Repository navigation
Update Qwen3.5 GB300 Dynamo-TRT AgentX recipes / 更新 Qwen3.5 GB300 Dynamo-TRT AgentX 配方 - #4452
RohitNagraj wants to merge 5 commits into
Conversation
更新 Qwen3.5 GB300 Dynamo-TRT AgentX 配方、TensorRT-LLM 镜像及 srt-slurm 版本。
|
Thanks for the contribution!
中文感谢你的贡献！
|
为 Qwen3.5 GB300 AgentX 变更日志条目填写 PR 链接。
RohitNagraj
added
the
non-canary-full-sweep-enabled
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
| - config-keys: | ||
| - qwen3.5-fp4-gb300-dynamo-trt-agentic-disagg | ||
| description: | ||
| - "Update Qwen3.5 GB300 Dynamo-TRT AgentX recipes, TensorRT-LLM image, and NUMA placement" | ||
| pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/XXX |
There was a problem hiding this comment.
🔴 Other srt-slurm configs pick up the srt-slurm v2.60.1 runtime with no sweep or changelog record. Only the Qwen3.5 key is covered. This entry lists only qwen3.5-fp4-gb300-dynamo-trt-agentic-disagg, and its description never mentions the pin bump. Fix: describe the upgrade from v2.59.1 to v2.60.1 in this PR's single block. Also list every config key whose run it changes, at least the ones with a decode-node frontend (such as the dsv4 trtllm gb300 AgentX recipe), or use "*" as the v2.43.4 pin-bump entry did.
Why this was flagged
The diff moves inferencex-e2e/utils/srt-slurm from ae3d687 (v2.59.1) to 3ce1c43 (v2.60.1). Every srt-slurm job clones that submodule (RECIPES.md:5). The PR description says this changes the runtime for all srt-slurm jobs, including post-eval routing for other recipes that put the frontend on a decode node. benchmarks/multi_node/srt-slurm-recipes/dsv4/trtllm/gb300-fp4/agentx/disagg-variants.yaml:147 is one such recipe (node: first_decode). infx.matrix.changed only snapshots configs and recipe dirs (revision.py:34-40), so it cannot detect a submodule bump. AGENTS.md:140 says to judge such changes by what they run, and the entry is what triggers result collection. With only the Qwen3.5 key listed (perf-changelog.yaml:11065), no sweep validates the other configs on the new runtime. The published results for those configs also have no record of the runtime change. The earlier v2.39.1 to v2.43.4 bump used config-keys: ["*"] and named the upgrade (perf-changelog.yaml:9344-9347).
Verification: The diff moves the submodule inferencex-e2e/utils/srt-slurm from ae3d6877c46f… to 3ce1c43677921dd4…, which every srt-slurm job clones (RECIPES.md:3). The only new changelog block (perf-changelog.yaml:11064-11068) lists only qwen3.5-fp4-gb300-dynamo-trt-agentic-disagg and never mentions the pin bump, though AGENTS.md:140 requires an entry covering the affected config keys.
将共享 srt-slurm v2.59.1 至 v2.60.1 的升级及评测位置调整写入变更日志，并使用通配符覆盖所有配置。
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=40025992272 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=40025992272 |
Resolve the perf-changelog.yaml append conflict by keeping main's entries and moving this PR's entry to the end. Picks up #4438, which fixes the single-node recipe matching that failed check-changelog. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ntX config A "*" entry plans every config, and the resulting search-space-config exceeds GitHub's job output size limit, failing setup. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Update
qwen3.5-fp4-gb300-dynamo-trt-agentic-disaggto ten recipe points using TensorRT-LLM1.3.0a0.dev202610210406and Dynamo1.6.0.dev20261010. The recipes select Qwen3.5 context/generation profiles, explicit NUMA affinity, MTP decode with draft length 6, and native host KV offload.Upgrade the shared srt-slurm pin from v2.59.1 to v2.60.1 for
numa_memory_bind: preferredand post-eval routing to the resolved public frontend. This changes the runtime used by all srt-slurm jobs, including eval placement for other recipes with a decode-node frontend. Both worker roles use iteration-statistics intervals of 20.AI model disclosure
claude-opus-5-5): recipe and configuration authoring.PR checklist and change type
Type of Change
Checklist
python -m infx.matrix.changed --base origin/mainlists affected config keys, I have appended one new entry covering them to the physical end ofinferencex-e2e/perf-changelog.yaml; if it lists none, I have added none. I have not edited historical entriesinferencex-e2e/perf-changelog.yamlentry, it carries exactly one primary sweep label:full-sweep-fail-fast(recommended),full-sweep-enabled, ornon-canary-full-sweep-enabled. Optional modifiersall-evals,evals-only, andagentx-fastrequire a primary label; the last two block reuse while applied.OWNER/MEMBER/COLLABORATOR) has commented/use <run_id>(or the legacy/reuse-sweep-run) on this PR. Do this only once there is a final full sweep that is all green with evals passing, since after this comment the primary sweep label will no longer automatically kick off new sweeps. Remove and re-add the primary sweep label to force a new sweep.中文
将
qwen3.5-fp4-gb300-dynamo-trt-agentic-disagg更新为十个配方配置点，使用 TensorRT-LLM1.3.0a0.dev202610210406和 Dynamo1.6.0.dev20261010。配方选择 Qwen3.5 context/generation 配置档，显式设置 NUMA 亲和性，decode 使用 draft 长度为 6 的 MTP，并启用原生主机 KV 卸载。将共享 srt-slurm 从 v2.59.1 升级至 v2.60.1，以支持
numa_memory_bind: preferred，并让 post-eval 连接到已解析的公共 frontend。此升级影响所有 srt-slurm 作业使用的运行时，也会改变其他将 frontend 放在 decode 节点上的配方的评测位置。prefill 和 decode 的迭代统计间隔均设为 20。改动类型为配置与文档更新。作者检查清单见上方折叠区。
AI 模型披露：Claude Opus 5.5（
claude-opus-5-5）用于配方和配置编写；Codex 中的 GPT-6 用于配置编辑、本地校验和 PR 准备，无法核实运行时的精确模型标识及版本。未使用委派代理。