Skip to content

[skip ci][Doc] Module design doc for diffusion runtime - #6440

Merged
hsliuustc0106 merged 9 commits into
vllm-project:mainfrom
fhfuih:diffusion-workrun-doc
Sep 7, 2026
Merged

hsliuustc0106 merged 9 commits into
vllm-project:mainfrom
fhfuih:diffusion-workrun-doc

Conversation

@fhfuih

@fhfuih fhfuih commented Aug 21, 2026

Copy link
Copy Markdown
Contributor

Purpose

Document diffusion runtime (from engine to model runner) in the new design module doc.

This PR does not change the draft status of this doc, so it can still receive other's completion before published.

Test Plan

NA. Doc only.

Copilot AI lite review requested due to automatic review settings August 21, 2026 06:18

Copilot AI 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.

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

@chatgpt-codex-connector

Copy link
Copy Markdown

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.
Credits must be used to enable repository wide code reviews.

@fhfuih

fhfuih commented Aug 21, 2026

Copy link
Copy Markdown
Contributor Author

CC @hsliuustc0106

@vllm-omni-review-bot

Copy link
Copy Markdown

This PR appears to belong to: docs/design/module/diffusion/diffusion_runtime.md.

Module owners: @Isotr0py @princepride @SamitHuang @fhfuih

@fhfuih, please review your own changes and leave a short self-review comment describing what you checked. PRs without author self-review may not be assigned a reviewer.

Please take a look when you have a chance. If you would like an automated review, mention @vllm-omni-review-bot in a comment.

Comment thread docs/design/module/diffusion/diffusion_runtime.md
Comment thread docs/design/module/diffusion/diffusion_runtime.md Outdated
@hsliuustc0106 hsliuustc0106 added the high priority high priority issue, needs to be done asap label Aug 21, 2026
@fhfuih
fhfuih force-pushed the diffusion-workrun-doc branch from a67ad76 to f7e9654 Compare August 25, 2026 09:34

@SamitHuang SamitHuang left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

Checked this draft against current vllm_omni/diffusion on main (UniProcDiffusionExecutor, scheduler, multiproc DLO dispatch, abort/cleanup). Three accuracy issues should be fixed before this is treated as the runtime source of truth: last_verified_commit points at a doc SHA from this PR, the DLO+AllGather exception is described as scheduler policy, and PREEMPTED is drawn as a normal engine path though only tests call preempt_request().

- diffusers.DiffusionPipeline
last_reviewed: 2026-07-16
last_reviewed: 2026-08-25
last_verified_commit: f7e96541dd11c21dce50fdcd9fc5b1269353ee2a

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

last_verified_commit is f7e96541, a documentation commit on this PR (and not even HEAD after 958f39c). Sibling module docs such as engine_orchestration.md point at the code tree that was checked. Point this at the diffusion runtime SHA you actually verified (UniProcDiffusionExecutor, dummy-run skip, #6413 cleanup), then update it whenever those sections change.

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

[P2] 19da238 still does not contain #6439 or #6580, while lines 351-357 describe both cleanup behaviors as current. Please either merge/rebase this document after those dependencies and set last_verified_commit to a tree that actually contains them, or mark those paragraphs as pending.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

[P2] 19da238 still does not contain #6439 or #6580, while lines 351-357 describe both cleanup behaviors as current. Please either merge/rebase this document after those dependencies and set last_verified_commit to a tree that actually contains them, or mark those paragraphs as pending.

I rephrased relevant paragraphs in the current doc and remove clearly affirmative tone that we does it. If this is to be merged before those two PRs, I think the current tone is more suitable (and you can open a new comment if you disagree, lest I miss the thread here)

If you plan to merge it after those two PRs, signal me to modify these paragraphs later if you feel needed

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

Answering the open question above: the requirement tone is fine to keep, but at current main nothing releases this bookkeeping at abort/consumer-drop time — release happens when the late OUTPUT_READY reaches the pump, and an unconsumed result is parked in _completed_outputs by design (#6023). The orphan-after-abort case is #6413, and #6253 / #6439 / #6580 are all still open. Suggest keeping the sentence but tagging it "(pending #6253/#6439/#6580)" so it does not read as current behavior.

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

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

Suggest keeping the sentence but tagging it "(pending #6253)" so it does not read as current behavior.

Done

Comment thread docs/design/module/diffusion/diffusion_runtime.md Outdated
Comment thread docs/design/module/diffusion/diffusion_runtime.md Outdated
@fhfuih
fhfuih force-pushed the diffusion-workrun-doc branch from 958f39c to d058fb1 Compare August 27, 2026 07:27
@fhfuih

fhfuih commented Aug 27, 2026

Copy link
Copy Markdown
Contributor Author

Checked this draft against current vllm_omni/diffusion on main (UniProcDiffusionExecutor, scheduler, multiproc DLO dispatch, abort/cleanup). Three accuracy issues should be fixed before this is treated as the runtime source of truth: last_verified_commit points at a doc SHA from this PR, the DLO+AllGather exception is described as scheduler policy, and PREEMPTED is drawn as a normal engine path though only tests call preempt_request().

Thanks for the careful review. Resolved all of them in the latest commit. Also rebased onto latest main

Comment thread docs/design/module/diffusion/diffusion_runtime.md Outdated
@fhfuih
fhfuih force-pushed the diffusion-workrun-doc branch from d058fb1 to 62f3c98 Compare August 31, 2026 06:33

@SamitHuang SamitHuang left a comment

Copy link
Copy Markdown
Collaborator

Choose a reason for hiding this comment

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

Re-checked the Aug 31 draft against current vllm_omni/diffusion on main (DiffusionExecutor.get_class, UniProcDiffusionExecutor, scheduler, DLO dispatch, abort/cleanup). PREEMPTED, COMPUTE_DONE ordering, and the executor-side DLO routing note from the last pass are fixed. Remaining accuracy gaps: last_verified_commit is 69 commits behind a tree that already changed runtime files, DLO concurrency is still under-specified, dummy-run skip is vaguer than the predicate in code, and INV-002 does not admit the per-rank request pick.

Comment thread docs/design/module/diffusion/diffusion_runtime.md Outdated
Comment thread docs/design/module/diffusion/diffusion_runtime.md Outdated
Comment thread docs/design/module/diffusion/diffusion_runtime.md
Comment thread docs/design/module/diffusion/diffusion_runtime.md Outdated
@hsliuustc0106 hsliuustc0106 added documentation Improvements or additions to documentation diffusion codes related to diffusion models labels Sep 1, 2026
Comment thread docs/design/module/diffusion/diffusion_runtime.md Outdated
Comment thread docs/design/module/diffusion/diffusion_runtime.md Outdated
Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Expand DLO DP concurrency (capacity, coalescing, homogeneity, rank pick),
name max_num_seqs as the scheduler capacity knob, document the exact
dummy-run skip predicate, and carve the per-rank AllGather dispatch out of
INV-002. Also restore the dropped Debugging map and Known limitations
sections.

Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Document diffusion_kv_metadata on NewRequestData, correct multiproc
shutdown so completed async outputs drop with the executor, and tag
abort/consumer-drop async-output reclaim as pending vllm-project#6253/vllm-project#6439/vllm-project#6580.

Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
@fhfuih
fhfuih force-pushed the diffusion-workrun-doc branch from 62f3c98 to c69b66f Compare September 3, 2026 09:59
@fhfuih

fhfuih commented Sep 3, 2026

Copy link
Copy Markdown
Contributor Author

@SamitHuang @hsliuustc0106 Resolved your recent reviews about DLO, new KV connector's metadata, and cleanup behaviors

Seems nothing to add to diffusion_runtime.md from today’s GMT+8 main commits till now.

Cleanup PRs (6253 / 6439 / 6580) are still open — no change to the pending tags.

@hsliuustc0106
hsliuustc0106 merged commit 4b96c57 into vllm-project:main Sep 7, 2026
3 checks passed
ZhengWG pushed a commit to ZhengWG/vllm-omni that referenced this pull request Sep 8, 2026
…6440)

Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Co-authored-by: Alicia <115451386+congw729@users.noreply.github.com>
Signed-off-by: ZhengWG <zwg0606@gmail.com>
haic0 pushed a commit to haic0/vllm-omni that referenced this pull request Sep 9, 2026
…6440)

Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Co-authored-by: Alicia <115451386+congw729@users.noreply.github.com>
wenj-yan pushed a commit to wenj-yan/vllm-omni that referenced this pull request Sep 9, 2026
…6440)

Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Co-authored-by: Alicia <115451386+congw729@users.noreply.github.com>
Signed-off-by: wenjie.yan <wenjyan@outlook.com>
khairulkabir1661 pushed a commit to khairulkabir1661/vllm-omni that referenced this pull request Sep 25, 2026
…6440)

Signed-off-by: Huang, Zeyu <11222265+fhfuih@users.noreply.github.com>
Co-authored-by: Alicia <115451386+congw729@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

diffusion codes related to diffusion models documentation Improvements or additions to documentation high priority high priority issue, needs to be done asap

Projects

None yet

Development

Successfully merging this pull request may close these issues.

6 participants