c128_state=0
+[2026-07-30 14:13:09 TP1 EP0] DSV4 pool sizes: full=1270016, swa=126976, c4=317504, c128=9922, c4_state=7936, c128_state=0
+[2026-07-30 14:13:09 TP3 EP0] DSV4 pool sizes: full=1270016, swa=126976, c4=317504, c128=9922, c4_state=7936, c128_state=0
+[2026-07-30 14:13:09 TP2 EP0] Initialize DeepSeekV4TokenToKVPool with max_num_reqs=256 swa_size=126976 c4_size=317504 c4_logical_size=317504 c128_size=9922 c4_state_pool_size=7936 c128_state_pool_size=32896
+[2026-07-30 14:13:09 TP4 EP0] Initialize DeepSeekV4TokenToKVPool with max_num_reqs=256 swa_size=126976 c4_size=317504 c4_logical_size=317504 c128_size=9922 c4_state_pool_size=7936 c128_state_pool_size=32896
+[2026-07-30 14:13:09 TP5 EP0] Initialize DeepSeekV4TokenToKVPool with max_num_reqs=256 swa_size=126976 c4_size=317504 c4_logical_size=317504 c128_size=9922 c4_state_pool_size=7936 c128_state_pool_size=32896
+[2026-07-30 14:13:09 TP7 EP0] Initialize DeepSeekV4TokenToKVPool with max_num_reqs=256 swa_size=126976 c4_size=317504 c4_logical_size=317504 c128_size=9922 c4_state_pool_size=7936 c128_state_pool_size=32896
+[2026-07-30 14:13:09 TP3 EP0] Initialize DeepSeekV4TokenToKVPool with max_num_reqs=256 swa_size=126976 c4_size=317504 c4_logical_size=317504 c128_size=9922 c4_state_pool_size=7936 c128_state_pool_size=32896
+[2026-07-30 14:13:09 TP6 EP0] Initialize DeepSeekV4TokenToKVPool with max_num_reqs=256 swa_size=126976 c4_size=317504 c4_logical_size=317504 c128_size=9922 c4_state_pool_size=7936 c128_state_pool_size=32896
+[2026-07-30 14:13:09 TP0 EP0] Initialize DeepSeekV4TokenToKVPool with max_num_reqs=256 swa_size=126976 c4_size=317504 c4_logical_size=317504 c128_size=9922 c4_state_pool_size=7936 c128_state_pool_size=32896
+[2026-07-30 14:13:09 TP1 EP0] Initialize DeepSeekV4TokenToKVPool with max_num_reqs=256 swa_size=126976 c4_size=317504 c4_logical_size=317504 c128_size=9922 c4_state_pool_size=7936 c128_state_pool_size=32896
+[2026-07-30 14:13:11 TP2 EP0] Memory pool end. avail mem=7.18 GB
+[2026-07-30 14:13:11 TP6 EP0] Memory pool end. avail mem=7.18 GB
+[2026-07-30 14:13:11 TP4 EP0] Memory pool end. avail mem=7.10 GB
+[2026-07-30 14:13:11 TP7 EP0] Memory pool end. avail mem=7.09 GB
+[2026-07-30 14:13:11 TP1 EP0] Memory pool end. avail mem=7.16 GB
+[2026-07-30 14:13:11 TP0 EP0] Memory pool end. avail mem=7.14 GB
+[2026-07-30 14:13:11 TP5 EP0] Memory pool end. avail mem=7.16 GB
+[2026-07-30 14:13:11 TP3 EP0] Memory pool end. avail mem=7.13 GB
+[2026-07-30 14:13:11 TP2 EP0] Using DeepseekV4AttnBackend for dsv4 attention backend (CUDA).
+[2026-07-30 14:13:11 TP6 EP0] Using DeepseekV4AttnBackend for dsv4 attention backend (CUDA).
+[2026-07-30 14:13:11 TP7 EP0] Using DeepseekV4AttnBackend for dsv4 attention backend (CUDA).
+[2026-07-30 14:13:11 TP4 EP0] Using DeepseekV4AttnBackend for dsv4 attention backend (CUDA).
+[2026-07-30 14:13:11 TP1 EP0] Using DeepseekV4AttnBackend for dsv4 attention backend (CUDA).
+[2026-07-30 14:13:11 TP3 EP0] Using DeepseekV4AttnBackend for dsv4 attention backend (CUDA).
+[2026-07-30 14:13:11 TP6 EP0] Running FlashInfer autotune with cache: /root/.cache/sglang/flashinfer/autotune/0.6.14/sm120/4f24e70399961c55/rank_tp6_pp0_dp0.json
+[2026-07-30 14:13:11 TP2 EP0] Running FlashInfer autotune with cache: /root/.cache/sglang/flashinfer/autotune/0.6.14/sm120/4f24e70399961c55/rank_tp2_pp0_dp0.json
+[2026-07-30 14:13:11 TP0 EP0] Using DeepseekV4AttnBackend for dsv4 attention backend (CUDA).
+[2026-07-30 14:13:11 TP5 EP0] Using DeepseekV4AttnBackend for dsv4 attention backend (CUDA).
+[2026-07-30 14:13:11 TP4 EP0] Running FlashInfer autotune with cache: /root/.cache/sglang/flashinfer/autotune/0.6.14/sm120/4f24e70399961c55/rank_tp4_pp0_dp0.json
+[2026-07-30 14:13:11 TP7 EP0] Running FlashInfer autotune with cache: /root/.cache/sglang/flashinfer/autotune/0.6.14/sm120/4f24e70399961c55/rank_tp7_pp0_dp0.json
+[2026-07-30 14:13:11 TP1 EP0] Running FlashInfer autotune with cache: /root/.cache/sglang/flashinfer/autotune/0.6.14/sm120/4f24e70399961c55/rank_tp1_pp0_dp0.json
+[2026-07-30 14:13:11 TP0 EP0] Running FlashInfer autotune with cache: /root/.cache/sglang/flashinfer/autotune/0.6.14/sm120/4f24e70399961c55/rank_tp0_pp0_dp0.json
+[2026-07-30 14:13:11 TP3 EP0] Running FlashInfer autotune with cache: /root/.cache/sglang/flashinfer/autotune/0.6.14/sm120/4f24e70399961c55/rank_tp3_pp0_dp0.json
+[2026-07-30 14:13:11 TP5 EP0] Running FlashInfer autotune with cache: /root/.cache/sglang/flashinfer/autotune/0.6.14/sm120/4f24e70399961c55/rank_tp5_pp0_dp0.json
+gpu1:237:237 [0] NCCL INFO Comm config Blocking set to 1
+gpu1:243:243 [6] NCCL INFO Comm config Blocking set to 1
+gpu1:241:241 [4] NCCL INFO Comm config Blocking set to 1
+gpu1:237:237 [0] NCCL INFO Initialized NET plugin IB
+gpu1:237:237 [0] NCCL INFO Assigned NET plugin IB to comm
+gpu1:237:237 [0] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu1:237:237 [0] NCCL INFO Using network IB
+gpu1:237:237 [0] NCCL INFO DMA-BUF is available on GPU device 0
+gpu1:237:237 [0] NCCL INFO ncclCommInitRankConfig comm 0x899bd480 rank 0 nranks 16 cudaDev 0 nvmlDev 0 busId 19000 commId 0x2f8347829d2566aa - Init START
+gpu1:244:244 [7] NCCL INFO Comm config Blocking set to 1
+gpu1:238:238 [1] NCCL INFO Comm config Blocking set to 1
+gpu1:241:241 [4] NCCL INFO Initialized NET plugin IB
+gpu1:241:241 [4] NCCL INFO Assigned NET plugin IB to comm
+gpu1:241:241 [4] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu1:241:241 [4] NCCL INFO Using network IB
+gpu1:241:241 [4] NCCL INFO DMA-BUF is available on GPU device 4
+gpu1:242:242 [5] NCCL INFO Comm config Blocking set to 1
+gpu1:241:241 [4] NCCL INFO ncclCommInitRankConfig comm 0x5d6feee0 rank 4 nranks 16 cudaDev 4 nvmlDev 4 busId aa000 commId 0x2f8347829d2566aa - Init START
+gpu1:238:238 [1] NCCL INFO Initialized NET plugin IB
+gpu1:238:238 [1] NCCL INFO Assigned NET plugin IB to comm
+gpu1:238:238 [1] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu1:238:238 [1] NCCL INFO Using network IB
+gpu1:238:238 [1] NCCL INFO DMA-BUF is available on GPU device 1
+gpu1:244:244 [7] NCCL INFO Initialized NET plugin IB
+gpu1:244:244 [7] NCCL INFO Assigned NET plugin IB to comm
+gpu1:242:242 [5] NCCL INFO Initialized NET plugin IB
+gpu1:244:244 [7] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu1:244:244 [7] NCCL INFO Using network IB
+gpu1:242:242 [5] NCCL INFO Assigned NET plugin IB to comm
+gpu1:242:242 [5] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu1:242:242 [5] NCCL INFO Using network IB
+gpu1:244:244 [7] NCCL INFO DMA-BUF is available on GPU device 7
+gpu1:242:242 [5] NCCL INFO DMA-BUF is available on GPU device 5
+gpu1:243:243 [6] NCCL INFO Initialized NET plugin IB
+gpu1:243:243 [6] NCCL INFO Assigned NET plugin IB to comm
+gpu1:243:243 [6] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu1:243:243 [6] NCCL INFO Using network IB
+gpu1:242:242 [5] NCCL INFO ncclCommInitRankConfig comm 0x48da7ad0 rank 5 nranks 16 cudaDev 5 nvmlDev 5 busId ab000 commId 0x2f8347829d2566aa - Init START
+gpu1:243:243 [6] NCCL INFO DMA-BUF is available on GPU device 6
+gpu1:238:238 [1] NCCL INFO ncclCommInitRankConfig comm 0x772fcba0 rank 1 nranks 16 cudaDev 1 nvmlDev 1 busId 1a000 commId 0x2f8347829d2566aa - Init START
+gpu1:244:244 [7] NCCL INFO ncclCommInitRankConfig comm 0x6f841220 rank 7 nranks 16 cudaDev 7 nvmlDev 7 busId af000 commId 0x2f8347829d2566aa - Init START
+gpu1:243:243 [6] NCCL INFO ncclCommInitRankConfig comm 0x5a031460 rank 6 nranks 16 cudaDev 6 nvmlDev 6 busId ae000 commId 0x2f8347829d2566aa - Init START
+gpu1:240:240 [3] NCCL INFO Comm config Blocking set to 1
+gpu1:240:240 [3] NCCL INFO Initialized NET plugin IB
+gpu1:240:240 [3] NCCL INFO Assigned NET plugin IB to comm
+gpu1:240:240 [3] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu1:240:240 [3] NCCL INFO Using network IB
+gpu1:240:240 [3] NCCL INFO DMA-BUF is available on GPU device 3
+gpu1:240:240 [3] NCCL INFO ncclCommInitRankConfig comm 0x7701d240 rank 3 nranks 16 cudaDev 3 nvmlDev 3 busId 1d000 commId 0x2f8347829d2566aa - Init START
+gpu1:239:239 [2] NCCL INFO Comm config Blocking set to 1
+gpu1:239:239 [2] NCCL INFO Initialized NET plugin IB
+gpu1:239:239 [2] NCCL INFO Assigned NET plugin IB to comm
+gpu1:239:239 [2] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu1:239:239 [2] NCCL INFO Using network IB
+gpu1:239:239 [2] NCCL INFO DMA-BUF is available on GPU device 2
+gpu1:239:239 [2] NCCL INFO ncclCommInitRankConfig comm 0x3db245d0 rank 2 nranks 16 cudaDev 2 nvmlDev 2 busId 1b000 commId 0x2f8347829d2566aa - Init START
+gpu1:244:244 [7] NCCL INFO Bootstrap timings total 0.016093 (create 0.000067, send 0.000227, recv 0.000288, ring 0.014919, delay 0.000000)
+gpu1:243:243 [6] NCCL INFO Bootstrap timings total 0.016052 (create 0.000059, send 0.000233, recv 0.000508, ring 0.014934, delay 0.000000)
+gpu1:241:241 [4] NCCL INFO Bootstrap timings total 0.017807 (create 0.000053, send 0.000238, recv 0.001559, ring 0.004630, delay 0.000000)
+gpu1:240:240 [3] NCCL INFO Bootstrap timings total 0.005339 (create 0.000055, send 0.000263, recv 0.000101, ring 0.000754, delay 0.000000)
+gpu1:242:242 [5] NCCL INFO Bootstrap timings total 0.016540 (create 0.000058, send 0.000234, recv 0.000829, ring 0.015094, delay 0.000000)
+gpu1:239:239 [2] NCCL INFO Bootstrap timings total 0.001757 (create 0.000058, send 0.000232, recv 0.000389, ring 0.000778, delay 0.000000)
+gpu1:237:237 [0] NCCL INFO Bootstrap timings total 0.019236 (create 0.000044, send 0.000239, recv 0.003113, ring 0.015555, delay 0.000000)
+gpu1:238:238 [1] NCCL INFO Bootstrap timings total 0.016565 (create 0.000063, send 0.000224, recv 0.015105, ring 0.000769, delay 0.000000)
+gpu1:243:243 [6] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 6 is 32-63,96-127. (GPU affinity = 32-63,96-127 ; CPU affinity = 0-127).
+gpu1:242:242 [5] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 5 is 32-63,96-127. (GPU affinity = 32-63,96-127 ; CPU affinity = 0-127).
+gpu1:239:239 [2] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 2 is 0-31,64-95. (GPU affinity = 0-31,64-95 ; CPU affinity = 0-127).
+gpu1:240:240 [3] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 3 is 0-31,64-95. (GPU affinity = 0-31,64-95 ; CPU affinity = 0-127).
+gpu1:237:237 [0] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 0 is 0-31,64-95. (GPU affinity = 0-31,64-95 ; CPU affinity = 0-127).
+gpu1:244:244 [7] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 7 is 32-63,96-127. (GPU affinity = 32-63,96-127 ; CPU affinity = 0-127).
+gpu1:238:238 [1] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 1 is 0-31,64-95. (GPU affinity = 0-31,64-95 ; CPU affinity = 0-127).
+gpu1:241:241 [4] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 4 is 32-63,96-127. (GPU affinity = 32-63,96-127 ; CPU affinity = 0-127).
+gpu1:241:241 [4] NCCL INFO comm 0x5d6feee0 rank 4 nRanks 16 nNodes 4 localRanks 4 localRank 0 MNNVL 0
+gpu1:244:244 [7] NCCL INFO comm 0x6f841220 rank 7 nRanks 16 nNodes 4 localRanks 4 localRank 3 MNNVL 0
+gpu1:242:242 [5] NCCL INFO comm 0x48da7ad0 rank 5 nRanks 16 nNodes 4 localRanks 4 localRank 1 MNNVL 0
+gpu1:240:240 [3] NCCL INFO comm 0x7701d240 rank 3 nRanks 16 nNodes 4 localRanks 4 localRank 3 MNNVL 0
+gpu1:244:244 [7] NCCL INFO Trees [0] -1/-1/-1->7->6 [1] 5/-1/-1->7->11 [2] -1/-1/-1->7->6 [3] 5/11/3->7->15
+gpu1:241:241 [4] NCCL INFO Trees [0] 5/-1/-1->4->8 [1] -1/-1/-1->4->6 [2] 5/8/0->4->12 [3] -1/-1/-1->4->6
+gpu1:244:244 [7] NCCL INFO P2P Chunksize set to 131072
+gpu1:241:241 [4] NCCL INFO P2P Chunksize set to 131072
+gpu1:244:244 [7] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu1:241:241 [4] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu1:243:243 [6] NCCL INFO comm 0x5a031460 rank 6 nRanks 16 nNodes 4 localRanks 4 localRank 2 MNNVL 0
+gpu1:239:239 [2] NCCL INFO comm 0x3db245d0 rank 2 nRanks 16 nNodes 4 localRanks 4 localRank 2 MNNVL 0
+gpu1:240:240 [3] NCCL INFO Trees [0] -1/-1/-1->3->2 [1] 1/11/-1->3->-1 [2] -1/-1/-1->3->2 [3] 1/-1/-1->3->7
+gpu1:240:240 [3] NCCL INFO P2P Chunksize set to 131072
+gpu1:239:239 [2] NCCL INFO Trees [0] 3/-1/-1->2->1 [1] 0/-1/-1->2->1 [2] 3/-1/-1->2->1 [3] 0/-1/-1->2->1
+gpu1:240:240 [3] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu1:239:239 [2] NCCL INFO P2P Chunksize set to 131072
+gpu1:237:237 [0] NCCL INFO comm 0x899bd480 rank 0 nRanks 16 nNodes 4 localRanks 4 localRank 0 MNNVL 0
+gpu1:243:243 [6] NCCL INFO Trees [0] 7/-1/-1->6->5 [1] 4/-1/-1->6->5 [2] 7/-1/-1->6->5 [3] 4/-1/-1->6->5
+gpu1:238:238 [1] NCCL INFO comm 0x772fcba0 rank 1 nRanks 16 nNodes 4 localRanks 4 localRank 1 MNNVL 0
+gpu1:239:239 [2] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu1:243:243 [6] NCCL INFO P2P Chunksize set to 131072
+gpu1:242:242 [5] NCCL INFO Trees [0] 6/-1/-1->5->4 [1] 6/-1/-1->5->7 [2] 6/-1/-1->5->4 [3] 6/-1/-1->5->7
+gpu1:242:242 [5] NCCL INFO P2P Chunksize set to 131072
+gpu1:243:243 [6] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu1:237:237 [0] NCCL INFO Channel 00/04 : 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+gpu1:237:237 [0] NCCL INFO Channel 01/04 : 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+gpu1:237:237 [0] NCCL INFO Channel 02/04 : 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+gpu1:237:237 [0] NCCL INFO Channel 03/04 : 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+gpu1:237:237 [0] NCCL INFO Trees [0] 1/8/-1->0->-1 [1] -1/-1/-1->0->2 [2] 1/-1/-1->0->4 [3] -1/-1/-1->0->2
+gpu1:242:242 [5] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu1:238:238 [1] NCCL INFO Trees [0] 2/-1/-1->1->0 [1] 2/-1/-1->1->3 [2] 2/-1/-1->1->0 [3] 2/-1/-1->1->3
+gpu1:237:237 [0] NCCL INFO P2P Chunksize set to 131072
+gpu1:238:238 [1] NCCL INFO P2P Chunksize set to 131072
+gpu1:237:237 [0] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu1:238:238 [1] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu1:244:2379 [0] NCCL INFO [Proxy Service] Device 7 CPU core 38
+gpu1:244:2382 [0] NCCL INFO [Proxy Service UDS] Device 7 CPU core 40
+gpu1:239:2381 [0] NCCL INFO [Proxy Service] Device 2 CPU core 28
+gpu1:241:2383 [0] NCCL INFO [Proxy Service] Device 4 CPU core 41
+gpu1:242:2384 [0] NCCL INFO [Proxy Service] Device 5 CPU core 105
+gpu1:240:2380 [0] NCCL INFO [Proxy Service] Device 3 CPU core 66
+gpu1:241:2388 [0] NCCL INFO [Proxy Service UDS] Device 4 CPU core 106
+gpu1:239:2385 [0] NCCL INFO [Proxy Service UDS] Device 2 CPU core 29
+gpu1:242:2389 [0] NCCL INFO [Proxy Service UDS] Device 5 CPU core 107
+gpu1:240:2386 [0] NCCL INFO [Proxy Service UDS] Device 3 CPU core 67
+gpu1:237:2390 [0] NCCL INFO [Proxy Service] Device 0 CPU core 6
+gpu1:238:2387 [0] NCCL INFO [Proxy Service] Device 1 CPU core 71
+gpu1:237:2392 [0] NCCL INFO [Proxy Service UDS] Device 0 CPU core 8
+gpu1:238:2391 [0] NCCL INFO [Proxy Service UDS] Device 1 CPU core 9
+gpu1:243:2393 [0] NCCL INFO [Proxy Service] Device 6 CPU core 116
+gpu1:243:2394 [0] NCCL INFO [Proxy Service UDS] Device 6 CPU core 117
+gpu1:238:238 [1] NCCL INFO Channel 00/0 : 1[1] -> 2[2] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 00/0 : 2[2] -> 3[3] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 00/0 : 6[6] -> 7[7] via P2P/IPC
+gpu1:238:238 [1] NCCL INFO Channel 01/0 : 1[1] -> 2[2] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 01/0 : 2[2] -> 3[3] via P2P/IPC
+gpu1:242:242 [5] NCCL INFO Channel 00/0 : 5[5] -> 6[6] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 02/0 : 2[2] -> 3[3] via P2P/IPC
+gpu1:238:238 [1] NCCL INFO Channel 02/0 : 1[1] -> 2[2] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 01/0 : 6[6] -> 7[7] via P2P/IPC
+gpu1:242:242 [5] NCCL INFO Channel 01/0 : 5[5] -> 6[6] via P2P/IPC
+gpu1:238:238 [1] NCCL INFO Channel 03/0 : 1[1] -> 2[2] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 02/0 : 6[6] -> 7[7] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 03/0 : 2[2] -> 3[3] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 03/0 : 6[6] -> 7[7] via P2P/IPC
+gpu1:242:242 [5] NCCL INFO Channel 02/0 : 5[5] -> 6[6] via P2P/IPC
+gpu1:242:242 [5] NCCL INFO Channel 03/0 : 5[5] -> 6[6] via P2P/IPC
+gpu1:237:2395 [0] NCCL INFO [Proxy Progress] Device 0 CPU core 10
+gpu1:237:237 [0] NCCL INFO Channel 00/0 : 15[7] -> 0[0] [receive] via NET/IB/0/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 01/0 : 15[7] -> 0[0] [receive] via NET/IB/0/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 02/0 : 15[7] -> 0[0] [receive] via NET/IB/0/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 03/0 : 15[7] -> 0[0] [receive] via NET/IB/0/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 00/0 : 0[0] -> 1[1] via P2P/IPC
+gpu1:237:237 [0] NCCL INFO Channel 01/0 : 0[0] -> 1[1] via P2P/IPC
+gpu1:241:2396 [0] NCCL INFO [Proxy Progress] Device 4 CPU core 108
+gpu1:241:241 [4] NCCL INFO Channel 00/0 : 3[3] -> 4[4] [receive] via NET/IB/1/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 02/0 : 0[0] -> 1[1] via P2P/IPC
+gpu1:241:241 [4] NCCL INFO Channel 01/0 : 3[3] -> 4[4] [receive] via NET/IB/1/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 02/0 : 3[3] -> 4[4] [receive] via NET/IB/1/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 03/0 : 0[0] -> 1[1] via P2P/IPC
+gpu1:241:241 [4] NCCL INFO Channel 03/0 : 3[3] -> 4[4] [receive] via NET/IB/1/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 00/0 : 4[4] -> 5[5] via P2P/IPC
+gpu1:241:241 [4] NCCL INFO Channel 01/0 : 4[4] -> 5[5] via P2P/IPC
+gpu1:241:241 [4] NCCL INFO Channel 02/0 : 4[4] -> 5[5] via P2P/IPC
+gpu1:244:2397 [0] NCCL INFO [Proxy Progress] Device 7 CPU core 109
+gpu1:244:244 [7] NCCL INFO Channel 00/0 : 7[7] -> 8[0] [send] via NET/IB/1/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 01/0 : 7[7] -> 8[0] [send] via NET/IB/1/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 02/0 : 7[7] -> 8[0] [send] via NET/IB/1/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 03/0 : 7[7] -> 8[0] [send] via NET/IB/1/GDRDMA
+gpu1:240:2398 [0] NCCL INFO [Proxy Progress] Device 3 CPU core 11
+gpu1:240:240 [3] NCCL INFO Channel 00/0 : 3[3] -> 4[4] [send] via NET/IB/0/GDRDMA
+gpu1:240:240 [3] NCCL INFO Channel 01/0 : 3[3] -> 4[4] [send] via NET/IB/0/GDRDMA
+gpu1:240:240 [3] NCCL INFO Channel 02/0 : 3[3] -> 4[4] [send] via NET/IB/0/GDRDMA
+gpu1:240:240 [3] NCCL INFO Channel 03/0 : 3[3] -> 4[4] [send] via NET/IB/0/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 03/0 : 4[4] -> 5[5] via P2P/IPC
+gpu1:241:241 [4] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu1:240:240 [3] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu1:244:244 [7] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu1:242:242 [5] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu1:241:241 [4] NCCL INFO Channel 01/0 : 4[4] -> 6[6] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu1:237:237 [0] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu1:239:239 [2] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu1:242:242 [5] NCCL INFO Channel 01/0 : 5[5] -> 7[7] via P2P/IPC
+gpu1:238:238 [1] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu1:237:237 [0] NCCL INFO Channel 01/0 : 0[0] -> 2[2] via P2P/IPC
+gpu1:238:238 [1] NCCL INFO Channel 01/0 : 1[1] -> 3[3] via P2P/IPC
+gpu1:237:237 [0] NCCL INFO Channel 03/0 : 0[0] -> 2[2] via P2P/IPC
+gpu1:238:238 [1] NCCL INFO Channel 03/0 : 1[1] -> 3[3] via P2P/IPC
+gpu1:241:241 [4] NCCL INFO Channel 03/0 : 4[4] -> 6[6] via P2P/IPC
+gpu1:242:242 [5] NCCL INFO Channel 03/0 : 5[5] -> 7[7] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 01/0 : 2[2] -> 0[0] via P2P/IPC
+gpu1:237:237 [0] NCCL INFO Channel 02/0 : 0[0] -> 4[4] [send] via NET/IB/0/GDRDMA
+gpu1:240:240 [3] NCCL INFO Channel 03/0 : 3[3] -> 7[7] [send] via NET/IB/0/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 03/0 : 3[3] -> 7[7] [receive] via NET/IB/1/GDRDMA
+gpu1:243:243 [6] NCCL INFO Channel 01/0 : 6[6] -> 4[4] via P2P/IPC
+gpu1:244:244 [7] NCCL INFO Channel 01/0 : 7[7] -> 11[3] [send] via NET/IB/1/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 03/0 : 7[7] -> 11[3] [send] via NET/IB/1/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 02/0 : 0[0] -> 4[4] [receive] via NET/IB/1/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 00/0 : 4[4] -> 8[0] [send] via NET/IB/1/GDRDMA
+gpu1:240:240 [3] NCCL INFO Channel 01/0 : 11[3] -> 3[3] [receive] via NET/IB/0/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 02/0 : 4[4] -> 8[0] [send] via NET/IB/1/GDRDMA
+gpu1:240:240 [3] NCCL INFO Channel 01/0 : 3[3] -> 11[3] [send] via NET/IB/0/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 03/0 : 15[7] -> 7[7] [receive] via NET/IB/1/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 03/0 : 7[7] -> 15[7] [send] via NET/IB/1/GDRDMA
+gpu1:239:239 [2] NCCL INFO Channel 03/0 : 2[2] -> 0[0] via P2P/IPC
+gpu1:237:237 [0] NCCL INFO Channel 00/0 : 8[0] -> 0[0] [receive] via NET/IB/0/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 00/0 : 0[0] -> 8[0] [send] via NET/IB/0/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 01/0 : 11[3] -> 7[7] [receive] via NET/IB/1/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 02/0 : 12[4] -> 4[4] [receive] via NET/IB/1/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 03/0 : 11[3] -> 7[7] [receive] via NET/IB/1/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 02/0 : 4[4] -> 12[4] [send] via NET/IB/1/GDRDMA
+gpu1:240:240 [3] NCCL INFO Channel 03/0 : 7[7] -> 3[3] [receive] via NET/IB/0/GDRDMA
+gpu1:243:243 [6] NCCL INFO Channel 03/0 : 6[6] -> 4[4] via P2P/IPC
+gpu1:244:244 [7] NCCL INFO Channel 03/0 : 7[7] -> 3[3] [send] via NET/IB/1/GDRDMA
+gpu1:237:237 [0] NCCL INFO Channel 02/0 : 4[4] -> 0[0] [receive] via NET/IB/0/GDRDMA
+gpu1:240:240 [3] NCCL INFO Channel 01/0 : 3[3] -> 1[1] via P2P/IPC
+gpu1:241:241 [4] NCCL INFO Channel 00/0 : 8[0] -> 4[4] [receive] via NET/IB/1/GDRDMA
+gpu1:241:241 [4] NCCL INFO Channel 02/0 : 8[0] -> 4[4] [receive] via NET/IB/1/GDRDMA
+gpu1:244:244 [7] NCCL INFO Channel 01/0 : 7[7] -> 5[5] via P2P/IPC
+gpu1:241:241 [4] NCCL INFO Channel 02/0 : 4[4] -> 0[0] [send] via NET/IB/1/GDRDMA
+gpu1:240:240 [3] NCCL INFO Channel 03/0 : 3[3] -> 1[1] via P2P/IPC
+gpu1:244:244 [7] NCCL INFO Channel 03/0 : 7[7] -> 5[5] via P2P/IPC
+gpu1:244:244 [7] NCCL INFO Channel 00/0 : 7[7] -> 6[6] via P2P/IPC
+gpu1:240:240 [3] NCCL INFO Channel 00/0 : 3[3] -> 2[2] via P2P/IPC
+gpu1:240:240 [3] NCCL INFO Channel 02/0 : 3[3] -> 2[2] via P2P/IPC
+gpu1:244:244 [7] NCCL INFO Channel 02/0 : 7[7] -> 6[6] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 00/0 : 6[6] -> 5[5] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 00/0 : 2[2] -> 1[1] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 01/0 : 6[6] -> 5[5] via P2P/IPC
+gpu1:238:238 [1] NCCL INFO Channel 00/0 : 1[1] -> 0[0] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 01/0 : 2[2] -> 1[1] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 02/0 : 6[6] -> 5[5] via P2P/IPC
+gpu1:238:238 [1] NCCL INFO Channel 02/0 : 1[1] -> 0[0] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 02/0 : 2[2] -> 1[1] via P2P/IPC
+gpu1:242:242 [5] NCCL INFO Channel 00/0 : 5[5] -> 4[4] via P2P/IPC
+gpu1:243:243 [6] NCCL INFO Channel 03/0 : 6[6] -> 5[5] via P2P/IPC
+gpu1:239:239 [2] NCCL INFO Channel 03/0 : 2[2] -> 1[1] via P2P/IPC
+gpu1:242:242 [5] NCCL INFO Channel 02/0 : 5[5] -> 4[4] via P2P/IPC
+gpu1:241:241 [4] NCCL INFO Connected all trees
+gpu1:242:242 [5] NCCL INFO Connected all trees
+gpu1:240:240 [3] NCCL INFO Connected all trees
+gpu1:239:239 [2] NCCL INFO Connected all trees
+gpu1:238:238 [1] NCCL INFO Connected all trees
+gpu1:244:244 [7] NCCL INFO Connected all trees
+gpu1:237:237 [0] NCCL INFO Connected all trees
+gpu1:243:243 [6] NCCL INFO Connected all trees
+gpu1:244:244 [7] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu1:244:244 [7] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu1:241:241 [4] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu1:241:241 [4] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu1:240:240 [3] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu1:240:240 [3] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu1:237:237 [0] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu1:237:237 [0] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu1:237:237 [0] NCCL INFO CC Off, workFifoBytes 1048576
+gpu1:243:2399 [0] NCCL INFO [Proxy Progress] Device 6 CPU core 54
+gpu1:239:2400 [0] NCCL INFO [Proxy Progress] Device 2 CPU core 29
+gpu1:242:2401 [0] NCCL INFO [Proxy Progress] Device 5 CPU core 110
+gpu1:238:2402 [0] NCCL INFO [Proxy Progress] Device 1 CPU core 12
+gpu1:239:239 [2] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu1:239:239 [2] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu1:243:243 [6] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu1:243:243 [6] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu1:242:242 [5] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu1:242:242 [5] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu1:238:238 [1] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu1:238:238 [1] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu1:239:239 [2] NCCL INFO ncclCommInitRankConfig comm 0x3db245d0 rank 2 nranks 16 cudaDev 2 nvmlDev 2 busId 1b000 commId 0x2f8347829d2566aa - Init COMPLETE
+gpu1:239:239 [2] NCCL INFO Init timings - ncclCommInitRankConfig: rank 2 nranks 16 total 0.32 (kernels 0.00, alloc 0.00, bootstrap 0.00, allgathers 0.04, topo 0.06, graphs 0.00, connections 0.21, rest 0.00)
+gpu1:237:237 [0] NCCL INFO ncclCommInitRankConfig comm 0x899bd480 rank 0 nranks 16 cudaDev 0 nvmlDev 0 busId 19000 commId 0x2f8347829d2566aa - Init COMPLETE
+gpu1:240:240 [3] NCCL INFO ncclCommInitRankConfig comm 0x7701d240 rank 3 nranks 16 cudaDev 3 nvmlDev 3 busId 1d000 commId 0x2f8347829d2566aa - Init COMPLETE
+gpu1:240:240 [3] NCCL INFO Init timings - ncclCommInitRankConfig: rank 3 nranks 16 total 0.32 (kernels 0.00, alloc 0.00, bootstrap 0.01, allgathers 0.03, topo 0.06, graphs 0.01, connections 0.16, rest 0.05)
+gpu1:237:237 [0] NCCL INFO Init timings - ncclCommInitRankConfig: rank 0 nranks 16 total 0.33 (kernels 0.00, alloc 0.00, bootstrap 0.02, allgathers 0.03, topo 0.06, graphs 0.01, connections 0.16, rest 0.05)
+gpu1:238:238 [1] NCCL INFO ncclCommInitRankConfig comm 0x772fcba0 rank 1 nranks 16 cudaDev 1 nvmlDev 1 busId 1a000 commId 0x2f8347829d2566aa - Init COMPLETE
+gpu1:238:238 [1] NCCL INFO Init timings - ncclCommInitRankConfig: rank 1 nranks 16 total 0.33 (kernels 0.00, alloc 0.00, bootstrap 0.02, allgathers 0.03, topo 0.06, graphs 0.01, connections 0.21, rest 0.00)
+gpu1:244:244 [7] NCCL INFO ncclCommInitRankConfig comm 0x6f841220 rank 7 nranks 16 cudaDev 7 nvmlDev 7 busId af000 commId 0x2f8347829d2566aa - Init COMPLETE
+gpu1:242:242 [5] NCCL INFO ncclCommInitRankConfig comm 0x48da7ad0 rank 5 nranks 16 cudaDev 5 nvmlDev 5 busId ab000 commId 0x2f8347829d2566aa - Init COMPLETE
+gpu1:244:244 [7] NCCL INFO Init timings - ncclCommInitRankConfig: rank 7 nranks 16 total 0.33 (kernels 0.00, alloc 0.00, bootstrap 0.02, allgathers 0.04, topo 0.06, graphs 0.00, connections 0.16, rest 0.05)
+gpu1:242:242 [5] NCCL INFO Init timings - ncclCommInitRankConfig: rank 5 nranks 16 total 0.33 (kernels 0.00, alloc 0.00, bootstrap 0.02, allgathers 0.04, topo 0.05, graphs 0.01, connections 0.21, rest 0.00)
+gpu1:241:241 [4] NCCL INFO ncclCommInitRankConfig comm 0x5d6feee0 rank 4 nranks 16 cudaDev 4 nvmlDev 4 busId aa000 commId 0x2f8347829d2566aa - Init COMPLETE
+gpu1:241:241 [4] NCCL INFO Init timings - ncclCommInitRankConfig: rank 4 nranks 16 total 0.33 (kernels 0.00, alloc 0.00, bootstrap 0.02, allgathers 0.00, topo 0.09, graphs 0.01, connections 0.16, rest 0.05)
+gpu1:243:243 [6] NCCL INFO ncclCommInitRankConfig comm 0x5a031460 rank 6 nranks 16 cudaDev 6 nvmlDev 6 busId ae000 commId 0x2f8347829d2566aa - Init COMPLETE
+gpu1:243:243 [6] NCCL INFO Init timings - ncclCommInitRankConfig: rank 6 nranks 16 total 0.33 (kernels 0.00, alloc 0.00, bootstrap 0.02, allgathers 0.04, topo 0.05, graphs 0.01, connections 0.21, rest 0.00)
+[TensorRT-LLM][INFO] Set logger level to INFO
+[TensorRT-LLM][INFO] Set logger level to INFO
+[TensorRT-LLM][INFO] Set logger level to INFO
+[TensorRT-LLM][INFO] Set logger level to INFO
+[TensorRT-LLM][INFO] Set logger level to INFO
+[TensorRT-LLM][INFO] Set logger level to INFO
+[TensorRT-LLM][INFO] Set logger level to INFO
+[TensorRT-LLM][INFO] Set logger level to INFO
+[2026-07-30 14:13:40 TP6 EP0] FlashInfer autotune completed.
+[2026-07-30 14:13:40 TP6 EP0] Disable prefill CUDA graph because cuda_graph_config resolved prefill.backend='disabled' (e.g. via --cuda-graph-backend-prefill=disabled or auto-disable rules).
+[2026-07-30 14:13:40 TP7 EP0] FlashInfer autotune completed.
+[2026-07-30 14:13:40 TP3 EP0] FlashInfer autotune completed.
+[2026-07-30 14:13:40 TP7 EP0] Disable prefill CUDA graph because cuda_graph_config resolved prefill.backend='disabled' (e.g. via --cuda-graph-backend-prefill=disabled or auto-disable rules).
+[2026-07-30 14:13:40 TP4 EP0] FlashInfer autotune completed.
+[2026-07-30 14:13:40 TP3 EP0] Disable prefill CUDA graph because cuda_graph_config resolved prefill.backend='disabled' (e.g. via --cuda-graph-backend-prefill=disabled or auto-disable rules).
+[2026-07-30 14:13:40 TP4 EP0] Disable prefill CUDA graph because cuda_graph_config resolved prefill.backend='disabled' (e.g. via --cuda-graph-backend-prefill=disabled or auto-disable rules).
+[2026-07-30 14:13:40 TP0 EP0] FlashInfer autotune completed.
+[2026-07-30 14:13:40 TP0 EP0] Disable prefill CUDA graph because cuda_graph_config resolved prefill.backend='disabled' (e.g. via --cuda-graph-backend-prefill=disabled or auto-disable rules).
+[2026-07-30 14:13:40 TP1 EP0] FlashInfer autotune completed.
+[2026-07-30 14:13:40 TP1 EP0] Disable prefill CUDA graph because cuda_graph_config resolved prefill.backend='disabled' (e.g. via --cuda-graph-backend-prefill=disabled or auto-disable rules).
+[2026-07-30 14:13:40 TP2 EP0] FlashInfer autotune completed.
+[2026-07-30 14:13:40 TP2 EP0] Disable prefill CUDA graph because cuda_graph_config resolved prefill.backend='disabled' (e.g. via --cuda-graph-backend-prefill=disabled or auto-disable rules).
+[2026-07-30 14:13:40 TP5 EP0] FlashInfer autotune completed.
+[2026-07-30 14:13:40 TP5 EP0] Disable prefill CUDA graph because cuda_graph_config resolved prefill.backend='disabled' (e.g. via --cuda-graph-backend-prefill=disabled or auto-disable rules).
+[2026-07-30 14:13:40 TP6 EP0] Capture target decode CUDA graph begin. backend=full, num_tokens_per_req=1, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], avail mem=6.86 GB
+[2026-07-30 14:13:40 TP7 EP0] Capture target decode CUDA graph begin. backend=full, num_tokens_per_req=1, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], avail mem=6.69 GB
+[2026-07-30 14:13:40 TP0 EP0] Capture target decode CUDA graph begin. backend=full, num_tokens_per_req=1, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], avail mem=6.80 GB
+[2026-07-30 14:13:40 TP3 EP0] Capture target decode CUDA graph begin. backend=full, num_tokens_per_req=1, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], avail mem=6.77 GB
+[2026-07-30 14:13:40 TP4 EP0] Capture target decode CUDA graph begin. backend=full, num_tokens_per_req=1, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], avail mem=6.72 GB
+[2026-07-30 14:13:40 TP1 EP0] Capture target decode CUDA graph begin. backend=full, num_tokens_per_req=1, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], avail mem=6.83 GB
+[2026-07-30 14:13:40 TP2 EP0] Capture target decode CUDA graph begin. backend=full, num_tokens_per_req=1, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], avail mem=6.86 GB
+[2026-07-30 14:13:40 TP5 EP0] Capture target decode CUDA graph begin. backend=full, num_tokens_per_req=1, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], avail mem=6.83 GB
+
0%| | 0/12 [00:00, ?it/s]
Capturing batches (bs=64 avail_mem=6.78 GB): 0%| | 0/12 [00:00, ?it/s]/usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:322: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
+ warnings.warn(
+/usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:322: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
+ warnings.warn(
+/usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:322: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
+ warnings.warn(
+/usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:322: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
+ warnings.warn(
+/usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:322: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
+ warnings.warn(
+/usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:322: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
+ warnings.warn(
+/usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:322: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
+ warnings.warn(
+/usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:322: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
+ warnings.warn(
+
Capturing batches (bs=64 avail_mem=6.78 GB): 8%|▊ | 1/12 [00:10<01:51, 10.12s/it]
Capturing batches (bs=56 avail_mem=6.02 GB): 8%|▊ | 1/12 [00:10<01:51, 10.12s/it]
Capturing batches (bs=56 avail_mem=6.02 GB): 17%|█▋ | 2/12 [00:14<01:10, 7.01s/it]
Capturing batches (bs=48 avail_mem=5.97 GB): 17%|█▋ | 2/12 [00:14<01:10, 7.01s/it]
Capturing batches (bs=48 avail_mem=5.97 GB): 25%|██▌ | 3/12 [00:22<01:04, 7.22s/it]
Capturing batches (bs=40 avail_mem=5.92 GB): 25%|██▌ | 3/12 [00:22<01:04, 7.22s/it]
Capturing batches (bs=40 avail_mem=5.92 GB): 33%|███▎ | 4/12 [00:29<00:58, 7.34s/it]
Capturing batches (bs=32 avail_mem=5.86 GB): 33%|███▎ | 4/12 [00:29<00:58, 7.34s/it]
Capturing batches (bs=32 avail_mem=5.86 GB): 42%|████▏ | 5/12 [00:37<00:51, 7.32s/it]
Capturing batches (bs=24 avail_mem=5.81 GB): 42%|████▏ | 5/12 [00:37<00:51, 7.32s/it]
Capturing batches (bs=24 avail_mem=5.81 GB): 50%|█████ | 6/12 [00:44<00:44, 7.34s/it]
Capturing batches (bs=16 avail_mem=5.76 GB): 50%|█████ | 6/12 [00:44<00:44, 7.34s/it]
Capturing batches (bs=16 avail_mem=5.76 GB): 58%|█████▊ | 7/12 [00:51<00:36, 7.35s/it]
Capturing batches (bs=12 avail_mem=5.70 GB): 58%|█████▊ | 7/12 [00:51<00:36, 7.35s/it]
Capturing batches (bs=12 avail_mem=5.70 GB): 67%|██████▋ | 8/12 [00:59<00:29, 7.36s/it]
Capturing batches (bs=8 avail_mem=5.65 GB): 67%|██████▋ | 8/12 [00:59<00:29, 7.36s/it]
Capturing batches (bs=8 avail_mem=5.65 GB): 75%|███████▌ | 9/12 [01:06<00:22, 7.36s/it]
Capturing batches (bs=4 avail_mem=5.60 GB): 75%|███████▌ | 9/12 [01:06<00:22, 7.36s/it]
Capturing batches (bs=4 avail_mem=5.60 GB): 83%|████████▎ | 10/12 [01:13<00:14, 7.33s/it]
Capturing batches (bs=2 avail_mem=5.54 GB): 83%|████████▎ | 10/12 [01:13<00:14, 7.33s/it]
Capturing batches (bs=2 avail_mem=5.54 GB): 92%|█████████▏| 11/12 [01:21<00:07, 7.46s/it]
Capturing batches (bs=1 avail_mem=5.49 GB): 92%|█████████▏| 11/12 [01:21<00:07, 7.46s/it]
Capturing batches (bs=1 avail_mem=5.49 GB): 100%|██████████| 12/12 [01:31<00:00, 8.03s/it]
Capturing batches (bs=1 avail_mem=5.49 GB): 100%|██████████| 12/12 [01:31<00:00, 7.59s/it]
+[2026-07-30 14:15:12 TP6 EP0] Capture target decode CUDA graph end. elapsed=92.29 s, mem usage=1.37 GB, avail mem=5.50 GB.
+[2026-07-30 14:15:12 TP6 EP0] Init Unified RadixTree with components (, )
+[2026-07-30 14:15:12 TP6 EP0] Tree cache initialized: source=default impl=UnifiedRadixCache hybrid_swa=True hybrid_ssm=False hierarchical=False streaming_wrapped=False
+[2026-07-30 14:15:12 TP2 EP0] Capture target decode CUDA graph end. elapsed=92.27 s, mem usage=1.37 GB, avail mem=5.50 GB.
+[2026-07-30 14:15:12 TP2 EP0] Init Unified RadixTree with components (, )
+[2026-07-30 14:15:12 TP2 EP0] Tree cache initialized: source=default impl=UnifiedRadixCache hybrid_swa=True hybrid_ssm=False hierarchical=False streaming_wrapped=False
+[2026-07-30 14:15:12 TP5 EP0] Capture target decode CUDA graph end. elapsed=92.24 s, mem usage=1.37 GB, avail mem=5.46 GB.
+[2026-07-30 14:15:12 TP5 EP0] Init Unified RadixTree with components (, )
+[2026-07-30 14:15:12 TP5 EP0] Tree cache initialized: source=default impl=UnifiedRadixCache hybrid_swa=True hybrid_ssm=False hierarchical=False streaming_wrapped=False
+[2026-07-30 14:15:12 TP7 EP0] Capture target decode CUDA graph end. elapsed=92.32 s, mem usage=1.37 GB, avail mem=5.32 GB.
+[2026-07-30 14:15:12 TP0 EP0] Capture target decode CUDA graph end. elapsed=92.32 s, mem usage=1.37 GB, avail mem=5.43 GB.
+[2026-07-30 14:15:12 TP3 EP0] Capture target decode CUDA graph end. elapsed=92.32 s, mem usage=1.37 GB, avail mem=5.40 GB.
+[2026-07-30 14:15:12 TP0 EP0] max_total_num_tokens=1270016, chunked_prefill_size=8192, max_prefill_tokens=16384, max_running_requests=256, context_len=1048576, available_gpu_mem=5.43 GB
+[2026-07-30 14:15:12 TP7 EP0] Init Unified RadixTree with components (, )
+[2026-07-30 14:15:12 TP7 EP0] Tree cache initialized: source=default impl=UnifiedRadixCache hybrid_swa=True hybrid_ssm=False hierarchical=False streaming_wrapped=False
+[2026-07-30 14:15:12 TP3 EP0] Init Unified RadixTree with components (, )
+[2026-07-30 14:15:12 TP0 EP0] Init Unified RadixTree with components (, )
+[2026-07-30 14:15:12 TP3 EP0] Tree cache initialized: source=default impl=UnifiedRadixCache hybrid_swa=True hybrid_ssm=False hierarchical=False streaming_wrapped=False
+[2026-07-30 14:15:12 TP4 EP0] Capture target decode CUDA graph end. elapsed=92.34 s, mem usage=1.37 GB, avail mem=5.35 GB.
+[2026-07-30 14:15:12 TP0 EP0] Tree cache initialized: source=default impl=UnifiedRadixCache hybrid_swa=True hybrid_ssm=False hierarchical=False streaming_wrapped=False
+[2026-07-30 14:15:12 TP4 EP0] Init Unified RadixTree with components (, )
+[2026-07-30 14:15:12 TP4 EP0] Tree cache initialized: source=default impl=UnifiedRadixCache hybrid_swa=True hybrid_ssm=False hierarchical=False streaming_wrapped=False
+[2026-07-30 14:15:12 TP1 EP0] Capture target decode CUDA graph end. elapsed=92.53 s, mem usage=1.37 GB, avail mem=5.46 GB.
+[2026-07-30 14:15:12 TP1 EP0] Init Unified RadixTree with components (, )
+[2026-07-30 14:15:12 TP1 EP0] Tree cache initialized: source=default impl=UnifiedRadixCache hybrid_swa=True hybrid_ssm=False hierarchical=False streaming_wrapped=False
+[2026-07-30 14:15:13] INFO: Started server process [1]
+[2026-07-30 14:15:13] INFO: Waiting for application startup.
+[2026-07-30 14:15:13] Using default chat sampling params from model generation config: {'temperature': 1.0, 'top_p': 1.0}
+[2026-07-30 14:15:13] INFO: Application startup complete.
+[2026-07-30 14:15:13] INFO: Uvicorn running on http://0.0.0.0:30002 (Press CTRL+C to quit)
+[2026-07-30 14:15:14] INFO: 10.101.0.11:38526 - "GET /health HTTP/1.1" 503 Service Unavailable
+[2026-07-30 14:15:14] INFO: 127.0.0.1:36774 - "GET /model_info HTTP/1.1" 200 OK
+[2026-07-30 14:15:18 TP0 EP0] Prefill batch, #new-seq: 1, #new-token: 256, #cached-token: 0, full token usage: 0.00, swa token usage: 0.00, #running-req: 0, #queue-req: 0, #pending-token: 0, cuda graph: False, input throughput (token/s): 43.97
+[2026-07-30 14:15:18] INFO: 127.0.0.1:36790 - "POST /generate HTTP/1.1" 200 OK
+[2026-07-30 14:15:18] The server is fired up and ready to roll!
+[2026-07-30 14:15:20 TP0 EP0] Prefill batch, #new-seq: 1, #new-token: 256, #cached-token: 0, full token usage: 0.00, swa token usage: 0.00, #running-req: 0, #queue-req: 0, #pending-token: 0, cuda graph: False, input throughput (token/s): 120.89
+[2026-07-30 14:15:20] INFO: 10.101.0.11:57594 - "GET /health HTTP/1.1" 200 OK
diff --git a/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/server/head_server_cmd.txt b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/server/head_server_cmd.txt
new file mode 100644
index 0000000..bd8d931
--- /dev/null
+++ b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/server/head_server_cmd.txt
@@ -0,0 +1 @@
+docker run -d --name dsv4pro_pro6000d_2node_sglang_tp16_quick_map_head --gpus all --network host --ipc host --shm-size 20g --ulimit memlock=-1 --ulimit stack=67108864 -v /data/hf_models/DeepSeek-V4-Pro:/data/hf_models/DeepSeek-V4-Pro:ro -v /data/hzy/sglang_cache/dsv4_pro_tp16:/root/.cache -e CUDA_VISIBLE_DEVICES=0\,1\,2\,3\,4\,5\,6\,7 -e PYTHONUNBUFFERED=1 -e HF_HUB_OFFLINE=1 -e TRANSFORMERS_OFFLINE=1 -e PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True -e NCCL_SOCKET_IFNAME=eth0 -e NCCL_IB_HCA==mlx5_0:1\,mlx5_3:1 -e NCCL_CROSS_NIC=1 -e NCCL_DEBUG=INFO -e SGLANG_SHARED_EXPERT_TP1=1 --device /dev/infiniband/rdma_cm --device /dev/infiniband/uverbs0 --device /dev/infiniband/uverbs3 --entrypoint python3 lmsysorg/sglang:nightly-dev-cu13-20260720-b3570a45 -m sglang.launch_server --model-path /data/hf_models/DeepSeek-V4-Pro --tp-size 16 --ep-size 2 --nnodes 2 --node-rank 0 --dist-init-addr 10.101.0.11:20002 --trust-remote-code --host 0.0.0.0 --port 30002 --mem-fraction-static 0.9 --cuda-graph-max-bs-decode 64 --max-running-requests 256
diff --git a/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/server/worker_nccl_transport.log b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/server/worker_nccl_transport.log
new file mode 100644
index 0000000..863ab77
--- /dev/null
+++ b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/server/worker_nccl_transport.log
@@ -0,0 +1,918 @@
+/usr/local/lib/python3.12/dist-packages/torchao/quantization/quant_api.py:1731: SyntaxWarning: invalid escape sequence '\.'
+ """Configuration class for applying different quantization configs to modules or parameters based on their fully qualified names (FQNs).
+/sgl-workspace/sglang/python/sglang/launch_server.py:56: UserWarning: 'python -m sglang.launch_server' is still supported, but 'sglang serve' is the recommended entrypoint.
+ Example: sglang serve --model-path [options]
+ warnings.warn(
+[2026-07-30 14:10:12] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:12] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:12] Breakable CUDA graph is incompatible with DeepSeek-V4 (heavy capture-pool memory pressure); disabling prefill CUDA graph.
+[2026-07-30 14:10:13] Use dsv4 attention backend for DeepseekV4ForCausalLM, setting page_size to 256.
+[2026-07-30 14:10:13] Setting swa_full_tokens_ratio to 0.1 for DeepseekV4ForCausalLM.
+[2026-07-30 14:10:13] Setting KV cache dtype to fp8_e4m3 for DeepseekV4ForCausalLM.
+[2026-07-30 14:10:13] Use flashinfer_mxfp4 as MoE runner backend on SM120 for DeepseekV4
+[2026-07-30 14:10:16] server_args=ServerArgs(model_path='/data/hf_models/DeepSeek-V4-Pro', tokenizer_path='/data/hf_models/DeepSeek-V4-Pro', tokenizer_mode='auto', tokenizer_backend='huggingface', tokenizer_worker_num=1, detokenizer_worker_num=1, skip_tokenizer_init=False, load_format='auto', model_loader_extra_config='{}', trust_remote_code=True, context_length=None, is_embedding=False, enable_multimodal=None, revision=None, model_impl='auto', model_config_parser='auto', json_model_override_args='{}', dtype='auto', quantization=None, quantization_param_path=None, kv_cache_dtype='fp8_e4m3', enable_fp32_lm_head=False, modelopt_quant=None, modelopt_checkpoint_restore_path=None, modelopt_checkpoint_save_path=None, modelopt_export_path=None, quantize_and_serve=False, rl_quant_profile=None, enable_tf32_matmul=False, mem_fraction_static=0.9, max_running_requests=256, max_queued_requests=None, max_total_tokens=None, chunked_prefill_size=8192, enable_dynamic_chunking=False, max_prefill_tokens=16384, prefill_max_requests=None, schedule_policy='fcfs', enable_priority_scheduling=False, disable_priority_preemption=False, default_priority_value=None, abort_on_priority_when_disabled=False, schedule_low_priority_values_first=False, priority_scheduling_preemption_threshold=10, retraction_policy='length', schedule_conservativeness=1.0, page_size=256, swa_full_tokens_ratio=0.1, disable_hybrid_swa_memory=False, radix_eviction_policy='lru', prefill_only_disable_kv_cache=False, disable_radix_cache=False, enable_page_major_kv_layout=False, enable_unified_memory=False, disable_chunked_prefix_cache=False, disable_overlap_schedule=False, num_continuous_decode_steps=1, scheduler_recv_interval=1, enable_mixed_chunk=False, nccl_port=None, dist_timeout=None, dist_init_addr='10.101.0.11:20002', nnodes=2, node_rank=1, tp_size=16, dcp_size=1, pp_size=1, pp_max_micro_batch_size=None, pp_async_batch_depth=0, dp_size=1, load_balance_method='round_robin', attn_cp_size=1, moe_dp_size=1, enable_prefill_cp=False, cp_strategy=None, enable_dsa_cache_layer_split=False, enable_dsa_prefill_context_parallel=False, dsa_prefill_cp_mode='round-robin-split', enable_prefill_context_parallel=False, prefill_cp_mode='in-seq-split', enable_dp_attention=False, enable_dp_attention_local_control_broadcast=False, enable_dp_lm_head=False, enable_attn_tp_input_scattered=False, disable_attn_tp_gather=False, enable_p2p_check=False, device='cuda', base_gpu_id=0, gpu_id_step=1, random_seed=916998606, watchdog_timeout=300, soft_watchdog_timeout=None, sleep_on_idle=False, use_ray=False, custom_sigquit_handler=None, numa_node=None, gc_threshold=None, host='0.0.0.0', port=30002, fastapi_root_path='', smg_grpc_mode=False, grpc_mode=False, grpc_port=None, skip_server_warmup=False, warmups=None, enable_http2=False, ssl_keyfile=None, ssl_certfile=None, ssl_ca_certs=None, ssl_keyfile_password=None, enable_ssl_refresh=False, api_key=None, admin_api_key=None, served_model_name='/data/hf_models/DeepSeek-V4-Pro', weight_version='default', chat_template=None, hf_chat_template_name=None, completion_template=None, file_storage_path='sglang_storage', enable_cache_report=False, reasoning_parser=None, default_chat_template_kwargs=None, strip_thinking_cache=False, enable_strict_thinking=False, tool_call_parser=None, tool_server=None, sampling_defaults='model', asr_max_buffer_seconds=60, asr_max_concurrent_sessions=32, preferred_sampling_params=None, allow_auto_truncate=False, stream_interval=1, batch_notify_size=16, stream_response_default_include_usage=False, incremental_streaming_output=False, enable_streaming_session=False, enable_session_radix_cache=False, log_level='info', log_level_http=None, log_requests=False, log_requests_level=2, log_requests_format='text', log_requests_target=None, uvicorn_access_log_exclude_prefixes=[], crash_dump_folder=None, show_time_cost=False, enable_metrics=False, smg_http_sidecar_port=None, enable_mfu_metrics=False, enable_metrics_for_all_schedulers=False, load_snapshot_publish_interval=15, tokenizer_metrics_custom_labels_header='x-custom-labels', tokenizer_metrics_allowed_custom_labels=None, extra_metric_labels=None, bucket_time_to_first_token=None, bucket_inter_token_latency=None, bucket_e2e_request_latency=None, prompt_tokens_buckets=None, generation_tokens_buckets=None, gc_warning_threshold_secs=0.0, decode_log_interval=40, enable_request_time_stats_logging=False, kv_events_config=None, enable_forward_pass_metrics=False, forward_pass_metrics_worker_id='', forward_pass_metrics_ipc_name=None, enable_trace=False, trace_modules='request', otlp_traces_endpoint='localhost:4317', export_metrics_to_file=False, export_metrics_to_file_dir=None, stat_loggers=None, constrained_json_whitespace_pattern=None, constrained_json_disable_any_whitespace=False, attention_backend='dsv4', decode_attention_backend=None, prefill_attention_backend=None, sampling_backend='flashinfer', grammar_backend='xgrammar', radix_cache_backend=None, mm_attention_backend=None, fp8_gemm_runner_backend='auto', fp4_gemm_runner_backend='auto', bf16_gemm_backend='auto', dsa_prefill_backend=None, dsa_decode_backend=None, dsa_paged_mqa_logits_backend='auto', dsa_topk_backend='sgl-kernel', disable_flashinfer_autotune=False, mamba_backend='triton', cuda_graph_config=CudaGraphConfig(decode=PhaseConfig(backend='full', max_bs=64, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], tc_compiler='eager', full_prefill_max_req=None), prefill=PhaseConfig(backend='disabled', max_bs=8192, bs=[4, 8, 12, 16, 20, 24, 28, 32, 48, 64, 80, 96, 112, 128, 144, 160, 176, 192, 208, 224, 240, 256, 288, 320, 352, 384, 416, 448, 480, 512, 576, 640, 704, 768, 832, 896, 960, 1024, 1280, 1536, 1792, 2048, 2304, 2560, 2816, 3072, 3328, 3584, 3840, 4096, 4608, 5120, 5632, 6144, 6656, 7168, 7680, 8192], tc_compiler='eager', full_prefill_max_req=None)), cuda_graph_backend_decode=None, cuda_graph_backend_prefill=None, cuda_graph_max_bs_decode=64, cuda_graph_max_bs_prefill=None, cuda_graph_bs_decode=None, cuda_graph_bs_prefill=None, cuda_graph_tc_compiler=None, disable_prefill_cuda_graph=False, disable_decode_cuda_graph=False, disable_cuda_graph=False, disable_cuda_graph_padding=False, enable_profile_cuda_graph=False, enable_cudagraph_gc=False, debug_cuda_graph=False, enable_layerwise_nvtx_marker=False, enable_nccl_nvls=False, enable_symm_mem=False, triton_attention_reduce_in_fp32=False, triton_attention_num_kv_splits=8, triton_attention_split_tile_size=None, flashinfer_mla_disable_ragged=False, enable_fused_qk_norm_rope=False, enable_precise_embedding_interpolation=False, enable_fused_moe_sum_all_reduce=False, enable_deepseek_v4_fp4_indexer=False, disable_custom_all_reduce=False, enable_mscclpp=False, enable_torch_symm_mem=False, pre_warm_nccl=False, enable_quant_communications=False, enable_flashinfer_allreduce_fusion=False, enforce_disable_flashinfer_allreduce_fusion=False, flashinfer_allreduce_fusion_backend=None, enable_aiter_allreduce_fusion=False, enable_torch_compile=False, enable_torch_compile_debug_mode=False, torch_compile_max_bs=32, torchao_config='', speculative_algorithm=None, speculative_draft_model_path=None, speculative_draft_model_revision=None, speculative_draft_load_format=None, speculative_num_steps=None, speculative_eagle_topk=None, speculative_num_draft_tokens=None, speculative_dflash_block_size=None, speculative_dspark_block_size=None, speculative_dspark_sps_table_path=None, speculative_dspark_confidence_sts_path=None, speculative_dspark_align_verify_tokens_to_graph_tier=False, speculative_accept_threshold_single=1.0, speculative_accept_threshold_acc=1.0, speculative_use_rejection_sampling=False, speculative_token_map=None, speculative_attention_mode='prefill', speculative_draft_attention_backend=None, speculative_draft_window_size=None, speculative_moe_runner_backend='flashinfer_mxfp4', speculative_moe_a2a_backend=None, speculative_draft_model_quantization=None, speculative_skip_dp_mlp_sync=False, enable_multi_layer_eagle=False, speculative_adaptive=False, speculative_adaptive_config=None, decoupled_spec_bind_endpoint=None, decoupled_spec_connect_endpoints=None, decoupled_spec_rank=None, decoupled_spec_role='null', spec_trace_dir=None, speculative_ngram_min_bfs_breadth=1, speculative_ngram_max_bfs_breadth=10, speculative_ngram_match_type='BFS', speculative_ngram_max_trie_depth=18, speculative_ngram_capacity=10000000, speculative_ngram_external_corpus_path=None, speculative_ngram_external_sam_budget=0, speculative_ngram_external_corpus_max_tokens=10000000, ep_size=2, moe_a2a_backend='none', moe_runner_backend='flashinfer_mxfp4', flashinfer_mxfp4_moe_precision='default', deepep_mode='auto', fuseep_mode=2, deepep_dispatcher_output_dtype='auto', ep_num_redundant_experts=0, ep_dispatch_algorithm=None, init_expert_location='trivial', enable_eplb=False, eplb_algorithm='auto', eplb_rebalance_num_iterations=1000, eplb_rebalance_layers_per_chunk=None, eplb_min_rebalancing_utilization_threshold=1.0, expert_distribution_recorder_mode=None, expert_distribution_recorder_buffer_size=1000, enable_expert_distribution_metrics=False, deepep_config=None, moe_dense_tp_size=None, elastic_ep_backend=None, enable_elastic_expert_backup=False, mooncake_ib_device=None, enable_waterfill=False, ep_join_mode=None, ep_join_rank_offset=0, elastic_ep_initial_size=None, max_ep_size=None, elastic_ep_scale_timeout=600, elastic_ep_rejoin=False, disable_flashinfer_cutlass_moe_fp4_allgather=False, disable_shared_experts_fusion=False, enforce_shared_experts_fusion=False, max_mamba_cache_size=None, mamba_ssm_dtype=None, enable_mamba_cache_stochastic_rounding=False, mamba_cache_philox_rounds=0, mamba_full_memory_ratio=0.9, mamba_radix_cache_strategy='auto', uses_mamba_radix_cache=False, mamba_track_interval=256, enable_int8_mamba_checkpoint=False, int8_mamba_ckpt_size=None, linear_attn_backend='triton', linear_attn_decode_backend=None, linear_attn_prefill_backend=None, enable_linear_replayssm=False, linear_replayssm_cache_len=16, enable_hierarchical_cache=False, hicache_ratio=2.0, hicache_size=0, hicache_write_policy='write_through', hicache_io_backend='kernel', hicache_mem_layout='page_first', hicache_storage_backend=None, hicache_storage_prefetch_policy='timeout', hicache_storage_backend_extra_config=None, enable_hisparse=False, hisparse_config=None, enable_broadcast_mm_inputs_process=False, enable_prefix_mm_cache=False, mm_enable_dp_encoder=False, mm_process_config={}, limit_mm_data_per_request=None, enable_mm_global_cache=False, disable_fast_image_processor=False, mm_feature_transport='cpu', keep_mm_feature_on_device=False, enable_lora=None, enable_lora_overlap_loading=None, max_lora_rank=None, lora_target_modules=None, lora_paths=None, max_loaded_loras=None, max_loras_per_batch=8, lora_eviction_policy='lru', lora_backend='csgmv', max_lora_chunk_size=16, experts_shared_outer_loras=None, lora_use_virtual_experts=False, lora_strict_loading=False, lora_drain_wait_threshold=0.0, enable_two_batch_overlap=False, enable_single_batch_overlap=False, tbo_token_distribution_threshold=0.48, cpu_offload_gb=0, offload_group_size=-1, offload_num_in_group=1, offload_prefetch_step=1, offload_mode='cpu', enable_lmcache=False, lmcache_config_file=None, enable_flexkv=False, flexkv_config_file=None, kt_weight_path=None, kt_method='AMXINT4', kt_cpuinfer=None, kt_threadpool_count=2, kt_num_gpu_experts=None, kt_max_deferred_experts_per_token=None, dllm_algorithm=None, dllm_algorithm_config=None, dllm_fdfo=True, disaggregation_mode='null', disaggregation_transfer_backend='mooncake', disaggregation_bootstrap_port=8998, disaggregation_ib_device=None, disaggregation_decode_enable_radix_cache=False, disaggregation_decode_enable_offload_kvcache=False, num_reserved_decode_tokens=512, disaggregation_decode_extra_slots=None, disaggregation_decode_polling_interval=1, optimistic_prefill_attempts=0, encoder_only=False, language_only=False, encoder_transfer_backend='zmq_to_scheduler', encoder_urls=[], encoder_bootstrap_port=8997, encoder_register_urls=[], enable_adaptive_dispatch_to_encoder=False, enable_pdmux=False, pdmux_config_path=None, sm_group_num=8, custom_weight_loader=[], weight_loader_disable_mmap=False, weight_loader_prefetch_checkpoints=False, weight_loader_prefetch_num_threads=4, weight_loader_drop_cache_after_load=False, remote_instance_weight_loader_seed_instance_ip=None, remote_instance_weight_loader_seed_instance_service_port=None, remote_instance_weight_loader_send_weights_group_ports=None, remote_instance_weight_loader_backend='nccl', remote_instance_weight_loader_start_seed_via_transfer_engine=False, engine_info_bootstrap_port=6789, modelexpress_config=None, download_dir=None, model_checksum=None, delete_ckpt_after_loading=False, decrypted_config_file=None, decrypted_draft_config_file=None, checkpoint_engine_wait_weights_before_ready=False, enable_prefill_delayer=False, prefill_delayer_max_delay_passes=30, prefill_delayer_token_usage_low_watermark=None, prefill_delayer_forward_passes_buckets=None, prefill_delayer_wait_seconds_buckets=None, prefill_delayer_queue_min_ratio=None, prefill_delayer_max_delay_ms=None, min_free_slots_delay=None, enable_deterministic_inference=False, rl_on_policy_target=None, kv_canary='none', kv_canary_real_data='none', kv_canary_sweep_interval=0, enable_dynamic_batch_tokenizer=False, dynamic_batch_tokenizer_batch_size=32, dynamic_batch_tokenizer_batch_timeout=0.002, enable_tokenizer_batch_encode=False, disable_tokenizer_batch_decode=False, debug_tensor_dump_output_folder=None, debug_tensor_dump_layers=None, debug_tensor_dump_input_file=None, enable_memory_saver=False, enable_weights_cpu_backup=False, enable_draft_weights_cpu_backup=False, enable_custom_logit_processor=False, enable_return_hidden_states=False, enable_return_routed_experts=False, enable_return_indexer_topk=False, disable_outlines_disk_cache=False, enable_mis=False, forward_hooks=None, msprobe_dump_config=None)
+[2026-07-30 14:10:16] User lacks permission to set NUMA affinity, skipping NUMA node configuration for GPU. If using docker, try adding --cap-add SYS_NICE to your docker run command.
+[2026-07-30 14:10:16] User lacks permission to set NUMA affinity, skipping NUMA node configuration for GPU. If using docker, try adding --cap-add SYS_NICE to your docker run command.
+[2026-07-30 14:10:16] User lacks permission to set NUMA affinity, skipping NUMA node configuration for GPU. If using docker, try adding --cap-add SYS_NICE to your docker run command.
+[2026-07-30 14:10:16] User lacks permission to set NUMA affinity, skipping NUMA node configuration for GPU. If using docker, try adding --cap-add SYS_NICE to your docker run command.
+[2026-07-30 14:10:16] User lacks permission to set NUMA affinity, skipping NUMA node configuration for GPU. If using docker, try adding --cap-add SYS_NICE to your docker run command.
+[2026-07-30 14:10:16] User lacks permission to set NUMA affinity, skipping NUMA node configuration for GPU. If using docker, try adding --cap-add SYS_NICE to your docker run command.
+[2026-07-30 14:10:16] User lacks permission to set NUMA affinity, skipping NUMA node configuration for GPU. If using docker, try adding --cap-add SYS_NICE to your docker run command.
+[2026-07-30 14:10:16] User lacks permission to set NUMA affinity, skipping NUMA node configuration for GPU. If using docker, try adding --cap-add SYS_NICE to your docker run command.
+[2026-07-30 14:10:29 TP9 EP1] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:29 TP9 EP1] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:30 TP14 EP1] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:30 TP14 EP1] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:30 TP15 EP1] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:30 TP15 EP1] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:30 TP12 EP1] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:30 TP12 EP1] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:30 TP13 EP1] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:30 TP13 EP1] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:30 TP8 EP1] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:30 TP8 EP1] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:30 TP11 EP1] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:30 TP11 EP1] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:30 TP10 EP1] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:30 TP10 EP1] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:31 TP9 EP1] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:32 TP9 EP1] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:32 TP9 EP1] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:32 TP9 EP1] Init torch distributed begin.
+[2026-07-30 14:10:32 TP14 EP1] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:32 TP15 EP1] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:32 TP12 EP1] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:32 TP13 EP1] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:32 TP8 EP1] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:32 TP14 EP1] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:32 TP14 EP1] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:32 TP15 EP1] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:32 TP15 EP1] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:32 TP11 EP1] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:32 TP10 EP1] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:10:32 TP12 EP1] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:32 TP12 EP1] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:32 TP13 EP1] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:32 TP13 EP1] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:32 TP14 EP1] Init torch distributed begin.
+[2026-07-30 14:10:32 TP15 EP1] Init torch distributed begin.
+[2026-07-30 14:10:32 TP12 EP1] Init torch distributed begin.
+[2026-07-30 14:10:32 TP13 EP1] Init torch distributed begin.
+[2026-07-30 14:10:32 TP8 EP1] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:32 TP8 EP1] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:33 TP10 EP1] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:33 TP10 EP1] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:33 TP8 EP1] Init torch distributed begin.
+[2026-07-30 14:10:33 TP11 EP1] Auto-detected DSV4 routed-expert layout: is_fp4_experts=True
+[2026-07-30 14:10:33 TP11 EP1] Hybrid swa model: self.hf_config.architectures=['DeepseekV4ForCausalLM']
+[2026-07-30 14:10:33 TP10 EP1] Init torch distributed begin.
+[2026-07-30 14:10:33 TP11 EP1] Init torch distributed begin.
+gpu3:236:236 [0] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so
+gpu3:236:236 [0] NCCL INFO cudaDriverVersion 13000
+gpu3:237:237 [1] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so
+gpu3:237:237 [1] NCCL INFO cudaDriverVersion 13000
+gpu3:238:238 [2] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so
+gpu3:238:238 [2] NCCL INFO cudaDriverVersion 13000
+gpu3:236:236 [0] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu3:237:237 [1] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu3:238:238 [2] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu3:236:236 [0] NCCL INFO Bootstrap: Using eth0:10.101.0.13<0>
+gpu3:236:236 [0] NCCL INFO NCCL version 2.28.9+cuda13.0
+gpu3:237:237 [1] NCCL INFO Bootstrap: Using eth0:10.101.0.13<0>
+gpu3:238:238 [2] NCCL INFO Bootstrap: Using eth0:10.101.0.13<0>
+gpu3:237:237 [1] NCCL INFO NCCL version 2.28.9+cuda13.0
+gpu3:238:238 [2] NCCL INFO NCCL version 2.28.9+cuda13.0
+gpu3:241:241 [5] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so
+gpu3:241:241 [5] NCCL INFO cudaDriverVersion 13000
+gpu3:239:239 [3] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so
+gpu3:239:239 [3] NCCL INFO cudaDriverVersion 13000
+gpu3:241:241 [5] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu3:239:239 [3] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu3:239:239 [3] NCCL INFO Bootstrap: Using eth0:10.101.0.13<0>
+gpu3:241:241 [5] NCCL INFO Bootstrap: Using eth0:10.101.0.13<0>
+gpu3:239:239 [3] NCCL INFO NCCL version 2.28.9+cuda13.0
+gpu3:241:241 [5] NCCL INFO NCCL version 2.28.9+cuda13.0
+gpu3:243:243 [7] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so
+gpu3:243:243 [7] NCCL INFO cudaDriverVersion 13000
+gpu3:240:240 [4] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so
+gpu3:240:240 [4] NCCL INFO cudaDriverVersion 13000
+gpu3:243:243 [7] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu3:240:240 [4] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu3:242:242 [6] NCCL INFO ENV/Plugin: Could not find: libnccl-env.so
+gpu3:242:242 [6] NCCL INFO cudaDriverVersion 13000
+gpu3:243:243 [7] NCCL INFO Bootstrap: Using eth0:10.101.0.13<0>
+gpu3:243:243 [7] NCCL INFO NCCL version 2.28.9+cuda13.0
+gpu3:240:240 [4] NCCL INFO Bootstrap: Using eth0:10.101.0.13<0>
+gpu3:240:240 [4] NCCL INFO NCCL version 2.28.9+cuda13.0
+gpu3:242:242 [6] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu3:242:242 [6] NCCL INFO Bootstrap: Using eth0:10.101.0.13<0>
+gpu3:242:242 [6] NCCL INFO NCCL version 2.28.9+cuda13.0
+gpu3:236:236 [0] NCCL INFO NET/Plugin: Could not find: libnccl-net.so
+gpu3:236:236 [0] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu3:236:236 [0] NCCL INFO NCCL_IB_HCA set to =mlx5_0:1,mlx5_3:1
+gpu3:238:238 [2] NCCL INFO NET/Plugin: Could not find: libnccl-net.so
+gpu3:239:239 [3] NCCL INFO NET/Plugin: Could not find: libnccl-net.so
+gpu3:237:237 [1] NCCL INFO NET/Plugin: Could not find: libnccl-net.so
+gpu3:238:238 [2] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu3:239:239 [3] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu3:237:237 [1] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu3:238:238 [2] NCCL INFO NCCL_IB_HCA set to =mlx5_0:1,mlx5_3:1
+gpu3:239:239 [3] NCCL INFO NCCL_IB_HCA set to =mlx5_0:1,mlx5_3:1
+gpu3:237:237 [1] NCCL INFO NCCL_IB_HCA set to =mlx5_0:1,mlx5_3:1
+gpu3:241:241 [5] NCCL INFO NET/Plugin: Could not find: libnccl-net.so
+gpu3:241:241 [5] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu3:241:241 [5] NCCL INFO NCCL_IB_HCA set to =mlx5_0:1,mlx5_3:1
+gpu3:237:237 [1] NCCL INFO NET/IB : Using [0]mlx5_0:1/RoCE [1]mlx5_3:1/RoCE [RO]; OOB eth0:10.101.0.13<0>
+gpu3:237:237 [1] NCCL INFO Initialized NET plugin IB
+gpu3:237:237 [1] NCCL INFO Assigned NET plugin IB to comm
+gpu3:237:237 [1] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu3:237:237 [1] NCCL INFO Using network IB
+gpu3:239:239 [3] NCCL INFO NET/IB : Using [0]mlx5_0:1/RoCE [1]mlx5_3:1/RoCE [RO]; OOB eth0:10.101.0.13<0>
+gpu3:239:239 [3] NCCL INFO Initialized NET plugin IB
+gpu3:238:238 [2] NCCL INFO NET/IB : Using [0]mlx5_0:1/RoCE [1]mlx5_3:1/RoCE [RO]; OOB eth0:10.101.0.13<0>
+gpu3:238:238 [2] NCCL INFO Initialized NET plugin IB
+gpu3:239:239 [3] NCCL INFO Assigned NET plugin IB to comm
+gpu3:239:239 [3] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu3:239:239 [3] NCCL INFO Using network IB
+gpu3:238:238 [2] NCCL INFO Assigned NET plugin IB to comm
+gpu3:238:238 [2] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu3:238:238 [2] NCCL INFO Using network IB
+gpu3:241:241 [5] NCCL INFO NET/IB : Using [0]mlx5_0:1/RoCE [1]mlx5_3:1/RoCE [RO]; OOB eth0:10.101.0.13<0>
+gpu3:241:241 [5] NCCL INFO Initialized NET plugin IB
+gpu3:241:241 [5] NCCL INFO Assigned NET plugin IB to comm
+gpu3:241:241 [5] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu3:241:241 [5] NCCL INFO Using network IB
+gpu3:236:236 [0] NCCL INFO NET/IB : Using [0]mlx5_0:1/RoCE [1]mlx5_3:1/RoCE [RO]; OOB eth0:10.101.0.13<0>
+gpu3:236:236 [0] NCCL INFO Initialized NET plugin IB
+gpu3:236:236 [0] NCCL INFO Assigned NET plugin IB to comm
+gpu3:236:236 [0] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu3:236:236 [0] NCCL INFO Using network IB
+gpu3:237:237 [1] NCCL INFO DMA-BUF is available on GPU device 1
+gpu3:237:237 [1] NCCL INFO ncclCommInitRank comm 0x1717b370 rank 9 nranks 16 cudaDev 1 nvmlDev 1 busId 1a000 commId 0x92ed644083b4525 - Init START
+gpu3:243:243 [7] NCCL INFO NET/Plugin: Could not find: libnccl-net.so
+gpu3:236:236 [0] NCCL INFO DMA-BUF is available on GPU device 0
+gpu3:236:236 [0] NCCL INFO ncclCommInitRank comm 0x2fca9260 rank 8 nranks 16 cudaDev 0 nvmlDev 0 busId 19000 commId 0x92ed644083b4525 - Init START
+gpu3:243:243 [7] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu3:243:243 [7] NCCL INFO NCCL_IB_HCA set to =mlx5_0:1,mlx5_3:1
+gpu3:238:238 [2] NCCL INFO DMA-BUF is available on GPU device 2
+gpu3:238:238 [2] NCCL INFO ncclCommInitRank comm 0x261e3fa0 rank 10 nranks 16 cudaDev 2 nvmlDev 2 busId 1b000 commId 0x92ed644083b4525 - Init START
+gpu3:237:237 [1] NCCL INFO RAS client listening socket at 127.0.0.1<28028>
+gpu3:239:239 [3] NCCL INFO DMA-BUF is available on GPU device 3
+gpu3:239:239 [3] NCCL INFO ncclCommInitRank comm 0x26b8f240 rank 11 nranks 16 cudaDev 3 nvmlDev 3 busId 1d000 commId 0x92ed644083b4525 - Init START
+gpu3:238:238 [2] NCCL INFO RAS client listening socket at 127.0.0.1<28028>
+gpu3:241:241 [5] NCCL INFO DMA-BUF is available on GPU device 5
+gpu3:241:241 [5] NCCL INFO ncclCommInitRank comm 0x52be2940 rank 13 nranks 16 cudaDev 5 nvmlDev 5 busId ab000 commId 0x92ed644083b4525 - Init START
+gpu3:243:243 [7] NCCL INFO NET/IB : Using [0]mlx5_0:1/RoCE [1]mlx5_3:1/RoCE [RO]; OOB eth0:10.101.0.13<0>
+gpu3:243:243 [7] NCCL INFO Initialized NET plugin IB
+gpu3:243:243 [7] NCCL INFO Assigned NET plugin IB to comm
+gpu3:243:243 [7] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu3:243:243 [7] NCCL INFO Using network IB
+gpu3:243:243 [7] NCCL INFO DMA-BUF is available on GPU device 7
+gpu3:243:243 [7] NCCL INFO ncclCommInitRank comm 0x420ab990 rank 15 nranks 16 cudaDev 7 nvmlDev 7 busId af000 commId 0x92ed644083b4525 - Init START
+gpu3:240:240 [4] NCCL INFO NET/Plugin: Could not find: libnccl-net.so
+gpu3:240:240 [4] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu3:240:240 [4] NCCL INFO NCCL_IB_HCA set to =mlx5_0:1,mlx5_3:1
+gpu3:242:242 [6] NCCL INFO NET/Plugin: Could not find: libnccl-net.so
+gpu3:242:242 [6] NCCL INFO NCCL_SOCKET_IFNAME set by environment to eth0
+gpu3:242:242 [6] NCCL INFO NCCL_IB_HCA set to =mlx5_0:1,mlx5_3:1
+gpu3:240:240 [4] NCCL INFO NET/IB : Using [0]mlx5_0:1/RoCE [1]mlx5_3:1/RoCE [RO]; OOB eth0:10.101.0.13<0>
+gpu3:240:240 [4] NCCL INFO Initialized NET plugin IB
+gpu3:240:240 [4] NCCL INFO Assigned NET plugin IB to comm
+gpu3:240:240 [4] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu3:240:240 [4] NCCL INFO Using network IB
+gpu3:240:240 [4] NCCL INFO DMA-BUF is available on GPU device 4
+gpu3:240:240 [4] NCCL INFO ncclCommInitRank comm 0x1e210850 rank 12 nranks 16 cudaDev 4 nvmlDev 4 busId aa000 commId 0x92ed644083b4525 - Init START
+gpu3:239:239 [3] NCCL INFO RAS client listening socket at 127.0.0.1<28028>
+gpu3:240:240 [4] NCCL INFO RAS client listening socket at 127.0.0.1<28028>
+gpu3:236:236 [0] NCCL INFO RAS client listening socket at 127.0.0.1<28028>
+gpu3:242:242 [6] NCCL INFO NET/IB : Using [0]mlx5_0:1/RoCE [1]mlx5_3:1/RoCE [RO]; OOB eth0:10.101.0.13<0>
+gpu3:242:242 [6] NCCL INFO Initialized NET plugin IB
+gpu3:242:242 [6] NCCL INFO Assigned NET plugin IB to comm
+gpu3:242:242 [6] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu3:242:242 [6] NCCL INFO Using network IB
+gpu3:242:242 [6] NCCL INFO DMA-BUF is available on GPU device 6
+gpu3:242:242 [6] NCCL INFO ncclCommInitRank comm 0x2e169bd0 rank 14 nranks 16 cudaDev 6 nvmlDev 6 busId ae000 commId 0x92ed644083b4525 - Init START
+gpu3:241:241 [5] NCCL INFO RAS client listening socket at 127.0.0.1<28028>
+gpu3:243:243 [7] NCCL INFO RAS client listening socket at 127.0.0.1<28028>
+gpu3:242:242 [6] NCCL INFO RAS client listening socket at 127.0.0.1<28028>
+gpu3:241:241 [5] NCCL INFO Bootstrap timings total 0.133899 (create 0.000030, send 0.000149, recv 0.078631, ring 0.054772, delay 0.000001)
+gpu3:240:240 [4] NCCL INFO Bootstrap timings total 0.109616 (create 0.000028, send 0.000115, recv 0.000453, ring 0.108739, delay 0.000001)
+gpu3:241:241 [5] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0.
+gpu3:240:240 [4] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0.
+gpu3:242:242 [6] NCCL INFO Bootstrap timings total 0.055548 (create 0.000034, send 0.000154, recv 0.000370, ring 0.054606, delay 0.000001)
+gpu3:243:243 [7] NCCL INFO Bootstrap timings total 0.127575 (create 0.000029, send 0.000147, recv 0.032944, ring 0.054672, delay 0.000001)
+gpu3:242:242 [6] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0.
+gpu3:243:243 [7] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0.
+gpu3:239:239 [3] NCCL INFO Bootstrap timings total 0.135653 (create 0.000024, send 0.000127, recv 0.026276, ring 0.108909, delay 0.000001)
+gpu3:239:239 [3] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0.
+gpu3:238:238 [2] NCCL INFO Bootstrap timings total 0.136863 (create 0.000029, send 0.000147, recv 0.001539, ring 0.134836, delay 0.000001)
+gpu3:237:237 [1] NCCL INFO Bootstrap timings total 0.145687 (create 0.000034, send 0.000195, recv 0.009176, ring 0.135969, delay 0.000001)
+gpu3:236:236 [0] NCCL INFO Bootstrap timings total 0.140719 (create 0.000027, send 0.000115, recv 0.000453, ring 0.101118, delay 0.000001)
+gpu3:238:238 [2] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0.
+gpu3:237:237 [1] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0.
+gpu3:236:236 [0] NCCL INFO NCCL_CUMEM_ENABLE set by environment to 0.
+gpu3:236:236 [0] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 0 is 0-31,64-95. (GPU affinity = 0-31,64-95 ; CPU affinity = 0-127).
+gpu3:236:236 [0] NCCL INFO NCCL_NVLS_ENABLE set by environment to 0.
+gpu3:236:236 [0] NCCL INFO NCCL_CROSS_NIC set by environment to 1.
+gpu3:243:243 [7] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 7 is 32-63,96-127. (GPU affinity = 32-63,96-127 ; CPU affinity = 0-127).
+gpu3:243:243 [7] NCCL INFO NCCL_NVLS_ENABLE set by environment to 0.
+gpu3:243:243 [7] NCCL INFO NCCL_CROSS_NIC set by environment to 1.
+gpu3:237:237 [1] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 1 is 0-31,64-95. (GPU affinity = 0-31,64-95 ; CPU affinity = 0-127).
+gpu3:237:237 [1] NCCL INFO NCCL_NVLS_ENABLE set by environment to 0.
+gpu3:237:237 [1] NCCL INFO NCCL_CROSS_NIC set by environment to 1.
+gpu3:241:241 [5] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 5 is 32-63,96-127. (GPU affinity = 32-63,96-127 ; CPU affinity = 0-127).
+gpu3:241:241 [5] NCCL INFO NCCL_NVLS_ENABLE set by environment to 0.
+gpu3:241:241 [5] NCCL INFO NCCL_CROSS_NIC set by environment to 1.
+gpu3:240:240 [4] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 4 is 32-63,96-127. (GPU affinity = 32-63,96-127 ; CPU affinity = 0-127).
+gpu3:240:240 [4] NCCL INFO NCCL_NVLS_ENABLE set by environment to 0.
+gpu3:240:240 [4] NCCL INFO NCCL_CROSS_NIC set by environment to 1.
+gpu3:238:238 [2] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 2 is 0-31,64-95. (GPU affinity = 0-31,64-95 ; CPU affinity = 0-127).
+gpu3:238:238 [2] NCCL INFO NCCL_NVLS_ENABLE set by environment to 0.
+gpu3:238:238 [2] NCCL INFO NCCL_CROSS_NIC set by environment to 1.
+gpu3:242:242 [6] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 6 is 32-63,96-127. (GPU affinity = 32-63,96-127 ; CPU affinity = 0-127).
+gpu3:242:242 [6] NCCL INFO NCCL_NVLS_ENABLE set by environment to 0.
+gpu3:242:242 [6] NCCL INFO NCCL_CROSS_NIC set by environment to 1.
+gpu3:239:239 [3] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 3 is 0-31,64-95. (GPU affinity = 0-31,64-95 ; CPU affinity = 0-127).
+gpu3:239:239 [3] NCCL INFO NCCL_NVLS_ENABLE set by environment to 0.
+gpu3:239:239 [3] NCCL INFO NCCL_CROSS_NIC set by environment to 1.
+gpu3:241:241 [5] NCCL INFO comm 0x52be2940 rank 13 nRanks 16 nNodes 4 localRanks 4 localRank 1 MNNVL 0
+gpu3:240:240 [4] NCCL INFO comm 0x1e210850 rank 12 nRanks 16 nNodes 4 localRanks 4 localRank 0 MNNVL 0
+gpu3:239:239 [3] NCCL INFO comm 0x26b8f240 rank 11 nRanks 16 nNodes 4 localRanks 4 localRank 3 MNNVL 0
+gpu3:241:241 [5] NCCL INFO Trees [0] 14/-1/-1->13->12 [1] 14/-1/-1->13->15 [2] 14/-1/-1->13->12 [3] 14/-1/-1->13->15
+gpu3:241:241 [5] NCCL INFO P2P Chunksize set to 131072
+gpu3:240:240 [4] NCCL INFO Trees [0] 13/-1/-1->12->8 [1] -1/-1/-1->12->14 [2] 13/4/-1->12->-1 [3] -1/-1/-1->12->14
+gpu3:240:240 [4] NCCL INFO P2P Chunksize set to 131072
+gpu3:242:242 [6] NCCL INFO comm 0x2e169bd0 rank 14 nRanks 16 nNodes 4 localRanks 4 localRank 2 MNNVL 0
+gpu3:239:239 [3] NCCL INFO Trees [0] -1/-1/-1->11->10 [1] 9/7/15->11->3 [2] -1/-1/-1->11->10 [3] 9/-1/-1->11->7
+gpu3:239:239 [3] NCCL INFO P2P Chunksize set to 131072
+gpu3:243:243 [7] NCCL INFO comm 0x420ab990 rank 15 nRanks 16 nNodes 4 localRanks 4 localRank 3 MNNVL 0
+gpu3:242:242 [6] NCCL INFO Trees [0] 15/-1/-1->14->13 [1] 12/-1/-1->14->13 [2] 15/-1/-1->14->13 [3] 12/-1/-1->14->13
+gpu3:243:243 [7] NCCL INFO Trees [0] -1/-1/-1->15->14 [1] 13/-1/-1->15->11 [2] -1/-1/-1->15->14 [3] 13/7/-1->15->-1
+gpu3:242:242 [6] NCCL INFO P2P Chunksize set to 131072
+gpu3:243:243 [7] NCCL INFO P2P Chunksize set to 131072
+gpu3:238:238 [2] NCCL INFO comm 0x261e3fa0 rank 10 nRanks 16 nNodes 4 localRanks 4 localRank 2 MNNVL 0
+gpu3:237:237 [1] NCCL INFO comm 0x1717b370 rank 9 nRanks 16 nNodes 4 localRanks 4 localRank 1 MNNVL 0
+gpu3:236:236 [0] NCCL INFO comm 0x2fca9260 rank 8 nRanks 16 nNodes 4 localRanks 4 localRank 0 MNNVL 0
+gpu3:238:238 [2] NCCL INFO Trees [0] 11/-1/-1->10->9 [1] 8/-1/-1->10->9 [2] 11/-1/-1->10->9 [3] 8/-1/-1->10->9
+gpu3:237:237 [1] NCCL INFO Trees [0] 10/-1/-1->9->8 [1] 10/-1/-1->9->11 [2] 10/-1/-1->9->8 [3] 10/-1/-1->9->11
+gpu3:238:238 [2] NCCL INFO P2P Chunksize set to 131072
+gpu3:237:237 [1] NCCL INFO P2P Chunksize set to 131072
+gpu3:236:236 [0] NCCL INFO Trees [0] 9/4/12->8->0 [1] -1/-1/-1->8->10 [2] 9/-1/-1->8->4 [3] -1/-1/-1->8->10
+gpu3:236:236 [0] NCCL INFO P2P Chunksize set to 131072
+gpu3:241:241 [5] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so
+gpu3:241:241 [5] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu3:243:243 [7] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so
+gpu3:242:242 [6] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so
+gpu3:243:243 [7] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu3:242:242 [6] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu3:240:240 [4] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so
+gpu3:240:240 [4] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu3:239:239 [3] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so
+gpu3:239:239 [3] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu3:237:237 [1] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so
+gpu3:238:238 [2] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so
+gpu3:241:1436 [0] NCCL INFO [Proxy Service] Device 5 CPU core 48
+gpu3:237:237 [1] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu3:238:238 [2] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu3:236:236 [0] NCCL INFO PROFILER/Plugin: Could not find: libnccl-profiler.so
+gpu3:236:236 [0] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu3:243:1437 [0] NCCL INFO [Proxy Service] Device 7 CPU core 114
+gpu3:242:1439 [0] NCCL INFO [Proxy Service] Device 6 CPU core 37
+gpu3:241:1440 [0] NCCL INFO [Proxy Service UDS] Device 5 CPU core 61
+gpu3:240:1438 [0] NCCL INFO [Proxy Service] Device 4 CPU core 60
+gpu3:243:1441 [0] NCCL INFO [Proxy Service UDS] Device 7 CPU core 62
+gpu3:242:1444 [0] NCCL INFO [Proxy Service UDS] Device 6 CPU core 114
+gpu3:240:1443 [0] NCCL INFO [Proxy Service UDS] Device 4 CPU core 121
+gpu3:237:1447 [0] NCCL INFO [Proxy Service] Device 1 CPU core 80
+gpu3:237:1448 [0] NCCL INFO [Proxy Service UDS] Device 1 CPU core 81
+gpu3:239:1442 [0] NCCL INFO [Proxy Service] Device 3 CPU core 30
+gpu3:238:1446 [0] NCCL INFO [Proxy Service] Device 2 CPU core 71
+gpu3:239:1445 [0] NCCL INFO [Proxy Service UDS] Device 3 CPU core 31
+gpu3:238:1449 [0] NCCL INFO [Proxy Service UDS] Device 2 CPU core 9
+gpu3:236:1450 [0] NCCL INFO [Proxy Service] Device 0 CPU core 8
+gpu3:236:1451 [0] NCCL INFO [Proxy Service UDS] Device 0 CPU core 10
+gpu3:242:242 [6] NCCL INFO Channel 00/0 : 14[6] -> 15[7] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Channel 00/0 : 10[2] -> 11[3] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 01/0 : 14[6] -> 15[7] via P2P/IPC
+gpu3:241:241 [5] NCCL INFO Channel 00/0 : 13[5] -> 14[6] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 02/0 : 14[6] -> 15[7] via P2P/IPC
+gpu3:237:237 [1] NCCL INFO Channel 00/0 : 9[1] -> 10[2] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Channel 01/0 : 10[2] -> 11[3] via P2P/IPC
+gpu3:241:241 [5] NCCL INFO Channel 01/0 : 13[5] -> 14[6] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 03/0 : 14[6] -> 15[7] via P2P/IPC
+gpu3:241:241 [5] NCCL INFO Channel 02/0 : 13[5] -> 14[6] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Channel 02/0 : 10[2] -> 11[3] via P2P/IPC
+gpu3:237:237 [1] NCCL INFO Channel 01/0 : 9[1] -> 10[2] via P2P/IPC
+gpu3:237:237 [1] NCCL INFO Channel 02/0 : 9[1] -> 10[2] via P2P/IPC
+gpu3:241:241 [5] NCCL INFO Channel 03/0 : 13[5] -> 14[6] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Channel 03/0 : 10[2] -> 11[3] via P2P/IPC
+gpu3:237:237 [1] NCCL INFO Channel 03/0 : 9[1] -> 10[2] via P2P/IPC
+gpu3:243:1452 [0] NCCL INFO [Proxy Progress] Device 7 CPU core 48
+gpu3:243:243 [7] NCCL INFO Channel 00/0 : 15[7] -> 0[0] [send] via NET/IB/1/GDRDMA
+gpu3:243:243 [7] NCCL INFO Channel 01/0 : 15[7] -> 0[0] [send] via NET/IB/1/GDRDMA
+gpu3:243:243 [7] NCCL INFO Channel 02/0 : 15[7] -> 0[0] [send] via NET/IB/1/GDRDMA
+gpu3:243:243 [7] NCCL INFO Channel 03/0 : 15[7] -> 0[0] [send] via NET/IB/1/GDRDMA
+gpu3:240:1453 [0] NCCL INFO [Proxy Progress] Device 4 CPU core 43
+gpu3:240:240 [4] NCCL INFO Channel 00/0 : 11[3] -> 12[4] [receive] via NET/IB/1/GDRDMA
+gpu3:240:240 [4] NCCL INFO Channel 01/0 : 11[3] -> 12[4] [receive] via NET/IB/1/GDRDMA
+gpu3:240:240 [4] NCCL INFO Channel 02/0 : 11[3] -> 12[4] [receive] via NET/IB/1/GDRDMA
+gpu3:240:240 [4] NCCL INFO Channel 03/0 : 11[3] -> 12[4] [receive] via NET/IB/1/GDRDMA
+gpu3:240:240 [4] NCCL INFO Channel 00/0 : 12[4] -> 13[5] via P2P/IPC
+gpu3:240:240 [4] NCCL INFO Channel 01/0 : 12[4] -> 13[5] via P2P/IPC
+gpu3:240:240 [4] NCCL INFO Channel 02/0 : 12[4] -> 13[5] via P2P/IPC
+gpu3:240:240 [4] NCCL INFO Channel 03/0 : 12[4] -> 13[5] via P2P/IPC
+gpu3:239:239 [3] NCCL INFO Channel 00/0 : 11[3] -> 12[4] [send] via NET/IB/0/GDRDMA
+gpu3:239:1454 [0] NCCL INFO [Proxy Progress] Device 3 CPU core 70
+gpu3:239:239 [3] NCCL INFO Channel 01/0 : 11[3] -> 12[4] [send] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 02/0 : 11[3] -> 12[4] [send] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 03/0 : 11[3] -> 12[4] [send] via NET/IB/0/GDRDMA
+gpu3:236:1455 [0] NCCL INFO [Proxy Progress] Device 0 CPU core 11
+gpu3:236:236 [0] NCCL INFO Channel 00/0 : 7[7] -> 8[0] [receive] via NET/IB/0/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 01/0 : 7[7] -> 8[0] [receive] via NET/IB/0/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 02/0 : 7[7] -> 8[0] [receive] via NET/IB/0/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 03/0 : 7[7] -> 8[0] [receive] via NET/IB/0/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 00/0 : 8[0] -> 9[1] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Channel 01/0 : 8[0] -> 9[1] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Channel 02/0 : 8[0] -> 9[1] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Channel 03/0 : 8[0] -> 9[1] via P2P/IPC
+gpu3:239:239 [3] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu3:240:240 [4] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu3:240:240 [4] NCCL INFO Channel 01/0 : 12[4] -> 14[6] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu3:241:241 [5] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu3:237:237 [1] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu3:241:241 [5] NCCL INFO Channel 01/0 : 13[5] -> 15[7] via P2P/IPC
+gpu3:243:243 [7] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu3:242:242 [6] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu3:237:237 [1] NCCL INFO Channel 01/0 : 9[1] -> 11[3] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu3:236:236 [0] NCCL INFO Channel 01/0 : 8[0] -> 10[2] via P2P/IPC
+gpu3:240:240 [4] NCCL INFO Channel 03/0 : 12[4] -> 14[6] via P2P/IPC
+gpu3:237:237 [1] NCCL INFO Channel 03/0 : 9[1] -> 11[3] via P2P/IPC
+gpu3:241:241 [5] NCCL INFO Channel 03/0 : 13[5] -> 15[7] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Channel 03/0 : 8[0] -> 10[2] via P2P/IPC
+gpu3:240:240 [4] NCCL INFO Channel 00/0 : 8[0] -> 12[4] [receive] via NET/IB/1/GDRDMA
+gpu3:238:238 [2] NCCL INFO Channel 01/0 : 10[2] -> 8[0] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Channel 00/0 : 4[4] -> 8[0] [receive] via NET/IB/0/GDRDMA
+gpu3:242:242 [6] NCCL INFO Channel 01/0 : 14[6] -> 12[4] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Channel 02/0 : 4[4] -> 8[0] [receive] via NET/IB/0/GDRDMA
+gpu3:238:238 [2] NCCL INFO Channel 03/0 : 10[2] -> 8[0] via P2P/IPC
+gpu3:243:243 [7] NCCL INFO Channel 01/0 : 11[3] -> 15[7] [receive] via NET/IB/1/GDRDMA
+gpu3:242:242 [6] NCCL INFO Channel 03/0 : 14[6] -> 12[4] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Channel 00/0 : 8[0] -> 12[4] [send] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 01/0 : 7[7] -> 11[3] [receive] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 03/0 : 7[7] -> 11[3] [receive] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 01/0 : 11[3] -> 15[7] [send] via NET/IB/0/GDRDMA
+gpu3:240:240 [4] NCCL INFO Channel 02/0 : 4[4] -> 12[4] [receive] via NET/IB/1/GDRDMA
+gpu3:240:240 [4] NCCL INFO Channel 02/0 : 12[4] -> 4[4] [send] via NET/IB/1/GDRDMA
+gpu3:243:243 [7] NCCL INFO Channel 03/0 : 7[7] -> 15[7] [receive] via NET/IB/1/GDRDMA
+gpu3:243:243 [7] NCCL INFO Channel 03/0 : 15[7] -> 7[7] [send] via NET/IB/1/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 01/0 : 3[3] -> 11[3] [receive] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 01/0 : 11[3] -> 3[3] [send] via NET/IB/0/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 00/0 : 0[0] -> 8[0] [receive] via NET/IB/0/GDRDMA
+gpu3:243:243 [7] NCCL INFO Channel 01/0 : 15[7] -> 11[3] [send] via NET/IB/1/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 00/0 : 8[0] -> 0[0] [send] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 01/0 : 15[7] -> 11[3] [receive] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 01/0 : 11[3] -> 7[7] [send] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 03/0 : 11[3] -> 7[7] [send] via NET/IB/0/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 00/0 : 12[4] -> 8[0] [receive] via NET/IB/0/GDRDMA
+gpu3:243:243 [7] NCCL INFO Channel 01/0 : 15[7] -> 13[5] via P2P/IPC
+gpu3:240:240 [4] NCCL INFO Channel 00/0 : 12[4] -> 8[0] [send] via NET/IB/1/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 00/0 : 8[0] -> 4[4] [send] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 01/0 : 11[3] -> 9[1] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Channel 02/0 : 8[0] -> 4[4] [send] via NET/IB/0/GDRDMA
+gpu3:243:243 [7] NCCL INFO Channel 03/0 : 15[7] -> 13[5] via P2P/IPC
+gpu3:239:239 [3] NCCL INFO Channel 03/0 : 11[3] -> 9[1] via P2P/IPC
+gpu3:243:243 [7] NCCL INFO Channel 00/0 : 15[7] -> 14[6] via P2P/IPC
+gpu3:243:243 [7] NCCL INFO Channel 02/0 : 15[7] -> 14[6] via P2P/IPC
+gpu3:239:239 [3] NCCL INFO Channel 00/0 : 11[3] -> 10[2] via P2P/IPC
+gpu3:239:239 [3] NCCL INFO Channel 02/0 : 11[3] -> 10[2] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 00/0 : 14[6] -> 13[5] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 01/0 : 14[6] -> 13[5] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 02/0 : 14[6] -> 13[5] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 03/0 : 14[6] -> 13[5] via P2P/IPC
+gpu3:241:241 [5] NCCL INFO Channel 00/0 : 13[5] -> 12[4] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Channel 00/0 : 10[2] -> 9[1] via P2P/IPC
+gpu3:237:237 [1] NCCL INFO Channel 00/0 : 9[1] -> 8[0] via P2P/IPC
+gpu3:241:241 [5] NCCL INFO Channel 02/0 : 13[5] -> 12[4] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Channel 01/0 : 10[2] -> 9[1] via P2P/IPC
+gpu3:237:237 [1] NCCL INFO Channel 02/0 : 9[1] -> 8[0] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Channel 02/0 : 10[2] -> 9[1] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Channel 03/0 : 10[2] -> 9[1] via P2P/IPC
+gpu3:243:243 [7] NCCL INFO Connected all trees
+gpu3:240:240 [4] NCCL INFO Connected all trees
+gpu3:239:239 [3] NCCL INFO Connected all trees
+gpu3:236:236 [0] NCCL INFO Connected all trees
+gpu3:241:241 [5] NCCL INFO Connected all trees
+gpu3:242:242 [6] NCCL INFO Connected all trees
+gpu3:237:237 [1] NCCL INFO Connected all trees
+gpu3:238:238 [2] NCCL INFO Connected all trees
+gpu3:239:239 [3] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so
+gpu3:243:243 [7] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so
+gpu3:239:239 [3] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu3:243:243 [7] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu3:239:239 [3] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu3:243:243 [7] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu3:240:240 [4] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so
+gpu3:240:240 [4] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu3:240:240 [4] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu3:236:236 [0] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so
+gpu3:236:236 [0] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu3:236:236 [0] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu3:242:1456 [0] NCCL INFO [Proxy Progress] Device 6 CPU core 102
+gpu3:242:242 [6] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so
+gpu3:242:242 [6] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu3:242:242 [6] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu3:237:1457 [0] NCCL INFO [Proxy Progress] Device 1 CPU core 82
+gpu3:237:237 [1] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so
+gpu3:237:237 [1] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu3:237:237 [1] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu3:241:1458 [0] NCCL INFO [Proxy Progress] Device 5 CPU core 115
+gpu3:238:1459 [0] NCCL INFO [Proxy Progress] Device 2 CPU core 76
+gpu3:241:241 [5] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so
+gpu3:241:241 [5] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu3:241:241 [5] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu3:238:238 [2] NCCL INFO TUNER/Plugin: Could not find: libnccl-tuner.so
+gpu3:238:238 [2] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu3:238:238 [2] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu3:241:241 [5] NCCL INFO ncclCommInitRank comm 0x52be2940 rank 13 nranks 16 cudaDev 5 nvmlDev 5 busId ab000 commId 0x92ed644083b4525 - Init COMPLETE
+gpu3:241:241 [5] NCCL INFO Init timings - ncclCommInitRank: rank 13 nranks 16 total 0.74 (kernels 0.31, alloc 0.03, bootstrap 0.13, allgathers 0.02, topo 0.02, graphs 0.00, connections 0.22, rest 0.00)
+gpu3:238:238 [2] NCCL INFO ncclCommInitRank comm 0x261e3fa0 rank 10 nranks 16 cudaDev 2 nvmlDev 2 busId 1b000 commId 0x92ed644083b4525 - Init COMPLETE
+gpu3:238:238 [2] NCCL INFO Init timings - ncclCommInitRank: rank 10 nranks 16 total 0.74 (kernels 0.31, alloc 0.03, bootstrap 0.14, allgathers 0.02, topo 0.03, graphs 0.00, connections 0.22, rest 0.00)
+gpu3:243:243 [7] NCCL INFO ncclCommInitRank comm 0x420ab990 rank 15 nranks 16 cudaDev 7 nvmlDev 7 busId af000 commId 0x92ed644083b4525 - Init COMPLETE
+gpu3:243:243 [7] NCCL INFO Init timings - ncclCommInitRank: rank 15 nranks 16 total 0.71 (kernels 0.30, alloc 0.01, bootstrap 0.13, allgathers 0.02, topo 0.02, graphs 0.00, connections 0.17, rest 0.05)
+gpu3:237:237 [1] NCCL INFO ncclCommInitRank comm 0x1717b370 rank 9 nranks 16 cudaDev 1 nvmlDev 1 busId 1a000 commId 0x92ed644083b4525 - Init COMPLETE
+gpu3:239:239 [3] NCCL INFO ncclCommInitRank comm 0x26b8f240 rank 11 nranks 16 cudaDev 3 nvmlDev 3 busId 1d000 commId 0x92ed644083b4525 - Init COMPLETE
+gpu3:237:237 [1] NCCL INFO Init timings - ncclCommInitRank: rank 9 nranks 16 total 0.74 (kernels 0.31, alloc 0.02, bootstrap 0.15, allgathers 0.02, topo 0.02, graphs 0.00, connections 0.22, rest 0.00)
+gpu3:239:239 [3] NCCL INFO Init timings - ncclCommInitRank: rank 11 nranks 16 total 0.74 (kernels 0.30, alloc 0.03, bootstrap 0.14, allgathers 0.02, topo 0.03, graphs 0.00, connections 0.17, rest 0.05)
+gpu3:236:236 [0] NCCL INFO ncclCommInitRank comm 0x2fca9260 rank 8 nranks 16 cudaDev 0 nvmlDev 0 busId 19000 commId 0x92ed644083b4525 - Init COMPLETE
+gpu3:236:236 [0] NCCL INFO Init timings - ncclCommInitRank: rank 8 nranks 16 total 0.74 (kernels 0.30, alloc 0.04, bootstrap 0.14, allgathers 0.02, topo 0.02, graphs 0.00, connections 0.17, rest 0.05)
+gpu3:242:242 [6] NCCL INFO ncclCommInitRank comm 0x2e169bd0 rank 14 nranks 16 cudaDev 6 nvmlDev 6 busId ae000 commId 0x92ed644083b4525 - Init COMPLETE
+gpu3:240:240 [4] NCCL INFO ncclCommInitRank comm 0x1e210850 rank 12 nranks 16 cudaDev 4 nvmlDev 4 busId aa000 commId 0x92ed644083b4525 - Init COMPLETE
+gpu3:242:242 [6] NCCL INFO Init timings - ncclCommInitRank: rank 14 nranks 16 total 0.71 (kernels 0.33, alloc 0.06, bootstrap 0.06, allgathers 0.02, topo 0.03, graphs 0.00, connections 0.19, rest 0.03)
+gpu3:240:240 [4] NCCL INFO Init timings - ncclCommInitRank: rank 12 nranks 16 total 0.71 (kernels 0.32, alloc 0.01, bootstrap 0.11, allgathers 0.02, topo 0.03, graphs 0.00, connections 0.17, rest 0.05)
+[2026-07-30 14:10:48 TP15 EP1] CustomAllreduce is disabled because this process group spans across nodes.
+[2026-07-30 14:10:48 TP14 EP1] CustomAllreduce is disabled because this process group spans across nodes.
+[2026-07-30 14:10:48 TP13 EP1] CustomAllreduce is disabled because this process group spans across nodes.
+[2026-07-30 14:10:48 TP12 EP1] CustomAllreduce is disabled because this process group spans across nodes.
+[2026-07-30 14:10:48 TP11 EP1] CustomAllreduce is disabled because this process group spans across nodes.
+[2026-07-30 14:10:48 TP10 EP1] CustomAllreduce is disabled because this process group spans across nodes.
+[2026-07-30 14:10:48 TP9 EP1] CustomAllreduce is disabled because this process group spans across nodes.
+[2026-07-30 14:10:48 TP8 EP1] CustomAllreduce is disabled because this process group spans across nodes.
+[2026-07-30 14:10:48 TP15 EP1] Init torch distributed ends. elapsed=15.56 s, mem usage=0.44 GB
+[2026-07-30 14:10:48 TP11 EP1] Init torch distributed ends. elapsed=15.20 s, mem usage=0.48 GB
+[2026-07-30 14:10:48 TP12 EP1] Init torch distributed ends. elapsed=15.49 s, mem usage=0.42 GB
+[2026-07-30 14:10:48 TP14 EP1] Init torch distributed ends. elapsed=15.58 s, mem usage=0.39 GB
+[2026-07-30 14:10:48 TP13 EP1] Init torch distributed ends. elapsed=15.49 s, mem usage=0.41 GB
+[2026-07-30 14:10:48 TP8 EP1] Init torch distributed ends. elapsed=15.34 s, mem usage=0.46 GB
+[2026-07-30 14:10:48 TP10 EP1] Init torch distributed ends. elapsed=15.23 s, mem usage=0.39 GB
+[2026-07-30 14:10:48 TP9 EP1] Init torch distributed ends. elapsed=16.03 s, mem usage=0.41 GB
+[2026-07-30 14:10:51 TP10 EP1] Load weight begin. avail mem=82.11 GB
+[2026-07-30 14:10:51 TP13 EP1] Load weight begin. avail mem=82.10 GB
+[2026-07-30 14:10:51 TP12 EP1] Load weight begin. avail mem=82.08 GB
+[2026-07-30 14:10:51 TP14 EP1] Load weight begin. avail mem=82.11 GB
+[2026-07-30 14:10:51 TP11 EP1] Load weight begin. avail mem=82.03 GB
+[2026-07-30 14:10:51 TP8 EP1] Load weight begin. avail mem=82.04 GB
+[2026-07-30 14:10:51 TP9 EP1] Load weight begin. avail mem=82.10 GB
+[2026-07-30 14:10:51 TP15 EP1] Load weight begin. avail mem=82.07 GB
+[2026-07-30 14:10:51 TP10 EP1] FlashInfer TRTLLM MoE deferred finalize is disabled (moe_runner_backend=flashinfer_mxfp4, quant_method=Mxfp4FlashinferCutlassMoEMethod).
+[2026-07-30 14:10:51 TP13 EP1] FlashInfer TRTLLM MoE deferred finalize is disabled (moe_runner_backend=flashinfer_mxfp4, quant_method=Mxfp4FlashinferCutlassMoEMethod).
+[2026-07-30 14:10:51 TP12 EP1] FlashInfer TRTLLM MoE deferred finalize is disabled (moe_runner_backend=flashinfer_mxfp4, quant_method=Mxfp4FlashinferCutlassMoEMethod).
+[2026-07-30 14:10:51 TP14 EP1] FlashInfer TRTLLM MoE deferred finalize is disabled (moe_runner_backend=flashinfer_mxfp4, quant_method=Mxfp4FlashinferCutlassMoEMethod).
+[2026-07-30 14:10:51 TP11 EP1] FlashInfer TRTLLM MoE deferred finalize is disabled (moe_runner_backend=flashinfer_mxfp4, quant_method=Mxfp4FlashinferCutlassMoEMethod).
+[2026-07-30 14:10:51 TP8 EP1] FlashInfer TRTLLM MoE deferred finalize is disabled (moe_runner_backend=flashinfer_mxfp4, quant_method=Mxfp4FlashinferCutlassMoEMethod).
+[2026-07-30 14:10:51 TP9 EP1] FlashInfer TRTLLM MoE deferred finalize is disabled (moe_runner_backend=flashinfer_mxfp4, quant_method=Mxfp4FlashinferCutlassMoEMethod).
+[2026-07-30 14:10:51 TP15 EP1] FlashInfer TRTLLM MoE deferred finalize is disabled (moe_runner_backend=flashinfer_mxfp4, quant_method=Mxfp4FlashinferCutlassMoEMethod).
+[2026-07-30 14:10:57 TP15 EP1] multimem all-gather disabled because the TP group spans across nodes.
+[2026-07-30 14:10:57 TP13 EP1] multimem all-gather disabled because the TP group spans across nodes.
+[2026-07-30 14:10:57 TP14 EP1] multimem all-gather disabled because the TP group spans across nodes.
+[2026-07-30 14:10:57 TP12 EP1] multimem all-gather disabled because the TP group spans across nodes.
+[2026-07-30 14:10:57 TP11 EP1] multimem all-gather disabled because the TP group spans across nodes.
+[2026-07-30 14:10:57 TP10 EP1] multimem all-gather disabled because the TP group spans across nodes.
+[2026-07-30 14:10:57 TP8 EP1] multimem all-gather disabled because the TP group spans across nodes.
+[2026-07-30 14:10:57 TP9 EP1] multimem all-gather disabled because the TP group spans across nodes.
+[2026-07-30 14:11:12 TP10 EP1] Execute dequant fp8 wo_a
+[2026-07-30 14:11:13 TP15 EP1] Execute dequant fp8 wo_a
+[2026-07-30 14:11:14 TP9 EP1] Execute dequant fp8 wo_a
+[2026-07-30 14:11:14 TP8 EP1] Execute dequant fp8 wo_a
+[2026-07-30 14:11:14 TP11 EP1] Execute dequant fp8 wo_a
+[2026-07-30 14:11:14 TP13 EP1] Execute dequant fp8 wo_a
+[2026-07-30 14:11:14 TP12 EP1] Execute dequant fp8 wo_a
+[2026-07-30 14:11:15 TP14 EP1] Execute dequant fp8 wo_a
+[2026-07-30 14:12:30 TP10 EP1] Using FP8 KV cache but no scaling factors provided. Defaulting to scaling factors of 1.0. This may lead to less accurate results!
+[2026-07-30 14:12:30 TP10 EP1] Load weight end. elapsed=98.93 s, type=DeepseekV4ForCausalLM, quant=fp8, fmt=e4m3, avail mem=25.24 GB, mem usage=56.87 GB.
+[2026-07-30 14:12:33 TP15 EP1] Using FP8 KV cache but no scaling factors provided. Defaulting to scaling factors of 1.0. This may lead to less accurate results!
+[2026-07-30 14:12:33 TP15 EP1] Load weight end. elapsed=101.92 s, type=DeepseekV4ForCausalLM, quant=fp8, fmt=e4m3, avail mem=25.19 GB, mem usage=56.87 GB.
+[2026-07-30 14:12:34 TP9 EP1] Using FP8 KV cache but no scaling factors provided. Defaulting to scaling factors of 1.0. This may lead to less accurate results!
+[2026-07-30 14:12:34 TP9 EP1] Load weight end. elapsed=103.00 s, type=DeepseekV4ForCausalLM, quant=fp8, fmt=e4m3, avail mem=25.23 GB, mem usage=56.87 GB.
+[2026-07-30 14:12:34 TP13 EP1] Using FP8 KV cache but no scaling factors provided. Defaulting to scaling factors of 1.0. This may lead to less accurate results!
+[2026-07-30 14:12:34 TP13 EP1] Load weight end. elapsed=103.30 s, type=DeepseekV4ForCausalLM, quant=fp8, fmt=e4m3, avail mem=25.23 GB, mem usage=56.87 GB.
+[2026-07-30 14:12:35 TP8 EP1] Using FP8 KV cache but no scaling factors provided. Defaulting to scaling factors of 1.0. This may lead to less accurate results!
+[2026-07-30 14:12:35 TP8 EP1] Load weight end. elapsed=103.94 s, type=DeepseekV4ForCausalLM, quant=fp8, fmt=e4m3, avail mem=25.17 GB, mem usage=56.87 GB.
+[2026-07-30 14:12:37 TP12 EP1] Using FP8 KV cache but no scaling factors provided. Defaulting to scaling factors of 1.0. This may lead to less accurate results!
+[2026-07-30 14:12:37 TP12 EP1] Load weight end. elapsed=105.58 s, type=DeepseekV4ForCausalLM, quant=fp8, fmt=e4m3, avail mem=25.21 GB, mem usage=56.87 GB.
+[2026-07-30 14:12:37 TP14 EP1] Using FP8 KV cache but no scaling factors provided. Defaulting to scaling factors of 1.0. This may lead to less accurate results!
+[2026-07-30 14:12:37 TP14 EP1] Load weight end. elapsed=106.39 s, type=DeepseekV4ForCausalLM, quant=fp8, fmt=e4m3, avail mem=25.24 GB, mem usage=56.87 GB.
+[2026-07-30 14:12:38 TP11 EP1] Using FP8 KV cache but no scaling factors provided. Defaulting to scaling factors of 1.0. This may lead to less accurate results!
+[2026-07-30 14:12:38 TP11 EP1] Load weight end. elapsed=107.41 s, type=DeepseekV4ForCausalLM, quant=fp8, fmt=e4m3, avail mem=25.16 GB, mem usage=56.87 GB.
+[2026-07-30 14:13:08 TP10 EP1] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:13:08 TP9 EP1] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:13:09 TP12 EP1] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:13:09 TP13 EP1] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:13:09 TP8 EP1] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:13:09 TP15 EP1] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:13:09 TP14 EP1] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:13:09 TP11 EP1] Tokenizer for /data/hf_models/DeepSeek-V4-Pro is still TokenizersBackend after retries with --trust-remote-code. Model-specific tokenizer attributes may be missing.
+[2026-07-30 14:13:09 TP15 EP1] DSV4 memory calculation: bytes_per_full_token=10993.84, available_bytes=16.93 GB, c128_state_fixed=3.92 GB, full_token=1270016
+[2026-07-30 14:13:09 TP15 EP1] DSV4 pool sizes: full=1270016, swa=126976, c4=317504, c128=9922, c4_state=7936, c128_state=0
+[2026-07-30 14:13:09 TP14 EP1] DSV4 memory calculation: bytes_per_full_token=10993.84, available_bytes=16.93 GB, c128_state_fixed=3.92 GB, full_token=1270016
+[2026-07-30 14:13:09 TP11 EP1] DSV4 memory calculation: bytes_per_full_token=10993.84, available_bytes=16.93 GB, c128_state_fixed=3.92 GB, full_token=1270016
+[2026-07-30 14:13:09 TP14 EP1] DSV4 pool sizes: full=1270016, swa=126976, c4=317504, c128=9922, c4_state=7936, c128_state=0
+[2026-07-30 14:13:09 TP11 EP1] DSV4 pool sizes: full=1270016, swa=126976, c4=317504, c128=9922, c4_state=7936, c128_state=0
+[2026-07-30 14:13:09 TP10 EP1] DSV4 memory calculation: bytes_per_full_token=10993.84, available_bytes=16.93 GB, c128_state_fixed=3.92 GB, full_token=1270016
+[2026-07-30 14:13:09 TP12 EP1] DSV4 memory calculation: bytes_per_full_token=10993.84, available_bytes=16.93 GB, c128_state_fixed=3.92 GB, full_token=1270016
+[2026-07-30 14:13:09 TP10 EP1] DSV4 pool sizes: full=1270016, swa=126976, c4=317504, c128=9922, c4_state=7936, c128_state=0
+[2026-07-30 14:13:09 TP9 EP1] DSV4 memory calculation: bytes_per_full_token=10993.84, available_bytes=16.93 GB, c128_state_fixed=3.92 GB, full_token=1270016
+[2026-07-30 14:13:09 TP9 EP1] DSV4 pool sizes: full=1270016, swa=126976, c4=317504, c128=9922, c4_state=7936, c128_state=0
+[2026-07-30 14:13:09 TP13 EP1] DSV4 memory calculation: bytes_per_full_token=10993.84, available_bytes=16.93 GB, c128_state_fixed=3.92 GB, full_token=1270016
+[2026-07-30 14:13:09 TP12 EP1] DSV4 pool sizes: full=1270016, swa=126976, c4=317504, c128=9922, c4_state=7936, c128_state=0
+[2026-07-30 14:13:09 TP13 EP1] DSV4 pool sizes: full=1270016, swa=126976, c4=317504, c128=9922, c4_state=7936, c128_state=0
+[2026-07-30 14:13:09 TP8 EP1] DSV4 memory calculation: bytes_per_full_token=10993.84, available_bytes=16.93 GB, c128_state_fixed=3.92 GB, full_token=1270016
+[2026-07-30 14:13:09 TP8 EP1] DSV4 pool sizes: full=1270016, swa=126976, c4=317504, c128=9922, c4_state=7936, c128_state=0
+[2026-07-30 14:13:09 TP10 EP1] Initialize DeepSeekV4TokenToKVPool with max_num_reqs=256 swa_size=126976 c4_size=317504 c4_logical_size=317504 c128_size=9922 c4_state_pool_size=7936 c128_state_pool_size=32896
+[2026-07-30 14:13:09 TP9 EP1] Initialize DeepSeekV4TokenToKVPool with max_num_reqs=256 swa_size=126976 c4_size=317504 c4_logical_size=317504 c128_size=9922 c4_state_pool_size=7936 c128_state_pool_size=32896
+[2026-07-30 14:13:09 TP12 EP1] Initialize DeepSeekV4TokenToKVPool with max_num_reqs=256 swa_size=126976 c4_size=317504 c4_logical_size=317504 c128_size=9922 c4_state_pool_size=7936 c128_state_pool_size=32896
+[2026-07-30 14:13:09 TP11 EP1] Initialize DeepSeekV4TokenToKVPool with max_num_reqs=256 swa_size=126976 c4_size=317504 c4_logical_size=317504 c128_size=9922 c4_state_pool_size=7936 c128_state_pool_size=32896
+[2026-07-30 14:13:09 TP13 EP1] Initialize DeepSeekV4TokenToKVPool with max_num_reqs=256 swa_size=126976 c4_size=317504 c4_logical_size=317504 c128_size=9922 c4_state_pool_size=7936 c128_state_pool_size=32896
+[2026-07-30 14:13:09 TP8 EP1] Initialize DeepSeekV4TokenToKVPool with max_num_reqs=256 swa_size=126976 c4_size=317504 c4_logical_size=317504 c128_size=9922 c4_state_pool_size=7936 c128_state_pool_size=32896
+[2026-07-30 14:13:09 TP15 EP1] Initialize DeepSeekV4TokenToKVPool with max_num_reqs=256 swa_size=126976 c4_size=317504 c4_logical_size=317504 c128_size=9922 c4_state_pool_size=7936 c128_state_pool_size=32896
+[2026-07-30 14:13:09 TP14 EP1] Initialize DeepSeekV4TokenToKVPool with max_num_reqs=256 swa_size=126976 c4_size=317504 c4_logical_size=317504 c128_size=9922 c4_state_pool_size=7936 c128_state_pool_size=32896
+[2026-07-30 14:13:11 TP10 EP1] Memory pool end. avail mem=7.18 GB
+[2026-07-30 14:13:11 TP9 EP1] Memory pool end. avail mem=7.16 GB
+[2026-07-30 14:13:11 TP8 EP1] Memory pool end. avail mem=7.10 GB
+[2026-07-30 14:13:11 TP13 EP1] Memory pool end. avail mem=7.16 GB
+[2026-07-30 14:13:11 TP14 EP1] Memory pool end. avail mem=7.18 GB
+[2026-07-30 14:13:11 TP11 EP1] Memory pool end. avail mem=7.09 GB
+[2026-07-30 14:13:11 TP12 EP1] Memory pool end. avail mem=7.14 GB
+[2026-07-30 14:13:11 TP15 EP1] Memory pool end. avail mem=7.13 GB
+[2026-07-30 14:13:11 TP10 EP1] Using DeepseekV4AttnBackend for dsv4 attention backend (CUDA).
+[2026-07-30 14:13:11 TP9 EP1] Using DeepseekV4AttnBackend for dsv4 attention backend (CUDA).
+[2026-07-30 14:13:11 TP10 EP1] Running FlashInfer autotune with cache: /root/.cache/sglang/flashinfer/autotune/0.6.14/sm120/4f24e70399961c55/rank_tp10_pp0_dp0.json
+[2026-07-30 14:13:11 TP9 EP1] Running FlashInfer autotune with cache: /root/.cache/sglang/flashinfer/autotune/0.6.14/sm120/4f24e70399961c55/rank_tp9_pp0_dp0.json
+[2026-07-30 14:13:11 TP13 EP1] Using DeepseekV4AttnBackend for dsv4 attention backend (CUDA).
+[2026-07-30 14:13:11 TP13 EP1] Running FlashInfer autotune with cache: /root/.cache/sglang/flashinfer/autotune/0.6.14/sm120/4f24e70399961c55/rank_tp13_pp0_dp0.json
+[2026-07-30 14:13:11 TP8 EP1] Using DeepseekV4AttnBackend for dsv4 attention backend (CUDA).
+[2026-07-30 14:13:11 TP8 EP1] Running FlashInfer autotune with cache: /root/.cache/sglang/flashinfer/autotune/0.6.14/sm120/4f24e70399961c55/rank_tp8_pp0_dp0.json
+[2026-07-30 14:13:11 TP11 EP1] Using DeepseekV4AttnBackend for dsv4 attention backend (CUDA).
+[2026-07-30 14:13:11 TP14 EP1] Using DeepseekV4AttnBackend for dsv4 attention backend (CUDA).
+[2026-07-30 14:13:11 TP11 EP1] Running FlashInfer autotune with cache: /root/.cache/sglang/flashinfer/autotune/0.6.14/sm120/4f24e70399961c55/rank_tp11_pp0_dp0.json
+[2026-07-30 14:13:11 TP14 EP1] Running FlashInfer autotune with cache: /root/.cache/sglang/flashinfer/autotune/0.6.14/sm120/4f24e70399961c55/rank_tp14_pp0_dp0.json
+[2026-07-30 14:13:11 TP12 EP1] Using DeepseekV4AttnBackend for dsv4 attention backend (CUDA).
+[2026-07-30 14:13:11 TP12 EP1] Running FlashInfer autotune with cache: /root/.cache/sglang/flashinfer/autotune/0.6.14/sm120/4f24e70399961c55/rank_tp12_pp0_dp0.json
+[2026-07-30 14:13:11 TP15 EP1] Using DeepseekV4AttnBackend for dsv4 attention backend (CUDA).
+[2026-07-30 14:13:11 TP15 EP1] Running FlashInfer autotune with cache: /root/.cache/sglang/flashinfer/autotune/0.6.14/sm120/4f24e70399961c55/rank_tp15_pp0_dp0.json
+gpu3:238:238 [2] NCCL INFO Comm config Blocking set to 1
+gpu3:239:239 [3] NCCL INFO Comm config Blocking set to 1
+gpu3:240:240 [4] NCCL INFO Comm config Blocking set to 1
+gpu3:241:241 [5] NCCL INFO Comm config Blocking set to 1
+gpu3:243:243 [7] NCCL INFO Comm config Blocking set to 1
+gpu3:242:242 [6] NCCL INFO Comm config Blocking set to 1
+gpu3:236:236 [0] NCCL INFO Comm config Blocking set to 1
+gpu3:237:237 [1] NCCL INFO Comm config Blocking set to 1
+gpu3:240:240 [4] NCCL INFO Initialized NET plugin IB
+gpu3:240:240 [4] NCCL INFO Assigned NET plugin IB to comm
+gpu3:240:240 [4] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu3:240:240 [4] NCCL INFO Using network IB
+gpu3:240:240 [4] NCCL INFO DMA-BUF is available on GPU device 4
+gpu3:240:240 [4] NCCL INFO ncclCommInitRankConfig comm 0x4833c320 rank 12 nranks 16 cudaDev 4 nvmlDev 4 busId aa000 commId 0x2f8347829d2566aa - Init START
+gpu3:238:238 [2] NCCL INFO Initialized NET plugin IB
+gpu3:238:238 [2] NCCL INFO Assigned NET plugin IB to comm
+gpu3:238:238 [2] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu3:239:239 [3] NCCL INFO Initialized NET plugin IB
+gpu3:238:238 [2] NCCL INFO Using network IB
+gpu3:239:239 [3] NCCL INFO Assigned NET plugin IB to comm
+gpu3:239:239 [3] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu3:239:239 [3] NCCL INFO Using network IB
+gpu3:241:241 [5] NCCL INFO Initialized NET plugin IB
+gpu3:238:238 [2] NCCL INFO DMA-BUF is available on GPU device 2
+gpu3:241:241 [5] NCCL INFO Assigned NET plugin IB to comm
+gpu3:241:241 [5] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu3:241:241 [5] NCCL INFO Using network IB
+gpu3:239:239 [3] NCCL INFO DMA-BUF is available on GPU device 3
+gpu3:241:241 [5] NCCL INFO DMA-BUF is available on GPU device 5
+gpu3:238:238 [2] NCCL INFO ncclCommInitRankConfig comm 0x4f6715d0 rank 10 nranks 16 cudaDev 2 nvmlDev 2 busId 1b000 commId 0x2f8347829d2566aa - Init START
+gpu3:236:236 [0] NCCL INFO Initialized NET plugin IB
+gpu3:236:236 [0] NCCL INFO Assigned NET plugin IB to comm
+gpu3:242:242 [6] NCCL INFO Initialized NET plugin IB
+gpu3:236:236 [0] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu3:236:236 [0] NCCL INFO Using network IB
+gpu3:242:242 [6] NCCL INFO Assigned NET plugin IB to comm
+gpu3:242:242 [6] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu3:242:242 [6] NCCL INFO Using network IB
+gpu3:236:236 [0] NCCL INFO DMA-BUF is available on GPU device 0
+gpu3:242:242 [6] NCCL INFO DMA-BUF is available on GPU device 6
+gpu3:241:241 [5] NCCL INFO ncclCommInitRankConfig comm 0x7708d890 rank 13 nranks 16 cudaDev 5 nvmlDev 5 busId ab000 commId 0x2f8347829d2566aa - Init START
+gpu3:239:239 [3] NCCL INFO ncclCommInitRankConfig comm 0x48c73380 rank 11 nranks 16 cudaDev 3 nvmlDev 3 busId 1d000 commId 0x2f8347829d2566aa - Init START
+gpu3:242:242 [6] NCCL INFO ncclCommInitRankConfig comm 0x5f758360 rank 14 nranks 16 cudaDev 6 nvmlDev 6 busId ae000 commId 0x2f8347829d2566aa - Init START
+gpu3:236:236 [0] NCCL INFO ncclCommInitRankConfig comm 0x5806f1e0 rank 8 nranks 16 cudaDev 0 nvmlDev 0 busId 19000 commId 0x2f8347829d2566aa - Init START
+gpu3:243:243 [7] NCCL INFO Initialized NET plugin IB
+gpu3:243:243 [7] NCCL INFO Assigned NET plugin IB to comm
+gpu3:243:243 [7] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu3:243:243 [7] NCCL INFO Using network IB
+gpu3:243:243 [7] NCCL INFO DMA-BUF is available on GPU device 7
+gpu3:243:243 [7] NCCL INFO ncclCommInitRankConfig comm 0x68dcd920 rank 15 nranks 16 cudaDev 7 nvmlDev 7 busId af000 commId 0x2f8347829d2566aa - Init START
+gpu3:237:237 [1] NCCL INFO Initialized NET plugin IB
+gpu3:237:237 [1] NCCL INFO Assigned NET plugin IB to comm
+gpu3:237:237 [1] NCCL INFO Assigned GIN plugin GIN_IB_GDAKI to comm
+gpu3:237:237 [1] NCCL INFO Using network IB
+gpu3:237:237 [1] NCCL INFO DMA-BUF is available on GPU device 1
+gpu3:237:237 [1] NCCL INFO ncclCommInitRankConfig comm 0x3b6c1b90 rank 9 nranks 16 cudaDev 1 nvmlDev 1 busId 1a000 commId 0x2f8347829d2566aa - Init START
+gpu3:236:236 [0] NCCL INFO Bootstrap timings total 0.018572 (create 0.000054, send 0.000208, recv 0.002101, ring 0.015113, delay 0.000000)
+gpu3:237:237 [1] NCCL INFO Bootstrap timings total 0.016925 (create 0.000055, send 0.000226, recv 0.000591, ring 0.015725, delay 0.000000)
+gpu3:239:239 [3] NCCL INFO Bootstrap timings total 0.018696 (create 0.000050, send 0.000275, recv 0.000679, ring 0.017365, delay 0.000000)
+gpu3:238:238 [2] NCCL INFO Bootstrap timings total 0.018889 (create 0.000058, send 0.000215, recv 0.000778, ring 0.015779, delay 0.000000)
+gpu3:241:241 [5] NCCL INFO Bootstrap timings total 0.018813 (create 0.000053, send 0.000311, recv 0.000930, ring 0.017141, delay 0.000000)
+gpu3:240:240 [4] NCCL INFO Bootstrap timings total 0.019323 (create 0.000041, send 0.000216, recv 0.001606, ring 0.017092, delay 0.000000)
+gpu3:243:243 [7] NCCL INFO Bootstrap timings total 0.017958 (create 0.000056, send 0.000257, recv 0.000576, ring 0.016703, delay 0.000000)
+gpu3:242:242 [6] NCCL INFO Bootstrap timings total 0.018714 (create 0.000054, send 0.000264, recv 0.001074, ring 0.016982, delay 0.000000)
+gpu3:238:238 [2] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 2 is 0-31,64-95. (GPU affinity = 0-31,64-95 ; CPU affinity = 0-127).
+gpu3:239:239 [3] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 3 is 0-31,64-95. (GPU affinity = 0-31,64-95 ; CPU affinity = 0-127).
+gpu3:240:240 [4] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 4 is 32-63,96-127. (GPU affinity = 32-63,96-127 ; CPU affinity = 0-127).
+gpu3:237:237 [1] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 1 is 0-31,64-95. (GPU affinity = 0-31,64-95 ; CPU affinity = 0-127).
+gpu3:242:242 [6] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 6 is 32-63,96-127. (GPU affinity = 32-63,96-127 ; CPU affinity = 0-127).
+gpu3:236:236 [0] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 0 is 0-31,64-95. (GPU affinity = 0-31,64-95 ; CPU affinity = 0-127).
+gpu3:241:241 [5] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 5 is 32-63,96-127. (GPU affinity = 32-63,96-127 ; CPU affinity = 0-127).
+gpu3:243:243 [7] NCCL INFO ncclTopoGetCpuAffinity: Affinity for GPU 7 is 32-63,96-127. (GPU affinity = 32-63,96-127 ; CPU affinity = 0-127).
+gpu3:240:240 [4] NCCL INFO comm 0x4833c320 rank 12 nRanks 16 nNodes 4 localRanks 4 localRank 0 MNNVL 0
+gpu3:239:239 [3] NCCL INFO comm 0x48c73380 rank 11 nRanks 16 nNodes 4 localRanks 4 localRank 3 MNNVL 0
+gpu3:241:241 [5] NCCL INFO comm 0x7708d890 rank 13 nRanks 16 nNodes 4 localRanks 4 localRank 1 MNNVL 0
+gpu3:240:240 [4] NCCL INFO Trees [0] 13/-1/-1->12->8 [1] -1/-1/-1->12->14 [2] 13/4/-1->12->-1 [3] -1/-1/-1->12->14
+gpu3:240:240 [4] NCCL INFO P2P Chunksize set to 131072
+gpu3:240:240 [4] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu3:238:238 [2] NCCL INFO comm 0x4f6715d0 rank 10 nRanks 16 nNodes 4 localRanks 4 localRank 2 MNNVL 0
+gpu3:237:237 [1] NCCL INFO comm 0x3b6c1b90 rank 9 nRanks 16 nNodes 4 localRanks 4 localRank 1 MNNVL 0
+gpu3:236:236 [0] NCCL INFO comm 0x5806f1e0 rank 8 nRanks 16 nNodes 4 localRanks 4 localRank 0 MNNVL 0
+gpu3:238:238 [2] NCCL INFO Trees [0] 11/-1/-1->10->9 [1] 8/-1/-1->10->9 [2] 11/-1/-1->10->9 [3] 8/-1/-1->10->9
+gpu3:237:237 [1] NCCL INFO Trees [0] 10/-1/-1->9->8 [1] 10/-1/-1->9->11 [2] 10/-1/-1->9->8 [3] 10/-1/-1->9->11
+gpu3:238:238 [2] NCCL INFO P2P Chunksize set to 131072
+gpu3:237:237 [1] NCCL INFO P2P Chunksize set to 131072
+gpu3:239:239 [3] NCCL INFO Trees [0] -1/-1/-1->11->10 [1] 9/7/15->11->3 [2] -1/-1/-1->11->10 [3] 9/-1/-1->11->7
+gpu3:237:237 [1] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu3:239:239 [3] NCCL INFO P2P Chunksize set to 131072
+gpu3:238:238 [2] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu3:242:242 [6] NCCL INFO comm 0x5f758360 rank 14 nRanks 16 nNodes 4 localRanks 4 localRank 2 MNNVL 0
+gpu3:241:241 [5] NCCL INFO Trees [0] 14/-1/-1->13->12 [1] 14/-1/-1->13->15 [2] 14/-1/-1->13->12 [3] 14/-1/-1->13->15
+gpu3:243:243 [7] NCCL INFO comm 0x68dcd920 rank 15 nRanks 16 nNodes 4 localRanks 4 localRank 3 MNNVL 0
+gpu3:241:241 [5] NCCL INFO P2P Chunksize set to 131072
+gpu3:239:239 [3] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu3:236:236 [0] NCCL INFO Trees [0] 9/4/12->8->0 [1] -1/-1/-1->8->10 [2] 9/-1/-1->8->4 [3] -1/-1/-1->8->10
+gpu3:236:236 [0] NCCL INFO P2P Chunksize set to 131072
+gpu3:241:241 [5] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu3:236:236 [0] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu3:242:242 [6] NCCL INFO Trees [0] 15/-1/-1->14->13 [1] 12/-1/-1->14->13 [2] 15/-1/-1->14->13 [3] 12/-1/-1->14->13
+gpu3:242:242 [6] NCCL INFO P2P Chunksize set to 131072
+gpu3:243:243 [7] NCCL INFO Trees [0] -1/-1/-1->15->14 [1] 13/-1/-1->15->11 [2] -1/-1/-1->15->14 [3] 13/7/-1->15->-1
+gpu3:242:242 [6] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu3:243:243 [7] NCCL INFO P2P Chunksize set to 131072
+gpu3:243:243 [7] NCCL INFO Check P2P Type isAllDirectP2p 1 directMode 0 isAllCudaP2p 1
+gpu3:238:2228 [0] NCCL INFO [Proxy Service] Device 2 CPU core 81
+gpu3:237:2229 [0] NCCL INFO [Proxy Service] Device 1 CPU core 18
+gpu3:238:2235 [0] NCCL INFO [Proxy Service UDS] Device 2 CPU core 19
+gpu3:243:2232 [0] NCCL INFO [Proxy Service] Device 7 CPU core 34
+gpu3:241:2231 [0] NCCL INFO [Proxy Service] Device 5 CPU core 116
+gpu3:242:2234 [0] NCCL INFO [Proxy Service] Device 6 CPU core 120
+gpu3:237:2236 [0] NCCL INFO [Proxy Service UDS] Device 1 CPU core 84
+gpu3:240:2237 [0] NCCL INFO [Proxy Service] Device 4 CPU core 36
+gpu3:243:2240 [0] NCCL INFO [Proxy Service UDS] Device 7 CPU core 101
+gpu3:241:2238 [0] NCCL INFO [Proxy Service UDS] Device 5 CPU core 53
+gpu3:239:2230 [0] NCCL INFO [Proxy Service] Device 3 CPU core 66
+gpu3:242:2241 [0] NCCL INFO [Proxy Service UDS] Device 6 CPU core 121
+gpu3:240:2242 [0] NCCL INFO [Proxy Service UDS] Device 4 CPU core 38
+gpu3:236:2233 [0] NCCL INFO [Proxy Service] Device 0 CPU core 3
+gpu3:239:2239 [0] NCCL INFO [Proxy Service UDS] Device 3 CPU core 68
+gpu3:236:2243 [0] NCCL INFO [Proxy Service UDS] Device 0 CPU core 5
+gpu3:238:238 [2] NCCL INFO Channel 00/0 : 10[2] -> 11[3] via P2P/IPC
+gpu3:237:237 [1] NCCL INFO Channel 00/0 : 9[1] -> 10[2] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Channel 01/0 : 10[2] -> 11[3] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 00/0 : 14[6] -> 15[7] via P2P/IPC
+gpu3:241:241 [5] NCCL INFO Channel 00/0 : 13[5] -> 14[6] via P2P/IPC
+gpu3:237:237 [1] NCCL INFO Channel 01/0 : 9[1] -> 10[2] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Channel 02/0 : 10[2] -> 11[3] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 01/0 : 14[6] -> 15[7] via P2P/IPC
+gpu3:241:241 [5] NCCL INFO Channel 01/0 : 13[5] -> 14[6] via P2P/IPC
+gpu3:241:241 [5] NCCL INFO Channel 02/0 : 13[5] -> 14[6] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 02/0 : 14[6] -> 15[7] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Channel 03/0 : 10[2] -> 11[3] via P2P/IPC
+gpu3:237:237 [1] NCCL INFO Channel 02/0 : 9[1] -> 10[2] via P2P/IPC
+gpu3:237:237 [1] NCCL INFO Channel 03/0 : 9[1] -> 10[2] via P2P/IPC
+gpu3:241:241 [5] NCCL INFO Channel 03/0 : 13[5] -> 14[6] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 03/0 : 14[6] -> 15[7] via P2P/IPC
+gpu3:240:2244 [0] NCCL INFO [Proxy Progress] Device 4 CPU core 103
+gpu3:240:240 [4] NCCL INFO Channel 00/0 : 11[3] -> 12[4] [receive] via NET/IB/1/GDRDMA
+gpu3:240:240 [4] NCCL INFO Channel 01/0 : 11[3] -> 12[4] [receive] via NET/IB/1/GDRDMA
+gpu3:240:240 [4] NCCL INFO Channel 02/0 : 11[3] -> 12[4] [receive] via NET/IB/1/GDRDMA
+gpu3:240:240 [4] NCCL INFO Channel 03/0 : 11[3] -> 12[4] [receive] via NET/IB/1/GDRDMA
+gpu3:240:240 [4] NCCL INFO Channel 00/0 : 12[4] -> 13[5] via P2P/IPC
+gpu3:236:2245 [0] NCCL INFO [Proxy Progress] Device 0 CPU core 7
+gpu3:236:236 [0] NCCL INFO Channel 00/0 : 7[7] -> 8[0] [receive] via NET/IB/0/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 01/0 : 7[7] -> 8[0] [receive] via NET/IB/0/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 02/0 : 7[7] -> 8[0] [receive] via NET/IB/0/GDRDMA
+gpu3:240:240 [4] NCCL INFO Channel 01/0 : 12[4] -> 13[5] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Channel 03/0 : 7[7] -> 8[0] [receive] via NET/IB/0/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 00/0 : 8[0] -> 9[1] via P2P/IPC
+gpu3:240:240 [4] NCCL INFO Channel 02/0 : 12[4] -> 13[5] via P2P/IPC
+gpu3:240:240 [4] NCCL INFO Channel 03/0 : 12[4] -> 13[5] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Channel 01/0 : 8[0] -> 9[1] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Channel 02/0 : 8[0] -> 9[1] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Channel 03/0 : 8[0] -> 9[1] via P2P/IPC
+gpu3:243:2246 [0] NCCL INFO [Proxy Progress] Device 7 CPU core 104
+gpu3:243:243 [7] NCCL INFO Channel 00/0 : 15[7] -> 0[0] [send] via NET/IB/1/GDRDMA
+gpu3:243:243 [7] NCCL INFO Channel 01/0 : 15[7] -> 0[0] [send] via NET/IB/1/GDRDMA
+gpu3:243:243 [7] NCCL INFO Channel 02/0 : 15[7] -> 0[0] [send] via NET/IB/1/GDRDMA
+gpu3:243:243 [7] NCCL INFO Channel 03/0 : 15[7] -> 0[0] [send] via NET/IB/1/GDRDMA
+gpu3:239:2247 [0] NCCL INFO [Proxy Progress] Device 3 CPU core 8
+gpu3:239:239 [3] NCCL INFO Channel 00/0 : 11[3] -> 12[4] [send] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 01/0 : 11[3] -> 12[4] [send] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 02/0 : 11[3] -> 12[4] [send] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 03/0 : 11[3] -> 12[4] [send] via NET/IB/0/GDRDMA
+gpu3:241:241 [5] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu3:240:240 [4] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu3:239:239 [3] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu3:240:240 [4] NCCL INFO Channel 01/0 : 12[4] -> 14[6] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu3:241:241 [5] NCCL INFO Channel 01/0 : 13[5] -> 15[7] via P2P/IPC
+gpu3:237:237 [1] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu3:236:236 [0] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu3:243:243 [7] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu3:242:242 [6] NCCL INFO Connected all rings, use ring PXN 0 GDR 1
+gpu3:237:237 [1] NCCL INFO Channel 01/0 : 9[1] -> 11[3] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Channel 01/0 : 8[0] -> 10[2] via P2P/IPC
+gpu3:241:241 [5] NCCL INFO Channel 03/0 : 13[5] -> 15[7] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Channel 03/0 : 8[0] -> 10[2] via P2P/IPC
+gpu3:237:237 [1] NCCL INFO Channel 03/0 : 9[1] -> 11[3] via P2P/IPC
+gpu3:240:240 [4] NCCL INFO Channel 03/0 : 12[4] -> 14[6] via P2P/IPC
+gpu3:243:243 [7] NCCL INFO Channel 01/0 : 11[3] -> 15[7] [receive] via NET/IB/1/GDRDMA
+gpu3:238:238 [2] NCCL INFO Channel 01/0 : 10[2] -> 8[0] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 01/0 : 14[6] -> 12[4] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Channel 00/0 : 4[4] -> 8[0] [receive] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 01/0 : 7[7] -> 11[3] [receive] via NET/IB/0/GDRDMA
+gpu3:240:240 [4] NCCL INFO Channel 00/0 : 8[0] -> 12[4] [receive] via NET/IB/1/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 02/0 : 4[4] -> 8[0] [receive] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 03/0 : 7[7] -> 11[3] [receive] via NET/IB/0/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 00/0 : 8[0] -> 12[4] [send] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 01/0 : 11[3] -> 15[7] [send] via NET/IB/0/GDRDMA
+gpu3:238:238 [2] NCCL INFO Channel 03/0 : 10[2] -> 8[0] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 03/0 : 14[6] -> 12[4] via P2P/IPC
+gpu3:240:240 [4] NCCL INFO Channel 02/0 : 4[4] -> 12[4] [receive] via NET/IB/1/GDRDMA
+gpu3:243:243 [7] NCCL INFO Channel 03/0 : 7[7] -> 15[7] [receive] via NET/IB/1/GDRDMA
+gpu3:240:240 [4] NCCL INFO Channel 02/0 : 12[4] -> 4[4] [send] via NET/IB/1/GDRDMA
+gpu3:243:243 [7] NCCL INFO Channel 03/0 : 15[7] -> 7[7] [send] via NET/IB/1/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 01/0 : 3[3] -> 11[3] [receive] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 01/0 : 11[3] -> 3[3] [send] via NET/IB/0/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 00/0 : 0[0] -> 8[0] [receive] via NET/IB/0/GDRDMA
+gpu3:243:243 [7] NCCL INFO Channel 01/0 : 15[7] -> 11[3] [send] via NET/IB/1/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 00/0 : 8[0] -> 0[0] [send] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 01/0 : 15[7] -> 11[3] [receive] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 01/0 : 11[3] -> 7[7] [send] via NET/IB/0/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 00/0 : 12[4] -> 8[0] [receive] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 03/0 : 11[3] -> 7[7] [send] via NET/IB/0/GDRDMA
+gpu3:236:236 [0] NCCL INFO Channel 00/0 : 8[0] -> 4[4] [send] via NET/IB/0/GDRDMA
+gpu3:240:240 [4] NCCL INFO Channel 00/0 : 12[4] -> 8[0] [send] via NET/IB/1/GDRDMA
+gpu3:243:243 [7] NCCL INFO Channel 01/0 : 15[7] -> 13[5] via P2P/IPC
+gpu3:236:236 [0] NCCL INFO Channel 02/0 : 8[0] -> 4[4] [send] via NET/IB/0/GDRDMA
+gpu3:239:239 [3] NCCL INFO Channel 01/0 : 11[3] -> 9[1] via P2P/IPC
+gpu3:243:243 [7] NCCL INFO Channel 03/0 : 15[7] -> 13[5] via P2P/IPC
+gpu3:239:239 [3] NCCL INFO Channel 03/0 : 11[3] -> 9[1] via P2P/IPC
+gpu3:243:243 [7] NCCL INFO Channel 00/0 : 15[7] -> 14[6] via P2P/IPC
+gpu3:243:243 [7] NCCL INFO Channel 02/0 : 15[7] -> 14[6] via P2P/IPC
+gpu3:239:239 [3] NCCL INFO Channel 00/0 : 11[3] -> 10[2] via P2P/IPC
+gpu3:239:239 [3] NCCL INFO Channel 02/0 : 11[3] -> 10[2] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Channel 00/0 : 10[2] -> 9[1] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 00/0 : 14[6] -> 13[5] via P2P/IPC
+gpu3:241:241 [5] NCCL INFO Channel 00/0 : 13[5] -> 12[4] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Channel 01/0 : 10[2] -> 9[1] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 01/0 : 14[6] -> 13[5] via P2P/IPC
+gpu3:241:241 [5] NCCL INFO Channel 02/0 : 13[5] -> 12[4] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 02/0 : 14[6] -> 13[5] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Channel 02/0 : 10[2] -> 9[1] via P2P/IPC
+gpu3:237:237 [1] NCCL INFO Channel 00/0 : 9[1] -> 8[0] via P2P/IPC
+gpu3:238:238 [2] NCCL INFO Channel 03/0 : 10[2] -> 9[1] via P2P/IPC
+gpu3:242:242 [6] NCCL INFO Channel 03/0 : 14[6] -> 13[5] via P2P/IPC
+gpu3:237:237 [1] NCCL INFO Channel 02/0 : 9[1] -> 8[0] via P2P/IPC
+gpu3:240:240 [4] NCCL INFO Connected all trees
+gpu3:241:241 [5] NCCL INFO Connected all trees
+gpu3:243:243 [7] NCCL INFO Connected all trees
+gpu3:242:242 [6] NCCL INFO Connected all trees
+gpu3:236:236 [0] NCCL INFO Connected all trees
+gpu3:237:237 [1] NCCL INFO Connected all trees
+gpu3:239:239 [3] NCCL INFO Connected all trees
+gpu3:238:238 [2] NCCL INFO Connected all trees
+gpu3:243:243 [7] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu3:243:243 [7] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu3:240:240 [4] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu3:240:240 [4] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu3:239:239 [3] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu3:239:239 [3] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu3:236:236 [0] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu3:236:236 [0] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu3:242:2248 [0] NCCL INFO [Proxy Progress] Device 6 CPU core 59
+gpu3:241:2249 [0] NCCL INFO [Proxy Progress] Device 5 CPU core 54
+gpu3:238:2250 [0] NCCL INFO [Proxy Progress] Device 2 CPU core 85
+gpu3:237:2251 [0] NCCL INFO [Proxy Progress] Device 1 CPU core 86
+gpu3:242:242 [6] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu3:242:242 [6] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu3:241:241 [5] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu3:241:241 [5] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu3:238:238 [2] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu3:238:238 [2] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu3:237:237 [1] NCCL INFO threadThresholds 8/8/64 | 128/8/64 | 512 | 512
+gpu3:237:237 [1] NCCL INFO 4 coll channels, 4 collnet channels, 0 nvls channels, 4 p2p channels, 1 p2p channels per peer
+gpu3:242:242 [6] NCCL INFO ncclCommInitRankConfig comm 0x5f758360 rank 14 nranks 16 cudaDev 6 nvmlDev 6 busId ae000 commId 0x2f8347829d2566aa - Init COMPLETE
+gpu3:240:240 [4] NCCL INFO ncclCommInitRankConfig comm 0x4833c320 rank 12 nranks 16 cudaDev 4 nvmlDev 4 busId aa000 commId 0x2f8347829d2566aa - Init COMPLETE
+gpu3:242:242 [6] NCCL INFO Init timings - ncclCommInitRankConfig: rank 14 nranks 16 total 0.33 (kernels 0.00, alloc 0.00, bootstrap 0.02, allgathers 0.05, topo 0.04, graphs 0.01, connections 0.21, rest 0.00)
+gpu3:240:240 [4] NCCL INFO Init timings - ncclCommInitRankConfig: rank 12 nranks 16 total 0.33 (kernels 0.00, alloc 0.00, bootstrap 0.02, allgathers 0.05, topo 0.04, graphs 0.01, connections 0.16, rest 0.05)
+gpu3:241:241 [5] NCCL INFO ncclCommInitRankConfig comm 0x7708d890 rank 13 nranks 16 cudaDev 5 nvmlDev 5 busId ab000 commId 0x2f8347829d2566aa - Init COMPLETE
+gpu3:243:243 [7] NCCL INFO ncclCommInitRankConfig comm 0x68dcd920 rank 15 nranks 16 cudaDev 7 nvmlDev 7 busId af000 commId 0x2f8347829d2566aa - Init COMPLETE
+gpu3:241:241 [5] NCCL INFO Init timings - ncclCommInitRankConfig: rank 13 nranks 16 total 0.33 (kernels 0.00, alloc 0.00, bootstrap 0.02, allgathers 0.05, topo 0.05, graphs 0.01, connections 0.21, rest 0.00)
+gpu3:243:243 [7] NCCL INFO Init timings - ncclCommInitRankConfig: rank 15 nranks 16 total 0.33 (kernels 0.00, alloc 0.00, bootstrap 0.02, allgathers 0.05, topo 0.05, graphs 0.01, connections 0.16, rest 0.05)
+gpu3:238:238 [2] NCCL INFO ncclCommInitRankConfig comm 0x4f6715d0 rank 10 nranks 16 cudaDev 2 nvmlDev 2 busId 1b000 commId 0x2f8347829d2566aa - Init COMPLETE
+gpu3:236:236 [0] NCCL INFO ncclCommInitRankConfig comm 0x5806f1e0 rank 8 nranks 16 cudaDev 0 nvmlDev 0 busId 19000 commId 0x2f8347829d2566aa - Init COMPLETE
+gpu3:238:238 [2] NCCL INFO Init timings - ncclCommInitRankConfig: rank 10 nranks 16 total 0.33 (kernels 0.00, alloc 0.00, bootstrap 0.02, allgathers 0.05, topo 0.04, graphs 0.01, connections 0.21, rest 0.00)
+gpu3:239:239 [3] NCCL INFO ncclCommInitRankConfig comm 0x48c73380 rank 11 nranks 16 cudaDev 3 nvmlDev 3 busId 1d000 commId 0x2f8347829d2566aa - Init COMPLETE
+gpu3:236:236 [0] NCCL INFO Init timings - ncclCommInitRankConfig: rank 8 nranks 16 total 0.33 (kernels 0.00, alloc 0.00, bootstrap 0.02, allgathers 0.05, topo 0.05, graphs 0.01, connections 0.16, rest 0.05)
+gpu3:237:237 [1] NCCL INFO ncclCommInitRankConfig comm 0x3b6c1b90 rank 9 nranks 16 cudaDev 1 nvmlDev 1 busId 1a000 commId 0x2f8347829d2566aa - Init COMPLETE
+gpu3:239:239 [3] NCCL INFO Init timings - ncclCommInitRankConfig: rank 11 nranks 16 total 0.33 (kernels 0.00, alloc 0.00, bootstrap 0.02, allgathers 0.05, topo 0.04, graphs 0.01, connections 0.16, rest 0.05)
+gpu3:237:237 [1] NCCL INFO Init timings - ncclCommInitRankConfig: rank 9 nranks 16 total 0.33 (kernels 0.00, alloc 0.00, bootstrap 0.02, allgathers 0.05, topo 0.04, graphs 0.00, connections 0.21, rest 0.00)
+[TensorRT-LLM][INFO] Set logger level to INFO
+[TensorRT-LLM][INFO] Set logger level to INFO
+[TensorRT-LLM][INFO] Set logger level to INFO
+[TensorRT-LLM][INFO] Set logger level to INFO
+[TensorRT-LLM][INFO] Set logger level to INFO
+[TensorRT-LLM][INFO] Set logger level to INFO
+[TensorRT-LLM][INFO] Set logger level to INFO
+[TensorRT-LLM][INFO] Set logger level to INFO
+[2026-07-30 14:13:40 TP11 EP1] FlashInfer autotune completed.
+[2026-07-30 14:13:40 TP11 EP1] Disable prefill CUDA graph because cuda_graph_config resolved prefill.backend='disabled' (e.g. via --cuda-graph-backend-prefill=disabled or auto-disable rules).
+[2026-07-30 14:13:40 TP14 EP1] FlashInfer autotune completed.
+[2026-07-30 14:13:40 TP14 EP1] Disable prefill CUDA graph because cuda_graph_config resolved prefill.backend='disabled' (e.g. via --cuda-graph-backend-prefill=disabled or auto-disable rules).
+[2026-07-30 14:13:40 TP9 EP1] FlashInfer autotune completed.
+[2026-07-30 14:13:40 TP9 EP1] Disable prefill CUDA graph because cuda_graph_config resolved prefill.backend='disabled' (e.g. via --cuda-graph-backend-prefill=disabled or auto-disable rules).
+[2026-07-30 14:13:40 TP10 EP1] FlashInfer autotune completed.
+[2026-07-30 14:13:40 TP10 EP1] Disable prefill CUDA graph because cuda_graph_config resolved prefill.backend='disabled' (e.g. via --cuda-graph-backend-prefill=disabled or auto-disable rules).
+[2026-07-30 14:13:40 TP8 EP1] FlashInfer autotune completed.
+[2026-07-30 14:13:40 TP8 EP1] Disable prefill CUDA graph because cuda_graph_config resolved prefill.backend='disabled' (e.g. via --cuda-graph-backend-prefill=disabled or auto-disable rules).
+[2026-07-30 14:13:40 TP15 EP1] FlashInfer autotune completed.
+[2026-07-30 14:13:40 TP15 EP1] Disable prefill CUDA graph because cuda_graph_config resolved prefill.backend='disabled' (e.g. via --cuda-graph-backend-prefill=disabled or auto-disable rules).
+[2026-07-30 14:13:40 TP13 EP1] FlashInfer autotune completed.
+[2026-07-30 14:13:40 TP13 EP1] Disable prefill CUDA graph because cuda_graph_config resolved prefill.backend='disabled' (e.g. via --cuda-graph-backend-prefill=disabled or auto-disable rules).
+[2026-07-30 14:13:40 TP12 EP1] FlashInfer autotune completed.
+[2026-07-30 14:13:40 TP12 EP1] Disable prefill CUDA graph because cuda_graph_config resolved prefill.backend='disabled' (e.g. via --cuda-graph-backend-prefill=disabled or auto-disable rules).
+[2026-07-30 14:13:40 TP14 EP1] Capture target decode CUDA graph begin. backend=full, num_tokens_per_req=1, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], avail mem=6.86 GB
+[2026-07-30 14:13:40 TP15 EP1] Capture target decode CUDA graph begin. backend=full, num_tokens_per_req=1, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], avail mem=6.77 GB
+[2026-07-30 14:13:40 TP12 EP1] Capture target decode CUDA graph begin. backend=full, num_tokens_per_req=1, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], avail mem=6.80 GB
+[2026-07-30 14:13:40 TP8 EP1] Capture target decode CUDA graph begin. backend=full, num_tokens_per_req=1, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], avail mem=6.72 GB
+[2026-07-30 14:13:40 TP11 EP1] Capture target decode CUDA graph begin. backend=full, num_tokens_per_req=1, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], avail mem=6.69 GB
+[2026-07-30 14:13:40 TP10 EP1] Capture target decode CUDA graph begin. backend=full, num_tokens_per_req=1, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], avail mem=6.86 GB
+[2026-07-30 14:13:40 TP9 EP1] Capture target decode CUDA graph begin. backend=full, num_tokens_per_req=1, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], avail mem=6.83 GB
+[2026-07-30 14:13:40 TP13 EP1] Capture target decode CUDA graph begin. backend=full, num_tokens_per_req=1, bs=[1, 2, 4, 8, 12, 16, 24, 32, 40, 48, 56, 64], avail mem=6.83 GB
+/usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:322: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
+ warnings.warn(
+/usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:322: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
+ warnings.warn(
+/usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:322: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
+ warnings.warn(
+/usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:322: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
+ warnings.warn(
+/usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:322: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
+ warnings.warn(
+/usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:322: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
+ warnings.warn(
+/usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:322: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
+ warnings.warn(
+/usr/local/lib/python3.12/dist-packages/torch/_inductor/compile_fx.py:322: UserWarning: TensorFloat32 tensor cores for float32 matrix multiplication available but not enabled. Consider setting `torch.set_float32_matmul_precision('high')` for better performance.
+ warnings.warn(
+[2026-07-30 14:15:12 TP15 EP1] Capture target decode CUDA graph end. elapsed=92.33 s, mem usage=1.37 GB, avail mem=5.40 GB.
+[2026-07-30 14:15:12 TP15 EP1] Init Unified RadixTree with components (, )
+[2026-07-30 14:15:12 TP15 EP1] Tree cache initialized: source=default impl=UnifiedRadixCache hybrid_swa=True hybrid_ssm=False hierarchical=False streaming_wrapped=False
+[2026-07-30 14:15:12 TP10 EP1] Capture target decode CUDA graph end. elapsed=92.33 s, mem usage=1.37 GB, avail mem=5.50 GB.
+[2026-07-30 14:15:12 TP10 EP1] Init Unified RadixTree with components (, )
+[2026-07-30 14:15:12 TP10 EP1] Tree cache initialized: source=default impl=UnifiedRadixCache hybrid_swa=True hybrid_ssm=False hierarchical=False streaming_wrapped=False
+[2026-07-30 14:15:12 TP9 EP1] Capture target decode CUDA graph end. elapsed=92.34 s, mem usage=1.37 GB, avail mem=5.46 GB.
+[2026-07-30 14:15:12 TP13 EP1] Capture target decode CUDA graph end. elapsed=92.33 s, mem usage=1.37 GB, avail mem=5.46 GB.
+[2026-07-30 14:15:12 TP9 EP1] Init Unified RadixTree with components (, )
+[2026-07-30 14:15:12 TP9 EP1] Tree cache initialized: source=default impl=UnifiedRadixCache hybrid_swa=True hybrid_ssm=False hierarchical=False streaming_wrapped=False
+[2026-07-30 14:15:12 TP13 EP1] Init Unified RadixTree with components (, )
+[2026-07-30 14:15:12 TP13 EP1] Tree cache initialized: source=default impl=UnifiedRadixCache hybrid_swa=True hybrid_ssm=False hierarchical=False streaming_wrapped=False
+[2026-07-30 14:15:12 TP12 EP1] Capture target decode CUDA graph end. elapsed=92.34 s, mem usage=1.37 GB, avail mem=5.43 GB.
+[2026-07-30 14:15:12 TP11 EP1] Capture target decode CUDA graph end. elapsed=92.35 s, mem usage=1.37 GB, avail mem=5.32 GB.
+[2026-07-30 14:15:12 TP12 EP1] Init Unified RadixTree with components (, )
+[2026-07-30 14:15:12 TP12 EP1] Tree cache initialized: source=default impl=UnifiedRadixCache hybrid_swa=True hybrid_ssm=False hierarchical=False streaming_wrapped=False
+[2026-07-30 14:15:12 TP11 EP1] Init Unified RadixTree with components (, )
+[2026-07-30 14:15:12 TP11 EP1] Tree cache initialized: source=default impl=UnifiedRadixCache hybrid_swa=True hybrid_ssm=False hierarchical=False streaming_wrapped=False
+[2026-07-30 14:15:12 TP14 EP1] Capture target decode CUDA graph end. elapsed=92.39 s, mem usage=1.37 GB, avail mem=5.50 GB.
+[2026-07-30 14:15:12 TP14 EP1] Init Unified RadixTree with components (, )
+[2026-07-30 14:15:12 TP14 EP1] Tree cache initialized: source=default impl=UnifiedRadixCache hybrid_swa=True hybrid_ssm=False hierarchical=False streaming_wrapped=False
+[2026-07-30 14:15:12 TP8 EP1] Capture target decode CUDA graph end. elapsed=92.40 s, mem usage=1.37 GB, avail mem=5.35 GB.
+[2026-07-30 14:15:12 TP8 EP1] Init Unified RadixTree with components (, )
+[2026-07-30 14:15:12 TP8 EP1] Tree cache initialized: source=default impl=UnifiedRadixCache hybrid_swa=True hybrid_ssm=False hierarchical=False streaming_wrapped=False
+[2026-07-30 14:15:12] Dummy health check server started in background thread at 0.0.0.0:30002
diff --git a/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/server/worker_server_cmd.txt b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/server/worker_server_cmd.txt
new file mode 100644
index 0000000..3b18997
--- /dev/null
+++ b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/server/worker_server_cmd.txt
@@ -0,0 +1 @@
+docker run -d --name dsv4pro_pro6000d_2node_sglang_tp16_quick_map_worker --gpus all --network host --ipc host --shm-size 20g --ulimit memlock=-1 --ulimit stack=67108864 -v /data/hf_models/DeepSeek-V4-Pro:/data/hf_models/DeepSeek-V4-Pro:ro -v /data/hzy/sglang_cache/dsv4_pro_tp16:/root/.cache -e CUDA_VISIBLE_DEVICES=0\,1\,2\,3\,4\,5\,6\,7 -e PYTHONUNBUFFERED=1 -e HF_HUB_OFFLINE=1 -e TRANSFORMERS_OFFLINE=1 -e PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True -e NCCL_SOCKET_IFNAME=eth0 -e NCCL_IB_HCA==mlx5_0:1\,mlx5_3:1 -e NCCL_CROSS_NIC=1 -e NCCL_DEBUG=INFO -e SGLANG_SHARED_EXPERT_TP1=1 --device /dev/infiniband/rdma_cm --device /dev/infiniband/uverbs0 --device /dev/infiniband/uverbs3 --entrypoint python3 lmsysorg/sglang:nightly-dev-cu13-20260720-b3570a45 -m sglang.launch_server --model-path /data/hf_models/DeepSeek-V4-Pro --tp-size 16 --ep-size 2 --nnodes 2 --node-rank 1 --dist-init-addr 10.101.0.11:20002 --trust-remote-code --host 0.0.0.0 --port 30002 --mem-fraction-static 0.9 --cuda-graph-max-bs-decode 64 --max-running-requests 256
diff --git a/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/summary.csv b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/summary.csv
new file mode 100644
index 0000000..07f1d0c
--- /dev/null
+++ b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/summary.csv
@@ -0,0 +1,13 @@
+run_id,suite,case_id,role,stage,repetition,isl,osl,concurrency,num_prompts,warmup_requests,status,error_type,exit_code,started_at,ended_at,elapsed_s,completed,failed,duration_s,actual_concurrency,peak_concurrent_requests,total_input_tokens,total_output_tokens,request_throughput,input_token_throughput,output_token_throughput,total_token_throughput,peak_output_token_throughput,e2e_mean_ms,e2e_p50_ms,e2e_p95_ms,e2e_p99_ms,ttft_mean_ms,ttft_p50_ms,ttft_p95_ms,ttft_p99_ms,tpot_mean_ms,tpot_p50_ms,tpot_p95_ms,tpot_p99_ms,itl_mean_ms,itl_p50_ms,itl_p95_ms,itl_p99_ms,bench_file,bench_log
+dsv4pro-phase1-full-20260730-220916,fixed,balanced_32k_to_1k_c8,,balanced,1,32768,1024,8,8,0,COMPLETED,,0,2026-07-30T22:28:41+0800,2026-07-30T22:31:22+0800,161.0,8,0,128.62811431800947,7.996993535425267,,262144,8192,0.06219480120979978,2037.9992460427193,63.68747643883498,2101.686722481554,,128579.77483438299,128578.26417451724,128588.8987001963,128589.40238882786,49665.467494261975,49563.69844998699,82083.64989476977,83746.94262376754,77.14008537646238,77.23808966229743,109.32309288714016,112.17132956471158,77.14005810044779,43.40646349010058,43.84165157971437,45.188912986195646,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/balanced_32k_to_1k_c8/rep1/bench.jsonl,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/balanced_32k_to_1k_c8/rep1/bench.log
+dsv4pro-phase1-full-20260730-220916,mixed,decode_control_1k_to_1k_c32,control,mixed_interference,1,1024,1024,32,64,1,COMPLETED,,0,2026-07-30T22:31:28+0800,2026-07-30T22:34:26+0800,178.0,64,0,143.81865712802391,31.98259215252821,,65536,65536,0.44500485040010307,455.68496680970554,455.68496680970554,911.3699336194111,,71870.21023202942,71867.41446750239,72007.63789927005,72012.73389439622,7370.452653370194,6938.839260517852,9442.636363586644,9443.245556704933,63.04961640142642,63.471088219448355,65.88327712923635,67.80910156527888,63.049588087913975,60.98580302204937,61.753195300116204,63.56921844591854,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_control_1k_to_1k_c32/rep1/bench.jsonl,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_control_1k_to_1k_c32/rep1/bench.log
+dsv4pro-phase1-full-20260730-220916,fixed,decode_latency_1k_to_1k_c1,,decode_latency,1,1024,1024,1,1,1,COMPLETED,,0,2026-07-30T22:21:44+0800,2026-07-30T22:22:50+0800,66.0,1,0,32.59594450396253,0.9987527008467322,,1024,1024,0.03067866310419184,31.414951018692445,31.414951018692445,62.82990203738489,,32555.28760998277,32555.28760998277,32555.28760998277,32555.28760998277,363.2723209448159,363.2723209448159,363.2723209448159,363.2723209448159,31.468245639333286,31.468245639333286,31.468245639333286,31.468245639333286,31.468210095789587,31.49253799347207,31.659404194215313,32.3411487112753,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_latency_1k_to_1k_c1/rep1/bench.jsonl,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_latency_1k_to_1k_c1/rep1/bench.log
+dsv4pro-phase1-full-20260730-220916,fixed,decode_throughput_1k_to_1k_c16,,decode_throughput,1,1024,1024,16,16,1,COMPLETED,,0,2026-07-30T22:22:56+0800,2026-07-30T22:24:25+0800,89.0,16,0,55.48426674498478,15.986115557256861,,16384,16384,0.28837003602370287,295.29091688827174,295.29091688827174,590.5818337765435,,55436.11873718692,55437.53465998452,55443.82167501317,55447.12010782387,4706.02819519263,4801.126986538293,4950.267374995747,4950.924708603998,49.58953132159755,49.49641186020903,50.01878498413362,50.915034893211796,49.58950283761395,49.32670452399179,49.8873026604997,52.25673661392648,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_throughput_1k_to_1k_c16/rep1/bench.jsonl,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_throughput_1k_to_1k_c16/rep1/bench.log
+dsv4pro-phase1-full-20260730-220916,fixed,decode_throughput_1k_to_1k_c32,,decode_throughput,1,1024,1024,32,32,1,COMPLETED,,0,2026-07-30T22:24:30+0800,2026-07-30T22:26:15+0800,105.0,32,0,70.97485623596003,31.975931620502816,,32768,32768,0.4508638931738606,461.68462661003326,461.68462661003326,923.3692532200665,,70921.47343050237,70921.31606096518,70933.18627287517,70935.98751158977,6650.527383468216,6178.440199000761,8021.852347906679,8023.242657405207,62.825949215087164,63.29258629570958,63.30661669023356,65.29997366317116,62.82592070965087,61.42424049903639,62.237933481810614,64.6031521842815,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_throughput_1k_to_1k_c32/rep1/bench.jsonl,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_throughput_1k_to_1k_c32/rep1/bench.log
+dsv4pro-phase1-full-20260730-220916,fixed,decode_throughput_1k_to_1k_c64,,decode_throughput,1,1024,1024,64,64,1,COMPLETED,,0,2026-07-30T22:26:20+0800,2026-07-30T22:28:36+0800,136.0,64,0,101.22691119497176,63.94603747793439,,65536,65536,0.6322429405825737,647.4167711565555,647.4167711565555,1294.833542313111,,101141.56026639376,101148.19569600513,101163.1970004586,101167.44335912867,7978.397482784203,6505.179281026358,12716.491229846724,12719.49685954547,91.06858532122146,92.52174521165054,93.44459645826957,94.39186767905878,91.06855870121748,86.24391999910586,87.80462378927041,90.7129056059057,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_throughput_1k_to_1k_c64/rep1/bench.jsonl,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_throughput_1k_to_1k_c64/rep1/bench.log
+dsv4pro-phase1-full-20260730-220916,mixed,decode_with_128k_prefill_1k_to_1k_c32,decode_background,mixed_interference,1,1024,1024,32,64,0,COMPLETED,,0,2026-07-30T22:34:26+0800,2026-07-30T22:38:09+0800,223.0,64,0,189.43961485201726,31.98654958092739,,65536,65536,0.3378385246929174,345.94664928554744,345.94664928554744,691.8932985710949,,94679.9942664975,94677.34680901049,117629.63820034638,117636.8834006862,7749.917391879535,7705.110649490962,10193.896457596566,10196.455221220385,84.97563721859039,86.47172644327328,109.7299457727384,110.56891953740994,84.97560842059728,60.77706499490887,61.61956858995836,64.70431641500909,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_with_128k_prefill_1k_to_1k_c32/rep1/bench.jsonl,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_with_128k_prefill_1k_to_1k_c32/rep1/bench.log
+dsv4pro-phase1-full-20260730-220916,mixed,long_prefill_injection_128k_to_1_c1,prefill_injection,mixed_interference,1,131072,1,1,1,0,COMPLETED,,0,2026-07-30T22:35:06+0800,2026-07-30T22:36:24+0800,78.0,1,0,45.83367280900711,0.9982227218104723,,131072,1,0.021818020217735695,2859.731545979053,0.021818020217735695,2859.7533639992707,,45752.21362197772,45752.21362197772,45752.21362197772,45752.21362197772,45752.13837000774,45752.13837000774,45752.13837000774,45752.13837000774,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/long_prefill_injection_128k_to_1_c1/rep1/bench.jsonl,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/long_prefill_injection_128k_to_1_c1/rep1/bench.log
+dsv4pro-phase1-full-20260730-220916,fixed,long_prefill_latency_128k_c1,,prefill_latency,1,131072,1,1,1,0,COMPLETED,,0,2026-07-30T22:16:51+0800,2026-07-30T22:18:12+0800,81.0,1,0,48.363155752944294,0.9996148667368967,,131072,1,0.020676897204730506,2710.162270418437,0.020676897204730506,2710.1829473156417,,48344.52949295519,48344.52949295519,48344.52949295519,48344.52949295519,48344.461318978574,48344.461318978574,48344.461318978574,48344.461318978574,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/long_prefill_latency_128k_c1/rep1/bench.jsonl,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/long_prefill_latency_128k_c1/rep1/bench.log
+dsv4pro-phase1-full-20260730-220916,fixed,mid_prefill_latency_32k_c1,,prefill_latency,1,32768,1,1,1,0,COMPLETED,,0,2026-07-30T22:16:00+0800,2026-07-30T22:16:46+0800,46.0,1,0,12.352426828991156,0.9985701137731423,,32768,1,0.08095575175988892,2652.75807366804,0.08095575175988892,2652.8390294198,,12334.764264000114,12334.764264000114,12334.764264000114,12334.764264000114,12334.68782599084,12334.68782599084,12334.68782599084,12334.68782599084,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/mid_prefill_latency_32k_c1/rep1/bench.jsonl,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/mid_prefill_latency_32k_c1/rep1/bench.log
+dsv4pro-phase1-full-20260730-220916,fixed,mid_prefill_throughput_32k_c16,,prefill_throughput,1,32768,1,16,16,0,COMPLETED,,0,2026-07-30T22:18:18+0800,2026-07-30T22:21:39+0800,201.0,16,0,168.4313552599633,8.771898294558559,,524288,16,0.09499418902914482,3112.7695861070174,0.09499418902914482,3112.8645802960464,,92341.41987219118,92749.04042950948,162087.50443853205,167117.798359727,92341.37187193119,92748.99350100895,162087.46085499297,167117.75140058598,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/mid_prefill_throughput_32k_c16/rep1/bench.jsonl,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/mid_prefill_throughput_32k_c16/rep1/bench.log
+dsv4pro-phase1-full-20260730-220916,fixed,short_prefill_latency_1k_c1,,prefill_latency,1,1024,1,1,1,1,COMPLETED,,0,2026-07-30T22:15:21+0800,2026-07-30T22:15:55+0800,34.0,1,0,0.5198853989713825,0.9654920680740482,,1024,1,1.9235008368739468,1969.6648569589215,1.9235008368739468,1971.5883577957954,,501.94522901438177,501.94522901438177,501.94522901438177,501.94522901438177,501.89953204244375,501.89953204244375,501.89953204244375,501.89953204244375,0.0,0.0,0.0,0.0,0.0,0.0,0.0,0.0,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/short_prefill_latency_1k_c1/rep1/bench.jsonl,/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/short_prefill_latency_1k_c1/rep1/bench.log
diff --git a/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/summary.jsonl b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/summary.jsonl
new file mode 100644
index 0000000..12f3d2c
--- /dev/null
+++ b/docs/dsv4pro_pro6000d_2node_sglang/results/dsv4pro-phase1-full-20260730-220916/summary.jsonl
@@ -0,0 +1,12 @@
+{"run_id": "dsv4pro-phase1-full-20260730-220916", "suite": "fixed", "case_id": "balanced_32k_to_1k_c8", "role": "", "stage": "balanced", "repetition": 1, "isl": 32768, "osl": 1024, "concurrency": 8, "num_prompts": 8, "warmup_requests": 0, "status": "COMPLETED", "error_type": "", "exit_code": 0, "started_at": "2026-07-30T22:28:41+0800", "ended_at": "2026-07-30T22:31:22+0800", "elapsed_s": 161.0, "completed": 8, "failed": 0, "duration_s": 128.62811431800947, "actual_concurrency": 7.996993535425267, "peak_concurrent_requests": null, "total_input_tokens": 262144, "total_output_tokens": 8192, "request_throughput": 0.06219480120979978, "input_token_throughput": 2037.9992460427193, "output_token_throughput": 63.68747643883498, "total_token_throughput": 2101.686722481554, "peak_output_token_throughput": null, "e2e_mean_ms": 128579.77483438299, "e2e_p50_ms": 128578.26417451724, "e2e_p95_ms": 128588.8987001963, "e2e_p99_ms": 128589.40238882786, "ttft_mean_ms": 49665.467494261975, "ttft_p50_ms": 49563.69844998699, "ttft_p95_ms": 82083.64989476977, "ttft_p99_ms": 83746.94262376754, "tpot_mean_ms": 77.14008537646238, "tpot_p50_ms": 77.23808966229743, "tpot_p95_ms": 109.32309288714016, "tpot_p99_ms": 112.17132956471158, "itl_mean_ms": 77.14005810044779, "itl_p50_ms": 43.40646349010058, "itl_p95_ms": 43.84165157971437, "itl_p99_ms": 45.188912986195646, "bench_file": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/balanced_32k_to_1k_c8/rep1/bench.jsonl", "bench_log": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/balanced_32k_to_1k_c8/rep1/bench.log"}
+{"run_id": "dsv4pro-phase1-full-20260730-220916", "suite": "mixed", "case_id": "decode_control_1k_to_1k_c32", "role": "control", "stage": "mixed_interference", "repetition": 1, "isl": 1024, "osl": 1024, "concurrency": 32, "num_prompts": 64, "warmup_requests": 1, "status": "COMPLETED", "error_type": "", "exit_code": 0, "started_at": "2026-07-30T22:31:28+0800", "ended_at": "2026-07-30T22:34:26+0800", "elapsed_s": 178.0, "completed": 64, "failed": 0, "duration_s": 143.81865712802391, "actual_concurrency": 31.98259215252821, "peak_concurrent_requests": null, "total_input_tokens": 65536, "total_output_tokens": 65536, "request_throughput": 0.44500485040010307, "input_token_throughput": 455.68496680970554, "output_token_throughput": 455.68496680970554, "total_token_throughput": 911.3699336194111, "peak_output_token_throughput": null, "e2e_mean_ms": 71870.21023202942, "e2e_p50_ms": 71867.41446750239, "e2e_p95_ms": 72007.63789927005, "e2e_p99_ms": 72012.73389439622, "ttft_mean_ms": 7370.452653370194, "ttft_p50_ms": 6938.839260517852, "ttft_p95_ms": 9442.636363586644, "ttft_p99_ms": 9443.245556704933, "tpot_mean_ms": 63.04961640142642, "tpot_p50_ms": 63.471088219448355, "tpot_p95_ms": 65.88327712923635, "tpot_p99_ms": 67.80910156527888, "itl_mean_ms": 63.049588087913975, "itl_p50_ms": 60.98580302204937, "itl_p95_ms": 61.753195300116204, "itl_p99_ms": 63.56921844591854, "bench_file": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_control_1k_to_1k_c32/rep1/bench.jsonl", "bench_log": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_control_1k_to_1k_c32/rep1/bench.log"}
+{"run_id": "dsv4pro-phase1-full-20260730-220916", "suite": "fixed", "case_id": "decode_latency_1k_to_1k_c1", "role": "", "stage": "decode_latency", "repetition": 1, "isl": 1024, "osl": 1024, "concurrency": 1, "num_prompts": 1, "warmup_requests": 1, "status": "COMPLETED", "error_type": "", "exit_code": 0, "started_at": "2026-07-30T22:21:44+0800", "ended_at": "2026-07-30T22:22:50+0800", "elapsed_s": 66.0, "completed": 1, "failed": 0, "duration_s": 32.59594450396253, "actual_concurrency": 0.9987527008467322, "peak_concurrent_requests": null, "total_input_tokens": 1024, "total_output_tokens": 1024, "request_throughput": 0.03067866310419184, "input_token_throughput": 31.414951018692445, "output_token_throughput": 31.414951018692445, "total_token_throughput": 62.82990203738489, "peak_output_token_throughput": null, "e2e_mean_ms": 32555.28760998277, "e2e_p50_ms": 32555.28760998277, "e2e_p95_ms": 32555.28760998277, "e2e_p99_ms": 32555.28760998277, "ttft_mean_ms": 363.2723209448159, "ttft_p50_ms": 363.2723209448159, "ttft_p95_ms": 363.2723209448159, "ttft_p99_ms": 363.2723209448159, "tpot_mean_ms": 31.468245639333286, "tpot_p50_ms": 31.468245639333286, "tpot_p95_ms": 31.468245639333286, "tpot_p99_ms": 31.468245639333286, "itl_mean_ms": 31.468210095789587, "itl_p50_ms": 31.49253799347207, "itl_p95_ms": 31.659404194215313, "itl_p99_ms": 32.3411487112753, "bench_file": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_latency_1k_to_1k_c1/rep1/bench.jsonl", "bench_log": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_latency_1k_to_1k_c1/rep1/bench.log"}
+{"run_id": "dsv4pro-phase1-full-20260730-220916", "suite": "fixed", "case_id": "decode_throughput_1k_to_1k_c16", "role": "", "stage": "decode_throughput", "repetition": 1, "isl": 1024, "osl": 1024, "concurrency": 16, "num_prompts": 16, "warmup_requests": 1, "status": "COMPLETED", "error_type": "", "exit_code": 0, "started_at": "2026-07-30T22:22:56+0800", "ended_at": "2026-07-30T22:24:25+0800", "elapsed_s": 89.0, "completed": 16, "failed": 0, "duration_s": 55.48426674498478, "actual_concurrency": 15.986115557256861, "peak_concurrent_requests": null, "total_input_tokens": 16384, "total_output_tokens": 16384, "request_throughput": 0.28837003602370287, "input_token_throughput": 295.29091688827174, "output_token_throughput": 295.29091688827174, "total_token_throughput": 590.5818337765435, "peak_output_token_throughput": null, "e2e_mean_ms": 55436.11873718692, "e2e_p50_ms": 55437.53465998452, "e2e_p95_ms": 55443.82167501317, "e2e_p99_ms": 55447.12010782387, "ttft_mean_ms": 4706.02819519263, "ttft_p50_ms": 4801.126986538293, "ttft_p95_ms": 4950.267374995747, "ttft_p99_ms": 4950.924708603998, "tpot_mean_ms": 49.58953132159755, "tpot_p50_ms": 49.49641186020903, "tpot_p95_ms": 50.01878498413362, "tpot_p99_ms": 50.915034893211796, "itl_mean_ms": 49.58950283761395, "itl_p50_ms": 49.32670452399179, "itl_p95_ms": 49.8873026604997, "itl_p99_ms": 52.25673661392648, "bench_file": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_throughput_1k_to_1k_c16/rep1/bench.jsonl", "bench_log": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_throughput_1k_to_1k_c16/rep1/bench.log"}
+{"run_id": "dsv4pro-phase1-full-20260730-220916", "suite": "fixed", "case_id": "decode_throughput_1k_to_1k_c32", "role": "", "stage": "decode_throughput", "repetition": 1, "isl": 1024, "osl": 1024, "concurrency": 32, "num_prompts": 32, "warmup_requests": 1, "status": "COMPLETED", "error_type": "", "exit_code": 0, "started_at": "2026-07-30T22:24:30+0800", "ended_at": "2026-07-30T22:26:15+0800", "elapsed_s": 105.0, "completed": 32, "failed": 0, "duration_s": 70.97485623596003, "actual_concurrency": 31.975931620502816, "peak_concurrent_requests": null, "total_input_tokens": 32768, "total_output_tokens": 32768, "request_throughput": 0.4508638931738606, "input_token_throughput": 461.68462661003326, "output_token_throughput": 461.68462661003326, "total_token_throughput": 923.3692532200665, "peak_output_token_throughput": null, "e2e_mean_ms": 70921.47343050237, "e2e_p50_ms": 70921.31606096518, "e2e_p95_ms": 70933.18627287517, "e2e_p99_ms": 70935.98751158977, "ttft_mean_ms": 6650.527383468216, "ttft_p50_ms": 6178.440199000761, "ttft_p95_ms": 8021.852347906679, "ttft_p99_ms": 8023.242657405207, "tpot_mean_ms": 62.825949215087164, "tpot_p50_ms": 63.29258629570958, "tpot_p95_ms": 63.30661669023356, "tpot_p99_ms": 65.29997366317116, "itl_mean_ms": 62.82592070965087, "itl_p50_ms": 61.42424049903639, "itl_p95_ms": 62.237933481810614, "itl_p99_ms": 64.6031521842815, "bench_file": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_throughput_1k_to_1k_c32/rep1/bench.jsonl", "bench_log": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_throughput_1k_to_1k_c32/rep1/bench.log"}
+{"run_id": "dsv4pro-phase1-full-20260730-220916", "suite": "fixed", "case_id": "decode_throughput_1k_to_1k_c64", "role": "", "stage": "decode_throughput", "repetition": 1, "isl": 1024, "osl": 1024, "concurrency": 64, "num_prompts": 64, "warmup_requests": 1, "status": "COMPLETED", "error_type": "", "exit_code": 0, "started_at": "2026-07-30T22:26:20+0800", "ended_at": "2026-07-30T22:28:36+0800", "elapsed_s": 136.0, "completed": 64, "failed": 0, "duration_s": 101.22691119497176, "actual_concurrency": 63.94603747793439, "peak_concurrent_requests": null, "total_input_tokens": 65536, "total_output_tokens": 65536, "request_throughput": 0.6322429405825737, "input_token_throughput": 647.4167711565555, "output_token_throughput": 647.4167711565555, "total_token_throughput": 1294.833542313111, "peak_output_token_throughput": null, "e2e_mean_ms": 101141.56026639376, "e2e_p50_ms": 101148.19569600513, "e2e_p95_ms": 101163.1970004586, "e2e_p99_ms": 101167.44335912867, "ttft_mean_ms": 7978.397482784203, "ttft_p50_ms": 6505.179281026358, "ttft_p95_ms": 12716.491229846724, "ttft_p99_ms": 12719.49685954547, "tpot_mean_ms": 91.06858532122146, "tpot_p50_ms": 92.52174521165054, "tpot_p95_ms": 93.44459645826957, "tpot_p99_ms": 94.39186767905878, "itl_mean_ms": 91.06855870121748, "itl_p50_ms": 86.24391999910586, "itl_p95_ms": 87.80462378927041, "itl_p99_ms": 90.7129056059057, "bench_file": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_throughput_1k_to_1k_c64/rep1/bench.jsonl", "bench_log": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_throughput_1k_to_1k_c64/rep1/bench.log"}
+{"run_id": "dsv4pro-phase1-full-20260730-220916", "suite": "mixed", "case_id": "decode_with_128k_prefill_1k_to_1k_c32", "role": "decode_background", "stage": "mixed_interference", "repetition": 1, "isl": 1024, "osl": 1024, "concurrency": 32, "num_prompts": 64, "warmup_requests": 0, "status": "COMPLETED", "error_type": "", "exit_code": 0, "started_at": "2026-07-30T22:34:26+0800", "ended_at": "2026-07-30T22:38:09+0800", "elapsed_s": 223.0, "completed": 64, "failed": 0, "duration_s": 189.43961485201726, "actual_concurrency": 31.98654958092739, "peak_concurrent_requests": null, "total_input_tokens": 65536, "total_output_tokens": 65536, "request_throughput": 0.3378385246929174, "input_token_throughput": 345.94664928554744, "output_token_throughput": 345.94664928554744, "total_token_throughput": 691.8932985710949, "peak_output_token_throughput": null, "e2e_mean_ms": 94679.9942664975, "e2e_p50_ms": 94677.34680901049, "e2e_p95_ms": 117629.63820034638, "e2e_p99_ms": 117636.8834006862, "ttft_mean_ms": 7749.917391879535, "ttft_p50_ms": 7705.110649490962, "ttft_p95_ms": 10193.896457596566, "ttft_p99_ms": 10196.455221220385, "tpot_mean_ms": 84.97563721859039, "tpot_p50_ms": 86.47172644327328, "tpot_p95_ms": 109.7299457727384, "tpot_p99_ms": 110.56891953740994, "itl_mean_ms": 84.97560842059728, "itl_p50_ms": 60.77706499490887, "itl_p95_ms": 61.61956858995836, "itl_p99_ms": 64.70431641500909, "bench_file": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_with_128k_prefill_1k_to_1k_c32/rep1/bench.jsonl", "bench_log": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/decode_with_128k_prefill_1k_to_1k_c32/rep1/bench.log"}
+{"run_id": "dsv4pro-phase1-full-20260730-220916", "suite": "mixed", "case_id": "long_prefill_injection_128k_to_1_c1", "role": "prefill_injection", "stage": "mixed_interference", "repetition": 1, "isl": 131072, "osl": 1, "concurrency": 1, "num_prompts": 1, "warmup_requests": 0, "status": "COMPLETED", "error_type": "", "exit_code": 0, "started_at": "2026-07-30T22:35:06+0800", "ended_at": "2026-07-30T22:36:24+0800", "elapsed_s": 78.0, "completed": 1, "failed": 0, "duration_s": 45.83367280900711, "actual_concurrency": 0.9982227218104723, "peak_concurrent_requests": null, "total_input_tokens": 131072, "total_output_tokens": 1, "request_throughput": 0.021818020217735695, "input_token_throughput": 2859.731545979053, "output_token_throughput": 0.021818020217735695, "total_token_throughput": 2859.7533639992707, "peak_output_token_throughput": null, "e2e_mean_ms": 45752.21362197772, "e2e_p50_ms": 45752.21362197772, "e2e_p95_ms": 45752.21362197772, "e2e_p99_ms": 45752.21362197772, "ttft_mean_ms": 45752.13837000774, "ttft_p50_ms": 45752.13837000774, "ttft_p95_ms": 45752.13837000774, "ttft_p99_ms": 45752.13837000774, "tpot_mean_ms": 0.0, "tpot_p50_ms": 0.0, "tpot_p95_ms": 0.0, "tpot_p99_ms": 0.0, "itl_mean_ms": 0.0, "itl_p50_ms": 0.0, "itl_p95_ms": 0.0, "itl_p99_ms": 0.0, "bench_file": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/long_prefill_injection_128k_to_1_c1/rep1/bench.jsonl", "bench_log": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/long_prefill_injection_128k_to_1_c1/rep1/bench.log"}
+{"run_id": "dsv4pro-phase1-full-20260730-220916", "suite": "fixed", "case_id": "long_prefill_latency_128k_c1", "role": "", "stage": "prefill_latency", "repetition": 1, "isl": 131072, "osl": 1, "concurrency": 1, "num_prompts": 1, "warmup_requests": 0, "status": "COMPLETED", "error_type": "", "exit_code": 0, "started_at": "2026-07-30T22:16:51+0800", "ended_at": "2026-07-30T22:18:12+0800", "elapsed_s": 81.0, "completed": 1, "failed": 0, "duration_s": 48.363155752944294, "actual_concurrency": 0.9996148667368967, "peak_concurrent_requests": null, "total_input_tokens": 131072, "total_output_tokens": 1, "request_throughput": 0.020676897204730506, "input_token_throughput": 2710.162270418437, "output_token_throughput": 0.020676897204730506, "total_token_throughput": 2710.1829473156417, "peak_output_token_throughput": null, "e2e_mean_ms": 48344.52949295519, "e2e_p50_ms": 48344.52949295519, "e2e_p95_ms": 48344.52949295519, "e2e_p99_ms": 48344.52949295519, "ttft_mean_ms": 48344.461318978574, "ttft_p50_ms": 48344.461318978574, "ttft_p95_ms": 48344.461318978574, "ttft_p99_ms": 48344.461318978574, "tpot_mean_ms": 0.0, "tpot_p50_ms": 0.0, "tpot_p95_ms": 0.0, "tpot_p99_ms": 0.0, "itl_mean_ms": 0.0, "itl_p50_ms": 0.0, "itl_p95_ms": 0.0, "itl_p99_ms": 0.0, "bench_file": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/long_prefill_latency_128k_c1/rep1/bench.jsonl", "bench_log": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/long_prefill_latency_128k_c1/rep1/bench.log"}
+{"run_id": "dsv4pro-phase1-full-20260730-220916", "suite": "fixed", "case_id": "mid_prefill_latency_32k_c1", "role": "", "stage": "prefill_latency", "repetition": 1, "isl": 32768, "osl": 1, "concurrency": 1, "num_prompts": 1, "warmup_requests": 0, "status": "COMPLETED", "error_type": "", "exit_code": 0, "started_at": "2026-07-30T22:16:00+0800", "ended_at": "2026-07-30T22:16:46+0800", "elapsed_s": 46.0, "completed": 1, "failed": 0, "duration_s": 12.352426828991156, "actual_concurrency": 0.9985701137731423, "peak_concurrent_requests": null, "total_input_tokens": 32768, "total_output_tokens": 1, "request_throughput": 0.08095575175988892, "input_token_throughput": 2652.75807366804, "output_token_throughput": 0.08095575175988892, "total_token_throughput": 2652.8390294198, "peak_output_token_throughput": null, "e2e_mean_ms": 12334.764264000114, "e2e_p50_ms": 12334.764264000114, "e2e_p95_ms": 12334.764264000114, "e2e_p99_ms": 12334.764264000114, "ttft_mean_ms": 12334.68782599084, "ttft_p50_ms": 12334.68782599084, "ttft_p95_ms": 12334.68782599084, "ttft_p99_ms": 12334.68782599084, "tpot_mean_ms": 0.0, "tpot_p50_ms": 0.0, "tpot_p95_ms": 0.0, "tpot_p99_ms": 0.0, "itl_mean_ms": 0.0, "itl_p50_ms": 0.0, "itl_p95_ms": 0.0, "itl_p99_ms": 0.0, "bench_file": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/mid_prefill_latency_32k_c1/rep1/bench.jsonl", "bench_log": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/mid_prefill_latency_32k_c1/rep1/bench.log"}
+{"run_id": "dsv4pro-phase1-full-20260730-220916", "suite": "fixed", "case_id": "mid_prefill_throughput_32k_c16", "role": "", "stage": "prefill_throughput", "repetition": 1, "isl": 32768, "osl": 1, "concurrency": 16, "num_prompts": 16, "warmup_requests": 0, "status": "COMPLETED", "error_type": "", "exit_code": 0, "started_at": "2026-07-30T22:18:18+0800", "ended_at": "2026-07-30T22:21:39+0800", "elapsed_s": 201.0, "completed": 16, "failed": 0, "duration_s": 168.4313552599633, "actual_concurrency": 8.771898294558559, "peak_concurrent_requests": null, "total_input_tokens": 524288, "total_output_tokens": 16, "request_throughput": 0.09499418902914482, "input_token_throughput": 3112.7695861070174, "output_token_throughput": 0.09499418902914482, "total_token_throughput": 3112.8645802960464, "peak_output_token_throughput": null, "e2e_mean_ms": 92341.41987219118, "e2e_p50_ms": 92749.04042950948, "e2e_p95_ms": 162087.50443853205, "e2e_p99_ms": 167117.798359727, "ttft_mean_ms": 92341.37187193119, "ttft_p50_ms": 92748.99350100895, "ttft_p95_ms": 162087.46085499297, "ttft_p99_ms": 167117.75140058598, "tpot_mean_ms": 0.0, "tpot_p50_ms": 0.0, "tpot_p95_ms": 0.0, "tpot_p99_ms": 0.0, "itl_mean_ms": 0.0, "itl_p50_ms": 0.0, "itl_p95_ms": 0.0, "itl_p99_ms": 0.0, "bench_file": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/mid_prefill_throughput_32k_c16/rep1/bench.jsonl", "bench_log": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/mid_prefill_throughput_32k_c16/rep1/bench.log"}
+{"run_id": "dsv4pro-phase1-full-20260730-220916", "suite": "fixed", "case_id": "short_prefill_latency_1k_c1", "role": "", "stage": "prefill_latency", "repetition": 1, "isl": 1024, "osl": 1, "concurrency": 1, "num_prompts": 1, "warmup_requests": 1, "status": "COMPLETED", "error_type": "", "exit_code": 0, "started_at": "2026-07-30T22:15:21+0800", "ended_at": "2026-07-30T22:15:55+0800", "elapsed_s": 34.0, "completed": 1, "failed": 0, "duration_s": 0.5198853989713825, "actual_concurrency": 0.9654920680740482, "peak_concurrent_requests": null, "total_input_tokens": 1024, "total_output_tokens": 1, "request_throughput": 1.9235008368739468, "input_token_throughput": 1969.6648569589215, "output_token_throughput": 1.9235008368739468, "total_token_throughput": 1971.5883577957954, "peak_output_token_throughput": null, "e2e_mean_ms": 501.94522901438177, "e2e_p50_ms": 501.94522901438177, "e2e_p95_ms": 501.94522901438177, "e2e_p99_ms": 501.94522901438177, "ttft_mean_ms": 501.89953204244375, "ttft_p50_ms": 501.89953204244375, "ttft_p95_ms": 501.89953204244375, "ttft_p99_ms": 501.89953204244375, "tpot_mean_ms": 0.0, "tpot_p50_ms": 0.0, "tpot_p95_ms": 0.0, "tpot_p99_ms": 0.0, "itl_mean_ms": 0.0, "itl_p50_ms": 0.0, "itl_p95_ms": 0.0, "itl_p99_ms": 0.0, "bench_file": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/short_prefill_latency_1k_c1/rep1/bench.jsonl", "bench_log": "/data/hzy/sskj/experiments/pro6000/dsv4pro_pro6000d_2node_sglang_tp16_quick_map/results/dsv4pro-phase1-full-20260730-220916/cases/short_prefill_latency_1k_c1/rep1/bench.log"}
diff --git a/docs/dsv4pro_pro6000d_2node_sglang/推理优化计划.html b/docs/dsv4pro_pro6000d_2node_sglang/推理优化计划.html
index f661f8f..64f4f9b 100644
--- a/docs/dsv4pro_pro6000d_2node_sglang/推理优化计划.html
+++ b/docs/dsv4pro_pro6000d_2node_sglang/推理优化计划.html
@@ -400,7 +400,7 @@
@@ -415,7 +415,7 @@
6000D 双机 DeepSeek-V4-Pro 推理优化计划
-适用环境:174.1.51.5 + 174.1.51.7,每台 8 张 RTX PRO 6000 Blackwell Server Edition
当前部署:DeepSeek-V4-Pro,16 张 GPU 组成一个完整实例
当前约束:模型暂时只能使用全部 16 张 GPU,无法额外复制一套模型进行 PD 分离
计划版本:2026-07-30 18:40 CST
+适用环境:174.1.51.5 + 174.1.51.7,每台 8 张 RTX PRO 6000 Blackwell Server Edition
当前部署:DeepSeek-V4-Pro,16 张 GPU 组成一个完整实例
当前约束:模型暂时只能使用全部 16 张 GPU,无法额外复制一套模型进行 PD 分离
计划版本:2026-07-30 22:55:37 CST
当前执行状态与阶段档案
@@ -434,12 +434,12 @@
| DeepSeek-V4-Pro / 双机 Pro6000D / SGLang TP16 快速性能地图 |
-网络错误已定位;RDMA fail-closed 代码与 dry-run 已完成,待真机 NET/IB 验证后重跑 |
+已完成;双 Rail NET/IB + GDRDMA,固定点 9/9、混合 A/B 3/3,总用时 28 分 36 秒 |
打开实施记录 |
| DeepSeek-V4-Pro / 双机 Pro6000D / SGLang Prefill 硬件指标归因 |
-设计已固化、代码尚未开始;等待修正后的 Phase 1 基线 |
+Phase 1 已交接,三个诊断负载已选定;等待阶段确认后开始代码实现 |
打开 Phase 2 档案 |
@@ -596,18 +596,17 @@ Nsight Compute 或专项 Microbenchmark
Phase 1:双机 SGLang TP16 快速性能地图实施记录。
代码采用单一 Shell 入口 run_quick_map.sh,不修改或调用旧的全天全量脚本。
-最终 Run 完成 1K、32K、128K 三个单请求 Prefill 点,输入吞吐分别为
-64.44、64.96、65.20 token/s。第 4 个 32K, C=16 点在稳定复现
-单序列 Chunk 推进后被主动中止;Manifest 状态为
-ABORTED_EARLY_FOR_PHASE2。Decode、Balanced 和混合 A/B 未执行。
+最终 Run dsv4pro-phase1-full-20260730-220916 已完成:
+Head 与 Worker 均通过双 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。
+单请求 32K/128K Prefill 输入吞吐为 2,652.76/2,710.16 token/s;
+1K → 1K Decode 的 Output TPS 从 C=1 的 31.41 提升到 C=64 的 647.42。
-旧脚本 TTFT 较短的问题已完成复核。新旧服务端关键运行参数相同;旧脚本固定
-warmup_requests=16、不清 Prefix Cache、按固定 Seed 递增长度,
-且正式 Run 前已有一次失败 Run 预热同一批请求。最小复现中,两次清 Cache 的
-1K 冷 TTFT 分别为 16.04 秒和 15.90 秒,OSL=128 时为 15.79 秒;旧产物的
-0.455 秒无法在冷缓存口径复现。因此当前 65 token/s 明确解释为完整冷 Prompt
-路径,Warm Prefix 性能后续单独做 A/B。
+最值得继续归因的现象有两个:32K Prefill 从 C=1 增至 C=16 时,
+聚合输入吞吐只提高 17.3%,TTFT P95 却从 12.335 秒增至 162.087 秒;
+在 C=32 Decode 中注入一个 128K Prefill 后,Output TPS 下降 24.08%,
+TPOT P95 增加 66.55%。Phase 2 将围绕这两个现象采集硬件时间序列。
5.1 当前固定快速矩阵
@@ -692,7 +691,11 @@ Nsight Compute 或专项 Microbenchmark
- 正式测量开始 10 秒后注入一个
128K → 1, C=1 长 Prefill;不额外执行 128K Warm-up。
- 比较 Output TPS、P95 TTFT、P95 TPOT 与 P95 E2E 的变化。
-该 A/B 在 Phase 1 中未执行,待 Prefill 异常完成归因后再决定是否重放。这是聚合级干扰探针,不声称具备逐请求时间线归因能力;后者在 Scheduler 与 Timeline 阶段实现。
+
+该 A/B 已完成:注入 128K Prefill 后,Decode Output TPS 从 455.68 降至
+345.95 token/s,TPOT P95 从 65.88 ms 增至 109.73 ms。它证明存在聚合级干扰,
+但逐请求调度时间线仍由后续阶段补齐。
+
5.3 本轮运行与停止策略
- 固定矩阵不做 Add-16 搜索,也不按 SLO 提前终止。
@@ -722,10 +725,10 @@ Nsight Compute 或专项 Microbenchmark
本阶段的设计、代码改动与结果同步维护在
-Phase 2:Prefill 硬件指标归因档案。第一轮只重放
-32K → 1, C=1,目标是在约 20 分钟内区分 GPU、CPU、双 Rail、
-频率节流和节点不均衡。该请求在测量前清 Prefix Cache;旧脚本中的热缓存
-TTFT 不作为本阶段参照值。
+Phase 2:Prefill 硬件指标归因档案。本阶段只重放三类代表负载:
+128K → 1, C=1、32K → 1, C=16,以及
+1K → 1K, C=32 的混合 A/B。目标是在约 30 分钟内区分 GPU、
+CPU、双 Rail、调度排队、频率节流和节点不均衡。
6.1 GPU
测试期间持续记录:
@@ -762,19 +765,11 @@ numastat -p <PID>
- CPU 空洞是否对应 GPU 空洞。
6.3 网络
-
-2026-07-30 控制组已确认:宿主机具备 mlx5_0/eth0 与
-mlx5_3/eth3 两条 400G Rail,但原服务容器没有
-/dev/infiniband,NCCL 实际使用 NET/Socket。
-当前第一优先级是校正容器数据面并重做 Phase 1,不再把约 65 token/s 当作模型或算子瓶颈。
-
-
-
-唯一启动入口现已在两端预检并透传
-rdma_cm/uverbs0/uverbs3,且服务健康后强制从两端 NCCL INFO
-日志确认 NET/IB 同时识别 mlx5_0/mlx5_3。代码、
-单测与 dry-run 已通过;真机服务尚未启动,因此这里仍不宣称 RDMA 已验证成功。
+Phase 1 正式 Run 已从两端 NCCL 日志确认
+NET/IB 同时识别 mlx5_0/mlx5_3,跨节点 Channel
+使用 GDRDMA。Phase 2 继续保留相同 fail-closed 门禁,并采集两条
+Rail 的流量和错误计数。
当前拓扑中需要分别观察两条 Compute Rail,确认:
@@ -795,10 +790,9 @@ NCCL_DEBUG_SUBSYS=INIT,NET,GRAPH,TUNING
该日志开销较高,不应在正式性能结果中长期启用。
6.4 NCCL_CROSS_NIC 快速 A/B
-旧 Phase 1 虽然设置了 NCCL_CROSS_NIC=1,但由于 NCCL 回退
-NET/Socket,该参数没有参与实际路径选择。先让容器真正使用
-mlx5_0/mlx5_3 的 NET/IB,再固定其余环境比较
-0/1/2,不能仅凭双 Rail 拓扑判断最优值。
+Phase 1 已证明 NCCL_CROSS_NIC=1 可以稳定完成端到端 Run,
+但这不代表它是性能最优值。固定其余环境后比较 0/1/2,
+不能仅凭双 Rail 拓扑判断。
- 分别执行相同消息范围的
all_reduce_perf,每个值至少重复 3 次,检查错误、algbw 和 busbw。
@@ -810,8 +804,8 @@ NCCL_DEBUG_SUBSYS=INIT,NET,GRAPH,TUNING
- Phase 1 保留一份不启用 Profiler 的端到端基线,避免 TPS 和时延被诊断工具污染。
- GPU、CPU 和网络的轻量采样可以伴随后续基线运行,但必须从 Case 开始前启动,并使用统一时间戳与 Case ID 对齐。
-- Phase 1 已提前结束,不把中途人工观察到的 GPU 数值伪装成完整硬件时间序列。
-- Phase 2 首轮独立重放 32K Prefill 并完整采集轻量指标;得到瓶颈方向后,再决定是否把 Phase 2 指标与 Phase 3 的短时间线放在同一诊断 Run。
+- Phase 1 已完整结束,其无 Profiler 结果作为后续 A/B 的请求层基线。
+- Phase 2 重放三个代表负载并完整采集轻量指标;得到瓶颈方向后,再决定 Phase 3 捕获哪段短时间线。
- 联合诊断 Run 的吞吐和时延只用于解释时间线;正式性能变化仍与 Phase 1 的无 Profiler 结果比较。
7. Phase 3:时间线 Profiling(Nsight Systems 为主)