perf(amd): switch DSV4 ATOM AgentX to native RCCL DEP / 将 DSV4 ATOM AgentX 切换到原生 RCCL DEP

[yhl-amd]

Summary / 概要

• Switch the DeepSeek-V4-Pro MI355X ATOM AgentX concurrency 48/64/96/128/256 band introduced by perf(amd): add DP-attention band to DSV4 ATOM MI355X AgentX MTP #2778 from DPA+TBO with the default MoE transport to native RCCL DEP: TP8, DPA8, EP8, RCCL all-to-all, TBO off, and EPLB off.

• Keep the concurrency 1/2/4/8/16 TP band unchanged at TP8/EP1 and retain its golden MTP acceptance length.

• Pin rocm/atom-dev:nightly_202609071454 (sha256:f252ec9a30e3d17b91f15887ea32baa1a0c5503b775f5adf42ff7cee8cfff769), built by the successful ROCm/ATOM release run at merge commit 5a9c2068 containing native RCCL DEP.

• Match the validated local c48 run: least-tokens scheduling, session affinity, ATOM_DP_LB_REQ_EQUIV=0, 32K state checkpoints, terminal-MTP optimizations, real MTP acceptance, and the validated AIPerf timeout/failure settings.

• 将 perf(amd): add DP-attention band to DSV4 ATOM MI355X AgentX MTP #2778 引入的 DeepSeek-V4-Pro MI355X ATOM AgentX c48/64/96/128/256 档位，从默认 MoE 通信的 DPA+TBO 切换为原生 RCCL DEP：TP8、DPA8、EP8、RCCL all-to-all，并关闭 TBO 和 EPLB。

• 保持 c1/2/4/8/16 档位为原来的 TP8/EP1，同时保留该档位的 MTP golden acceptance length，不改变低并发测试语义。

• 固定使用 rocm/atom-dev:nightly_202609071454（sha256:f252ec9a30e3d17b91f15887ea32baa1a0c5503b775f5adf42ff7cee8cfff769）。该镜像由成功的 ROCm/ATOM 发布任务基于包含原生 RCCL DEP 的合并提交 5a9c2068 构建。

• 对齐本地已验证的 c48 参数：least-tokens 调度、session affinity、ATOM_DP_LB_REQ_EQUIV=0、32K state checkpoint、terminal-MTP 优化、真实 MTP 接受率，以及已验证的 AIPerf 超时和失败阈值。

Local validation / 本地验证

Validated source run:
atom-native-rccl-dpa8-ep8-eplboff-agentx-c48-ci3600-realmtp-fp4idx-r4-20260902

• 3,920 completed requests, 0 request errors
• 1,066.71 output tokens/s
• 96.55% prompt cache-read rate
• Median ITL: 19.12 ms

本地 c48 实跑完成 3,920 个请求且请求错误数为 0；输出吞吐为 1,066.71 token/s，prompt cache-read rate 为 96.55%，ITL 中位数为 19.12 ms。

Test plan / 测试计划

• bash -n benchmarks/single_node/agentic/dsv4_fp4_mi355x_atom_mtp.sh

• Parse configs/amd-master.yaml, configs/runners.yaml, and perf-changelog.yaml with PyYAML

• git diff --check

• Generate the exact dsv4-fp4-mi355x-atom-agentic-mtp matrix with generate_sweep_configs.py

• Verify TP points are c1/2/4/8/16 with EP1 and DEP points are c48/64/96/128/256 with EP8+DPA

• Bash、YAML 和补丁格式检查通过

• 仓库矩阵生成器成功展开全部十个测试点，并确认低并发 TP 档和高并发 DEP 档的边界正确

References / 参考

• Base change / 基础改动: perf(amd): add DP-attention band to DSV4 ATOM MI355X AgentX MTP #2778
• Native RCCL DEP / 原生 RCCL DEP: [MoE] Add native RCCL DEP transport ROCm/ATOM#2110
• Image release / 镜像发布: https://github.com/ROCm/ATOM/actions/runs/34129583845

---

Note

Medium Risk
Benchmark recipe and cluster image pin change how high-concurrency AgentX runs are launched (MoE/DPA transport and acceptance-length behavior), which can shift reported throughput without affecting production serving code.

Overview
Moves the DeepSeek-V4-Pro MI355X ATOM AgentX high-concurrency sweep (c48–256) from DPA + TBO with TP-only MoE to native RCCL DEP: config now uses TP8 / DPA / EP8, pins rocm/atom-dev:nightly_202609071454, and the benchmark script wires expert parallel when EP_SIZE > 1 plus DEP server flags (--all2all-backend rccl, least-tokens load balance, standard MoE backend) with session affinity and stricter EP_SIZE == TP checks.

The low-concurrency TP band (c1–16, EP1, no DPA) is unchanged for throughput semantics: golden synthetic MTP acceptance applies only there; DEP and eval use real MTP acceptance. TBO/NUMA queue env vars and --enable-tbo are dropped from the DEP path; server launch adds --data-parallel-size 1 and splits EP_ARGS / DEP_ARGS instead of a single DPA+TBO arg list. perf-changelog documents the image bump and transport switch.

^{Reviewed by Cursor Bugbot for commit 5c91ae8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit 252e081. Configure here.}

[cursor]

Validated DEP knobs were reverted

High Severity

The DEP band now sets ATOM_DP_LB_REQ_EQUIV=512 and the old DPA prefill-delayer knobs, dropping the validated ATOM_DP_LB_REQ_EQUIV=0 and terminal-MTP flags. Least-tokens routing and MTP no longer match the claimed c48 run, so the high-concurrency band will not reproduce that configuration.

 

^{Reviewed by Cursor Bugbot for commit 252e081. Configure here.}

[yhl-amd]

Superseded by #2912, whose head branch is hosted directly in the upstream SemiAnalysisAI/InferenceX repository. The commit SHA is unchanged.