Skip to content

[VideoGenX] sweep Wan2.2 B200 on 1, 2, 4 and 8 GPUs / [VideoGenX] 在 B200 上以 1、2、4、8 卡扫描 Wan2.2 - #5084

Merged
edwingao28 merged 2 commits into
mainfrom
feat/wan22-b200-gpu-sweep
Nov 15, 2026
Merged

edwingao28 merged 2 commits into
mainfrom
feat/wan22-b200-gpu-sweep

Conversation

@edwingao28

@edwingao28 edwingao28 commented Nov 15, 2026 •

Copy link
Copy Markdown
Collaborator

What changed

wan22-bf16-b200-vllm-omni-video only swept 8 GPUs, so the VideoGenX dashboard shows a single Wan2.2 B200 point instead of a GPU-count curve like the H3 B200 configs (1/2/4/8). This adds 1, 2 and 4 GPU entries, each with its own recipe, because the existing recipe hardcodes CFG 2 × Ulysses 4:

GPUs Recipe CFG parallel Ulysses VAE patch parallel
1 serving-client-g1.yaml 1 1 off
2 serving-client-g2.yaml 2 1 2
4 serving-client-g4.yaml 2 2 4
8 serving-client.yaml (unchanged) 2 4 8

Everything else (image digest, wan22.yaml release-v1 workload, prompt-embedding cache off, tiled VAE) matches the 8-GPU recipe. Video fragments do not accept override_* variant bundles, hence separate files.

Verification

  • python -m infx.matrix.plan against main yields 4 video jobs (g1, g2, g4, g8) with the expected recipes.
  • pytest infx/tests/matrix: passing.
  • GPU sweep: pending on this PR.

Add 1, 2 and 4 GPU entries to wan22-bf16-b200-vllm-omni-video, each with
its own recipe: CFG 1 x Ulysses 1 on one GPU, CFG 2 x Ulysses 1 on two and
CFG 2 x Ulysses 2 on four. The 8-GPU recipe is unchanged.

Co-authored-by: Perplexity <noreply@perplexity.ai>
@edwingao28
edwingao28 requested a review from a team November 15, 2026 02:53
@edwingao28 edwingao28 added the full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended) label Nov 15, 2026
@github-actions

Copy link
Copy Markdown
Contributor

Thanks for the contribution!

  • Review: If this PR changes files owned by someone other than a repository admin or @SemiAnalysisAI/core, ask one eligible CODEOWNER to complete the latest PR_REVIEW_CHECKLIST.md before contacting a core maintainer on Slack. Follow the template exactly, including As a PR reviewer and CODEOWNER, I have reviewed this and have, so sign-off verification triggers.
  • PR verification: Sweeps run only when the PR appends an inferencex-e2e/perf-changelog.yaml entry and carries exactly one primary label: full-sweep-fail-fast (strongly recommended; canary plus per-matrix fail-fast), full-sweep-enabled (canary; matrix jobs continue after a failure), or non-canary-full-sweep-enabled (no canary or fail-fast). The modifiers all-evals, evals-only, and agentx-fast require a primary label. On fork PRs, a maintainer applies the label. See sweep labels and reuse.
  • After merging: PR authors must ensure all GitHub Actions jobs pass. Transient failures often pass on rerun; see how to rerun failed jobs.
中文

感谢你的贡献!

  • **审阅:**如果 PR 修改的文件归属于仓库管理员及 @SemiAnalysisAI/core 之外的 CODEOWNER,请先联系一位有资格的 CODEOWNER 填写最新的 PR_REVIEW_CHECKLIST.md,再通过 Slack 联系核心维护者。必须严格遵循模板,并保留 As a PR reviewer and CODEOWNER, I have reviewed this and have,才能触发签核验证。
  • **PR 验证:**只有当 PR 在 inferencex-e2e/perf-changelog.yaml 末尾追加了条目,并且恰好带有一个主标签时,才会运行扫描:full-sweep-fail-fast(强烈推荐;canary 加逐矩阵 fail-fast)、full-sweep-enabled(有 canary;矩阵任务在失败后继续运行)或 non-canary-full-sweep-enabled(无 canary,也无 fail-fast)。修饰标签 all-evals、evals-only 和 agentx-fast 必须与主标签一起使用。fork PR 的标签由维护者添加。参见扫描标签与复用。
  • **合并后:**PR 作者必须确保所有 GitHub Actions 任务通过。临时性失败通常可以通过重新运行恢复;参见重新运行失败任务的说明。

Co-authored-by: Perplexity <noreply@perplexity.ai>

@claude claude Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Nothing blocking. The comments below are optional suggestions. There is no need to push a fix for them before merging.

Comment thread inferencex-e2e/perf-changelog.yaml Outdated
description:
- "Sweep the B200 Wan2.2-T2V-A14B vLLM-Omni baseline on 1, 2, 4 and 8 GPUs at C1 on the same wan22.yaml release-v1 workload, so the dashboard shows a GPU-count curve."
- "Deployment per count: 1 GPU runs both CFG branches with Ulysses 1 and no VAE patch parallelism; 2 GPUs CFG parallel 2 x Ulysses 1; 4 GPUs CFG parallel 2 x Ulysses 2; 8 GPUs keeps CFG parallel 2 x Ulysses 4. Tiled VAE patch parallelism spans every GPU from 2 GPUs up."
pr-link: TBD

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🟡 (optional) The changelog check on this PR rejects the new entry, so CI fails and the Wan2.2 sweep never gets collected. The new block's pr-link: TBD is not an accepted placeholder: on PR runs validate_added_pr_link accepts only the canonical PR URL or an XXX placeholder. Fix: set the pr-link to https://github.com/SemiAnalysisAI/InferenceX/pull/<this PR number>, or to XXX until the number is known.

Why this was flagged

The new entry at inferencex-e2e/perf-changelog.yaml:12900 has pr-link: TBD. On a PR run, compare_entries passes each appended entry to validate_added_pr_link (inferencex-e2e/infx/workflows/validate_perf_changelog.py:158-159). That function raises ChangelogValidationError when the link is neither in PR_LINK_PLACEHOLDERS ("XXX" or https://github.com/SemiAnalysisAI/InferenceX/pull/XXX, line 21-24) nor the expected canonical URL (line 112-116). The error message is "new PR entry must use ... or an XXX placeholder; found 'TBD'". The changelog validation therefore fails for this PR. Without a valid entry, the 1/2/4/8-GPU sweep this PR exists to trigger is never collected or ingested. Every earlier entry in the file uses a canonical link, so this is the only TBD in the file.

Verification: nit. Trigger: someone tries to merge this PR through the documented merge_with_reuse path before pr-link: TBD is replaced. PR CI does not reject TBD. merge_with_reuse.py:676 calls canonicalize_changelog(pr, repo). That goes through canonicalize_appended_links to compare_entries(base, head, pr_number) (prepare_perf_changelog_merge.py:88), which raises on 'TBD'.

@github-actions

Copy link
Copy Markdown
Contributor

@edwingao28

edwingao28 commented Nov 15, 2026 •

Copy link
Copy Markdown
Collaborator Author

/use 43669876970

@edwingao28
edwingao28 merged commit 879bb37 into main Nov 15, 2026
40 checks passed
@edwingao28
edwingao28 deleted the feat/wan22-b200-gpu-sweep branch November 15, 2026 04:38
xinli-sw added a commit that referenced this pull request Nov 15, 2026
…ecode

Resolve the perf-changelog.yaml append conflict by keeping main's new
entries (#5081, #5084) and moving this PR's entry to the end, so the
changelog check sees additions only.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fridge003 added a commit that referenced this pull request Nov 16, 2026
…ernel-c1

Resolve the perf-changelog.yaml tail conflict: keep main's new entries
(#5043, #5081, #5084, #5082, #5089) byte for byte and append this PR's
Muse Spark GB300 megakernel entry after them.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

full-sweep-fail-fast Full sweep with canary gate; first failure cancels the rest of that matrix (recommended)

Projects

Status: Done

Development

Successfully merging this pull request may close these issues.

1 participant