[AMD][AgentX] DSv4 MI355X agentic disagg: add EAGLE/MTP speculative-decoding arm / DSv4 MI355X 智能体分离式推理：新增 EAGLE/MTP 投机解码分支

[ichbinblau]

Summary

DSv4 FP4 MI355X SGLang PD-disaggregated agentic-coding benchmark: HiCache sizing/NCCL/router fixes, the TP8/EP8/DPA hicache search-space arm, and a new EAGLE/MTP speculative-decoding arm. Originally split as #2308 (fixes) stacked under #2309 (MTP), now combined into a single PR targeting main directly.

What's changed

HiCache sizing (ratio vs size)

• dsv4_fp4_mi355x_sglang-disagg.sh: set HICACHE_RATIO=4 and FORCE_HICACHE_RATIO=1 so this recipe explicitly opts into ratio-based sizing (DeepSeek-V4's hybrid HiCache pool rejects --hicache-size).
• job.slurm: forward FORCE_HICACHE_RATIO into the container's HiCache env file (hicache_mc_<job>.env) alongside the other HICACHE_* vars -- it was previously silently dropped, so the override never reached the server.

NCCL / decode cuda-graph crash workaround

• env.sh: add TORCH_NCCL_BLOCKING_WAIT/NCCL_BLOCKING_WAIT=1 to avoid a ROCm 7.2 HIP-runtime watchdog race during decode cuda-graph capture ([AMD] qwen3-235b-mxfp4 nightly aborts at CUDA-graph capture: NCCL watchdog hipEventQuery on capturing stream (hipErrorCapturedEvent) sgl-project/sglang#29235, [Block Issue][hip graph]: HIP error: operation not permitted when stream is capturing ROCm/hip#3876, [ROCm] Avoid watchdog event queries during graph capture pytorch/pytorch#176251); NCCL_DEBUG=INFO left commented out (log-spam reduction).
• models.yaml: disable prefill cuda-graph (disable_cuda_graph: true) for DeepSeek-V4-Pro-AgentX, drop --enable-two-batch-overlap from dp_flags, lower prefill mem_fraction_static 0.8->0.72 and no-dp max_running_requests 128->64; remove the redundant disable_cuda_graph: false from decode (already the code default).
• Recipe image bump: lmsysorg/sglang-rocm:v0.5.14-rocm720-mi35x-20260710 -> v0.5.15.post1-rocm720-mi35x-20260719.

RDMA QoS/DCQCN + GPU-sanity pre-flight gates (job.slurm)

• New helpers/rdma_check.sh: validates NIC QoS (PFC/DSCP) and DCQCN configuration on every allocated node via srun before any container/GPU time is spent. Override with SKIP_RDMA_CHECK=1.
• New helpers/gpu_sanity.sh: after container pre-clean/teardown, fails fast (naming the offending process) if a bare, non-containerized process is still holding GPU VRAM.

Router policy fix + DPA (ep=8/dp-attn) hicache arm

• server_sglang.sh (agentic branch only -- DI branch untouched): fixes a bug where the script read ROUTER_PREFILL_POLICY but recipe scripts export PREFILL_ROUTER_POLICY, so the recipe's policy override was silently ignored and the router always fell back to the hardcoded default. Also switches the default prefill router policy cache_aware -> consistent_hashing, adds --dp-aware to the router flags, and comments out the separate --decode-policy override (superseded by --dp-aware).
• Adds PREFILL_DP_ATTN/DECODE_DP_ATTN to the env-var whitelist forwarded into the sibling client container, alongside the existing PREFILL_ENABLE_DP/DECODE_ENABLE_DP.
• configs/amd-master.yaml: bump dsv4-fp4-mi355x-sglang-disagg-agentic-hicache image to v0.5.15.post1-rocm720-mi35x-20260719; re-enable (previously commented out) the TP8/EP8/DPA (ep=8, dp-attn=true) hicache search-space arm at conc-list: [64, 96, 128].
• trace_replay.sh: when PREFILL_ENABLE_DP=true, export AIPERF_HTTP_X_SMG_ROUTING_KEY_FROM_CORRELATION_ID=true so aiperf's HTTP routing key is derived from the correlation ID under DP-attention.

EAGLE/MTP speculative-decoding arm

• models.yaml: add mtp_flags: "--speculative-algorithm EAGLE --speculative-eagle-topk 1" to DeepSeek-V4-Pro-AgentX. DeepSeek-V4's MTP head is a native NextN/EAGLE draft shipped with the model, so SGLang defaults the draft path to the main checkpoint -- no --speculative-draft-model-path needed. --speculative-num-steps/--speculative-num-draft-tokens are already derived from DECODE_MTP_SIZE in server_sglang.sh; mtp_flags is only appended to the decode server config (MTP is decode-only).
• configs/amd-master.yaml: new dsv4-fp4-mi355x-sglang-disagg-agentic-hicache-mtp config-key with two search-space arms mirroring the non-MTP hicache config's topologies (spec-decoding: "mtp", DECODE_MTP_SIZE=3): TP8/EP1 no-DP at conc-list: [2, 4, 8, 16, 32], and TP8/EP8/DPA at conc-list: [64, 96, 128].

perf-changelog.yaml

Two appended entries (tail of file, per the append-only convention), one per config-key touched, both linked to this PR: dsv4-fp4-mi355x-sglang-disagg-agentic-hicache (image bump / HiCache / NCCL / router fixes / DPA arm) and dsv4-fp4-mi355x-sglang-disagg-agentic-hicache-mtp (new EAGLE/MTP arm).

Test plan

• bash -n on all touched shell scripts (env.sh, job.slurm, server_sglang.sh, trace_replay.sh, setup_deps.sh, helpers/gpu_sanity.sh, helpers/rdma_check.sh) -- syntax OK
• configs/amd-master.yaml, benchmarks/multi_node/amd_utils/models.yaml, perf-changelog.yaml -- YAML parses cleanly
• generate_sweep_configs.py full-sweep resolves all 16 dsv4-fp4-mi355x-sglang-disagg-agentic-hicache* configs correctly: 8 non-MTP (TP8/EP1 conc 2/4/8/16/32, TP8/EP8/DPA conc 64/96/128) + 8 MTP twins with identical topologies plus spec-decoding: mtp
• Diffed against main to confirm no unintended regression in the IS_AGENTIC/DI router split or the router-readiness canary (wait_or_die)
• CI: e2e Test - dsv4-fp4-mi355x-sglang-disagg-agentic-hicache and ...-hicache-mtp (multi-node, mi355x-amds)

中文说明

DSv4 FP4 MI355X SGLang 预填充/解码分离式（disaggregated）智能体编码（agentic-coding）基准测试：修复 HiCache 容量配置、NCCL 以及路由策略问题，重新启用 TP8/EP8/DPA 的 hicache 搜索空间分支，并新增一个 EAGLE/MTP 投机解码（speculative decoding）分支。原本拆分为 #2308（修复项）叠加在 #2309（MTP）之上，现已合并为一个直接面向 main 的单一 PR。

HiCache 容量配置（ratio 与 size）

• dsv4_fp4_mi355x_sglang-disagg.sh：设置 HICACHE_RATIO=4 和 FORCE_HICACHE_RATIO=1，使该配方显式启用基于比例（ratio-based）的容量配置（DeepSeek-V4 的混合 HiCache 池不支持 --hicache-size）。
• job.slurm：将 FORCE_HICACHE_RATIO 转发到容器的 HiCache 环境变量文件（hicache_mc_<job>.env）中，此前该变量被静默丢弃，导致覆盖配置无法生效。

NCCL / 解码 CUDA-graph 崩溃的规避方案

• env.sh：新增 TORCH_NCCL_BLOCKING_WAIT/NCCL_BLOCKING_WAIT=1，以规避 ROCm 7.2 在解码 CUDA-graph 捕获阶段出现的 HIP 运行时看门狗（watchdog）竞态问题；NCCL_DEBUG=INFO 保持注释状态以减少日志噪音。
• models.yaml：为 DeepSeek-V4-Pro-AgentX 禁用预填充 CUDA-graph（disable_cuda_graph: true），从 dp_flags 中移除 --enable-two-batch-overlap，将预填充 mem_fraction_static 从 0.8 降至 0.72，非 DP 场景下 max_running_requests 从 128 降至 64；移除解码侧多余的 disable_cuda_graph: false（已是代码默认值）。
• 镜像版本升级：lmsysorg/sglang-rocm:v0.5.14-rocm720-mi35x-20260710 -> v0.5.15.post1-rocm720-mi35x-20260719。

RDMA QoS/DCQCN 与 GPU 健康检查前置校验（job.slurm）

• 新增 helpers/rdma_check.sh：在占用任何容器/GPU 时间之前，通过 srun 校验每个分配节点的网卡 QoS（PFC/DSCP）与 DCQCN 配置；可通过 SKIP_RDMA_CHECK=1 跳过。
• 新增 helpers/gpu_sanity.sh：容器预清理/收尾之后，若仍有裸机（非容器化）进程占用 GPU 显存，则快速失败并报告具体进程。

路由策略修复 + DPA（ep=8/dp-attn）hicache 分支

• server_sglang.sh（仅智能体分支，DI 分支未改动）：修复脚本读取 ROUTER_PREFILL_POLICY 而配方实际导出的是 PREFILL_ROUTER_POLICY 的问题，此前该覆盖配置被静默忽略，路由器始终回退到硬编码默认值。同时将默认预填充路由策略由 cache_aware 切换为 consistent_hashing，为路由参数新增 --dp-aware，并注释掉原有的 --decode-policy 覆盖项（已被 --dp-aware 取代）。
• 将 PREFILL_DP_ATTN/DECODE_DP_ATTN 加入转发到客户端容器的环境变量白名单，与现有的 PREFILL_ENABLE_DP/DECODE_ENABLE_DP 并列。
• configs/amd-master.yaml：将 dsv4-fp4-mi355x-sglang-disagg-agentic-hicache 镜像升级至 v0.5.15.post1-rocm720-mi35x-20260719；重新启用此前被注释的 TP8/EP8/DPA（ep=8, dp-attn=true）hicache 搜索空间分支，并发列表为 [64, 96, 128]。
• trace_replay.sh：当 PREFILL_ENABLE_DP=true 时，导出 AIPERF_HTTP_X_SMG_ROUTING_KEY_FROM_CORRELATION_ID=true，使 aiperf 的 HTTP 路由键在 DP-attention 下基于关联 ID（correlation ID）派生。

EAGLE/MTP 投机解码分支

• models.yaml：为 DeepSeek-V4-Pro-AgentX 新增 mtp_flags: "--speculative-algorithm EAGLE --speculative-eagle-topk 1"。DeepSeek-V4 的 MTP 头是模型自带的原生 NextN/EAGLE 草稿模型，因此 SGLang 默认将草稿路径指向主检查点，无需 --speculative-draft-model-path。--speculative-num-steps/--speculative-num-draft-tokens 已在 server_sglang.sh 中由 DECODE_MTP_SIZE 派生；mtp_flags 仅追加到解码侧服务器配置中（MTP 仅作用于解码）。
• configs/amd-master.yaml：新增 dsv4-fp4-mi355x-sglang-disagg-agentic-hicache-mtp 配置项，包含两个搜索空间分支，拓扑结构与非 MTP 版本的 hicache 配置一致（spec-decoding: "mtp"，DECODE_MTP_SIZE=3）：TP8/EP1 非 DP，并发列表为 [2, 4, 8, 16, 32]；以及 TP8/EP8/DPA，并发列表为 [64, 96, 128]。

perf-changelog.yaml

在文件末尾追加两条记录（遵循只追加约定），分别对应本次改动涉及的两个配置项，均关联本 PR：dsv4-fp4-mi355x-sglang-disagg-agentic-hicache（镜像升级/HiCache/NCCL/路由修复/DPA 分支）与 dsv4-fp4-mi355x-sglang-disagg-agentic-hicache-mtp（新增 EAGLE/MTP 分支）。

测试计划

• 对所有改动的 shell 脚本（env.sh、job.slurm、server_sglang.sh、trace_replay.sh、setup_deps.sh、helpers/gpu_sanity.sh、helpers/rdma_check.sh）执行 bash -n —— 语法正确
• configs/amd-master.yaml、benchmarks/multi_node/amd_utils/models.yaml、perf-changelog.yaml —— YAML 解析无误
• generate_sweep_configs.py full-sweep 正确解析全部 16 个 dsv4-fp4-mi355x-sglang-disagg-agentic-hicache* 配置：8 个非 MTP（TP8/EP1 并发 2/4/8/16/32，TP8/EP8/DPA 并发 64/96/128）+ 8 个拓扑相同并带 spec-decoding: mtp 的 MTP 对应版本
• 与 main 分支对比，确认 IS_AGENTIC/DI 路由拆分逻辑及路由就绪校验（wait_or_die）未出现意外回归
• CI：e2e Test - dsv4-fp4-mi355x-sglang-disagg-agentic-hicache 与 ...-hicache-mtp（多节点，mi355x-amds）

Made with Cursor

[github-actions]

Thanks for the contribution! Please reach out to respective companies' CODEOWNER to fill in the latest PR_REVIEW_CHECKLIST.md before pinging core maintainer on Slack for review. In order for the signoff PR check bot to trigger, you must follow the PR_REVIEW_CHECKLIST.md template correctly, including the phrase As a PR reviewer and CODEOWNER, I have reviewed this and have.

For PR verification, add the full-sweep-fail-fast label (strongly recommended) to this PR — the benchmark sweep only runs on labeled PRs. Use full-sweep-enabled only if you need matrix jobs to keep running past a failure.

PR authors are responsible for ensuring that after merging, all GitHub Action jobs fully pass. A lot of the time, failures are just flakes and simply re-running the failed jobs will fix it. See GitHub's docs on re-running failed jobs

---

感谢你的贡献！请联系相应公司的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md，然后再在 Slack 上联系核心维护者进行审阅。为了触发 signoff PR 检查机器人，你必须正确遵循 PR_REVIEW_CHECKLIST.md 模板，包括保留英文语句 As a PR reviewer and CODEOWNER, I have reviewed this and have。

如需进行 PR 验证，请为此 PR 添加 full-sweep-fail-fast 标签（强烈推荐）— 基准测试 sweep 仅在带有标签的 PR 上运行。仅当需要矩阵任务在失败后继续运行时才使用 full-sweep-enabled。

PR 作者有责任确保合并后所有 GitHub Action 任务完全通过。 很多时候失败只是偶发抖动（flake），重新运行失败的任务即可解决。参见 GitHub 关于重新运行失败任务的文档