Use MXFP4 weights for PR 2139 GB200 vLLM recipes - #2185
Merged
Ankur-singh merged 1 commit intoJul 14, 2026
Conversation
Add a dedicated MXFP4 model-path alias for the five standard GB200 Dynamo-vLLM recipes while preserving the existing NVFP4 mapping for all other configurations.
![claude[bot]](https://avatars.githubusercontent.com/in/1236702?s=60&v=4){kind=small}
Ankur-singh
merged commit Jul 14, 2026
316ae1e
into
SemiAnalysisAI:dsv4-fp4-gb200-dynamo-vllm-config-refresh
Ankur-singh
added a commit
that referenced
this pull request
Jul 30, 2026
…GB200 Dynamo vLLM 推理栈 (#2139) * perf(dsv4): refresh GB200 Dynamo vLLM stack Update the vLLM image and Dynamo wheel, and enable multiple frontends across all five GB200 DeepSeek-V4-Pro topologies. 中文：更新 vLLM 镜像和 Dynamo wheel，并为 DeepSeek-V4-Pro 的全部五种 GB200 拓扑启用多个前端。 * chore: link performance changelog to PR #2139 Backfill the canonical pull request URL after creating the draft PR. 中文：创建草稿 PR 后，将正式 PR 链接回填至性能变更日志。 * perf(dsv4): use MessagePack for Dynamo request plane Set DYN_REQUEST_PLANE_CODEC=msgpack on every GB200 Dynamo frontend, prefill worker, and decode worker. 中文：为所有 GB200 Dynamo 前端、预填充 worker 和解码 worker 设置 DYN_REQUEST_PLANE_CODEC=msgpack。 * fix(dsv4): reduce GB200 1P4D decode memory Lower decode gpu-memory-utilization from 0.9 to 0.85 for the GB200 low-middle 1P4D TP8 recipe after the PR sweep OOM. 中文：针对 PR 扫描中出现的显存不足问题，将 GB200 低中段 1P4D TP8 配置的解码 gpu-memory-utilization 从 0.9 降至 0.85。 * Use MXFP4 weights for PR 2139 GB200 vLLM recipes (#2185) Add a dedicated MXFP4 model-path alias for the five standard GB200 Dynamo-vLLM recipes while preserving the existing NVFP4 mapping for all other configurations. * Align PR 2139 GB200 recipe identity versions (#2209) Update the five standard GB200 Dynamo-vLLM recipe identities to the active container, Dynamo 1.3.0.dev1, and vLLM 0.1.dev18219+g4ba0a72a4. * Use DeepGEMM-timeout vLLM setup for GB200 DSV4 (#2248) * dsv4-fp4-gb200-dynamo-vllm: vLLM 0.25.1 container, remove setup_script (drop DeepGEMM-timeout patch) --------- Co-authored-by: jthomson04 <jwillthomson19@gmail.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
deepseek-v4-pro-mxfp4model-path alias for/mnt/lustre01/models/DeepSeek-V4-Pro.deepseek-v4-protoDeepSeek-V4-Pro-NVFP4mapping for MTP2, agentic, and every other configuration.Validation
bash -n runners/launch_gb200-nv.shperf-changelog.yaml,configs/nvidia-master.yaml, and all DeepSeek-V4 vLLM recipe YAML files (29 files total).python -m pytest utils/matrix_logic/ -q— 198 passed.utils/validate_perf_changelog.pypassed against PR 2139's current main merge-base.Parent PR state
PR 2139 currently has a pre-existing merge conflict with
main, so validation directly against the latestorigin/mainremains blocked until the parent branch is synchronized. This child PR intentionally does not rebase or otherwise modify the parent branch.