Skip to content

knowledge(vllm-omni): 1 change(s) from intake #409

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
tzhouam merged 1 commit into from
Oct 24, 2026
Merged

knowledge(vllm-omni): 1 change(s) from intake #409

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
11 changes: 10 additions & 1 deletion knowledge/repos/vllm-omni/models/qwen-image/rules.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,7 +4,7 @@ created: 2026-09-02
updated: 2026-10-24
type: rule
tags: [vllm-omni, models, diffusion]
sources: ["PR #5887", tests/e2e/accuracy/test_qwen_image.py, "PR #6110", vllm_omni/diffusion/models/qwen_image/qwen_image_transformer.py, "PR #5586", "PR #7513", "PR #8699", "PR #9272", "PR #9268"]
sources: ["PR #5887", tests/e2e/accuracy/test_qwen_image.py, "PR #6110", vllm_omni/diffusion/models/qwen_image/qwen_image_transformer.py, "PR #5586", "PR #7513", "PR #8699", "PR #9272", "PR #9268", "PR #8868"]
confidence: high
---

Expand Down Expand Up @@ -80,3 +80,12 @@ confidence: high
- 验收：`test_amd_pipeline.py` 断言 grade/pool/mirror/timeout、两个 env 值、`full_model and diffusion and rocm and MI325 and cards_1` 标记、节点守卫文案、`setup_reference.sh` 与 BUILDKITE_COMMIT 读取，并断言 "Qwen-Image 2.1 Accuracy" 不出现在 ready/merge pipeline；`test_vae.py` 以 monkeypatch 分别证明 determinism scope 仅 CUDA 生效且退出后还原、MIOpen 路径保持默认策略。gfx942 IT2I XFAIL（SSIM 0.9834/PSNR 32.30）与 T2I 通过（0.9968/44.63）是 MI300 nightly 的历史证据，不构成对 MIOpen 非确定性的定量刻画。^[PR #9268]

<!-- kb:rule status=active since=v0.30.0 -->

## VLLM-OMNI-PR8868-QWEN21-VAE-TILE-STARTS — VAE tiled decode 的 tile 起点必须均匀铺开且尾 tile 贴边结束

- 触发：修改 Qwen-Image-2.1 VAE 的 tiling/tile 布局——`vllm_omni/diffusion/models/qwen_image_21/autoencoder_kl_qwenimage21.py` 的 `_tile_starts`、tiled decode 循环或 `tile_sample_min`/`tile_sample_stride` 语义，或 `vllm_omni/diffusion/distributed/autoencoders/autoencoder_kl_qwenimage21.py` 的 `tile_split`/`tile_merge` 与 `tile_spec` 中的 `starts_*`/`overlap_*`/`cells_per_latent` 字段。
- 强制：轴向 tile 起点必须统一由 `_tile_starts(size, tile_min, stride)` 一类布局产生，并保持不变量：首起点为 0、严格递增、尾 tile 恰好在边界结束（`starts[-1] == size - tile_min`）、相邻间距落在 `[1, stride]` 且整数网格内尽量均匀（max-min ≤ 1）、轴尺寸 ≤ tile_min 时单 tile 覆盖；每个 tile 都保持完整 decode window。分布式 `tile_split` 的 overlap 必须由相邻起点的实际差值乘 `cells_per_latent`（`spatial_compression_ratio // (patch_size or 1)`，unpatchify 前的折算）推导并写入 `tile_spec`，`tile_merge` 从 `tile_spec` 读回同一组 starts/overlap，grid 形状等于两轴起点数、坐标使用起点索引。

@chatgpt-codex-connector chatgpt-codex-connector Bot Oct 24, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

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

P2 Badge Derive overlap from tile width minus the start spacing

When adjacent starts are not half a tile apart, the prescribed calculation produces the tile spacing rather than the overlap. For example, with tile_min=8 and starts [0,5,10], adjacent tiles overlap by 3 latent cells, not 5. The overlap should be (tile_min - (starts[i] - starts[i-1])) * cells_per_latent. Following the current mandatory rule would give distributed blending/merging incorrect overlap widths. The acceptance example [0,4,8] masks the error because its spacing and overlap happen to be equal.

Useful? React with 👍 / 👎.

- 禁止：回到 `range(0, size, stride)` 固定步长铺 tile 使尾 tile 被图像边界截短——被截断的尾 tile 以更小窗口参与 causal conv 解码，与邻 tile 产生系统性亮度偏移，blend 掩盖不了，接缝位置恰为尾 tile 起点 + blend 宽度（#8850：tile 512/stride 384 时，高 1024、宽 1024/1408/1440/1792 出现约 8-11 px、+6/+7 luma 的固定位置竖条纹，与 prompt/seed 无关；stride 整除或尾 tile 恰为完整窗口时不可见）。不得把关闭 tiling 当作修复方向（4096x2048 untiled 时 VAE 每 rank 峰值约 74.5 GB，tiled 约 24 GiB，#8850）。
- 验收：布局单测对 1..256 等全尺寸范围断言上述不变量（含单 tile 分支与间距均匀性）；以非 stride 整除的 latent 尺寸（如 tile 8、stride 6 下的 16x22、22x16）断言所有解码 tile 窗口等于完整 tile 尺寸；分布式路径回归 tile_split/tile_merge 与整体 tiled_decode 结果一致（16x16 latent、8x8 tile、stride 6 时起点为 [0,4,8]，9 个 task、3x3 grid）；端到端可用列中值相对两侧 6..20 像素邻域中值偏差 ≥3 的检测器确认无接缝。^[PR #8868]

<!-- kb:rule status=active since=v0.30.0 -->
Loading