Repository navigation
[feat] add H3 video config on H200 / [feat] 在 H200 上新增 H3 视频配置 - #4970
edwingao28 wants to merge 3 commits into
Conversation
Mirror the SGLang H200 H3 deployment (h3-v2.yaml release-v1, C1) with vLLM-Omni on the same GPU counts, so both engines bind one plan hash. Co-authored-by: Perplexity <noreply@perplexity.ai>
|
Thanks for the contribution!
中文感谢你的贡献!
|
Co-authored-by: Perplexity <noreply@perplexity.ai>
There was a problem hiding this comment.
Apart from the inline finding, I checked a few other things and found no problems. The changelog says the prompt-embedding cache is off, and that's accurate: the H200 recipe already sets OMNI_DIFFUSION_PROMPT_EMBED_CACHE: "0" in a line this PR doesn't touch. The revision bump to 5917730a… only affects this H200 recipe, which no config key used until now. The B200 recipe already uses that revision. Apart from gpu_type, the updated H200 recipe is now byte-identical to the B200 one, {gpus_per_node} substitution included.
Extended reasoning...
The PR adds one config key, h3-bf16-h200-vllm-omni-video, and makes the old, unused H200 vLLM-Omni recipe match the B200 recipe. It also adds a changelog block and updates three docs. No security-sensitive code is involved. CODEOWNERS covers nvidia-master.yaml, the h200-* recipe tree, MODELS.md and video-generation/, so owners need to review it. Per the repo's rules, the new benchmark also needs an accepted sweep before merge.
This review covers commit a30585e, which is no longer the latest commit on this pull request; later commits are not covered by it.
|
View unofficial run (performance): https://inferencex.semianalysis.com/inference?unofficialRun=42763517224 View unofficial run (accuracy): https://inferencex.semianalysis.com/evaluation?unofficialRun=42763517224 |
|
/use 42763517224 |
|
@edwingao28 staged run 42763517224. This run has no staged performance curves. This run remains available across future |
Summary
H3 video results on H200 come from one engine, SGLang. This adds vLLM-Omni on the same H200 workload so the two can be compared, as #4874 did on B200:
h3-bf16-h200-vllm-omni-videoruns vLLM-Omni v0.32.0, pinned by its linux/amd64 digest, on 1, 2, 4 and 8 H200 GPUs with the spec and set the SGLang H200 key uses (h3-v2.yamlrelease-v1, C1). The bound plan hash equals SGLang's (a5c800f1…) on every point.Changes:
h3/vllm-omni/h200-bf16/serving-client.yaml: it existed from [VideoGenX] run native video sweeps from YAML recipes / [VideoGenX] 使用 YAML 配方运行原生视频扫描 #2916 with no config key and was stale (hard-coded 4 GPUs, the v0.30.0 engine revision, no AdaLN setting, 900 s readiness). It now matches the B200 recipe: Ulysses across all GPUs (--usp N,--tensor-parallel-size 1), text-encoder TP and tiled VAE patch parallelism of the same degree, and the v0.32.0 revision.OMNI_DIFFUSION_PROMPT_EMBED_CACHE=0), cross-request AdaLN projection reuse off (minimax_h3_adaln_cache: false), and the runner's x264 options reach the encoder throughvideo_codec_options.perf-changelog.yamlblock for the new key; MODELS.md, video-generation.md and SUBMITTING.md list the row.Risks: the 1-GPU point runs without offload on 141 GB, as SGLang's does; the sweep is the first vLLM-Omni run on H200.
Delivery: needs the PR full sweep on H200 (4 points).
中文
概要
H200 上的 H3 视频结果目前只有 SGLang 一个引擎。本 PR 与 #4874 在 B200 上的做法相同,在同一 H200 负载上新增 vLLM-Omni 以便比较:
h3-bf16-h200-vllm-omni-video在 1、2、4、8 张 H200 上运行按 linux/amd64 digest 固定的 vLLM-Omni v0.32.0,spec 和 set 与 SGLang H200 key 相同(h3-v2.yamlrelease-v1,C1)。每个点位绑定的 plan hash 都与 SGLang 相同(a5c800f1…)。改动:
h3/vllm-omni/h200-bf16/serving-client.yaml:该 recipe 来自 [VideoGenX] run native video sweeps from YAML recipes / [VideoGenX] 使用 YAML 配方运行原生视频扫描 #2916,没有对应的 config key 且已过时(固定 4 GPU、v0.30.0 引擎版本、缺少 AdaLN 设置、900 秒就绪超时)。现与 B200 recipe 对齐:Ulysses 覆盖全部 GPU,text-encoder TP 和 tiled VAE patch 并行度相同,并使用 v0.32.0 版本。OMNI_DIFFUSION_PROMPT_EMBED_CACHE=0),关闭跨请求 AdaLN 投影复用(minimax_h3_adaln_cache: false),runner 的 x264 参数通过video_codec_options传给编码器。perf-changelog.yaml中为新 key 新增一个条目;MODELS.md、video-generation.md 和 SUBMITTING.md 已列出该行。风险: 与 SGLang 一样,单卡点位在 141 GB 上不使用 offload;本次 sweep 是 vLLM-Omni 在 H200 上的首次运行。
交付: 需要在 H200 上运行 PR full sweep(4 个点位)。
AI model disclosure
Prepared by an AI agent running in Perplexity Computer (orchestrator model: Claude Opus 5.5, as exposed by the runtime); the delegated agent that wrote the recipes, config and this description runs in the same session, and its exact model identifier could not be verified.
Created with Perplexity