[AMD][AgentX] GLM-5.2 FP4 MI355X sglang-disagg HiCache MTP / [AMD][AgentX] MI355X 上 GLM-5.2 FP4 分离式 HiCache MTP - #2679
Conversation
…er config Introduce 1P1D TP8 disaggregated agentic bring-up for amd/GLM-5.2-MXFP4 on MI355X (models.yaml AgentX entry, launcher recipe, amd-master hicache sweep). Scope ionic MSN-safe MoRI IO QP tuning to GLM-5.2-MXFP4 with DISAGG=true in env.sh; keep HiCache on ratio for GLM-5.2 DSA pools. Pin sglang-rocm v0.5.16-rocm720-mi35x-20260729; improve SLURM_REUSE nodelist resolution. 中文:新增 GLM-5.2 MXFP4 MI355X 分离式 SGLang agentic 启动脚本与 amd-master 配置(1P1D TP8、HiCache);在 env.sh 中仅对 GLM-5.2-MXFP4 且 DISAGG 开启时应用 MoRI IO ionic 调参;HiCache 对 GLM-5.2 使用 ratio sizing;镜像 v0.5.16;并改进 SLURM 复用作业的节点列表解析。 Co-authored-by: Cursor <cursoragent@cursor.com>
Document validated 1P1D L2 settings (ratio 1.5, selective write-through, page_first_direct) in glm5.2-fp4-mi355x-sglang-disagg-agentic-hicache prefill additional-settings for sweep/CI parity with cluster bring-up. 中文:在 amd-master 的 GLM-5.2 分离式 agentic HiCache 配置中写入经 1P1D 验证的 L2 参数(ratio 1.5、write_through_selective、page_first_direct 等), 便于扫描与 CI 和集群 bring-up 一致。 Co-authored-by: Cursor <cursoragent@cursor.com>
…e config Set MI355X amd-aim MODEL_PATH/MODEL_DIR to /it-share/hf_cache in the GLM sglang-disagg recipe so job.slurm does not fall back to /it-share/data. Write IBDEVICES and MORI_RDMA_TC into hicache_mc.env for sglang containers. 中文:在 GLM 分离式 recipe 中默认使用 /it-share/hf_cache,避免 job.slurm 误用 /it-share/data;并通过 hicache_mc.env 向容器传递 IBDEVICES 与 MORI_RDMA_TC。 Co-authored-by: Cursor <cursoragent@cursor.com>
Use GLM52_MODEL_ROOT (default /it-share/hf_cache) for MODEL_PATH so
launch_mi355x-amds.sh cannot override with /it-share/data via ${VAR:-}.
中文:GLM recipe 通过 GLM52_MODEL_ROOT 固定权重根目录(默认 /it-share/hf_cache),
避免 CI launcher 已设置的 MODEL_PATH=/it-share/data 被 ${MODEL_PATH:-} 保留而导致找错路径。
Co-authored-by: Cursor <cursoragent@cursor.com>
Align GLM disagg recipe and amd-master comments with launch_mi355x-amds.sh and job.slurm mia1* MODEL_DIR; remove hf_cache / GLM52_MODEL_ROOT override. 中文:GLM 分离式 recipe 默认使用 /it-share/data,与 CI launcher 及 job.slurm 一致; 不再强制 /it-share/hf_cache。 Co-authored-by: Cursor <cursoragent@cursor.com>
Enable built-in MTP/EAGLE on the MI355X GLM-5.2 disagg HiCache agentic path: amd-master spec-decoding mtp with DECODE_MTP_SIZE=2, EAGLE mtp_flags and decode mem_fraction_static 0.85 in models.yaml, and agentic script MTP block with simulated acceptance for non-eval throughput runs. 中文:为 MI355X GLM-5.2 分离式 HiCache AgentX 配置启用内置 MTP/EAGLE 投机解码:在 amd-master 中改为 mtp 与 DECODE_MTP_SIZE=2,在 models.yaml 中补充 EAGLE mtp_flags 并将 decode 静态显存比例调至 0.85,在 agentic 脚本中增加 MTP 逻辑(非 eval 吞吐跑使用模拟接受长度)。 Co-authored-by: Cursor <cursoragent@cursor.com>
中文:在 perf-changelog.yaml 末尾追加 glm5.2-fp4-mi355x-sglang-disagg-agentic-hicache 条目。 Co-authored-by: Cursor <cursoragent@cursor.com>
|
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 For PR verification, add the 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 如需进行 PR 验证,请为此 PR 添加 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 |
…e pr-link 中文:将 perf-changelog.yaml 中该条目的 pr-link 填为 #2679。 Co-authored-by: Cursor <cursoragent@cursor.com>
Summary
glm5.2-fp4-mi355x-sglang-disagg-agentic-hicachetoamd-master.yaml: GLM-5.2 MXFP4 on MI355X, 1P1D TP8 disaggregated SGLang (sglang-disagg, MoRI P2P), agentic-coding,conc-list[2, 4, 8, 16, 32], DRAM HiCache KV offload, built-in MTP (spec-decoding: mtp,DECODE_MTP_SIZE=2).benchmarks/multi_node/agentic/glm5.2_fp4_mi355x_sglang-disagg.sh(CI-style sibling of the DSv4 MI355X disagg recipe): env-driven submit viasubmit.sh, HiCache defaults, MTP gate + simulated acceptance for non-eval throughput.models.yamlGLM-5.2-MXFP4-AgentX: EAGLEmtp_flags(--speculative-algorithm EAGLE --speculative-eagle-topk 1); decodemem_fraction_static0.85.env.sh: ionic MSN-safe MoRI IO QP tuning scoped toGLM-5.2-MXFP4+DISAGG=true.job.slurm/ recipe: default model root/it-share/data(mia1*); forwardIBDEVICES/MORI_RDMA_TCintohicache_mc.env.additional-settings:HICACHE_RATIO=1.5,write_through_selective,page_first_direct,directIO,page_size=1.golden_al_distribution/glm5.2_mtp.yaml(thinking_on,num_speculative_tokens=2→ AL 2.50:SGLANG_SIMULATE_ACC_LEN=2.50);EVAL_ONLY=trueruns use real draft-model MTP.perf-changelog.yamlentry forglm5.2-fp4-mi355x-sglang-disagg-agentic-hicache.Image:
lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260729Model:
amd/GLM-5.2-MXFP4KV offload: SGLang HiCache (DRAM,
kv-offload-backend: hicache,dram-utilization: 0.80)Topology: 1P1D TP8 / EP1, MoRI disagg transfer; MTP size 2 (EAGLE)
Test plan
full-sweep-fail-fast(orsweep-enabledfor trim-conc); green sweep on this PR/reuse-sweep-runafter full green with evalsRelated
golden_al_distribution/glm5.2_mtp.yaml(from [Klaud Cold] feat(golden-al): add GLM-5.2 MTP golden AL distribution / 新增 GLM-5.2 MTP 黄金接受长度(AL)分布 #2460)中文说明
amd-master.yaml新增glm5.2-fp4-mi355x-sglang-disagg-agentic-hicache:MI355X 上 GLM-5.2 MXFP4 的 1P1D TP8 分离式 SGLang(sglang-disagg,MoRI P2P),agentic-coding 场景,conc-list为 [2, 4, 8, 16, 32],DRAM HiCache KV 卸载,内置 MTP(spec-decoding: mtp,DECODE_MTP_SIZE=2)。benchmarks/multi_node/agentic/glm5.2_fp4_mi355x_sglang-disagg.sh(对标 DSv4 MI355X 分离式 recipe 的 CI 入口):由环境变量驱动、经submit.sh提交;含 HiCache 默认参数,以及 MTP 门控与非 eval 吞吐的模拟接受长度。models.yaml的GLM-5.2-MXFP4-AgentX:补充 EAGLEmtp_flags(--speculative-algorithm EAGLE --speculative-eagle-topk 1);decodemem_fraction_static调至 0.85。env.sh:仅在GLM-5.2-MXFP4且DISAGG=true时启用 ionic MSN-safe MoRI IO QP 调参。job.slurm/ recipe:默认模型根目录/it-share/data(mia1*);经hicache_mc.env向容器转发IBDEVICES/MORI_RDMA_TC。additional-settings中固化经验证的 HiCache L2 参数:HICACHE_RATIO=1.5、write_through_selective、page_first_direct、directIO、page_size=1。golden_al_distribution/glm5.2_mtp.yaml的合成接受长度(thinking_on、num_speculative_tokens=2→ AL 2.50:SGLANG_SIMULATE_ACC_LEN=2.50);EVAL_ONLY=true时走真实 draft-model MTP。perf-changelog.yaml末尾追加glm5.2-fp4-mi355x-sglang-disagg-agentic-hicache条目。镜像:
lmsysorg/sglang-rocm:v0.5.16-rocm720-mi35x-20260729模型:
amd/GLM-5.2-MXFP4KV 卸载: SGLang HiCache(DRAM,
kv-offload-backend: hicache,dram-utilization: 0.80)拓扑: 1P1D TP8 / EP1,MoRI 分离式传输;MTP size 2(EAGLE)
测试计划
full-sweep-fail-fast(或sweep-enabled做 trim-conc);在本 PR 上跑出绿 sweep/reuse-sweep-run相关
golden_al_distribution/glm5.2_mtp.yaml(来自 [Klaud Cold] feat(golden-al): add GLM-5.2 MTP golden AL distribution / 新增 GLM-5.2 MTP 黄金接受长度(AL)分布 #2460)